blob: 1ff8dcebf7c6edf33abb7eb2583d5b6eab37ce29 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Dynamic DMA mapping support.
3 *
John W. Linville569c8bf2005-09-29 14:45:24 -07004 * This implementation is for IA-64 and EM64T platforms that do not support
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 * 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.
John W. Linville569c8bf2005-09-29 14:45:24 -070014 * 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070017 */
18
19#include <linux/cache.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070020#include <linux/dma-mapping.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <linux/mm.h>
22#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/spinlock.h>
24#include <linux/string.h>
25#include <linux/types.h>
26#include <linux/ctype.h>
27
28#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <asm/dma.h>
Tony Luck17e5ad62005-09-29 15:52:13 -070030#include <asm/scatterlist.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031
32#include <linux/init.h>
33#include <linux/bootmem.h>
34
35#define OFFSET(val,align) ((unsigned long) \
36 ( (val) & ( (align) - 1)))
37
38#define SG_ENT_VIRT_ADDRESS(sg) (page_address((sg)->page) + (sg)->offset)
39#define SG_ENT_PHYS_ADDRESS(SG) virt_to_phys(SG_ENT_VIRT_ADDRESS(SG))
40
41/*
42 * Maximum allowable number of contiguous slabs to map,
43 * must be a power of 2. What is the appropriate value ?
44 * The complexity of {map,unmap}_single is linearly dependent on this value.
45 */
46#define IO_TLB_SEGSIZE 128
47
48/*
49 * log of the size of each IO TLB slab. The number of slabs is command line
50 * controllable.
51 */
52#define IO_TLB_SHIFT 11
53
Alex Williamson0b9afed2005-09-06 11:20:49 -060054#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
55
56/*
57 * Minimum IO TLB size to bother booting with. Systems with mainly
58 * 64bit capable cards will only lightly use the swiotlb. If we can't
59 * allocate a contiguous 1MB, we're probably in trouble anyway.
60 */
61#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
62
John W. Linvillede69e0f2005-09-29 14:44:57 -070063/*
64 * Enumeration for sync targets
65 */
66enum dma_sync_target {
67 SYNC_FOR_CPU = 0,
68 SYNC_FOR_DEVICE = 1,
69};
70
Linus Torvalds1da177e2005-04-16 15:20:36 -070071int swiotlb_force;
72
73/*
74 * Used to do a quick range check in swiotlb_unmap_single and
75 * swiotlb_sync_single_*, to see if the memory was in fact allocated by this
76 * API.
77 */
78static char *io_tlb_start, *io_tlb_end;
79
80/*
81 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
82 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
83 */
84static unsigned long io_tlb_nslabs;
85
86/*
87 * When the IOMMU overflows we return a fallback buffer. This sets the size.
88 */
89static unsigned long io_tlb_overflow = 32*1024;
90
91void *io_tlb_overflow_buffer;
92
93/*
94 * This is a free list describing the number of free entries available from
95 * each index
96 */
97static unsigned int *io_tlb_list;
98static unsigned int io_tlb_index;
99
100/*
101 * We need to save away the original address corresponding to a mapped entry
102 * for the sync operations.
103 */
104static unsigned char **io_tlb_orig_addr;
105
106/*
107 * Protect the above data structures in the map and unmap calls
108 */
109static DEFINE_SPINLOCK(io_tlb_lock);
110
111static int __init
112setup_io_tlb_npages(char *str)
113{
114 if (isdigit(*str)) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700115 io_tlb_nslabs = simple_strtoul(str, &str, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116 /* avoid tail segment of size < IO_TLB_SEGSIZE */
117 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
118 }
119 if (*str == ',')
120 ++str;
121 if (!strcmp(str, "force"))
122 swiotlb_force = 1;
123 return 1;
124}
125__setup("swiotlb=", setup_io_tlb_npages);
126/* make io_tlb_overflow tunable too? */
127
128/*
129 * Statically reserve bounce buffer space and initialize bounce buffer data
Tony Luck17e5ad62005-09-29 15:52:13 -0700130 * structures for the software IO TLB used to implement the DMA API.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 */
132void
133swiotlb_init_with_default_size (size_t default_size)
134{
135 unsigned long i;
136
137 if (!io_tlb_nslabs) {
Alex Williamsone8579e72005-08-04 13:06:00 -0700138 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
140 }
141
142 /*
143 * Get IO TLB memory from the low pages
144 */
Yasunori Goto281dd252005-10-19 15:52:18 -0700145 io_tlb_start = alloc_bootmem_low_pages_limit(io_tlb_nslabs *
146 (1 << IO_TLB_SHIFT), 0x100000000);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 if (!io_tlb_start)
148 panic("Cannot allocate SWIOTLB buffer");
149 io_tlb_end = io_tlb_start + io_tlb_nslabs * (1 << IO_TLB_SHIFT);
150
151 /*
152 * Allocate and initialize the free list array. This array is used
153 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
154 * between io_tlb_start and io_tlb_end.
