blob: 566791b3f583112bbcd1cd8410047b6349fe7017 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Dynamic DMA mapping support.
3 *
4 * This implementation is for IA-64 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 */
16
17#include <linux/cache.h>
18#include <linux/mm.h>
19#include <linux/module.h>
20#include <linux/pci.h>
21#include <linux/spinlock.h>
22#include <linux/string.h>
23#include <linux/types.h>
24#include <linux/ctype.h>
25
26#include <asm/io.h>
27#include <asm/pci.h>
28#include <asm/dma.h>
29
30#include <linux/init.h>
31#include <linux/bootmem.h>
32
33#define OFFSET(val,align) ((unsigned long) \
34 ( (val) & ( (align) - 1)))
35
36#define SG_ENT_VIRT_ADDRESS(sg) (page_address((sg)->page) + (sg)->offset)
37#define SG_ENT_PHYS_ADDRESS(SG) virt_to_phys(SG_ENT_VIRT_ADDRESS(SG))
38
39/*
40 * Maximum allowable number of contiguous slabs to map,
41 * must be a power of 2. What is the appropriate value ?
42 * The complexity of {map,unmap}_single is linearly dependent on this value.
43 */
44#define IO_TLB_SEGSIZE 128
45
46/*
47 * log of the size of each IO TLB slab. The number of slabs is command line
48 * controllable.
49 */
50#define IO_TLB_SHIFT 11
51
Alex Williamson0b9afed2005-09-06 11:20:49 -060052#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
53
54/*
55 * Minimum IO TLB size to bother booting with. Systems with mainly
56 * 64bit capable cards will only lightly use the swiotlb. If we can't
57 * allocate a contiguous 1MB, we're probably in trouble anyway.
58 */
59#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
60
Linus Torvalds1da177e2005-04-16 15:20:36 -070061int swiotlb_force;
62
63/*
64 * Used to do a quick range check in swiotlb_unmap_single and
65 * swiotlb_sync_single_*, to see if the memory was in fact allocated by this
66 * API.
67 */
68static char *io_tlb_start, *io_tlb_end;
69
70/*
71 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
72 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
73 */
74static unsigned long io_tlb_nslabs;
75
76/*
77 * When the IOMMU overflows we return a fallback buffer. This sets the size.
78 */
79static unsigned long io_tlb_overflow = 32*1024;
80
81void *io_tlb_overflow_buffer;
82
83/*
84 * This is a free list describing the number of free entries available from
85 * each index
86 */
87static unsigned int *io_tlb_list;
88static unsigned int io_tlb_index;
89
90/*
91 * We need to save away the original address corresponding to a mapped entry
92 * for the sync operations.
93 */
94static unsigned char **io_tlb_orig_addr;
95
96/*
97 * Protect the above data structures in the map and unmap calls
98 */
99static DEFINE_SPINLOCK(io_tlb_lock);
100
101static int __init
102setup_io_tlb_npages(char *str)
103{
104 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700105 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106 /* avoid tail segment of size < IO_TLB_SEGSIZE */
107 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
108 }
109 if (*str == ',')
110 ++str;
111 if (!strcmp(str, "force"))
112 swiotlb_force = 1;
113 return 1;
114}
115__setup("swiotlb=", setup_io_tlb_npages);
116/* make io_tlb_overflow tunable too? */
117
118/*
119 * Statically reserve bounce buffer space and initialize bounce buffer data
120 * structures for the software IO TLB used to implement the PCI DMA API.
121 */
122void
123swiotlb_init_with_default_size (size_t default_size)
124{
125 unsigned long i;
126
127 if (!io_tlb_nslabs) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700128 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
130 }
131
132 /*
133 * Get IO TLB memory from the low pages
134 */
135 io_tlb_start = alloc_bootmem_low_pages(io_tlb_nslabs *
136 (1 << IO_TLB_SHIFT));
137 if (!io_tlb_start)
138 panic("Cannot allocate SWIOTLB buffer");
139 io_tlb_end = io_tlb_start + io_tlb_nslabs * (1 << IO_TLB_SHIFT);
140
141 /*
142 * Allocate and initialize the free list array. This array is used
143 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
144 * between io_tlb_start and io_tlb_end.
