Version:  2.0.40 2.2.26 2.4.37 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2 4.3 4.4 4.5 4.6 4.7

Linux/mm/page_alloc.c

  1 /*
  2  *  linux/mm/page_alloc.c
  3  *
  4  *  Manages the free list, the system allocates free pages here.
  5  *  Note that kmalloc() lives in slab.c
  6  *
  7  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  8  *  Swap reorganised 29.12.95, Stephen Tweedie
  9  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
 10  *  Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
 11  *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
 12  *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
 13  *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
 14  *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
 15  */
 16 
 17 #include <linux/stddef.h>
 18 #include <linux/mm.h>
 19 #include <linux/swap.h>
 20 #include <linux/interrupt.h>
 21 #include <linux/pagemap.h>
 22 #include <linux/jiffies.h>
 23 #include <linux/bootmem.h>
 24 #include <linux/memblock.h>
 25 #include <linux/compiler.h>
 26 #include <linux/kernel.h>
 27 #include <linux/kmemcheck.h>
 28 #include <linux/kasan.h>
 29 #include <linux/module.h>
 30 #include <linux/suspend.h>
 31 #include <linux/pagevec.h>
 32 #include <linux/blkdev.h>
 33 #include <linux/slab.h>
 34 #include <linux/ratelimit.h>
 35 #include <linux/oom.h>
 36 #include <linux/notifier.h>
 37 #include <linux/topology.h>
 38 #include <linux/sysctl.h>
 39 #include <linux/cpu.h>
 40 #include <linux/cpuset.h>
 41 #include <linux/memory_hotplug.h>
 42 #include <linux/nodemask.h>
 43 #include <linux/vmalloc.h>
 44 #include <linux/vmstat.h>
 45 #include <linux/mempolicy.h>
 46 #include <linux/memremap.h>
 47 #include <linux/stop_machine.h>
 48 #include <linux/sort.h>
 49 #include <linux/pfn.h>
 50 #include <linux/backing-dev.h>
 51 #include <linux/fault-inject.h>
 52 #include <linux/page-isolation.h>
 53 #include <linux/page_ext.h>
 54 #include <linux/debugobjects.h>
 55 #include <linux/kmemleak.h>
 56 #include <linux/compaction.h>
 57 #include <trace/events/kmem.h>
 58 #include <linux/prefetch.h>
 59 #include <linux/mm_inline.h>
 60 #include <linux/migrate.h>
 61 #include <linux/page_ext.h>
 62 #include <linux/hugetlb.h>
 63 #include <linux/sched/rt.h>
 64 #include <linux/page_owner.h>
 65 #include <linux/kthread.h>
 66 
 67 #include <asm/sections.h>
 68 #include <asm/tlbflush.h>
 69 #include <asm/div64.h>
 70 #include "internal.h"
 71 
 72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
 73 static DEFINE_MUTEX(pcp_batch_high_lock);
 74 #define MIN_PERCPU_PAGELIST_FRACTION    (8)
 75 
 76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
 77 DEFINE_PER_CPU(int, numa_node);
 78 EXPORT_PER_CPU_SYMBOL(numa_node);
 79 #endif
 80 
 81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
 82 /*
 83  * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
 84  * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
 85  * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
 86  * defined in <linux/topology.h>.
 87  */
 88 DEFINE_PER_CPU(int, _numa_mem_);                /* Kernel "local memory" node */
 89 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
 90 int _node_numa_mem_[MAX_NUMNODES];
 91 #endif
 92 
 93 /*
 94  * Array of node states.
 95  */
 96 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
 97         [N_POSSIBLE] = NODE_MASK_ALL,
 98         [N_ONLINE] = { { [0] = 1UL } },
 99 #ifndef CONFIG_NUMA
100         [N_NORMAL_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102         [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #endif
104 #ifdef CONFIG_MOVABLE_NODE
105         [N_MEMORY] = { { [0] = 1UL } },
106 #endif
107         [N_CPU] = { { [0] = 1UL } },
108 #endif  /* NUMA */
109 };
110 EXPORT_SYMBOL(node_states);
111 
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock);
114 
115 unsigned long totalram_pages __read_mostly;
116 unsigned long totalreserve_pages __read_mostly;
117 unsigned long totalcma_pages __read_mostly;
118 
119 int percpu_pagelist_fraction;
120 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
121 
122 /*
123  * A cached value of the page's pageblock's migratetype, used when the page is
124  * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125  * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126  * Also the migratetype set in the page does not necessarily match the pcplist
127  * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128  * other index - this ensures that it will be put on the correct CMA freelist.
129  */
130 static inline int get_pcppage_migratetype(struct page *page)
131 {
132         return page->index;
133 }
134 
135 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
136 {
137         page->index = migratetype;
138 }
139 
140 #ifdef CONFIG_PM_SLEEP
141 /*
142  * The following functions are used by the suspend/hibernate code to temporarily
143  * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144  * while devices are suspended.  To avoid races with the suspend/hibernate code,
145  * they should always be called with pm_mutex held (gfp_allowed_mask also should
146  * only be modified with pm_mutex held, unless the suspend/hibernate code is
147  * guaranteed not to run in parallel with that modification).
148  */
149 
150 static gfp_t saved_gfp_mask;
151 
152 void pm_restore_gfp_mask(void)
153 {
154         WARN_ON(!mutex_is_locked(&pm_mutex));
155         if (saved_gfp_mask) {
156                 gfp_allowed_mask = saved_gfp_mask;
157                 saved_gfp_mask = 0;
158         }
159 }
160 
161 void pm_restrict_gfp_mask(void)
162 {
163         WARN_ON(!mutex_is_locked(&pm_mutex));
164         WARN_ON(saved_gfp_mask);
165         saved_gfp_mask = gfp_allowed_mask;
166         gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
167 }
168 
169 bool pm_suspended_storage(void)
170 {
171         if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
172                 return false;
173         return true;
174 }
175 #endif /* CONFIG_PM_SLEEP */
176 
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly;
179 #endif
180 
181 static void __free_pages_ok(struct page *page, unsigned int order);
182 
183 /*
184  * results with 256, 32 in the lowmem_reserve sysctl:
185  *      1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186  *      1G machine -> (16M dma, 784M normal, 224M high)
187  *      NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188  *      HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189  *      HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
190  *
191  * TBD: should special case ZONE_DMA32 machines here - in those we normally
192  * don't need any ZONE_NORMAL reservation
193  */
194 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
195 #ifdef CONFIG_ZONE_DMA
196          256,
197 #endif
198 #ifdef CONFIG_ZONE_DMA32
199          256,
200 #endif
201 #ifdef CONFIG_HIGHMEM
202          32,
203 #endif
204          32,
205 };
206 
207 EXPORT_SYMBOL(totalram_pages);
208 
209 static char * const zone_names[MAX_NR_ZONES] = {
210 #ifdef CONFIG_ZONE_DMA
211          "DMA",
212 #endif
213 #ifdef CONFIG_ZONE_DMA32
214          "DMA32",
215 #endif
216          "Normal",
217 #ifdef CONFIG_HIGHMEM
218          "HighMem",
219 #endif
220          "Movable",
221 #ifdef CONFIG_ZONE_DEVICE
222          "Device",
223 #endif
224 };
225 
226 char * const migratetype_names[MIGRATE_TYPES] = {
227         "Unmovable",
228         "Movable",
229         "Reclaimable",
230         "HighAtomic",
231 #ifdef CONFIG_CMA
232         "CMA",
233 #endif
234 #ifdef CONFIG_MEMORY_ISOLATION
235         "Isolate",
236 #endif
237 };
238 
239 compound_page_dtor * const compound_page_dtors[] = {
240         NULL,
241         free_compound_page,
242 #ifdef CONFIG_HUGETLB_PAGE
243         free_huge_page,
244 #endif
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
246         free_transhuge_page,
247 #endif
248 };
249 
250 int min_free_kbytes = 1024;
251 int user_min_free_kbytes = -1;
252 int watermark_scale_factor = 10;
253 
254 static unsigned long __meminitdata nr_kernel_pages;
255 static unsigned long __meminitdata nr_all_pages;
256 static unsigned long __meminitdata dma_reserve;
257 
258 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
259 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
260 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
261 static unsigned long __initdata required_kernelcore;
262 static unsigned long __initdata required_movablecore;
263 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
264 static bool mirrored_kernelcore;
265 
266 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
267 int movable_zone;
268 EXPORT_SYMBOL(movable_zone);
269 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
270 
271 #if MAX_NUMNODES > 1
272 int nr_node_ids __read_mostly = MAX_NUMNODES;
273 int nr_online_nodes __read_mostly = 1;
274 EXPORT_SYMBOL(nr_node_ids);
275 EXPORT_SYMBOL(nr_online_nodes);
276 #endif
277 
278 int page_group_by_mobility_disabled __read_mostly;
279 
280 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
281 static inline void reset_deferred_meminit(pg_data_t *pgdat)
282 {
283         pgdat->first_deferred_pfn = ULONG_MAX;
284 }
285 
286 /* Returns true if the struct page for the pfn is uninitialised */
287 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
288 {
289         int nid = early_pfn_to_nid(pfn);
290 
291         if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
292                 return true;
293 
294         return false;
295 }
296 
297 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
298 {
299         if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
300                 return true;
301 
302         return false;
303 }
304 
305 /*
306  * Returns false when the remaining initialisation should be deferred until
307  * later in the boot cycle when it can be parallelised.
308  */
309 static inline bool update_defer_init(pg_data_t *pgdat,
310                                 unsigned long pfn, unsigned long zone_end,
311                                 unsigned long *nr_initialised)
312 {
313         unsigned long max_initialise;
314 
315         /* Always populate low zones for address-contrained allocations */
316         if (zone_end < pgdat_end_pfn(pgdat))
317                 return true;
318         /*
319          * Initialise at least 2G of a node but also take into account that
320          * two large system hashes that can take up 1GB for 0.25TB/node.
321          */
322         max_initialise = max(2UL << (30 - PAGE_SHIFT),
323                 (pgdat->node_spanned_pages >> 8));
324 
325         (*nr_initialised)++;
326         if ((*nr_initialised > max_initialise) &&
327             (pfn & (PAGES_PER_SECTION - 1)) == 0) {
328                 pgdat->first_deferred_pfn = pfn;
329                 return false;
330         }
331 
332         return true;
333 }
334 #else
335 static inline void reset_deferred_meminit(pg_data_t *pgdat)
336 {
337 }
338 
339 static inline bool early_page_uninitialised(unsigned long pfn)
340 {
341         return false;
342 }
343 
344 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
345 {
346         return false;
347 }
348 
349 static inline bool update_defer_init(pg_data_t *pgdat,
350                                 unsigned long pfn, unsigned long zone_end,
351                                 unsigned long *nr_initialised)
352 {
353         return true;
354 }
355 #endif
356 
357 /* Return a pointer to the bitmap storing bits affecting a block of pages */
358 static inline unsigned long *get_pageblock_bitmap(struct page *page,
359                                                         unsigned long pfn)
360 {
361 #ifdef CONFIG_SPARSEMEM
362         return __pfn_to_section(pfn)->pageblock_flags;
363 #else
364         return page_zone(page)->pageblock_flags;
365 #endif /* CONFIG_SPARSEMEM */
366 }
367 
368 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
369 {
370 #ifdef CONFIG_SPARSEMEM
371         pfn &= (PAGES_PER_SECTION-1);
372         return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
373 #else
374         pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
375         return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
376 #endif /* CONFIG_SPARSEMEM */
377 }
378 
379 /**
380  * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
381  * @page: The page within the block of interest
382  * @pfn: The target page frame number
383  * @end_bitidx: The last bit of interest to retrieve
384  * @mask: mask of bits that the caller is interested in
385  *
386  * Return: pageblock_bits flags
387  */
388 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
389                                         unsigned long pfn,
390                                         unsigned long end_bitidx,
391                                         unsigned long mask)
392 {
393         unsigned long *bitmap;
394         unsigned long bitidx, word_bitidx;
395         unsigned long word;
396 
397         bitmap = get_pageblock_bitmap(page, pfn);
398         bitidx = pfn_to_bitidx(page, pfn);
399         word_bitidx = bitidx / BITS_PER_LONG;
400         bitidx &= (BITS_PER_LONG-1);
401 
402         word = bitmap[word_bitidx];
403         bitidx += end_bitidx;
404         return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
405 }
406 
407 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
408                                         unsigned long end_bitidx,
409                                         unsigned long mask)
410 {
411         return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
412 }
413 
414 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
415 {
416         return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
417 }
418 
419 /**
420  * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
421  * @page: The page within the block of interest
422  * @flags: The flags to set
423  * @pfn: The target page frame number
424  * @end_bitidx: The last bit of interest
425  * @mask: mask of bits that the caller is interested in
426  */
427 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
428                                         unsigned long pfn,
429                                         unsigned long end_bitidx,
430                                         unsigned long mask)
431 {
432         unsigned long *bitmap;
433         unsigned long bitidx, word_bitidx;
434         unsigned long old_word, word;
435 
436         BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
437 
438         bitmap = get_pageblock_bitmap(page, pfn);
439         bitidx = pfn_to_bitidx(page, pfn);
440         word_bitidx = bitidx / BITS_PER_LONG;
441         bitidx &= (BITS_PER_LONG-1);
442 
443         VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
444 
445         bitidx += end_bitidx;
446         mask <<= (BITS_PER_LONG - bitidx - 1);
447         flags <<= (BITS_PER_LONG - bitidx - 1);
448 
449         word = READ_ONCE(bitmap[word_bitidx]);
450         for (;;) {
451                 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
452                 if (word == old_word)
453                         break;
454                 word = old_word;
455         }
456 }
457 
458 void set_pageblock_migratetype(struct page *page, int migratetype)
459 {
460         if (unlikely(page_group_by_mobility_disabled &&
461                      migratetype < MIGRATE_PCPTYPES))
462                 migratetype = MIGRATE_UNMOVABLE;
463 
464         set_pageblock_flags_group(page, (unsigned long)migratetype,
465                                         PB_migrate, PB_migrate_end);
466 }
467 
468 #ifdef CONFIG_DEBUG_VM
469 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
470 {
471         int ret = 0;
472         unsigned seq;
473         unsigned long pfn = page_to_pfn(page);
474         unsigned long sp, start_pfn;
475 
476         do {
477                 seq = zone_span_seqbegin(zone);
478                 start_pfn = zone->zone_start_pfn;
479                 sp = zone->spanned_pages;
480                 if (!zone_spans_pfn(zone, pfn))
481                         ret = 1;
482         } while (zone_span_seqretry(zone, seq));
483 
484         if (ret)
485                 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
486                         pfn, zone_to_nid(zone), zone->name,
487                         start_pfn, start_pfn + sp);
488 
489         return ret;
490 }
491 
492 static int page_is_consistent(struct zone *zone, struct page *page)
493 {
494         if (!pfn_valid_within(page_to_pfn(page)))
495                 return 0;
496         if (zone != page_zone(page))
497                 return 0;
498 
499         return 1;
500 }
501 /*
502  * Temporary debugging check for pages not lying within a given zone.
503  */
504 static int bad_range(struct zone *zone, struct page *page)
505 {
506         if (page_outside_zone_boundaries(zone, page))
507                 return 1;
508         if (!page_is_consistent(zone, page))
509                 return 1;
510 
511         return 0;
512 }
513 #else
514 static inline int bad_range(struct zone *zone, struct page *page)
515 {
516         return 0;
517 }
518 #endif
519 
520 static void bad_page(struct page *page, const char *reason,
521                 unsigned long bad_flags)
522 {
523         static unsigned long resume;
524         static unsigned long nr_shown;
525         static unsigned long nr_unshown;
526 
527         /*
528          * Allow a burst of 60 reports, then keep quiet for that minute;
529          * or allow a steady drip of one report per second.
530          */
531         if (nr_shown == 60) {
532                 if (time_before(jiffies, resume)) {
533                         nr_unshown++;
534                         goto out;
535                 }
536                 if (nr_unshown) {
537                         pr_alert(
538                               "BUG: Bad page state: %lu messages suppressed\n",
539                                 nr_unshown);
540                         nr_unshown = 0;
541                 }
542                 nr_shown = 0;
543         }
544         if (nr_shown++ == 0)
545                 resume = jiffies + 60 * HZ;
546 
547         pr_alert("BUG: Bad page state in process %s  pfn:%05lx\n",
548                 current->comm, page_to_pfn(page));
549         __dump_page(page, reason);
550         bad_flags &= page->flags;
551         if (bad_flags)
552                 pr_alert("bad because of flags: %#lx(%pGp)\n",
553                                                 bad_flags, &bad_flags);
554         dump_page_owner(page);
555 
556         print_modules();
557         dump_stack();
558 out:
559         /* Leave bad fields for debug, except PageBuddy could make trouble */
560         page_mapcount_reset(page); /* remove PageBuddy */
561         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
562 }
563 
564 /*
565  * Higher-order pages are called "compound pages".  They are structured thusly:
566  *
567  * The first PAGE_SIZE page is called the "head page" and have PG_head set.
568  *
569  * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
570  * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
571  *
572  * The first tail page's ->compound_dtor holds the offset in array of compound
573  * page destructors. See compound_page_dtors.
574  *
575  * The first tail page's ->compound_order holds the order of allocation.
576  * This usage means that zero-order pages may not be compound.
577  */
578 
579 void free_compound_page(struct page *page)
580 {
581         __free_pages_ok(page, compound_order(page));
582 }
583 
584 void prep_compound_page(struct page *page, unsigned int order)
585 {
586         int i;
587         int nr_pages = 1 << order;
588 
589         set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
590         set_compound_order(page, order);
591         __SetPageHead(page);
592         for (i = 1; i < nr_pages; i++) {
593                 struct page *p = page + i;
594                 set_page_count(p, 0);
595                 p->mapping = TAIL_MAPPING;
596                 set_compound_head(p, page);
597         }
598         atomic_set(compound_mapcount_ptr(page), -1);
599 }
600 
601 #ifdef CONFIG_DEBUG_PAGEALLOC
602 unsigned int _debug_guardpage_minorder;
603 bool _debug_pagealloc_enabled __read_mostly
604                         = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
605 EXPORT_SYMBOL(_debug_pagealloc_enabled);
606 bool _debug_guardpage_enabled __read_mostly;
607 
608 static int __init early_debug_pagealloc(char *buf)
609 {
610         if (!buf)
611                 return -EINVAL;
612         return kstrtobool(buf, &_debug_pagealloc_enabled);
613 }
614 early_param("debug_pagealloc", early_debug_pagealloc);
615 
616 static bool need_debug_guardpage(void)
617 {
618         /* If we don't use debug_pagealloc, we don't need guard page */
619         if (!debug_pagealloc_enabled())
620                 return false;
621 
622         return true;
623 }
624 
625 static void init_debug_guardpage(void)
626 {
627         if (!debug_pagealloc_enabled())
628                 return;
629 
630         _debug_guardpage_enabled = true;
631 }
632 
633 struct page_ext_operations debug_guardpage_ops = {
634         .need = need_debug_guardpage,
635         .init = init_debug_guardpage,
636 };
637 
638 static int __init debug_guardpage_minorder_setup(char *buf)
639 {
640         unsigned long res;
641 
642         if (kstrtoul(buf, 10, &res) < 0 ||  res > MAX_ORDER / 2) {
643                 pr_err("Bad debug_guardpage_minorder value\n");
644                 return 0;
645         }
646         _debug_guardpage_minorder = res;
647         pr_info("Setting debug_guardpage_minorder to %lu\n", res);
648         return 0;
649 }
650 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
651 
652 static inline void set_page_guard(struct zone *zone, struct page *page,
653                                 unsigned int order, int migratetype)
654 {
655         struct page_ext *page_ext;
656 
657         if (!debug_guardpage_enabled())
658                 return;
659 
660         page_ext = lookup_page_ext(page);
661         if (unlikely(!page_ext))
662                 return;
663 
664         __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
665 
666         INIT_LIST_HEAD(&page->lru);
667         set_page_private(page, order);
668         /* Guard pages are not available for any usage */
669         __mod_zone_freepage_state(zone, -(1 << order), migratetype);
670 }
671 
672 static inline void clear_page_guard(struct zone *zone, struct page *page,
673                                 unsigned int order, int migratetype)
674 {
675         struct page_ext *page_ext;
676 
677         if (!debug_guardpage_enabled())
678                 return;
679 
680         page_ext = lookup_page_ext(page);
681         if (unlikely(!page_ext))
682                 return;
683 
684         __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
685 
686         set_page_private(page, 0);
687         if (!is_migrate_isolate(migratetype))
688                 __mod_zone_freepage_state(zone, (1 << order), migratetype);
689 }
690 #else
691 struct page_ext_operations debug_guardpage_ops = { NULL, };
692 static inline void set_page_guard(struct zone *zone, struct page *page,
693                                 unsigned int order, int migratetype) {}
694 static inline void clear_page_guard(struct zone *zone, struct page *page,
695                                 unsigned int order, int migratetype) {}
696 #endif
697 
698 static inline void set_page_order(struct page *page, unsigned int order)
699 {
700         set_page_private(page, order);
701         __SetPageBuddy(page);
702 }
703 
704 static inline void rmv_page_order(struct page *page)
705 {
706         __ClearPageBuddy(page);
707         set_page_private(page, 0);
708 }
709 
710 /*
711  * This function checks whether a page is free && is the buddy
712  * we can do coalesce a page and its buddy if
713  * (a) the buddy is not in a hole &&
714  * (b) the buddy is in the buddy system &&
715  * (c) a page and its buddy have the same order &&
716  * (d) a page and its buddy are in the same zone.
717  *
718  * For recording whether a page is in the buddy system, we set ->_mapcount
719  * PAGE_BUDDY_MAPCOUNT_VALUE.
720  * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
721  * serialized by zone->lock.
722  *
723  * For recording page's order, we use page_private(page).
724  */
725 static inline int page_is_buddy(struct page *page, struct page *buddy,
726                                                         unsigned int order)
727 {
728         if (!pfn_valid_within(page_to_pfn(buddy)))
729                 return 0;
730 
731         if (page_is_guard(buddy) && page_order(buddy) == order) {
732                 if (page_zone_id(page) != page_zone_id(buddy))
733                         return 0;
734 
735                 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
736 
737                 return 1;
738         }
739 
740         if (PageBuddy(buddy) && page_order(buddy) == order) {
741                 /*
742                  * zone check is done late to avoid uselessly
743                  * calculating zone/node ids for pages that could
744                  * never merge.
745                  */
746                 if (page_zone_id(page) != page_zone_id(buddy))
747                         return 0;
748 
749                 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
750 
751                 return 1;
752         }
753         return 0;
754 }
755 
756 /*
757  * Freeing function for a buddy system allocator.
758  *
759  * The concept of a buddy system is to maintain direct-mapped table
760  * (containing bit values) for memory blocks of various "orders".
761  * The bottom level table contains the map for the smallest allocatable
762  * units of memory (here, pages), and each level above it describes
763  * pairs of units from the levels below, hence, "buddies".
764  * At a high level, all that happens here is marking the table entry
765  * at the bottom level available, and propagating the changes upward
766  * as necessary, plus some accounting needed to play nicely with other
767  * parts of the VM system.
768  * At each level, we keep a list of pages, which are heads of continuous
769  * free pages of length of (1 << order) and marked with _mapcount
770  * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
771  * field.
772  * So when we are allocating or freeing one, we can derive the state of the
773  * other.  That is, if we allocate a small block, and both were
774  * free, the remainder of the region must be split into blocks.
775  * If a block is freed, and its buddy is also free, then this
776  * triggers coalescing into a block of larger size.
777  *
778  * -- nyc
779  */
780 
781 static inline void __free_one_page(struct page *page,
782                 unsigned long pfn,
783                 struct zone *zone, unsigned int order,
784                 int migratetype)
785 {
786         unsigned long page_idx;
787         unsigned long combined_idx;
788         unsigned long uninitialized_var(buddy_idx);
789         struct page *buddy;
790         unsigned int max_order;
791 
792         max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
793 
794         VM_BUG_ON(!zone_is_initialized(zone));
795         VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
796 
797         VM_BUG_ON(migratetype == -1);
798         if (likely(!is_migrate_isolate(migratetype)))
799                 __mod_zone_freepage_state(zone, 1 << order, migratetype);
800 
801         page_idx = pfn & ((1 << MAX_ORDER) - 1);
802 
803         VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
804         VM_BUG_ON_PAGE(bad_range(zone, page), page);
805 
806 continue_merging:
807         while (order < max_order - 1) {
808                 buddy_idx = __find_buddy_index(page_idx, order);
809                 buddy = page + (buddy_idx - page_idx);
810                 if (!page_is_buddy(page, buddy, order))
811                         goto done_merging;
812                 /*
813                  * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
814                  * merge with it and move up one order.
815                  */
816                 if (page_is_guard(buddy)) {
817                         clear_page_guard(zone, buddy, order, migratetype);
818                 } else {
819                         list_del(&buddy->lru);
820                         zone->free_area[order].nr_free--;
821                         rmv_page_order(buddy);
822                 }
823                 combined_idx = buddy_idx & page_idx;
824                 page = page + (combined_idx - page_idx);
825                 page_idx = combined_idx;
826                 order++;
827         }
828         if (max_order < MAX_ORDER) {
829                 /* If we are here, it means order is >= pageblock_order.
830                  * We want to prevent merge between freepages on isolate
831                  * pageblock and normal pageblock. Without this, pageblock
832                  * isolation could cause incorrect freepage or CMA accounting.
833                  *
834                  * We don't want to hit this code for the more frequent
835                  * low-order merging.
836                  */
837                 if (unlikely(has_isolate_pageblock(zone))) {
838                         int buddy_mt;
839 
840                         buddy_idx = __find_buddy_index(page_idx, order);
841                         buddy = page + (buddy_idx - page_idx);
842                         buddy_mt = get_pageblock_migratetype(buddy);
843 
844                         if (migratetype != buddy_mt
845                                         && (is_migrate_isolate(migratetype) ||
846                                                 is_migrate_isolate(buddy_mt)))
847                                 goto done_merging;
848                 }
849                 max_order++;
850                 goto continue_merging;
851         }
852 
853 done_merging:
854         set_page_order(page, order);
855 
856         /*
857          * If this is not the largest possible page, check if the buddy
858          * of the next-highest order is free. If it is, it's possible
859          * that pages are being freed that will coalesce soon. In case,
860          * that is happening, add the free page to the tail of the list
861          * so it's less likely to be used soon and more likely to be merged
862          * as a higher order page
863          */
864         if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
865                 struct page *higher_page, *higher_buddy;
866                 combined_idx = buddy_idx & page_idx;
867                 higher_page = page + (combined_idx - page_idx);
868                 buddy_idx = __find_buddy_index(combined_idx, order + 1);
869                 higher_buddy = higher_page + (buddy_idx - combined_idx);
870                 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
871                         list_add_tail(&page->lru,
872                                 &zone->free_area[order].free_list[migratetype]);
873                         goto out;
874                 }
875         }
876 
877         list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
878 out:
879         zone->free_area[order].nr_free++;
880 }
881 
882 /*
883  * A bad page could be due to a number of fields. Instead of multiple branches,
884  * try and check multiple fields with one check. The caller must do a detailed
885  * check if necessary.
886  */
887 static inline bool page_expected_state(struct page *page,
888                                         unsigned long check_flags)
889 {
890         if (unlikely(atomic_read(&page->_mapcount) != -1))
891                 return false;
892 
893         if (unlikely((unsigned long)page->mapping |
894                         page_ref_count(page) |
895 #ifdef CONFIG_MEMCG
896                         (unsigned long)page->mem_cgroup |
897 #endif
898                         (page->flags & check_flags)))
899                 return false;
900 
901         return true;
902 }
903 
904 static void free_pages_check_bad(struct page *page)
905 {
906         const char *bad_reason;
907         unsigned long bad_flags;
908 
909         bad_reason = NULL;
910         bad_flags = 0;
911 
912         if (unlikely(atomic_read(&page->_mapcount) != -1))
913                 bad_reason = "nonzero mapcount";
914         if (unlikely(page->mapping != NULL))
915                 bad_reason = "non-NULL mapping";
916         if (unlikely(page_ref_count(page) != 0))
917                 bad_reason = "nonzero _refcount";
918         if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
919                 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
920                 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
921         }
922 #ifdef CONFIG_MEMCG
923         if (unlikely(page->mem_cgroup))
924                 bad_reason = "page still charged to cgroup";
925 #endif
926         bad_page(page, bad_reason, bad_flags);
927 }
928 
929 static inline int free_pages_check(struct page *page)
930 {
931         if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
932                 return 0;
933 
934         /* Something has gone sideways, find it */
935         free_pages_check_bad(page);
936         return 1;
937 }
938 
939 static int free_tail_pages_check(struct page *head_page, struct page *page)
940 {
941         int ret = 1;
942 
943         /*
944          * We rely page->lru.next never has bit 0 set, unless the page
945          * is PageTail(). Let's make sure that's true even for poisoned ->lru.
946          */
947         BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
948 
949         if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
950                 ret = 0;
951                 goto out;
952         }
953         switch (page - head_page) {
954         case 1:
955                 /* the first tail page: ->mapping is compound_mapcount() */
956                 if (unlikely(compound_mapcount(page))) {
957                         bad_page(page, "nonzero compound_mapcount", 0);
958                         goto out;
959                 }
960                 break;
961         case 2:
962                 /*
963                  * the second tail page: ->mapping is
964                  * page_deferred_list().next -- ignore value.
965                  */
966                 break;
967         default:
968                 if (page->mapping != TAIL_MAPPING) {
969                         bad_page(page, "corrupted mapping in tail page", 0);
970                         goto out;
971                 }
972                 break;
973         }
974         if (unlikely(!PageTail(page))) {
975                 bad_page(page, "PageTail not set", 0);
976                 goto out;
977         }
978         if (unlikely(compound_head(page) != head_page)) {
979                 bad_page(page, "compound_head not consistent", 0);
980                 goto out;
981         }
982         ret = 0;
983 out:
984         page->mapping = NULL;
985         clear_compound_head(page);
986         return ret;
987 }
988 
989 static __always_inline bool free_pages_prepare(struct page *page,
990                                         unsigned int order, bool check_free)
991 {
992         int bad = 0;
993 
994         VM_BUG_ON_PAGE(PageTail(page), page);
995 
996         trace_mm_page_free(page, order);
997         kmemcheck_free_shadow(page, order);
998 
999         /*
1000          * Check tail pages before head page information is cleared to
1001          * avoid checking PageCompound for order-0 pages.
