Version:  2.0.40 2.2.26 2.4.37 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 4.8 4.9 4.10

Linux/mm/percpu.c

  1 /*
  2  * mm/percpu.c - percpu memory allocator
  3  *
  4  * Copyright (C) 2009           SUSE Linux Products GmbH
  5  * Copyright (C) 2009           Tejun Heo <tj@kernel.org>
  6  *
  7  * This file is released under the GPLv2.
  8  *
  9  * This is percpu allocator which can handle both static and dynamic
 10  * areas.  Percpu areas are allocated in chunks.  Each chunk is
 11  * consisted of boot-time determined number of units and the first
 12  * chunk is used for static percpu variables in the kernel image
 13  * (special boot time alloc/init handling necessary as these areas
 14  * need to be brought up before allocation services are running).
 15  * Unit grows as necessary and all units grow or shrink in unison.
 16  * When a chunk is filled up, another chunk is allocated.
 17  *
 18  *  c0                           c1                         c2
 19  *  -------------------          -------------------        ------------
 20  * | u0 | u1 | u2 | u3 |        | u0 | u1 | u2 | u3 |      | u0 | u1 | u
 21  *  -------------------  ......  -------------------  ....  ------------
 22  *
 23  * Allocation is done in offset-size areas of single unit space.  Ie,
 24  * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
 25  * c1:u1, c1:u2 and c1:u3.  On UMA, units corresponds directly to
 26  * cpus.  On NUMA, the mapping can be non-linear and even sparse.
 27  * Percpu access can be done by configuring percpu base registers
 28  * according to cpu to unit mapping and pcpu_unit_size.
 29  *
 30  * There are usually many small percpu allocations many of them being
 31  * as small as 4 bytes.  The allocator organizes chunks into lists
 32  * according to free size and tries to allocate from the fullest one.
 33  * Each chunk keeps the maximum contiguous area size hint which is
 34  * guaranteed to be equal to or larger than the maximum contiguous
 35  * area in the chunk.  This helps the allocator not to iterate the
 36  * chunk maps unnecessarily.
 37  *
 38  * Allocation state in each chunk is kept using an array of integers
 39  * on chunk->map.  A positive value in the map represents a free
 40  * region and negative allocated.  Allocation inside a chunk is done
 41  * by scanning this map sequentially and serving the first matching
 42  * entry.  This is mostly copied from the percpu_modalloc() allocator.
 43  * Chunks can be determined from the address using the index field
 44  * in the page struct. The index field contains a pointer to the chunk.
 45  *
 46  * To use this allocator, arch code should do the followings.
 47  *
 48  * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
 49  *   regular address to percpu pointer and back if they need to be
 50  *   different from the default
 51  *
 52  * - use pcpu_setup_first_chunk() during percpu area initialization to
 53  *   setup the first chunk containing the kernel static percpu area
 54  */
 55 
 56 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 57 
 58 #include <linux/bitmap.h>
 59 #include <linux/bootmem.h>
 60 #include <linux/err.h>
 61 #include <linux/list.h>
 62 #include <linux/log2.h>
 63 #include <linux/mm.h>
 64 #include <linux/module.h>
 65 #include <linux/mutex.h>
 66 #include <linux/percpu.h>
 67 #include <linux/pfn.h>
 68 #include <linux/slab.h>
 69 #include <linux/spinlock.h>
 70 #include <linux/vmalloc.h>
 71 #include <linux/workqueue.h>
 72 #include <linux/kmemleak.h>
 73 
 74 #include <asm/cacheflush.h>
 75 #include <asm/sections.h>
 76 #include <asm/tlbflush.h>
 77 #include <asm/io.h>
 78 
 79 #define PCPU_SLOT_BASE_SHIFT            5       /* 1-31 shares the same slot */
 80 #define PCPU_DFL_MAP_ALLOC              16      /* start a map with 16 ents */
 81 #define PCPU_ATOMIC_MAP_MARGIN_LOW      32
 82 #define PCPU_ATOMIC_MAP_MARGIN_HIGH     64
 83 #define PCPU_EMPTY_POP_PAGES_LOW        2
 84 #define PCPU_EMPTY_POP_PAGES_HIGH       4
 85 
 86 #ifdef CONFIG_SMP
 87 /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
 88 #ifndef __addr_to_pcpu_ptr
 89 #define __addr_to_pcpu_ptr(addr)                                        \
 90         (void __percpu *)((unsigned long)(addr) -                       \
 91                           (unsigned long)pcpu_base_addr +               \
 92                           (unsigned long)__per_cpu_start)
 93 #endif
 94 #ifndef __pcpu_ptr_to_addr
 95 #define __pcpu_ptr_to_addr(ptr)                                         \
 96         (void __force *)((unsigned long)(ptr) +                         \
 97                          (unsigned long)pcpu_base_addr -                \
 98                          (unsigned long)__per_cpu_start)
 99 #endif
100 #else   /* CONFIG_SMP */
101 /* on UP, it's always identity mapped */
102 #define __addr_to_pcpu_ptr(addr)        (void __percpu *)(addr)
103 #define __pcpu_ptr_to_addr(ptr)         (void __force *)(ptr)
104 #endif  /* CONFIG_SMP */
105 
106 struct pcpu_chunk {
107         struct list_head        list;           /* linked to pcpu_slot lists */
108         int                     free_size;      /* free bytes in the chunk */
109         int                     contig_hint;    /* max contiguous size hint */
110         void                    *base_addr;     /* base address of this chunk */
111 
112         int                     map_used;       /* # of map entries used before the sentry */
113         int                     map_alloc;      /* # of map entries allocated */
114         int                     *map;           /* allocation map */
115         struct list_head        map_extend_list;/* on pcpu_map_extend_chunks */
116 
117         void                    *data;          /* chunk data */
118         int                     first_free;     /* no free below this */
119         bool                    immutable;      /* no [de]population allowed */
120         int                     nr_populated;   /* # of populated pages */
121         unsigned long           populated[];    /* populated bitmap */
122 };
123 
124 static int pcpu_unit_pages __read_mostly;
125 static int pcpu_unit_size __read_mostly;
126 static int pcpu_nr_units __read_mostly;
127 static int pcpu_atom_size __read_mostly;
128 static int pcpu_nr_slots __read_mostly;
129 static size_t pcpu_chunk_struct_size __read_mostly;
130 
131 /* cpus with the lowest and highest unit addresses */
132 static unsigned int pcpu_low_unit_cpu __read_mostly;
133 static unsigned int pcpu_high_unit_cpu __read_mostly;
134 
135 /* the address of the first chunk which starts with the kernel static area */
136 void *pcpu_base_addr __read_mostly;
137 EXPORT_SYMBOL_GPL(pcpu_base_addr);
138 
139 static const int *pcpu_unit_map __read_mostly;          /* cpu -> unit */
140 const unsigned long *pcpu_unit_offsets __read_mostly;   /* cpu -> unit offset */
141 
142 /* group information, used for vm allocation */
143 static int pcpu_nr_groups __read_mostly;
144 static const unsigned long *pcpu_group_offsets __read_mostly;
145 static const size_t *pcpu_group_sizes __read_mostly;
146 
147 /*
148  * The first chunk which always exists.  Note that unlike other
149  * chunks, this one can be allocated and mapped in several different
150  * ways and thus often doesn't live in the vmalloc area.
151  */
152 static struct pcpu_chunk *pcpu_first_chunk;
153 
154 /*
155  * Optional reserved chunk.  This chunk reserves part of the first
156  * chunk and serves it for reserved allocations.  The amount of
157  * reserved offset is in pcpu_reserved_chunk_limit.  When reserved
158  * area doesn't exist, the following variables contain NULL and 0
159  * respectively.
160  */
161 static struct pcpu_chunk *pcpu_reserved_chunk;
162 static int pcpu_reserved_chunk_limit;
163 
164 static DEFINE_SPINLOCK(pcpu_lock);      /* all internal data structures */
165 static DEFINE_MUTEX(pcpu_alloc_mutex);  /* chunk create/destroy, [de]pop, map ext */
166 
167 static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
168 
169 /* chunks which need their map areas extended, protected by pcpu_lock */
170 static LIST_HEAD(pcpu_map_extend_chunks);
171 
172 /*
173  * The number of empty populated pages, protected by pcpu_lock.  The
174  * reserved chunk doesn't contribute to the count.
175  */
176 static int pcpu_nr_empty_pop_pages;
177 
178 /*
179  * Balance work is used to populate or destroy chunks asynchronously.  We
180  * try to keep the number of populated free pages between
181  * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
182  * empty chunk.
183  */
184 static void pcpu_balance_workfn(struct work_struct *work);
185 static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
186 static bool pcpu_async_enabled __read_mostly;
187 static bool pcpu_atomic_alloc_failed;
188 
189 static void pcpu_schedule_balance_work(void)
190 {
191         if (pcpu_async_enabled)
192                 schedule_work(&pcpu_balance_work);
193 }
194 
195 static bool pcpu_addr_in_first_chunk(void *addr)
196 {
197         void *first_start = pcpu_first_chunk->base_addr;
198 
199         return addr >= first_start && addr < first_start + pcpu_unit_size;
200 }
201 
202 static bool pcpu_addr_in_reserved_chunk(void *addr)
203 {
204         void *first_start = pcpu_first_chunk->base_addr;
205 
206         return addr >= first_start &&
207                 addr < first_start + pcpu_reserved_chunk_limit;
208 }
209 
210 static int __pcpu_size_to_slot(int size)
211 {
212         int highbit = fls(size);        /* size is in bytes */
213         return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
214 }
215 
216 static int pcpu_size_to_slot(int size)
217 {
218         if (size == pcpu_unit_size)
219                 return pcpu_nr_slots - 1;
220         return __pcpu_size_to_slot(size);
221 }
222 
223 static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
224 {
225         if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
226                 return 0;
227 
228         return pcpu_size_to_slot(chunk->free_size);
229 }
230 
231 /* set the pointer to a chunk in a page struct */
232 static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
233 {
234         page->index = (unsigned long)pcpu;
235 }
236 
237 /* obtain pointer to a chunk from a page struct */
238 static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
239 {
240         return (struct pcpu_chunk *)page->index;
241 }
242 
243 static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
244 {
245         return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
246 }
247 
248 static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
249                                      unsigned int cpu, int page_idx)
250 {
251         return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
252                 (page_idx << PAGE_SHIFT);
253 }
254 
255 static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
256                                            int *rs, int *re, int end)
257 {
258         *rs = find_next_zero_bit(chunk->populated, end, *rs);
259         *re = find_next_bit(chunk->populated, end, *rs + 1);
260 }
261 
262 static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
263                                          int *rs, int *re, int end)
264 {
265         *rs = find_next_bit(chunk->populated, end, *rs);
266         *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
267 }
268 
269 /*
270  * (Un)populated page region iterators.  Iterate over (un)populated
271  * page regions between @start and @end in @chunk.  @rs and @re should
272  * be integer variables and will be set to start and end page index of
273  * the current region.
