Loading...
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
56enum swiotlb_force 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 */
63static 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 */
69static unsigned long io_tlb_nslabs;
70
71/*
72 * When the IOMMU overflows we return a fallback buffer. This sets the size.
73 */
74static unsigned long io_tlb_overflow = 32*1024;
75
76static 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 */
82static unsigned int *io_tlb_list;
83static unsigned int io_tlb_index;
84
85/*
86 * Max segment that we can provide which (if pages are contingous) will
87 * not be bounced (unless SWIOTLB_FORCE is set).
88 */
89unsigned int max_segment;
90
91/*
92 * We need to save away the original address corresponding to a mapped entry
93 * for the sync operations.
94 */
95#define INVALID_PHYS_ADDR (~(phys_addr_t)0)
96static phys_addr_t *io_tlb_orig_addr;
97
98/*
99 * Protect the above data structures in the map and unmap calls
100 */
101static DEFINE_SPINLOCK(io_tlb_lock);
102
103static int late_alloc;
104
105static int __init
106setup_io_tlb_npages(char *str)
107{
108 if (isdigit(*str)) {
109 io_tlb_nslabs = simple_strtoul(str, &str, 0);
110 /* avoid tail segment of size < IO_TLB_SEGSIZE */
111 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
112 }
113 if (*str == ',')
114 ++str;
115 if (!strcmp(str, "force")) {
116 swiotlb_force = SWIOTLB_FORCE;
117 } else if (!strcmp(str, "noforce")) {
118 swiotlb_force = SWIOTLB_NO_FORCE;
119 io_tlb_nslabs = 1;
120 }
121
122 return 0;
123}
124early_param("swiotlb", setup_io_tlb_npages);
125/* make io_tlb_overflow tunable too? */
126
127unsigned long swiotlb_nr_tbl(void)
128{
129 return io_tlb_nslabs;
130}
131EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
132
133unsigned int swiotlb_max_segment(void)
134{
135 return max_segment;
136}
137EXPORT_SYMBOL_GPL(swiotlb_max_segment);
138
139void swiotlb_set_max_segment(unsigned int val)
140{
141 if (swiotlb_force == SWIOTLB_FORCE)
142 max_segment = 1;
143 else
144 max_segment = rounddown(val, PAGE_SIZE);
145}
146
147/* default to 64MB */
148#define IO_TLB_DEFAULT_SIZE (64UL<<20)
149unsigned long swiotlb_size_or_default(void)
150{
151 unsigned long size;
152
153 size = io_tlb_nslabs << IO_TLB_SHIFT;
154
155 return size ? size : (IO_TLB_DEFAULT_SIZE);
156}
157
158/* Note that this doesn't work with highmem page */
159static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
160 volatile void *address)
161{
162 return phys_to_dma(hwdev, virt_to_phys(address));
163}
164
165static bool no_iotlb_memory;
166
167void swiotlb_print_info(void)
168{
169 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
170 unsigned char *vstart, *vend;
171
172 if (no_iotlb_memory) {
173 pr_warn("software IO TLB: No low mem\n");
174 return;
175 }
176
177 vstart = phys_to_virt(io_tlb_start);
178 vend = phys_to_virt(io_tlb_end);
179
180 printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
181 (unsigned long long)io_tlb_start,
182 (unsigned long long)io_tlb_end,
183 bytes >> 20, vstart, vend - 1);
184}
185
186int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
187{
188 void *v_overflow_buffer;
189 unsigned long i, bytes;
190
191 bytes = nslabs << IO_TLB_SHIFT;
192
193 io_tlb_nslabs = nslabs;
194 io_tlb_start = __pa(tlb);
195 io_tlb_end = io_tlb_start + bytes;
196
197 /*
198 * Get the overflow emergency buffer
199 */
200 v_overflow_buffer = memblock_virt_alloc_low_nopanic(
201 PAGE_ALIGN(io_tlb_overflow),
202 PAGE_SIZE);
203 if (!v_overflow_buffer)
204 return -ENOMEM;
205
206 io_tlb_overflow_buffer = __pa(v_overflow_buffer);
207
208 /*
209 * Allocate and initialize the free list array. This array is used
210 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
211 * between io_tlb_start and io_tlb_end.
212 */
213 io_tlb_list = memblock_virt_alloc(
214 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
215 PAGE_SIZE);
216 io_tlb_orig_addr = memblock_virt_alloc(
217 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
218 PAGE_SIZE);
219 for (i = 0; i < io_tlb_nslabs; i++) {
220 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
221 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
222 }
223 io_tlb_index = 0;
224
225 if (verbose)
226 swiotlb_print_info();
227
228 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
229 return 0;
230}
231
232/*
233 * Statically reserve bounce buffer space and initialize bounce buffer data
234 * structures for the software IO TLB used to implement the DMA API.
235 */
236void __init
237swiotlb_init(int verbose)
238{
239 size_t default_size = IO_TLB_DEFAULT_SIZE;
240 unsigned char *vstart;
241 unsigned long bytes;
242
243 if (!io_tlb_nslabs) {
244 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
245 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
246 }
247
248 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
249
250 /* Get IO TLB memory from the low pages */
251 vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
252 if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
253 return;
254
255 if (io_tlb_start)
256 memblock_free_early(io_tlb_start,
257 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
258 pr_warn("Cannot allocate SWIOTLB buffer");
259 no_iotlb_memory = true;
260}
261
262/*
263 * Systems with larger DMA zones (those that don't support ISA) can
264 * initialize the swiotlb later using the slab allocator if needed.
265 * This should be just like above, but with some error catching.