1002          */
1003         if (unlikely(order)) {
1004                 bool compound = PageCompound(page);
1005                 int i;
1006 
1007                 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1008 
1009                 for (i = 1; i < (1 << order); i++) {
1010                         if (compound)
1011                                 bad += free_tail_pages_check(page, page + i);
1012                         if (unlikely(free_pages_check(page + i))) {
1013                                 bad++;
1014                                 continue;
1015                         }
1016                         (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1017                 }
1018         }
1019         if (PageAnonHead(page))
1020                 page->mapping = NULL;
1021         if (check_free)
1022                 bad += free_pages_check(page);
1023         if (bad)
1024                 return false;
1025 
1026         page_cpupid_reset_last(page);
1027         page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1028         reset_page_owner(page, order);
1029 
1030         if (!PageHighMem(page)) {
1031                 debug_check_no_locks_freed(page_address(page),
1032                                            PAGE_SIZE << order);
1033                 debug_check_no_obj_freed(page_address(page),
1034                                            PAGE_SIZE << order);
1035         }
1036         arch_free_page(page, order);
1037         kernel_poison_pages(page, 1 << order, 0);
1038         kernel_map_pages(page, 1 << order, 0);
1039         kasan_free_pages(page, order);
1040 
1041         return true;
1042 }
1043 
1044 #ifdef CONFIG_DEBUG_VM
1045 static inline bool free_pcp_prepare(struct page *page)
1046 {
1047         return free_pages_prepare(page, 0, true);
1048 }
1049 
1050 static inline bool bulkfree_pcp_prepare(struct page *page)
1051 {
1052         return false;
1053 }
1054 #else
1055 static bool free_pcp_prepare(struct page *page)
1056 {
1057         return free_pages_prepare(page, 0, false);
1058 }
1059 
1060 static bool bulkfree_pcp_prepare(struct page *page)
1061 {
1062         return free_pages_check(page);
1063 }
1064 #endif /* CONFIG_DEBUG_VM */
1065 
1066 /*
1067  * Frees a number of pages from the PCP lists
1068  * Assumes all pages on list are in same zone, and of same order.
1069  * count is the number of pages to free.
1070  *
1071  * If the zone was previously in an "all pages pinned" state then look to
1072  * see if this freeing clears that state.
1073  *
1074  * And clear the zone's pages_scanned counter, to hold off the "all pages are
1075  * pinned" detection logic.
1076  */
1077 static void free_pcppages_bulk(struct zone *zone, int count,
1078                                         struct per_cpu_pages *pcp)
1079 {
1080         int migratetype = 0;
1081         int batch_free = 0;
1082         unsigned long nr_scanned;
1083         bool isolated_pageblocks;
1084 
1085         spin_lock(&zone->lock);
1086         isolated_pageblocks = has_isolate_pageblock(zone);
1087         nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1088         if (nr_scanned)
1089                 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1090 
1091         while (count) {
1092                 struct page *page;
1093                 struct list_head *list;
1094 
1095                 /*
1096                  * Remove pages from lists in a round-robin fashion. A
1097                  * batch_free count is maintained that is incremented when an
1098                  * empty list is encountered.  This is so more pages are freed
1099                  * off fuller lists instead of spinning excessively around empty
1100                  * lists
1101                  */
1102                 do {
1103                         batch_free++;
1104                         if (++migratetype == MIGRATE_PCPTYPES)
1105                                 migratetype = 0;
1106                         list = &pcp->lists[migratetype];
1107                 } while (list_empty(list));
1108 
1109                 /* This is the only non-empty list. Free them all. */
1110                 if (batch_free == MIGRATE_PCPTYPES)
1111                         batch_free = count;
1112 
1113                 do {
1114                         int mt; /* migratetype of the to-be-freed page */
1115 
1116                         page = list_last_entry(list, struct page, lru);
1117                         /* must delete as __free_one_page list manipulates */
1118                         list_del(&page->lru);
1119 
1120                         mt = get_pcppage_migratetype(page);
1121                         /* MIGRATE_ISOLATE page should not go to pcplists */
1122                         VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1123                         /* Pageblock could have been isolated meanwhile */
1124                         if (unlikely(isolated_pageblocks))
1125                                 mt = get_pageblock_migratetype(page);
1126 
1127                         if (bulkfree_pcp_prepare(page))
1128                                 continue;
1129 
1130                         __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1131                         trace_mm_page_pcpu_drain(page, 0, mt);
1132                 } while (--count && --batch_free && !list_empty(list));
1133         }
1134         spin_unlock(&zone->lock);
1135 }
1136 
1137 static void free_one_page(struct zone *zone,
1138                                 struct page *page, unsigned long pfn,
1139                                 unsigned int order,
1140                                 int migratetype)
1141 {
1142         unsigned long nr_scanned;
1143         spin_lock(&zone->lock);
1144         nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
1145         if (nr_scanned)
1146                 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
1147 
1148         if (unlikely(has_isolate_pageblock(zone) ||
1149                 is_migrate_isolate(migratetype))) {
1150                 migratetype = get_pfnblock_migratetype(page, pfn);
1151         }
1152         __free_one_page(page, pfn, zone, order, migratetype);
1153         spin_unlock(&zone->lock);
1154 }
1155 
1156 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1157                                 unsigned long zone, int nid)
1158 {
1159         set_page_links(page, zone, nid, pfn);
1160         init_page_count(page);
1161         page_mapcount_reset(page);
1162         page_cpupid_reset_last(page);
1163 
1164         INIT_LIST_HEAD(&page->lru);
1165 #ifdef WANT_PAGE_VIRTUAL
1166         /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1167         if (!is_highmem_idx(zone))
1168                 set_page_address(page, __va(pfn << PAGE_SHIFT));
1169 #endif
1170 }
1171 
1172 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1173                                         int nid)
1174 {
1175         return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1176 }
1177 
1178 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1179 static void init_reserved_page(unsigned long pfn)
1180 {
1181         pg_data_t *pgdat;
1182         int nid, zid;
1183 
1184         if (!early_page_uninitialised(pfn))
1185                 return;
1186 
1187         nid = early_pfn_to_nid(pfn);
1188         pgdat = NODE_DATA(nid);
1189 
1190         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1191                 struct zone *zone = &pgdat->node_zones[zid];
1192 
1193                 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1194                         break;
1195         }
1196         __init_single_pfn(pfn, zid, nid);
1197 }
1198 #else
1199 static inline void init_reserved_page(unsigned long pfn)
1200 {
1201 }
1202 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1203 
1204 /*
1205  * Initialised pages do not have PageReserved set. This function is
1206  * called for each range allocated by the bootmem allocator and
1207  * marks the pages PageReserved. The remaining valid pages are later
1208  * sent to the buddy page allocator.
1209  */
1210 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1211 {
1212         unsigned long start_pfn = PFN_DOWN(start);
1213         unsigned long end_pfn = PFN_UP(end);
1214 
1215         for (; start_pfn < end_pfn; start_pfn++) {
1216                 if (pfn_valid(start_pfn)) {
1217                         struct page *page = pfn_to_page(start_pfn);
1218 
1219                         init_reserved_page(start_pfn);
1220 
1221                         /* Avoid false-positive PageTail() */
1222                         INIT_LIST_HEAD(&page->lru);
1223 
1224                         SetPageReserved(page);
1225                 }
1226         }
1227 }
1228 
1229 static void __free_pages_ok(struct page *page, unsigned int order)
1230 {
1231         unsigned long flags;
1232         int migratetype;
1233         unsigned long pfn = page_to_pfn(page);
1234 
1235         if (!free_pages_prepare(page, order, true))
1236                 return;
1237 
1238         migratetype = get_pfnblock_migratetype(page, pfn);
1239         local_irq_save(flags);
1240         __count_vm_events(PGFREE, 1 << order);
1241         free_one_page(page_zone(page), page, pfn, order, migratetype);
1242         local_irq_restore(flags);
1243 }
1244 
1245 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1246 {
1247         unsigned int nr_pages = 1 << order;
1248         struct page *p = page;
1249         unsigned int loop;
1250 
1251         prefetchw(p);
1252         for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1253                 prefetchw(p + 1);
1254                 __ClearPageReserved(p);
1255                 set_page_count(p, 0);
1256         }
1257         __ClearPageReserved(p);
1258         set_page_count(p, 0);
1259 
1260         page_zone(page)->managed_pages += nr_pages;
1261         set_page_refcounted(page);
1262         __free_pages(page, order);
1263 }
1264 
1265 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1266         defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1267 
1268 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1269 
1270 int __meminit early_pfn_to_nid(unsigned long pfn)
1271 {
1272         static DEFINE_SPINLOCK(early_pfn_lock);
1273         int nid;
1274 
1275         spin_lock(&early_pfn_lock);
1276         nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1277         if (nid < 0)
1278                 nid = first_online_node;
1279         spin_unlock(&early_pfn_lock);
1280 
1281         return nid;
1282 }
1283 #endif
1284 
1285 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1286 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1287                                         struct mminit_pfnnid_cache *state)
1288 {
1289         int nid;
1290 
1291         nid = __early_pfn_to_nid(pfn, state);
1292         if (nid >= 0 && nid != node)
1293                 return false;
1294         return true;
1295 }
1296 
1297 /* Only safe to use early in boot when initialisation is single-threaded */
1298 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1299 {
1300         return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1301 }
1302 
1303 #else
1304 
1305 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1306 {
1307         return true;
1308 }
1309 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1310                                         struct mminit_pfnnid_cache *state)
1311 {
1312         return true;
1313 }
1314 #endif
1315 
1316 
1317 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1318                                                         unsigned int order)
1319 {
1320         if (early_page_uninitialised(pfn))
1321                 return;
1322         return __free_pages_boot_core(page, order);
1323 }
1324 
1325 /*
1326  * Check that the whole (or subset of) a pageblock given by the interval of
1327  * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1328  * with the migration of free compaction scanner. The scanners then need to
1329  * use only pfn_valid_within() check for arches that allow holes within
1330  * pageblocks.
1331  *
1332  * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1333  *
1334  * It's possible on some configurations to have a setup like node0 node1 node0
1335  * i.e. it's possible that all pages within a zones range of pages do not
1336  * belong to a single zone. We assume that a border between node0 and node1
1337  * can occur within a single pageblock, but not a node0 node1 node0
1338  * interleaving within a single pageblock. It is therefore sufficient to check
1339  * the first and last page of a pageblock and avoid checking each individual
1340  * page in a pageblock.
1341  */
1342 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1343                                      unsigned long end_pfn, struct zone *zone)
1344 {
1345         struct page *start_page;
1346         struct page *end_page;
1347 
1348         /* end_pfn is one past the range we are checking */
1349         end_pfn--;
1350 
1351         if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1352                 return NULL;
1353 
1354         start_page = pfn_to_page(start_pfn);
1355 
1356         if (page_zone(start_page) != zone)
1357                 return NULL;
1358 
1359         end_page = pfn_to_page(end_pfn);
1360 
1361         /* This gives a shorter code than deriving page_zone(end_page) */
1362         if (page_zone_id(start_page) != page_zone_id(end_page))
1363                 return NULL;
1364 
1365         return start_page;
1366 }
1367 
1368 void set_zone_contiguous(struct zone *zone)
1369 {
1370         unsigned long block_start_pfn = zone->zone_start_pfn;
1371         unsigned long block_end_pfn;
1372 
1373         block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1374         for (; block_start_pfn < zone_end_pfn(zone);
1375                         block_start_pfn = block_end_pfn,
1376                          block_end_pfn += pageblock_nr_pages) {
1377 
1378                 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1379 
1380                 if (!__pageblock_pfn_to_page(block_start_pfn,
1381                                              block_end_pfn, zone))
1382                         return;
1383         }
1384 
1385         /* We confirm that there is no hole */
1386         zone->contiguous = true;
1387 }
1388 
1389 void clear_zone_contiguous(struct zone *zone)
1390 {
1391         zone->contiguous = false;
1392 }
1393 
1394 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1395 static void __init deferred_free_range(struct page *page,
1396                                         unsigned long pfn, int nr_pages)
1397 {
1398         int i;
1399 
1400         if (!page)
1401                 return;
1402 
1403         /* Free a large naturally-aligned chunk if possible */
1404         if (nr_pages == MAX_ORDER_NR_PAGES &&
1405             (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1406                 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1407                 __free_pages_boot_core(page, MAX_ORDER-1);
1408                 return;
1409         }
1410 
1411         for (i = 0; i < nr_pages; i++, page++)
1412                 __free_pages_boot_core(page, 0);
1413 }
1414 
1415 /* Completion tracking for deferred_init_memmap() threads */
1416 static atomic_t pgdat_init_n_undone __initdata;
1417 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1418 
1419 static inline void __init pgdat_init_report_one_done(void)
1420 {
1421         if (atomic_dec_and_test(&pgdat_init_n_undone))
1422                 complete(&pgdat_init_all_done_comp);
1423 }
1424 
1425 /* Initialise remaining memory on a node */
1426 static int __init deferred_init_memmap(void *data)
1427 {
1428         pg_data_t *pgdat = data;
1429         int nid = pgdat->node_id;
1430         struct mminit_pfnnid_cache nid_init_state = { };
1431         unsigned long start = jiffies;
1432         unsigned long nr_pages = 0;
1433         unsigned long walk_start, walk_end;
1434         int i, zid;
1435         struct zone *zone;
1436         unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1437         const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1438 
1439         if (first_init_pfn == ULONG_MAX) {
1440                 pgdat_init_report_one_done();
1441                 return 0;
1442         }
1443 
1444         /* Bind memory initialisation thread to a local node if possible */
1445         if (!cpumask_empty(cpumask))
1446                 set_cpus_allowed_ptr(current, cpumask);
1447 
1448         /* Sanity check boundaries */
1449         BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1450         BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1451         pgdat->first_deferred_pfn = ULONG_MAX;
1452 
1453         /* Only the highest zone is deferred so find it */
1454         for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1455                 zone = pgdat->node_zones + zid;
1456                 if (first_init_pfn < zone_end_pfn(zone))
1457                         break;
1458         }
1459 
1460         for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1461                 unsigned long pfn, end_pfn;
1462                 struct page *page = NULL;
1463                 struct page *free_base_page = NULL;
1464                 unsigned long free_base_pfn = 0;
1465                 int nr_to_free = 0;
1466 
1467                 end_pfn = min(walk_end, zone_end_pfn(zone));
1468                 pfn = first_init_pfn;
1469                 if (pfn < walk_start)
1470                         pfn = walk_start;
1471                 if (pfn < zone->zone_start_pfn)
1472                         pfn = zone->zone_start_pfn;
1473 
1474                 for (; pfn < end_pfn; pfn++) {
1475                         if (!pfn_valid_within(pfn))
1476                                 goto free_range;
1477 
1478                         /*
1479                          * Ensure pfn_valid is checked every
1480                          * MAX_ORDER_NR_PAGES for memory holes
1481                          */
1482                         if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1483                                 if (!pfn_valid(pfn)) {
1484                                         page = NULL;
1485                                         goto free_range;
1486                                 }
1487                         }
1488 
1489                         if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1490                                 page = NULL;
1491                                 goto free_range;
1492                         }
1493 
1494                         /* Minimise pfn page lookups and scheduler checks */
1495                         if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1496                                 page++;
1497                         } else {
1498                                 nr_pages += nr_to_free;
1499                                 deferred_free_range(free_base_page,
1500                                                 free_base_pfn, nr_to_free);
1501                                 free_base_page = NULL;
1502                                 free_base_pfn = nr_to_free = 0;
1503 
1504                                 page = pfn_to_page(pfn);
1505                                 cond_resched();
1506                         }
1507 
1508                         if (page->flags) {
1509                                 VM_BUG_ON(page_zone(page) != zone);
1510                                 goto free_range;
1511                         }
1512 
1513                         __init_single_page(page, pfn, zid, nid);
1514                         if (!free_base_page) {
1515                                 free_base_page = page;
1516                                 free_base_pfn = pfn;
1517                                 nr_to_free = 0;
1518                         }
1519                         nr_to_free++;
1520 
1521                         /* Where possible, batch up pages for a single free */
1522                         continue;
1523 free_range:
1524                         /* Free the current block of pages to allocator */
1525                         nr_pages += nr_to_free;
1526                         deferred_free_range(free_base_page, free_base_pfn,
1527                                                                 nr_to_free);
1528                         free_base_page = NULL;
1529                         free_base_pfn = nr_to_free = 0;
1530                 }
1531 
1532                 first_init_pfn = max(end_pfn, first_init_pfn);
1533         }
1534 
1535         /* Sanity check that the next zone really is unpopulated */
1536         WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1537 
1538         pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1539                                         jiffies_to_msecs(jiffies - start));
1540 
1541         pgdat_init_report_one_done();
1542         return 0;
1543 }
1544 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1545 
1546 void __init page_alloc_init_late(void)
1547 {
1548         struct zone *zone;
1549 
1550 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1551         int nid;
1552 
1553         /* There will be num_node_state(N_MEMORY) threads */
1554         atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1555         for_each_node_state(nid, N_MEMORY) {
1556                 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1557         }
1558 
1559         /* Block until all are initialised */
1560         wait_for_completion(&pgdat_init_all_done_comp);
1561 
1562         /* Reinit limits that are based on free pages after the kernel is up */
1563         files_maxfiles_init();
1564 #endif
1565 
1566         for_each_populated_zone(zone)
1567                 set_zone_contiguous(zone);
1568 }
1569 
1570 #ifdef CONFIG_CMA
1571 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1572 void __init init_cma_reserved_pageblock(struct page *page)
1573 {
1574         unsigned i = pageblock_nr_pages;
1575         struct page *p = page;
1576 
1577         do {
1578                 __ClearPageReserved(p);
1579                 set_page_count(p, 0);
1580         } while (++p, --i);
1581 
1582         set_pageblock_migratetype(page, MIGRATE_CMA);
1583 
1584         if (pageblock_order >= MAX_ORDER) {
1585                 i = pageblock_nr_pages;
1586                 p = page;
1587                 do {
1588                         set_page_refcounted(p);
1589                         __free_pages(p, MAX_ORDER - 1);
1590                         p += MAX_ORDER_NR_PAGES;
1591                 } while (i -= MAX_ORDER_NR_PAGES);
1592         } else {
1593                 set_page_refcounted(page);
1594                 __free_pages(page, pageblock_order);
1595         }
1596 
1597         adjust_managed_page_count(page, pageblock_nr_pages);
1598 }
1599 #endif
1600 
1601 /*
1602  * The order of subdivision here is critical for the IO subsystem.
1603  * Please do not alter this order without good reasons and regression
1604  * testing. Specifically, as large blocks of memory are subdivided,
1605  * the order in which smaller blocks are delivered depends on the order
1606  * they're subdivided in this function. This is the primary factor
1607  * influencing the order in which pages are delivered to the IO
1608  * subsystem according to empirical testing, and this is also justified
1609  * by considering the behavior of a buddy system containing a single
1610  * large block of memory acted on by a series of small allocations.
1611  * This behavior is a critical factor in sglist merging's success.
1612  *
1613  * -- nyc
1614  */
1615 static inline void expand(struct zone *zone, struct page *page,
1616         int low, int high, struct free_area *area,
1617         int migratetype)
1618 {
1619         unsigned long size = 1 << high;
1620 
1621         while (high > low) {
1622                 area--;
1623                 high--;
1624                 size >>= 1;
1625                 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1626 
1627                 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1628                         debug_guardpage_enabled() &&
1629                         high < debug_guardpage_minorder()) {
1630                         /*
1631                          * Mark as guard pages (or page), that will allow to
1632                          * merge back to allocator when buddy will be freed.
1633                          * Corresponding page table entries will not be touched,
1634                          * pages will stay not present in virtual address space
1635                          */
1636                         set_page_guard(zone, &page[size], high, migratetype);
1637                         continue;
1638                 }
1639                 list_add(&page[size].lru, &area->free_list[migratetype]);
1640                 area->nr_free++;
1641                 set_page_order(&page[size], high);
1642         }
1643 }
1644 
1645 static void check_new_page_bad(struct page *page)
1646 {
1647         const char *bad_reason = NULL;
1648         unsigned long bad_flags = 0;
1649 
1650         if (unlikely(atomic_read(&page->_mapcount) != -1))
1651                 bad_reason = "nonzero mapcount";
1652         if (unlikely(page->mapping != NULL))
1653                 bad_reason = "non-NULL mapping";
1654         if (unlikely(page_ref_count(page) != 0))
1655                 bad_reason = "nonzero _count";
1656         if (unlikely(page->flags & __PG_HWPOISON)) {
1657                 bad_reason = "HWPoisoned (hardware-corrupted)";
1658                 bad_flags = __PG_HWPOISON;
1659                 /* Don't complain about hwpoisoned pages */
1660                 page_mapcount_reset(page); /* remove PageBuddy */
1661                 return;
1662         }
1663         if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1664                 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1665                 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1666         }
1667 #ifdef CONFIG_MEMCG
1668         if (unlikely(page->mem_cgroup))
1669                 bad_reason = "page still charged to cgroup";
1670 #endif
1671         bad_page(page, bad_reason, bad_flags);
1672 }
1673 
1674 /*
1675  * This page is about to be returned from the page allocator
1676  */
1677 static inline int check_new_page(struct page *page)
1678 {
1679         if (likely(page_expected_state(page,
1680                                 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1681                 return 0;
1682 
1683         check_new_page_bad(page);
1684         return 1;
1685 }
1686 
1687 static inline bool free_pages_prezeroed(bool poisoned)
1688 {
1689         return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1690                 page_poisoning_enabled() && poisoned;
1691 }
1692 
1693 #ifdef CONFIG_DEBUG_VM
1694 static bool check_pcp_refill(struct page *page)
1695 {
1696         return false;
1697 }
1698 
1699 static bool check_new_pcp(struct page *page)
1700 {
1701         return check_new_page(page);
1702 }
1703 #else
1704 static bool check_pcp_refill(struct page *page)
1705 {
1706         return check_new_page(page);
1707 }
1708 static bool check_new_pcp(struct page *page)
1709 {
1710         return false;
1711 }
1712 #endif /* CONFIG_DEBUG_VM */
1713 
1714 static bool check_new_pages(struct page *page, unsigned int order)
1715 {
1716         int i;
1717         for (i = 0; i < (1 << order); i++) {
1718                 struct page *p = page + i;
1719 
1720                 if (unlikely(check_new_page(p)))
1721                         return true;
1722         }
1723 
1724         return false;
1725 }
1726 
1727 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1728                                                         unsigned int alloc_flags)
1729 {
1730         int i;
1731         bool poisoned = true;
1732 
1733         for (i = 0; i < (1 << order); i++) {
1734                 struct page *p = page + i;
1735                 if (poisoned)
1736                         poisoned &= page_is_poisoned(p);
1737         }
1738 
1739         set_page_private(page, 0);
1740         set_page_refcounted(page);
1741 
1742         arch_alloc_page(page, order);
1743         kernel_map_pages(page, 1 << order, 1);
1744         kernel_poison_pages(page, 1 << order, 1);
1745         kasan_alloc_pages(page, order);
1746 
1747         if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1748                 for (i = 0; i < (1 << order); i++)
1749                         clear_highpage(page + i);
1750 
1751         if (order && (gfp_flags & __GFP_COMP))
1752                 prep_compound_page(page, order);
1753 
1754         set_page_owner(page, order, gfp_flags);
1755 
1756         /*
1757          * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1758          * allocate the page. The expectation is that the caller is taking
1759          * steps that will free more memory. The caller should avoid the page
1760          * being used for !PFMEMALLOC purposes.
1761          */
1762         if (alloc_flags & ALLOC_NO_WATERMARKS)
1763                 set_page_pfmemalloc(page);
1764         else
1765                 clear_page_pfmemalloc(page);
1766 }
1767 
1768 /*
1769  * Go through the free lists for the given migratetype and remove
1770  * the smallest available page from the freelists
1771  */
1772 static inline
1773 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1774                                                 int migratetype)
1775 {
1776         unsigned int current_order;
1777         struct free_area *area;
1778         struct page *page;
1779 
1780         /* Find a page of the appropriate size in the preferred list */
1781         for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1782                 area = &(zone->free_area[current_order]);
1783                 page = list_first_entry_or_null(&area->free_list[migratetype],
1784                                                         struct page, lru);
1785                 if (!page)
1786                         continue;
1787                 list_del(&page->lru);
1788                 rmv_page_order(page);
1789                 area->nr_free--;
1790                 expand(zone, page, order, current_order, area, migratetype);
1791                 set_pcppage_migratetype(page, migratetype);
1792                 return page;
1793         }
1794 
1795         return NULL;
1796 }
1797 
1798 
1799 /*
1800  * This array describes the order lists are fallen back to when
1801  * the free lists for the desirable migrate type are depleted
1802  */
1803 static int fallbacks[MIGRATE_TYPES][4] = {
1804         [MIGRATE_UNMOVABLE]   = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE,   MIGRATE_TYPES },
1805         [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE,   MIGRATE_MOVABLE,   MIGRATE_TYPES },
1806         [MIGRATE_MOVABLE]     = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1807 #ifdef CONFIG_CMA
1808         [MIGRATE_CMA]         = { MIGRATE_TYPES }, /* Never used */
1809 #endif
1810 #ifdef CONFIG_MEMORY_ISOLATION
1811         [MIGRATE_ISOLATE]     = { MIGRATE_TYPES }, /* Never used */
1812 #endif
1813 };
1814 
1815 #ifdef CONFIG_CMA
1816 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1817                                         unsigned int order)
1818 {
1819         return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1820 }
1821 #else
1822 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1823                                         unsigned int order) { return NULL; }
1824 #endif
1825 
1826 /*
1827  * Move the free pages in a range to the free lists of the requested type.
1828  * Note that start_page and end_pages are not aligned on a pageblock
1829  * boundary. If alignment is required, use move_freepages_block()
1830  */
1831 int move_freepages(struct zone *zone,
1832                           struct page *start_page, struct page *end_page,
1833                           int migratetype)
1834 {
1835         struct page *page;
1836         unsigned int order;
1837         int pages_moved = 0;
1838 
1839 #ifndef CONFIG_HOLES_IN_ZONE
1840         /*
1841          * page_zone is not safe to call in this context when
1842          * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1843          * anyway as we check zone boundaries in move_freepages_block().
1844          * Remove at a later date when no bug reports exist related to
1845          * grouping pages by mobility
1846          */
1847         VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1848 #endif
1849 
1850         for (page = start_page; page <= end_page;) {
1851                 /* Make sure we are not inadvertently changing nodes */
1852                 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1853 
1854                 if (!pfn_valid_within(page_to_pfn(page))) {
1855                         page++;
1856                         continue;
1857                 }
1858 
1859                 if (!PageBuddy(page)) {
1860                         page++;
1861                         continue;
1862                 }
1863 
1864                 order = page_order(page);
1865                 list_move(&page->lru,
1866                           &zone->free_area[order].free_list[migratetype]);
1867                 page += 1 << order;
1868                 pages_moved += 1 << order;
1869         }
1870 
1871         return pages_moved;
1872 }
1873 
1874 int move_freepages_block(struct zone *zone, struct page *page,
1875                                 int migratetype)
1876 {
1877         unsigned long start_pfn, end_pfn;
1878         struct page *start_page, *end_page;
1879 
1880         start_pfn = page_to_pfn(page);
1881         start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1882         start_page = pfn_to_page(start_pfn);
1883         end_page = start_page + pageblock_nr_pages - 1;
1884         end_pfn = start_pfn + pageblock_nr_pages - 1;
1885 
1886         /* Do not cross zone boundaries */
1887         if (!zone_spans_pfn(zone, start_pfn))
1888                 start_page = page;
1889         if (!zone_spans_pfn(zone, end_pfn))
1890                 return 0;
1891 
1892         return move_freepages(zone, start_page, end_page, migratetype);
1893 }
1894 
1895 static void change_pageblock_range(struct page *pageblock_page,
1896                                         int start_order, int migratetype)
1897 {
1898         int nr_pageblocks = 1 << (start_order - pageblock_order);
1899 
1900         while (nr_pageblocks--) {
1901                 set_pageblock_migratetype(pageblock_page, migratetype);
1902                 pageblock_page += pageblock_nr_pages;
1903         }
1904 }
1905 
1906 /*
1907  * When we are falling back to another migratetype during allocation, try to
1908  * steal extra free pages from the same pageblocks to satisfy further
1909  * allocations, instead of polluting multiple pageblocks.
1910  *
1911  * If we are stealing a relatively large buddy page, it is likely there will
1912  * be more free pages in the pageblock, so try to steal them all. For
1913  * reclaimable and unmovable allocations, we steal regardless of page size,
1914  * as fragmentation caused by those allocations polluting movable pageblocks
1915  * is worse than movable allocations stealing from unmovable and reclaimable
1916  * pageblocks.
1917  */
1918 static bool can_steal_fallback(unsigned int order, int start_mt)
1919 {
1920         /*
1921          * Leaving this order check is intended, although there is
1922          * relaxed order check in next check. The reason is that
1923          * we can actually steal whole pageblock if this condition met,
1924          * but, below check doesn't guarantee it and that is just heuristic
1925          * so could be changed anytime.
1926          */
1927         if (order >= pageblock_order)
1928                 return true;
1929 
1930         if (order >= pageblock_order / 2 ||
1931                 start_mt == MIGRATE_RECLAIMABLE ||
1932                 start_mt == MIGRATE_UNMOVABLE ||
1933                 page_group_by_mobility_disabled)
1934                 return true;
1935 
1936         return false;
1937 }
1938 
1939 /*
1940  * This function implements actual steal behaviour. If order is large enough,
1941  * we can steal whole pageblock. If not, we first move freepages in this
1942  * pageblock and check whether half of pages are moved or not. If half of
1943  * pages are moved, we can change migratetype of pageblock and permanently
1944  * use it's pages as requested migratetype in the future.
1945  */
1946 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1947                                                           int start_type)
1948 {
1949         unsigned int current_order = page_order(page);
1950         int pages;
1951 
1952         /* Take ownership for orders >= pageblock_order */
1953         if (current_order >= pageblock_order) {
1954                 change_pageblock_range(page, current_order, start_type);
1955                 return;
1956         }
1957 
1958         pages = move_freepages_block(zone, page, start_type);
1959 
1960         /* Claim the whole block if over half of it is free */
1961         if (pages >= (1 << (pageblock_order-1)) ||
1962                         page_group_by_mobility_disabled)
1963                 set_pageblock_migratetype(page, start_type);
1964 }
1965 
1966 /*
1967  * Check whether there is a suitable fallback freepage with requested order.
1968  * If only_stealable is true, this function returns fallback_mt only if
1969  * we can steal other freepages all together. This would help to reduce
1970  * fragmentation due to mixed migratetype pages in one pageblock.
