Version:  2.0.40 2.2.26 2.4.37 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 4.0 4.1 4.2

Linux/lib/swiotlb.c

  1 /*
  2  * Dynamic DMA mapping support.
  3  *
  4  * This implementation is a fallback for platforms that do not support
  5  * I/O TLBs (aka DMA address translation hardware).
  6  * Copyright (C) 2000 Asit Mallick <Asit.K.Mallick@intel.com>
  7  * Copyright (C) 2000 Goutham Rao <goutham.rao@intel.com>
  8  * Copyright (C) 2000, 2003 Hewlett-Packard Co
  9  *      David Mosberger-Tang <davidm@hpl.hp.com>
 10  *
 11  * 03/05/07 davidm      Switch from PCI-DMA to generic device DMA API.
 12  * 00/12/13 davidm      Rename to swiotlb.c and add mark_clean() to avoid
 13  *                      unnecessary i-cache flushing.
 14  * 04/07/.. ak          Better overflow handling. Assorted fixes.
 15  * 05/09/10 linville    Add support for syncing ranges, support syncing for
 16  *                      DMA_BIDIRECTIONAL mappings, miscellaneous cleanup.
 17  * 08/12/11 beckyb      Add highmem support
 18  */
 19 
 20 #include <linux/cache.h>
 21 #include <linux/dma-mapping.h>
 22 #include <linux/mm.h>
 23 #include <linux/export.h>
 24 #include <linux/spinlock.h>
 25 #include <linux/string.h>
 26 #include <linux/swiotlb.h>
 27 #include <linux/pfn.h>
 28 #include <linux/types.h>
 29 #include <linux/ctype.h>
 30 #include <linux/highmem.h>
 31 #include <linux/gfp.h>
 32 #include <linux/scatterlist.h>
 33 
 34 #include <asm/io.h>
 35 #include <asm/dma.h>
 36 
 37 #include <linux/init.h>
 38 #include <linux/bootmem.h>
 39 #include <linux/iommu-helper.h>
 40 
 41 #define CREATE_TRACE_POINTS
 42 #include <trace/events/swiotlb.h>
 43 
 44 #define OFFSET(val,align) ((unsigned long)      \
 45                            ( (val) & ( (align) - 1)))
 46 
 47 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
 48 
 49 /*
 50  * Minimum IO TLB size to bother booting with.  Systems with mainly
 51  * 64bit capable cards will only lightly use the swiotlb.  If we can't
 52  * allocate a contiguous 1MB, we're probably in trouble anyway.
 53  */
 54 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
 55 
 56 int swiotlb_force;
 57 
 58 /*
 59  * Used to do a quick range check in swiotlb_tbl_unmap_single and
 60  * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
 61  * API.
 62  */
 63 static phys_addr_t io_tlb_start, io_tlb_end;
 64 
 65 /*
 66  * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
 67  * io_tlb_end.  This is command line adjustable via setup_io_tlb_npages.
 68  */
 69 static unsigned long io_tlb_nslabs;
 70 
 71 /*
 72  * When the IOMMU overflows we return a fallback buffer. This sets the size.
 73  */
 74 static unsigned long io_tlb_overflow = 32*1024;
 75 
 76 static phys_addr_t io_tlb_overflow_buffer;
 77 
 78 /*
 79  * This is a free list describing the number of free entries available from
 80  * each index
 81  */
 82 static unsigned int *io_tlb_list;
 83 static unsigned int io_tlb_index;
 84 
 85 /*
 86  * We need to save away the original address corresponding to a mapped entry
 87  * for the sync operations.