266 */
267int
268swiotlb_late_init_with_default_size(size_t default_size)
269{
270 unsigned long bytes, req_nslabs = io_tlb_nslabs;
271 unsigned char *vstart = NULL;
272 unsigned int order;
273 int rc = 0;
274
275 if (!io_tlb_nslabs) {
276 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
277 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
278 }
279
280 /*
281 * Get IO TLB memory from the low pages
282 */
283 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
284 io_tlb_nslabs = SLABS_PER_PAGE << order;
285 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
286
287 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
288 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
289 order);
290 if (vstart)
291 break;
292 order--;
293 }
294
295 if (!vstart) {
296 io_tlb_nslabs = req_nslabs;
297 return -ENOMEM;
298 }
299 if (order != get_order(bytes)) {
300 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
301 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
302 io_tlb_nslabs = SLABS_PER_PAGE << order;
303 }
304 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
305 if (rc)
306 free_pages((unsigned long)vstart, order);
307
308 return rc;
309}
310
311int
312swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
313{
314 unsigned long i, bytes;
315 unsigned char *v_overflow_buffer;
316
317 bytes = nslabs << IO_TLB_SHIFT;
318
319 io_tlb_nslabs = nslabs;
320 io_tlb_start = virt_to_phys(tlb);
321 io_tlb_end = io_tlb_start + bytes;
322
323 memset(tlb, 0, bytes);
324
325 /*
326 * Get the overflow emergency buffer
327 */
328 v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
329 get_order(io_tlb_overflow));
330 if (!v_overflow_buffer)
331 goto cleanup2;
332
333 io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
334
335 /*
336 * Allocate and initialize the free list array. This array is used
337 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
338 * between io_tlb_start and io_tlb_end.
339 */
340 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
341 get_order(io_tlb_nslabs * sizeof(int)));
342 if (!io_tlb_list)
343 goto cleanup3;
344
345 io_tlb_orig_addr = (phys_addr_t *)
346 __get_free_pages(GFP_KERNEL,
347 get_order(io_tlb_nslabs *
348 sizeof(phys_addr_t)));
349 if (!io_tlb_orig_addr)
350 goto cleanup4;
351
352 for (i = 0; i < io_tlb_nslabs; i++) {
353 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
354 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
355 }
356 io_tlb_index = 0;
357
358 swiotlb_print_info();
359
360 late_alloc = 1;
361
362 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
363
364 return 0;
365
366cleanup4:
367 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
368 sizeof(int)));
369 io_tlb_list = NULL;
370cleanup3:
371 free_pages((unsigned long)v_overflow_buffer,
372 get_order(io_tlb_overflow));
373 io_tlb_overflow_buffer = 0;
374cleanup2:
375 io_tlb_end = 0;
376 io_tlb_start = 0;
377 io_tlb_nslabs = 0;
378 max_segment = 0;
379 return -ENOMEM;
380}
381
382void __init swiotlb_free(void)
383{
384 if (!io_tlb_orig_addr)
385 return;
386
387 if (late_alloc) {
388 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
389 get_order(io_tlb_overflow));
390 free_pages((unsigned long)io_tlb_orig_addr,
391 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
392 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
393 sizeof(int)));
394 free_pages((unsigned long)phys_to_virt(io_tlb_start),
395 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
396 } else {
397 memblock_free_late(io_tlb_overflow_buffer,
398 PAGE_ALIGN(io_tlb_overflow));
399 memblock_free_late(__pa(io_tlb_orig_addr),
400 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
401 memblock_free_late(__pa(io_tlb_list),
402 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
403 memblock_free_late(io_tlb_start,
404 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
405 }
406 io_tlb_nslabs = 0;
407 max_segment = 0;
408}
409
410int is_swiotlb_buffer(phys_addr_t paddr)
411{
412 return paddr >= io_tlb_start && paddr < io_tlb_end;
413}
414
415/*
416 * Bounce: copy the swiotlb buffer back to the original dma location
417 */
418static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
419 size_t size, enum dma_data_direction dir)
420{
421 unsigned long pfn = PFN_DOWN(orig_addr);
422 unsigned char *vaddr = phys_to_virt(tlb_addr);
423
424 if (PageHighMem(pfn_to_page(pfn))) {
425 /* The buffer does not have a mapping. Map it in and copy */
426 unsigned int offset = orig_addr & ~PAGE_MASK;
427 char *buffer;
428 unsigned int sz = 0;
429 unsigned long flags;
430
431 while (size) {
432 sz = min_t(size_t, PAGE_SIZE - offset, size);
433
434 local_irq_save(flags);
435 buffer = kmap_atomic(pfn_to_page(pfn));
436 if (dir == DMA_TO_DEVICE)
437 memcpy(vaddr, buffer + offset, sz);
438 else
439 memcpy(buffer + offset, vaddr, sz);
440 kunmap_atomic(buffer);
441 local_irq_restore(flags);
442
443 size -= sz;
444 pfn++;
445 vaddr += sz;
446 offset = 0;
447 }
448 } else if (dir == DMA_TO_DEVICE) {
449 memcpy(vaddr, phys_to_virt(orig_addr), size);
450 } else {
451 memcpy(phys_to_virt(orig_addr), vaddr, size);
452 }
453}
454
455phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
456 dma_addr_t tbl_dma_addr,
457 phys_addr_t orig_addr, size_t size,
458 enum dma_data_direction dir,
459 unsigned long attrs)
460{
461 unsigned long flags;
462 phys_addr_t tlb_addr;
463 unsigned int nslots, stride, index, wrap;
464 int i;
465 unsigned long mask;
466 unsigned long offset_slots;
467 unsigned long max_slots;
468
469 if (no_iotlb_memory)
470 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
471
472 mask = dma_get_seg_boundary(hwdev);
473
474 tbl_dma_addr &= mask;
475
476 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
477
478 /*
479 * Carefully handle integer overflow which can occur when mask == ~0UL.
480 */
481 max_slots = mask + 1
482 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
483 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
484
485 /*
486 * For mappings greater than or equal to a page, we limit the stride
487 * (and hence alignment) to a page size.
488 */
489 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
490 if (size >= PAGE_SIZE)
491 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
492 else
493 stride = 1;
494
495 BUG_ON(!nslots);
496
497 /*
498 * Find suitable number of IO TLB entries size that will fit this
499 * request and allocate a buffer from that IO TLB pool.