1971  */
1972 int find_suitable_fallback(struct free_area *area, unsigned int order,
1973                         int migratetype, bool only_stealable, bool *can_steal)
1974 {
1975         int i;
1976         int fallback_mt;
1977 
1978         if (area->nr_free == 0)
1979                 return -1;
1980 
1981         *can_steal = false;
1982         for (i = 0;; i++) {
1983                 fallback_mt = fallbacks[migratetype][i];
1984                 if (fallback_mt == MIGRATE_TYPES)
1985                         break;
1986 
1987                 if (list_empty(&area->free_list[fallback_mt]))
1988                         continue;
1989 
1990                 if (can_steal_fallback(order, migratetype))
1991                         *can_steal = true;
1992 
1993                 if (!only_stealable)
1994                         return fallback_mt;
1995 
1996                 if (*can_steal)
1997                         return fallback_mt;
1998         }
1999 
2000         return -1;
2001 }
2002 
2003 /*
2004  * Reserve a pageblock for exclusive use of high-order atomic allocations if
2005  * there are no empty page blocks that contain a page with a suitable order
2006  */
2007 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2008                                 unsigned int alloc_order)
2009 {
2010         int mt;
2011         unsigned long max_managed, flags;
2012 
2013         /*
2014          * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2015          * Check is race-prone but harmless.
2016          */
2017         max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2018         if (zone->nr_reserved_highatomic >= max_managed)
2019                 return;
2020 
2021         spin_lock_irqsave(&zone->lock, flags);
2022 
2023         /* Recheck the nr_reserved_highatomic limit under the lock */
2024         if (zone->nr_reserved_highatomic >= max_managed)
2025                 goto out_unlock;
2026 
2027         /* Yoink! */
2028         mt = get_pageblock_migratetype(page);
2029         if (mt != MIGRATE_HIGHATOMIC &&
2030                         !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2031                 zone->nr_reserved_highatomic += pageblock_nr_pages;
2032                 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2033                 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2034         }
2035 
2036 out_unlock:
2037         spin_unlock_irqrestore(&zone->lock, flags);
2038 }
2039 
2040 /*
2041  * Used when an allocation is about to fail under memory pressure. This
2042  * potentially hurts the reliability of high-order allocations when under
2043  * intense memory pressure but failed atomic allocations should be easier
2044  * to recover from than an OOM.
2045  */
2046 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2047 {
2048         struct zonelist *zonelist = ac->zonelist;
2049         unsigned long flags;
2050         struct zoneref *z;
2051         struct zone *zone;
2052         struct page *page;
2053         int order;
2054 
2055         for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2056                                                                 ac->nodemask) {
2057                 /* Preserve at least one pageblock */
2058                 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2059                         continue;
2060 
2061                 spin_lock_irqsave(&zone->lock, flags);
2062                 for (order = 0; order < MAX_ORDER; order++) {
2063                         struct free_area *area = &(zone->free_area[order]);
2064 
2065                         page = list_first_entry_or_null(
2066                                         &area->free_list[MIGRATE_HIGHATOMIC],
2067                                         struct page, lru);
2068                         if (!page)
2069                                 continue;
2070 
2071                         /*
2072                          * It should never happen but changes to locking could
2073                          * inadvertently allow a per-cpu drain to add pages
2074                          * to MIGRATE_HIGHATOMIC while unreserving so be safe
2075                          * and watch for underflows.
2076                          */
2077                         zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
2078                                 zone->nr_reserved_highatomic);
2079 
2080                         /*
2081                          * Convert to ac->migratetype and avoid the normal
2082                          * pageblock stealing heuristics. Minimally, the caller
2083                          * is doing the work and needs the pages. More
2084                          * importantly, if the block was always converted to
2085                          * MIGRATE_UNMOVABLE or another type then the number
2086                          * of pageblocks that cannot be completely freed
2087                          * may increase.
2088                          */
2089                         set_pageblock_migratetype(page, ac->migratetype);
2090                         move_freepages_block(zone, page, ac->migratetype);
2091                         spin_unlock_irqrestore(&zone->lock, flags);
2092                         return;
2093                 }
2094                 spin_unlock_irqrestore(&zone->lock, flags);
2095         }
2096 }
2097 
2098 /* Remove an element from the buddy allocator from the fallback list */
2099 static inline struct page *
2100 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2101 {
2102         struct free_area *area;
2103         unsigned int current_order;
2104         struct page *page;
2105         int fallback_mt;
2106         bool can_steal;
2107 
2108         /* Find the largest possible block of pages in the other list */
2109         for (current_order = MAX_ORDER-1;
2110                                 current_order >= order && current_order <= MAX_ORDER-1;
2111                                 --current_order) {
2112                 area = &(zone->free_area[current_order]);
2113                 fallback_mt = find_suitable_fallback(area, current_order,
2114                                 start_migratetype, false, &can_steal);
2115                 if (fallback_mt == -1)
2116                         continue;
2117 
2118                 page = list_first_entry(&area->free_list[fallback_mt],
2119                                                 struct page, lru);
2120                 if (can_steal)
2121                         steal_suitable_fallback(zone, page, start_migratetype);
2122 
2123                 /* Remove the page from the freelists */
2124                 area->nr_free--;
2125                 list_del(&page->lru);
2126                 rmv_page_order(page);
2127 
2128                 expand(zone, page, order, current_order, area,
2129                                         start_migratetype);
2130                 /*
2131                  * The pcppage_migratetype may differ from pageblock's
2132                  * migratetype depending on the decisions in
2133                  * find_suitable_fallback(). This is OK as long as it does not
2134                  * differ for MIGRATE_CMA pageblocks. Those can be used as
2135                  * fallback only via special __rmqueue_cma_fallback() function
2136                  */
2137                 set_pcppage_migratetype(page, start_migratetype);
2138 
2139                 trace_mm_page_alloc_extfrag(page, order, current_order,
2140                         start_migratetype, fallback_mt);
2141 
2142                 return page;
2143         }
2144 
2145         return NULL;
2146 }
2147 
2148 /*
2149  * Do the hard work of removing an element from the buddy allocator.
2150  * Call me with the zone->lock already held.
2151  */
2152 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2153                                 int migratetype)
2154 {
2155         struct page *page;
2156 
2157         page = __rmqueue_smallest(zone, order, migratetype);
2158         if (unlikely(!page)) {
2159                 if (migratetype == MIGRATE_MOVABLE)
2160                         page = __rmqueue_cma_fallback(zone, order);
2161 
2162                 if (!page)
2163                         page = __rmqueue_fallback(zone, order, migratetype);
2164         }
2165 
2166         trace_mm_page_alloc_zone_locked(page, order, migratetype);
2167         return page;
2168 }
2169 
2170 /*
2171  * Obtain a specified number of elements from the buddy allocator, all under
2172  * a single hold of the lock, for efficiency.  Add them to the supplied list.
2173  * Returns the number of new pages which were placed at *list.
2174  */
2175 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2176                         unsigned long count, struct list_head *list,
2177                         int migratetype, bool cold)
2178 {
2179         int i;
2180 
2181         spin_lock(&zone->lock);
2182         for (i = 0; i < count; ++i) {
2183                 struct page *page = __rmqueue(zone, order, migratetype);
2184                 if (unlikely(page == NULL))
2185                         break;
2186 
2187                 if (unlikely(check_pcp_refill(page)))
2188                         continue;
2189 
2190                 /*
2191                  * Split buddy pages returned by expand() are received here
2192                  * in physical page order. The page is added to the callers and
2193                  * list and the list head then moves forward. From the callers
2194                  * perspective, the linked list is ordered by page number in
2195                  * some conditions. This is useful for IO devices that can
2196                  * merge IO requests if the physical pages are ordered
2197                  * properly.
2198                  */
2199                 if (likely(!cold))
2200                         list_add(&page->lru, list);
2201                 else
2202                         list_add_tail(&page->lru, list);
2203                 list = &page->lru;
2204                 if (is_migrate_cma(get_pcppage_migratetype(page)))
2205                         __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2206                                               -(1 << order));
2207         }
2208         __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2209         spin_unlock(&zone->lock);
2210         return i;
2211 }
2212 
2213 #ifdef CONFIG_NUMA
2214 /*
2215  * Called from the vmstat counter updater to drain pagesets of this
2216  * currently executing processor on remote nodes after they have
2217  * expired.
2218  *
2219  * Note that this function must be called with the thread pinned to
2220  * a single processor.
2221  */
2222 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2223 {
2224         unsigned long flags;
2225         int to_drain, batch;
2226 
2227         local_irq_save(flags);
2228         batch = READ_ONCE(pcp->batch);
2229         to_drain = min(pcp->count, batch);
2230         if (to_drain > 0) {
2231                 free_pcppages_bulk(zone, to_drain, pcp);
2232                 pcp->count -= to_drain;
2233         }
2234         local_irq_restore(flags);
2235 }
2236 #endif
2237 
2238 /*
2239  * Drain pcplists of the indicated processor and zone.
2240  *
2241  * The processor must either be the current processor and the
2242  * thread pinned to the current processor or a processor that
2243  * is not online.
2244  */
2245 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2246 {
2247         unsigned long flags;
2248         struct per_cpu_pageset *pset;
2249         struct per_cpu_pages *pcp;
2250 
2251         local_irq_save(flags);
2252         pset = per_cpu_ptr(zone->pageset, cpu);
2253 
2254         pcp = &pset->pcp;
2255         if (pcp->count) {
2256                 free_pcppages_bulk(zone, pcp->count, pcp);
2257                 pcp->count = 0;
2258         }
2259         local_irq_restore(flags);
2260 }
2261 
2262 /*
2263  * Drain pcplists of all zones on the indicated processor.
2264  *
2265  * The processor must either be the current processor and the
2266  * thread pinned to the current processor or a processor that
2267  * is not online.
2268  */
2269 static void drain_pages(unsigned int cpu)
2270 {
2271         struct zone *zone;
2272 
2273         for_each_populated_zone(zone) {
2274                 drain_pages_zone(cpu, zone);
2275         }
2276 }
2277 
2278 /*
2279  * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2280  *
2281  * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2282  * the single zone's pages.
2283  */
2284 void drain_local_pages(struct zone *zone)
2285 {
2286         int cpu = smp_processor_id();
2287 
2288         if (zone)
2289                 drain_pages_zone(cpu, zone);
2290         else
2291                 drain_pages(cpu);
2292 }
2293 
2294 /*
2295  * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2296  *
2297  * When zone parameter is non-NULL, spill just the single zone's pages.
2298  *
2299  * Note that this code is protected against sending an IPI to an offline
2300  * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2301  * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2302  * nothing keeps CPUs from showing up after we populated the cpumask and
2303  * before the call to on_each_cpu_mask().
2304  */
2305 void drain_all_pages(struct zone *zone)
2306 {
2307         int cpu;
2308 
2309         /*
2310          * Allocate in the BSS so we wont require allocation in
2311          * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2312          */
2313         static cpumask_t cpus_with_pcps;
2314 
2315         /*
2316          * We don't care about racing with CPU hotplug event
2317          * as offline notification will cause the notified
2318          * cpu to drain that CPU pcps and on_each_cpu_mask
2319          * disables preemption as part of its processing
2320          */
2321         for_each_online_cpu(cpu) {
2322                 struct per_cpu_pageset *pcp;
2323                 struct zone *z;
2324                 bool has_pcps = false;
2325 
2326                 if (zone) {
2327                         pcp = per_cpu_ptr(zone->pageset, cpu);
2328                         if (pcp->pcp.count)
2329                                 has_pcps = true;
2330                 } else {
2331                         for_each_populated_zone(z) {
2332                                 pcp = per_cpu_ptr(z->pageset, cpu);
2333                                 if (pcp->pcp.count) {
2334                                         has_pcps = true;
2335                                         break;
2336                                 }
2337                         }
2338                 }
2339 
2340                 if (has_pcps)
2341                         cpumask_set_cpu(cpu, &cpus_with_pcps);
2342                 else
2343                         cpumask_clear_cpu(cpu, &cpus_with_pcps);
2344         }
2345         on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2346                                                                 zone, 1);
2347 }
2348 
2349 #ifdef CONFIG_HIBERNATION
2350 
2351 void mark_free_pages(struct zone *zone)
2352 {
2353         unsigned long pfn, max_zone_pfn;
2354         unsigned long flags;
2355         unsigned int order, t;
2356         struct page *page;
2357 
2358         if (zone_is_empty(zone))
2359                 return;
2360 
2361         spin_lock_irqsave(&zone->lock, flags);
2362 
2363         max_zone_pfn = zone_end_pfn(zone);
2364         for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2365                 if (pfn_valid(pfn)) {
2366                         page = pfn_to_page(pfn);
2367 
2368                         if (page_zone(page) != zone)
2369                                 continue;
2370 
2371                         if (!swsusp_page_is_forbidden(page))
2372                                 swsusp_unset_page_free(page);
2373                 }
2374 
2375         for_each_migratetype_order(order, t) {
2376                 list_for_each_entry(page,
2377                                 &zone->free_area[order].free_list[t], lru) {
2378                         unsigned long i;
2379 
2380                         pfn = page_to_pfn(page);
2381                         for (i = 0; i < (1UL << order); i++)
2382                                 swsusp_set_page_free(pfn_to_page(pfn + i));
2383                 }
2384         }
2385         spin_unlock_irqrestore(&zone->lock, flags);
2386 }
2387 #endif /* CONFIG_PM */
2388 
2389 /*
2390  * Free a 0-order page
2391  * cold == true ? free a cold page : free a hot page
2392  */
2393 void free_hot_cold_page(struct page *page, bool cold)
2394 {
2395         struct zone *zone = page_zone(page);
2396         struct per_cpu_pages *pcp;
2397         unsigned long flags;
2398         unsigned long pfn = page_to_pfn(page);
2399         int migratetype;
2400 
2401         if (!free_pcp_prepare(page))
2402                 return;
2403 
2404         migratetype = get_pfnblock_migratetype(page, pfn);
2405         set_pcppage_migratetype(page, migratetype);
2406         local_irq_save(flags);
2407         __count_vm_event(PGFREE);
2408 
2409         /*
2410          * We only track unmovable, reclaimable and movable on pcp lists.
2411          * Free ISOLATE pages back to the allocator because they are being
2412          * offlined but treat RESERVE as movable pages so we can get those
2413          * areas back if necessary. Otherwise, we may have to free
2414          * excessively into the page allocator
2415          */
2416         if (migratetype >= MIGRATE_PCPTYPES) {
2417                 if (unlikely(is_migrate_isolate(migratetype))) {
2418                         free_one_page(zone, page, pfn, 0, migratetype);
2419                         goto out;
2420                 }
2421                 migratetype = MIGRATE_MOVABLE;
2422         }
2423 
2424         pcp = &this_cpu_ptr(zone->pageset)->pcp;
2425         if (!cold)
2426                 list_add(&page->lru, &pcp->lists[migratetype]);
2427         else
2428                 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2429         pcp->count++;
2430         if (pcp->count >= pcp->high) {
2431                 unsigned long batch = READ_ONCE(pcp->batch);
2432                 free_pcppages_bulk(zone, batch, pcp);
2433                 pcp->count -= batch;
2434         }
2435 
2436 out:
2437         local_irq_restore(flags);
2438 }
2439 
2440 /*
2441  * Free a list of 0-order pages
2442  */
2443 void free_hot_cold_page_list(struct list_head *list, bool cold)
2444 {
2445         struct page *page, *next;
2446 
2447         list_for_each_entry_safe(page, next, list, lru) {
2448                 trace_mm_page_free_batched(page, cold);
2449                 free_hot_cold_page(page, cold);
2450         }
2451 }
2452 
2453 /*
2454  * split_page takes a non-compound higher-order page, and splits it into
2455  * n (1<<order) sub-pages: page[0..n]
2456  * Each sub-page must be freed individually.
2457  *
2458  * Note: this is probably too low level an operation for use in drivers.
2459  * Please consult with lkml before using this in your driver.
2460  */
2461 void split_page(struct page *page, unsigned int order)
2462 {
2463         int i;
2464         gfp_t gfp_mask;
2465 
2466         VM_BUG_ON_PAGE(PageCompound(page), page);
2467         VM_BUG_ON_PAGE(!page_count(page), page);
2468 
2469 #ifdef CONFIG_KMEMCHECK
2470         /*
2471          * Split shadow pages too, because free(page[0]) would
2472          * otherwise free the whole shadow.
2473          */
2474         if (kmemcheck_page_is_tracked(page))
2475                 split_page(virt_to_page(page[0].shadow), order);
2476 #endif
2477 
2478         gfp_mask = get_page_owner_gfp(page);
2479         set_page_owner(page, 0, gfp_mask);
2480         for (i = 1; i < (1 << order); i++) {
2481                 set_page_refcounted(page + i);
2482                 set_page_owner(page + i, 0, gfp_mask);
2483         }
2484 }
2485 EXPORT_SYMBOL_GPL(split_page);
2486 
2487 int __isolate_free_page(struct page *page, unsigned int order)
2488 {
2489         unsigned long watermark;
2490         struct zone *zone;
2491         int mt;
2492 
2493         BUG_ON(!PageBuddy(page));
2494 
2495         zone = page_zone(page);
2496         mt = get_pageblock_migratetype(page);
2497 
2498         if (!is_migrate_isolate(mt)) {
2499                 /* Obey watermarks as if the page was being allocated */
2500                 watermark = low_wmark_pages(zone) + (1 << order);
2501                 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2502                         return 0;
2503 
2504                 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2505         }
2506 
2507         /* Remove page from free list */
2508         list_del(&page->lru);
2509         zone->free_area[order].nr_free--;
2510         rmv_page_order(page);
2511 
2512         set_page_owner(page, order, __GFP_MOVABLE);
2513 
2514         /* Set the pageblock if the isolated page is at least a pageblock */
2515         if (order >= pageblock_order - 1) {
2516                 struct page *endpage = page + (1 << order) - 1;
2517                 for (; page < endpage; page += pageblock_nr_pages) {
2518                         int mt = get_pageblock_migratetype(page);
2519                         if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2520                                 set_pageblock_migratetype(page,
2521                                                           MIGRATE_MOVABLE);
2522                 }
2523         }
2524 
2525 
2526         return 1UL << order;
2527 }
2528 
2529 /*
2530  * Similar to split_page except the page is already free. As this is only
2531  * being used for migration, the migratetype of the block also changes.
2532  * As this is called with interrupts disabled, the caller is responsible
2533  * for calling arch_alloc_page() and kernel_map_page() after interrupts
2534  * are enabled.
2535  *
2536  * Note: this is probably too low level an operation for use in drivers.
2537  * Please consult with lkml before using this in your driver.
2538  */
2539 int split_free_page(struct page *page)
2540 {
2541         unsigned int order;
2542         int nr_pages;
2543 
2544         order = page_order(page);
2545 
2546         nr_pages = __isolate_free_page(page, order);
2547         if (!nr_pages)
2548                 return 0;
2549 
2550         /* Split into individual pages */
2551         set_page_refcounted(page);
2552         split_page(page, order);
2553         return nr_pages;
2554 }
2555 
2556 /*
2557  * Update NUMA hit/miss statistics
2558  *
2559  * Must be called with interrupts disabled.
2560  *
2561  * When __GFP_OTHER_NODE is set assume the node of the preferred
2562  * zone is the local node. This is useful for daemons who allocate
2563  * memory on behalf of other processes.
2564  */
2565 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2566                                                                 gfp_t flags)
2567 {
2568 #ifdef CONFIG_NUMA
2569         int local_nid = numa_node_id();
2570         enum zone_stat_item local_stat = NUMA_LOCAL;
2571 
2572         if (unlikely(flags & __GFP_OTHER_NODE)) {
2573                 local_stat = NUMA_OTHER;
2574                 local_nid = preferred_zone->node;
2575         }
2576 
2577         if (z->node == local_nid) {
2578                 __inc_zone_state(z, NUMA_HIT);
2579                 __inc_zone_state(z, local_stat);
2580         } else {
2581                 __inc_zone_state(z, NUMA_MISS);
2582                 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2583         }
2584 #endif
2585 }
2586 
2587 /*
2588  * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2589  */
2590 static inline
2591 struct page *buffered_rmqueue(struct zone *preferred_zone,
2592                         struct zone *zone, unsigned int order,
2593                         gfp_t gfp_flags, unsigned int alloc_flags,
2594                         int migratetype)
2595 {
2596         unsigned long flags;
2597         struct page *page;
2598         bool cold = ((gfp_flags & __GFP_COLD) != 0);
2599 
2600         if (likely(order == 0)) {
2601                 struct per_cpu_pages *pcp;
2602                 struct list_head *list;
2603 
2604                 local_irq_save(flags);
2605                 do {
2606                         pcp = &this_cpu_ptr(zone->pageset)->pcp;
2607                         list = &pcp->lists[migratetype];
2608                         if (list_empty(list)) {
2609                                 pcp->count += rmqueue_bulk(zone, 0,
2610                                                 pcp->batch, list,
2611                                                 migratetype, cold);
2612                                 if (unlikely(list_empty(list)))
2613                                         goto failed;
2614                         }
2615 
2616                         if (cold)
2617                                 page = list_last_entry(list, struct page, lru);
2618                         else
2619                                 page = list_first_entry(list, struct page, lru);
2620 
2621                         __dec_zone_state(zone, NR_ALLOC_BATCH);
2622                         list_del(&page->lru);
2623                         pcp->count--;
2624 
2625                 } while (check_new_pcp(page));
2626         } else {
2627                 /*
2628                  * We most definitely don't want callers attempting to
2629                  * allocate greater than order-1 page units with __GFP_NOFAIL.
2630                  */
2631                 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2632                 spin_lock_irqsave(&zone->lock, flags);
2633 
2634                 do {
2635                         page = NULL;
2636                         if (alloc_flags & ALLOC_HARDER) {
2637                                 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2638                                 if (page)
2639                                         trace_mm_page_alloc_zone_locked(page, order, migratetype);
2640                         }
2641                         if (!page)
2642                                 page = __rmqueue(zone, order, migratetype);
2643                 } while (page && check_new_pages(page, order));
2644                 spin_unlock(&zone->lock);
2645                 if (!page)
2646                         goto failed;
2647                 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2648                 __mod_zone_freepage_state(zone, -(1 << order),
2649                                           get_pcppage_migratetype(page));
2650         }
2651 
2652         if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2653             !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2654                 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2655 
2656         __count_zone_vm_events(PGALLOC, zone, 1 << order);
2657         zone_statistics(preferred_zone, zone, gfp_flags);
2658         local_irq_restore(flags);
2659 
2660         VM_BUG_ON_PAGE(bad_range(zone, page), page);
2661         return page;
2662 
2663 failed:
2664         local_irq_restore(flags);
2665         return NULL;
2666 }
2667 
2668 #ifdef CONFIG_FAIL_PAGE_ALLOC
2669 
2670 static struct {
2671         struct fault_attr attr;
2672 
2673         bool ignore_gfp_highmem;
2674         bool ignore_gfp_reclaim;
2675         u32 min_order;
2676 } fail_page_alloc = {
2677         .attr = FAULT_ATTR_INITIALIZER,
2678         .ignore_gfp_reclaim = true,
2679         .ignore_gfp_highmem = true,
2680         .min_order = 1,
2681 };
2682 
2683 static int __init setup_fail_page_alloc(char *str)
2684 {
2685         return setup_fault_attr(&fail_page_alloc.attr, str);
2686 }
2687 __setup("fail_page_alloc=", setup_fail_page_alloc);
2688 
2689 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2690 {
2691         if (order < fail_page_alloc.min_order)
2692                 return false;
2693         if (gfp_mask & __GFP_NOFAIL)
2694                 return false;
2695         if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2696                 return false;
2697         if (fail_page_alloc.ignore_gfp_reclaim &&
2698                         (gfp_mask & __GFP_DIRECT_RECLAIM))
2699                 return false;
2700 
2701         return should_fail(&fail_page_alloc.attr, 1 << order);
2702 }
2703 
2704 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2705 
2706 static int __init fail_page_alloc_debugfs(void)
2707 {
2708         umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2709         struct dentry *dir;
2710 
2711         dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2712                                         &fail_page_alloc.attr);
2713         if (IS_ERR(dir))
2714                 return PTR_ERR(dir);
2715 
2716         if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2717                                 &fail_page_alloc.ignore_gfp_reclaim))
2718                 goto fail;
2719         if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2720                                 &fail_page_alloc.ignore_gfp_highmem))
2721                 goto fail;
2722         if (!debugfs_create_u32("min-order", mode, dir,
2723                                 &fail_page_alloc.min_order))
2724                 goto fail;
2725 
2726         return 0;
2727 fail:
2728         debugfs_remove_recursive(dir);
2729 
2730         return -ENOMEM;
2731 }
2732 
2733 late_initcall(fail_page_alloc_debugfs);
2734 
2735 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2736 
2737 #else /* CONFIG_FAIL_PAGE_ALLOC */
2738 
2739 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2740 {
2741         return false;
2742 }
2743 
2744 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2745 
2746 /*
2747  * Return true if free base pages are above 'mark'. For high-order checks it
2748  * will return true of the order-0 watermark is reached and there is at least
2749  * one free page of a suitable size. Checking now avoids taking the zone lock
2750  * to check in the allocation paths if no pages are free.
2751  */
2752 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2753                          int classzone_idx, unsigned int alloc_flags,
2754                          long free_pages)
2755 {
2756         long min = mark;
2757         int o;
2758         const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2759 
2760         /* free_pages may go negative - that's OK */
2761         free_pages -= (1 << order) - 1;
2762 
2763         if (alloc_flags & ALLOC_HIGH)
2764                 min -= min / 2;
2765 
2766         /*
2767          * If the caller does not have rights to ALLOC_HARDER then subtract
2768          * the high-atomic reserves. This will over-estimate the size of the
2769          * atomic reserve but it avoids a search.
2770          */
2771         if (likely(!alloc_harder))
2772                 free_pages -= z->nr_reserved_highatomic;
2773         else
2774                 min -= min / 4;
2775 
2776 #ifdef CONFIG_CMA
2777         /* If allocation can't use CMA areas don't use free CMA pages */
2778         if (!(alloc_flags & ALLOC_CMA))
2779                 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2780 #endif
2781 
2782         /*
2783          * Check watermarks for an order-0 allocation request. If these
2784          * are not met, then a high-order request also cannot go ahead
2785          * even if a suitable page happened to be free.
2786          */
2787         if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2788                 return false;
2789 
2790         /* If this is an order-0 request then the watermark is fine */
2791         if (!order)
2792                 return true;
2793 
2794         /* For a high-order request, check at least one suitable page is free */
2795         for (o = order; o < MAX_ORDER; o++) {
2796                 struct free_area *area = &z->free_area[o];
2797                 int mt;
2798 
2799                 if (!area->nr_free)
2800                         continue;
2801 
2802                 if (alloc_harder)
2803                         return true;
2804 
2805                 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2806                         if (!list_empty(&area->free_list[mt]))
2807                                 return true;
2808                 }
2809 
2810 #ifdef CONFIG_CMA
2811                 if ((alloc_flags & ALLOC_CMA) &&
2812                     !list_empty(&area->free_list[MIGRATE_CMA])) {
2813                         return true;
2814                 }
2815 #endif
2816         }
2817         return false;
2818 }
2819 
2820 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2821                       int classzone_idx, unsigned int alloc_flags)
2822 {
2823         return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2824                                         zone_page_state(z, NR_FREE_PAGES));
2825 }
2826 
2827 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2828                 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2829 {
2830         long free_pages = zone_page_state(z, NR_FREE_PAGES);
2831         long cma_pages = 0;
2832 
2833 #ifdef CONFIG_CMA
2834         /* If allocation can't use CMA areas don't use free CMA pages */
2835         if (!(alloc_flags & ALLOC_CMA))
2836                 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2837 #endif
2838 
2839         /*
2840          * Fast check for order-0 only. If this fails then the reserves
2841          * need to be calculated. There is a corner case where the check
2842          * passes but only the high-order atomic reserve are free. If
2843          * the caller is !atomic then it'll uselessly search the free
2844          * list. That corner case is then slower but it is harmless.
2845          */
2846         if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2847                 return true;
2848 
2849         return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2850                                         free_pages);
2851 }
2852 
2853 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2854                         unsigned long mark, int classzone_idx)
2855 {
2856         long free_pages = zone_page_state(z, NR_FREE_PAGES);
2857 
2858         if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2859                 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2860 
2861         return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2862                                                                 free_pages);
2863 }
2864 
2865 #ifdef CONFIG_NUMA
2866 static bool zone_local(struct zone *local_zone, struct zone *zone)
2867 {
2868         return local_zone->node == zone->node;
2869 }
2870 
2871 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2872 {
2873         return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2874                                 RECLAIM_DISTANCE;
2875 }
2876 #else   /* CONFIG_NUMA */
2877 static bool zone_local(struct zone *local_zone, struct zone *zone)
2878 {
2879         return true;
2880 }
2881 
2882 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2883 {
2884         return true;
2885 }
2886 #endif  /* CONFIG_NUMA */
2887 
2888 static void reset_alloc_batches(struct zone *preferred_zone)
2889 {
2890         struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2891 
2892         do {
2893                 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2894                         high_wmark_pages(zone) - low_wmark_pages(zone) -
2895                         atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2896                 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2897         } while (zone++ != preferred_zone);
2898 }
2899 
2900 /*
2901  * get_page_from_freelist goes through the zonelist trying to allocate
2902  * a page.
2903  */
2904 static struct page *
2905 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2906                                                 const struct alloc_context *ac)
2907 {
2908         struct zoneref *z = ac->preferred_zoneref;
2909         struct zone *zone;
2910         bool fair_skipped = false;
2911         bool apply_fair = (alloc_flags & ALLOC_FAIR);
2912 
2913 zonelist_scan:
2914         /*
2915          * Scan zonelist, looking for a zone with enough free.
2916          * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2917          */
2918         for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2919                                                                 ac->nodemask) {
2920                 struct page *page;
2921                 unsigned long mark;
2922 
2923                 if (cpusets_enabled() &&
2924                         (alloc_flags & ALLOC_CPUSET) &&
2925                         !__cpuset_zone_allowed(zone, gfp_mask))
2926                                 continue;
2927                 /*
2928                  * Distribute pages in proportion to the individual
2929                  * zone size to ensure fair page aging.  The zone a
2930                  * page was allocated in should have no effect on the
2931                  * time the page has in memory before being reclaimed.
2932                  */
2933                 if (apply_fair) {
2934                         if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2935                                 fair_skipped = true;
2936                                 continue;
2937                         }
2938                         if (!zone_local(ac->preferred_zoneref->zone, zone)) {
2939                                 if (fair_skipped)
2940                                         goto reset_fair;
2941                                 apply_fair = false;
2942                         }
2943                 }
2944                 /*
2945                  * When allocating a page cache page for writing, we
2946                  * want to get it from a zone that is within its dirty
2947                  * limit, such that no single zone holds more than its
2948                  * proportional share of globally allowed dirty pages.
2949                  * The dirty limits take into account the zone's
2950                  * lowmem reserves and high watermark so that kswapd
2951                  * should be able to balance it without having to
2952                  * write pages from its LRU list.
2953                  *
2954                  * This may look like it could increase pressure on
2955                  * lower zones by failing allocations in higher zones
2956                  * before they are full.  But the pages that do spill
2957                  * over are limited as the lower zones are protected
2958                  * by this very same mechanism.  It should not become
2959                  * a practical burden to them.