 88  */
 89 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
 90 static phys_addr_t *io_tlb_orig_addr;
 91 
 92 /*
 93  * Protect the above data structures in the map and unmap calls
 94  */
 95 static DEFINE_SPINLOCK(io_tlb_lock);
 96 
 97 static int late_alloc;
 98 
 99 static int __init
100 setup_io_tlb_npages(char *str)
101 {
102         if (isdigit(*str)) {
103                 io_tlb_nslabs = simple_strtoul(str, &str, 0);
104                 /* avoid tail segment of size < IO_TLB_SEGSIZE */
105                 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
106         }
107         if (*str == ',')
108                 ++str;
109         if (!strcmp(str, "force"))
110                 swiotlb_force = 1;
111 
112         return 0;
113 }
114 early_param("swiotlb", setup_io_tlb_npages);
115 /* make io_tlb_overflow tunable too? */
116 
117 unsigned long swiotlb_nr_tbl(void)
118 {
119         return io_tlb_nslabs;
120 }
121 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
122 
123 /* default to 64MB */
124 #define IO_TLB_DEFAULT_SIZE (64UL<<20)
125 unsigned long swiotlb_size_or_default(void)
126 {
127         unsigned long size;
128 
129         size = io_tlb_nslabs << IO_TLB_SHIFT;
130 
131         return size ? size : (IO_TLB_DEFAULT_SIZE);
132 }
133 
134 /* Note that this doesn't work with highmem page */
135 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
136                                       volatile void *address)
137 {
138         return phys_to_dma(hwdev, virt_to_phys(address));
139 }
140 
141 static bool no_iotlb_memory;
142 
143 void swiotlb_print_info(void)
144 {
145         unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
146         unsigned char *vstart, *vend;
147 
148         if (no_iotlb_memory) {
149                 pr_warn("software IO TLB: No low mem\n");
150                 return;
151         }
152 
153         vstart = phys_to_virt(io_tlb_start);
154         vend = phys_to_virt(io_tlb_end);
155 
156         printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
157                (unsigned long long)io_tlb_start,
158                (unsigned long long)io_tlb_end,
159                bytes >> 20, vstart, vend - 1);
160 }
161 
162 int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
163 {
164         void *v_overflow_buffer;
165         unsigned long i, bytes;
166 
167         bytes = nslabs << IO_TLB_SHIFT;
168 
169         io_tlb_nslabs = nslabs;
170         io_tlb_start = __pa(tlb);
171         io_tlb_end = io_tlb_start + bytes;
172 
173         /*
174          * Get the overflow emergency buffer
175          */
176         v_overflow_buffer = memblock_virt_alloc_low_nopanic(
177                                                 PAGE_ALIGN(io_tlb_overflow),
178                                                 PAGE_SIZE);
179         if (!v_overflow_buffer)
180                 return -ENOMEM;
181 
182         io_tlb_overflow_buffer = __pa(v_overflow_buffer);
183 
184         /*
185          * Allocate and initialize the free list array.  This array is used
186          * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
187          * between io_tlb_start and io_tlb_end.
188          */
189         io_tlb_list = memblock_virt_alloc(
190                                 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
191                                 PAGE_SIZE);
192         io_tlb_orig_addr = memblock_virt_alloc(
193                                 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
194                                 PAGE_SIZE);
195         for (i = 0; i < io_tlb_nslabs; i++) {
196                 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
197                 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
198         }
199         io_tlb_index = 0;
200 
201         if (verbose)
202                 swiotlb_print_info();
203 
204         return 0;
205 }
206 
207 /*
208  * Statically reserve bounce buffer space and initialize bounce buffer data
209  * structures for the software IO TLB used to implement the DMA API.
210  */
211 void  __init
212 swiotlb_init(int verbose)
213 {
214         size_t default_size = IO_TLB_DEFAULT_SIZE;
215         unsigned char *vstart;
216         unsigned long bytes;
217 
218         if (!io_tlb_nslabs) {
219                 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
220                 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
221         }
222 
223         bytes = io_tlb_nslabs << IO_TLB_SHIFT;
224 
225         /* Get IO TLB memory from the low pages */
226         vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
227         if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
228                 return;
229 
230         if (io_tlb_start)
231                 memblock_free_early(io_tlb_start,
232                                     PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
233         pr_warn("Cannot allocate SWIOTLB buffer");
234         no_iotlb_memory = true;
235 }
236 
237 /*
238  * Systems with larger DMA zones (those that don't support ISA) can
239  * initialize the swiotlb later using the slab allocator if needed.
240  * This should be just like above, but with some error catching.