500 */
501 spin_lock_irqsave(&io_tlb_lock, flags);
502 index = ALIGN(io_tlb_index, stride);
503 if (index >= io_tlb_nslabs)
504 index = 0;
505 wrap = index;
506
507 do {
508 while (iommu_is_span_boundary(index, nslots, offset_slots,
509 max_slots)) {
510 index += stride;
511 if (index >= io_tlb_nslabs)
512 index = 0;
513 if (index == wrap)
514 goto not_found;
515 }
516
517 /*
518 * If we find a slot that indicates we have 'nslots' number of
519 * contiguous buffers, we allocate the buffers from that slot
520 * and mark the entries as '0' indicating unavailable.
521 */
522 if (io_tlb_list[index] >= nslots) {
523 int count = 0;
524
525 for (i = index; i < (int) (index + nslots); i++)
526 io_tlb_list[i] = 0;
527 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
528 io_tlb_list[i] = ++count;
529 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
530
531 /*
532 * Update the indices to avoid searching in the next
533 * round.
534 */
535 io_tlb_index = ((index + nslots) < io_tlb_nslabs
536 ? (index + nslots) : 0);
537
538 goto found;
539 }
540 index += stride;
541 if (index >= io_tlb_nslabs)
542 index = 0;
543 } while (index != wrap);
544
545not_found:
546 spin_unlock_irqrestore(&io_tlb_lock, flags);
547 if (printk_ratelimit())
548 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
549 return SWIOTLB_MAP_ERROR;
550found:
551 spin_unlock_irqrestore(&io_tlb_lock, flags);
552
553 /*
554 * Save away the mapping from the original address to the DMA address.
555 * This is needed when we sync the memory. Then we sync the buffer if
556 * needed.
557 */
558 for (i = 0; i < nslots; i++)
559 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
560 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
561 (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
562 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
563
564 return tlb_addr;
565}
566EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
567
568/*
569 * Allocates bounce buffer and returns its kernel virtual address.
570 */
571
572static phys_addr_t
573map_single(struct device *hwdev, phys_addr_t phys, size_t size,
574 enum dma_data_direction dir, unsigned long attrs)
575{
576 dma_addr_t start_dma_addr;
577
578 if (swiotlb_force == SWIOTLB_NO_FORCE) {
579 dev_warn_ratelimited(hwdev, "Cannot do DMA to address %pa\n",
580 &phys);
581 return SWIOTLB_MAP_ERROR;
582 }
583
584 start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
585 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size,
586 dir, attrs);
587}
588
589/*
590 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
591 */
592void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
593 size_t size, enum dma_data_direction dir,
594 unsigned long attrs)
595{
596 unsigned long flags;
597 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
598 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
599 phys_addr_t orig_addr = io_tlb_orig_addr[index];
600
601 /*
602 * First, sync the memory before unmapping the entry
603 */
604 if (orig_addr != INVALID_PHYS_ADDR &&
605 !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
606 ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
607 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
608
609 /*
610 * Return the buffer to the free list by setting the corresponding
611 * entries to indicate the number of contiguous entries available.
612 * While returning the entries to the free list, we merge the entries
613 * with slots below and above the pool being returned.
614 */
615 spin_lock_irqsave(&io_tlb_lock, flags);
616 {
617 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
618 io_tlb_list[index + nslots] : 0);
619 /*
620 * Step 1: return the slots to the free list, merging the
621 * slots with superceeding slots
622 */
623 for (i = index + nslots - 1; i >= index; i--) {
624 io_tlb_list[i] = ++count;
625 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
626 }
627 /*
628 * Step 2: merge the returned slots with the preceding slots,
629 * if available (non zero)
630 */
631 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
632 io_tlb_list[i] = ++count;
633 }
634 spin_unlock_irqrestore(&io_tlb_lock, flags);
635}
636EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
637
638void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
639 size_t size, enum dma_data_direction dir,
640 enum dma_sync_target target)
641{
642 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
643 phys_addr_t orig_addr = io_tlb_orig_addr[index];
644
645 if (orig_addr == INVALID_PHYS_ADDR)
646 return;
647 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
648
649 switch (target) {
650 case SYNC_FOR_CPU:
651 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
652 swiotlb_bounce(orig_addr, tlb_addr,
653 size, DMA_FROM_DEVICE);
654 else
655 BUG_ON(dir != DMA_TO_DEVICE);
656 break;
657 case SYNC_FOR_DEVICE:
658 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
659 swiotlb_bounce(orig_addr, tlb_addr,
660 size, DMA_TO_DEVICE);
661 else
662 BUG_ON(dir != DMA_FROM_DEVICE);
663 break;
664 default:
665 BUG();
666 }
667}
668EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
669
670void *
671swiotlb_alloc_coherent(struct device *hwdev, size_t size,
672 dma_addr_t *dma_handle, gfp_t flags)
673{
674 dma_addr_t dev_addr;
675 void *ret;
676 int order = get_order(size);
677 u64 dma_mask = DMA_BIT_MASK(32);
678
679 if (hwdev && hwdev->coherent_dma_mask)
680 dma_mask = hwdev->coherent_dma_mask;
681
682 ret = (void *)__get_free_pages(flags, order);
683 if (ret) {
684 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
685 if (dev_addr + size - 1 > dma_mask) {
686 /*
687 * The allocated memory isn't reachable by the device.
688 */
689 free_pages((unsigned long) ret, order);
690 ret = NULL;
691 }
692 }
693 if (!ret) {
694 /*
695 * We are either out of memory or the device can't DMA to
696 * GFP_DMA memory; fall back on map_single(), which
697 * will grab memory from the lowest available address range.
698 */
699 phys_addr_t paddr = map_single(hwdev, 0, size,
700 DMA_FROM_DEVICE, 0);
701 if (paddr == SWIOTLB_MAP_ERROR)
702 goto err_warn;
703
704 ret = phys_to_virt(paddr);
705 dev_addr = phys_to_dma(hwdev, paddr);
706
707 /* Confirm address can be DMA'd by device */
708 if (dev_addr + size - 1 > dma_mask) {
709 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
710 (unsigned long long)dma_mask,
711 (unsigned long long)dev_addr);
712
713 /*
714 * DMA_TO_DEVICE to avoid memcpy in unmap_single.