2960                  *
2961                  * XXX: For now, allow allocations to potentially
2962                  * exceed the per-zone dirty limit in the slowpath
2963                  * (spread_dirty_pages unset) before going into reclaim,
2964                  * which is important when on a NUMA setup the allowed
2965                  * zones are together not big enough to reach the
2966                  * global limit.  The proper fix for these situations
2967                  * will require awareness of zones in the
2968                  * dirty-throttling and the flusher threads.
2969                  */
2970                 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2971                         continue;
2972 
2973                 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2974                 if (!zone_watermark_fast(zone, order, mark,
2975                                        ac_classzone_idx(ac), alloc_flags)) {
2976                         int ret;
2977 
2978                         /* Checked here to keep the fast path fast */
2979                         BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2980                         if (alloc_flags & ALLOC_NO_WATERMARKS)
2981                                 goto try_this_zone;
2982 
2983                         if (zone_reclaim_mode == 0 ||
2984                             !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2985                                 continue;
2986 
2987                         ret = zone_reclaim(zone, gfp_mask, order);
2988                         switch (ret) {
2989                         case ZONE_RECLAIM_NOSCAN:
2990                                 /* did not scan */
2991                                 continue;
2992                         case ZONE_RECLAIM_FULL:
2993                                 /* scanned but unreclaimable */
2994                                 continue;
2995                         default:
2996                                 /* did we reclaim enough */
2997                                 if (zone_watermark_ok(zone, order, mark,
2998                                                 ac_classzone_idx(ac), alloc_flags))
2999                                         goto try_this_zone;
3000 
3001                                 continue;
3002                         }
3003                 }
3004 
3005 try_this_zone:
3006                 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
3007                                 gfp_mask, alloc_flags, ac->migratetype);
3008                 if (page) {
3009                         prep_new_page(page, order, gfp_mask, alloc_flags);
3010 
3011                         /*
3012                          * If this is a high-order atomic allocation then check
3013                          * if the pageblock should be reserved for the future
3014                          */
3015                         if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3016                                 reserve_highatomic_pageblock(page, zone, order);
3017 
3018                         return page;
3019                 }
3020         }
3021 
3022         /*
3023          * The first pass makes sure allocations are spread fairly within the
3024          * local node.  However, the local node might have free pages left
3025          * after the fairness batches are exhausted, and remote zones haven't
3026          * even been considered yet.  Try once more without fairness, and
3027          * include remote zones now, before entering the slowpath and waking
3028          * kswapd: prefer spilling to a remote zone over swapping locally.
3029          */
3030         if (fair_skipped) {
3031 reset_fair:
3032                 apply_fair = false;
3033                 fair_skipped = false;
3034                 reset_alloc_batches(ac->preferred_zoneref->zone);
3035                 z = ac->preferred_zoneref;
3036                 goto zonelist_scan;
3037         }
3038 
3039         return NULL;
3040 }
3041 
3042 /*
3043  * Large machines with many possible nodes should not always dump per-node
3044  * meminfo in irq context.
3045  */
3046 static inline bool should_suppress_show_mem(void)
3047 {
3048         bool ret = false;
3049 
3050 #if NODES_SHIFT > 8
3051         ret = in_interrupt();
3052 #endif
3053         return ret;
3054 }
3055 
3056 static DEFINE_RATELIMIT_STATE(nopage_rs,
3057                 DEFAULT_RATELIMIT_INTERVAL,
3058                 DEFAULT_RATELIMIT_BURST);
3059 
3060 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
3061 {
3062         unsigned int filter = SHOW_MEM_FILTER_NODES;
3063 
3064         if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3065             debug_guardpage_minorder() > 0)
3066                 return;
3067 
3068         /*
3069          * This documents exceptions given to allocations in certain
3070          * contexts that are allowed to allocate outside current's set
3071          * of allowed nodes.
3072          */
3073         if (!(gfp_mask & __GFP_NOMEMALLOC))
3074                 if (test_thread_flag(TIF_MEMDIE) ||
3075                     (current->flags & (PF_MEMALLOC | PF_EXITING)))
3076                         filter &= ~SHOW_MEM_FILTER_NODES;
3077         if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3078                 filter &= ~SHOW_MEM_FILTER_NODES;
3079 
3080         if (fmt) {
3081                 struct va_format vaf;
3082                 va_list args;
3083 
3084                 va_start(args, fmt);
3085 
3086                 vaf.fmt = fmt;
3087                 vaf.va = &args;
3088 
3089                 pr_warn("%pV", &vaf);
3090 
3091                 va_end(args);
3092         }
3093 
3094         pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
3095                 current->comm, order, gfp_mask, &gfp_mask);
3096         dump_stack();
3097         if (!should_suppress_show_mem())
3098                 show_mem(filter);
3099 }
3100 
3101 static inline struct page *
3102 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3103         const struct alloc_context *ac, unsigned long *did_some_progress)
3104 {
3105         struct oom_control oc = {
3106                 .zonelist = ac->zonelist,
3107                 .nodemask = ac->nodemask,
3108                 .gfp_mask = gfp_mask,
3109                 .order = order,
3110         };
3111         struct page *page;
3112 
3113         *did_some_progress = 0;
3114 
3115         /*
3116          * Acquire the oom lock.  If that fails, somebody else is
3117          * making progress for us.
3118          */
3119         if (!mutex_trylock(&oom_lock)) {
3120                 *did_some_progress = 1;
3121                 schedule_timeout_uninterruptible(1);
3122                 return NULL;
3123         }
3124 
3125         /*
3126          * Go through the zonelist yet one more time, keep very high watermark
3127          * here, this is only to catch a parallel oom killing, we must fail if
3128          * we're still under heavy pressure.
3129          */
3130         page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3131                                         ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3132         if (page)
3133                 goto out;
3134 
3135         if (!(gfp_mask & __GFP_NOFAIL)) {
3136                 /* Coredumps can quickly deplete all memory reserves */
3137                 if (current->flags & PF_DUMPCORE)
3138                         goto out;
3139                 /* The OOM killer will not help higher order allocs */
3140                 if (order > PAGE_ALLOC_COSTLY_ORDER)
3141                         goto out;
3142                 /* The OOM killer does not needlessly kill tasks for lowmem */
3143                 if (ac->high_zoneidx < ZONE_NORMAL)
3144                         goto out;
3145                 if (pm_suspended_storage())
3146                         goto out;
3147                 /*
3148                  * XXX: GFP_NOFS allocations should rather fail than rely on
3149                  * other request to make a forward progress.
3150                  * We are in an unfortunate situation where out_of_memory cannot
3151                  * do much for this context but let's try it to at least get
3152                  * access to memory reserved if the current task is killed (see
3153                  * out_of_memory). Once filesystems are ready to handle allocation
3154                  * failures more gracefully we should just bail out here.
3155                  */
3156 
3157                 /* The OOM killer may not free memory on a specific node */
3158                 if (gfp_mask & __GFP_THISNODE)
3159                         goto out;
3160         }
3161         /* Exhausted what can be done so it's blamo time */
3162         if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3163                 *did_some_progress = 1;
3164 
3165                 if (gfp_mask & __GFP_NOFAIL) {
3166                         page = get_page_from_freelist(gfp_mask, order,
3167                                         ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3168                         /*
3169                          * fallback to ignore cpuset restriction if our nodes
3170                          * are depleted
3171                          */
3172                         if (!page)
3173                                 page = get_page_from_freelist(gfp_mask, order,
3174                                         ALLOC_NO_WATERMARKS, ac);
3175                 }
3176         }
3177 out:
3178         mutex_unlock(&oom_lock);
3179         return page;
3180 }
3181 
3182 
3183 /*
3184  * Maximum number of compaction retries wit a progress before OOM
3185  * killer is consider as the only way to move forward.
3186  */
3187 #define MAX_COMPACT_RETRIES 16
3188 
3189 #ifdef CONFIG_COMPACTION
3190 /* Try memory compaction for high-order allocations before reclaim */
3191 static struct page *
3192 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3193                 unsigned int alloc_flags, const struct alloc_context *ac,
3194                 enum migrate_mode mode, enum compact_result *compact_result)
3195 {
3196         struct page *page;
3197         int contended_compaction;
3198 
3199         if (!order)
3200                 return NULL;
3201 
3202         current->flags |= PF_MEMALLOC;
3203         *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3204                                                 mode, &contended_compaction);
3205         current->flags &= ~PF_MEMALLOC;
3206 
3207         if (*compact_result <= COMPACT_INACTIVE)
3208                 return NULL;
3209 
3210         /*
3211          * At least in one zone compaction wasn't deferred or skipped, so let's
3212          * count a compaction stall
3213          */
3214         count_vm_event(COMPACTSTALL);
3215 
3216         page = get_page_from_freelist(gfp_mask, order,
3217                                         alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3218 
3219         if (page) {
3220                 struct zone *zone = page_zone(page);
3221 
3222                 zone->compact_blockskip_flush = false;
3223                 compaction_defer_reset(zone, order, true);
3224                 count_vm_event(COMPACTSUCCESS);
3225                 return page;
3226         }
3227 
3228         /*
3229          * It's bad if compaction run occurs and fails. The most likely reason
3230          * is that pages exist, but not enough to satisfy watermarks.
3231          */
3232         count_vm_event(COMPACTFAIL);
3233 
3234         /*
3235          * In all zones where compaction was attempted (and not
3236          * deferred or skipped), lock contention has been detected.
3237          * For THP allocation we do not want to disrupt the others
3238          * so we fallback to base pages instead.
3239          */
3240         if (contended_compaction == COMPACT_CONTENDED_LOCK)
3241                 *compact_result = COMPACT_CONTENDED;
3242 
3243         /*
3244          * If compaction was aborted due to need_resched(), we do not
3245          * want to further increase allocation latency, unless it is
3246          * khugepaged trying to collapse.
3247          */
3248         if (contended_compaction == COMPACT_CONTENDED_SCHED
3249                 && !(current->flags & PF_KTHREAD))
3250                 *compact_result = COMPACT_CONTENDED;
3251 
3252         cond_resched();
3253 
3254         return NULL;
3255 }
3256 
3257 static inline bool
3258 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3259                      enum compact_result compact_result, enum migrate_mode *migrate_mode,
3260                      int compaction_retries)
3261 {
3262         int max_retries = MAX_COMPACT_RETRIES;
3263 
3264         if (!order)
3265                 return false;
3266 
3267         /*
3268          * compaction considers all the zone as desperately out of memory
3269          * so it doesn't really make much sense to retry except when the
3270          * failure could be caused by weak migration mode.
3271          */
3272         if (compaction_failed(compact_result)) {
3273                 if (*migrate_mode == MIGRATE_ASYNC) {
3274                         *migrate_mode = MIGRATE_SYNC_LIGHT;
3275                         return true;
3276                 }
3277                 return false;
3278         }
3279 
3280         /*
3281          * make sure the compaction wasn't deferred or didn't bail out early
3282          * due to locks contention before we declare that we should give up.
3283          * But do not retry if the given zonelist is not suitable for
3284          * compaction.
3285          */
3286         if (compaction_withdrawn(compact_result))
3287                 return compaction_zonelist_suitable(ac, order, alloc_flags);
3288 
3289         /*
3290          * !costly requests are much more important than __GFP_REPEAT
3291          * costly ones because they are de facto nofail and invoke OOM
3292          * killer to move on while costly can fail and users are ready
3293          * to cope with that. 1/4 retries is rather arbitrary but we
3294          * would need much more detailed feedback from compaction to
3295          * make a better decision.
3296          */
3297         if (order > PAGE_ALLOC_COSTLY_ORDER)
3298                 max_retries /= 4;
3299         if (compaction_retries <= max_retries)
3300                 return true;
3301 
3302         return false;
3303 }
3304 #else
3305 static inline struct page *
3306 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3307                 unsigned int alloc_flags, const struct alloc_context *ac,
3308                 enum migrate_mode mode, enum compact_result *compact_result)
3309 {
3310         *compact_result = COMPACT_SKIPPED;
3311         return NULL;
3312 }
3313 
3314 static inline bool
3315 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3316                      enum compact_result compact_result,
3317                      enum migrate_mode *migrate_mode,
3318                      int compaction_retries)
3319 {
3320         struct zone *zone;
3321         struct zoneref *z;
3322 
3323         if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3324                 return false;
3325 
3326         /*
3327          * There are setups with compaction disabled which would prefer to loop
3328          * inside the allocator rather than hit the oom killer prematurely.
3329          * Let's give them a good hope and keep retrying while the order-0
3330          * watermarks are OK.
3331          */
3332         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3333                                         ac->nodemask) {
3334                 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3335                                         ac_classzone_idx(ac), alloc_flags))
3336                         return true;
3337         }
3338         return false;
3339 }
3340 #endif /* CONFIG_COMPACTION */
3341 
3342 /* Perform direct synchronous page reclaim */
3343 static int
3344 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3345                                         const struct alloc_context *ac)
3346 {
3347         struct reclaim_state reclaim_state;
3348         int progress;
3349 
3350         cond_resched();
3351 
3352         /* We now go into synchronous reclaim */
3353         cpuset_memory_pressure_bump();
3354         current->flags |= PF_MEMALLOC;
3355         lockdep_set_current_reclaim_state(gfp_mask);
3356         reclaim_state.reclaimed_slab = 0;
3357         current->reclaim_state = &reclaim_state;
3358 
3359         progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3360                                                                 ac->nodemask);
3361 
3362         current->reclaim_state = NULL;
3363         lockdep_clear_current_reclaim_state();
3364         current->flags &= ~PF_MEMALLOC;
3365 
3366         cond_resched();
3367 
3368         return progress;
3369 }
3370 
3371 /* The really slow allocator path where we enter direct reclaim */
3372 static inline struct page *
3373 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3374                 unsigned int alloc_flags, const struct alloc_context *ac,
3375                 unsigned long *did_some_progress)
3376 {
3377         struct page *page = NULL;
3378         bool drained = false;
3379 
3380         *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3381         if (unlikely(!(*did_some_progress)))
3382                 return NULL;
3383 
3384 retry:
3385         page = get_page_from_freelist(gfp_mask, order,
3386                                         alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3387 
3388         /*
3389          * If an allocation failed after direct reclaim, it could be because
3390          * pages are pinned on the per-cpu lists or in high alloc reserves.
3391          * Shrink them them and try again
3392          */
3393         if (!page && !drained) {
3394                 unreserve_highatomic_pageblock(ac);
3395                 drain_all_pages(NULL);
3396                 drained = true;
3397                 goto retry;
3398         }
3399 
3400         return page;
3401 }
3402 
3403 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3404 {
3405         struct zoneref *z;
3406         struct zone *zone;
3407 
3408         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3409                                                 ac->high_zoneidx, ac->nodemask)
3410                 wakeup_kswapd(zone, order, ac_classzone_idx(ac));
3411 }
3412 
3413 static inline unsigned int
3414 gfp_to_alloc_flags(gfp_t gfp_mask)
3415 {
3416         unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3417 
3418         /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3419         BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3420 
3421         /*
3422          * The caller may dip into page reserves a bit more if the caller
3423          * cannot run direct reclaim, or if the caller has realtime scheduling
3424          * policy or is asking for __GFP_HIGH memory.  GFP_ATOMIC requests will
3425          * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3426          */
3427         alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3428 
3429         if (gfp_mask & __GFP_ATOMIC) {
3430                 /*
3431                  * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3432                  * if it can't schedule.
3433                  */
3434                 if (!(gfp_mask & __GFP_NOMEMALLOC))
3435                         alloc_flags |= ALLOC_HARDER;
3436                 /*
3437                  * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3438                  * comment for __cpuset_node_allowed().
3439                  */
3440                 alloc_flags &= ~ALLOC_CPUSET;
3441         } else if (unlikely(rt_task(current)) && !in_interrupt())
3442                 alloc_flags |= ALLOC_HARDER;
3443 
3444         if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3445                 if (gfp_mask & __GFP_MEMALLOC)
3446                         alloc_flags |= ALLOC_NO_WATERMARKS;
3447                 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3448                         alloc_flags |= ALLOC_NO_WATERMARKS;
3449                 else if (!in_interrupt() &&
3450                                 ((current->flags & PF_MEMALLOC) ||
3451                                  unlikely(test_thread_flag(TIF_MEMDIE))))
3452                         alloc_flags |= ALLOC_NO_WATERMARKS;
3453         }
3454 #ifdef CONFIG_CMA
3455         if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3456                 alloc_flags |= ALLOC_CMA;
3457 #endif
3458         return alloc_flags;
3459 }
3460 
3461 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3462 {
3463         return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3464 }
3465 
3466 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3467 {
3468         return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3469 }
3470 
3471 /*
3472  * Maximum number of reclaim retries without any progress before OOM killer
3473  * is consider as the only way to move forward.
3474  */
3475 #define MAX_RECLAIM_RETRIES 16
3476 
3477 /*
3478  * Checks whether it makes sense to retry the reclaim to make a forward progress
3479  * for the given allocation request.
3480  * The reclaim feedback represented by did_some_progress (any progress during
3481  * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3482  * any progress in a row) is considered as well as the reclaimable pages on the
3483  * applicable zone list (with a backoff mechanism which is a function of
3484  * no_progress_loops).
3485  *
3486  * Returns true if a retry is viable or false to enter the oom path.
3487  */
3488 static inline bool
3489 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3490                      struct alloc_context *ac, int alloc_flags,
3491                      bool did_some_progress, int no_progress_loops)
3492 {
3493         struct zone *zone;
3494         struct zoneref *z;
3495 
3496         /*
3497          * Make sure we converge to OOM if we cannot make any progress
3498          * several times in the row.
3499          */
3500         if (no_progress_loops > MAX_RECLAIM_RETRIES)
3501                 return false;
3502 
3503         /*
3504          * Keep reclaiming pages while there is a chance this will lead somewhere.
3505          * If none of the target zones can satisfy our allocation request even
3506          * if all reclaimable pages are considered then we are screwed and have
3507          * to go OOM.
3508          */
3509         for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3510                                         ac->nodemask) {
3511                 unsigned long available;
3512                 unsigned long reclaimable;
3513 
3514                 available = reclaimable = zone_reclaimable_pages(zone);
3515                 available -= DIV_ROUND_UP(no_progress_loops * available,
3516                                           MAX_RECLAIM_RETRIES);
3517                 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3518 
3519                 /*
3520                  * Would the allocation succeed if we reclaimed the whole
3521                  * available?
3522                  */
3523                 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3524                                 ac_classzone_idx(ac), alloc_flags, available)) {
3525                         /*
3526                          * If we didn't make any progress and have a lot of
3527                          * dirty + writeback pages then we should wait for
3528                          * an IO to complete to slow down the reclaim and
3529                          * prevent from pre mature OOM
3530                          */
3531                         if (!did_some_progress) {
3532                                 unsigned long writeback;
3533                                 unsigned long dirty;
3534 
3535                                 writeback = zone_page_state_snapshot(zone,
3536                                                                      NR_WRITEBACK);
3537                                 dirty = zone_page_state_snapshot(zone, NR_FILE_DIRTY);
3538 
3539                                 if (2*(writeback + dirty) > reclaimable) {
3540                                         congestion_wait(BLK_RW_ASYNC, HZ/10);
3541                                         return true;
3542                                 }
3543                         }
3544 
3545                         /*
3546                          * Memory allocation/reclaim might be called from a WQ
3547                          * context and the current implementation of the WQ
3548                          * concurrency control doesn't recognize that
3549                          * a particular WQ is congested if the worker thread is
3550                          * looping without ever sleeping. Therefore we have to
3551                          * do a short sleep here rather than calling
3552                          * cond_resched().
3553                          */
3554                         if (current->flags & PF_WQ_WORKER)
3555                                 schedule_timeout_uninterruptible(1);
3556                         else
3557                                 cond_resched();
3558 
3559                         return true;
3560                 }
3561         }
3562 
3563         return false;
3564 }
3565 
3566 static inline struct page *
3567 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3568                                                 struct alloc_context *ac)
3569 {
3570         bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3571         struct page *page = NULL;
3572         unsigned int alloc_flags;
3573         unsigned long did_some_progress;
3574         enum migrate_mode migration_mode = MIGRATE_ASYNC;
3575         enum compact_result compact_result;
3576         int compaction_retries = 0;
3577         int no_progress_loops = 0;
3578 
3579         /*
3580          * In the slowpath, we sanity check order to avoid ever trying to
3581          * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3582          * be using allocators in order of preference for an area that is
3583          * too large.
3584          */
3585         if (order >= MAX_ORDER) {
3586                 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3587                 return NULL;
3588         }
3589 
3590         /*
3591          * We also sanity check to catch abuse of atomic reserves being used by
3592          * callers that are not in atomic context.
3593          */
3594         if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3595                                 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3596                 gfp_mask &= ~__GFP_ATOMIC;
3597 
3598 retry:
3599         if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3600                 wake_all_kswapds(order, ac);
3601 
3602         /*
3603          * OK, we're below the kswapd watermark and have kicked background
3604          * reclaim. Now things get more complex, so set up alloc_flags according
3605          * to how we want to proceed.
3606          */
3607         alloc_flags = gfp_to_alloc_flags(gfp_mask);
3608 
3609         /*
3610          * Reset the zonelist iterators if memory policies can be ignored.
3611          * These allocations are high priority and system rather than user
3612          * orientated.
3613          */
3614         if ((alloc_flags & ALLOC_NO_WATERMARKS) || !(alloc_flags & ALLOC_CPUSET)) {
3615                 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3616                 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3617                                         ac->high_zoneidx, ac->nodemask);
3618         }
3619 
3620         /* This is the last chance, in general, before the goto nopage. */
3621         page = get_page_from_freelist(gfp_mask, order,
3622                                 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3623         if (page)
3624                 goto got_pg;
3625 
3626         /* Allocate without watermarks if the context allows */
3627         if (alloc_flags & ALLOC_NO_WATERMARKS) {
3628                 page = get_page_from_freelist(gfp_mask, order,
3629                                                 ALLOC_NO_WATERMARKS, ac);
3630                 if (page)
3631                         goto got_pg;
3632         }
3633 
3634         /* Caller is not willing to reclaim, we can't balance anything */
3635         if (!can_direct_reclaim) {
3636                 /*
3637                  * All existing users of the __GFP_NOFAIL are blockable, so warn
3638                  * of any new users that actually allow this type of allocation
3639                  * to fail.
3640                  */
3641                 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3642                 goto nopage;
3643         }
3644 
3645         /* Avoid recursion of direct reclaim */
3646         if (current->flags & PF_MEMALLOC) {
3647                 /*
3648                  * __GFP_NOFAIL request from this context is rather bizarre
3649                  * because we cannot reclaim anything and only can loop waiting
3650                  * for somebody to do a work for us.
3651                  */
3652                 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3653                         cond_resched();
3654                         goto retry;
3655                 }
3656                 goto nopage;
3657         }
3658 
3659         /* Avoid allocations with no watermarks from looping endlessly */
3660         if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3661                 goto nopage;
3662 
3663         /*
3664          * Try direct compaction. The first pass is asynchronous. Subsequent
3665          * attempts after direct reclaim are synchronous
3666          */
3667         page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3668                                         migration_mode,
3669                                         &compact_result);
3670         if (page)
3671                 goto got_pg;
3672 
3673         /* Checks for THP-specific high-order allocations */
3674         if (is_thp_gfp_mask(gfp_mask)) {
3675                 /*
3676                  * If compaction is deferred for high-order allocations, it is
3677                  * because sync compaction recently failed. If this is the case
3678                  * and the caller requested a THP allocation, we do not want
3679                  * to heavily disrupt the system, so we fail the allocation
3680                  * instead of entering direct reclaim.
3681                  */
3682                 if (compact_result == COMPACT_DEFERRED)
3683                         goto nopage;
3684 
3685                 /*
3686                  * Compaction is contended so rather back off than cause
3687                  * excessive stalls.
3688                  */
3689                 if(compact_result == COMPACT_CONTENDED)
3690                         goto nopage;
3691         }
3692 
3693         if (order && compaction_made_progress(compact_result))
3694                 compaction_retries++;
3695 
3696         /* Try direct reclaim and then allocating */
3697         page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3698                                                         &did_some_progress);
3699         if (page)
3700                 goto got_pg;
3701 
3702         /* Do not loop if specifically requested */
3703         if (gfp_mask & __GFP_NORETRY)
3704                 goto noretry;
3705 
3706         /*
3707          * Do not retry costly high order allocations unless they are
3708          * __GFP_REPEAT
3709          */
3710         if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3711                 goto noretry;
3712 
3713         /*
3714          * Costly allocations might have made a progress but this doesn't mean
3715          * their order will become available due to high fragmentation so
3716          * always increment the no progress counter for them
3717          */
3718         if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3719                 no_progress_loops = 0;
3720         else
3721                 no_progress_loops++;
3722 
3723         if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3724                                  did_some_progress > 0, no_progress_loops))
3725                 goto retry;
3726 
3727         /*
3728          * It doesn't make any sense to retry for the compaction if the order-0
3729          * reclaim is not able to make any progress because the current
3730          * implementation of the compaction depends on the sufficient amount
3731          * of free memory (see __compaction_suitable)
3732          */
3733         if (did_some_progress > 0 &&
3734                         should_compact_retry(ac, order, alloc_flags,
3735                                 compact_result, &migration_mode,
3736                                 compaction_retries))
3737                 goto retry;
3738 
3739         /* Reclaim has failed us, start killing things */
3740         page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3741         if (page)
3742                 goto got_pg;
3743 
3744         /* Retry as long as the OOM killer is making progress */
3745         if (did_some_progress) {
3746                 no_progress_loops = 0;
3747                 goto retry;
3748         }
3749 
3750 noretry:
3751         /*
3752          * High-order allocations do not necessarily loop after direct reclaim
3753          * and reclaim/compaction depends on compaction being called after
3754          * reclaim so call directly if necessary.
3755          * It can become very expensive to allocate transparent hugepages at
3756          * fault, so use asynchronous memory compaction for THP unless it is
3757          * khugepaged trying to collapse. All other requests should tolerate
3758          * at least light sync migration.
3759          */
3760         if (is_thp_gfp_mask(gfp_mask) && !(current->flags & PF_KTHREAD))
3761                 migration_mode = MIGRATE_ASYNC;
3762         else
3763                 migration_mode = MIGRATE_SYNC_LIGHT;
3764         page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3765                                             ac, migration_mode,
3766                                             &compact_result);
3767         if (page)
3768                 goto got_pg;
3769 nopage:
3770         warn_alloc_failed(gfp_mask, order, NULL);
3771 got_pg:
3772         return page;
3773 }
3774 
3775 /*
3776  * This is the 'heart' of the zoned buddy allocator.
3777  */
3778 struct page *
3779 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3780                         struct zonelist *zonelist, nodemask_t *nodemask)
3781 {
3782         struct page *page;
3783         unsigned int cpuset_mems_cookie;
3784         unsigned int alloc_flags = ALLOC_WMARK_LOW|ALLOC_FAIR;
3785         gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3786         struct alloc_context ac = {
3787                 .high_zoneidx = gfp_zone(gfp_mask),
3788                 .zonelist = zonelist,
3789                 .nodemask = nodemask,
3790                 .migratetype = gfpflags_to_migratetype(gfp_mask),
3791         };
3792 
3793         if (cpusets_enabled()) {
3794                 alloc_mask |= __GFP_HARDWALL;
3795                 alloc_flags |= ALLOC_CPUSET;
3796                 if (!ac.nodemask)
3797                         ac.nodemask = &cpuset_current_mems_allowed;
3798         }
3799 
3800         gfp_mask &= gfp_allowed_mask;
3801 
3802         lockdep_trace_alloc(gfp_mask);
3803 
3804         might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3805 
3806         if (should_fail_alloc_page(gfp_mask, order))
3807                 return NULL;
3808 
3809         /*
3810          * Check the zones suitable for the gfp_mask contain at least one
3811          * valid zone. It's possible to have an empty zonelist as a result
3812          * of __GFP_THISNODE and a memoryless node
3813          */
3814         if (unlikely(!zonelist->_zonerefs->zone))
3815                 return NULL;
3816 
3817         if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3818                 alloc_flags |= ALLOC_CMA;
3819 
3820 retry_cpuset:
3821         cpuset_mems_cookie = read_mems_allowed_begin();
3822 
3823         /* Dirty zone balancing only done in the fast path */
3824         ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3825 
3826         /*
3827          * The preferred zone is used for statistics but crucially it is
3828          * also used as the starting point for the zonelist iterator. It
3829          * may get reset for allocations that ignore memory policies.
3830          */
3831         ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3832                                         ac.high_zoneidx, ac.nodemask);
3833         if (!ac.preferred_zoneref) {
3834                 page = NULL;
3835                 goto no_zone;
3836         }
3837 
3838         /* First allocation attempt */
3839         page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3840         if (likely(page))
3841                 goto out;
3842 
3843         /*
3844          * Runtime PM, block IO and its error handling path can deadlock
3845          * because I/O on the device might not complete.
3846          */
3847         alloc_mask = memalloc_noio_flags(gfp_mask);
3848         ac.spread_dirty_pages = false;
3849 
3850         /*
3851          * Restore the original nodemask if it was potentially replaced with
3852          * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3853          */
3854         if (cpusets_enabled())
3855                 ac.nodemask = nodemask;
3856         page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3857 
3858 no_zone:
3859         /*
3860          * When updating a task's mems_allowed, it is possible to race with
3861          * parallel threads in such a way that an allocation can fail while
3862          * the mask is being updated. If a page allocation is about to fail,
3863          * check if the cpuset changed during allocation and if so, retry.
3864          */
3865         if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) {
3866                 alloc_mask = gfp_mask;
3867                 goto retry_cpuset;
3868         }
3869 
3870 out:
3871         if (kmemcheck_enabled && page)
3872                 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3873 
3874         trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3875 
3876         return page;
3877 }
3878 EXPORT_SYMBOL(__alloc_pages_nodemask);
3879 
3880 /*
3881  * Common helper functions.
3882  */
3883 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3884 {
3885         struct page *page;
3886 
3887         /*
3888          * __get_free_pages() returns a 32-bit address, which cannot represent
3889          * a highmem page
3890          */
3891         VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3892 
3893         page = alloc_pages(gfp_mask, order);
3894         if (!page)
3895                 return 0;
3896         return (unsigned long) page_address(page);
3897 }
3898 EXPORT_SYMBOL(__get_free_pages);
3899 
3900 unsigned long get_zeroed_page(gfp_t gfp_mask)
3901 {
3902         return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3903 }
3904 EXPORT_SYMBOL(get_zeroed_page);
3905 
3906 void __free_pages(struct page *page, unsigned int order)
3907 {
3908         if (put_page_testzero(page)) {
3909                 if (order == 0)
3910                         free_hot_cold_page(page, false);
3911                 else
3912                         __free_pages_ok(page, order);
3913         }
3914 }
3915 
3916 EXPORT_SYMBOL(__free_pages);
3917 
3918 void free_pages(unsigned long addr, unsigned int order)
3919 {
3920         if (addr != 0) {
3921                 VM_BUG_ON(!virt_addr_valid((void *)addr));
3922                 __free_pages(virt_to_page((void *)addr), order);
3923         }
3924 }
3925 
3926 EXPORT_SYMBOL(free_pages);
3927 
3928 /*
3929  * Page Fragment:
3930  *  An arbitrary-length arbitrary-offset area of memory which resides
3931  *  within a 0 or higher order page.  Multiple fragments within that page
3932  *  are individually refcounted, in the page's reference counter.