241  */
242 int
243 swiotlb_late_init_with_default_size(size_t default_size)
244 {
245         unsigned long bytes, req_nslabs = io_tlb_nslabs;
246         unsigned char *vstart = NULL;
247         unsigned int order;
248         int rc = 0;
249 
250         if (!io_tlb_nslabs) {
251                 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
252                 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
253         }
254 
255         /*
256          * Get IO TLB memory from the low pages
257          */
258         order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
259         io_tlb_nslabs = SLABS_PER_PAGE << order;
260         bytes = io_tlb_nslabs << IO_TLB_SHIFT;
261 
262         while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
263                 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
264                                                   order);
265                 if (vstart)
266                         break;
267                 order--;
268         }
269 
270         if (!vstart) {
271                 io_tlb_nslabs = req_nslabs;
272                 return -ENOMEM;
273         }
274         if (order != get_order(bytes)) {
275                 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
276                        "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
277                 io_tlb_nslabs = SLABS_PER_PAGE << order;
278         }
279         rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
280         if (rc)
281                 free_pages((unsigned long)vstart, order);
282         return rc;
283 }
284 
285 int
286 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
287 {
288         unsigned long i, bytes;
289         unsigned char *v_overflow_buffer;
290 
291         bytes = nslabs << IO_TLB_SHIFT;
292 
293         io_tlb_nslabs = nslabs;
294         io_tlb_start = virt_to_phys(tlb);
295         io_tlb_end = io_tlb_start + bytes;
296 
297         memset(tlb, 0, bytes);
298 
299         /*
300          * Get the overflow emergency buffer
301          */
302         v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
303                                                      get_order(io_tlb_overflow));
304         if (!v_overflow_buffer)
305                 goto cleanup2;
306 
307         io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
308 
309         /*
310          * Allocate and initialize the free list array.  This array is used
311          * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
312          * between io_tlb_start and io_tlb_end.
313          */
314         io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
315                                       get_order(io_tlb_nslabs * sizeof(int)));
316         if (!io_tlb_list)
317                 goto cleanup3;
318 
319         io_tlb_orig_addr = (phys_addr_t *)
320                 __get_free_pages(GFP_KERNEL,
321                                  get_order(io_tlb_nslabs *
322                                            sizeof(phys_addr_t)));
323         if (!io_tlb_orig_addr)
324                 goto cleanup4;
325 
326         for (i = 0; i < io_tlb_nslabs; i++) {
327                 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
328                 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
329         }
330         io_tlb_index = 0;
331 
332         swiotlb_print_info();
333 
334         late_alloc = 1;
335 
336         return 0;
337 
338 cleanup4:
339         free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
340                                                          sizeof(int)));
341         io_tlb_list = NULL;
342 cleanup3:
343         free_pages((unsigned long)v_overflow_buffer,
344                    get_order(io_tlb_overflow));
345         io_tlb_overflow_buffer = 0;
346 cleanup2:
347         io_tlb_end = 0;
348         io_tlb_start = 0;
349         io_tlb_nslabs = 0;
350         return -ENOMEM;
351 }
352 
353 void __init swiotlb_free(void)
354 {
355         if (!io_tlb_orig_addr)
356                 return;
357 
358         if (late_alloc) {
359                 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
360                            get_order(io_tlb_overflow));
361                 free_pages((unsigned long)io_tlb_orig_addr,
362                            get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
363                 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
364                                                                  sizeof(int)));
365                 free_pages((unsigned long)phys_to_virt(io_tlb_start),
366                            get_order(io_tlb_nslabs << IO_TLB_SHIFT));
367         } else {
368                 memblock_free_late(io_tlb_overflow_buffer,
369                                    PAGE_ALIGN(io_tlb_overflow));
370                 memblock_free_late(__pa(io_tlb_orig_addr),
371                                    PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
372                 memblock_free_late(__pa(io_tlb_list),
373                                    PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
374                 memblock_free_late(io_tlb_start,
375                                    PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
376         }
377         io_tlb_nslabs = 0;
378 }
379 
380 int is_swiotlb_buffer(phys_addr_t paddr)
381 {
382         return paddr >= io_tlb_start && paddr < io_tlb_end;
383 }
384 
385 /*
386  * Bounce: copy the swiotlb buffer back to the original dma location
387  */
388 static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
389                            size_t size, enum dma_data_direction dir)
390 {
391         unsigned long pfn = PFN_DOWN(orig_addr);
392         unsigned char *vaddr = phys_to_virt(tlb_addr);
393 
394         if (PageHighMem(pfn_to_page(pfn))) {
395                 /* The buffer does not have a mapping.  Map it in and copy */
396                 unsigned int offset = orig_addr & ~PAGE_MASK;
397                 char *buffer;
398                 unsigned int sz = 0;
399                 unsigned long flags;
400 
401                 while (size) {
402                         sz = min_t(size_t, PAGE_SIZE - offset, size);
403 
404                         local_irq_save(flags);
405                         buffer = kmap_atomic(pfn_to_page(pfn));
406                         if (dir == DMA_TO_DEVICE)
407                                 memcpy(vaddr, buffer + offset, sz);
408                         else
409                                 memcpy(buffer + offset, vaddr, sz);
410                         kunmap_atomic(buffer);
411                         local_irq_restore(flags);
412 
413                         size -= sz;
414                         pfn++;
415                         vaddr += sz;
416                         offset = 0;
417                 }
418         } else if (dir == DMA_TO_DEVICE) {
419                 memcpy(vaddr, phys_to_virt(orig_addr), size);
420         } else {
421                 memcpy(phys_to_virt(orig_addr), vaddr, size);
422         }
423 }
424 
425 phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
426                                    dma_addr_t tbl_dma_addr,
427                                    phys_addr_t orig_addr, size_t size,
428                                    enum dma_data_direction dir)
429 {
430         unsigned long flags;
431         phys_addr_t tlb_addr;
432         unsigned int nslots, stride, index, wrap;
433         int i;
434         unsigned long mask;
435         unsigned long offset_slots;
436         unsigned long max_slots;
437 
438         if (no_iotlb_memory)
439                 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
440 
441         mask = dma_get_seg_boundary(hwdev);
442 
443         tbl_dma_addr &= mask;
444 
445         offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
446 
447         /*
448          * Carefully handle integer overflow which can occur when mask == ~0UL.