274  */
275 #define pcpu_for_each_unpop_region(chunk, rs, re, start, end)               \
276         for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
277              (rs) < (re);                                                   \
278              (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
279 
280 #define pcpu_for_each_pop_region(chunk, rs, re, start, end)                 \
281         for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end));   \
282              (rs) < (re);                                                   \
283              (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
284 
285 /**
286  * pcpu_mem_zalloc - allocate memory
287  * @size: bytes to allocate
288  *
289  * Allocate @size bytes.  If @size is smaller than PAGE_SIZE,
290  * kzalloc() is used; otherwise, vzalloc() is used.  The returned
291  * memory is always zeroed.
292  *
293  * CONTEXT:
294  * Does GFP_KERNEL allocation.
295  *
296  * RETURNS:
297  * Pointer to the allocated area on success, NULL on failure.
298  */
299 static void *pcpu_mem_zalloc(size_t size)
300 {
301         if (WARN_ON_ONCE(!slab_is_available()))
302                 return NULL;
303 
304         if (size <= PAGE_SIZE)
305                 return kzalloc(size, GFP_KERNEL);
306         else
307                 return vzalloc(size);
308 }
309 
310 /**
311  * pcpu_mem_free - free memory
312  * @ptr: memory to free
313  *
314  * Free @ptr.  @ptr should have been allocated using pcpu_mem_zalloc().
315  */
316 static void pcpu_mem_free(void *ptr)
317 {
318         kvfree(ptr);
319 }
320 
321 /**
322  * pcpu_count_occupied_pages - count the number of pages an area occupies
323  * @chunk: chunk of interest
324  * @i: index of the area in question
325  *
326  * Count the number of pages chunk's @i'th area occupies.  When the area's
327  * start and/or end address isn't aligned to page boundary, the straddled
328  * page is included in the count iff the rest of the page is free.
329  */
330 static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
331 {
332         int off = chunk->map[i] & ~1;
333         int end = chunk->map[i + 1] & ~1;
334 
335         if (!PAGE_ALIGNED(off) && i > 0) {
336                 int prev = chunk->map[i - 1];
337 
338                 if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
339                         off = round_down(off, PAGE_SIZE);
340         }
341 
342         if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
343                 int next = chunk->map[i + 1];
344                 int nend = chunk->map[i + 2] & ~1;
345 
346                 if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
347                         end = round_up(end, PAGE_SIZE);
348         }
349 
350         return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
351 }
352 
353 /**
354  * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
355  * @chunk: chunk of interest
356  * @oslot: the previous slot it was on
357  *
358  * This function is called after an allocation or free changed @chunk.
359  * New slot according to the changed state is determined and @chunk is
360  * moved to the slot.  Note that the reserved chunk is never put on
361  * chunk slots.
362  *
363  * CONTEXT:
364  * pcpu_lock.
365  */
366 static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
367 {
368         int nslot = pcpu_chunk_slot(chunk);
369 
370         if (chunk != pcpu_reserved_chunk && oslot != nslot) {
371                 if (oslot < nslot)
372                         list_move(&chunk->list, &pcpu_slot[nslot]);
373                 else
374                         list_move_tail(&chunk->list, &pcpu_slot[nslot]);
375         }
376 }
377 
378 /**
379  * pcpu_need_to_extend - determine whether chunk area map needs to be extended
380  * @chunk: chunk of interest
381  * @is_atomic: the allocation context
382  *
383  * Determine whether area map of @chunk needs to be extended.  If
384  * @is_atomic, only the amount necessary for a new allocation is
385  * considered; however, async extension is scheduled if the left amount is
386  * low.  If !@is_atomic, it aims for more empty space.  Combined, this
387  * ensures that the map is likely to have enough available space to
388  * accomodate atomic allocations which can't extend maps directly.
389  *
390  * CONTEXT:
391  * pcpu_lock.
392  *
393  * RETURNS:
394  * New target map allocation length if extension is necessary, 0
395  * otherwise.
396  */
397 static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
398 {
399         int margin, new_alloc;
400 
401         lockdep_assert_held(&pcpu_lock);
402 
403         if (is_atomic) {
404                 margin = 3;
405 
406                 if (chunk->map_alloc <
407                     chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) {
408                         if (list_empty(&chunk->map_extend_list)) {
409                                 list_add_tail(&chunk->map_extend_list,
410                                               &pcpu_map_extend_chunks);
411                                 pcpu_schedule_balance_work();
412                         }
413                 }
414         } else {
415                 margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
416         }
417 
418         if (chunk->map_alloc >= chunk->map_used + margin)
419                 return 0;
420 
421         new_alloc = PCPU_DFL_MAP_ALLOC;
422         while (new_alloc < chunk->map_used + margin)
423                 new_alloc *= 2;
424 
425         return new_alloc;
426 }
427 
428 /**
429  * pcpu_extend_area_map - extend area map of a chunk
430  * @chunk: chunk of interest
431  * @new_alloc: new target allocation length of the area map
432  *
433  * Extend area map of @chunk to have @new_alloc entries.
434  *
435  * CONTEXT:
436  * Does GFP_KERNEL allocation.  Grabs and releases pcpu_lock.
437  *
438  * RETURNS:
439  * 0 on success, -errno on failure.
440  */
441 static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
442 {
443         int *old = NULL, *new = NULL;
444         size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
445         unsigned long flags;
446 
447         lockdep_assert_held(&pcpu_alloc_mutex);
448 
449         new = pcpu_mem_zalloc(new_size);
450         if (!new)
451                 return -ENOMEM;
452 
453         /* acquire pcpu_lock and switch to new area map */
454         spin_lock_irqsave(&pcpu_lock, flags);
455 
456         if (new_alloc <= chunk->map_alloc)
457                 goto out_unlock;
458 
459         old_size = chunk->map_alloc * sizeof(chunk->map[0]);
460         old = chunk->map;
461 
462         memcpy(new, old, old_size);
463 
464         chunk->map_alloc = new_alloc;
465         chunk->map = new;
466         new = NULL;
467 
468 out_unlock:
469         spin_unlock_irqrestore(&pcpu_lock, flags);
470 
471         /*
472          * pcpu_mem_free() might end up calling vfree() which uses
473          * IRQ-unsafe lock and thus can't be called under pcpu_lock.
474          */
475         pcpu_mem_free(old);
476         pcpu_mem_free(new);
477 
478         return 0;
479 }
480 
481 /**
482  * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
483  * @chunk: chunk the candidate area belongs to
484  * @off: the offset to the start of the candidate area
485  * @this_size: the size of the candidate area
486  * @size: the size of the target allocation
487  * @align: the alignment of the target allocation
488  * @pop_only: only allocate from already populated region
489  *
490  * We're trying to allocate @size bytes aligned at @align.  @chunk's area
491  * at @off sized @this_size is a candidate.  This function determines
492  * whether the target allocation fits in the candidate area and returns the
493  * number of bytes to pad after @off.  If the target area doesn't fit, -1
494  * is returned.
495  *
496  * If @pop_only is %true, this function only considers the already
497  * populated part of the candidate area.
498  */
499 static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
500                             int size, int align, bool pop_only)
501 {
502         int cand_off = off;
503 
504         while (true) {
505                 int head = ALIGN(cand_off, align) - off;
506                 int page_start, page_end, rs, re;
507 
508                 if (this_size < head + size)
509                         return -1;
510 
511                 if (!pop_only)
512                         return head;
513 
514                 /*
515                  * If the first unpopulated page is beyond the end of the
516                  * allocation, the whole allocation is populated;
517                  * otherwise, retry from the end of the unpopulated area.
518                  */
519                 page_start = PFN_DOWN(head + off);
520                 page_end = PFN_UP(head + off + size);
521 
522                 rs = page_start;
523                 pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
524                 if (rs >= page_end)
525                         return head;
526                 cand_off = re * PAGE_SIZE;
527         }
528 }
529 
530 /**
531  * pcpu_alloc_area - allocate area from a pcpu_chunk
532  * @chunk: chunk of interest
533  * @size: wanted size in bytes
534  * @align: wanted align
535  * @pop_only: allocate only from the populated area
536  * @occ_pages_p: out param for the number of pages the area occupies
537  *
538  * Try to allocate @size bytes area aligned at @align from @chunk.
539  * Note that this function only allocates the offset.  It doesn't
540  * populate or map the area.
541  *
542  * @chunk->map must have at least two free slots.
543  *
544  * CONTEXT:
545  * pcpu_lock.
546  *
547  * RETURNS:
548  * Allocated offset in @chunk on success, -1 if no matching area is
549  * found.
550  */
551 static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
552                            bool pop_only, int *occ_pages_p)
553 {
554         int oslot = pcpu_chunk_slot(chunk);
555         int max_contig = 0;
556         int i, off;
557         bool seen_free = false;
558         int *p;
559 
560         for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
561                 int head, tail;
562                 int this_size;
563 
564                 off = *p;
565                 if (off & 1)
566                         continue;
567 
568                 this_size = (p[1] & ~1) - off;
569 
570                 head = pcpu_fit_in_area(chunk, off, this_size, size, align,
571                                         pop_only);
572                 if (head < 0) {
573                         if (!seen_free) {
574                                 chunk->first_free = i;
575                                 seen_free = true;
576                         }
577                         max_contig = max(this_size, max_contig);
578                         continue;
579                 }
580 
581                 /*
582                  * If head is small or the previous block is free,
583                  * merge'em.  Note that 'small' is defined as smaller
584                  * than sizeof(int), which is very small but isn't too
585                  * uncommon for percpu allocations.