145 */
146 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
147 for (i = 0; i < io_tlb_nslabs; i++)
148 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
149 io_tlb_index = 0;
150 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(char *));
151
152 /*
153 * Get the overflow emergency buffer
154 */
155 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
156 printk(KERN_INFO "Placing software IO TLB between 0x%lx - 0x%lx\n",
157 virt_to_phys(io_tlb_start), virt_to_phys(io_tlb_end));
158}
159
160void
161swiotlb_init (void)
162{
163 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
164}
165
Alex Williamson0b9afed2005-09-06 11:20:49 -0600166/*
167 * Systems with larger DMA zones (those that don't support ISA) can
168 * initialize the swiotlb later using the slab allocator if needed.
169 * This should be just like above, but with some error catching.
170 */
171int
172swiotlb_late_init_with_default_size (size_t default_size)
173{
174 unsigned long i, req_nslabs = io_tlb_nslabs;
175 unsigned int order;
176
177 if (!io_tlb_nslabs) {
178 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
179 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
180 }
181
182 /*
183 * Get IO TLB memory from the low pages
184 */
185 order = get_order(io_tlb_nslabs * (1 << IO_TLB_SHIFT));
186 io_tlb_nslabs = SLABS_PER_PAGE << order;
187
188 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
189 io_tlb_start = (char *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
190 order);
191 if (io_tlb_start)
192 break;
193 order--;
194 }
195
196 if (!io_tlb_start)
197 goto cleanup1;
198
199 if (order != get_order(io_tlb_nslabs * (1 << IO_TLB_SHIFT))) {
200 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
201 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
202 io_tlb_nslabs = SLABS_PER_PAGE << order;
203 }
204 io_tlb_end = io_tlb_start + io_tlb_nslabs * (1 << IO_TLB_SHIFT);
205 memset(io_tlb_start, 0, io_tlb_nslabs * (1 << IO_TLB_SHIFT));
206
207 /*
208 * Allocate and initialize the free list array. This array is used
209 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
210 * between io_tlb_start and io_tlb_end.
211 */
212 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
213 get_order(io_tlb_nslabs * sizeof(int)));
214 if (!io_tlb_list)
215 goto cleanup2;
216
217 for (i = 0; i < io_tlb_nslabs; i++)
218 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
219 io_tlb_index = 0;
220
221 io_tlb_orig_addr = (unsigned char **)__get_free_pages(GFP_KERNEL,
222 get_order(io_tlb_nslabs * sizeof(char *)));
223 if (!io_tlb_orig_addr)
224 goto cleanup3;
225
226 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(char *));
227
228 /*
229 * Get the overflow emergency buffer
230 */
231 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
232 get_order(io_tlb_overflow));
233 if (!io_tlb_overflow_buffer)
234 goto cleanup4;
235
236 printk(KERN_INFO "Placing %ldMB software IO TLB between 0x%lx - "
237 "0x%lx\n", (io_tlb_nslabs * (1 << IO_TLB_SHIFT)) >> 20,
238 virt_to_phys(io_tlb_start), virt_to_phys(io_tlb_end));
239
240 return 0;
241
242cleanup4:
243 free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs *
244 sizeof(char *)));
245 io_tlb_orig_addr = NULL;
246cleanup3:
247 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
248 sizeof(int)));
249 io_tlb_list = NULL;
250 io_tlb_end = NULL;
251cleanup2:
252 free_pages((unsigned long)io_tlb_start, order);
253 io_tlb_start = NULL;
254cleanup1:
255 io_tlb_nslabs = req_nslabs;
256 return -ENOMEM;
257}
258
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259static inline int
260address_needs_mapping(struct device *hwdev, dma_addr_t addr)
261{
262 dma_addr_t mask = 0xffffffff;
263 /* If the device has a mask, use it, otherwise default to 32 bits */
264 if (hwdev && hwdev->dma_mask)
265 mask = *hwdev->dma_mask;
266 return (addr & ~mask) != 0;
267}
268
269/*
270 * Allocates bounce buffer and returns its kernel virtual address.