449          */
450         max_slots = mask + 1
451                     ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
452                     : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
453 
454         /*
455          * For mappings greater than a page, we limit the stride (and
456          * hence alignment) to a page size.
457          */
458         nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
459         if (size > PAGE_SIZE)
460                 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
461         else
462                 stride = 1;
463 
464         BUG_ON(!nslots);
465 
466         /*
467          * Find suitable number of IO TLB entries size that will fit this
468          * request and allocate a buffer from that IO TLB pool.
469          */
470         spin_lock_irqsave(&io_tlb_lock, flags);
471         index = ALIGN(io_tlb_index, stride);
472         if (index >= io_tlb_nslabs)
473                 index = 0;
474         wrap = index;
475 
476         do {
477                 while (iommu_is_span_boundary(index, nslots, offset_slots,
478                                               max_slots)) {
479                         index += stride;
480                         if (index >= io_tlb_nslabs)
481                                 index = 0;
482                         if (index == wrap)
483                                 goto not_found;
484                 }
485 
486                 /*
487                  * If we find a slot that indicates we have 'nslots' number of
488                  * contiguous buffers, we allocate the buffers from that slot
489                  * and mark the entries as '' indicating unavailable.
490                  */
491                 if (io_tlb_list[index] >= nslots) {
492                         int count = 0;
493 
494                         for (i = index; i < (int) (index + nslots); i++)
495                                 io_tlb_list[i] = 0;
496                         for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
497                                 io_tlb_list[i] = ++count;
498                         tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
499 
500                         /*
501                          * Update the indices to avoid searching in the next
502                          * round.
503                          */
504                         io_tlb_index = ((index + nslots) < io_tlb_nslabs
505                                         ? (index + nslots) : 0);
506 
507                         goto found;
508                 }
509                 index += stride;
510                 if (index >= io_tlb_nslabs)
511                         index = 0;
512         } while (index != wrap);
513 
514 not_found:
515         spin_unlock_irqrestore(&io_tlb_lock, flags);
516         if (printk_ratelimit())
517                 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
518         return SWIOTLB_MAP_ERROR;
519 found:
520         spin_unlock_irqrestore(&io_tlb_lock, flags);
521 
522         /*
523          * Save away the mapping from the original address to the DMA address.
524          * This is needed when we sync the memory.  Then we sync the buffer if
525          * needed.
526          */
527         for (i = 0; i < nslots; i++)
528                 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
529         if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
530                 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
531 
532         return tlb_addr;
533 }
534 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
535 
536 /*
537  * Allocates bounce buffer and returns its kernel virtual address.
538  */
539 
540 static phys_addr_t
541 map_single(struct device *hwdev, phys_addr_t phys, size_t size,
542            enum dma_data_direction dir)
543 {
544         dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
545 
546         return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
547 }
548 
549 /*
550  * dma_addr is the kernel virtual address of the bounce buffer to unmap.