3933  *
3934  * The page_frag functions below provide a simple allocation framework for
3935  * page fragments.  This is used by the network stack and network device
3936  * drivers to provide a backing region of memory for use as either an
3937  * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3938  */
3939 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3940                                        gfp_t gfp_mask)
3941 {
3942         struct page *page = NULL;
3943         gfp_t gfp = gfp_mask;
3944 
3945 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3946         gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3947                     __GFP_NOMEMALLOC;
3948         page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3949                                 PAGE_FRAG_CACHE_MAX_ORDER);
3950         nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3951 #endif
3952         if (unlikely(!page))
3953                 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3954 
3955         nc->va = page ? page_address(page) : NULL;
3956 
3957         return page;
3958 }
3959 
3960 void *__alloc_page_frag(struct page_frag_cache *nc,
3961                         unsigned int fragsz, gfp_t gfp_mask)
3962 {
3963         unsigned int size = PAGE_SIZE;
3964         struct page *page;
3965         int offset;
3966 
3967         if (unlikely(!nc->va)) {
3968 refill:
3969                 page = __page_frag_refill(nc, gfp_mask);
3970                 if (!page)
3971                         return NULL;
3972 
3973 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3974                 /* if size can vary use size else just use PAGE_SIZE */
3975                 size = nc->size;
3976 #endif
3977                 /* Even if we own the page, we do not use atomic_set().
3978                  * This would break get_page_unless_zero() users.
3979                  */
3980                 page_ref_add(page, size - 1);
3981 
3982                 /* reset page count bias and offset to start of new frag */
3983                 nc->pfmemalloc = page_is_pfmemalloc(page);
3984                 nc->pagecnt_bias = size;
3985                 nc->offset = size;
3986         }
3987 
3988         offset = nc->offset - fragsz;
3989         if (unlikely(offset < 0)) {
3990                 page = virt_to_page(nc->va);
3991 
3992                 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3993                         goto refill;
3994 
3995 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3996                 /* if size can vary use size else just use PAGE_SIZE */
3997                 size = nc->size;
3998 #endif
3999                 /* OK, page count is 0, we can safely set it */
4000                 set_page_count(page, size);
4001 
4002                 /* reset page count bias and offset to start of new frag */
4003                 nc->pagecnt_bias = size;
4004                 offset = size - fragsz;
4005         }
4006 
4007         nc->pagecnt_bias--;
4008         nc->offset = offset;
4009 
4010         return nc->va + offset;
4011 }
4012 EXPORT_SYMBOL(__alloc_page_frag);
4013 
4014 /*
4015  * Frees a page fragment allocated out of either a compound or order 0 page.
4016  */
4017 void __free_page_frag(void *addr)
4018 {
4019         struct page *page = virt_to_head_page(addr);
4020 
4021         if (unlikely(put_page_testzero(page)))
4022                 __free_pages_ok(page, compound_order(page));
4023 }
4024 EXPORT_SYMBOL(__free_page_frag);
4025 
4026 /*
4027  * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
4028  * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
4029  * equivalent to alloc_pages.
4030  *
4031  * It should be used when the caller would like to use kmalloc, but since the
4032  * allocation is large, it has to fall back to the page allocator.
4033  */
4034 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
4035 {
4036         struct page *page;
4037 
4038         page = alloc_pages(gfp_mask, order);
4039         if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4040                 __free_pages(page, order);
4041                 page = NULL;
4042         }
4043         return page;
4044 }
4045 
4046 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
4047 {
4048         struct page *page;
4049 
4050         page = alloc_pages_node(nid, gfp_mask, order);
4051         if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
4052                 __free_pages(page, order);
4053                 page = NULL;
4054         }
4055         return page;
4056 }
4057 
4058 /*
4059  * __free_kmem_pages and free_kmem_pages will free pages allocated with
4060  * alloc_kmem_pages.
4061  */
4062 void __free_kmem_pages(struct page *page, unsigned int order)
4063 {
4064         memcg_kmem_uncharge(page, order);
4065         __free_pages(page, order);
4066 }
4067 
4068 void free_kmem_pages(unsigned long addr, unsigned int order)
4069 {
4070         if (addr != 0) {
4071                 VM_BUG_ON(!virt_addr_valid((void *)addr));
4072                 __free_kmem_pages(virt_to_page((void *)addr), order);
4073         }
4074 }
4075 
4076 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4077                 size_t size)
4078 {
4079         if (addr) {
4080                 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4081                 unsigned long used = addr + PAGE_ALIGN(size);
4082 
4083                 split_page(virt_to_page((void *)addr), order);
4084                 while (used < alloc_end) {
4085                         free_page(used);
4086                         used += PAGE_SIZE;
4087                 }
4088         }
4089         return (void *)addr;
4090 }
4091 
4092 /**
4093  * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4094  * @size: the number of bytes to allocate
4095  * @gfp_mask: GFP flags for the allocation
4096  *
4097  * This function is similar to alloc_pages(), except that it allocates the
4098  * minimum number of pages to satisfy the request.  alloc_pages() can only
4099  * allocate memory in power-of-two pages.
4100  *
4101  * This function is also limited by MAX_ORDER.
4102  *
4103  * Memory allocated by this function must be released by free_pages_exact().
4104  */
4105 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4106 {
4107         unsigned int order = get_order(size);
4108         unsigned long addr;
4109 
4110         addr = __get_free_pages(gfp_mask, order);
4111         return make_alloc_exact(addr, order, size);
4112 }
4113 EXPORT_SYMBOL(alloc_pages_exact);
4114 
4115 /**
4116  * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4117  *                         pages on a node.
4118  * @nid: the preferred node ID where memory should be allocated
4119  * @size: the number of bytes to allocate
4120  * @gfp_mask: GFP flags for the allocation
4121  *
4122  * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4123  * back.
4124  */
4125 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4126 {
4127         unsigned int order = get_order(size);
4128         struct page *p = alloc_pages_node(nid, gfp_mask, order);
4129         if (!p)
4130                 return NULL;
4131         return make_alloc_exact((unsigned long)page_address(p), order, size);
4132 }
4133 
4134 /**
4135  * free_pages_exact - release memory allocated via alloc_pages_exact()
4136  * @virt: the value returned by alloc_pages_exact.
4137  * @size: size of allocation, same value as passed to alloc_pages_exact().
4138  *
4139  * Release the memory allocated by a previous call to alloc_pages_exact.
4140  */
4141 void free_pages_exact(void *virt, size_t size)
4142 {
4143         unsigned long addr = (unsigned long)virt;
4144         unsigned long end = addr + PAGE_ALIGN(size);
4145 
4146         while (addr < end) {
4147                 free_page(addr);
4148                 addr += PAGE_SIZE;
4149         }
4150 }
4151 EXPORT_SYMBOL(free_pages_exact);
4152 
4153 /**
4154  * nr_free_zone_pages - count number of pages beyond high watermark
4155  * @offset: The zone index of the highest zone
4156  *
4157  * nr_free_zone_pages() counts the number of counts pages which are beyond the
4158  * high watermark within all zones at or below a given zone index.  For each
4159  * zone, the number of pages is calculated as:
4160  *     managed_pages - high_pages
4161  */
4162 static unsigned long nr_free_zone_pages(int offset)
4163 {
4164         struct zoneref *z;
4165         struct zone *zone;
4166 
4167         /* Just pick one node, since fallback list is circular */
4168         unsigned long sum = 0;
4169 
4170         struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4171 
4172         for_each_zone_zonelist(zone, z, zonelist, offset) {
4173                 unsigned long size = zone->managed_pages;
4174                 unsigned long high = high_wmark_pages(zone);
4175                 if (size > high)
4176                         sum += size - high;
4177         }
4178 
4179         return sum;
4180 }
4181 
4182 /**
4183  * nr_free_buffer_pages - count number of pages beyond high watermark
4184  *
4185  * nr_free_buffer_pages() counts the number of pages which are beyond the high
4186  * watermark within ZONE_DMA and ZONE_NORMAL.
4187  */
4188 unsigned long nr_free_buffer_pages(void)
4189 {
4190         return nr_free_zone_pages(gfp_zone(GFP_USER));
4191 }
4192 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4193 
4194 /**
4195  * nr_free_pagecache_pages - count number of pages beyond high watermark
4196  *
4197  * nr_free_pagecache_pages() counts the number of pages which are beyond the
4198  * high watermark within all zones.
4199  */
4200 unsigned long nr_free_pagecache_pages(void)
4201 {
4202         return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4203 }
4204 
4205 static inline void show_node(struct zone *zone)
4206 {
4207         if (IS_ENABLED(CONFIG_NUMA))
4208                 printk("Node %d ", zone_to_nid(zone));
4209 }
4210 
4211 long si_mem_available(void)
4212 {
4213         long available;
4214         unsigned long pagecache;
4215         unsigned long wmark_low = 0;
4216         unsigned long pages[NR_LRU_LISTS];
4217         struct zone *zone;
4218         int lru;
4219 
4220         for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4221                 pages[lru] = global_page_state(NR_LRU_BASE + lru);
4222 
4223         for_each_zone(zone)
4224                 wmark_low += zone->watermark[WMARK_LOW];
4225 
4226         /*
4227          * Estimate the amount of memory available for userspace allocations,
4228          * without causing swapping.
4229          */
4230         available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4231 
4232         /*
4233          * Not all the page cache can be freed, otherwise the system will
4234          * start swapping. Assume at least half of the page cache, or the
4235          * low watermark worth of cache, needs to stay.
4236          */
4237         pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4238         pagecache -= min(pagecache / 2, wmark_low);
4239         available += pagecache;
4240 
4241         /*
4242          * Part of the reclaimable slab consists of items that are in use,
4243          * and cannot be freed. Cap this estimate at the low watermark.
4244          */
4245         available += global_page_state(NR_SLAB_RECLAIMABLE) -
4246                      min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4247 
4248         if (available < 0)
4249                 available = 0;
4250         return available;
4251 }
4252 EXPORT_SYMBOL_GPL(si_mem_available);
4253 
4254 void si_meminfo(struct sysinfo *val)
4255 {
4256         val->totalram = totalram_pages;
4257         val->sharedram = global_page_state(NR_SHMEM);
4258         val->freeram = global_page_state(NR_FREE_PAGES);
4259         val->bufferram = nr_blockdev_pages();
4260         val->totalhigh = totalhigh_pages;
4261         val->freehigh = nr_free_highpages();
4262         val->mem_unit = PAGE_SIZE;
4263 }
4264 
4265 EXPORT_SYMBOL(si_meminfo);
4266 
4267 #ifdef CONFIG_NUMA
4268 void si_meminfo_node(struct sysinfo *val, int nid)
4269 {
4270         int zone_type;          /* needs to be signed */
4271         unsigned long managed_pages = 0;
4272         unsigned long managed_highpages = 0;
4273         unsigned long free_highpages = 0;
4274         pg_data_t *pgdat = NODE_DATA(nid);
4275 
4276         for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4277                 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4278         val->totalram = managed_pages;
4279         val->sharedram = node_page_state(nid, NR_SHMEM);
4280         val->freeram = node_page_state(nid, NR_FREE_PAGES);
4281 #ifdef CONFIG_HIGHMEM
4282         for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4283                 struct zone *zone = &pgdat->node_zones[zone_type];
4284 
4285                 if (is_highmem(zone)) {
4286                         managed_highpages += zone->managed_pages;
4287                         free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4288                 }
4289         }
4290         val->totalhigh = managed_highpages;
4291         val->freehigh = free_highpages;
4292 #else
4293         val->totalhigh = managed_highpages;
4294         val->freehigh = free_highpages;
4295 #endif
4296         val->mem_unit = PAGE_SIZE;
4297 }
4298 #endif
4299 
4300 /*
4301  * Determine whether the node should be displayed or not, depending on whether
4302  * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4303  */
4304 bool skip_free_areas_node(unsigned int flags, int nid)
4305 {
4306         bool ret = false;
4307         unsigned int cpuset_mems_cookie;
4308 
4309         if (!(flags & SHOW_MEM_FILTER_NODES))
4310                 goto out;
4311 
4312         do {
4313                 cpuset_mems_cookie = read_mems_allowed_begin();
4314                 ret = !node_isset(nid, cpuset_current_mems_allowed);
4315         } while (read_mems_allowed_retry(cpuset_mems_cookie));
4316 out:
4317         return ret;
4318 }
4319 
4320 #define K(x) ((x) << (PAGE_SHIFT-10))
4321 
4322 static void show_migration_types(unsigned char type)
4323 {
4324         static const char types[MIGRATE_TYPES] = {
4325                 [MIGRATE_UNMOVABLE]     = 'U',
4326                 [MIGRATE_MOVABLE]       = 'M',
4327                 [MIGRATE_RECLAIMABLE]   = 'E',
4328                 [MIGRATE_HIGHATOMIC]    = 'H',
4329 #ifdef CONFIG_CMA
4330                 [MIGRATE_CMA]           = 'C',
4331 #endif
4332 #ifdef CONFIG_MEMORY_ISOLATION
4333                 [MIGRATE_ISOLATE]       = 'I',
4334 #endif
4335         };
4336         char tmp[MIGRATE_TYPES + 1];
4337         char *p = tmp;
4338         int i;
4339 
4340         for (i = 0; i < MIGRATE_TYPES; i++) {
4341                 if (type & (1 << i))
4342                         *p++ = types[i];
4343         }
4344 
4345         *p = '\0';
4346         printk("(%s) ", tmp);
4347 }
4348 
4349 /*
4350  * Show free area list (used inside shift_scroll-lock stuff)
4351  * We also calculate the percentage fragmentation. We do this by counting the
4352  * memory on each free list with the exception of the first item on the list.
4353  *
4354  * Bits in @filter:
4355  * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4356  *   cpuset.
4357  */
4358 void show_free_areas(unsigned int filter)
4359 {
4360         unsigned long free_pcp = 0;
4361         int cpu;
4362         struct zone *zone;
4363 
4364         for_each_populated_zone(zone) {
4365                 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4366                         continue;
4367 
4368                 for_each_online_cpu(cpu)
4369                         free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4370         }
4371 
4372         printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4373                 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4374                 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4375                 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4376                 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4377                 " free:%lu free_pcp:%lu free_cma:%lu\n",
4378                 global_page_state(NR_ACTIVE_ANON),
4379                 global_page_state(NR_INACTIVE_ANON),
4380                 global_page_state(NR_ISOLATED_ANON),
4381                 global_page_state(NR_ACTIVE_FILE),
4382                 global_page_state(NR_INACTIVE_FILE),
4383                 global_page_state(NR_ISOLATED_FILE),
4384                 global_page_state(NR_UNEVICTABLE),
4385                 global_page_state(NR_FILE_DIRTY),
4386                 global_page_state(NR_WRITEBACK),
4387                 global_page_state(NR_UNSTABLE_NFS),
4388                 global_page_state(NR_SLAB_RECLAIMABLE),
4389                 global_page_state(NR_SLAB_UNRECLAIMABLE),
4390                 global_page_state(NR_FILE_MAPPED),
4391                 global_page_state(NR_SHMEM),
4392                 global_page_state(NR_PAGETABLE),
4393                 global_page_state(NR_BOUNCE),
4394                 global_page_state(NR_FREE_PAGES),
4395                 free_pcp,
4396                 global_page_state(NR_FREE_CMA_PAGES));
4397 
4398         for_each_populated_zone(zone) {
4399                 int i;
4400 
4401                 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4402                         continue;
4403 
4404                 free_pcp = 0;
4405                 for_each_online_cpu(cpu)
4406                         free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4407 
4408                 show_node(zone);
4409                 printk("%s"
4410                         " free:%lukB"
4411                         " min:%lukB"
4412                         " low:%lukB"
4413                         " high:%lukB"
4414                         " active_anon:%lukB"
4415                         " inactive_anon:%lukB"
4416                         " active_file:%lukB"
4417                         " inactive_file:%lukB"
4418                         " unevictable:%lukB"
4419                         " isolated(anon):%lukB"
4420                         " isolated(file):%lukB"
4421                         " present:%lukB"
4422                         " managed:%lukB"
4423                         " mlocked:%lukB"
4424                         " dirty:%lukB"
4425                         " writeback:%lukB"
4426                         " mapped:%lukB"
4427                         " shmem:%lukB"
4428                         " slab_reclaimable:%lukB"
4429                         " slab_unreclaimable:%lukB"
4430                         " kernel_stack:%lukB"
4431                         " pagetables:%lukB"
4432                         " unstable:%lukB"
4433                         " bounce:%lukB"
4434                         " free_pcp:%lukB"
4435                         " local_pcp:%ukB"
4436                         " free_cma:%lukB"
4437                         " writeback_tmp:%lukB"
4438                         " pages_scanned:%lu"
4439                         " all_unreclaimable? %s"
4440                         "\n",
4441                         zone->name,
4442                         K(zone_page_state(zone, NR_FREE_PAGES)),
4443                         K(min_wmark_pages(zone)),
4444                         K(low_wmark_pages(zone)),
4445                         K(high_wmark_pages(zone)),
4446                         K(zone_page_state(zone, NR_ACTIVE_ANON)),
4447                         K(zone_page_state(zone, NR_INACTIVE_ANON)),
4448                         K(zone_page_state(zone, NR_ACTIVE_FILE)),
4449                         K(zone_page_state(zone, NR_INACTIVE_FILE)),
4450                         K(zone_page_state(zone, NR_UNEVICTABLE)),
4451                         K(zone_page_state(zone, NR_ISOLATED_ANON)),
4452                         K(zone_page_state(zone, NR_ISOLATED_FILE)),
4453                         K(zone->present_pages),
4454                         K(zone->managed_pages),
4455                         K(zone_page_state(zone, NR_MLOCK)),
4456                         K(zone_page_state(zone, NR_FILE_DIRTY)),
4457                         K(zone_page_state(zone, NR_WRITEBACK)),
4458                         K(zone_page_state(zone, NR_FILE_MAPPED)),
4459                         K(zone_page_state(zone, NR_SHMEM)),
4460                         K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4461                         K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4462                         zone_page_state(zone, NR_KERNEL_STACK) *
4463                                 THREAD_SIZE / 1024,
4464                         K(zone_page_state(zone, NR_PAGETABLE)),
4465                         K(zone_page_state(zone, NR_UNSTABLE_NFS)),
4466                         K(zone_page_state(zone, NR_BOUNCE)),
4467                         K(free_pcp),
4468                         K(this_cpu_read(zone->pageset->pcp.count)),
4469                         K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
4470                         K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
4471                         K(zone_page_state(zone, NR_PAGES_SCANNED)),
4472                         (!zone_reclaimable(zone) ? "yes" : "no")
4473                         );
4474                 printk("lowmem_reserve[]:");
4475                 for (i = 0; i < MAX_NR_ZONES; i++)
4476                         printk(" %ld", zone->lowmem_reserve[i]);
4477                 printk("\n");
4478         }
4479 
4480         for_each_populated_zone(zone) {
4481                 unsigned int order;
4482                 unsigned long nr[MAX_ORDER], flags, total = 0;
4483                 unsigned char types[MAX_ORDER];
4484 
4485                 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4486                         continue;
4487                 show_node(zone);
4488                 printk("%s: ", zone->name);
4489 
4490                 spin_lock_irqsave(&zone->lock, flags);
4491                 for (order = 0; order < MAX_ORDER; order++) {
4492                         struct free_area *area = &zone->free_area[order];
4493                         int type;
4494 
4495                         nr[order] = area->nr_free;
4496                         total += nr[order] << order;
4497 
4498                         types[order] = 0;
4499                         for (type = 0; type < MIGRATE_TYPES; type++) {
4500                                 if (!list_empty(&area->free_list[type]))
4501                                         types[order] |= 1 << type;
4502                         }
4503                 }
4504                 spin_unlock_irqrestore(&zone->lock, flags);
4505                 for (order = 0; order < MAX_ORDER; order++) {
4506                         printk("%lu*%lukB ", nr[order], K(1UL) << order);
4507                         if (nr[order])
4508                                 show_migration_types(types[order]);
4509                 }
4510                 printk("= %lukB\n", K(total));
4511         }
4512 
4513         hugetlb_show_meminfo();
4514 
4515         printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4516 
4517         show_swap_cache_info();
4518 }
4519 
4520 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4521 {
4522         zoneref->zone = zone;
4523         zoneref->zone_idx = zone_idx(zone);
4524 }
4525 
4526 /*
4527  * Builds allocation fallback zone lists.
4528  *
4529  * Add all populated zones of a node to the zonelist.
4530  */
4531 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4532                                 int nr_zones)
4533 {
4534         struct zone *zone;
4535         enum zone_type zone_type = MAX_NR_ZONES;
4536 
4537         do {
4538                 zone_type--;
4539                 zone = pgdat->node_zones + zone_type;
4540                 if (populated_zone(zone)) {
4541                         zoneref_set_zone(zone,
4542                                 &zonelist->_zonerefs[nr_zones++]);
4543                         check_highest_zone(zone_type);
4544                 }
4545         } while (zone_type);
4546 
4547         return nr_zones;
4548 }
4549 
4550 
4551 /*
4552  *  zonelist_order:
4553  *  0 = automatic detection of better ordering.
4554  *  1 = order by ([node] distance, -zonetype)
4555  *  2 = order by (-zonetype, [node] distance)
4556  *
4557  *  If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4558  *  the same zonelist. So only NUMA can configure this param.
4559  */
4560 #define ZONELIST_ORDER_DEFAULT  0
4561 #define ZONELIST_ORDER_NODE     1
4562 #define ZONELIST_ORDER_ZONE     2
4563 
4564 /* zonelist order in the kernel.
4565  * set_zonelist_order() will set this to NODE or ZONE.
4566  */
4567 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4568 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4569 
4570 
4571 #ifdef CONFIG_NUMA
4572 /* The value user specified ....changed by config */
4573 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4574 /* string for sysctl */
4575 #define NUMA_ZONELIST_ORDER_LEN 16
4576 char numa_zonelist_order[16] = "default";
4577 
4578 /*
4579  * interface for configure zonelist ordering.
4580  * command line option "numa_zonelist_order"
4581  *      = "[dD]efault   - default, automatic configuration.
4582  *      = "[nN]ode      - order by node locality, then by zone within node
4583  *      = "[zZ]one      - order by zone, then by locality within zone
4584  */
4585 
4586 static int __parse_numa_zonelist_order(char *s)
4587 {
4588         if (*s == 'd' || *s == 'D') {
4589                 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4590         } else if (*s == 'n' || *s == 'N') {
4591                 user_zonelist_order = ZONELIST_ORDER_NODE;
4592         } else if (*s == 'z' || *s == 'Z') {
4593                 user_zonelist_order = ZONELIST_ORDER_ZONE;
4594         } else {
4595                 pr_warn("Ignoring invalid numa_zonelist_order value:  %s\n", s);
4596                 return -EINVAL;
4597         }
4598         return 0;
4599 }
4600 
4601 static __init int setup_numa_zonelist_order(char *s)
4602 {
4603         int ret;
4604 
4605         if (!s)
4606                 return 0;
4607 
4608         ret = __parse_numa_zonelist_order(s);
4609         if (ret == 0)
4610                 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4611 
4612         return ret;
4613 }
4614 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4615 
4616 /*
4617  * sysctl handler for numa_zonelist_order
4618  */
4619 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4620                 void __user *buffer, size_t *length,
4621                 loff_t *ppos)
4622 {
4623         char saved_string[NUMA_ZONELIST_ORDER_LEN];
4624         int ret;
4625         static DEFINE_MUTEX(zl_order_mutex);
4626 
4627         mutex_lock(&zl_order_mutex);
4628         if (write) {
4629                 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4630                         ret = -EINVAL;
4631                         goto out;
4632                 }
4633                 strcpy(saved_string, (char *)table->data);
4634         }
4635         ret = proc_dostring(table, write, buffer, length, ppos);
4636         if (ret)
4637                 goto out;
4638         if (write) {
4639                 int oldval = user_zonelist_order;
4640 
4641                 ret = __parse_numa_zonelist_order((char *)table->data);
4642                 if (ret) {
4643                         /*
4644                          * bogus value.  restore saved string
4645                          */
4646                         strncpy((char *)table->data, saved_string,
4647                                 NUMA_ZONELIST_ORDER_LEN);
4648                         user_zonelist_order = oldval;
4649                 } else if (oldval != user_zonelist_order) {
4650                         mutex_lock(&zonelists_mutex);
4651                         build_all_zonelists(NULL, NULL);
4652                         mutex_unlock(&zonelists_mutex);
4653                 }
4654         }
4655 out:
4656         mutex_unlock(&zl_order_mutex);
4657         return ret;
4658 }
4659 
4660 
4661 #define MAX_NODE_LOAD (nr_online_nodes)
4662 static int node_load[MAX_NUMNODES];
4663 
4664 /**
4665  * find_next_best_node - find the next node that should appear in a given node's fallback list
4666  * @node: node whose fallback list we're appending
4667  * @used_node_mask: nodemask_t of already used nodes
4668  *
4669  * We use a number of factors to determine which is the next node that should
4670  * appear on a given node's fallback list.  The node should not have appeared
4671  * already in @node's fallback list, and it should be the next closest node
4672  * according to the distance array (which contains arbitrary distance values
4673  * from each node to each node in the system), and should also prefer nodes
4674  * with no CPUs, since presumably they'll have very little allocation pressure
4675  * on them otherwise.
4676  * It returns -1 if no node is found.
4677  */
4678 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4679 {
4680         int n, val;
4681         int min_val = INT_MAX;
4682         int best_node = NUMA_NO_NODE;
4683         const struct cpumask *tmp = cpumask_of_node(0);
4684 
4685         /* Use the local node if we haven't already */
4686         if (!node_isset(node, *used_node_mask)) {
4687                 node_set(node, *used_node_mask);
4688                 return node;
4689         }
4690 
4691         for_each_node_state(n, N_MEMORY) {
4692 
4693                 /* Don't want a node to appear more than once */
4694                 if (node_isset(n, *used_node_mask))
4695                         continue;
4696 
4697                 /* Use the distance array to find the distance */
4698                 val = node_distance(node, n);
4699 
4700                 /* Penalize nodes under us ("prefer the next node") */
4701                 val += (n < node);
4702 
4703                 /* Give preference to headless and unused nodes */
4704                 tmp = cpumask_of_node(n);
4705                 if (!cpumask_empty(tmp))
4706                         val += PENALTY_FOR_NODE_WITH_CPUS;
4707 
4708                 /* Slight preference for less loaded node */
4709                 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4710                 val += node_load[n];
4711 
4712                 if (val < min_val) {
4713                         min_val = val;
4714                         best_node = n;
4715                 }
4716         }
4717 
4718         if (best_node >= 0)
4719                 node_set(best_node, *used_node_mask);
4720 
4721         return best_node;
4722 }
4723 
4724 
4725 /*
4726  * Build zonelists ordered by node and zones within node.
4727  * This results in maximum locality--normal zone overflows into local
4728  * DMA zone, if any--but risks exhausting DMA zone.
4729  */
4730 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4731 {
4732         int j;
4733         struct zonelist *zonelist;
4734 
4735         zonelist = &pgdat->node_zonelists[0];
4736         for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4737                 ;
4738         j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4739         zonelist->_zonerefs[j].zone = NULL;
4740         zonelist->_zonerefs[j].zone_idx = 0;
4741 }
4742 
4743 /*
4744  * Build gfp_thisnode zonelists
4745  */
4746 static void build_thisnode_zonelists(pg_data_t *pgdat)
4747 {
4748         int j;
4749         struct zonelist *zonelist;
4750 
4751         zonelist = &pgdat->node_zonelists[1];
4752         j = build_zonelists_node(pgdat, zonelist, 0);
4753         zonelist->_zonerefs[j].zone = NULL;
4754         zonelist->_zonerefs[j].zone_idx = 0;
4755 }
4756 
4757 /*
4758  * Build zonelists ordered by zone and nodes within zones.
4759  * This results in conserving DMA zone[s] until all Normal memory is
4760  * exhausted, but results in overflowing to remote node while memory
4761  * may still exist in local DMA zone.
4762  */
4763 static int node_order[MAX_NUMNODES];
4764 
4765 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4766 {
4767         int pos, j, node;
4768         int zone_type;          /* needs to be signed */
4769         struct zone *z;
4770         struct zonelist *zonelist;
4771 
4772         zonelist = &pgdat->node_zonelists[0];
4773         pos = 0;
4774         for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4775                 for (j = 0; j < nr_nodes; j++) {
4776                         node = node_order[j];
4777                         z = &NODE_DATA(node)->node_zones[zone_type];
4778                         if (populated_zone(z)) {
4779                                 zoneref_set_zone(z,
4780                                         &zonelist->_zonerefs[pos++]);
4781                                 check_highest_zone(zone_type);
4782                         }
4783                 }
4784         }
4785         zonelist->_zonerefs[pos].zone = NULL;
4786         zonelist->_zonerefs[pos].zone_idx = 0;
4787 }
4788 
4789 #if defined(CONFIG_64BIT)
4790 /*
4791  * Devices that require DMA32/DMA are relatively rare and do not justify a
4792  * penalty to every machine in case the specialised case applies. Default
4793  * to Node-ordering on 64-bit NUMA machines
4794  */
4795 static int default_zonelist_order(void)
4796 {
4797         return ZONELIST_ORDER_NODE;
4798 }
4799 #else
4800 /*
4801  * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4802  * by the kernel. If processes running on node 0 deplete the low memory zone
4803  * then reclaim will occur more frequency increasing stalls and potentially
4804  * be easier to OOM if a large percentage of the zone is under writeback or
4805  * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4806  * Hence, default to zone ordering on 32-bit.
