Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
swiotlb.c
Go to the documentation of this file.
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 <[email protected]>
7  * Copyright (C) 2000 Goutham Rao <[email protected]>
8  * Copyright (C) 2000, 2003 Hewlett-Packard Co
9  * David Mosberger-Tang <[email protected]>
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 
33 #include <asm/io.h>
34 #include <asm/dma.h>
35 #include <asm/scatterlist.h>
36 
37 #include <linux/init.h>
38 #include <linux/bootmem.h>
39 #include <linux/iommu-helper.h>
40 
41 #define OFFSET(val,align) ((unsigned long) \
42  ( (val) & ( (align) - 1)))
43 
44 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
45 
46 /*
47  * Minimum IO TLB size to bother booting with. Systems with mainly
48  * 64bit capable cards will only lightly use the swiotlb. If we can't
49  * allocate a contiguous 1MB, we're probably in trouble anyway.
50  */
51 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
52 
54 
55 /*
56  * Used to do a quick range check in swiotlb_tbl_unmap_single and
57  * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
58  * API.
59  */
60 static char *io_tlb_start, *io_tlb_end;
61 
62 /*
63  * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
64  * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
65  */
66 static unsigned long io_tlb_nslabs;
67 
68 /*
69  * When the IOMMU overflows we return a fallback buffer. This sets the size.
70  */
71 static unsigned long io_tlb_overflow = 32*1024;
72 
73 static void *io_tlb_overflow_buffer;
74 
75 /*
76  * This is a free list describing the number of free entries available from
77  * each index
78  */
79 static unsigned int *io_tlb_list;
80 static unsigned int io_tlb_index;
81 
82 /*
83  * We need to save away the original address corresponding to a mapped entry
84  * for the sync operations.
85  */
86 static phys_addr_t *io_tlb_orig_addr;
87 
88 /*
89  * Protect the above data structures in the map and unmap calls
90  */
91 static DEFINE_SPINLOCK(io_tlb_lock);
92 
93 static int late_alloc;
94 
95 static int __init
96 setup_io_tlb_npages(char *str)
97 {
98  if (isdigit(*str)) {
99  io_tlb_nslabs = simple_strtoul(str, &str, 0);
100  /* avoid tail segment of size < IO_TLB_SEGSIZE */
101  io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
102  }
103  if (*str == ',')
104  ++str;
105  if (!strcmp(str, "force"))
106  swiotlb_force = 1;
107 
108  return 1;
109 }
110 __setup("swiotlb=", setup_io_tlb_npages);
111 /* make io_tlb_overflow tunable too? */
112 
113 unsigned long swiotlb_nr_tbl(void)
114 {
115  return io_tlb_nslabs;
116 }
118 /* Note that this doesn't work with highmem page */
119 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
120  volatile void *address)
121 {
122  return phys_to_dma(hwdev, virt_to_phys(address));
123 }
124 
126 {
127  unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
129 
130  pstart = virt_to_phys(io_tlb_start);
131  pend = virt_to_phys(io_tlb_end);
132 
133  printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
134  (unsigned long long)pstart, (unsigned long long)pend - 1,
135  bytes >> 20, io_tlb_start, io_tlb_end - 1);
136 }
137 
138 void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
139 {
140  unsigned long i, bytes;
141 
142  bytes = nslabs << IO_TLB_SHIFT;
143 
144  io_tlb_nslabs = nslabs;
145  io_tlb_start = tlb;
146  io_tlb_end = io_tlb_start + bytes;
147 
148  /*
149  * Allocate and initialize the free list array. This array is used
150  * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
151  * between io_tlb_start and io_tlb_end.
152  */
153  io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
154  for (i = 0; i < io_tlb_nslabs; i++)
155  io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
156  io_tlb_index = 0;
157  io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
158 
159  /*
160  * Get the overflow emergency buffer
161  */
162  io_tlb_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
163  if (!io_tlb_overflow_buffer)
164  panic("Cannot allocate SWIOTLB overflow buffer!\n");
165  if (verbose)
167 }
168 
169 /*
170  * Statically reserve bounce buffer space and initialize bounce buffer data
171  * structures for the software IO TLB used to implement the DMA API.