271 */
272static void *
273map_single(struct device *hwdev, char *buffer, size_t size, int dir)
274{
275 unsigned long flags;
276 char *dma_addr;
277 unsigned int nslots, stride, index, wrap;
278 int i;
279
280 /*
281 * For mappings greater than a page, we limit the stride (and
282 * hence alignment) to a page size.
283 */
284 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
285 if (size > PAGE_SIZE)
286 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
287 else
288 stride = 1;
289
290 if (!nslots)
291 BUG();
292
293 /*
294 * Find suitable number of IO TLB entries size that will fit this
295 * request and allocate a buffer from that IO TLB pool.
296 */
297 spin_lock_irqsave(&io_tlb_lock, flags);
298 {
299 wrap = index = ALIGN(io_tlb_index, stride);
300
301 if (index >= io_tlb_nslabs)
302 wrap = index = 0;
303
304 do {
305 /*
306 * If we find a slot that indicates we have 'nslots'
307 * number of contiguous buffers, we allocate the
308 * buffers from that slot and mark the entries as '0'
309 * indicating unavailable.
310 */
311 if (io_tlb_list[index] >= nslots) {
312 int count = 0;
313
314 for (i = index; i < (int) (index + nslots); i++)
315 io_tlb_list[i] = 0;
316 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
317 io_tlb_list[i] = ++count;
318 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
319
320 /*
321 * Update the indices to avoid searching in
322 * the next round.
323 */
324 io_tlb_index = ((index + nslots) < io_tlb_nslabs
325 ? (index + nslots) : 0);
326
327 goto found;
328 }
329 index += stride;
330 if (index >= io_tlb_nslabs)
331 index = 0;
332 } while (index != wrap);
333
334 spin_unlock_irqrestore(&io_tlb_lock, flags);
335 return NULL;
336 }
337 found:
338 spin_unlock_irqrestore(&io_tlb_lock, flags);
339
340 /*
341 * Save away the mapping from the original address to the DMA address.
342 * This is needed when we sync the memory. Then we sync the buffer if
343 * needed.
344 */
345 io_tlb_orig_addr[index] = buffer;
346 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
347 memcpy(dma_addr, buffer, size);
348
349 return dma_addr;
350}
351
352/*
353 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
354 */
355static void
356unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
357{
358 unsigned long flags;
359 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
360 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
361 char *buffer = io_tlb_orig_addr[index];
362
363 /*
364 * First, sync the memory before unmapping the entry
365 */
366 if (buffer && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
367 /*
368 * bounce... copy the data back into the original buffer * and
369 * delete the bounce buffer.
370 */
371 memcpy(buffer, dma_addr, size);
372
373 /*
374 * Return the buffer to the free list by setting the corresponding
375 * entries to indicate the number of contigous entries available.
376 * While returning the entries to the free list, we merge the entries
377 * with slots below and above the pool being returned.
378 */
379 spin_lock_irqsave(&io_tlb_lock, flags);
380 {
381 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
382 io_tlb_list[index + nslots] : 0);
383 /*
384 * Step 1: return the slots to the free list, merging the
385 * slots with superceeding slots
386 */
387 for (i = index + nslots - 1; i >= index; i--)
388 io_tlb_list[i] = ++count;
389 /*
390 * Step 2: merge the returned slots with the preceding slots,
391 * if available (non zero)
392 */
393 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
394 io_tlb_list[i] = ++count;
395 }
396 spin_unlock_irqrestore(&io_tlb_lock, flags);
397}
398
399static void
400sync_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
401{
402 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
403 char *buffer = io_tlb_orig_addr[index];
404
405 /*
406 * bounce... copy the data back into/from the original buffer
407 * XXX How do you handle DMA_BIDIRECTIONAL here ?
408 */
409 if (dir == DMA_FROM_DEVICE)
410 memcpy(buffer, dma_addr, size);
411 else if (dir == DMA_TO_DEVICE)
412 memcpy(dma_addr, buffer, size);
413 else
414 BUG();
415}
416
417void *
418swiotlb_alloc_coherent(struct device *hwdev, size_t size,
419 dma_addr_t *dma_handle, int flags)
420{
421 unsigned long dev_addr;
422 void *ret;
423 int order = get_order(size);
424
425 /*
426 * XXX fix me: the DMA API should pass us an explicit DMA mask
427 * instead, or use ZONE_DMA32 (ia64 overloads ZONE_DMA to be a ~32
428 * bit range instead of a 16MB one).