4807  */
4808 static int default_zonelist_order(void)
4809 {
4810         return ZONELIST_ORDER_ZONE;
4811 }
4812 #endif /* CONFIG_64BIT */
4813 
4814 static void set_zonelist_order(void)
4815 {
4816         if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4817                 current_zonelist_order = default_zonelist_order();
4818         else
4819                 current_zonelist_order = user_zonelist_order;
4820 }
4821 
4822 static void build_zonelists(pg_data_t *pgdat)
4823 {
4824         int i, node, load;
4825         nodemask_t used_mask;
4826         int local_node, prev_node;
4827         struct zonelist *zonelist;
4828         unsigned int order = current_zonelist_order;
4829 
4830         /* initialize zonelists */
4831         for (i = 0; i < MAX_ZONELISTS; i++) {
4832                 zonelist = pgdat->node_zonelists + i;
4833                 zonelist->_zonerefs[0].zone = NULL;
4834                 zonelist->_zonerefs[0].zone_idx = 0;
4835         }
4836 
4837         /* NUMA-aware ordering of nodes */
4838         local_node = pgdat->node_id;
4839         load = nr_online_nodes;
4840         prev_node = local_node;
4841         nodes_clear(used_mask);
4842 
4843         memset(node_order, 0, sizeof(node_order));
4844         i = 0;
4845 
4846         while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4847                 /*
4848                  * We don't want to pressure a particular node.
4849                  * So adding penalty to the first node in same
4850                  * distance group to make it round-robin.
4851                  */
4852                 if (node_distance(local_node, node) !=
4853                     node_distance(local_node, prev_node))
4854                         node_load[node] = load;
4855 
4856                 prev_node = node;
4857                 load--;
4858                 if (order == ZONELIST_ORDER_NODE)
4859                         build_zonelists_in_node_order(pgdat, node);
4860                 else
4861                         node_order[i++] = node; /* remember order */
4862         }
4863 
4864         if (order == ZONELIST_ORDER_ZONE) {
4865                 /* calculate node order -- i.e., DMA last! */
4866                 build_zonelists_in_zone_order(pgdat, i);
4867         }
4868 
4869         build_thisnode_zonelists(pgdat);
4870 }
4871 
4872 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4873 /*
4874  * Return node id of node used for "local" allocations.
4875  * I.e., first node id of first zone in arg node's generic zonelist.
4876  * Used for initializing percpu 'numa_mem', which is used primarily
4877  * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4878  */
4879 int local_memory_node(int node)
4880 {
4881         struct zoneref *z;
4882 
4883         z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4884                                    gfp_zone(GFP_KERNEL),
4885                                    NULL);
4886         return z->zone->node;
4887 }
4888 #endif
4889 
4890 #else   /* CONFIG_NUMA */
4891 
4892 static void set_zonelist_order(void)
4893 {
4894         current_zonelist_order = ZONELIST_ORDER_ZONE;
4895 }
4896 
4897 static void build_zonelists(pg_data_t *pgdat)
4898 {
4899         int node, local_node;
4900         enum zone_type j;
4901         struct zonelist *zonelist;
4902 
4903         local_node = pgdat->node_id;
4904 
4905         zonelist = &pgdat->node_zonelists[0];
4906         j = build_zonelists_node(pgdat, zonelist, 0);
4907 
4908         /*
4909          * Now we build the zonelist so that it contains the zones
4910          * of all the other nodes.
4911          * We don't want to pressure a particular node, so when
4912          * building the zones for node N, we make sure that the
4913          * zones coming right after the local ones are those from
4914          * node N+1 (modulo N)
4915          */
4916         for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4917                 if (!node_online(node))
4918                         continue;
4919                 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4920         }
4921         for (node = 0; node < local_node; node++) {
4922                 if (!node_online(node))
4923                         continue;
4924                 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4925         }
4926 
4927         zonelist->_zonerefs[j].zone = NULL;
4928         zonelist->_zonerefs[j].zone_idx = 0;
4929 }
4930 
4931 #endif  /* CONFIG_NUMA */
4932 
4933 /*
4934  * Boot pageset table. One per cpu which is going to be used for all
4935  * zones and all nodes. The parameters will be set in such a way
4936  * that an item put on a list will immediately be handed over to
4937  * the buddy list. This is safe since pageset manipulation is done
4938  * with interrupts disabled.
4939  *
4940  * The boot_pagesets must be kept even after bootup is complete for
4941  * unused processors and/or zones. They do play a role for bootstrapping
4942  * hotplugged processors.
4943  *
4944  * zoneinfo_show() and maybe other functions do
4945  * not check if the processor is online before following the pageset pointer.
4946  * Other parts of the kernel may not check if the zone is available.
4947  */
4948 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4949 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4950 static void setup_zone_pageset(struct zone *zone);
4951 
4952 /*
4953  * Global mutex to protect against size modification of zonelists
4954  * as well as to serialize pageset setup for the new populated zone.
4955  */
4956 DEFINE_MUTEX(zonelists_mutex);
4957 
4958 /* return values int ....just for stop_machine() */
4959 static int __build_all_zonelists(void *data)
4960 {
4961         int nid;
4962         int cpu;
4963         pg_data_t *self = data;
4964 
4965 #ifdef CONFIG_NUMA
4966         memset(node_load, 0, sizeof(node_load));
4967 #endif
4968 
4969         if (self && !node_online(self->node_id)) {
4970                 build_zonelists(self);
4971         }
4972 
4973         for_each_online_node(nid) {
4974                 pg_data_t *pgdat = NODE_DATA(nid);
4975 
4976                 build_zonelists(pgdat);
4977         }
4978 
4979         /*
4980          * Initialize the boot_pagesets that are going to be used
4981          * for bootstrapping processors. The real pagesets for
4982          * each zone will be allocated later when the per cpu
4983          * allocator is available.
4984          *
4985          * boot_pagesets are used also for bootstrapping offline
4986          * cpus if the system is already booted because the pagesets
4987          * are needed to initialize allocators on a specific cpu too.
4988          * F.e. the percpu allocator needs the page allocator which
4989          * needs the percpu allocator in order to allocate its pagesets
4990          * (a chicken-egg dilemma).
4991          */
4992         for_each_possible_cpu(cpu) {
4993                 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4994 
4995 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4996                 /*
4997                  * We now know the "local memory node" for each node--
4998                  * i.e., the node of the first zone in the generic zonelist.
4999                  * Set up numa_mem percpu variable for on-line cpus.  During
5000                  * boot, only the boot cpu should be on-line;  we'll init the
5001                  * secondary cpus' numa_mem as they come on-line.  During
5002                  * node/memory hotplug, we'll fixup all on-line cpus.
5003                  */
5004                 if (cpu_online(cpu))
5005                         set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5006 #endif
5007         }
5008 
5009         return 0;
5010 }
5011 
5012 static noinline void __init
5013 build_all_zonelists_init(void)
5014 {
5015         __build_all_zonelists(NULL);
5016         mminit_verify_zonelist();
5017         cpuset_init_current_mems_allowed();
5018 }
5019 
5020 /*
5021  * Called with zonelists_mutex held always
5022  * unless system_state == SYSTEM_BOOTING.
5023  *
5024  * __ref due to (1) call of __meminit annotated setup_zone_pageset
5025  * [we're only called with non-NULL zone through __meminit paths] and
5026  * (2) call of __init annotated helper build_all_zonelists_init
5027  * [protected by SYSTEM_BOOTING].
5028  */
5029 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5030 {
5031         set_zonelist_order();
5032 
5033         if (system_state == SYSTEM_BOOTING) {
5034                 build_all_zonelists_init();
5035         } else {
5036 #ifdef CONFIG_MEMORY_HOTPLUG
5037                 if (zone)
5038                         setup_zone_pageset(zone);
5039 #endif
5040                 /* we have to stop all cpus to guarantee there is no user
5041                    of zonelist */
5042                 stop_machine(__build_all_zonelists, pgdat, NULL);
5043                 /* cpuset refresh routine should be here */
5044         }
5045         vm_total_pages = nr_free_pagecache_pages();
5046         /*
5047          * Disable grouping by mobility if the number of pages in the
5048          * system is too low to allow the mechanism to work. It would be
5049          * more accurate, but expensive to check per-zone. This check is
5050          * made on memory-hotadd so a system can start with mobility
5051          * disabled and enable it later
5052          */
5053         if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5054                 page_group_by_mobility_disabled = 1;
5055         else
5056                 page_group_by_mobility_disabled = 0;
5057 
5058         pr_info("Built %i zonelists in %s order, mobility grouping %s.  Total pages: %ld\n",
5059                 nr_online_nodes,
5060                 zonelist_order_name[current_zonelist_order],
5061                 page_group_by_mobility_disabled ? "off" : "on",
5062                 vm_total_pages);
5063 #ifdef CONFIG_NUMA
5064         pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5065 #endif
5066 }
5067 
5068 /*
5069  * Helper functions to size the waitqueue hash table.
5070  * Essentially these want to choose hash table sizes sufficiently
5071  * large so that collisions trying to wait on pages are rare.
5072  * But in fact, the number of active page waitqueues on typical
5073  * systems is ridiculously low, less than 200. So this is even
5074  * conservative, even though it seems large.
5075  *
5076  * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
5077  * waitqueues, i.e. the size of the waitq table given the number of pages.
5078  */
5079 #define PAGES_PER_WAITQUEUE     256
5080 
5081 #ifndef CONFIG_MEMORY_HOTPLUG
5082 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5083 {
5084         unsigned long size = 1;
5085 
5086         pages /= PAGES_PER_WAITQUEUE;
5087 
5088         while (size < pages)
5089                 size <<= 1;
5090 
5091         /*
5092          * Once we have dozens or even hundreds of threads sleeping
5093          * on IO we've got bigger problems than wait queue collision.
5094          * Limit the size of the wait table to a reasonable size.
5095          */
5096         size = min(size, 4096UL);
5097 
5098         return max(size, 4UL);
5099 }
5100 #else
5101 /*
5102  * A zone's size might be changed by hot-add, so it is not possible to determine
5103  * a suitable size for its wait_table.  So we use the maximum size now.
5104  *
5105  * The max wait table size = 4096 x sizeof(wait_queue_head_t).   ie:
5106  *
5107  *    i386 (preemption config)    : 4096 x 16 = 64Kbyte.
5108  *    ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
5109  *    ia64, x86-64 (preemption)   : 4096 x 24 = 96Kbyte.
5110  *
5111  * The maximum entries are prepared when a zone's memory is (512K + 256) pages
5112  * or more by the traditional way. (See above).  It equals:
5113  *
5114  *    i386, x86-64, powerpc(4K page size) : =  ( 2G + 1M)byte.
5115  *    ia64(16K page size)                 : =  ( 8G + 4M)byte.
5116  *    powerpc (64K page size)             : =  (32G +16M)byte.
5117  */
5118 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
5119 {
5120         return 4096UL;
5121 }
5122 #endif
5123 
5124 /*
5125  * This is an integer logarithm so that shifts can be used later
5126  * to extract the more random high bits from the multiplicative
5127  * hash function before the remainder is taken.
5128  */
5129 static inline unsigned long wait_table_bits(unsigned long size)
5130 {
5131         return ffz(~size);
5132 }
5133 
5134 /*
5135  * Initially all pages are reserved - free ones are freed
5136  * up by free_all_bootmem() once the early boot process is
5137  * done. Non-atomic initialization, single-pass.
5138  */
5139 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5140                 unsigned long start_pfn, enum memmap_context context)
5141 {
5142         struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5143         unsigned long end_pfn = start_pfn + size;
5144         pg_data_t *pgdat = NODE_DATA(nid);
5145         unsigned long pfn;
5146         unsigned long nr_initialised = 0;
5147 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5148         struct memblock_region *r = NULL, *tmp;
5149 #endif
5150 
5151         if (highest_memmap_pfn < end_pfn - 1)
5152                 highest_memmap_pfn = end_pfn - 1;
5153 
5154         /*
5155          * Honor reservation requested by the driver for this ZONE_DEVICE
5156          * memory
5157          */
5158         if (altmap && start_pfn == altmap->base_pfn)
5159                 start_pfn += altmap->reserve;
5160 
5161         for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5162                 /*
5163                  * There can be holes in boot-time mem_map[]s handed to this
5164                  * function.  They do not exist on hotplugged memory.
5165                  */
5166                 if (context != MEMMAP_EARLY)
5167                         goto not_early;
5168 
5169                 if (!early_pfn_valid(pfn))
5170                         continue;
5171                 if (!early_pfn_in_nid(pfn, nid))
5172                         continue;
5173                 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5174                         break;
5175 
5176 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5177                 /*
5178                  * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
5179                  * from zone_movable_pfn[nid] to end of each node should be
5180                  * ZONE_MOVABLE not ZONE_NORMAL. skip it.
5181                  */
5182                 if (!mirrored_kernelcore && zone_movable_pfn[nid])
5183                         if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
5184                                 continue;
5185 
5186                 /*
5187                  * Check given memblock attribute by firmware which can affect
5188                  * kernel memory layout.  If zone==ZONE_MOVABLE but memory is
5189                  * mirrored, it's an overlapped memmap init. skip it.
5190                  */
5191                 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5192                         if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5193                                 for_each_memblock(memory, tmp)
5194                                         if (pfn < memblock_region_memory_end_pfn(tmp))
5195                                                 break;
5196                                 r = tmp;
5197                         }
5198                         if (pfn >= memblock_region_memory_base_pfn(r) &&
5199                             memblock_is_mirror(r)) {
5200                                 /* already initialized as NORMAL */
5201                                 pfn = memblock_region_memory_end_pfn(r);
5202                                 continue;
5203                         }
5204                 }
5205 #endif
5206 
5207 not_early:
5208                 /*
5209                  * Mark the block movable so that blocks are reserved for
5210                  * movable at startup. This will force kernel allocations
5211                  * to reserve their blocks rather than leaking throughout
5212                  * the address space during boot when many long-lived
5213                  * kernel allocations are made.
5214                  *
5215                  * bitmap is created for zone's valid pfn range. but memmap
5216                  * can be created for invalid pages (for alignment)
5217                  * check here not to call set_pageblock_migratetype() against
5218                  * pfn out of zone.
5219                  */
5220                 if (!(pfn & (pageblock_nr_pages - 1))) {
5221                         struct page *page = pfn_to_page(pfn);
5222 
5223                         __init_single_page(page, pfn, zone, nid);
5224                         set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5225                 } else {
5226                         __init_single_pfn(pfn, zone, nid);
5227                 }
5228         }
5229 }
5230 
5231 static void __meminit zone_init_free_lists(struct zone *zone)
5232 {
5233         unsigned int order, t;
5234         for_each_migratetype_order(order, t) {
5235                 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5236                 zone->free_area[order].nr_free = 0;
5237         }
5238 }
5239 
5240 #ifndef __HAVE_ARCH_MEMMAP_INIT
5241 #define memmap_init(size, nid, zone, start_pfn) \
5242         memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5243 #endif
5244 
5245 static int zone_batchsize(struct zone *zone)
5246 {
5247 #ifdef CONFIG_MMU
5248         int batch;
5249 
5250         /*
5251          * The per-cpu-pages pools are set to around 1000th of the
5252          * size of the zone.  But no more than 1/2 of a meg.
5253          *
5254          * OK, so we don't know how big the cache is.  So guess.
5255          */
5256         batch = zone->managed_pages / 1024;
5257         if (batch * PAGE_SIZE > 512 * 1024)
5258                 batch = (512 * 1024) / PAGE_SIZE;
5259         batch /= 4;             /* We effectively *= 4 below */
5260         if (batch < 1)
5261                 batch = 1;
5262 
5263         /*
5264          * Clamp the batch to a 2^n - 1 value. Having a power
5265          * of 2 value was found to be more likely to have
5266          * suboptimal cache aliasing properties in some cases.
5267          *
5268          * For example if 2 tasks are alternately allocating
5269          * batches of pages, one task can end up with a lot
5270          * of pages of one half of the possible page colors
5271          * and the other with pages of the other colors.
5272          */
5273         batch = rounddown_pow_of_two(batch + batch/2) - 1;
5274 
5275         return batch;
5276 
5277 #else
5278         /* The deferral and batching of frees should be suppressed under NOMMU
5279          * conditions.
5280          *
5281          * The problem is that NOMMU needs to be able to allocate large chunks
5282          * of contiguous memory as there's no hardware page translation to
5283          * assemble apparent contiguous memory from discontiguous pages.
5284          *
5285          * Queueing large contiguous runs of pages for batching, however,
5286          * causes the pages to actually be freed in smaller chunks.  As there
5287          * can be a significant delay between the individual batches being
5288          * recycled, this leads to the once large chunks of space being
5289          * fragmented and becoming unavailable for high-order allocations.
5290          */
5291         return 0;
5292 #endif
5293 }
5294 
5295 /*
5296  * pcp->high and pcp->batch values are related and dependent on one another:
5297  * ->batch must never be higher then ->high.
5298  * The following function updates them in a safe manner without read side
5299  * locking.
5300  *
5301  * Any new users of pcp->batch and pcp->high should ensure they can cope with
5302  * those fields changing asynchronously (acording the the above rule).
5303  *
5304  * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5305  * outside of boot time (or some other assurance that no concurrent updaters
5306  * exist).
5307  */
5308 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5309                 unsigned long batch)
5310 {
5311        /* start with a fail safe value for batch */
5312         pcp->batch = 1;
5313         smp_wmb();
5314 
5315        /* Update high, then batch, in order */
5316         pcp->high = high;
5317         smp_wmb();
5318 
5319         pcp->batch = batch;
5320 }
5321 
5322 /* a companion to pageset_set_high() */
5323 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5324 {
5325         pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5326 }
5327 
5328 static void pageset_init(struct per_cpu_pageset *p)
5329 {
5330         struct per_cpu_pages *pcp;
5331         int migratetype;
5332 
5333         memset(p, 0, sizeof(*p));
5334 
5335         pcp = &p->pcp;
5336         pcp->count = 0;
5337         for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5338                 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5339 }
5340 
5341 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5342 {
5343         pageset_init(p);
5344         pageset_set_batch(p, batch);
5345 }
5346 
5347 /*
5348  * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5349  * to the value high for the pageset p.
5350  */
5351 static void pageset_set_high(struct per_cpu_pageset *p,
5352                                 unsigned long high)
5353 {
5354         unsigned long batch = max(1UL, high / 4);
5355         if ((high / 4) > (PAGE_SHIFT * 8))
5356                 batch = PAGE_SHIFT * 8;
5357 
5358         pageset_update(&p->pcp, high, batch);
5359 }
5360 
5361 static void pageset_set_high_and_batch(struct zone *zone,
5362                                        struct per_cpu_pageset *pcp)
5363 {
5364         if (percpu_pagelist_fraction)
5365                 pageset_set_high(pcp,
5366                         (zone->managed_pages /
5367                                 percpu_pagelist_fraction));
5368         else
5369                 pageset_set_batch(pcp, zone_batchsize(zone));
5370 }
5371 
5372 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5373 {
5374         struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5375 
5376         pageset_init(pcp);
5377         pageset_set_high_and_batch(zone, pcp);
5378 }
5379 
5380 static void __meminit setup_zone_pageset(struct zone *zone)
5381 {
5382         int cpu;
5383         zone->pageset = alloc_percpu(struct per_cpu_pageset);
5384         for_each_possible_cpu(cpu)
5385                 zone_pageset_init(zone, cpu);
5386 }
5387 
5388 /*
5389  * Allocate per cpu pagesets and initialize them.
5390  * Before this call only boot pagesets were available.
5391  */
5392 void __init setup_per_cpu_pageset(void)
5393 {
5394         struct zone *zone;
5395 
5396         for_each_populated_zone(zone)
5397                 setup_zone_pageset(zone);
5398 }
5399 
5400 static noinline __init_refok
5401 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
5402 {
5403         int i;
5404         size_t alloc_size;
5405 
5406         /*
5407          * The per-page waitqueue mechanism uses hashed waitqueues
5408          * per zone.
5409          */
5410         zone->wait_table_hash_nr_entries =
5411                  wait_table_hash_nr_entries(zone_size_pages);
5412         zone->wait_table_bits =
5413                 wait_table_bits(zone->wait_table_hash_nr_entries);
5414         alloc_size = zone->wait_table_hash_nr_entries
5415                                         * sizeof(wait_queue_head_t);
5416 
5417         if (!slab_is_available()) {
5418                 zone->wait_table = (wait_queue_head_t *)
5419                         memblock_virt_alloc_node_nopanic(
5420                                 alloc_size, zone->zone_pgdat->node_id);
5421         } else {
5422                 /*
5423                  * This case means that a zone whose size was 0 gets new memory
5424                  * via memory hot-add.
5425                  * But it may be the case that a new node was hot-added.  In
5426                  * this case vmalloc() will not be able to use this new node's
5427                  * memory - this wait_table must be initialized to use this new
5428                  * node itself as well.
5429                  * To use this new node's memory, further consideration will be
5430                  * necessary.
5431                  */
5432                 zone->wait_table = vmalloc(alloc_size);
5433         }
5434         if (!zone->wait_table)
5435                 return -ENOMEM;
5436 
5437         for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
5438                 init_waitqueue_head(zone->wait_table + i);
5439 
5440         return 0;
5441 }
5442 
5443 static __meminit void zone_pcp_init(struct zone *zone)
5444 {
5445         /*
5446          * per cpu subsystem is not up at this point. The following code
5447          * relies on the ability of the linker to provide the
5448          * offset of a (static) per cpu variable into the per cpu area.
5449          */
5450         zone->pageset = &boot_pageset;
5451 
5452         if (populated_zone(zone))
5453                 printk(KERN_DEBUG "  %s zone: %lu pages, LIFO batch:%u\n",
5454                         zone->name, zone->present_pages,
5455                                          zone_batchsize(zone));
5456 }
5457 
5458 int __meminit init_currently_empty_zone(struct zone *zone,
5459                                         unsigned long zone_start_pfn,
5460                                         unsigned long size)
5461 {
5462         struct pglist_data *pgdat = zone->zone_pgdat;
5463         int ret;
5464         ret = zone_wait_table_init(zone, size);
5465         if (ret)
5466                 return ret;
5467         pgdat->nr_zones = zone_idx(zone) + 1;
5468 
5469         zone->zone_start_pfn = zone_start_pfn;
5470 
5471         mminit_dprintk(MMINIT_TRACE, "memmap_init",
5472                         "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5473                         pgdat->node_id,
5474                         (unsigned long)zone_idx(zone),
5475                         zone_start_pfn, (zone_start_pfn + size));
5476 
5477         zone_init_free_lists(zone);
5478 
5479         return 0;
5480 }
5481 
5482 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5483 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5484 
5485 /*
5486  * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5487  */
5488 int __meminit __early_pfn_to_nid(unsigned long pfn,
5489                                         struct mminit_pfnnid_cache *state)
5490 {
5491         unsigned long start_pfn, end_pfn;
5492         int nid;
5493 
5494         if (state->last_start <= pfn && pfn < state->last_end)
5495                 return state->last_nid;
5496 
5497         nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5498         if (nid != -1) {
5499                 state->last_start = start_pfn;
5500                 state->last_end = end_pfn;
5501                 state->last_nid = nid;
5502         }
5503 
5504         return nid;
5505 }
5506 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5507 
5508 /**
5509  * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5510  * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5511  * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5512  *
5513  * If an architecture guarantees that all ranges registered contain no holes
5514  * and may be freed, this this function may be used instead of calling
5515  * memblock_free_early_nid() manually.
5516  */
5517 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5518 {
5519         unsigned long start_pfn, end_pfn;
5520         int i, this_nid;
5521 
5522         for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5523                 start_pfn = min(start_pfn, max_low_pfn);
5524                 end_pfn = min(end_pfn, max_low_pfn);
5525 
5526                 if (start_pfn < end_pfn)
5527                         memblock_free_early_nid(PFN_PHYS(start_pfn),
5528                                         (end_pfn - start_pfn) << PAGE_SHIFT,
5529                                         this_nid);
5530         }
5531 }
5532 
5533 /**
5534  * sparse_memory_present_with_active_regions - Call memory_present for each active range
5535  * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5536  *
5537  * If an architecture guarantees that all ranges registered contain no holes and may
5538  * be freed, this function may be used instead of calling memory_present() manually.
5539  */
5540 void __init sparse_memory_present_with_active_regions(int nid)
5541 {
5542         unsigned long start_pfn, end_pfn;
5543         int i, this_nid;
5544 
5545         for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5546                 memory_present(this_nid, start_pfn, end_pfn);
5547 }
5548 
5549 /**
5550  * get_pfn_range_for_nid - Return the start and end page frames for a node
5551  * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5552  * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5553  * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5554  *
5555  * It returns the start and end page frame of a node based on information
5556  * provided by memblock_set_node(). If called for a node
5557  * with no available memory, a warning is printed and the start and end
5558  * PFNs will be 0.
5559  */
5560 void __meminit get_pfn_range_for_nid(unsigned int nid,
5561                         unsigned long *start_pfn, unsigned long *end_pfn)
5562 {
5563         unsigned long this_start_pfn, this_end_pfn;
5564         int i;
5565 
5566         *start_pfn = -1UL;
5567         *end_pfn = 0;
5568 
5569         for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5570                 *start_pfn = min(*start_pfn, this_start_pfn);
5571                 *end_pfn = max(*end_pfn, this_end_pfn);
5572         }
5573 
5574         if (*start_pfn == -1UL)
5575                 *start_pfn = 0;
5576 }
5577 
5578 /*
5579  * This finds a zone that can be used for ZONE_MOVABLE pages. The
5580  * assumption is made that zones within a node are ordered in monotonic
5581  * increasing memory addresses so that the "highest" populated zone is used
5582  */
5583 static void __init find_usable_zone_for_movable(void)
5584 {
5585         int zone_index;
5586         for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5587                 if (zone_index == ZONE_MOVABLE)
5588                         continue;
5589 
5590                 if (arch_zone_highest_possible_pfn[zone_index] >
5591                                 arch_zone_lowest_possible_pfn[zone_index])
5592                         break;
5593         }
5594 
5595         VM_BUG_ON(zone_index == -1);
5596         movable_zone = zone_index;
5597 }
5598 
5599 /*
5600  * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5601  * because it is sized independent of architecture. Unlike the other zones,
5602  * the starting point for ZONE_MOVABLE is not fixed. It may be different
5603  * in each node depending on the size of each node and how evenly kernelcore
5604  * is distributed. This helper function adjusts the zone ranges
5605  * provided by the architecture for a given node by using the end of the
5606  * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5607  * zones within a node are in order of monotonic increases memory addresses
5608  */
5609 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5610                                         unsigned long zone_type,
5611                                         unsigned long node_start_pfn,
5612                                         unsigned long node_end_pfn,
5613                                         unsigned long *zone_start_pfn,
5614                                         unsigned long *zone_end_pfn)
5615 {
5616         /* Only adjust if ZONE_MOVABLE is on this node */
5617         if (zone_movable_pfn[nid]) {
5618                 /* Size ZONE_MOVABLE */
5619                 if (zone_type == ZONE_MOVABLE) {
5620                         *zone_start_pfn = zone_movable_pfn[nid];
5621                         *zone_end_pfn = min(node_end_pfn,
5622                                 arch_zone_highest_possible_pfn[movable_zone]);
5623 
5624                 /* Check if this whole range is within ZONE_MOVABLE */
5625                 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5626                         *zone_start_pfn = *zone_end_pfn;
5627         }
5628 }
5629 
5630 /*
5631  * Return the number of pages a zone spans in a node, including holes
5632  * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5633  */
5634 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5635                                         unsigned long zone_type,
5636                                         unsigned long node_start_pfn,
5637                                         unsigned long node_end_pfn,
5638                                         unsigned long *zone_start_pfn,
5639                                         unsigned long *zone_end_pfn,
5640                                         unsigned long *ignored)
5641 {
5642         /* When hotadd a new node from cpu_up(), the node should be empty */
5643         if (!node_start_pfn && !node_end_pfn)
5644                 return 0;
5645 
5646         /* Get the start and end of the zone */
5647         *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5648         *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5649         adjust_zone_range_for_zone_movable(nid, zone_type,
5650                                 node_start_pfn, node_end_pfn,
5651                                 zone_start_pfn, zone_end_pfn);
5652 
5653         /* Check that this node has pages within the zone's required range */
5654         if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5655                 return 0;
5656 
5657         /* Move the zone boundaries inside the node if necessary */
5658         *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5659         *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5660 
5661         /* Return the spanned pages */
5662         return *zone_end_pfn - *zone_start_pfn;
5663 }
5664 
5665 /*
5666  * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5667  * then all holes in the requested range will be accounted for.
5668  */
5669 unsigned long __meminit __absent_pages_in_range(int nid,
5670                                 unsigned long range_start_pfn,
5671                                 unsigned long range_end_pfn)
5672 {
5673         unsigned long nr_absent = range_end_pfn - range_start_pfn;
5674         unsigned long start_pfn, end_pfn;
5675         int i;
5676 
5677         for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5678                 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5679                 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5680                 nr_absent -= end_pfn - start_pfn;
5681         }
5682         return nr_absent;
5683 }
5684 
5685 /**
5686  * absent_pages_in_range - Return number of page frames in holes within a range
5687  * @start_pfn: The start PFN to start searching for holes
5688  * @end_pfn: The end PFN to stop searching for holes
5689  *
5690  * It returns the number of pages frames in memory holes within a range.
5691  */
5692 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5693                                                         unsigned long end_pfn)
5694 {
5695         return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5696 }
5697 
5698 /* Return the number of page frames in holes in a zone on a node */
5699 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5700                                         unsigned long zone_type,
5701                                         unsigned long node_start_pfn,
5702                                         unsigned long node_end_pfn,
5703                                         unsigned long *ignored)
5704 {
5705         unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5706         unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5707         unsigned long zone_start_pfn, zone_end_pfn;
5708         unsigned long nr_absent;
5709 
5710         /* When hotadd a new node from cpu_up(), the node should be empty */
5711         if (!node_start_pfn && !node_end_pfn)
5712                 return 0;
5713 
5714         zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5715         zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5716 
5717         adjust_zone_range_for_zone_movable(nid, zone_type,
5718                         node_start_pfn, node_end_pfn,
5719                         &zone_start_pfn, &zone_end_pfn);
5720         nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5721 
5722         /*
5723          * ZONE_MOVABLE handling.
5724          * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5725          * and vice versa.