715 * The DMA_ATTR_SKIP_CPU_SYNC is optional.
716 */
717 swiotlb_tbl_unmap_single(hwdev, paddr,
718 size, DMA_TO_DEVICE,
719 DMA_ATTR_SKIP_CPU_SYNC);
720 goto err_warn;
721 }
722 }
723
724 *dma_handle = dev_addr;
725 memset(ret, 0, size);
726
727 return ret;
728
729err_warn:
730 pr_warn("swiotlb: coherent allocation failed for device %s size=%zu\n",
731 dev_name(hwdev), size);
732 dump_stack();
733
734 return NULL;
735}
736EXPORT_SYMBOL(swiotlb_alloc_coherent);
737
738void
739swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
740 dma_addr_t dev_addr)
741{
742 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
743
744 WARN_ON(irqs_disabled());
745 if (!is_swiotlb_buffer(paddr))
746 free_pages((unsigned long)vaddr, get_order(size));
747 else
748 /*
749 * DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single.
750 * DMA_ATTR_SKIP_CPU_SYNC is optional.
751 */
752 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE,
753 DMA_ATTR_SKIP_CPU_SYNC);
754}
755EXPORT_SYMBOL(swiotlb_free_coherent);
756
757static void
758swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
759 int do_panic)
760{
761 if (swiotlb_force == SWIOTLB_NO_FORCE)
762 return;
763
764 /*
765 * Ran out of IOMMU space for this operation. This is very bad.
766 * Unfortunately the drivers cannot handle this operation properly.
767 * unless they check for dma_mapping_error (most don't)
768 * When the mapping is small enough return a static buffer to limit
769 * the damage, or panic when the transfer is too big.
770 */
771 dev_err_ratelimited(dev, "DMA: Out of SW-IOMMU space for %zu bytes\n",
772 size);
773
774 if (size <= io_tlb_overflow || !do_panic)
775 return;
776
777 if (dir == DMA_BIDIRECTIONAL)
778 panic("DMA: Random memory could be DMA accessed\n");
779 if (dir == DMA_FROM_DEVICE)
780 panic("DMA: Random memory could be DMA written\n");
781 if (dir == DMA_TO_DEVICE)
782 panic("DMA: Random memory could be DMA read\n");
783}
784
785/*
786 * Map a single buffer of the indicated size for DMA in streaming mode. The
787 * physical address to use is returned.
788 *
789 * Once the device is given the dma address, the device owns this memory until
790 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
791 */
792dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
793 unsigned long offset, size_t size,
794 enum dma_data_direction dir,
795 unsigned long attrs)
796{
797 phys_addr_t map, phys = page_to_phys(page) + offset;
798 dma_addr_t dev_addr = phys_to_dma(dev, phys);
799
800 BUG_ON(dir == DMA_NONE);
801 /*
802 * If the address happens to be in the device's DMA window,
803 * we can safely return the device addr and not worry about bounce
804 * buffering it.
805 */
806 if (dma_capable(dev, dev_addr, size) && swiotlb_force != SWIOTLB_FORCE)
807 return dev_addr;
808
809 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
810
811 /* Oh well, have to allocate and map a bounce buffer. */
812 map = map_single(dev, phys, size, dir, attrs);
813 if (map == SWIOTLB_MAP_ERROR) {
814 swiotlb_full(dev, size, dir, 1);
815 return phys_to_dma(dev, io_tlb_overflow_buffer);
816 }
817
818 dev_addr = phys_to_dma(dev, map);
819
820 /* Ensure that the address returned is DMA'ble */
821 if (dma_capable(dev, dev_addr, size))
822 return dev_addr;
823
824 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
825 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
826
827 return phys_to_dma(dev, io_tlb_overflow_buffer);
828}
829EXPORT_SYMBOL_GPL(swiotlb_map_page);
830
831/*
832 * Unmap a single streaming mode DMA translation. The dma_addr and size must
833 * match what was provided for in a previous swiotlb_map_page call. All
834 * other usages are undefined.
835 *
836 * After this call, reads by the cpu to the buffer are guaranteed to see
837 * whatever the device wrote there.
838 */
839static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
840 size_t size, enum dma_data_direction dir,
841 unsigned long attrs)
842{
843 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
844
845 BUG_ON(dir == DMA_NONE);
846
847 if (is_swiotlb_buffer(paddr)) {
848 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
849 return;
850 }
851
852 if (dir != DMA_FROM_DEVICE)
853 return;
854
855 /*
856 * phys_to_virt doesn't work with hihgmem page but we could
857 * call dma_mark_clean() with hihgmem page here. However, we
858 * are fine since dma_mark_clean() is null on POWERPC. We can
859 * make dma_mark_clean() take a physical address if necessary.
860 */
861 dma_mark_clean(phys_to_virt(paddr), size);
862}
863
864void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
865 size_t size, enum dma_data_direction dir,
866 unsigned long attrs)
867{
868 unmap_single(hwdev, dev_addr, size, dir, attrs);
869}
870EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
871
872/*
873 * Make physical memory consistent for a single streaming mode DMA translation
874 * after a transfer.
875 *
876 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
877 * using the cpu, yet do not wish to teardown the dma mapping, you must
878 * call this function before doing so. At the next point you give the dma
879 * address back to the card, you must first perform a
880 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
881 */
882static void
883swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
884 size_t size, enum dma_data_direction dir,
885 enum dma_sync_target target)
886{
887 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
888
889 BUG_ON(dir == DMA_NONE);
890
891 if (is_swiotlb_buffer(paddr)) {
892 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
893 return;
894 }
895
896 if (dir != DMA_FROM_DEVICE)
897 return;
898
899 dma_mark_clean(phys_to_virt(paddr), size);
900}
901
902void
903swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
904 size_t size, enum dma_data_direction dir)
905{
906 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
907}
908EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
909
910void
911swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
912 size_t size, enum dma_data_direction dir)
913{
914 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
915}
916EXPORT_SYMBOL(swiotlb_sync_single_for_device);
917
918/*
919 * Map a set of buffers described by scatterlist in streaming mode for DMA.