429 */
430 flags |= GFP_DMA;
431
432 ret = (void *)__get_free_pages(flags, order);
433 if (ret && address_needs_mapping(hwdev, virt_to_phys(ret))) {
434 /*
435 * The allocated memory isn't reachable by the device.
436 * Fall back on swiotlb_map_single().
437 */
438 free_pages((unsigned long) ret, order);
439 ret = NULL;
440 }
441 if (!ret) {
442 /*
443 * We are either out of memory or the device can't DMA
444 * to GFP_DMA memory; fall back on
445 * swiotlb_map_single(), which will grab memory from
446 * the lowest available address range.
447 */
448 dma_addr_t handle;
449 handle = swiotlb_map_single(NULL, NULL, size, DMA_FROM_DEVICE);
450 if (dma_mapping_error(handle))
451 return NULL;
452
453 ret = phys_to_virt(handle);
454 }
455
456 memset(ret, 0, size);
457 dev_addr = virt_to_phys(ret);
458
459 /* Confirm address can be DMA'd by device */
460 if (address_needs_mapping(hwdev, dev_addr)) {
461 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016lx\n",
462 (unsigned long long)*hwdev->dma_mask, dev_addr);
463 panic("swiotlb_alloc_coherent: allocated memory is out of "
464 "range for device");
465 }
466 *dma_handle = dev_addr;
467 return ret;
468}
469
470void
471swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
472 dma_addr_t dma_handle)
473{
474 if (!(vaddr >= (void *)io_tlb_start
475 && vaddr < (void *)io_tlb_end))
476 free_pages((unsigned long) vaddr, get_order(size));
477 else
478 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
479 swiotlb_unmap_single (hwdev, dma_handle, size, DMA_TO_DEVICE);
480}
481
482static void
483swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
484{
485 /*
486 * Ran out of IOMMU space for this operation. This is very bad.
487 * Unfortunately the drivers cannot handle this operation properly.
488 * unless they check for pci_dma_mapping_error (most don't)
489 * When the mapping is small enough return a static buffer to limit
490 * the damage, or panic when the transfer is too big.
491 */
492 printk(KERN_ERR "PCI-DMA: Out of SW-IOMMU space for %lu bytes at "
493 "device %s\n", size, dev ? dev->bus_id : "?");
494
495 if (size > io_tlb_overflow && do_panic) {
496 if (dir == PCI_DMA_FROMDEVICE || dir == PCI_DMA_BIDIRECTIONAL)
497 panic("PCI-DMA: Memory would be corrupted\n");
498 if (dir == PCI_DMA_TODEVICE || dir == PCI_DMA_BIDIRECTIONAL)
499 panic("PCI-DMA: Random memory would be DMAed\n");
500 }
501}
502
503/*
504 * Map a single buffer of the indicated size for DMA in streaming mode. The
505 * PCI address to use is returned.
506 *
507 * Once the device is given the dma address, the device owns this memory until
508 * either swiotlb_unmap_single or swiotlb_dma_sync_single is performed.
509 */
510dma_addr_t
511swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir)
512{
513 unsigned long dev_addr = virt_to_phys(ptr);
514 void *map;
515
516 if (dir == DMA_NONE)
517 BUG();
518 /*
519 * If the pointer passed in happens to be in the device's DMA window,
520 * we can safely return the device addr and not worry about bounce
521 * buffering it.
522 */
523 if (!address_needs_mapping(hwdev, dev_addr) && !swiotlb_force)
524 return dev_addr;
525
526 /*
527 * Oh well, have to allocate and map a bounce buffer.