172  */
173 static void __init
174 swiotlb_init_with_default_size(size_t default_size, int verbose)
175 {
176  unsigned long bytes;
177 
178  if (!io_tlb_nslabs) {
179  io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
180  io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
181  }
182 
183  bytes = io_tlb_nslabs << IO_TLB_SHIFT;
184 
185  /*
186  * Get IO TLB memory from the low pages
187  */
188  io_tlb_start = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
189  if (!io_tlb_start)
190  panic("Cannot allocate SWIOTLB buffer");
191 
192  swiotlb_init_with_tbl(io_tlb_start, io_tlb_nslabs, verbose);
193 }
194 
195 void __init
196 swiotlb_init(int verbose)
197 {
198  swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
199 }
200 
201 /*
202  * Systems with larger DMA zones (those that don't support ISA) can
203  * initialize the swiotlb later using the slab allocator if needed.
204  * This should be just like above, but with some error catching.
205  */
206 int
208 {
209  unsigned long bytes, req_nslabs = io_tlb_nslabs;
210  unsigned int order;
211  int rc = 0;
212 
213  if (!io_tlb_nslabs) {
214  io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
215  io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
216  }
217 
218  /*
219  * Get IO TLB memory from the low pages
220  */
221  order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
222  io_tlb_nslabs = SLABS_PER_PAGE << order;
223  bytes = io_tlb_nslabs << IO_TLB_SHIFT;
224 
225  while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
226  io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
227  order);
228  if (io_tlb_start)
229  break;
230  order--;
231  }
232 
233  if (!io_tlb_start) {
234  io_tlb_nslabs = req_nslabs;
235  return -ENOMEM;
236  }
237  if (order != get_order(bytes)) {
238  printk(KERN_WARNING "Warning: only able to allocate %ld MB "
239  "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
240  io_tlb_nslabs = SLABS_PER_PAGE << order;
241  }
242  rc = swiotlb_late_init_with_tbl(io_tlb_start, io_tlb_nslabs);
243  if (rc)
244  free_pages((unsigned long)io_tlb_start, order);
245  return rc;
246 }
247 
248 int
249 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
250 {
251  unsigned long i, bytes;
252 
253  bytes = nslabs << IO_TLB_SHIFT;
254 
255  io_tlb_nslabs = nslabs;
256  io_tlb_start = tlb;
257  io_tlb_end = io_tlb_start + bytes;
258 
259  memset(io_tlb_start, 0, bytes);
260 
261  /*
262  * Allocate and initialize the free list array. This array is used
263  * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
264  * between io_tlb_start and io_tlb_end.
265  */
266  io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
267  get_order(io_tlb_nslabs * sizeof(int)));
268  if (!io_tlb_list)
269  goto cleanup2;
270 
271  for (i = 0; i < io_tlb_nslabs; i++)
272  io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
273  io_tlb_index = 0;
274 
275  io_tlb_orig_addr = (phys_addr_t *)
277  get_order(io_tlb_nslabs *
278  sizeof(phys_addr_t)));
279  if (!io_tlb_orig_addr)
280  goto cleanup3;
281 
282  memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
283 
284  /*
285  * Get the overflow emergency buffer
286  */
287  io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
288  get_order(io_tlb_overflow));
289  if (!io_tlb_overflow_buffer)
290  goto cleanup4;
291 
293 
294  late_alloc = 1;
295 
296  return 0;
297 
298 cleanup4:
299  free_pages((unsigned long)io_tlb_orig_addr,
300  get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
301  io_tlb_orig_addr = NULL;
302 cleanup3:
303  free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
304  sizeof(int)));
305  io_tlb_list = NULL;
306 cleanup2:
307  io_tlb_end = NULL;
308  io_tlb_start = NULL;
309  io_tlb_nslabs = 0;
310  return -ENOMEM;
311 }
312 
314 {
315  if (!io_tlb_overflow_buffer)
316  return;
317 
318  if (late_alloc) {
319  free_pages((unsigned long)io_tlb_overflow_buffer,
320  get_order(io_tlb_overflow));
321  free_pages((unsigned long)io_tlb_orig_addr,