920 * This is the scatter-gather version of the above swiotlb_map_page
921 * interface. Here the scatter gather list elements are each tagged with the
922 * appropriate dma address and length. They are obtained via
923 * sg_dma_{address,length}(SG).
924 *
925 * NOTE: An implementation may be able to use a smaller number of
926 * DMA address/length pairs than there are SG table elements.
927 * (for example via virtual mapping capabilities)
928 * The routine returns the number of addr/length pairs actually
929 * used, at most nents.
930 *
931 * Device ownership issues as mentioned above for swiotlb_map_page are the
932 * same here.
933 */
934int
935swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
936 enum dma_data_direction dir, unsigned long attrs)
937{
938 struct scatterlist *sg;
939 int i;
940
941 BUG_ON(dir == DMA_NONE);
942
943 for_each_sg(sgl, sg, nelems, i) {
944 phys_addr_t paddr = sg_phys(sg);
945 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
946
947 if (swiotlb_force == SWIOTLB_FORCE ||
948 !dma_capable(hwdev, dev_addr, sg->length)) {
949 phys_addr_t map = map_single(hwdev, sg_phys(sg),
950 sg->length, dir, attrs);
951 if (map == SWIOTLB_MAP_ERROR) {
952 /* Don't panic here, we expect map_sg users
953 to do proper error handling. */
954 swiotlb_full(hwdev, sg->length, dir, 0);
955 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
956 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
957 attrs);
958 sg_dma_len(sgl) = 0;
959 return 0;
960 }
961 sg->dma_address = phys_to_dma(hwdev, map);
962 } else
963 sg->dma_address = dev_addr;
964 sg_dma_len(sg) = sg->length;
965 }
966 return nelems;
967}
968EXPORT_SYMBOL(swiotlb_map_sg_attrs);
969
970/*
971 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
972 * concerning calls here are the same as for swiotlb_unmap_page() above.
973 */
974void
975swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
976 int nelems, enum dma_data_direction dir,
977 unsigned long attrs)
978{
979 struct scatterlist *sg;
980 int i;
981
982 BUG_ON(dir == DMA_NONE);
983
984 for_each_sg(sgl, sg, nelems, i)
985 unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir,
986 attrs);
987}
988EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
989
990/*
991 * Make physical memory consistent for a set of streaming mode DMA translations
992 * after a transfer.
993 *
994 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
995 * and usage.
996 */
997static void
998swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
999 int nelems, enum dma_data_direction dir,
1000 enum dma_sync_target target)
1001{
1002 struct scatterlist *sg;
1003 int i;
1004
1005 for_each_sg(sgl, sg, nelems, i)
1006 swiotlb_sync_single(hwdev, sg->dma_address,
1007 sg_dma_len(sg), dir, target);
1008}
1009
1010void
1011swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
1012 int nelems, enum dma_data_direction dir)
1013{
1014 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
1015}
1016EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
1017
1018void
1019swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
1020 int nelems, enum dma_data_direction dir)
1021{
1022 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
1023}
1024EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
1025
1026int
1027swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
1028{
1029 return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
1030}
1031EXPORT_SYMBOL(swiotlb_dma_mapping_error);
1032
1033/*
1034 * Return whether the given device DMA address mask can be supported
1035 * properly. For example, if your device can only drive the low 24-bits
1036 * during bus mastering, then you would pass 0x00ffffff as the mask to
1037 * this function.
1038 */
1039int
1040swiotlb_dma_supported(struct device *hwdev, u64 mask)
1041{
1042 return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
1043}
1044EXPORT_SYMBOL(swiotlb_dma_supported);
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
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
53int swiotlb_force;
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 */
60static 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 */
66static unsigned long io_tlb_nslabs;
67
68/*
69 * When the IOMMU overflows we return a fallback buffer. This sets the size.
70 */
71static unsigned long io_tlb_overflow = 32*1024;
72
73static 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 */
79static unsigned int *io_tlb_list;
80static 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 */
86static phys_addr_t *io_tlb_orig_addr;
87
88/*
89 * Protect the above data structures in the map and unmap calls
90 */
91static DEFINE_SPINLOCK(io_tlb_lock);
92
93static int late_alloc;
94
95static int __init
96setup_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
113unsigned long swiotlb_nr_tbl(void)
114{
115 return io_tlb_nslabs;
116}
117EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
118/* Note that this doesn't work with highmem page */
119static 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
125void swiotlb_print_info(void)
126{
127 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
128 phys_addr_t pstart, pend;
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
138void __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)
166 swiotlb_print_info();
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 */
173void __init
174swiotlb_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
195void __init
196swiotlb_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 */
206int
207swiotlb_late_init_with_default_size(size_t default_size)
208{
209 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
210 unsigned int order;
211
212 if (!io_tlb_nslabs) {
213 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
214 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
215 }
216
217 /*
218 * Get IO TLB memory from the low pages
219 */
220 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
221 io_tlb_nslabs = SLABS_PER_PAGE << order;
222 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
223
224 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
225 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
226 order);
227 if (io_tlb_start)
228 break;
229 order--;
230 }
231
232 if (!io_tlb_start)
233 goto cleanup1;
234
235 if (order != get_order(bytes)) {
236 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
237 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
238 io_tlb_nslabs = SLABS_PER_PAGE << order;
239 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
240 }
241 io_tlb_end = io_tlb_start + bytes;
242 memset(io_tlb_start, 0, bytes);
243
244 /*
245 * Allocate and initialize the free list array. This array is used
246 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
247 * between io_tlb_start and io_tlb_end.