528 */
529 map = map_single(hwdev, ptr, size, dir);
530 if (!map) {
531 swiotlb_full(hwdev, size, dir, 1);
532 map = io_tlb_overflow_buffer;
533 }
534
535 dev_addr = virt_to_phys(map);
536
537 /*
538 * Ensure that the address returned is DMA'ble
539 */
540 if (address_needs_mapping(hwdev, dev_addr))
541 panic("map_single: bounce buffer is not DMA'ble");
542
543 return dev_addr;
544}
545
546/*
547 * Since DMA is i-cache coherent, any (complete) pages that were written via
548 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
549 * flush them when they get mapped into an executable vm-area.
550 */
551static void
552mark_clean(void *addr, size_t size)
553{
554 unsigned long pg_addr, end;
555
556 pg_addr = PAGE_ALIGN((unsigned long) addr);
557 end = (unsigned long) addr + size;
558 while (pg_addr + PAGE_SIZE <= end) {
559 struct page *page = virt_to_page(pg_addr);
560 set_bit(PG_arch_1, &page->flags);
561 pg_addr += PAGE_SIZE;
562 }
563}
564
565/*
566 * Unmap a single streaming mode DMA translation. The dma_addr and size must
567 * match what was provided for in a previous swiotlb_map_single call. All
568 * other usages are undefined.
569 *
570 * After this call, reads by the cpu to the buffer are guaranteed to see
571 * whatever the device wrote there.
572 */
573void
574swiotlb_unmap_single(struct device *hwdev, dma_addr_t dev_addr, size_t size,
575 int dir)
576{
577 char *dma_addr = phys_to_virt(dev_addr);
578
579 if (dir == DMA_NONE)
580 BUG();
581 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
582 unmap_single(hwdev, dma_addr, size, dir);
583 else if (dir == DMA_FROM_DEVICE)
584 mark_clean(dma_addr, size);
585}
586
587/*
588 * Make physical memory consistent for a single streaming mode DMA translation
589 * after a transfer.
590 *
591 * If you perform a swiotlb_map_single() but wish to interrogate the buffer
592 * using the cpu, yet do not wish to teardown the PCI dma mapping, you must
593 * call this function before doing so. At the next point you give the PCI dma
594 * address back to the card, you must first perform a
595 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
596 */
John W. Linville8270f3f2005-09-29 14:43:32 -0700597static inline void
598swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
599 size_t size, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600{
601 char *dma_addr = phys_to_virt(dev_addr);
602
603 if (dir == DMA_NONE)
604 BUG();
605 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
606 sync_single(hwdev, dma_addr, size, dir);
607 else if (dir == DMA_FROM_DEVICE)
608 mark_clean(dma_addr, size);
609}
610
611void
John W. Linville8270f3f2005-09-29 14:43:32 -0700612swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
613 size_t size, int dir)
614{
615 swiotlb_sync_single(hwdev, dev_addr, size, dir);
616}
617
618void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
620 size_t size, int dir)
621{
John W. Linville8270f3f2005-09-29 14:43:32 -0700622 swiotlb_sync_single(hwdev, dev_addr, size, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623}
624
625/*
John W. Linville878a97c2005-09-29 14:44:23 -0700626 * Same as above, but for a sub-range of the mapping.
627 */
628static inline void
629swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
630 unsigned long offset, size_t size, int dir)
631{
632 char *dma_addr = phys_to_virt(dev_addr) + offset;
633
634 if (dir == DMA_NONE)
635 BUG();
636 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
637 sync_single(hwdev, dma_addr, size, dir);
638 else if (dir == DMA_FROM_DEVICE)
639 mark_clean(dma_addr, size);
640}
641
642void
643swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
644 unsigned long offset, size_t size, int dir)
645{
646 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir);
647}
648
649void
650swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
651 unsigned long offset, size_t size, int dir)
652{
653 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir);
654}
655
656/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 * Map a set of buffers described by scatterlist in streaming mode for DMA.
658 * This is the scatter-gather version of the above swiotlb_map_single
659 * interface. Here the scatter gather list elements are each tagged with the
660 * appropriate dma address and length. They are obtained via
661 * sg_dma_{address,length}(SG).
662 *
663 * NOTE: An implementation may be able to use a smaller number of
664 * DMA address/length pairs than there are SG table elements.
665 * (for example via virtual mapping capabilities)
666 * The routine returns the number of addr/length pairs actually
667 * used, at most nents.