5726          */
5727         if (zone_movable_pfn[nid]) {
5728                 if (mirrored_kernelcore) {
5729                         unsigned long start_pfn, end_pfn;
5730                         struct memblock_region *r;
5731 
5732                         for_each_memblock(memory, r) {
5733                                 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5734                                                   zone_start_pfn, zone_end_pfn);
5735                                 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5736                                                 zone_start_pfn, zone_end_pfn);
5737 
5738                                 if (zone_type == ZONE_MOVABLE &&
5739                                     memblock_is_mirror(r))
5740                                         nr_absent += end_pfn - start_pfn;
5741 
5742                                 if (zone_type == ZONE_NORMAL &&
5743                                     !memblock_is_mirror(r))
5744                                         nr_absent += end_pfn - start_pfn;
5745                         }
5746                 } else {
5747                         if (zone_type == ZONE_NORMAL)
5748                                 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5749                 }
5750         }
5751 
5752         return nr_absent;
5753 }
5754 
5755 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5756 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5757                                         unsigned long zone_type,
5758                                         unsigned long node_start_pfn,
5759                                         unsigned long node_end_pfn,
5760                                         unsigned long *zone_start_pfn,
5761                                         unsigned long *zone_end_pfn,
5762                                         unsigned long *zones_size)
5763 {
5764         unsigned int zone;
5765 
5766         *zone_start_pfn = node_start_pfn;
5767         for (zone = 0; zone < zone_type; zone++)
5768                 *zone_start_pfn += zones_size[zone];
5769 
5770         *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5771 
5772         return zones_size[zone_type];
5773 }
5774 
5775 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5776                                                 unsigned long zone_type,
5777                                                 unsigned long node_start_pfn,
5778                                                 unsigned long node_end_pfn,
5779                                                 unsigned long *zholes_size)
5780 {
5781         if (!zholes_size)
5782                 return 0;
5783 
5784         return zholes_size[zone_type];
5785 }
5786 
5787 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5788 
5789 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5790                                                 unsigned long node_start_pfn,
5791                                                 unsigned long node_end_pfn,
5792                                                 unsigned long *zones_size,
5793                                                 unsigned long *zholes_size)
5794 {
5795         unsigned long realtotalpages = 0, totalpages = 0;
5796         enum zone_type i;
5797 
5798         for (i = 0; i < MAX_NR_ZONES; i++) {
5799                 struct zone *zone = pgdat->node_zones + i;
5800                 unsigned long zone_start_pfn, zone_end_pfn;
5801                 unsigned long size, real_size;
5802 
5803                 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5804                                                   node_start_pfn,
5805                                                   node_end_pfn,
5806                                                   &zone_start_pfn,
5807                                                   &zone_end_pfn,
5808                                                   zones_size);
5809                 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5810                                                   node_start_pfn, node_end_pfn,
5811                                                   zholes_size);
5812                 if (size)
5813                         zone->zone_start_pfn = zone_start_pfn;
5814                 else
5815                         zone->zone_start_pfn = 0;
5816                 zone->spanned_pages = size;
5817                 zone->present_pages = real_size;
5818 
5819                 totalpages += size;
5820                 realtotalpages += real_size;
5821         }
5822 
5823         pgdat->node_spanned_pages = totalpages;
5824         pgdat->node_present_pages = realtotalpages;
5825         printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5826                                                         realtotalpages);
5827 }
5828 
5829 #ifndef CONFIG_SPARSEMEM
5830 /*
5831  * Calculate the size of the zone->blockflags rounded to an unsigned long
5832  * Start by making sure zonesize is a multiple of pageblock_order by rounding
5833  * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5834  * round what is now in bits to nearest long in bits, then return it in
5835  * bytes.
5836  */
5837 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5838 {
5839         unsigned long usemapsize;
5840 
5841         zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5842         usemapsize = roundup(zonesize, pageblock_nr_pages);
5843         usemapsize = usemapsize >> pageblock_order;
5844         usemapsize *= NR_PAGEBLOCK_BITS;
5845         usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5846 
5847         return usemapsize / 8;
5848 }
5849 
5850 static void __init setup_usemap(struct pglist_data *pgdat,
5851                                 struct zone *zone,
5852                                 unsigned long zone_start_pfn,
5853                                 unsigned long zonesize)
5854 {
5855         unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5856         zone->pageblock_flags = NULL;
5857         if (usemapsize)
5858                 zone->pageblock_flags =
5859                         memblock_virt_alloc_node_nopanic(usemapsize,
5860                                                          pgdat->node_id);
5861 }
5862 #else
5863 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5864                                 unsigned long zone_start_pfn, unsigned long zonesize) {}
5865 #endif /* CONFIG_SPARSEMEM */
5866 
5867 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5868 
5869 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5870 void __paginginit set_pageblock_order(void)
5871 {
5872         unsigned int order;
5873 
5874         /* Check that pageblock_nr_pages has not already been setup */
5875         if (pageblock_order)
5876                 return;
5877 
5878         if (HPAGE_SHIFT > PAGE_SHIFT)
5879                 order = HUGETLB_PAGE_ORDER;
5880         else
5881                 order = MAX_ORDER - 1;
5882 
5883         /*
5884          * Assume the largest contiguous order of interest is a huge page.
5885          * This value may be variable depending on boot parameters on IA64 and
5886          * powerpc.
5887          */
5888         pageblock_order = order;
5889 }
5890 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5891 
5892 /*
5893  * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5894  * is unused as pageblock_order is set at compile-time. See
5895  * include/linux/pageblock-flags.h for the values of pageblock_order based on
5896  * the kernel config
5897  */
5898 void __paginginit set_pageblock_order(void)
5899 {
5900 }
5901 
5902 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5903 
5904 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5905                                                    unsigned long present_pages)
5906 {
5907         unsigned long pages = spanned_pages;
5908 
5909         /*
5910          * Provide a more accurate estimation if there are holes within
5911          * the zone and SPARSEMEM is in use. If there are holes within the
5912          * zone, each populated memory region may cost us one or two extra
5913          * memmap pages due to alignment because memmap pages for each
5914          * populated regions may not naturally algined on page boundary.
5915          * So the (present_pages >> 4) heuristic is a tradeoff for that.
5916          */
5917         if (spanned_pages > present_pages + (present_pages >> 4) &&
5918             IS_ENABLED(CONFIG_SPARSEMEM))
5919                 pages = present_pages;
5920 
5921         return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5922 }
5923 
5924 /*
5925  * Set up the zone data structures:
5926  *   - mark all pages reserved
5927  *   - mark all memory queues empty
5928  *   - clear the memory bitmaps
5929  *
5930  * NOTE: pgdat should get zeroed by caller.
5931  */
5932 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5933 {
5934         enum zone_type j;
5935         int nid = pgdat->node_id;
5936         int ret;
5937 
5938         pgdat_resize_init(pgdat);
5939 #ifdef CONFIG_NUMA_BALANCING
5940         spin_lock_init(&pgdat->numabalancing_migrate_lock);
5941         pgdat->numabalancing_migrate_nr_pages = 0;
5942         pgdat->numabalancing_migrate_next_window = jiffies;
5943 #endif
5944 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5945         spin_lock_init(&pgdat->split_queue_lock);
5946         INIT_LIST_HEAD(&pgdat->split_queue);
5947         pgdat->split_queue_len = 0;
5948 #endif
5949         init_waitqueue_head(&pgdat->kswapd_wait);
5950         init_waitqueue_head(&pgdat->pfmemalloc_wait);
5951 #ifdef CONFIG_COMPACTION
5952         init_waitqueue_head(&pgdat->kcompactd_wait);
5953 #endif
5954         pgdat_page_ext_init(pgdat);
5955 
5956         for (j = 0; j < MAX_NR_ZONES; j++) {
5957                 struct zone *zone = pgdat->node_zones + j;
5958                 unsigned long size, realsize, freesize, memmap_pages;
5959                 unsigned long zone_start_pfn = zone->zone_start_pfn;
5960 
5961                 size = zone->spanned_pages;
5962                 realsize = freesize = zone->present_pages;
5963 
5964                 /*
5965                  * Adjust freesize so that it accounts for how much memory
5966                  * is used by this zone for memmap. This affects the watermark
5967                  * and per-cpu initialisations
5968                  */
5969                 memmap_pages = calc_memmap_size(size, realsize);
5970                 if (!is_highmem_idx(j)) {
5971                         if (freesize >= memmap_pages) {
5972                                 freesize -= memmap_pages;
5973                                 if (memmap_pages)
5974                                         printk(KERN_DEBUG
5975                                                "  %s zone: %lu pages used for memmap\n",
5976                                                zone_names[j], memmap_pages);
5977                         } else
5978                                 pr_warn("  %s zone: %lu pages exceeds freesize %lu\n",
5979                                         zone_names[j], memmap_pages, freesize);
5980                 }
5981 
5982                 /* Account for reserved pages */
5983                 if (j == 0 && freesize > dma_reserve) {
5984                         freesize -= dma_reserve;
5985                         printk(KERN_DEBUG "  %s zone: %lu pages reserved\n",
5986                                         zone_names[0], dma_reserve);
5987                 }
5988 
5989                 if (!is_highmem_idx(j))
5990                         nr_kernel_pages += freesize;
5991                 /* Charge for highmem memmap if there are enough kernel pages */
5992                 else if (nr_kernel_pages > memmap_pages * 2)
5993                         nr_kernel_pages -= memmap_pages;
5994                 nr_all_pages += freesize;
5995 
5996                 /*
5997                  * Set an approximate value for lowmem here, it will be adjusted
5998                  * when the bootmem allocator frees pages into the buddy system.
5999                  * And all highmem pages will be managed by the buddy system.
6000                  */
6001                 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6002 #ifdef CONFIG_NUMA
6003                 zone->node = nid;
6004                 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
6005                                                 / 100;
6006                 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
6007 #endif
6008                 zone->name = zone_names[j];
6009                 spin_lock_init(&zone->lock);
6010                 spin_lock_init(&zone->lru_lock);
6011                 zone_seqlock_init(zone);
6012                 zone->zone_pgdat = pgdat;
6013                 zone_pcp_init(zone);
6014 
6015                 /* For bootup, initialized properly in watermark setup */
6016                 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
6017 
6018                 lruvec_init(&zone->lruvec);
6019                 if (!size)
6020                         continue;
6021 
6022                 set_pageblock_order();
6023                 setup_usemap(pgdat, zone, zone_start_pfn, size);
6024                 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
6025                 BUG_ON(ret);
6026                 memmap_init(size, nid, j, zone_start_pfn);
6027         }
6028 }
6029 
6030 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
6031 {
6032         unsigned long __maybe_unused start = 0;
6033         unsigned long __maybe_unused offset = 0;
6034 
6035         /* Skip empty nodes */
6036         if (!pgdat->node_spanned_pages)
6037                 return;
6038 
6039 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6040         start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6041         offset = pgdat->node_start_pfn - start;
6042         /* ia64 gets its own node_mem_map, before this, without bootmem */
6043         if (!pgdat->node_mem_map) {
6044                 unsigned long size, end;
6045                 struct page *map;
6046 
6047                 /*
6048                  * The zone's endpoints aren't required to be MAX_ORDER
6049                  * aligned but the node_mem_map endpoints must be in order
6050                  * for the buddy allocator to function correctly.
6051                  */
6052                 end = pgdat_end_pfn(pgdat);
6053                 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6054                 size =  (end - start) * sizeof(struct page);
6055                 map = alloc_remap(pgdat->node_id, size);
6056                 if (!map)
6057                         map = memblock_virt_alloc_node_nopanic(size,
6058                                                                pgdat->node_id);
6059                 pgdat->node_mem_map = map + offset;
6060         }
6061 #ifndef CONFIG_NEED_MULTIPLE_NODES
6062         /*
6063          * With no DISCONTIG, the global mem_map is just set as node 0's
6064          */
6065         if (pgdat == NODE_DATA(0)) {
6066                 mem_map = NODE_DATA(0)->node_mem_map;
6067 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6068                 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6069                         mem_map -= offset;
6070 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6071         }
6072 #endif
6073 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6074 }
6075 
6076 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6077                 unsigned long node_start_pfn, unsigned long *zholes_size)
6078 {
6079         pg_data_t *pgdat = NODE_DATA(nid);
6080         unsigned long start_pfn = 0;
6081         unsigned long end_pfn = 0;
6082 
6083         /* pg_data_t should be reset to zero when it's allocated */
6084         WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
6085 
6086         reset_deferred_meminit(pgdat);
6087         pgdat->node_id = nid;
6088         pgdat->node_start_pfn = node_start_pfn;
6089 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6090         get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6091         pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6092                 (u64)start_pfn << PAGE_SHIFT,
6093                 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6094 #else
6095         start_pfn = node_start_pfn;
6096 #endif
6097         calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6098                                   zones_size, zholes_size);
6099 
6100         alloc_node_mem_map(pgdat);
6101 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6102         printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6103                 nid, (unsigned long)pgdat,
6104                 (unsigned long)pgdat->node_mem_map);
6105 #endif
6106 
6107         free_area_init_core(pgdat);
6108 }
6109 
6110 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6111 
6112 #if MAX_NUMNODES > 1
6113 /*
6114  * Figure out the number of possible node ids.
6115  */
6116 void __init setup_nr_node_ids(void)
6117 {
6118         unsigned int highest;
6119 
6120         highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6121         nr_node_ids = highest + 1;
6122 }
6123 #endif
6124 
6125 /**
6126  * node_map_pfn_alignment - determine the maximum internode alignment
6127  *
6128  * This function should be called after node map is populated and sorted.
6129  * It calculates the maximum power of two alignment which can distinguish
6130  * all the nodes.
6131  *
6132  * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6133  * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)).  If the
6134  * nodes are shifted by 256MiB, 256MiB.  Note that if only the last node is
6135  * shifted, 1GiB is enough and this function will indicate so.
6136  *
6137  * This is used to test whether pfn -> nid mapping of the chosen memory
6138  * model has fine enough granularity to avoid incorrect mapping for the
6139  * populated node map.
6140  *
6141  * Returns the determined alignment in pfn's.  0 if there is no alignment
6142  * requirement (single node).
6143  */
6144 unsigned long __init node_map_pfn_alignment(void)
6145 {
6146         unsigned long accl_mask = 0, last_end = 0;
6147         unsigned long start, end, mask;
6148         int last_nid = -1;
6149         int i, nid;
6150 
6151         for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6152                 if (!start || last_nid < 0 || last_nid == nid) {
6153                         last_nid = nid;
6154                         last_end = end;
6155                         continue;
6156                 }
6157 
6158                 /*
6159                  * Start with a mask granular enough to pin-point to the
6160                  * start pfn and tick off bits one-by-one until it becomes
6161                  * too coarse to separate the current node from the last.
6162                  */
6163                 mask = ~((1 << __ffs(start)) - 1);
6164                 while (mask && last_end <= (start & (mask << 1)))
6165                         mask <<= 1;
6166 
6167                 /* accumulate all internode masks */
6168                 accl_mask |= mask;
6169         }
6170 
6171         /* convert mask to number of pages */
6172         return ~accl_mask + 1;
6173 }
6174 
6175 /* Find the lowest pfn for a node */
6176 static unsigned long __init find_min_pfn_for_node(int nid)
6177 {
6178         unsigned long min_pfn = ULONG_MAX;
6179         unsigned long start_pfn;
6180         int i;
6181 
6182         for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6183                 min_pfn = min(min_pfn, start_pfn);
6184 
6185         if (min_pfn == ULONG_MAX) {
6186                 pr_warn("Could not find start_pfn for node %d\n", nid);
6187                 return 0;
6188         }
6189 
6190         return min_pfn;
6191 }
6192 
6193 /**
6194  * find_min_pfn_with_active_regions - Find the minimum PFN registered
6195  *
6196  * It returns the minimum PFN based on information provided via
6197  * memblock_set_node().
6198  */
6199 unsigned long __init find_min_pfn_with_active_regions(void)
6200 {
6201         return find_min_pfn_for_node(MAX_NUMNODES);
6202 }
6203 
6204 /*
6205  * early_calculate_totalpages()
6206  * Sum pages in active regions for movable zone.
6207  * Populate N_MEMORY for calculating usable_nodes.
6208  */
6209 static unsigned long __init early_calculate_totalpages(void)
6210 {
6211         unsigned long totalpages = 0;
6212         unsigned long start_pfn, end_pfn;
6213         int i, nid;
6214 
6215         for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6216                 unsigned long pages = end_pfn - start_pfn;
6217 
6218                 totalpages += pages;
6219                 if (pages)
6220                         node_set_state(nid, N_MEMORY);
6221         }
6222         return totalpages;
6223 }
6224 
6225 /*
6226  * Find the PFN the Movable zone begins in each node. Kernel memory
6227  * is spread evenly between nodes as long as the nodes have enough
6228  * memory. When they don't, some nodes will have more kernelcore than
6229  * others
6230  */
6231 static void __init find_zone_movable_pfns_for_nodes(void)
6232 {
6233         int i, nid;
6234         unsigned long usable_startpfn;
6235         unsigned long kernelcore_node, kernelcore_remaining;
6236         /* save the state before borrow the nodemask */
6237         nodemask_t saved_node_state = node_states[N_MEMORY];
6238         unsigned long totalpages = early_calculate_totalpages();
6239         int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6240         struct memblock_region *r;
6241 
6242         /* Need to find movable_zone earlier when movable_node is specified. */
6243         find_usable_zone_for_movable();
6244 
6245         /*
6246          * If movable_node is specified, ignore kernelcore and movablecore
6247          * options.
6248          */
6249         if (movable_node_is_enabled()) {
6250                 for_each_memblock(memory, r) {
6251                         if (!memblock_is_hotpluggable(r))
6252                                 continue;
6253 
6254                         nid = r->nid;
6255 
6256                         usable_startpfn = PFN_DOWN(r->base);
6257                         zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6258                                 min(usable_startpfn, zone_movable_pfn[nid]) :
6259                                 usable_startpfn;
6260                 }
6261 
6262                 goto out2;
6263         }
6264 
6265         /*
6266          * If kernelcore=mirror is specified, ignore movablecore option
6267          */
6268         if (mirrored_kernelcore) {
6269                 bool mem_below_4gb_not_mirrored = false;
6270 
6271                 for_each_memblock(memory, r) {
6272                         if (memblock_is_mirror(r))
6273                                 continue;
6274 
6275                         nid = r->nid;
6276 
6277                         usable_startpfn = memblock_region_memory_base_pfn(r);
6278 
6279                         if (usable_startpfn < 0x100000) {
6280                                 mem_below_4gb_not_mirrored = true;
6281                                 continue;
6282                         }
6283 
6284                         zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6285                                 min(usable_startpfn, zone_movable_pfn[nid]) :
6286                                 usable_startpfn;
6287                 }
6288 
6289                 if (mem_below_4gb_not_mirrored)
6290                         pr_warn("This configuration results in unmirrored kernel memory.");
6291 
6292                 goto out2;
6293         }
6294 
6295         /*
6296          * If movablecore=nn[KMG] was specified, calculate what size of
6297          * kernelcore that corresponds so that memory usable for
6298          * any allocation type is evenly spread. If both kernelcore
6299          * and movablecore are specified, then the value of kernelcore
6300          * will be used for required_kernelcore if it's greater than
6301          * what movablecore would have allowed.
6302          */
6303         if (required_movablecore) {
6304                 unsigned long corepages;
6305 
6306                 /*
6307                  * Round-up so that ZONE_MOVABLE is at least as large as what
6308                  * was requested by the user
6309                  */
6310                 required_movablecore =
6311                         roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6312                 required_movablecore = min(totalpages, required_movablecore);
6313                 corepages = totalpages - required_movablecore;
6314 
6315                 required_kernelcore = max(required_kernelcore, corepages);
6316         }
6317 
6318         /*
6319          * If kernelcore was not specified or kernelcore size is larger
6320          * than totalpages, there is no ZONE_MOVABLE.
6321          */
6322         if (!required_kernelcore || required_kernelcore >= totalpages)
6323                 goto out;
6324 
6325         /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6326         usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6327 
6328 restart:
6329         /* Spread kernelcore memory as evenly as possible throughout nodes */
6330         kernelcore_node = required_kernelcore / usable_nodes;
6331         for_each_node_state(nid, N_MEMORY) {
6332                 unsigned long start_pfn, end_pfn;
6333 
6334                 /*
6335                  * Recalculate kernelcore_node if the division per node
6336                  * now exceeds what is necessary to satisfy the requested
6337                  * amount of memory for the kernel
6338                  */
6339                 if (required_kernelcore < kernelcore_node)
6340                         kernelcore_node = required_kernelcore / usable_nodes;
6341 
6342                 /*
6343                  * As the map is walked, we track how much memory is usable
6344                  * by the kernel using kernelcore_remaining. When it is
6345                  * 0, the rest of the node is usable by ZONE_MOVABLE
6346                  */
6347                 kernelcore_remaining = kernelcore_node;
6348 
6349                 /* Go through each range of PFNs within this node */
6350                 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6351                         unsigned long size_pages;
6352 
6353                         start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6354                         if (start_pfn >= end_pfn)
6355                                 continue;
6356 
6357                         /* Account for what is only usable for kernelcore */
6358                         if (start_pfn < usable_startpfn) {
6359                                 unsigned long kernel_pages;
6360                                 kernel_pages = min(end_pfn, usable_startpfn)
6361                                                                 - start_pfn;
6362 
6363                                 kernelcore_remaining -= min(kernel_pages,
6364                                                         kernelcore_remaining);
6365                                 required_kernelcore -= min(kernel_pages,
6366                                                         required_kernelcore);
6367 
6368                                 /* Continue if range is now fully accounted */
6369                                 if (end_pfn <= usable_startpfn) {
6370 
6371                                         /*
6372                                          * Push zone_movable_pfn to the end so
6373                                          * that if we have to rebalance
6374                                          * kernelcore across nodes, we will
6375                                          * not double account here
6376                                          */
6377                                         zone_movable_pfn[nid] = end_pfn;
6378                                         continue;
6379                                 }
6380                                 start_pfn = usable_startpfn;
6381                         }
6382 
6383                         /*
6384                          * The usable PFN range for ZONE_MOVABLE is from
6385                          * start_pfn->end_pfn. Calculate size_pages as the
6386                          * number of pages used as kernelcore
6387                          */
6388                         size_pages = end_pfn - start_pfn;
6389                         if (size_pages > kernelcore_remaining)
6390                                 size_pages = kernelcore_remaining;
6391                         zone_movable_pfn[nid] = start_pfn + size_pages;
6392 
6393                         /*
6394                          * Some kernelcore has been met, update counts and
6395                          * break if the kernelcore for this node has been
6396                          * satisfied
6397                          */
6398                         required_kernelcore -= min(required_kernelcore,
6399                                                                 size_pages);
6400                         kernelcore_remaining -= size_pages;
6401                         if (!kernelcore_remaining)
6402                                 break;
6403                 }
6404         }
6405 
6406         /*
6407          * If there is still required_kernelcore, we do another pass with one
6408          * less node in the count. This will push zone_movable_pfn[nid] further
6409          * along on the nodes that still have memory until kernelcore is
6410          * satisfied
6411          */
6412         usable_nodes--;
6413         if (usable_nodes && required_kernelcore > usable_nodes)
6414                 goto restart;
6415 
6416 out2:
6417         /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6418         for (nid = 0; nid < MAX_NUMNODES; nid++)
6419                 zone_movable_pfn[nid] =
6420                         roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6421 
6422 out:
6423         /* restore the node_state */
6424         node_states[N_MEMORY] = saved_node_state;
6425 }
6426 
6427 /* Any regular or high memory on that node ? */
6428 static void check_for_memory(pg_data_t *pgdat, int nid)
6429 {
6430         enum zone_type zone_type;
6431 
6432         if (N_MEMORY == N_NORMAL_MEMORY)
6433                 return;
6434 
6435         for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6436                 struct zone *zone = &pgdat->node_zones[zone_type];
6437                 if (populated_zone(zone)) {
6438                         node_set_state(nid, N_HIGH_MEMORY);
6439                         if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6440                             zone_type <= ZONE_NORMAL)
6441                                 node_set_state(nid, N_NORMAL_MEMORY);
6442                         break;
6443                 }
6444         }
6445 }
6446 
6447 /**
6448  * free_area_init_nodes - Initialise all pg_data_t and zone data
6449  * @max_zone_pfn: an array of max PFNs for each zone
6450  *
6451  * This will call free_area_init_node() for each active node in the system.
6452  * Using the page ranges provided by memblock_set_node(), the size of each
6453  * zone in each node and their holes is calculated. If the maximum PFN
6454  * between two adjacent zones match, it is assumed that the zone is empty.
6455  * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6456  * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6457  * starts where the previous one ended. For example, ZONE_DMA32 starts
6458  * at arch_max_dma_pfn.
6459  */
6460 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6461 {
6462         unsigned long start_pfn, end_pfn;
6463         int i, nid;
6464 
6465         /* Record where the zone boundaries are */
6466         memset(arch_zone_lowest_possible_pfn, 0,
6467                                 sizeof(arch_zone_lowest_possible_pfn));
6468         memset(arch_zone_highest_possible_pfn, 0,
6469                                 sizeof(arch_zone_highest_possible_pfn));
6470         arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
6471         arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
6472         for (i = 1; i < MAX_NR_ZONES; i++) {
6473                 if (i == ZONE_MOVABLE)
6474                         continue;
6475                 arch_zone_lowest_possible_pfn[i] =
6476                         arch_zone_highest_possible_pfn[i-1];
6477                 arch_zone_highest_possible_pfn[i] =
6478                         max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
6479         }
6480         arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6481         arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6482 
6483         /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6484         memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6485         find_zone_movable_pfns_for_nodes();
6486 
6487         /* Print out the zone ranges */
6488         pr_info("Zone ranges:\n");
6489         for (i = 0; i < MAX_NR_ZONES; i++) {
6490                 if (i == ZONE_MOVABLE)
6491                         continue;
6492                 pr_info("  %-8s ", zone_names[i]);
6493                 if (arch_zone_lowest_possible_pfn[i] ==
6494                                 arch_zone_highest_possible_pfn[i])
6495                         pr_cont("empty\n");
6496                 else
6497                         pr_cont("[mem %#018Lx-%#018Lx]\n",
6498                                 (u64)arch_zone_lowest_possible_pfn[i]
6499                                         << PAGE_SHIFT,
6500                                 ((u64)arch_zone_highest_possible_pfn[i]
6501                                         << PAGE_SHIFT) - 1);
6502         }
6503 
6504         /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6505         pr_info("Movable zone start for each node\n");
6506         for (i = 0; i < MAX_NUMNODES; i++) {
6507                 if (zone_movable_pfn[i])
6508                         pr_info("  Node %d: %#018Lx\n", i,
6509                                (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6510         }
6511 
6512         /* Print out the early node map */
6513         pr_info("Early memory node ranges\n");
6514         for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6515                 pr_info("  node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6516                         (u64)start_pfn << PAGE_SHIFT,
6517                         ((u64)end_pfn << PAGE_SHIFT) - 1);
6518 
6519         /* Initialise every node */
6520         mminit_verify_pageflags_layout();
6521         setup_nr_node_ids();
6522         for_each_online_node(nid) {
6523                 pg_data_t *pgdat = NODE_DATA(nid);
6524                 free_area_init_node(nid, NULL,
6525                                 find_min_pfn_for_node(nid), NULL);
6526 
6527                 /* Any memory on that node */
6528                 if (pgdat->node_present_pages)
6529                         node_set_state(nid, N_MEMORY);
6530                 check_for_memory(pgdat, nid);
6531         }
6532 }
6533 
6534 static int __init cmdline_parse_core(char *p, unsigned long *core)
6535 {
6536         unsigned long long coremem;
6537         if (!p)
6538                 return -EINVAL;
6539 
6540         coremem = memparse(p, &p);
6541         *core = coremem >> PAGE_SHIFT;
6542 
6543         /* Paranoid check that UL is enough for the coremem value */
6544         WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6545 
6546         return 0;
6547 }
6548 
6549 /*
6550  * kernelcore=size sets the amount of memory for use for allocations that
6551  * cannot be reclaimed or migrated.
6552  */
6553 static int __init cmdline_parse_kernelcore(char *p)
6554 {
6555         /* parse kernelcore=mirror */
6556         if (parse_option_str(p, "mirror")) {
6557                 mirrored_kernelcore = true;
6558                 return 0;
6559         }
6560 
6561         return cmdline_parse_core(p, &required_kernelcore);
6562 }
6563 
6564 /*
6565  * movablecore=size sets the amount of memory for use for allocations that
6566  * can be reclaimed or migrated.
6567  */
6568 static int __init cmdline_parse_movablecore(char *p)
6569 {
6570         return cmdline_parse_core(p, &required_movablecore);
6571 }
6572 
6573 early_param("kernelcore", cmdline_parse_kernelcore);
6574 early_param("movablecore", cmdline_parse_movablecore);
6575 
6576 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6577 
6578 void adjust_managed_page_count(struct page *page, long count)
6579 {
6580         spin_lock(&managed_page_count_lock);
6581         page_zone(page)->managed_pages += count;
6582         totalram_pages += count;
6583 #ifdef CONFIG_HIGHMEM
6584         if (PageHighMem(page))
6585                 totalhigh_pages += count;
6586 #endif
6587         spin_unlock(&managed_page_count_lock);
6588 }
6589 EXPORT_SYMBOL(adjust_managed_page_count);
6590 
6591 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6592 {
6593         void *pos;
6594         unsigned long pages = 0;
6595 
6596         start = (void *)PAGE_ALIGN((unsigned long)start);
6597         end = (void *)((unsigned long)end & PAGE_MASK);
6598         for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6599                 if ((unsigned int)poison <= 0xFF)
6600                         memset(pos, poison, PAGE_SIZE);
6601                 free_reserved_page(virt_to_page(pos));
6602         }
6603 
6604         if (pages && s)
6605                 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6606                         s, pages << (PAGE_SHIFT - 10), start, end);
6607 
6608         return pages;
6609 }
6610 EXPORT_SYMBOL(free_reserved_area);
6611 
6612 #ifdef  CONFIG_HIGHMEM
6613 void free_highmem_page(struct page *page)
6614 {
6615         __free_reserved_page(page);
6616         totalram_pages++;
6617         page_zone(page)->managed_pages++;
6618         totalhigh_pages++;
6619 }
6620 #endif
6621 
6622 
6623 void __init mem_init_print_info(const char *str)
6624 {
6625         unsigned long physpages, codesize, datasize, rosize, bss_size;
6626         unsigned long init_code_size, init_data_size;
6627 
6628         physpages = get_num_physpages();
6629         codesize = _etext - _stext;
6630         datasize = _edata - _sdata;
6631         rosize = __end_rodata - __start_rodata;
6632         bss_size = __bss_stop - __bss_start;
6633         init_data_size = __init_end - __init_begin;
6634         init_code_size = _einittext - _sinittext;
6635 
6636         /*
6637          * Detect special cases and adjust section sizes accordingly:
6638          * 1) .init.* may be embedded into .data sections
6639          * 2) .init.text.* may be out of [__init_begin, __init_end],
6640          *    please refer to arch/tile/kernel/vmlinux.lds.S.
6641          * 3) .rodata.* may be embedded into .text or .data sections.