551  */
552 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
553                               size_t size, enum dma_data_direction dir)
554 {
555         unsigned long flags;
556         int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
557         int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
558         phys_addr_t orig_addr = io_tlb_orig_addr[index];
559 
560         /*
561          * First, sync the memory before unmapping the entry
562          */
563         if (orig_addr != INVALID_PHYS_ADDR &&
564             ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
565                 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
566 
567         /*
568          * Return the buffer to the free list by setting the corresponding
569          * entries to indicate the number of contiguous entries available.
570          * While returning the entries to the free list, we merge the entries
571          * with slots below and above the pool being returned.
572          */
573         spin_lock_irqsave(&io_tlb_lock, flags);
574         {
575                 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
576                          io_tlb_list[index + nslots] : 0);
577                 /*
578                  * Step 1: return the slots to the free list, merging the
579                  * slots with superceeding slots
580                  */
581                 for (i = index + nslots - 1; i >= index; i--) {
582                         io_tlb_list[i] = ++count;
583                         io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
584                 }
585                 /*
586                  * Step 2: merge the returned slots with the preceding slots,
587                  * if available (non zero)
588                  */
589                 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
590                         io_tlb_list[i] = ++count;
591         }
592         spin_unlock_irqrestore(&io_tlb_lock, flags);
593 }
594 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
595 
596 void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
597                              size_t size, enum dma_data_direction dir,
598                              enum dma_sync_target target)
599 {
600         int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
601         phys_addr_t orig_addr = io_tlb_orig_addr[index];
602 
603         if (orig_addr == INVALID_PHYS_ADDR)
604                 return;
605         orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
606 
607         switch (target) {
608         case SYNC_FOR_CPU:
609                 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
610                         swiotlb_bounce(orig_addr, tlb_addr,
611                                        size, DMA_FROM_DEVICE);
612                 else
613                         BUG_ON(dir != DMA_TO_DEVICE);
614                 break;
615         case SYNC_FOR_DEVICE:
616                 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
617                         swiotlb_bounce(orig_addr, tlb_addr,
618                                        size, DMA_TO_DEVICE);
619                 else
620                         BUG_ON(dir != DMA_FROM_DEVICE);
621                 break;
622         default:
623                 BUG();
624         }
625 }
626 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
627 
628 void *
629 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
630                        dma_addr_t *dma_handle, gfp_t flags)
631 {
632         dma_addr_t dev_addr;
633         void *ret;
634         int order = get_order(size);
635         u64 dma_mask = DMA_BIT_MASK(32);
636 
637         if (hwdev && hwdev->coherent_dma_mask)
638                 dma_mask = hwdev->coherent_dma_mask;
639 
640         ret = (void *)__get_free_pages(flags, order);
641         if (ret) {
642                 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
643                 if (dev_addr + size - 1 > dma_mask) {
644                         /*
645                          * The allocated memory isn't reachable by the device.
646                          */
647                         free_pages((unsigned long) ret, order);
648                         ret = NULL;
649                 }
650         }
651         if (!ret) {
652                 /*
653                  * We are either out of memory or the device can't DMA to
654                  * GFP_DMA memory; fall back on map_single(), which
655                  * will grab memory from the lowest available address range.
656                  */
657                 phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
658                 if (paddr == SWIOTLB_MAP_ERROR)
659                         goto err_warn;
660 
661                 ret = phys_to_virt(paddr);
662                 dev_addr = phys_to_dma(hwdev, paddr);
663 
664                 /* Confirm address can be DMA'd by device */
665                 if (dev_addr + size - 1 > dma_mask) {
666                         printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
667                                (unsigned long long)dma_mask,
668                                (unsigned long long)dev_addr);
669 
670                         /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
671                         swiotlb_tbl_unmap_single(hwdev, paddr,
672                                                  size, DMA_TO_DEVICE);
673                         goto err_warn;
674                 }
675         }
676 
677         *dma_handle = dev_addr;
678         memset(ret, 0, size);
679 
680         return ret;
681 
682 err_warn:
683         pr_warn("swiotlb: coherent allocation failed for device %s size=%zu\n",
684                 dev_name(hwdev), size);
685         dump_stack();
686 
687         return NULL;
688 }
689 EXPORT_SYMBOL(swiotlb_alloc_coherent);
690 
691 void
692 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
693                       dma_addr_t dev_addr)
694 {
695         phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
696 
697         WARN_ON(irqs_disabled());
698         if (!is_swiotlb_buffer(paddr))
699                 free_pages((unsigned long)vaddr, get_order(size));
700         else
701                 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
702                 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
703 }
704 EXPORT_SYMBOL(swiotlb_free_coherent);
705 
706 static void
707 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
708              int do_panic)
709 {
710         /*
711          * Ran out of IOMMU space for this operation. This is very bad.