155 */
156 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
157 for (i = 0; i < io_tlb_nslabs; i++)
158 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
159 io_tlb_index = 0;
160 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(char *));
161
162 /*
163 * Get the overflow emergency buffer
164 */
165 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
166 printk(KERN_INFO "Placing software IO TLB between 0x%lx - 0x%lx\n",
167 virt_to_phys(io_tlb_start), virt_to_phys(io_tlb_end));
168}
169
170void
171swiotlb_init (void)
172{
173 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
174}
175
Alex Williamson0b9afed2005-09-06 11:20:49 -0600176/*
177 * Systems with larger DMA zones (those that don't support ISA) can
178 * initialize the swiotlb later using the slab allocator if needed.
179 * This should be just like above, but with some error catching.
180 */
181int
182swiotlb_late_init_with_default_size (size_t default_size)
183{
184 unsigned long i, req_nslabs = io_tlb_nslabs;
185 unsigned int order;
186
187 if (!io_tlb_nslabs) {
188 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
189 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
190 }
191
192 /*
193 * Get IO TLB memory from the low pages
194 */
195 order = get_order(io_tlb_nslabs * (1 << IO_TLB_SHIFT));
196 io_tlb_nslabs = SLABS_PER_PAGE << order;
197
198 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
199 io_tlb_start = (char *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
200 order);
201 if (io_tlb_start)
202 break;
203 order--;
204 }
205
206 if (!io_tlb_start)
207 goto cleanup1;
208
209 if (order != get_order(io_tlb_nslabs * (1 << IO_TLB_SHIFT))) {
210 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
211 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
212 io_tlb_nslabs = SLABS_PER_PAGE << order;
213 }
214 io_tlb_end = io_tlb_start + io_tlb_nslabs * (1 << IO_TLB_SHIFT);
215 memset(io_tlb_start, 0, io_tlb_nslabs * (1 << IO_TLB_SHIFT));
216
217 /*
218 * Allocate and initialize the free list array. This array is used
219 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
220 * between io_tlb_start and io_tlb_end.
221 */
222 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
223 get_order(io_tlb_nslabs * sizeof(int)));
224 if (!io_tlb_list)
225 goto cleanup2;
226
227 for (i = 0; i < io_tlb_nslabs; i++)
228 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
229 io_tlb_index = 0;
230
231 io_tlb_orig_addr = (unsigned char **)__get_free_pages(GFP_KERNEL,
232 get_order(io_tlb_nslabs * sizeof(char *)));
233 if (!io_tlb_orig_addr)
234 goto cleanup3;
235
236 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(char *));
237
238 /*
239 * Get the overflow emergency buffer
240 */
241 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
242 get_order(io_tlb_overflow));
243 if (!io_tlb_overflow_buffer)
244 goto cleanup4;
245
246 printk(KERN_INFO "Placing %ldMB software IO TLB between 0x%lx - "
247 "0x%lx\n", (io_tlb_nslabs * (1 << IO_TLB_SHIFT)) >> 20,
248 virt_to_phys(io_tlb_start), virt_to_phys(io_tlb_end));
249
250 return 0;
251
252cleanup4:
253 free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs *
254 sizeof(char *)));
255 io_tlb_orig_addr = NULL;
256cleanup3:
257 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
258 sizeof(int)));
259 io_tlb_list = NULL;
260 io_tlb_end = NULL;
261cleanup2:
262 free_pages((unsigned long)io_tlb_start, order);
263 io_tlb_start = NULL;
264cleanup1:
265 io_tlb_nslabs = req_nslabs;
266 return -ENOMEM;
267}
268
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269static inline int
270address_needs_mapping(struct device *hwdev, dma_addr_t addr)
271{
272 dma_addr_t mask = 0xffffffff;
273 /* If the device has a mask, use it, otherwise default to 32 bits */
274 if (hwdev && hwdev->dma_mask)
275 mask = *hwdev->dma_mask;
276 return (addr & ~mask) != 0;
277}
278
279/*
280 * Allocates bounce buffer and returns its kernel virtual address.