668 *
669 * Device ownership issues as mentioned above for swiotlb_map_single are the
670 * same here.
671 */
672int
673swiotlb_map_sg(struct device *hwdev, struct scatterlist *sg, int nelems,
674 int dir)
675{
676 void *addr;
677 unsigned long dev_addr;
678 int i;
679
680 if (dir == DMA_NONE)
681 BUG();
682
683 for (i = 0; i < nelems; i++, sg++) {
684 addr = SG_ENT_VIRT_ADDRESS(sg);
685 dev_addr = virt_to_phys(addr);
686 if (swiotlb_force || address_needs_mapping(hwdev, dev_addr)) {
687 sg->dma_address = (dma_addr_t) virt_to_phys(map_single(hwdev, addr, sg->length, dir));
688 if (!sg->dma_address) {
689 /* Don't panic here, we expect map_sg users
690 to do proper error handling. */
691 swiotlb_full(hwdev, sg->length, dir, 0);
692 swiotlb_unmap_sg(hwdev, sg - i, i, dir);
693 sg[0].dma_length = 0;
694 return 0;
695 }
696 } else
697 sg->dma_address = dev_addr;
698 sg->dma_length = sg->length;
699 }
700 return nelems;
701}
702
703/*
704 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
705 * concerning calls here are the same as for swiotlb_unmap_single() above.
706 */
707void
708swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sg, int nelems,
709 int dir)
710{
711 int i;
712
713 if (dir == DMA_NONE)
714 BUG();
715
716 for (i = 0; i < nelems; i++, sg++)
717 if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
718 unmap_single(hwdev, (void *) phys_to_virt(sg->dma_address), sg->dma_length, dir);
719 else if (dir == DMA_FROM_DEVICE)
720 mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length);
721}
722
723/*
724 * Make physical memory consistent for a set of streaming mode DMA translations
725 * after a transfer.
726 *
727 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
728 * and usage.
729 */
John W. Linville8270f3f2005-09-29 14:43:32 -0700730static inline void
731swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sg,
732 int nelems, int dir)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733{
734 int i;
735
736 if (dir == DMA_NONE)
737 BUG();
738
739 for (i = 0; i < nelems; i++, sg++)
740 if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
741 sync_single(hwdev, (void *) sg->dma_address,
742 sg->dma_length, dir);
743}
744
745void
John W. Linville8270f3f2005-09-29 14:43:32 -0700746swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
747 int nelems, int dir)
748{
749 swiotlb_sync_sg(hwdev, sg, nelems, dir);
750}
751
752void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
754 int nelems, int dir)
755{
John W. Linville8270f3f2005-09-29 14:43:32 -0700756 swiotlb_sync_sg(hwdev, sg, nelems, dir);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757}
758
759int
760swiotlb_dma_mapping_error(dma_addr_t dma_addr)
761{
762 return (dma_addr == virt_to_phys(io_tlb_overflow_buffer));
763}
764
765/*
766 * Return whether the given PCI device DMA address mask can be supported
767 * properly. For example, if your device can only drive the low 24-bits
768 * during PCI bus mastering, then you would pass 0x00ffffff as the mask to
769 * this function.
770 */
771int
772swiotlb_dma_supported (struct device *hwdev, u64 mask)
773{
774 return (virt_to_phys (io_tlb_end) - 1) <= mask;
775}
776
777EXPORT_SYMBOL(swiotlb_init);
778EXPORT_SYMBOL(swiotlb_map_single);
779EXPORT_SYMBOL(swiotlb_unmap_single);
780EXPORT_SYMBOL(swiotlb_map_sg);
781EXPORT_SYMBOL(swiotlb_unmap_sg);
782EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
783EXPORT_SYMBOL(swiotlb_sync_single_for_device);
John W. Linville878a97c2005-09-29 14:44:23 -0700784EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
785EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
787EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
788EXPORT_SYMBOL(swiotlb_dma_mapping_error);
789EXPORT_SYMBOL(swiotlb_alloc_coherent);
790EXPORT_SYMBOL(swiotlb_free_coherent);
791EXPORT_SYMBOL(swiotlb_dma_supported);