712          * Unfortunately the drivers cannot handle this operation properly.
713          * unless they check for dma_mapping_error (most don't)
714          * When the mapping is small enough return a static buffer to limit
715          * the damage, or panic when the transfer is too big.
716          */
717         printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
718                "device %s\n", size, dev ? dev_name(dev) : "?");
719 
720         if (size <= io_tlb_overflow || !do_panic)
721                 return;
722 
723         if (dir == DMA_BIDIRECTIONAL)
724                 panic("DMA: Random memory could be DMA accessed\n");
725         if (dir == DMA_FROM_DEVICE)
726                 panic("DMA: Random memory could be DMA written\n");
727         if (dir == DMA_TO_DEVICE)
728                 panic("DMA: Random memory could be DMA read\n");
729 }
730 
731 /*
732  * Map a single buffer of the indicated size for DMA in streaming mode.  The
733  * physical address to use is returned.
734  *
735  * Once the device is given the dma address, the device owns this memory until
736  * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
737  */
738 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
739                             unsigned long offset, size_t size,
740                             enum dma_data_direction dir,
741                             struct dma_attrs *attrs)
742 {
743         phys_addr_t map, phys = page_to_phys(page) + offset;
744         dma_addr_t dev_addr = phys_to_dma(dev, phys);
745 
746         BUG_ON(dir == DMA_NONE);
747         /*
748          * If the address happens to be in the device's DMA window,
749          * we can safely return the device addr and not worry about bounce
750          * buffering it.
751          */
752         if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
753                 return dev_addr;
754 
755         trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
756 
757         /* Oh well, have to allocate and map a bounce buffer. */
758         map = map_single(dev, phys, size, dir);
759         if (map == SWIOTLB_MAP_ERROR) {
760                 swiotlb_full(dev, size, dir, 1);
761                 return phys_to_dma(dev, io_tlb_overflow_buffer);
762         }
763 
764         dev_addr = phys_to_dma(dev, map);
765 
766         /* Ensure that the address returned is DMA'ble */
767         if (!dma_capable(dev, dev_addr, size)) {
768                 swiotlb_tbl_unmap_single(dev, map, size, dir);
769                 return phys_to_dma(dev, io_tlb_overflow_buffer);
770         }
771 
772         return dev_addr;
773 }
774 EXPORT_SYMBOL_GPL(swiotlb_map_page);
775 
776 /*
777  * Unmap a single streaming mode DMA translation.  The dma_addr and size must
778  * match what was provided for in a previous swiotlb_map_page call.  All
779  * other usages are undefined.
780  *
781  * After this call, reads by the cpu to the buffer are guaranteed to see
782  * whatever the device wrote there.
783  */
784 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
785                          size_t size, enum dma_data_direction dir)
786 {
787         phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
788 
789         BUG_ON(dir == DMA_NONE);
790 
791         if (is_swiotlb_buffer(paddr)) {
792                 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
793                 return;
794         }
795 
796         if (dir != DMA_FROM_DEVICE)
797                 return;
798 
799         /*
800          * phys_to_virt doesn't work with hihgmem page but we could
801          * call dma_mark_clean() with hihgmem page here. However, we
802          * are fine since dma_mark_clean() is null on POWERPC. We can
803          * make dma_mark_clean() take a physical address if necessary.
804          */
805         dma_mark_clean(phys_to_virt(paddr), size);
806 }
807 
808 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
809                         size_t size, enum dma_data_direction dir,
810                         struct dma_attrs *attrs)
811 {
812         unmap_single(hwdev, dev_addr, size, dir);
813 }
814 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
815 
816 /*
817  * Make physical memory consistent for a single streaming mode DMA translation
818  * after a transfer.