586                  */
587                 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
588                         *p = off += head;
589                         if (p[-1] & 1)
590                                 chunk->free_size -= head;
591                         else
592                                 max_contig = max(*p - p[-1], max_contig);
593                         this_size -= head;
594                         head = 0;
595                 }
596 
597                 /* if tail is small, just keep it around */
598                 tail = this_size - head - size;
599                 if (tail < sizeof(int)) {
600                         tail = 0;
601                         size = this_size - head;
602                 }
603 
604                 /* split if warranted */
605                 if (head || tail) {
606                         int nr_extra = !!head + !!tail;
607 
608                         /* insert new subblocks */
609                         memmove(p + nr_extra + 1, p + 1,
610                                 sizeof(chunk->map[0]) * (chunk->map_used - i));
611                         chunk->map_used += nr_extra;
612 
613                         if (head) {
614                                 if (!seen_free) {
615                                         chunk->first_free = i;
616                                         seen_free = true;
617                                 }
618                                 *++p = off += head;
619                                 ++i;
620                                 max_contig = max(head, max_contig);
621                         }
622                         if (tail) {
623                                 p[1] = off + size;
624                                 max_contig = max(tail, max_contig);
625                         }
626                 }
627 
628                 if (!seen_free)
629                         chunk->first_free = i + 1;
630 
631                 /* update hint and mark allocated */
632                 if (i + 1 == chunk->map_used)
633                         chunk->contig_hint = max_contig; /* fully scanned */
634                 else
635                         chunk->contig_hint = max(chunk->contig_hint,
636                                                  max_contig);
637 
638                 chunk->free_size -= size;
639                 *p |= 1;
640 
641                 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
642                 pcpu_chunk_relocate(chunk, oslot);
643                 return off;
644         }
645 
646         chunk->contig_hint = max_contig;        /* fully scanned */
647         pcpu_chunk_relocate(chunk, oslot);
648 
649         /* tell the upper layer that this chunk has no matching area */
650         return -1;
651 }
652 
653 /**
654  * pcpu_free_area - free area to a pcpu_chunk
655  * @chunk: chunk of interest
656  * @freeme: offset of area to free
657  * @occ_pages_p: out param for the number of pages the area occupies
658  *
659  * Free area starting from @freeme to @chunk.  Note that this function
660  * only modifies the allocation map.  It doesn't depopulate or unmap
661  * the area.
662  *
663  * CONTEXT:
664  * pcpu_lock.
665  */
666 static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
667                            int *occ_pages_p)
668 {
669         int oslot = pcpu_chunk_slot(chunk);
670         int off = 0;
671         unsigned i, j;
672         int to_free = 0;
673         int *p;
674 
675         freeme |= 1;    /* we are searching for <given offset, in use> pair */
676 
677         i = 0;
678         j = chunk->map_used;
679         while (i != j) {
680                 unsigned k = (i + j) / 2;
681                 off = chunk->map[k];
682                 if (off < freeme)
683                         i = k + 1;
684                 else if (off > freeme)
685                         j = k;
686                 else
687                         i = j = k;
688         }
689         BUG_ON(off != freeme);
690 
691         if (i < chunk->first_free)
692                 chunk->first_free = i;
693 
694         p = chunk->map + i;
695         *p = off &= ~1;
696         chunk->free_size += (p[1] & ~1) - off;
697 
698         *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
699 
700         /* merge with next? */
701         if (!(p[1] & 1))
702                 to_free++;
703         /* merge with previous? */
704         if (i > 0 && !(p[-1] & 1)) {
705                 to_free++;
706                 i--;
707                 p--;
708         }
709         if (to_free) {
710                 chunk->map_used -= to_free;
711                 memmove(p + 1, p + 1 + to_free,
712                         (chunk->map_used - i) * sizeof(chunk->map[0]));
713         }
714 
715         chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
716         pcpu_chunk_relocate(chunk, oslot);
717 }
718 
719 static struct pcpu_chunk *pcpu_alloc_chunk(void)
720 {
721         struct pcpu_chunk *chunk;
722 
723         chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
724         if (!chunk)
725                 return NULL;
726 
727         chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
728                                                 sizeof(chunk->map[0]));
729         if (!chunk->map) {
730                 pcpu_mem_free(chunk);
731                 return NULL;
732         }
733 
734         chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
735         chunk->map[0] = 0;
736         chunk->map[1] = pcpu_unit_size | 1;
737         chunk->map_used = 1;
738 
739         INIT_LIST_HEAD(&chunk->list);
740         INIT_LIST_HEAD(&chunk->map_extend_list);
741         chunk->free_size = pcpu_unit_size;
742         chunk->contig_hint = pcpu_unit_size;
743 
744         return chunk;
745 }
746 
747 static void pcpu_free_chunk(struct pcpu_chunk *chunk)
748 {
749         if (!chunk)
750                 return;
751         pcpu_mem_free(chunk->map);
752         pcpu_mem_free(chunk);
753 }
754 
755 /**
756  * pcpu_chunk_populated - post-population bookkeeping
757  * @chunk: pcpu_chunk which got populated
758  * @page_start: the start page
759  * @page_end: the end page
760  *
761  * Pages in [@page_start,@page_end) have been populated to @chunk.  Update
762  * the bookkeeping information accordingly.  Must be called after each
763  * successful population.
764  */
765 static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
766                                  int page_start, int page_end)
767 {
768         int nr = page_end - page_start;
769 
770         lockdep_assert_held(&pcpu_lock);
771 
772         bitmap_set(chunk->populated, page_start, nr);
773         chunk->nr_populated += nr;
774         pcpu_nr_empty_pop_pages += nr;
775 }
776 
777 /**
778  * pcpu_chunk_depopulated - post-depopulation bookkeeping
779  * @chunk: pcpu_chunk which got depopulated
780  * @page_start: the start page
781  * @page_end: the end page
782  *
783  * Pages in [@page_start,@page_end) have been depopulated from @chunk.
784  * Update the bookkeeping information accordingly.  Must be called after
785  * each successful depopulation.
786  */
787 static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
788                                    int page_start, int page_end)
789 {
790         int nr = page_end - page_start;
791 
792         lockdep_assert_held(&pcpu_lock);
793 
794         bitmap_clear(chunk->populated, page_start, nr);
795         chunk->nr_populated -= nr;
796         pcpu_nr_empty_pop_pages -= nr;
797 }
798 
799 /*
800  * Chunk management implementation.
801  *
802  * To allow different implementations, chunk alloc/free and
803  * [de]population are implemented in a separate file which is pulled
804  * into this file and compiled together.  The following functions
805  * should be implemented.
806  *
807  * pcpu_populate_chunk          - populate the specified range of a chunk
808  * pcpu_depopulate_chunk        - depopulate the specified range of a chunk
809  * pcpu_create_chunk            - create a new chunk
810  * pcpu_destroy_chunk           - destroy a chunk, always preceded by full depop
811  * pcpu_addr_to_page            - translate address to physical address
812  * pcpu_verify_alloc_info       - check alloc_info is acceptable during init
813  */
814 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
815 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
816 static struct pcpu_chunk *pcpu_create_chunk(void);
817 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
818 static struct page *pcpu_addr_to_page(void *addr);
819 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
820 
821 #ifdef CONFIG_NEED_PER_CPU_KM
822 #include "percpu-km.c"
823 #else
824 #include "percpu-vm.c"
825 #endif
826 
827 /**
828  * pcpu_chunk_addr_search - determine chunk containing specified address
829  * @addr: address for which the chunk needs to be determined.
830  *
831  * RETURNS:
832  * The address of the found chunk.
833  */
834 static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
835 {
836         /* is it in the first chunk? */
837         if (pcpu_addr_in_first_chunk(addr)) {
838                 /* is it in the reserved area? */
839                 if (pcpu_addr_in_reserved_chunk(addr))
840                         return pcpu_reserved_chunk;
841                 return pcpu_first_chunk;
842         }
843 
844         /*
845          * The address is relative to unit0 which might be unused and
846          * thus unmapped.  Offset the address to the unit space of the
847          * current processor before looking it up in the vmalloc
848          * space.  Note that any possible cpu id can be used here, so
849          * there's no need to worry about preemption or cpu hotplug.
850          */
851         addr += pcpu_unit_offsets[raw_smp_processor_id()];
852         return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
853 }
854 
855 /**
856  * pcpu_alloc - the percpu allocator
857  * @size: size of area to allocate in bytes
858  * @align: alignment of area (max PAGE_SIZE)
859  * @reserved: allocate from the reserved chunk if available
860  * @gfp: allocation flags
861  *
862  * Allocate percpu area of @size bytes aligned at @align.  If @gfp doesn't
863  * contain %GFP_KERNEL, the allocation is atomic.
864  *
865  * RETURNS:
866  * Percpu pointer to the allocated area on success, NULL on failure.
867  */
868 static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
869                                  gfp_t gfp)
870 {
871         static int warn_limit = 10;
872         struct pcpu_chunk *chunk;
873         const char *err;
874         bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
875         int occ_pages = 0;
876         int slot, off, new_alloc, cpu, ret;
877         unsigned long flags;
878         void __percpu *ptr;
879 
880         /*
881          * We want the lowest bit of offset available for in-use/free
882          * indicator, so force >= 16bit alignment and make size even.