281 */
282static void *
283map_single(struct device *hwdev, char *buffer, size_t size, int dir)
284{
285 unsigned long flags;
286 char *dma_addr;
287 unsigned int nslots, stride, index, wrap;
288 int i;
289
290 /*
291 * For mappings greater than a page, we limit the stride (and
292 * hence alignment) to a page size.
293 */
294 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
295 if (size > PAGE_SIZE)
296 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
297 else
298 stride = 1;
299
300 if (!nslots)
301 BUG();
302
303 /*
304 * Find suitable number of IO TLB entries size that will fit this
305 * request and allocate a buffer from that IO TLB pool.
306 */
307 spin_lock_irqsave(&io_tlb_lock, flags);
308 {
309 wrap = index = ALIGN(io_tlb_index, stride);
310
311 if (index >= io_tlb_nslabs)
312 wrap = index = 0;
313
314 do {
315 /*
316 * If we find a slot that indicates we have 'nslots'
317 * number of contiguous buffers, we allocate the
318 * buffers from that slot and mark the entries as '0'
319 * indicating unavailable.
320 */
321 if (io_tlb_list[index] >= nslots) {
322 int count = 0;
323
324 for (i = index; i < (int) (index + nslots); i++)
325 io_tlb_list[i] = 0;
326 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
327 io_tlb_list[i] = ++count;
328 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
329
330 /*
331 * Update the indices to avoid searching in
332 * the next round.
333 */
334 io_tlb_index = ((index + nslots) < io_tlb_nslabs
335 ? (index + nslots) : 0);
336
337 goto found;
338 }
339 index += stride;
340 if (index >= io_tlb_nslabs)
341 index = 0;
342 } while (index != wrap);
343
344 spin_unlock_irqrestore(&io_tlb_lock, flags);
345 return NULL;
346 }
347 found:
348 spin_unlock_irqrestore(&io_tlb_lock, flags);
349
350 /*
351 * Save away the mapping from the original address to the DMA address.
352 * This is needed when we sync the memory. Then we sync the buffer if
353 * needed.
354 */
355 io_tlb_orig_addr[index] = buffer;
356 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
357 memcpy(dma_addr, buffer, size);
358
359 return dma_addr;
360}
361
362/*
363 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
364 */
365static void
366unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
367{
368 unsigned long flags;
369 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
370 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
371 char *buffer = io_tlb_orig_addr[index];
372
373 /*
374 * First, sync the memory before unmapping the entry
375 */
376 if (buffer && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
377 /*
378 * bounce... copy the data back into the original buffer * and
379 * delete the bounce buffer.
380 */
381 memcpy(buffer, dma_addr, size);
382
383 /*
384 * Return the buffer to the free list by setting the corresponding
385 * entries to indicate the number of contigous entries available.
386 * While returning the entries to the free list, we merge the entries
387 * with slots below and above the pool being returned.