248 */
249 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
250 get_order(io_tlb_nslabs * sizeof(int)));
251 if (!io_tlb_list)
252 goto cleanup2;
253
254 for (i = 0; i < io_tlb_nslabs; i++)
255 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
256 io_tlb_index = 0;
257
258 io_tlb_orig_addr = (phys_addr_t *)
259 __get_free_pages(GFP_KERNEL,
260 get_order(io_tlb_nslabs *
261 sizeof(phys_addr_t)));
262 if (!io_tlb_orig_addr)
263 goto cleanup3;
264
265 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
266
267 /*
268 * Get the overflow emergency buffer
269 */
270 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
271 get_order(io_tlb_overflow));
272 if (!io_tlb_overflow_buffer)
273 goto cleanup4;
274
275 swiotlb_print_info();
276
277 late_alloc = 1;
278
279 return 0;
280
281cleanup4:
282 free_pages((unsigned long)io_tlb_orig_addr,
283 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
284 io_tlb_orig_addr = NULL;
285cleanup3:
286 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
287 sizeof(int)));
288 io_tlb_list = NULL;
289cleanup2:
290 io_tlb_end = NULL;
291 free_pages((unsigned long)io_tlb_start, order);
292 io_tlb_start = NULL;
293cleanup1:
294 io_tlb_nslabs = req_nslabs;
295 return -ENOMEM;
296}
297
298void __init swiotlb_free(void)
299{
300 if (!io_tlb_overflow_buffer)
301 return;
302
303 if (late_alloc) {
304 free_pages((unsigned long)io_tlb_overflow_buffer,
305 get_order(io_tlb_overflow));
306 free_pages((unsigned long)io_tlb_orig_addr,
307 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
308 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
309 sizeof(int)));
310 free_pages((unsigned long)io_tlb_start,
311 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
312 } else {
313 free_bootmem_late(__pa(io_tlb_overflow_buffer),
314 PAGE_ALIGN(io_tlb_overflow));
315 free_bootmem_late(__pa(io_tlb_orig_addr),
316 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
317 free_bootmem_late(__pa(io_tlb_list),
318 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
319 free_bootmem_late(__pa(io_tlb_start),
320 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
321 }
322 io_tlb_nslabs = 0;
323}
324
325static int is_swiotlb_buffer(phys_addr_t paddr)
326{
327 return paddr >= virt_to_phys(io_tlb_start) &&
328 paddr < virt_to_phys(io_tlb_end);
329}
330
331/*
332 * Bounce: copy the swiotlb buffer back to the original dma location
333 */
334void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
335 enum dma_data_direction dir)
336{
337 unsigned long pfn = PFN_DOWN(phys);
338
339 if (PageHighMem(pfn_to_page(pfn))) {
340 /* The buffer does not have a mapping. Map it in and copy */
341 unsigned int offset = phys & ~PAGE_MASK;
342 char *buffer;
343 unsigned int sz = 0;
344 unsigned long flags;
345
346 while (size) {
347 sz = min_t(size_t, PAGE_SIZE - offset, size);
348
349 local_irq_save(flags);
350 buffer = kmap_atomic(pfn_to_page(pfn));
351 if (dir == DMA_TO_DEVICE)
352 memcpy(dma_addr, buffer + offset, sz);
353 else
354 memcpy(buffer + offset, dma_addr, sz);
355 kunmap_atomic(buffer);
356 local_irq_restore(flags);
357
358 size -= sz;
359 pfn++;
360 dma_addr += sz;
361 offset = 0;
362 }
363 } else {
364 if (dir == DMA_TO_DEVICE)
365 memcpy(dma_addr, phys_to_virt(phys), size);
366 else
367 memcpy(phys_to_virt(phys), dma_addr, size);
368 }
369}
370EXPORT_SYMBOL_GPL(swiotlb_bounce);
371
372void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
373 phys_addr_t phys, size_t size,
374 enum dma_data_direction dir)
375{
376 unsigned long flags;
377 char *dma_addr;
378 unsigned int nslots, stride, index, wrap;
379 int i;
380 unsigned long mask;
381 unsigned long offset_slots;
382 unsigned long max_slots;
383
384 mask = dma_get_seg_boundary(hwdev);
385
386 tbl_dma_addr &= mask;
387
388 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
389
390 /*
391 * Carefully handle integer overflow which can occur when mask == ~0UL.
392 */
393 max_slots = mask + 1
394 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
395 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
396
397 /*
398 * For mappings greater than a page, we limit the stride (and
399 * hence alignment) to a page size.
400 */
401 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
402 if (size > PAGE_SIZE)
403 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
404 else
405 stride = 1;
406
407 BUG_ON(!nslots);
408
409 /*
410 * Find suitable number of IO TLB entries size that will fit this
411 * request and allocate a buffer from that IO TLB pool.
412 */
413 spin_lock_irqsave(&io_tlb_lock, flags);
414 index = ALIGN(io_tlb_index, stride);
415 if (index >= io_tlb_nslabs)
416 index = 0;
417 wrap = index;
418
419 do {
420 while (iommu_is_span_boundary(index, nslots, offset_slots,
421 max_slots)) {
422 index += stride;
423 if (index >= io_tlb_nslabs)
424 index = 0;
425 if (index == wrap)
426 goto not_found;
427 }
428
429 /*
430 * If we find a slot that indicates we have 'nslots' number of
431 * contiguous buffers, we allocate the buffers from that slot
432 * and mark the entries as '0' indicating unavailable.
433 */
434 if (io_tlb_list[index] >= nslots) {
435 int count = 0;
436
437 for (i = index; i < (int) (index + nslots); i++)
438 io_tlb_list[i] = 0;
439 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
440 io_tlb_list[i] = ++count;
441 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
442
443 /*
444 * Update the indices to avoid searching in the next
445 * round.
446 */
447 io_tlb_index = ((index + nslots) < io_tlb_nslabs
448 ? (index + nslots) : 0);
449
450 goto found;
451 }
452 index += stride;
453 if (index >= io_tlb_nslabs)
454 index = 0;
455 } while (index != wrap);
456
457not_found:
458 spin_unlock_irqrestore(&io_tlb_lock, flags);
459 return NULL;
460found:
461 spin_unlock_irqrestore(&io_tlb_lock, flags);
462
463 /*
464 * Save away the mapping from the original address to the DMA address.
465 * This is needed when we sync the memory. Then we sync the buffer if
466 * needed.