883          */
884         if (unlikely(align < 2))
885                 align = 2;
886 
887         size = ALIGN(size, 2);
888 
889         if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
890                      !is_power_of_2(align))) {
891                 WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
892                      size, align);
893                 return NULL;
894         }
895 
896         if (!is_atomic)
897                 mutex_lock(&pcpu_alloc_mutex);
898 
899         spin_lock_irqsave(&pcpu_lock, flags);
900 
901         /* serve reserved allocations from the reserved chunk if available */
902         if (reserved && pcpu_reserved_chunk) {
903                 chunk = pcpu_reserved_chunk;
904 
905                 if (size > chunk->contig_hint) {
906                         err = "alloc from reserved chunk failed";
907                         goto fail_unlock;
908                 }
909 
910                 while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
911                         spin_unlock_irqrestore(&pcpu_lock, flags);
912                         if (is_atomic ||
913                             pcpu_extend_area_map(chunk, new_alloc) < 0) {
914                                 err = "failed to extend area map of reserved chunk";
915                                 goto fail;
916                         }
917                         spin_lock_irqsave(&pcpu_lock, flags);
918                 }
919 
920                 off = pcpu_alloc_area(chunk, size, align, is_atomic,
921                                       &occ_pages);
922                 if (off >= 0)
923                         goto area_found;
924 
925                 err = "alloc from reserved chunk failed";
926                 goto fail_unlock;
927         }
928 
929 restart:
930         /* search through normal chunks */
931         for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
932                 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
933                         if (size > chunk->contig_hint)
934                                 continue;
935 
936                         new_alloc = pcpu_need_to_extend(chunk, is_atomic);
937                         if (new_alloc) {
938                                 if (is_atomic)
939                                         continue;
940                                 spin_unlock_irqrestore(&pcpu_lock, flags);
941                                 if (pcpu_extend_area_map(chunk,
942                                                          new_alloc) < 0) {
943                                         err = "failed to extend area map";
944                                         goto fail;
945                                 }
946                                 spin_lock_irqsave(&pcpu_lock, flags);
947                                 /*
948                                  * pcpu_lock has been dropped, need to
949                                  * restart cpu_slot list walking.
950                                  */
951                                 goto restart;
952                         }
953 
954                         off = pcpu_alloc_area(chunk, size, align, is_atomic,
955                                               &occ_pages);
956                         if (off >= 0)
957                                 goto area_found;
958                 }
959         }
960 
961         spin_unlock_irqrestore(&pcpu_lock, flags);
962 
963         /*
964          * No space left.  Create a new chunk.  We don't want multiple
965          * tasks to create chunks simultaneously.  Serialize and create iff
966          * there's still no empty chunk after grabbing the mutex.
967          */
968         if (is_atomic)
969                 goto fail;
970 
971         if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
972                 chunk = pcpu_create_chunk();
973                 if (!chunk) {
974                         err = "failed to allocate new chunk";
975                         goto fail;
976                 }
977 
978                 spin_lock_irqsave(&pcpu_lock, flags);
979                 pcpu_chunk_relocate(chunk, -1);
980         } else {
981                 spin_lock_irqsave(&pcpu_lock, flags);
982         }
983 
984         goto restart;
985 
986 area_found:
987         spin_unlock_irqrestore(&pcpu_lock, flags);
988 
989         /* populate if not all pages are already there */
990         if (!is_atomic) {
991                 int page_start, page_end, rs, re;
992 
993                 page_start = PFN_DOWN(off);
994                 page_end = PFN_UP(off + size);
995 
996                 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
997                         WARN_ON(chunk->immutable);
998 
999                         ret = pcpu_populate_chunk(chunk, rs, re);
1000 
1001                         spin_lock_irqsave(&pcpu_lock, flags);
1002                         if (ret) {
1003                                 pcpu_free_area(chunk, off, &occ_pages);
1004                                 err = "failed to populate";
1005                                 goto fail_unlock;
1006                         }
1007                         pcpu_chunk_populated(chunk, rs, re);
1008                         spin_unlock_irqrestore(&pcpu_lock, flags);
1009                 }
1010 
1011                 mutex_unlock(&pcpu_alloc_mutex);
1012         }
1013 
1014         if (chunk != pcpu_reserved_chunk)
1015                 pcpu_nr_empty_pop_pages -= occ_pages;
1016 
1017         if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1018                 pcpu_schedule_balance_work();
1019 
1020         /* clear the areas and return address relative to base address */
1021         for_each_possible_cpu(cpu)
1022                 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1023 
1024         ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1025         kmemleak_alloc_percpu(ptr, size, gfp);
1026         return ptr;
1027 
1028 fail_unlock:
1029         spin_unlock_irqrestore(&pcpu_lock, flags);
1030 fail:
1031         if (!is_atomic && warn_limit) {
1032                 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1033                         size, align, is_atomic, err);
1034                 dump_stack();
1035                 if (!--warn_limit)
1036                         pr_info("limit reached, disable warning\n");
1037         }
1038         if (is_atomic) {
1039                 /* see the flag handling in pcpu_blance_workfn() */
1040                 pcpu_atomic_alloc_failed = true;
1041                 pcpu_schedule_balance_work();
1042         } else {
1043                 mutex_unlock(&pcpu_alloc_mutex);
1044         }
1045         return NULL;
1046 }
1047 
1048 /**
1049  * __alloc_percpu_gfp - allocate dynamic percpu area
1050  * @size: size of area to allocate in bytes
1051  * @align: alignment of area (max PAGE_SIZE)
1052  * @gfp: allocation flags
1053  *
1054  * Allocate zero-filled percpu area of @size bytes aligned at @align.  If
1055  * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1056  * be called from any context but is a lot more likely to fail.
1057  *
1058  * RETURNS:
1059  * Percpu pointer to the allocated area on success, NULL on failure.
1060  */
1061 void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1062 {
1063         return pcpu_alloc(size, align, false, gfp);
1064 }
1065 EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1066 
1067 /**
1068  * __alloc_percpu - allocate dynamic percpu area
1069  * @size: size of area to allocate in bytes
1070  * @align: alignment of area (max PAGE_SIZE)
1071  *
1072  * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1073  */
1074 void __percpu *__alloc_percpu(size_t size, size_t align)
1075 {
1076         return pcpu_alloc(size, align, false, GFP_KERNEL);
1077 }
1078 EXPORT_SYMBOL_GPL(__alloc_percpu);
1079 
1080 /**
1081  * __alloc_reserved_percpu - allocate reserved percpu area
1082  * @size: size of area to allocate in bytes
1083  * @align: alignment of area (max PAGE_SIZE)
1084  *
1085  * Allocate zero-filled percpu area of @size bytes aligned at @align
1086  * from reserved percpu area if arch has set it up; otherwise,
1087  * allocation is served from the same dynamic area.  Might sleep.
1088  * Might trigger writeouts.
1089  *
1090  * CONTEXT:
1091  * Does GFP_KERNEL allocation.
1092  *
1093  * RETURNS:
1094  * Percpu pointer to the allocated area on success, NULL on failure.
1095  */
1096 void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1097 {
1098         return pcpu_alloc(size, align, true, GFP_KERNEL);
1099 }
1100 
1101 /**
1102  * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1103  * @work: unused
1104  *
1105  * Reclaim all fully free chunks except for the first one.
1106  */
1107 static void pcpu_balance_workfn(struct work_struct *work)
1108 {
1109         LIST_HEAD(to_free);
1110         struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1111         struct pcpu_chunk *chunk, *next;
1112         int slot, nr_to_pop, ret;
1113 
1114         /*
1115          * There's no reason to keep around multiple unused chunks and VM
1116          * areas can be scarce.  Destroy all free chunks except for one.
1117          */
1118         mutex_lock(&pcpu_alloc_mutex);
1119         spin_lock_irq(&pcpu_lock);
1120 
1121         list_for_each_entry_safe(chunk, next, free_head, list) {
1122                 WARN_ON(chunk->immutable);
1123 
1124                 /* spare the first one */
1125                 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1126                         continue;
1127 
1128                 list_del_init(&chunk->map_extend_list);
1129                 list_move(&chunk->list, &to_free);
1130         }
1131 
1132         spin_unlock_irq(&pcpu_lock);
1133 
1134         list_for_each_entry_safe(chunk, next, &to_free, list) {
1135                 int rs, re;
1136 
1137                 pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1138                         pcpu_depopulate_chunk(chunk, rs, re);
1139                         spin_lock_irq(&pcpu_lock);
1140                         pcpu_chunk_depopulated(chunk, rs, re);
1141                         spin_unlock_irq(&pcpu_lock);
1142                 }
1143                 pcpu_destroy_chunk(chunk);
1144         }
1145 
1146         /* service chunks which requested async area map extension */
1147         do {
1148                 int new_alloc = 0;
1149 
1150                 spin_lock_irq(&pcpu_lock);
1151 
1152                 chunk = list_first_entry_or_null(&pcpu_map_extend_chunks,
1153                                         struct pcpu_chunk, map_extend_list);
1154                 if (chunk) {
1155                         list_del_init(&chunk->map_extend_list);
1156                         new_alloc = pcpu_need_to_extend(chunk, false);
1157                 }
1158 
1159                 spin_unlock_irq(&pcpu_lock);
1160 
1161                 if (new_alloc)
1162                         pcpu_extend_area_map(chunk, new_alloc);
1163         } while (chunk);
1164 
1165         /*
1166          * Ensure there are certain number of free populated pages for
1167          * atomic allocs.  Fill up from the most packed so that atomic
1168          * allocs don't increase fragmentation.  If atomic allocation
1169          * failed previously, always populate the maximum amount.  This
1170          * should prevent atomic allocs larger than PAGE_SIZE from keeping
1171          * failing indefinitely; however, large atomic allocs are not
1172          * something we support properly and can be highly unreliable and
1173          * inefficient.