388 */
389 spin_lock_irqsave(&io_tlb_lock, flags);
390 {
391 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
392 io_tlb_list[index + nslots] : 0);
393 /*
394 * Step 1: return the slots to the free list, merging the
395 * slots with superceeding slots
396 */
397 for (i = index + nslots - 1; i >= index; i--)
398 io_tlb_list[i] = ++count;
399 /*
400 * Step 2: merge the returned slots with the preceding slots,
401 * if available (non zero)
402 */
403 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
404 io_tlb_list[i] = ++count;
405 }
406 spin_unlock_irqrestore(&io_tlb_lock, flags);
407}
408
409static void
John W. Linvillede69e0f2005-09-29 14:44:57 -0700410sync_single(struct device *hwdev, char *dma_addr, size_t size,
411 int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412{
413 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
414 char *buffer = io_tlb_orig_addr[index];
415
John W. Linvillede69e0f2005-09-29 14:44:57 -0700416 switch (target) {
417 case SYNC_FOR_CPU:
418 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
419 memcpy(buffer, dma_addr, size);
420 else if (dir != DMA_TO_DEVICE)
421 BUG();
422 break;
423 case SYNC_FOR_DEVICE:
424 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
425 memcpy(dma_addr, buffer, size);
426 else if (dir != DMA_FROM_DEVICE)
427 BUG();
428 break;
429 default:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430 BUG();
John W. Linvillede69e0f2005-09-29 14:44:57 -0700431 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432}
433
434void *
435swiotlb_alloc_coherent(struct device *hwdev, size_t size,
Al Viro06a54492005-10-21 03:21:03 -0400436 dma_addr_t *dma_handle, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
438 unsigned long dev_addr;
439 void *ret;
440 int order = get_order(size);
441
442 /*
443 * XXX fix me: the DMA API should pass us an explicit DMA mask
444 * instead, or use ZONE_DMA32 (ia64 overloads ZONE_DMA to be a ~32
445 * bit range instead of a 16MB one).
446 */
447 flags |= GFP_DMA;
448
449 ret = (void *)__get_free_pages(flags, order);
450 if (ret && address_needs_mapping(hwdev, virt_to_phys(ret))) {
451 /*
452 * The allocated memory isn't reachable by the device.
453 * Fall back on swiotlb_map_single().
454 */
455 free_pages((unsigned long) ret, order);
456 ret = NULL;
457 }
458 if (!ret) {
459 /*
460 * We are either out of memory or the device can't DMA
461 * to GFP_DMA memory; fall back on
462 * swiotlb_map_single(), which will grab memory from
463 * the lowest available address range.
464 */
465 dma_addr_t handle;
466 handle = swiotlb_map_single(NULL, NULL, size, DMA_FROM_DEVICE);
467 if (dma_mapping_error(handle))
468 return NULL;
469
470 ret = phys_to_virt(handle);
471 }
472
473 memset(ret, 0, size);
474 dev_addr = virt_to_phys(ret);
475
476 /* Confirm address can be DMA'd by device */
477 if (address_needs_mapping(hwdev, dev_addr)) {
478 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016lx\n",
479 (unsigned long long)*hwdev->dma_mask, dev_addr);
480 panic("swiotlb_alloc_coherent: allocated memory is out of "
481 "range for device");
482 }
483 *dma_handle = dev_addr;
484 return ret;
485}
486
487void
488swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
489 dma_addr_t dma_handle)
490{
491 if (!(vaddr >= (void *)io_tlb_start
492 && vaddr < (void *)io_tlb_end))
493 free_pages((unsigned long) vaddr, get_order(size));
494 else
495 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
496 swiotlb_unmap_single (hwdev, dma_handle, size, DMA_TO_DEVICE);
497}
498
499static void
500swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
501{
502 /*
503 * Ran out of IOMMU space for this operation. This is very bad.
504 * Unfortunately the drivers cannot handle this operation properly.
Tony Luck17e5ad62005-09-29 15:52:13 -0700505 * unless they check for dma_mapping_error (most don't)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 * When the mapping is small enough return a static buffer to limit
507 * the damage, or panic when the transfer is too big.
508 */
Tony Luck17e5ad62005-09-29 15:52:13 -0700509 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %lu bytes at "
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 "device %s\n", size, dev ? dev->bus_id : "?");
511
512 if (size > io_tlb_overflow && do_panic) {
Tony Luck17e5ad62005-09-29 15:52:13 -0700513 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
514 panic("DMA: Memory would be corrupted\n");
515 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
516 panic("DMA: Random memory would be DMAed\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 }
518}
519
520/*
521 * Map a single buffer of the indicated size for DMA in streaming mode. The
Tony Luck17e5ad62005-09-29 15:52:13 -0700522 * physical address to use is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 *
524 * Once the device is given the dma address, the device owns this memory until
525 * either swiotlb_unmap_single or swiotlb_dma_sync_single is performed.
526 */
527dma_addr_t
528swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir)
529{
530 unsigned long dev_addr = virt_to_phys(ptr);
531 void *map;
532
533 if (dir == DMA_NONE)
534 BUG();
535 /*
536 * If the pointer passed in happens to be in the device's DMA window,
537 * we can safely return the device addr and not worry about bounce
538 * buffering it.