6642          */
6643 #define adj_init_size(start, end, size, pos, adj) \
6644         do { \
6645                 if (start <= pos && pos < end && size > adj) \
6646                         size -= adj; \
6647         } while (0)
6648 
6649         adj_init_size(__init_begin, __init_end, init_data_size,
6650                      _sinittext, init_code_size);
6651         adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6652         adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6653         adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6654         adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6655 
6656 #undef  adj_init_size
6657 
6658         pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6659 #ifdef  CONFIG_HIGHMEM
6660                 ", %luK highmem"
6661 #endif
6662                 "%s%s)\n",
6663                 nr_free_pages() << (PAGE_SHIFT - 10),
6664                 physpages << (PAGE_SHIFT - 10),
6665                 codesize >> 10, datasize >> 10, rosize >> 10,
6666                 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6667                 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6668                 totalcma_pages << (PAGE_SHIFT - 10),
6669 #ifdef  CONFIG_HIGHMEM
6670                 totalhigh_pages << (PAGE_SHIFT - 10),
6671 #endif
6672                 str ? ", " : "", str ? str : "");
6673 }
6674 
6675 /**
6676  * set_dma_reserve - set the specified number of pages reserved in the first zone
6677  * @new_dma_reserve: The number of pages to mark reserved
6678  *
6679  * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6680  * In the DMA zone, a significant percentage may be consumed by kernel image
6681  * and other unfreeable allocations which can skew the watermarks badly. This
6682  * function may optionally be used to account for unfreeable pages in the
6683  * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6684  * smaller per-cpu batchsize.
6685  */
6686 void __init set_dma_reserve(unsigned long new_dma_reserve)
6687 {
6688         dma_reserve = new_dma_reserve;
6689 }
6690 
6691 void __init free_area_init(unsigned long *zones_size)
6692 {
6693         free_area_init_node(0, zones_size,
6694                         __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6695 }
6696 
6697 static int page_alloc_cpu_notify(struct notifier_block *self,
6698                                  unsigned long action, void *hcpu)
6699 {
6700         int cpu = (unsigned long)hcpu;
6701 
6702         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6703                 lru_add_drain_cpu(cpu);
6704                 drain_pages(cpu);
6705 
6706                 /*
6707                  * Spill the event counters of the dead processor
6708                  * into the current processors event counters.
6709                  * This artificially elevates the count of the current
6710                  * processor.
6711                  */
6712                 vm_events_fold_cpu(cpu);
6713 
6714                 /*
6715                  * Zero the differential counters of the dead processor
6716                  * so that the vm statistics are consistent.
6717                  *
6718                  * This is only okay since the processor is dead and cannot
6719                  * race with what we are doing.
6720                  */
6721                 cpu_vm_stats_fold(cpu);
6722         }
6723         return NOTIFY_OK;
6724 }
6725 
6726 void __init page_alloc_init(void)
6727 {
6728         hotcpu_notifier(page_alloc_cpu_notify, 0);
6729 }
6730 
6731 /*
6732  * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6733  *      or min_free_kbytes changes.
6734  */
6735 static void calculate_totalreserve_pages(void)
6736 {
6737         struct pglist_data *pgdat;
6738         unsigned long reserve_pages = 0;
6739         enum zone_type i, j;
6740 
6741         for_each_online_pgdat(pgdat) {
6742                 for (i = 0; i < MAX_NR_ZONES; i++) {
6743                         struct zone *zone = pgdat->node_zones + i;
6744                         long max = 0;
6745 
6746                         /* Find valid and maximum lowmem_reserve in the zone */
6747                         for (j = i; j < MAX_NR_ZONES; j++) {
6748                                 if (zone->lowmem_reserve[j] > max)
6749                                         max = zone->lowmem_reserve[j];
6750                         }
6751 
6752                         /* we treat the high watermark as reserved pages. */
6753                         max += high_wmark_pages(zone);
6754 
6755                         if (max > zone->managed_pages)
6756                                 max = zone->managed_pages;
6757 
6758                         zone->totalreserve_pages = max;
6759 
6760                         reserve_pages += max;
6761                 }
6762         }
6763         totalreserve_pages = reserve_pages;
6764 }
6765 
6766 /*
6767  * setup_per_zone_lowmem_reserve - called whenever
6768  *      sysctl_lowmem_reserve_ratio changes.  Ensures that each zone
6769  *      has a correct pages reserved value, so an adequate number of
6770  *      pages are left in the zone after a successful __alloc_pages().
6771  */
6772 static void setup_per_zone_lowmem_reserve(void)
6773 {
6774         struct pglist_data *pgdat;
6775         enum zone_type j, idx;
6776 
6777         for_each_online_pgdat(pgdat) {
6778                 for (j = 0; j < MAX_NR_ZONES; j++) {
6779                         struct zone *zone = pgdat->node_zones + j;
6780                         unsigned long managed_pages = zone->managed_pages;
6781 
6782                         zone->lowmem_reserve[j] = 0;
6783 
6784                         idx = j;
6785                         while (idx) {
6786                                 struct zone *lower_zone;
6787 
6788                                 idx--;
6789 
6790                                 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6791                                         sysctl_lowmem_reserve_ratio[idx] = 1;
6792 
6793                                 lower_zone = pgdat->node_zones + idx;
6794                                 lower_zone->lowmem_reserve[j] = managed_pages /
6795                                         sysctl_lowmem_reserve_ratio[idx];
6796                                 managed_pages += lower_zone->managed_pages;
6797                         }
6798                 }
6799         }
6800 
6801         /* update totalreserve_pages */
6802         calculate_totalreserve_pages();
6803 }
6804 
6805 static void __setup_per_zone_wmarks(void)
6806 {
6807         unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6808         unsigned long lowmem_pages = 0;
6809         struct zone *zone;
6810         unsigned long flags;
6811 
6812         /* Calculate total number of !ZONE_HIGHMEM pages */
6813         for_each_zone(zone) {
6814                 if (!is_highmem(zone))
6815                         lowmem_pages += zone->managed_pages;
6816         }
6817 
6818         for_each_zone(zone) {
6819                 u64 tmp;
6820 
6821                 spin_lock_irqsave(&zone->lock, flags);
6822                 tmp = (u64)pages_min * zone->managed_pages;
6823                 do_div(tmp, lowmem_pages);
6824                 if (is_highmem(zone)) {
6825                         /*
6826                          * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6827                          * need highmem pages, so cap pages_min to a small
6828                          * value here.
6829                          *
6830                          * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6831                          * deltas control asynch page reclaim, and so should
6832                          * not be capped for highmem.
6833                          */
6834                         unsigned long min_pages;
6835 
6836                         min_pages = zone->managed_pages / 1024;
6837                         min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6838                         zone->watermark[WMARK_MIN] = min_pages;
6839                 } else {
6840                         /*
6841                          * If it's a lowmem zone, reserve a number of pages
6842                          * proportionate to the zone's size.
6843                          */
6844                         zone->watermark[WMARK_MIN] = tmp;
6845                 }
6846 
6847                 /*
6848                  * Set the kswapd watermarks distance according to the
6849                  * scale factor in proportion to available memory, but
6850                  * ensure a minimum size on small systems.
6851                  */
6852                 tmp = max_t(u64, tmp >> 2,
6853                             mult_frac(zone->managed_pages,
6854                                       watermark_scale_factor, 10000));
6855 
6856                 zone->watermark[WMARK_LOW]  = min_wmark_pages(zone) + tmp;
6857                 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6858 
6859                 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6860                         high_wmark_pages(zone) - low_wmark_pages(zone) -
6861                         atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6862 
6863                 spin_unlock_irqrestore(&zone->lock, flags);
6864         }
6865 
6866         /* update totalreserve_pages */
6867         calculate_totalreserve_pages();
6868 }
6869 
6870 /**
6871  * setup_per_zone_wmarks - called when min_free_kbytes changes
6872  * or when memory is hot-{added|removed}
6873  *
6874  * Ensures that the watermark[min,low,high] values for each zone are set
6875  * correctly with respect to min_free_kbytes.
6876  */
6877 void setup_per_zone_wmarks(void)
6878 {
6879         mutex_lock(&zonelists_mutex);
6880         __setup_per_zone_wmarks();
6881         mutex_unlock(&zonelists_mutex);
6882 }
6883 
6884 /*
6885  * Initialise min_free_kbytes.
6886  *
6887  * For small machines we want it small (128k min).  For large machines
6888  * we want it large (64MB max).  But it is not linear, because network
6889  * bandwidth does not increase linearly with machine size.  We use
6890  *
6891  *      min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6892  *      min_free_kbytes = sqrt(lowmem_kbytes * 16)
6893  *
6894  * which yields
6895  *
6896  * 16MB:        512k
6897  * 32MB:        724k
6898  * 64MB:        1024k
6899  * 128MB:       1448k
6900  * 256MB:       2048k
6901  * 512MB:       2896k
6902  * 1024MB:      4096k
6903  * 2048MB:      5792k
6904  * 4096MB:      8192k
6905  * 8192MB:      11584k
6906  * 16384MB:     16384k
6907  */
6908 int __meminit init_per_zone_wmark_min(void)
6909 {
6910         unsigned long lowmem_kbytes;
6911         int new_min_free_kbytes;
6912 
6913         lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6914         new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6915 
6916         if (new_min_free_kbytes > user_min_free_kbytes) {
6917                 min_free_kbytes = new_min_free_kbytes;
6918                 if (min_free_kbytes < 128)
6919                         min_free_kbytes = 128;
6920                 if (min_free_kbytes > 65536)
6921                         min_free_kbytes = 65536;
6922         } else {
6923                 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6924                                 new_min_free_kbytes, user_min_free_kbytes);
6925         }
6926         setup_per_zone_wmarks();
6927         refresh_zone_stat_thresholds();
6928         setup_per_zone_lowmem_reserve();
6929         return 0;
6930 }
6931 core_initcall(init_per_zone_wmark_min)
6932 
6933 /*
6934  * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6935  *      that we can call two helper functions whenever min_free_kbytes
6936  *      changes.
6937  */
6938 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6939         void __user *buffer, size_t *length, loff_t *ppos)
6940 {
6941         int rc;
6942 
6943         rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6944         if (rc)
6945                 return rc;
6946 
6947         if (write) {
6948                 user_min_free_kbytes = min_free_kbytes;
6949                 setup_per_zone_wmarks();
6950         }
6951         return 0;
6952 }
6953 
6954 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6955         void __user *buffer, size_t *length, loff_t *ppos)
6956 {
6957         int rc;
6958 
6959         rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6960         if (rc)
6961                 return rc;
6962 
6963         if (write)
6964                 setup_per_zone_wmarks();
6965 
6966         return 0;
6967 }
6968 
6969 #ifdef CONFIG_NUMA
6970 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6971         void __user *buffer, size_t *length, loff_t *ppos)
6972 {
6973         struct zone *zone;
6974         int rc;
6975 
6976         rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6977         if (rc)
6978                 return rc;
6979 
6980         for_each_zone(zone)
6981                 zone->min_unmapped_pages = (zone->managed_pages *
6982                                 sysctl_min_unmapped_ratio) / 100;
6983         return 0;
6984 }
6985 
6986 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6987         void __user *buffer, size_t *length, loff_t *ppos)
6988 {
6989         struct zone *zone;
6990         int rc;
6991 
6992         rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6993         if (rc)
6994                 return rc;
6995 
6996         for_each_zone(zone)
6997                 zone->min_slab_pages = (zone->managed_pages *
6998                                 sysctl_min_slab_ratio) / 100;
6999         return 0;
7000 }
7001 #endif
7002 
7003 /*
7004  * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7005  *      proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7006  *      whenever sysctl_lowmem_reserve_ratio changes.
7007  *
7008  * The reserve ratio obviously has absolutely no relation with the
7009  * minimum watermarks. The lowmem reserve ratio can only make sense
7010  * if in function of the boot time zone sizes.
7011  */
7012 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7013         void __user *buffer, size_t *length, loff_t *ppos)
7014 {
7015         proc_dointvec_minmax(table, write, buffer, length, ppos);
7016         setup_per_zone_lowmem_reserve();
7017         return 0;
7018 }
7019 
7020 /*
7021  * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7022  * cpu.  It is the fraction of total pages in each zone that a hot per cpu
7023  * pagelist can have before it gets flushed back to buddy allocator.
7024  */
7025 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7026         void __user *buffer, size_t *length, loff_t *ppos)
7027 {
7028         struct zone *zone;
7029         int old_percpu_pagelist_fraction;
7030         int ret;
7031 
7032         mutex_lock(&pcp_batch_high_lock);
7033         old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7034 
7035         ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7036         if (!write || ret < 0)
7037                 goto out;
7038 
7039         /* Sanity checking to avoid pcp imbalance */
7040         if (percpu_pagelist_fraction &&
7041             percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7042                 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7043                 ret = -EINVAL;
7044                 goto out;
7045         }
7046 
7047         /* No change? */
7048         if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7049                 goto out;
7050 
7051         for_each_populated_zone(zone) {
7052                 unsigned int cpu;
7053 
7054                 for_each_possible_cpu(cpu)
7055                         pageset_set_high_and_batch(zone,
7056                                         per_cpu_ptr(zone->pageset, cpu));
7057         }
7058 out:
7059         mutex_unlock(&pcp_batch_high_lock);
7060         return ret;
7061 }
7062 
7063 #ifdef CONFIG_NUMA
7064 int hashdist = HASHDIST_DEFAULT;
7065 
7066 static int __init set_hashdist(char *str)
7067 {
7068         if (!str)
7069                 return 0;
7070         hashdist = simple_strtoul(str, &str, 0);
7071         return 1;
7072 }
7073 __setup("hashdist=", set_hashdist);
7074 #endif
7075 
7076 /*
7077  * allocate a large system hash table from bootmem
7078  * - it is assumed that the hash table must contain an exact power-of-2
7079  *   quantity of entries
7080  * - limit is the number of hash buckets, not the total allocation size
7081  */
7082 void *__init alloc_large_system_hash(const char *tablename,
7083                                      unsigned long bucketsize,
7084                                      unsigned long numentries,
7085                                      int scale,
7086                                      int flags,
7087                                      unsigned int *_hash_shift,
7088                                      unsigned int *_hash_mask,
7089                                      unsigned long low_limit,
7090                                      unsigned long high_limit)
7091 {
7092         unsigned long long max = high_limit;
7093         unsigned long log2qty, size;
7094         void *table = NULL;
7095 
7096         /* allow the kernel cmdline to have a say */
7097         if (!numentries) {
7098                 /* round applicable memory size up to nearest megabyte */
7099                 numentries = nr_kernel_pages;
7100 
7101                 /* It isn't necessary when PAGE_SIZE >= 1MB */
7102                 if (PAGE_SHIFT < 20)
7103                         numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7104 
7105                 /* limit to 1 bucket per 2^scale bytes of low memory */
7106                 if (scale > PAGE_SHIFT)
7107                         numentries >>= (scale - PAGE_SHIFT);
7108                 else
7109                         numentries <<= (PAGE_SHIFT - scale);
7110 
7111                 /* Make sure we've got at least a 0-order allocation.. */
7112                 if (unlikely(flags & HASH_SMALL)) {
7113                         /* Makes no sense without HASH_EARLY */
7114                         WARN_ON(!(flags & HASH_EARLY));
7115                         if (!(numentries >> *_hash_shift)) {
7116                                 numentries = 1UL << *_hash_shift;
7117                                 BUG_ON(!numentries);
7118                         }
7119                 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7120                         numentries = PAGE_SIZE / bucketsize;
7121         }
7122         numentries = roundup_pow_of_two(numentries);
7123 
7124         /* limit allocation size to 1/16 total memory by default */
7125         if (max == 0) {
7126                 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7127                 do_div(max, bucketsize);
7128         }
7129         max = min(max, 0x80000000ULL);
7130 
7131         if (numentries < low_limit)
7132                 numentries = low_limit;
7133         if (numentries > max)
7134                 numentries = max;
7135 
7136         log2qty = ilog2(numentries);
7137 
7138         do {
7139                 size = bucketsize << log2qty;
7140                 if (flags & HASH_EARLY)
7141                         table = memblock_virt_alloc_nopanic(size, 0);
7142                 else if (hashdist)
7143                         table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7144                 else {
7145                         /*
7146                          * If bucketsize is not a power-of-two, we may free
7147                          * some pages at the end of hash table which
7148                          * alloc_pages_exact() automatically does
7149                          */
7150                         if (get_order(size) < MAX_ORDER) {
7151                                 table = alloc_pages_exact(size, GFP_ATOMIC);
7152                                 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7153                         }
7154                 }
7155         } while (!table && size > PAGE_SIZE && --log2qty);
7156 
7157         if (!table)
7158                 panic("Failed to allocate %s hash table\n", tablename);
7159 
7160         pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7161                 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7162 
7163         if (_hash_shift)
7164                 *_hash_shift = log2qty;
7165         if (_hash_mask)
7166                 *_hash_mask = (1 << log2qty) - 1;
7167 
7168         return table;
7169 }
7170 
7171 /*
7172  * This function checks whether pageblock includes unmovable pages or not.
7173  * If @count is not zero, it is okay to include less @count unmovable pages
7174  *
7175  * PageLRU check without isolation or lru_lock could race so that
7176  * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7177  * expect this function should be exact.
7178  */
7179 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7180                          bool skip_hwpoisoned_pages)
7181 {
7182         unsigned long pfn, iter, found;
7183         int mt;
7184 
7185         /*
7186          * For avoiding noise data, lru_add_drain_all() should be called
7187          * If ZONE_MOVABLE, the zone never contains unmovable pages
7188          */
7189         if (zone_idx(zone) == ZONE_MOVABLE)
7190                 return false;
7191         mt = get_pageblock_migratetype(page);
7192         if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7193                 return false;
7194 
7195         pfn = page_to_pfn(page);
7196         for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7197                 unsigned long check = pfn + iter;
7198 
7199                 if (!pfn_valid_within(check))
7200                         continue;
7201 
7202                 page = pfn_to_page(check);
7203 
7204                 /*
7205                  * Hugepages are not in LRU lists, but they're movable.
7206                  * We need not scan over tail pages bacause we don't
7207                  * handle each tail page individually in migration.
7208                  */
7209                 if (PageHuge(page)) {
7210                         iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7211                         continue;
7212                 }
7213 
7214                 /*
7215                  * We can't use page_count without pin a page
7216                  * because another CPU can free compound page.
7217                  * This check already skips compound tails of THP
7218                  * because their page->_refcount is zero at all time.
7219                  */
7220                 if (!page_ref_count(page)) {
7221                         if (PageBuddy(page))
7222                                 iter += (1 << page_order(page)) - 1;
7223                         continue;
7224                 }
7225 
7226                 /*
7227                  * The HWPoisoned page may be not in buddy system, and
7228                  * page_count() is not 0.
7229                  */
7230                 if (skip_hwpoisoned_pages && PageHWPoison(page))
7231                         continue;
7232 
7233                 if (!PageLRU(page))
7234                         found++;
7235                 /*
7236                  * If there are RECLAIMABLE pages, we need to check
7237                  * it.  But now, memory offline itself doesn't call
7238                  * shrink_node_slabs() and it still to be fixed.
7239                  */
7240                 /*
7241                  * If the page is not RAM, page_count()should be 0.
7242                  * we don't need more check. This is an _used_ not-movable page.
7243                  *
7244                  * The problematic thing here is PG_reserved pages. PG_reserved
7245                  * is set to both of a memory hole page and a _used_ kernel
7246                  * page at boot.
7247                  */
7248                 if (found > count)
7249                         return true;
7250         }
7251         return false;
7252 }
7253 
7254 bool is_pageblock_removable_nolock(struct page *page)
7255 {
7256         struct zone *zone;
7257         unsigned long pfn;
7258 
7259         /*
7260          * We have to be careful here because we are iterating over memory
7261          * sections which are not zone aware so we might end up outside of
7262          * the zone but still within the section.
7263          * We have to take care about the node as well. If the node is offline
7264          * its NODE_DATA will be NULL - see page_zone.
7265          */
7266         if (!node_online(page_to_nid(page)))
7267                 return false;
7268 
7269         zone = page_zone(page);
7270         pfn = page_to_pfn(page);
7271         if (!zone_spans_pfn(zone, pfn))
7272                 return false;
7273 
7274         return !has_unmovable_pages(zone, page, 0, true);
7275 }
7276 
7277 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7278 
7279 static unsigned long pfn_max_align_down(unsigned long pfn)
7280 {
7281         return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7282                              pageblock_nr_pages) - 1);
7283 }
7284 
7285 static unsigned long pfn_max_align_up(unsigned long pfn)
7286 {
7287         return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7288                                 pageblock_nr_pages));
7289 }
7290 
7291 /* [start, end) must belong to a single zone. */
7292 static int __alloc_contig_migrate_range(struct compact_control *cc,
7293                                         unsigned long start, unsigned long end)
7294 {
7295         /* This function is based on compact_zone() from compaction.c. */
7296         unsigned long nr_reclaimed;
7297         unsigned long pfn = start;
7298         unsigned int tries = 0;
7299         int ret = 0;
7300 
7301         migrate_prep();
7302 
7303         while (pfn < end || !list_empty(&cc->migratepages)) {
7304                 if (fatal_signal_pending(current)) {
7305                         ret = -EINTR;
7306                         break;
7307                 }
7308 
7309                 if (list_empty(&cc->migratepages)) {
7310                         cc->nr_migratepages = 0;
7311                         pfn = isolate_migratepages_range(cc, pfn, end);
7312                         if (!pfn) {
7313                                 ret = -EINTR;
7314                                 break;
7315                         }
7316                         tries = 0;
7317                 } else if (++tries == 5) {
7318                         ret = ret < 0 ? ret : -EBUSY;
7319                         break;
7320                 }
7321 
7322                 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7323                                                         &cc->migratepages);
7324                 cc->nr_migratepages -= nr_reclaimed;
7325 
7326                 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7327                                     NULL, 0, cc->mode, MR_CMA);
7328         }
7329         if (ret < 0) {
7330                 putback_movable_pages(&cc->migratepages);
7331                 return ret;
7332         }
7333         return 0;
7334 }
7335 
7336 /**
7337  * alloc_contig_range() -- tries to allocate given range of pages
7338  * @start:      start PFN to allocate
7339  * @end:        one-past-the-last PFN to allocate
7340  * @migratetype:        migratetype of the underlaying pageblocks (either
7341  *                      #MIGRATE_MOVABLE or #MIGRATE_CMA).  All pageblocks
7342  *                      in range must have the same migratetype and it must
7343  *                      be either of the two.
7344  *
7345  * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7346  * aligned, however it's the caller's responsibility to guarantee that
7347  * we are the only thread that changes migrate type of pageblocks the
7348  * pages fall in.
7349  *
7350  * The PFN range must belong to a single zone.
7351  *
7352  * Returns zero on success or negative error code.  On success all
7353  * pages which PFN is in [start, end) are allocated for the caller and
7354  * need to be freed with free_contig_range().
7355  */
7356 int alloc_contig_range(unsigned long start, unsigned long end,
7357                        unsigned migratetype)
7358 {
7359         unsigned long outer_start, outer_end;
7360         unsigned int order;
7361         int ret = 0;
7362 
7363         struct compact_control cc = {
7364                 .nr_migratepages = 0,
7365                 .order = -1,
7366                 .zone = page_zone(pfn_to_page(start)),
7367                 .mode = MIGRATE_SYNC,
7368                 .ignore_skip_hint = true,
7369         };
7370         INIT_LIST_HEAD(&cc.migratepages);
7371 
7372         /*
7373          * What we do here is we mark all pageblocks in range as
7374          * MIGRATE_ISOLATE.  Because pageblock and max order pages may
7375          * have different sizes, and due to the way page allocator
7376          * work, we align the range to biggest of the two pages so
7377          * that page allocator won't try to merge buddies from
7378          * different pageblocks and change MIGRATE_ISOLATE to some
7379          * other migration type.
7380          *
7381          * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7382          * migrate the pages from an unaligned range (ie. pages that
7383          * we are interested in).  This will put all the pages in
7384          * range back to page allocator as MIGRATE_ISOLATE.
7385          *
7386          * When this is done, we take the pages in range from page
7387          * allocator removing them from the buddy system.  This way
7388          * page allocator will never consider using them.
7389          *
7390          * This lets us mark the pageblocks back as
7391          * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7392          * aligned range but not in the unaligned, original range are
7393          * put back to page allocator so that buddy can use them.
7394          */
7395 
7396         ret = start_isolate_page_range(pfn_max_align_down(start),
7397                                        pfn_max_align_up(end), migratetype,
7398                                        false);
7399         if (ret)
7400                 return ret;
7401 
7402         /*
7403          * In case of -EBUSY, we'd like to know which page causes problem.
7404          * So, just fall through. We will check it in test_pages_isolated().
7405          */
7406         ret = __alloc_contig_migrate_range(&cc, start, end);
7407         if (ret && ret != -EBUSY)
7408                 goto done;
7409 
7410         /*
7411          * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7412          * aligned blocks that are marked as MIGRATE_ISOLATE.  What's
7413          * more, all pages in [start, end) are free in page allocator.
7414          * What we are going to do is to allocate all pages from
7415          * [start, end) (that is remove them from page allocator).
7416          *
7417          * The only problem is that pages at the beginning and at the
7418          * end of interesting range may be not aligned with pages that
7419          * page allocator holds, ie. they can be part of higher order
7420          * pages.  Because of this, we reserve the bigger range and
7421          * once this is done free the pages we are not interested in.
7422          *
7423          * We don't have to hold zone->lock here because the pages are
7424          * isolated thus they won't get removed from buddy.
7425          */
7426 
7427         lru_add_drain_all();
7428         drain_all_pages(cc.zone);
7429 
7430         order = 0;
7431         outer_start = start;
7432         while (!PageBuddy(pfn_to_page(outer_start))) {
7433                 if (++order >= MAX_ORDER) {
7434                         outer_start = start;
7435                         break;
7436                 }
7437                 outer_start &= ~0UL << order;
7438         }
7439 
7440         if (outer_start != start) {
7441                 order = page_order(pfn_to_page(outer_start));
7442 
7443                 /*
7444                  * outer_start page could be small order buddy page and
7445                  * it doesn't include start page. Adjust outer_start
7446                  * in this case to report failed page properly
7447                  * on tracepoint in test_pages_isolated()
7448                  */
7449                 if (outer_start + (1UL << order) <= start)
7450                         outer_start = start;
7451         }
7452 
7453         /* Make sure the range is really isolated. */
7454         if (test_pages_isolated(outer_start, end, false)) {
7455                 pr_info("%s: [%lx, %lx) PFNs busy\n",
7456                         __func__, outer_start, end);
7457                 ret = -EBUSY;
7458                 goto done;
7459         }
7460 
7461         /* Grab isolated pages from freelists. */
7462         outer_end = isolate_freepages_range(&cc, outer_start, end);
7463         if (!outer_end) {
7464                 ret = -EBUSY;
7465                 goto done;
7466         }
7467 
7468         /* Free head and tail (if any) */
7469         if (start != outer_start)
7470                 free_contig_range(outer_start, start - outer_start);
7471         if (end != outer_end)
7472                 free_contig_range(end, outer_end - end);
7473 
7474 done:
7475         undo_isolate_page_range(pfn_max_align_down(start),
7476                                 pfn_max_align_up(end), migratetype);
7477         return ret;
7478 }
7479 
7480 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7481 {
7482         unsigned int count = 0;
7483 
7484         for (; nr_pages--; pfn++) {
7485                 struct page *page = pfn_to_page(pfn);
7486 
7487                 count += page_count(page) != 1;
7488                 __free_page(page);
7489         }
7490         WARN(count != 0, "%d pages are still in use!\n", count);
7491 }
7492 #endif
7493 
7494 #ifdef CONFIG_MEMORY_HOTPLUG
7495 /*
7496  * The zone indicated has a new number of managed_pages; batch sizes and percpu
7497  * page high values need to be recalulated.
7498  */
7499 void __meminit zone_pcp_update(struct zone *zone)
7500 {
7501         unsigned cpu;
7502         mutex_lock(&pcp_batch_high_lock);
7503         for_each_possible_cpu(cpu)
7504                 pageset_set_high_and_batch(zone,
7505                                 per_cpu_ptr(zone->pageset, cpu));
7506         mutex_unlock(&pcp_batch_high_lock);
7507 }
7508 #endif
7509 
7510 void zone_pcp_reset(struct zone *zone)
7511 {
7512         unsigned long flags;
7513         int cpu;
7514         struct per_cpu_pageset *pset;
7515 
7516         /* avoid races with drain_pages()  */
7517         local_irq_save(flags);
7518         if (zone->pageset != &boot_pageset) {
7519                 for_each_online_cpu(cpu) {
7520                         pset = per_cpu_ptr(zone->pageset, cpu);
7521                         drain_zonestat(zone, pset);
7522                 }
7523                 free_percpu(zone->pageset);
7524                 zone->pageset = &boot_pageset;
7525         }
7526         local_irq_restore(flags);
7527 }
7528 
7529 #ifdef CONFIG_MEMORY_HOTREMOVE
7530 /*
7531  * All pages in the range must be in a single zone and isolated
7532  * before calling this.
7533  */
7534 void
7535 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7536 {
7537         struct page *page;
7538         struct zone *zone;
7539         unsigned int order, i;
7540         unsigned long pfn;
7541         unsigned long flags;
7542         /* find the first valid pfn */
7543         for (pfn = start_pfn; pfn < end_pfn; pfn++)
7544                 if (pfn_valid(pfn))
7545                         break;
7546         if (pfn == end_pfn)
7547                 return;
7548         zone = page_zone(pfn_to_page(pfn));
7549         spin_lock_irqsave(&zone->lock, flags);
7550         pfn = start_pfn;
7551         while (pfn < end_pfn) {
7552                 if (!pfn_valid(pfn)) {
7553                         pfn++;
7554                         continue;
7555                 }
7556                 page = pfn_to_page(pfn);
7557                 /*
7558                  * The HWPoisoned page may be not in buddy system, and
7559                  * page_count() is not 0.
7560                  */
7561                 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7562                         pfn++;
7563                         SetPageReserved(page);
7564                         continue;
7565                 }
7566 
7567                 BUG_ON(page_count(page));
7568                 BUG_ON(!PageBuddy(page));
7569                 order = page_order(page);
7570 #ifdef CONFIG_DEBUG_VM
7571                 pr_info("remove from free list %lx %d %lx\n",
7572                         pfn, 1 << order, end_pfn);
7573 #endif
7574                 list_del(&page->lru);
7575                 rmv_page_order(page);
7576                 zone->free_area[order].nr_free--;
7577                 for (i = 0; i < (1 << order); i++)
7578                         SetPageReserved((page+i));
7579                 pfn += (1 << order);
7580         }
7581         spin_unlock_irqrestore(&zone->lock, flags);
7582 }
7583 #endif
7584 
7585 bool is_free_buddy_page(struct page *page)
7586 {
7587         struct zone *zone = page_zone(page);
7588         unsigned long pfn = page_to_pfn(page);
7589         unsigned long flags;
7590         unsigned int order;
7591 
7592         spin_lock_irqsave(&zone->lock, flags);
7593         for (order = 0; order < MAX_ORDER; order++) {
7594                 struct page *page_head = page - (pfn & ((1 << order) - 1));
7595 
7596                 if (PageBuddy(page_head) && page_order(page_head) >= order)
7597                         break;
7598         }
7599         spin_unlock_irqrestore(&zone->lock, flags);
7600 
7601         return order < MAX_ORDER;
7602 }
7603 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us