467 */
468 for (i = 0; i < nslots; i++)
469 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
470 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
471 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
472
473 return dma_addr;
474}
475EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
476
477/*
478 * Allocates bounce buffer and returns its kernel virtual address.
479 */
480
481static void *
482map_single(struct device *hwdev, phys_addr_t phys, size_t size,
483 enum dma_data_direction dir)
484{
485 dma_addr_t start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start);
486
487 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
488}
489
490/*
491 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
492 */
493void
494swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
495 enum dma_data_direction dir)
496{
497 unsigned long flags;
498 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
499 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
500 phys_addr_t phys = io_tlb_orig_addr[index];
501
502 /*
503 * First, sync the memory before unmapping the entry
504 */
505 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
506 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
507
508 /*
509 * Return the buffer to the free list by setting the corresponding
510 * entries to indicate the number of contiguous entries available.
511 * While returning the entries to the free list, we merge the entries
512 * with slots below and above the pool being returned.
513 */
514 spin_lock_irqsave(&io_tlb_lock, flags);
515 {
516 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
517 io_tlb_list[index + nslots] : 0);
518 /*
519 * Step 1: return the slots to the free list, merging the
520 * slots with superceeding slots
521 */
522 for (i = index + nslots - 1; i >= index; i--)
523 io_tlb_list[i] = ++count;
524 /*
525 * Step 2: merge the returned slots with the preceding slots,
526 * if available (non zero)
527 */
528 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
529 io_tlb_list[i] = ++count;
530 }
531 spin_unlock_irqrestore(&io_tlb_lock, flags);
532}
533EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
534
535void
536swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
537 enum dma_data_direction dir,
538 enum dma_sync_target target)
539{
540 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
541 phys_addr_t phys = io_tlb_orig_addr[index];
542
543 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
544
545 switch (target) {
546 case SYNC_FOR_CPU:
547 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
548 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
549 else
550 BUG_ON(dir != DMA_TO_DEVICE);
551 break;
552 case SYNC_FOR_DEVICE:
553 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
554 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
555 else
556 BUG_ON(dir != DMA_FROM_DEVICE);
557 break;
558 default:
559 BUG();
560 }
561}
562EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
563
564void *
565swiotlb_alloc_coherent(struct device *hwdev, size_t size,
566 dma_addr_t *dma_handle, gfp_t flags)
567{
568 dma_addr_t dev_addr;
569 void *ret;
570 int order = get_order(size);
571 u64 dma_mask = DMA_BIT_MASK(32);
572
573 if (hwdev && hwdev->coherent_dma_mask)
574 dma_mask = hwdev->coherent_dma_mask;
575
576 ret = (void *)__get_free_pages(flags, order);
577 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
578 /*
579 * The allocated memory isn't reachable by the device.
580 */
581 free_pages((unsigned long) ret, order);
582 ret = NULL;
583 }
584 if (!ret) {
585 /*
586 * We are either out of memory or the device can't DMA to
587 * GFP_DMA memory; fall back on map_single(), which
588 * will grab memory from the lowest available address range.
589 */
590 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
591 if (!ret)
592 return NULL;
593 }
594
595 memset(ret, 0, size);
596 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
597
598 /* Confirm address can be DMA'd by device */
599 if (dev_addr + size - 1 > dma_mask) {
600 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
601 (unsigned long long)dma_mask,
602 (unsigned long long)dev_addr);
603
604 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
605 swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
606 return NULL;
607 }
608 *dma_handle = dev_addr;
609 return ret;
610}
611EXPORT_SYMBOL(swiotlb_alloc_coherent);
612
613void
614swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
615 dma_addr_t dev_addr)
616{
617 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
618
619 WARN_ON(irqs_disabled());
620 if (!is_swiotlb_buffer(paddr))
621 free_pages((unsigned long)vaddr, get_order(size));
622 else
623 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
624 swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
625}
626EXPORT_SYMBOL(swiotlb_free_coherent);
627
628static void
629swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
630 int do_panic)
631{
632 /*
633 * Ran out of IOMMU space for this operation. This is very bad.
634 * Unfortunately the drivers cannot handle this operation properly.
635 * unless they check for dma_mapping_error (most don't)
636 * When the mapping is small enough return a static buffer to limit
637 * the damage, or panic when the transfer is too big.
638 */
639 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
640 "device %s\n", size, dev ? dev_name(dev) : "?");
641
642 if (size <= io_tlb_overflow || !do_panic)
643 return;
644
645 if (dir == DMA_BIDIRECTIONAL)
646 panic("DMA: Random memory could be DMA accessed\n");
647 if (dir == DMA_FROM_DEVICE)
648 panic("DMA: Random memory could be DMA written\n");
649 if (dir == DMA_TO_DEVICE)
650 panic("DMA: Random memory could be DMA read\n");
651}
652
653/*
654 * Map a single buffer of the indicated size for DMA in streaming mode. The
655 * physical address to use is returned.
656 *
657 * Once the device is given the dma address, the device owns this memory until
658 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
659 */
660dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
661 unsigned long offset, size_t size,
662 enum dma_data_direction dir,
663 struct dma_attrs *attrs)
664{
665 phys_addr_t phys = page_to_phys(page) + offset;
666 dma_addr_t dev_addr = phys_to_dma(dev, phys);
667 void *map;
668
669 BUG_ON(dir == DMA_NONE);
670 /*
671 * If the address happens to be in the device's DMA window,
672 * we can safely return the device addr and not worry about bounce
673 * buffering it.
674 */
675 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
676 return dev_addr;
677
678 /*
679 * Oh well, have to allocate and map a bounce buffer.
680 */
681 map = map_single(dev, phys, size, dir);
682 if (!map) {
683 swiotlb_full(dev, size, dir, 1);
684 map = io_tlb_overflow_buffer;
685 }
686
687 dev_addr = swiotlb_virt_to_bus(dev, map);
688
689 /*
690 * Ensure that the address returned is DMA'ble
691 */
692 if (!dma_capable(dev, dev_addr, size)) {
693 swiotlb_tbl_unmap_single(dev, map, size, dir);
694 dev_addr = swiotlb_virt_to_bus(dev, io_tlb_overflow_buffer);
695 }
696
697 return dev_addr;
698}
699EXPORT_SYMBOL_GPL(swiotlb_map_page);
700
701/*
702 * Unmap a single streaming mode DMA translation. The dma_addr and size must
703 * match what was provided for in a previous swiotlb_map_page call. All
704 * other usages are undefined.