322  get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
323  free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
324  sizeof(int)));
325  free_pages((unsigned long)io_tlb_start,
326  get_order(io_tlb_nslabs << IO_TLB_SHIFT));
327  } else {
328  free_bootmem_late(__pa(io_tlb_overflow_buffer),
329  PAGE_ALIGN(io_tlb_overflow));
330  free_bootmem_late(__pa(io_tlb_orig_addr),
331  PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
332  free_bootmem_late(__pa(io_tlb_list),
333  PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
334  free_bootmem_late(__pa(io_tlb_start),
335  PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
336  }
337  io_tlb_nslabs = 0;
338 }
339 
340 static int is_swiotlb_buffer(phys_addr_t paddr)
341 {
342  return paddr >= virt_to_phys(io_tlb_start) &&
343  paddr < virt_to_phys(io_tlb_end);
344 }
345 
346 /*
347  * Bounce: copy the swiotlb buffer back to the original dma location
348  */
350  enum dma_data_direction dir)
351 {
352  unsigned long pfn = PFN_DOWN(phys);
353 
354  if (PageHighMem(pfn_to_page(pfn))) {
355  /* The buffer does not have a mapping. Map it in and copy */
356  unsigned int offset = phys & ~PAGE_MASK;
357  char *buffer;
358  unsigned int sz = 0;
359  unsigned long flags;
360 
361  while (size) {
362  sz = min_t(size_t, PAGE_SIZE - offset, size);
363 
364  local_irq_save(flags);
365  buffer = kmap_atomic(pfn_to_page(pfn));
366  if (dir == DMA_TO_DEVICE)
367  memcpy(dma_addr, buffer + offset, sz);
368  else
369  memcpy(buffer + offset, dma_addr, sz);
370  kunmap_atomic(buffer);
371  local_irq_restore(flags);
372 
373  size -= sz;
374  pfn++;
375  dma_addr += sz;
376  offset = 0;
377  }
378  } else {
379  if (dir == DMA_TO_DEVICE)
380  memcpy(dma_addr, phys_to_virt(phys), size);
381  else
382  memcpy(phys_to_virt(phys), dma_addr, size);
383  }
384 }
386 
387 void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
388  phys_addr_t phys, size_t size,
389  enum dma_data_direction dir)
390 {
391  unsigned long flags;
392  char *dma_addr;
393  unsigned int nslots, stride, index, wrap;
394  int i;
395  unsigned long mask;
396  unsigned long offset_slots;
397  unsigned long max_slots;
398 
399  mask = dma_get_seg_boundary(hwdev);
400 
401  tbl_dma_addr &= mask;
402 
403  offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
404 
405  /*
406  * Carefully handle integer overflow which can occur when mask == ~0UL.
407  */
408  max_slots = mask + 1
409  ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
410  : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
411 
412  /*
413  * For mappings greater than a page, we limit the stride (and
414  * hence alignment) to a page size.
415  */
416  nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
417  if (size > PAGE_SIZE)
418  stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
419  else
420  stride = 1;
421 
422  BUG_ON(!nslots);
423 
424  /*
425  * Find suitable number of IO TLB entries size that will fit this
426  * request and allocate a buffer from that IO TLB pool.
427  */
428  spin_lock_irqsave(&io_tlb_lock, flags);
429  index = ALIGN(io_tlb_index, stride);
430  if (index >= io_tlb_nslabs)
431  index = 0;
432  wrap = index;
433 
434  do {
435  while (iommu_is_span_boundary(index, nslots, offset_slots,
436  max_slots)) {
437  index += stride;
438  if (index >= io_tlb_nslabs)
439  index = 0;
440  if (index == wrap)
441  goto not_found;
442  }
443 
444  /*
445  * If we find a slot that indicates we have 'nslots' number of
446  * contiguous buffers, we allocate the buffers from that slot
447  * and mark the entries as '0' indicating unavailable.
448  */
449  if (io_tlb_list[index] >= nslots) {
450  int count = 0;
451 
452  for (i = index; i < (int) (index + nslots); i++)
453  io_tlb_list[i] = 0;
454  for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
455  io_tlb_list[i] = ++count;
456  dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
457 
458  /*
459  * Update the indices to avoid searching in the next
460  * round.