1174          */
1175 retry_pop:
1176         if (pcpu_atomic_alloc_failed) {
1177                 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1178                 /* best effort anyway, don't worry about synchronization */
1179                 pcpu_atomic_alloc_failed = false;
1180         } else {
1181                 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1182                                   pcpu_nr_empty_pop_pages,
1183                                   0, PCPU_EMPTY_POP_PAGES_HIGH);
1184         }
1185 
1186         for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1187                 int nr_unpop = 0, rs, re;
1188 
1189                 if (!nr_to_pop)
1190                         break;
1191 
1192                 spin_lock_irq(&pcpu_lock);
1193                 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1194                         nr_unpop = pcpu_unit_pages - chunk->nr_populated;
1195                         if (nr_unpop)
1196                                 break;
1197                 }
1198                 spin_unlock_irq(&pcpu_lock);
1199 
1200                 if (!nr_unpop)
1201                         continue;
1202 
1203                 /* @chunk can't go away while pcpu_alloc_mutex is held */
1204                 pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1205                         int nr = min(re - rs, nr_to_pop);
1206 
1207                         ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1208                         if (!ret) {
1209                                 nr_to_pop -= nr;
1210                                 spin_lock_irq(&pcpu_lock);
1211                                 pcpu_chunk_populated(chunk, rs, rs + nr);
1212                                 spin_unlock_irq(&pcpu_lock);
1213                         } else {
1214                                 nr_to_pop = 0;
1215                         }
1216 
1217                         if (!nr_to_pop)
1218                                 break;
1219                 }
1220         }
1221 
1222         if (nr_to_pop) {
1223                 /* ran out of chunks to populate, create a new one and retry */
1224                 chunk = pcpu_create_chunk();
1225                 if (chunk) {
1226                         spin_lock_irq(&pcpu_lock);
1227                         pcpu_chunk_relocate(chunk, -1);
1228                         spin_unlock_irq(&pcpu_lock);
1229                         goto retry_pop;
1230                 }
1231         }
1232 
1233         mutex_unlock(&pcpu_alloc_mutex);
1234 }
1235 
1236 /**
1237  * free_percpu - free percpu area
1238  * @ptr: pointer to area to free
1239  *
1240  * Free percpu area @ptr.
1241  *
1242  * CONTEXT:
1243  * Can be called from atomic context.
1244  */
1245 void free_percpu(void __percpu *ptr)
1246 {
1247         void *addr;
1248         struct pcpu_chunk *chunk;
1249         unsigned long flags;
1250         int off, occ_pages;
1251 
1252         if (!ptr)
1253                 return;
1254 
1255         kmemleak_free_percpu(ptr);
1256 
1257         addr = __pcpu_ptr_to_addr(ptr);
1258 
1259         spin_lock_irqsave(&pcpu_lock, flags);
1260 
1261         chunk = pcpu_chunk_addr_search(addr);
1262         off = addr - chunk->base_addr;
1263 
1264         pcpu_free_area(chunk, off, &occ_pages);
1265 
1266         if (chunk != pcpu_reserved_chunk)
1267                 pcpu_nr_empty_pop_pages += occ_pages;
1268 
1269         /* if there are more than one fully free chunks, wake up grim reaper */
1270         if (chunk->free_size == pcpu_unit_size) {
1271                 struct pcpu_chunk *pos;
1272 
1273                 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1274                         if (pos != chunk) {
1275                                 pcpu_schedule_balance_work();
1276                                 break;
1277                         }
1278         }
1279 
1280         spin_unlock_irqrestore(&pcpu_lock, flags);
1281 }
1282 EXPORT_SYMBOL_GPL(free_percpu);
1283 
1284 /**
1285  * is_kernel_percpu_address - test whether address is from static percpu area
1286  * @addr: address to test
1287  *
1288  * Test whether @addr belongs to in-kernel static percpu area.  Module
1289  * static percpu areas are not considered.  For those, use
1290  * is_module_percpu_address().
1291  *
1292  * RETURNS:
1293  * %true if @addr is from in-kernel static percpu area, %false otherwise.
1294  */
1295 bool is_kernel_percpu_address(unsigned long addr)
1296 {
1297 #ifdef CONFIG_SMP
1298         const size_t static_size = __per_cpu_end - __per_cpu_start;
1299         void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1300         unsigned int cpu;
1301 
1302         for_each_possible_cpu(cpu) {
1303                 void *start = per_cpu_ptr(base, cpu);
1304 
1305                 if ((void *)addr >= start && (void *)addr < start + static_size)
1306                         return true;
1307         }
1308 #endif
1309         /* on UP, can't distinguish from other static vars, always false */
1310         return false;
1311 }
1312 
1313 /**
1314  * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1315  * @addr: the address to be converted to physical address
1316  *
1317  * Given @addr which is dereferenceable address obtained via one of
1318  * percpu access macros, this function translates it into its physical
1319  * address.  The caller is responsible for ensuring @addr stays valid
1320  * until this function finishes.
1321  *
1322  * percpu allocator has special setup for the first chunk, which currently
1323  * supports either embedding in linear address space or vmalloc mapping,
1324  * and, from the second one, the backing allocator (currently either vm or
1325  * km) provides translation.
1326  *
1327  * The addr can be translated simply without checking if it falls into the
1328  * first chunk. But the current code reflects better how percpu allocator
1329  * actually works, and the verification can discover both bugs in percpu
1330  * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1331  * code.
1332  *
1333  * RETURNS:
1334  * The physical address for @addr.
1335  */
1336 phys_addr_t per_cpu_ptr_to_phys(void *addr)
1337 {
1338         void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1339         bool in_first_chunk = false;
1340         unsigned long first_low, first_high;
1341         unsigned int cpu;
1342 
1343         /*
1344          * The following test on unit_low/high isn't strictly
1345          * necessary but will speed up lookups of addresses which
1346          * aren't in the first chunk.
1347          */
1348         first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1349         first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1350                                      pcpu_unit_pages);
1351         if ((unsigned long)addr >= first_low &&
1352             (unsigned long)addr < first_high) {
1353                 for_each_possible_cpu(cpu) {
1354                         void *start = per_cpu_ptr(base, cpu);
1355 
1356                         if (addr >= start && addr < start + pcpu_unit_size) {
1357                                 in_first_chunk = true;
1358                                 break;
1359                         }
1360                 }
1361         }
1362 
1363         if (in_first_chunk) {
1364                 if (!is_vmalloc_addr(addr))
1365                         return __pa(addr);
1366                 else
1367                         return page_to_phys(vmalloc_to_page(addr)) +
1368                                offset_in_page(addr);
1369         } else
1370                 return page_to_phys(pcpu_addr_to_page(addr)) +
1371                        offset_in_page(addr);
1372 }
1373 
1374 /**
1375  * pcpu_alloc_alloc_info - allocate percpu allocation info
1376  * @nr_groups: the number of groups
1377  * @nr_units: the number of units
1378  *
1379  * Allocate ai which is large enough for @nr_groups groups containing
1380  * @nr_units units.  The returned ai's groups[0].cpu_map points to the
1381  * cpu_map array which is long enough for @nr_units and filled with
1382  * NR_CPUS.  It's the caller's responsibility to initialize cpu_map
1383  * pointer of other groups.
1384  *
1385  * RETURNS:
1386  * Pointer to the allocated pcpu_alloc_info on success, NULL on
1387  * failure.
1388  */
1389 struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1390                                                       int nr_units)
1391 {
1392         struct pcpu_alloc_info *ai;
1393         size_t base_size, ai_size;
1394         void *ptr;
1395         int unit;
1396 
1397         base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1398                           __alignof__(ai->groups[0].cpu_map[0]));
1399         ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1400 
1401         ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1402         if (!ptr)
1403                 return NULL;
1404         ai = ptr;
1405         ptr += base_size;
1406 
1407         ai->groups[0].cpu_map = ptr;
1408 
1409         for (unit = 0; unit < nr_units; unit++)
1410                 ai->groups[0].cpu_map[unit] = NR_CPUS;
1411 
1412         ai->nr_groups = nr_groups;
1413         ai->__ai_size = PFN_ALIGN(ai_size);
1414 
1415         return ai;
1416 }
1417 
1418 /**
1419  * pcpu_free_alloc_info - free percpu allocation info
1420  * @ai: pcpu_alloc_info to free
1421  *
1422  * Free @ai which was allocated by pcpu_alloc_alloc_info().
1423  */
1424 void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1425 {
1426         memblock_free_early(__pa(ai), ai->__ai_size);
1427 }
1428 
1429 /**
1430  * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1431  * @lvl: loglevel
1432  * @ai: allocation info to dump
1433  *
1434  * Print out information about @ai using loglevel @lvl.
1435  */
1436 static void pcpu_dump_alloc_info(const char *lvl,
1437                                  const struct pcpu_alloc_info *ai)
1438 {
1439         int group_width = 1, cpu_width = 1, width;
1440         char empty_str[] = "--------";
1441         int alloc = 0, alloc_end = 0;
1442         int group, v;
1443         int upa, apl;   /* units per alloc, allocs per line */
1444 
1445         v = ai->nr_groups;
1446         while (v /= 10)
1447                 group_width++;
1448 
1449         v = num_possible_cpus();
1450         while (v /= 10)
1451                 cpu_width++;
1452         empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1453 
1454         upa = ai->alloc_size / ai->unit_size;
1455         width = upa * (cpu_width + 1) + group_width + 3;
1456         apl = rounddown_pow_of_two(max(60 / width, 1));
1457 
1458         printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1459                lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1460                ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1461 
1462         for (group = 0; group < ai->nr_groups; group++) {
1463                 const struct pcpu_group_info *gi = &ai->groups[group];
1464                 int unit = 0, unit_end = 0;
1465 
1466                 BUG_ON(gi->nr_units % upa);
1467                 for (alloc_end += gi->nr_units / upa;
1468                      alloc < alloc_end; alloc++) {
1469                         if (!(alloc % apl)) {
1470                                 pr_cont("\n");
1471                                 printk("%spcpu-alloc: ", lvl);
1472                         }
1473                         pr_cont("[%0*d] ", group_width, group);
1474 
1475                         for (unit_end += upa; unit < unit_end; unit++)
1476                                 if (gi->cpu_map[unit] != NR_CPUS)
1477                                         pr_cont("%0*d ",
1478                                                 cpu_width, gi->cpu_map[unit]);
1479                                 else
1480                                         pr_cont("%s ", empty_str);
1481                 }
1482         }
1483         pr_cont("\n");
1484 }
1485 
1486 /**
1487  * pcpu_setup_first_chunk - initialize the first percpu chunk
1488  * @ai: pcpu_alloc_info describing how to percpu area is shaped
1489  * @base_addr: mapped address
1490  *
1491  * Initialize the first percpu chunk which contains the kernel static
1492  * perpcu area.  This function is to be called from arch percpu area
1493  * setup path.