819  *
820  * If you perform a swiotlb_map_page() but wish to interrogate the buffer
821  * using the cpu, yet do not wish to teardown the dma mapping, you must
822  * call this function before doing so.  At the next point you give the dma
823  * address back to the card, you must first perform a
824  * swiotlb_dma_sync_for_device, and then the device again owns the buffer
825  */
826 static void
827 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
828                     size_t size, enum dma_data_direction dir,
829                     enum dma_sync_target target)
830 {
831         phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
832 
833         BUG_ON(dir == DMA_NONE);
834 
835         if (is_swiotlb_buffer(paddr)) {
836                 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
837                 return;
838         }
839 
840         if (dir != DMA_FROM_DEVICE)
841                 return;
842 
843         dma_mark_clean(phys_to_virt(paddr), size);
844 }
845 
846 void
847 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
848                             size_t size, enum dma_data_direction dir)
849 {
850         swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
851 }
852 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
853 
854 void
855 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
856                                size_t size, enum dma_data_direction dir)
857 {
858         swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
859 }
860 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
861 
862 /*
863  * Map a set of buffers described by scatterlist in streaming mode for DMA.
864  * This is the scatter-gather version of the above swiotlb_map_page
865  * interface.  Here the scatter gather list elements are each tagged with the
866  * appropriate dma address and length.  They are obtained via
867  * sg_dma_{address,length}(SG).
868  *
869  * NOTE: An implementation may be able to use a smaller number of
870  *       DMA address/length pairs than there are SG table elements.
871  *       (for example via virtual mapping capabilities)
872  *       The routine returns the number of addr/length pairs actually
873  *       used, at most nents.
874  *
875  * Device ownership issues as mentioned above for swiotlb_map_page are the
876  * same here.
877  */
878 int
879 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
880                      enum dma_data_direction dir, struct dma_attrs *attrs)
881 {
882         struct scatterlist *sg;
883         int i;
884 
885         BUG_ON(dir == DMA_NONE);
886 
887         for_each_sg(sgl, sg, nelems, i) {
888                 phys_addr_t paddr = sg_phys(sg);
889                 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
890 
891                 if (swiotlb_force ||
892                     !dma_capable(hwdev, dev_addr, sg->length)) {
893                         phys_addr_t map = map_single(hwdev, sg_phys(sg),
894                                                      sg->length, dir);
895                         if (map == SWIOTLB_MAP_ERROR) {
896                                 /* Don't panic here, we expect map_sg users
897                                    to do proper error handling. */
898                                 swiotlb_full(hwdev, sg->length, dir, 0);
899                                 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
900                                                        attrs);
901                                 sg_dma_len(sgl) = 0;
902                                 return 0;
903                         }
904                         sg->dma_address = phys_to_dma(hwdev, map);
905                 } else
906                         sg->dma_address = dev_addr;
907                 sg_dma_len(sg) = sg->length;
908         }
909         return nelems;
910 }
911 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
912 
913 int
914 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
915                enum dma_data_direction dir)
916 {
917         return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
918 }
919 EXPORT_SYMBOL(swiotlb_map_sg);
920 
921 /*
922  * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
923  * concerning calls here are the same as for swiotlb_unmap_page() above.
924  */
925 void
926 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
927                        int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
928 {
929         struct scatterlist *sg;
930         int i;
931 
932         BUG_ON(dir == DMA_NONE);
933 
934         for_each_sg(sgl, sg, nelems, i)
935                 unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
936 
937 }
938 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
939 
940 void
941 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
942                  enum dma_data_direction dir)
943 {
944         return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
945 }
946 EXPORT_SYMBOL(swiotlb_unmap_sg);
947 
948 /*
949  * Make physical memory consistent for a set of streaming mode DMA translations
950  * after a transfer.
951  *
952  * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
953  * and usage.
954  */
955 static void
956 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
957                 int nelems, enum dma_data_direction dir,
958                 enum dma_sync_target target)
959 {
960         struct scatterlist *sg;
961         int i;
962 
963         for_each_sg(sgl, sg, nelems, i)
964                 swiotlb_sync_single(hwdev, sg->dma_address,
965                                     sg_dma_len(sg), dir, target);
966 }
967 
968 void
969 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
970                         int nelems, enum dma_data_direction dir)
971 {
972         swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
973 }
974 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
975 
976 void
977 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
978                            int nelems, enum dma_data_direction dir)
979 {
980         swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
981 }
982 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
983 
984 int
985 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
986 {
987         return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
988 }
989 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
990 
991 /*
992  * Return whether the given device DMA address mask can be supported
993  * properly.  For example, if your device can only drive the low 24-bits
994  * during bus mastering, then you would pass 0x00ffffff as the mask to
995  * this function.
996  */
997 int
998 swiotlb_dma_supported(struct device *hwdev, u64 mask)
999 {
1000         return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
1001 }
1002 EXPORT_SYMBOL(swiotlb_dma_supported);
1003 

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