461  */
462  io_tlb_index = ((index + nslots) < io_tlb_nslabs
463  ? (index + nslots) : 0);
464 
465  goto found;
466  }
467  index += stride;
468  if (index >= io_tlb_nslabs)
469  index = 0;
470  } while (index != wrap);
471 
472 not_found:
473  spin_unlock_irqrestore(&io_tlb_lock, flags);
474  return NULL;
475 found:
476  spin_unlock_irqrestore(&io_tlb_lock, flags);
477 
478  /*
479  * Save away the mapping from the original address to the DMA address.
480  * This is needed when we sync the memory. Then we sync the buffer if
481  * needed.
482  */
483  for (i = 0; i < nslots; i++)
484  io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
485  if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
486  swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
487 
488  return dma_addr;
489 }
491 
492 /*
493  * Allocates bounce buffer and returns its kernel virtual address.
494  */
495 
496 static void *
497 map_single(struct device *hwdev, phys_addr_t phys, size_t size,
498  enum dma_data_direction dir)
499 {
500  dma_addr_t start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start);
501 
502  return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
503 }
504 
505 /*
506  * dma_addr is the kernel virtual address of the bounce buffer to unmap.
507  */
508 void
509 swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
510  enum dma_data_direction dir)
511 {
512  unsigned long flags;
513  int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
514  int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
515  phys_addr_t phys = io_tlb_orig_addr[index];
516 
517  /*
518  * First, sync the memory before unmapping the entry
519  */
520  if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
521  swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
522 
523  /*
524  * Return the buffer to the free list by setting the corresponding
525  * entries to indicate the number of contiguous entries available.
526  * While returning the entries to the free list, we merge the entries
527  * with slots below and above the pool being returned.
528  */
529  spin_lock_irqsave(&io_tlb_lock, flags);
530  {
531  count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
532  io_tlb_list[index + nslots] : 0);
533  /*
534  * Step 1: return the slots to the free list, merging the
535  * slots with superceeding slots
536  */
537  for (i = index + nslots - 1; i >= index; i--)
538  io_tlb_list[i] = ++count;
539  /*
540  * Step 2: merge the returned slots with the preceding slots,
541  * if available (non zero)
542  */
543  for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
544  io_tlb_list[i] = ++count;
545  }
546  spin_unlock_irqrestore(&io_tlb_lock, flags);
547 }
549 
550 void
551 swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
552  enum dma_data_direction dir,
553  enum dma_sync_target target)
554 {
555  int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
556  phys_addr_t phys = io_tlb_orig_addr[index];
557 
558  phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
559 
560  switch (target) {
561  case SYNC_FOR_CPU:
562  if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
563  swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
564  else
565  BUG_ON(dir != DMA_TO_DEVICE);
566  break;
567  case SYNC_FOR_DEVICE:
568  if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
569  swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
570  else
571  BUG_ON(dir != DMA_FROM_DEVICE);
572  break;
573  default:
574  BUG();
575  }
576 }
578 
579 void *
580 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
582 {
584  void *ret;
585  int order = get_order(size);
586  u64 dma_mask = DMA_BIT_MASK(32);
587 
588  if (hwdev && hwdev->coherent_dma_mask)
589  dma_mask = hwdev->coherent_dma_mask;
590 
591  ret = (void *)__get_free_pages(flags, order);
592  if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
593  /*
594  * The allocated memory isn't reachable by the device.
595  */
596  free_pages((unsigned long) ret, order);
597  ret = NULL;
598  }
599  if (!ret) {
600  /*
601  * We are either out of memory or the device can't DMA to
602  * GFP_DMA memory; fall back on map_single(), which
603  * will grab memory from the lowest available address range.