1494  *
1495  * @ai contains all information necessary to initialize the first
1496  * chunk and prime the dynamic percpu allocator.
1497  *
1498  * @ai->static_size is the size of static percpu area.
1499  *
1500  * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1501  * reserve after the static area in the first chunk.  This reserves
1502  * the first chunk such that it's available only through reserved
1503  * percpu allocation.  This is primarily used to serve module percpu
1504  * static areas on architectures where the addressing model has
1505  * limited offset range for symbol relocations to guarantee module
1506  * percpu symbols fall inside the relocatable range.
1507  *
1508  * @ai->dyn_size determines the number of bytes available for dynamic
1509  * allocation in the first chunk.  The area between @ai->static_size +
1510  * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1511  *
1512  * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1513  * and equal to or larger than @ai->static_size + @ai->reserved_size +
1514  * @ai->dyn_size.
1515  *
1516  * @ai->atom_size is the allocation atom size and used as alignment
1517  * for vm areas.
1518  *
1519  * @ai->alloc_size is the allocation size and always multiple of
1520  * @ai->atom_size.  This is larger than @ai->atom_size if
1521  * @ai->unit_size is larger than @ai->atom_size.
1522  *
1523  * @ai->nr_groups and @ai->groups describe virtual memory layout of
1524  * percpu areas.  Units which should be colocated are put into the
1525  * same group.  Dynamic VM areas will be allocated according to these
1526  * groupings.  If @ai->nr_groups is zero, a single group containing
1527  * all units is assumed.
1528  *
1529  * The caller should have mapped the first chunk at @base_addr and
1530  * copied static data to each unit.
1531  *
1532  * If the first chunk ends up with both reserved and dynamic areas, it
1533  * is served by two chunks - one to serve the core static and reserved
1534  * areas and the other for the dynamic area.  They share the same vm
1535  * and page map but uses different area allocation map to stay away
1536  * from each other.  The latter chunk is circulated in the chunk slots
1537  * and available for dynamic allocation like any other chunks.
1538  *
1539  * RETURNS:
1540  * 0 on success, -errno on failure.
1541  */
1542 int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1543                                   void *base_addr)
1544 {
1545         static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1546         static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1547         size_t dyn_size = ai->dyn_size;
1548         size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1549         struct pcpu_chunk *schunk, *dchunk = NULL;
1550         unsigned long *group_offsets;
1551         size_t *group_sizes;
1552         unsigned long *unit_off;
1553         unsigned int cpu;
1554         int *unit_map;
1555         int group, unit, i;
1556 
1557 #define PCPU_SETUP_BUG_ON(cond) do {                                    \
1558         if (unlikely(cond)) {                                           \
1559                 pr_emerg("failed to initialize, %s\n", #cond);          \
1560                 pr_emerg("cpu_possible_mask=%*pb\n",                    \
1561                          cpumask_pr_args(cpu_possible_mask));           \
1562                 pcpu_dump_alloc_info(KERN_EMERG, ai);                   \
1563                 BUG();                                                  \
1564         }                                                               \
1565 } while (0)
1566 
1567         /* sanity checks */
1568         PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1569 #ifdef CONFIG_SMP
1570         PCPU_SETUP_BUG_ON(!ai->static_size);
1571         PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
1572 #endif
1573         PCPU_SETUP_BUG_ON(!base_addr);
1574         PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
1575         PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1576         PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
1577         PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1578         PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1579         PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1580 
1581         /* process group information and build config tables accordingly */
1582         group_offsets = memblock_virt_alloc(ai->nr_groups *
1583                                              sizeof(group_offsets[0]), 0);
1584         group_sizes = memblock_virt_alloc(ai->nr_groups *
1585                                            sizeof(group_sizes[0]), 0);
1586         unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1587         unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1588 
1589         for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1590                 unit_map[cpu] = UINT_MAX;
1591 
1592         pcpu_low_unit_cpu = NR_CPUS;
1593         pcpu_high_unit_cpu = NR_CPUS;
1594 
1595         for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1596                 const struct pcpu_group_info *gi = &ai->groups[group];
1597 
1598                 group_offsets[group] = gi->base_offset;
1599                 group_sizes[group] = gi->nr_units * ai->unit_size;
1600 
1601                 for (i = 0; i < gi->nr_units; i++) {
1602                         cpu = gi->cpu_map[i];
1603                         if (cpu == NR_CPUS)
1604                                 continue;
1605 
1606                         PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
1607                         PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1608                         PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1609 
1610                         unit_map[cpu] = unit + i;
1611                         unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1612 
1613                         /* determine low/high unit_cpu */
1614                         if (pcpu_low_unit_cpu == NR_CPUS ||
1615                             unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1616                                 pcpu_low_unit_cpu = cpu;
1617                         if (pcpu_high_unit_cpu == NR_CPUS ||
1618                             unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1619                                 pcpu_high_unit_cpu = cpu;
1620                 }
1621         }
1622         pcpu_nr_units = unit;
1623 
1624         for_each_possible_cpu(cpu)
1625                 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1626 
1627         /* we're done parsing the input, undefine BUG macro and dump config */
1628 #undef PCPU_SETUP_BUG_ON
1629         pcpu_dump_alloc_info(KERN_DEBUG, ai);
1630 
1631         pcpu_nr_groups = ai->nr_groups;
1632         pcpu_group_offsets = group_offsets;
1633         pcpu_group_sizes = group_sizes;
1634         pcpu_unit_map = unit_map;
1635         pcpu_unit_offsets = unit_off;
1636 
1637         /* determine basic parameters */
1638         pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1639         pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1640         pcpu_atom_size = ai->atom_size;
1641         pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1642                 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1643 
1644         /*
1645          * Allocate chunk slots.  The additional last slot is for
1646          * empty chunks.
1647          */
1648         pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1649         pcpu_slot = memblock_virt_alloc(
1650                         pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1651         for (i = 0; i < pcpu_nr_slots; i++)
1652                 INIT_LIST_HEAD(&pcpu_slot[i]);
1653 
1654         /*
1655          * Initialize static chunk.  If reserved_size is zero, the
1656          * static chunk covers static area + dynamic allocation area
1657          * in the first chunk.  If reserved_size is not zero, it
1658          * covers static area + reserved area (mostly used for module
1659          * static percpu allocation).
1660          */
1661         schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1662         INIT_LIST_HEAD(&schunk->list);
1663         INIT_LIST_HEAD(&schunk->map_extend_list);
1664         schunk->base_addr = base_addr;
1665         schunk->map = smap;
1666         schunk->map_alloc = ARRAY_SIZE(smap);
1667         schunk->immutable = true;
1668         bitmap_fill(schunk->populated, pcpu_unit_pages);
1669         schunk->nr_populated = pcpu_unit_pages;
1670 
1671         if (ai->reserved_size) {
1672                 schunk->free_size = ai->reserved_size;
1673                 pcpu_reserved_chunk = schunk;
1674                 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1675         } else {
1676                 schunk->free_size = dyn_size;
1677                 dyn_size = 0;                   /* dynamic area covered */
1678         }
1679         schunk->contig_hint = schunk->free_size;
1680 
1681         schunk->map[0] = 1;
1682         schunk->map[1] = ai->static_size;
1683         schunk->map_used = 1;
1684         if (schunk->free_size)
1685                 schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size;
1686         schunk->map[schunk->map_used] |= 1;
1687 
1688         /* init dynamic chunk if necessary */
1689         if (dyn_size) {
1690                 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1691                 INIT_LIST_HEAD(&dchunk->list);
1692                 INIT_LIST_HEAD(&dchunk->map_extend_list);
1693                 dchunk->base_addr = base_addr;
1694                 dchunk->map = dmap;
1695                 dchunk->map_alloc = ARRAY_SIZE(dmap);
1696                 dchunk->immutable = true;
1697                 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1698                 dchunk->nr_populated = pcpu_unit_pages;
1699 
1700                 dchunk->contig_hint = dchunk->free_size = dyn_size;
1701                 dchunk->map[0] = 1;
1702                 dchunk->map[1] = pcpu_reserved_chunk_limit;
1703                 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1704                 dchunk->map_used = 2;
1705         }
1706 
1707         /* link the first chunk in */
1708         pcpu_first_chunk = dchunk ?: schunk;
1709         pcpu_nr_empty_pop_pages +=
1710                 pcpu_count_occupied_pages(pcpu_first_chunk, 1);
1711         pcpu_chunk_relocate(pcpu_first_chunk, -1);
1712 
1713         /* we're done */
1714         pcpu_base_addr = base_addr;
1715         return 0;
1716 }
1717 
1718 #ifdef CONFIG_SMP
1719 
1720 const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1721         [PCPU_FC_AUTO]  = "auto",
1722         [PCPU_FC_EMBED] = "embed",
1723         [PCPU_FC_PAGE]  = "page",
1724 };
1725 
1726 enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1727 
1728 static int __init percpu_alloc_setup(char *str)
1729 {
1730         if (!str)
1731                 return -EINVAL;
1732 
1733         if (0)
1734                 /* nada */;
1735 #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1736         else if (!strcmp(str, "embed"))
1737                 pcpu_chosen_fc = PCPU_FC_EMBED;
1738 #endif
1739 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1740         else if (!strcmp(str, "page"))
1741                 pcpu_chosen_fc = PCPU_FC_PAGE;
1742 #endif
1743         else
1744                 pr_warn("unknown allocator %s specified\n", str);
1745 
1746         return 0;
1747 }
1748 early_param("percpu_alloc", percpu_alloc_setup);
1749 
1750 /*
1751  * pcpu_embed_first_chunk() is used by the generic percpu setup.
1752  * Build it if needed by the arch config or the generic setup is going
1753  * to be used.