705 *
706 * After this call, reads by the cpu to the buffer are guaranteed to see
707 * whatever the device wrote there.
708 */
709static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
710 size_t size, enum dma_data_direction dir)
711{
712 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
713
714 BUG_ON(dir == DMA_NONE);
715
716 if (is_swiotlb_buffer(paddr)) {
717 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
718 return;
719 }
720
721 if (dir != DMA_FROM_DEVICE)
722 return;
723
724 /*
725 * phys_to_virt doesn't work with hihgmem page but we could
726 * call dma_mark_clean() with hihgmem page here. However, we
727 * are fine since dma_mark_clean() is null on POWERPC. We can
728 * make dma_mark_clean() take a physical address if necessary.
729 */
730 dma_mark_clean(phys_to_virt(paddr), size);
731}
732
733void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
734 size_t size, enum dma_data_direction dir,
735 struct dma_attrs *attrs)
736{
737 unmap_single(hwdev, dev_addr, size, dir);
738}
739EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
740
741/*
742 * Make physical memory consistent for a single streaming mode DMA translation
743 * after a transfer.
744 *
745 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
746 * using the cpu, yet do not wish to teardown the dma mapping, you must
747 * call this function before doing so. At the next point you give the dma
748 * address back to the card, you must first perform a
749 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
750 */
751static void
752swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
753 size_t size, enum dma_data_direction dir,
754 enum dma_sync_target target)
755{
756 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
757
758 BUG_ON(dir == DMA_NONE);
759
760 if (is_swiotlb_buffer(paddr)) {
761 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
762 target);
763 return;
764 }
765
766 if (dir != DMA_FROM_DEVICE)
767 return;
768
769 dma_mark_clean(phys_to_virt(paddr), size);
770}
771
772void
773swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
774 size_t size, enum dma_data_direction dir)
775{
776 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
777}
778EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
779
780void
781swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
782 size_t size, enum dma_data_direction dir)
783{
784 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
785}
786EXPORT_SYMBOL(swiotlb_sync_single_for_device);
787
788/*
789 * Map a set of buffers described by scatterlist in streaming mode for DMA.
790 * This is the scatter-gather version of the above swiotlb_map_page
791 * interface. Here the scatter gather list elements are each tagged with the
792 * appropriate dma address and length. They are obtained via
793 * sg_dma_{address,length}(SG).
794 *
795 * NOTE: An implementation may be able to use a smaller number of
796 * DMA address/length pairs than there are SG table elements.
797 * (for example via virtual mapping capabilities)
798 * The routine returns the number of addr/length pairs actually
799 * used, at most nents.
800 *
801 * Device ownership issues as mentioned above for swiotlb_map_page are the
802 * same here.
803 */
804int
805swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
806 enum dma_data_direction dir, struct dma_attrs *attrs)
807{
808 struct scatterlist *sg;
809 int i;
810
811 BUG_ON(dir == DMA_NONE);
812
813 for_each_sg(sgl, sg, nelems, i) {
814 phys_addr_t paddr = sg_phys(sg);
815 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
816
817 if (swiotlb_force ||
818 !dma_capable(hwdev, dev_addr, sg->length)) {
819 void *map = map_single(hwdev, sg_phys(sg),
820 sg->length, dir);
821 if (!map) {
822 /* Don't panic here, we expect map_sg users
823 to do proper error handling. */
824 swiotlb_full(hwdev, sg->length, dir, 0);
825 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
826 attrs);
827 sgl[0].dma_length = 0;
828 return 0;
829 }
830 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
831 } else
832 sg->dma_address = dev_addr;
833 sg->dma_length = sg->length;
834 }
835 return nelems;
836}
837EXPORT_SYMBOL(swiotlb_map_sg_attrs);
838
839int
840swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
841 enum dma_data_direction dir)
842{
843 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
844}
845EXPORT_SYMBOL(swiotlb_map_sg);
846
847/*
848 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
849 * concerning calls here are the same as for swiotlb_unmap_page() above.
850 */
851void
852swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
853 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
854{
855 struct scatterlist *sg;
856 int i;
857
858 BUG_ON(dir == DMA_NONE);
859
860 for_each_sg(sgl, sg, nelems, i)
861 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
862
863}
864EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
865
866void
867swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
868 enum dma_data_direction dir)
869{
870 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
871}
872EXPORT_SYMBOL(swiotlb_unmap_sg);
873
874/*
875 * Make physical memory consistent for a set of streaming mode DMA translations
876 * after a transfer.
877 *
878 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
879 * and usage.
880 */
881static void
882swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
883 int nelems, enum dma_data_direction dir,
884 enum dma_sync_target target)
885{
886 struct scatterlist *sg;
887 int i;
888
889 for_each_sg(sgl, sg, nelems, i)
890 swiotlb_sync_single(hwdev, sg->dma_address,
891 sg->dma_length, dir, target);
892}
893
894void
895swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
896 int nelems, enum dma_data_direction dir)
897{
898 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
899}
900EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
901
902void
903swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
904 int nelems, enum dma_data_direction dir)
905{
906 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
907}
908EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
909
910int
911swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
912{
913 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
914}
915EXPORT_SYMBOL(swiotlb_dma_mapping_error);
916
917/*
918 * Return whether the given device DMA address mask can be supported
919 * properly. For example, if your device can only drive the low 24-bits
920 * during bus mastering, then you would pass 0x00ffffff as the mask to
921 * this function.
922 */
923int
924swiotlb_dma_supported(struct device *hwdev, u64 mask)
925{
926 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
927}
928EXPORT_SYMBOL(swiotlb_dma_supported);