604  */
605  ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
606  if (!ret)
607  return NULL;
608  }
609 
610  memset(ret, 0, size);
611  dev_addr = swiotlb_virt_to_bus(hwdev, ret);
612 
613  /* Confirm address can be DMA'd by device */
614  if (dev_addr + size - 1 > dma_mask) {
615  printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
616  (unsigned long long)dma_mask,
617  (unsigned long long)dev_addr);
618 
619  /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
620  swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
621  return NULL;
622  }
623  *dma_handle = dev_addr;
624  return ret;
625 }
627 
628 void
629 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
630  dma_addr_t dev_addr)
631 {
632  phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
633 
635  if (!is_swiotlb_buffer(paddr))
636  free_pages((unsigned long)vaddr, get_order(size));
637  else
638  /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
639  swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
640 }
642 
643 static void
644 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
645  int do_panic)
646 {
647  /*
648  * Ran out of IOMMU space for this operation. This is very bad.
649  * Unfortunately the drivers cannot handle this operation properly.
650  * unless they check for dma_mapping_error (most don't)
651  * When the mapping is small enough return a static buffer to limit
652  * the damage, or panic when the transfer is too big.
653  */
654  printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
655  "device %s\n", size, dev ? dev_name(dev) : "?");
656 
657  if (size <= io_tlb_overflow || !do_panic)
658  return;
659 
660  if (dir == DMA_BIDIRECTIONAL)
661  panic("DMA: Random memory could be DMA accessed\n");
662  if (dir == DMA_FROM_DEVICE)
663  panic("DMA: Random memory could be DMA written\n");
664  if (dir == DMA_TO_DEVICE)
665  panic("DMA: Random memory could be DMA read\n");
666 }
667 
668 /*
669  * Map a single buffer of the indicated size for DMA in streaming mode. The
670  * physical address to use is returned.
671  *
672  * Once the device is given the dma address, the device owns this memory until
673  * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
674  */
676  unsigned long offset, size_t size,
677  enum dma_data_direction dir,
678  struct dma_attrs *attrs)
679 {
681  dma_addr_t dev_addr = phys_to_dma(dev, phys);
682  void *map;
683 
684  BUG_ON(dir == DMA_NONE);
685  /*
686  * If the address happens to be in the device's DMA window,
687  * we can safely return the device addr and not worry about bounce
688  * buffering it.
689  */
690  if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
691  return dev_addr;
692 
693  /*
694  * Oh well, have to allocate and map a bounce buffer.
695  */
696  map = map_single(dev, phys, size, dir);
697  if (!map) {
698  swiotlb_full(dev, size, dir, 1);
699  map = io_tlb_overflow_buffer;
700  }
701 
702  dev_addr = swiotlb_virt_to_bus(dev, map);
703 
704  /*
705  * Ensure that the address returned is DMA'ble
706  */
707  if (!dma_capable(dev, dev_addr, size)) {
708  swiotlb_tbl_unmap_single(dev, map, size, dir);
709  dev_addr = swiotlb_virt_to_bus(dev, io_tlb_overflow_buffer);
710  }
711 
712  return dev_addr;
713 }
715 
716 /*
717  * Unmap a single streaming mode DMA translation. The dma_addr and size must
718  * match what was provided for in a previous swiotlb_map_page call. All
719  * other usages are undefined.
720  *
721  * After this call, reads by the cpu to the buffer are guaranteed to see
722  * whatever the device wrote there.
723  */
724 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
725  size_t size, enum dma_data_direction dir)
726 {
727  phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
728 
729  BUG_ON(dir == DMA_NONE);
730 
731  if (is_swiotlb_buffer(paddr)) {
732  swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
733  return;
734  }
735 
736  if (dir != DMA_FROM_DEVICE)
737  return;
738 
739  /*
740  * phys_to_virt doesn't work with hihgmem page but we could
741  * call dma_mark_clean() with hihgmem page here. However, we
742  * are fine since dma_mark_clean() is null on POWERPC. We can
743  * make dma_mark_clean() take a physical address if necessary.
744  */
745  dma_mark_clean(phys_to_virt(paddr), size);
746 }
747 
748 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
749  size_t size, enum dma_data_direction dir,
750  struct dma_attrs *attrs)
751 {
752  unmap_single(hwdev, dev_addr, size, dir);
753 }
755 
756 /*
757  * Make physical memory consistent for a single streaming mode DMA translation
758  * after a transfer.