1754  */
1755 #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1756         !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1757 #define BUILD_EMBED_FIRST_CHUNK
1758 #endif
1759 
1760 /* build pcpu_page_first_chunk() iff needed by the arch config */
1761 #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1762 #define BUILD_PAGE_FIRST_CHUNK
1763 #endif
1764 
1765 /* pcpu_build_alloc_info() is used by both embed and page first chunk */
1766 #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1767 /**
1768  * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1769  * @reserved_size: the size of reserved percpu area in bytes
1770  * @dyn_size: minimum free size for dynamic allocation in bytes
1771  * @atom_size: allocation atom size
1772  * @cpu_distance_fn: callback to determine distance between cpus, optional
1773  *
1774  * This function determines grouping of units, their mappings to cpus
1775  * and other parameters considering needed percpu size, allocation
1776  * atom size and distances between CPUs.
1777  *
1778  * Groups are always multiples of atom size and CPUs which are of
1779  * LOCAL_DISTANCE both ways are grouped together and share space for
1780  * units in the same group.  The returned configuration is guaranteed
1781  * to have CPUs on different nodes on different groups and >=75% usage
1782  * of allocated virtual address space.
1783  *
1784  * RETURNS:
1785  * On success, pointer to the new allocation_info is returned.  On
1786  * failure, ERR_PTR value is returned.
1787  */
1788 static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1789                                 size_t reserved_size, size_t dyn_size,
1790                                 size_t atom_size,
1791                                 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1792 {
1793         static int group_map[NR_CPUS] __initdata;
1794         static int group_cnt[NR_CPUS] __initdata;
1795         const size_t static_size = __per_cpu_end - __per_cpu_start;
1796         int nr_groups = 1, nr_units = 0;
1797         size_t size_sum, min_unit_size, alloc_size;
1798         int upa, max_upa, uninitialized_var(best_upa);  /* units_per_alloc */
1799         int last_allocs, group, unit;
1800         unsigned int cpu, tcpu;
1801         struct pcpu_alloc_info *ai;
1802         unsigned int *cpu_map;
1803 
1804         /* this function may be called multiple times */
1805         memset(group_map, 0, sizeof(group_map));
1806         memset(group_cnt, 0, sizeof(group_cnt));
1807 
1808         /* calculate size_sum and ensure dyn_size is enough for early alloc */
1809         size_sum = PFN_ALIGN(static_size + reserved_size +
1810                             max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1811         dyn_size = size_sum - static_size - reserved_size;
1812 
1813         /*
1814          * Determine min_unit_size, alloc_size and max_upa such that
1815          * alloc_size is multiple of atom_size and is the smallest
1816          * which can accommodate 4k aligned segments which are equal to
1817          * or larger than min_unit_size.
1818          */
1819         min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1820 
1821         alloc_size = roundup(min_unit_size, atom_size);
1822         upa = alloc_size / min_unit_size;
1823         while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1824                 upa--;
1825         max_upa = upa;
1826 
1827         /* group cpus according to their proximity */
1828         for_each_possible_cpu(cpu) {
1829                 group = 0;
1830         next_group:
1831                 for_each_possible_cpu(tcpu) {
1832                         if (cpu == tcpu)
1833                                 break;
1834                         if (group_map[tcpu] == group && cpu_distance_fn &&
1835                             (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1836                              cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1837                                 group++;
1838                                 nr_groups = max(nr_groups, group + 1);
1839                                 goto next_group;
1840                         }
1841                 }
1842                 group_map[cpu] = group;
1843                 group_cnt[group]++;
1844         }
1845 
1846         /*
1847          * Expand unit size until address space usage goes over 75%
1848          * and then as much as possible without using more address
1849          * space.
1850          */
1851         last_allocs = INT_MAX;
1852         for (upa = max_upa; upa; upa--) {
1853                 int allocs = 0, wasted = 0;
1854 
1855                 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1856                         continue;
1857 
1858                 for (group = 0; group < nr_groups; group++) {
1859                         int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1860                         allocs += this_allocs;
1861                         wasted += this_allocs * upa - group_cnt[group];
1862                 }
1863 
1864                 /*
1865                  * Don't accept if wastage is over 1/3.  The
1866                  * greater-than comparison ensures upa==1 always
1867                  * passes the following check.
1868                  */
1869                 if (wasted > num_possible_cpus() / 3)
1870                         continue;
1871 
1872                 /* and then don't consume more memory */
1873                 if (allocs > last_allocs)
1874                         break;
1875                 last_allocs = allocs;
1876                 best_upa = upa;
1877         }
1878         upa = best_upa;
1879 
1880         /* allocate and fill alloc_info */
1881         for (group = 0; group < nr_groups; group++)
1882                 nr_units += roundup(group_cnt[group], upa);
1883 
1884         ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1885         if (!ai)
1886                 return ERR_PTR(-ENOMEM);
1887         cpu_map = ai->groups[0].cpu_map;
1888 
1889         for (group = 0; group < nr_groups; group++) {
1890                 ai->groups[group].cpu_map = cpu_map;
1891                 cpu_map += roundup(group_cnt[group], upa);
1892         }
1893 
1894         ai->static_size = static_size;
1895         ai->reserved_size = reserved_size;
1896         ai->dyn_size = dyn_size;
1897         ai->unit_size = alloc_size / upa;
1898         ai->atom_size = atom_size;
1899         ai->alloc_size = alloc_size;
1900 
1901         for (group = 0, unit = 0; group_cnt[group]; group++) {
1902                 struct pcpu_group_info *gi = &ai->groups[group];
1903 
1904                 /*
1905                  * Initialize base_offset as if all groups are located
1906                  * back-to-back.  The caller should update this to
1907                  * reflect actual allocation.
1908                  */
1909                 gi->base_offset = unit * ai->unit_size;
1910 
1911                 for_each_possible_cpu(cpu)
1912                         if (group_map[cpu] == group)
1913                                 gi->cpu_map[gi->nr_units++] = cpu;
1914                 gi->nr_units = roundup(gi->nr_units, upa);
1915                 unit += gi->nr_units;
1916         }
1917         BUG_ON(unit != nr_units);
1918 
1919         return ai;
1920 }
1921 #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1922 
1923 #if defined(BUILD_EMBED_FIRST_CHUNK)
1924 /**
1925  * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1926  * @reserved_size: the size of reserved percpu area in bytes
1927  * @dyn_size: minimum free size for dynamic allocation in bytes
1928  * @atom_size: allocation atom size
1929  * @cpu_distance_fn: callback to determine distance between cpus, optional
1930  * @alloc_fn: function to allocate percpu page
1931  * @free_fn: function to free percpu page
1932  *
1933  * This is a helper to ease setting up embedded first percpu chunk and
1934  * can be called where pcpu_setup_first_chunk() is expected.
1935  *
1936  * If this function is used to setup the first chunk, it is allocated
1937  * by calling @alloc_fn and used as-is without being mapped into
1938  * vmalloc area.  Allocations are always whole multiples of @atom_size
1939  * aligned to @atom_size.
1940  *
1941  * This enables the first chunk to piggy back on the linear physical
1942  * mapping which often uses larger page size.  Please note that this
1943  * can result in very sparse cpu->unit mapping on NUMA machines thus
1944  * requiring large vmalloc address space.  Don't use this allocator if
1945  * vmalloc space is not orders of magnitude larger than distances
1946  * between node memory addresses (ie. 32bit NUMA machines).
1947  *
1948  * @dyn_size specifies the minimum dynamic area size.
1949  *
1950  * If the needed size is smaller than the minimum or specified unit
1951  * size, the leftover is returned using @free_fn.
1952  *
1953  * RETURNS:
1954  * 0 on success, -errno on failure.
1955  */
1956 int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1957                                   size_t atom_size,
1958                                   pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1959                                   pcpu_fc_alloc_fn_t alloc_fn,
1960                                   pcpu_fc_free_fn_t free_fn)
1961 {
1962         void *base = (void *)ULONG_MAX;
1963         void **areas = NULL;
1964         struct pcpu_alloc_info *ai;
1965         size_t size_sum, areas_size;
1966         unsigned long max_distance;
1967         int group, i, highest_group, rc;
1968 
1969         ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1970                                    cpu_distance_fn);
1971         if (IS_ERR(ai))
1972                 return PTR_ERR(ai);
1973 
1974         size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1975         areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1976 
1977         areas = memblock_virt_alloc_nopanic(areas_size, 0);
1978         if (!areas) {
1979                 rc = -ENOMEM;
1980                 goto out_free;
1981         }
1982 
1983         /* allocate, copy and determine base address & max_distance */
1984         highest_group = 0;
1985         for (group = 0; group < ai->nr_groups; group++) {
1986                 struct pcpu_group_info *gi = &ai->groups[group];
1987                 unsigned int cpu = NR_CPUS;
1988                 void *ptr;
1989 
1990                 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1991                         cpu = gi->cpu_map[i];
1992                 BUG_ON(cpu == NR_CPUS);
1993 
1994                 /* allocate space for the whole group */
1995                 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1996                 if (!ptr) {
1997                         rc = -ENOMEM;
1998                         goto out_free_areas;
1999                 }
2000                 /* kmemleak tracks the percpu allocations separately */
2001                 kmemleak_free(ptr);
2002                 areas[group] = ptr;
2003 
2004                 base = min(ptr, base);
2005                 if (ptr > areas[highest_group])
2006                         highest_group = group;
2007         }
2008         max_distance = areas[highest_group] - base;
2009         max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2010 
2011         /* warn if maximum distance is further than 75% of vmalloc space */
2012         if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2013                 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2014                                 max_distance, VMALLOC_TOTAL);
2015 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2016                 /* and fail if we have fallback */
2017                 rc = -EINVAL;
2018                 goto out_free_areas;
2019 #endif
2020         }
2021 
2022         /*
2023          * Copy data and free unused parts.  This should happen after all
2024          * allocations are complete; otherwise, we may end up with
2025          * overlapping groups.
2026          */
2027         for (group = 0; group < ai->nr_groups; group++) {
2028                 struct pcpu_group_info *gi = &ai->groups[group];
2029                 void *ptr = areas[group];
2030 
2031                 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2032                         if (gi->cpu_map[i] == NR_CPUS) {
2033                                 /* unused unit, free whole */
2034                                 free_fn(ptr, ai->unit_size);
2035                                 continue;
2036                         }
2037                         /* copy and return the unused part */
2038                         memcpy(ptr, __per_cpu_load, ai->static_size);
2039                         free_fn(ptr + size_sum, ai->unit_size - size_sum);
2040                 }
2041         }
2042 
2043         /* base address is now known, determine group base offsets */
2044         for (group = 0; group < ai->nr_groups; group++) {
2045                 ai->groups[group].base_offset = areas[group] - base;
2046         }
2047 
2048         pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2049                 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2050                 ai->dyn_size, ai->unit_size);
2051 
2052         rc = pcpu_setup_first_chunk(ai, base);
2053         goto out_free;
2054 
2055 out_free_areas:
2056         for (group = 0; group < ai->nr_groups; group++)
2057                 if (areas[group])
2058                         free_fn(areas[group],
2059                                 ai->groups[group].nr_units * ai->unit_size);
2060 out_free:
2061         pcpu_free_alloc_info(ai);
2062         if (areas)
2063                 memblock_free_early(__pa(areas), areas_size);
2064         return rc;
2065 }
2066 #endif /* BUILD_EMBED_FIRST_CHUNK */
2067 
2068 #ifdef BUILD_PAGE_FIRST_CHUNK
2069 /**
2070  * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2071  * @reserved_size: the size of reserved percpu area in bytes
2072  * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2073  * @free_fn: function to free percpu page, always called with PAGE_SIZE
2074  * @populate_pte_fn: function to populate pte
2075  *
2076  * This is a helper to ease setting up page-remapped first percpu
2077  * chunk and can be called where pcpu_setup_first_chunk() is expected.
2078  *
2079  * This is the basic allocator.  Static percpu area is allocated
2080  * page-by-page into vmalloc area.
2081  *
2082  * RETURNS:
2083  * 0 on success, -errno on failure.
2084  */
2085 int __init pcpu_page_first_chunk(size_t reserved_size,
2086                                  pcpu_fc_alloc_fn_t alloc_fn,
2087                                  pcpu_fc_free_fn_t free_fn,
2088                                  pcpu_fc_populate_pte_fn_t populate_pte_fn)
2089 {
2090         static struct vm_struct vm;
2091         struct pcpu_alloc_info *ai;
2092         char psize_str[16];
2093         int unit_pages;
2094         size_t pages_size;
2095         struct page **pages;
2096         int unit, i, j, rc;
2097         int upa;
2098         int nr_g0_units;
2099 
2100         snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2101 
2102         ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2103         if (IS_ERR(ai))
2104                 return PTR_ERR(ai);
2105         BUG_ON(ai->nr_groups != 1);
2106         upa = ai->alloc_size/ai->unit_size;
2107         nr_g0_units = roundup(num_possible_cpus(), upa);
2108         if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) {
2109                 pcpu_free_alloc_info(ai);
2110                 return -EINVAL;
2111         }
2112 
2113         unit_pages = ai->unit_size >> PAGE_SHIFT;
2114 
2115         /* unaligned allocations can't be freed, round up to page size */
2116         pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2117                                sizeof(pages[0]));
2118         pages = memblock_virt_alloc(pages_size, 0);
2119 
2120         /* allocate pages */
2121         j = 0;
2122         for (unit = 0; unit < num_possible_cpus(); unit++) {
2123                 unsigned int cpu = ai->groups[0].cpu_map[unit];
2124                 for (i = 0; i < unit_pages; i++) {
2125                         void *ptr;
2126 
2127                         ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2128                         if (!ptr) {
2129                                 pr_warn("failed to allocate %s page for cpu%u\n",
2130                                                 psize_str, cpu);
2131                                 goto enomem;
2132                         }
2133                         /* kmemleak tracks the percpu allocations separately */
2134                         kmemleak_free(ptr);
2135                         pages[j++] = virt_to_page(ptr);
2136                 }
2137         }
2138 
2139         /* allocate vm area, map the pages and copy static data */
2140         vm.flags = VM_ALLOC;
2141         vm.size = num_possible_cpus() * ai->unit_size;
2142         vm_area_register_early(&vm, PAGE_SIZE);
2143 
2144         for (unit = 0; unit < num_possible_cpus(); unit++) {
2145                 unsigned long unit_addr =
2146                         (unsigned long)vm.addr + unit * ai->unit_size;
2147 
2148                 for (i = 0; i < unit_pages; i++)
2149                         populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2150 
2151                 /* pte already populated, the following shouldn't fail */
2152                 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2153                                       unit_pages);
2154                 if (rc < 0)
2155                         panic("failed to map percpu area, err=%d\n", rc);
2156 
2157                 /*
2158                  * FIXME: Archs with virtual cache should flush local
2159                  * cache for the linear mapping here - something
2160                  * equivalent to flush_cache_vmap() on the local cpu.
2161                  * flush_cache_vmap() can't be used as most supporting
2162                  * data structures are not set up yet.
2163                  */
2164 
2165                 /* copy static data */
2166                 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2167         }
2168 
2169         /* we're ready, commit */
2170         pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2171                 unit_pages, psize_str, vm.addr, ai->static_size,
2172                 ai->reserved_size, ai->dyn_size);
2173 
2174         rc = pcpu_setup_first_chunk(ai, vm.addr);
2175         goto out_free_ar;
2176 
2177 enomem:
2178         while (--j >= 0)
2179                 free_fn(page_address(pages[j]), PAGE_SIZE);
2180         rc = -ENOMEM;
2181 out_free_ar:
2182         memblock_free_early(__pa(pages), pages_size);
2183         pcpu_free_alloc_info(ai);
2184         return rc;
2185 }
2186 #endif /* BUILD_PAGE_FIRST_CHUNK */
2187 
2188 #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2189 /*
2190  * Generic SMP percpu area setup.
2191  *
2192  * The embedding helper is used because its behavior closely resembles
2193  * the original non-dynamic generic percpu area setup.  This is
2194  * important because many archs have addressing restrictions and might
2195  * fail if the percpu area is located far away from the previous
2196  * location.  As an added bonus, in non-NUMA cases, embedding is
2197  * generally a good idea TLB-wise because percpu area can piggy back
2198  * on the physical linear memory mapping which uses large page
2199  * mappings on applicable archs.
2200  */
2201 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2202 EXPORT_SYMBOL(__per_cpu_offset);
2203 
2204 static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2205                                        size_t align)
2206 {
2207         return  memblock_virt_alloc_from_nopanic(
2208                         size, align, __pa(MAX_DMA_ADDRESS));
2209 }
2210 
2211 static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2212 {
2213         memblock_free_early(__pa(ptr), size);
2214 }
2215 
2216 void __init setup_per_cpu_areas(void)
2217 {
2218         unsigned long delta;
2219         unsigned int cpu;
2220         int rc;
2221 
2222         /*
2223          * Always reserve area for module percpu variables.  That's
2224          * what the legacy allocator did.
2225          */
2226         rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2227                                     PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2228                                     pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2229         if (rc < 0)
2230                 panic("Failed to initialize percpu areas.");
2231 
2232         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2233         for_each_possible_cpu(cpu)
2234                 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2235 }
2236 #endif  /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2237 
2238 #else   /* CONFIG_SMP */
2239 
2240 /*
2241  * UP percpu area setup.
2242  *
2243  * UP always uses km-based percpu allocator with identity mapping.
2244  * Static percpu variables are indistinguishable from the usual static
2245  * variables and don't require any special preparation.
2246  */
2247 void __init setup_per_cpu_areas(void)
2248 {
2249         const size_t unit_size =
2250                 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2251                                          PERCPU_DYNAMIC_RESERVE));
2252         struct pcpu_alloc_info *ai;
2253         void *fc;
2254 
2255         ai = pcpu_alloc_alloc_info(1, 1);
2256         fc = memblock_virt_alloc_from_nopanic(unit_size,
2257                                               PAGE_SIZE,
2258                                               __pa(MAX_DMA_ADDRESS));
2259         if (!ai || !fc)
2260                 panic("Failed to allocate memory for percpu areas.");
2261         /* kmemleak tracks the percpu allocations separately */
2262         kmemleak_free(fc);
2263 
2264         ai->dyn_size = unit_size;
2265         ai->unit_size = unit_size;
2266         ai->atom_size = unit_size;
2267         ai->alloc_size = unit_size;
2268         ai->groups[0].nr_units = 1;
2269         ai->groups[0].cpu_map[0] = 0;
2270 
2271         if (pcpu_setup_first_chunk(ai, fc) < 0)
2272                 panic("Failed to initialize percpu areas.");
2273 }
2274 
2275 #endif  /* CONFIG_SMP */
2276 
2277 /*
2278  * First and reserved chunks are initialized with temporary allocation
2279  * map in initdata so that they can be used before slab is online.
2280  * This function is called after slab is brought up and replaces those
2281  * with properly allocated maps.
2282  */
2283 void __init percpu_init_late(void)
2284 {
2285         struct pcpu_chunk *target_chunks[] =
2286                 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2287         struct pcpu_chunk *chunk;
2288         unsigned long flags;
2289         int i;
2290 
2291         for (i = 0; (chunk = target_chunks[i]); i++) {
2292                 int *map;
2293                 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2294 
2295                 BUILD_BUG_ON(size > PAGE_SIZE);
2296 
2297                 map = pcpu_mem_zalloc(size);
2298                 BUG_ON(!map);
2299 
2300                 spin_lock_irqsave(&pcpu_lock, flags);
2301                 memcpy(map, chunk->map, size);
2302                 chunk->map = map;
2303                 spin_unlock_irqrestore(&pcpu_lock, flags);
2304         }
2305 }
2306 
2307 /*
2308  * Percpu allocator is initialized early during boot when neither slab or
2309  * workqueue is available.  Plug async management until everything is up
2310  * and running.
2311  */
2312 static int __init percpu_enable_async(void)
2313 {
2314         pcpu_async_enabled = true;
2315         return 0;
2316 }
2317 subsys_initcall(percpu_enable_async);
2318 

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