759  *
760  * If you perform a swiotlb_map_page() but wish to interrogate the buffer
761  * using the cpu, yet do not wish to teardown the dma mapping, you must
762  * call this function before doing so. At the next point you give the dma
763  * address back to the card, you must first perform a
764  * swiotlb_dma_sync_for_device, and then the device again owns the buffer
765  */
766 static void
767 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
768  size_t size, enum dma_data_direction dir,
769  enum dma_sync_target target)
770 {
771  phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
772 
773  BUG_ON(dir == DMA_NONE);
774 
775  if (is_swiotlb_buffer(paddr)) {
776  swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
777  target);
778  return;
779  }
780 
781  if (dir != DMA_FROM_DEVICE)
782  return;
783 
784  dma_mark_clean(phys_to_virt(paddr), size);
785 }
786 
787 void
789  size_t size, enum dma_data_direction dir)
790 {
791  swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
792 }
794 
795 void
797  size_t size, enum dma_data_direction dir)
798 {
799  swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
800 }
802 
803 /*
804  * Map a set of buffers described by scatterlist in streaming mode for DMA.
805  * This is the scatter-gather version of the above swiotlb_map_page
806  * interface. Here the scatter gather list elements are each tagged with the
807  * appropriate dma address and length. They are obtained via
808  * sg_dma_{address,length}(SG).
809  *
810  * NOTE: An implementation may be able to use a smaller number of
811  * DMA address/length pairs than there are SG table elements.
812  * (for example via virtual mapping capabilities)
813  * The routine returns the number of addr/length pairs actually
814  * used, at most nents.
815  *
816  * Device ownership issues as mentioned above for swiotlb_map_page are the
817  * same here.
818  */
819 int
820 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
821  enum dma_data_direction dir, struct dma_attrs *attrs)
822 {
823  struct scatterlist *sg;
824  int i;
825 
826  BUG_ON(dir == DMA_NONE);
827 
828  for_each_sg(sgl, sg, nelems, i) {
829  phys_addr_t paddr = sg_phys(sg);
830  dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
831 
832  if (swiotlb_force ||
833  !dma_capable(hwdev, dev_addr, sg->length)) {
834  void *map = map_single(hwdev, sg_phys(sg),
835  sg->length, dir);
836  if (!map) {
837  /* Don't panic here, we expect map_sg users
838  to do proper error handling. */
839  swiotlb_full(hwdev, sg->length, dir, 0);
840  swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
841  attrs);
842  sgl[0].dma_length = 0;
843  return 0;
844  }
845  sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
846  } else
847  sg->dma_address = dev_addr;
848  sg->dma_length = sg->length;
849  }
850  return nelems;
851 }
853 
854 int
855 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
856  enum dma_data_direction dir)
857 {
858  return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
859 }
861 
862 /*
863  * Unmap a set of streaming mode DMA translations. Again, cpu read rules
864  * concerning calls here are the same as for swiotlb_unmap_page() above.
865  */
866 void
868  int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
869 {
870  struct scatterlist *sg;
871  int i;
872 
873  BUG_ON(dir == DMA_NONE);
874 
875  for_each_sg(sgl, sg, nelems, i)
876  unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
877 
878 }
880 
881 void
882 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
883  enum dma_data_direction dir)
884 {
885  return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
886 }
888 
889 /*
890  * Make physical memory consistent for a set of streaming mode DMA translations
891  * after a transfer.
892  *
893  * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
894  * and usage.
895  */
896 static void
897 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
898  int nelems, enum dma_data_direction dir,
899  enum dma_sync_target target)
900 {
901  struct scatterlist *sg;
902  int i;
903 
904  for_each_sg(sgl, sg, nelems, i)
905  swiotlb_sync_single(hwdev, sg->dma_address,
906  sg->dma_length, dir, target);
907 }
908 
909 void
911  int nelems, enum dma_data_direction dir)
912 {
913  swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
914 }
916 
917 void
919  int nelems, enum dma_data_direction dir)
920 {
921  swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
922 }
924 
925 int
927 {
928  return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
929 }
931 
932 /*
933  * Return whether the given device DMA address mask can be supported
934  * properly. For example, if your device can only drive the low 24-bits
935  * during bus mastering, then you would pass 0x00ffffff as the mask to
936  * this function.
937  */
938 int
940 {
941  return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
942 }