539 */
540 if (!address_needs_mapping(hwdev, dev_addr) && !swiotlb_force)
541 return dev_addr;
542
543 /*
544 * Oh well, have to allocate and map a bounce buffer.
545 */
546 map = map_single(hwdev, ptr, size, dir);
547 if (!map) {
548 swiotlb_full(hwdev, size, dir, 1);
549 map = io_tlb_overflow_buffer;
550 }
551
552 dev_addr = virt_to_phys(map);
553
554 /*
555 * Ensure that the address returned is DMA'ble
556 */
557 if (address_needs_mapping(hwdev, dev_addr))
558 panic("map_single: bounce buffer is not DMA'ble");
559
560 return dev_addr;
561}
562
563/*
564 * Since DMA is i-cache coherent, any (complete) pages that were written via
565 * DMA can be marked as "clean" so that lazy_mmu_prot_update() doesn't have to
566 * flush them when they get mapped into an executable vm-area.
567 */
568static void
569mark_clean(void *addr, size_t size)
570{
571 unsigned long pg_addr, end;
572
573 pg_addr = PAGE_ALIGN((unsigned long) addr);
574 end = (unsigned long) addr + size;
575 while (pg_addr + PAGE_SIZE <= end) {
576 struct page *page = virt_to_page(pg_addr);
577 set_bit(PG_arch_1, &page->flags);
578 pg_addr += PAGE_SIZE;
579 }
580}
581
582/*
583 * Unmap a single streaming mode DMA translation. The dma_addr and size must
584 * match what was provided for in a previous swiotlb_map_single call. All
585 * other usages are undefined.
586 *
587 * After this call, reads by the cpu to the buffer are guaranteed to see
588 * whatever the device wrote there.
589 */
590void
591swiotlb_unmap_single(struct device *hwdev, dma_addr_t dev_addr, size_t size,
592 int dir)
593{
594 char *dma_addr = phys_to_virt(dev_addr);
595
596 if (dir == DMA_NONE)
597 BUG();
598 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
599 unmap_single(hwdev, dma_addr, size, dir);
600 else if (dir == DMA_FROM_DEVICE)
601 mark_clean(dma_addr, size);
602}
603
604/*
605 * Make physical memory consistent for a single streaming mode DMA translation
606 * after a transfer.
607 *
608 * If you perform a swiotlb_map_single() but wish to interrogate the buffer
Tony Luck17e5ad62005-09-29 15:52:13 -0700609 * using the cpu, yet do not wish to teardown the dma mapping, you must
610 * call this function before doing so. At the next point you give the dma
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 * address back to the card, you must first perform a
612 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
613 */
John W. Linville8270f3f2005-09-29 14:43:32 -0700614static inline void
615swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700616 size_t size, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617{
618 char *dma_addr = phys_to_virt(dev_addr);
619
620 if (dir == DMA_NONE)
621 BUG();
622 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
John W. Linvillede69e0f2005-09-29 14:44:57 -0700623 sync_single(hwdev, dma_addr, size, dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 else if (dir == DMA_FROM_DEVICE)
625 mark_clean(dma_addr, size);
626}
627
628void
John W. Linville8270f3f2005-09-29 14:43:32 -0700629swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
630 size_t size, int dir)
631{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700632 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700633}
634
635void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
637 size_t size, int dir)
638{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700639 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640}
641
642/*
John W. Linville878a97c2005-09-29 14:44:23 -0700643 * Same as above, but for a sub-range of the mapping.
644 */
645static inline void
646swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700647 unsigned long offset, size_t size,
648 int dir, int target)
John W. Linville878a97c2005-09-29 14:44:23 -0700649{
650 char *dma_addr = phys_to_virt(dev_addr) + offset;
651
652 if (dir == DMA_NONE)
653 BUG();
654 if (dma_addr >= io_tlb_start && dma_addr < io_tlb_end)
John W. Linvillede69e0f2005-09-29 14:44:57 -0700655 sync_single(hwdev, dma_addr, size, dir, target);
John W. Linville878a97c2005-09-29 14:44:23 -0700656 else if (dir == DMA_FROM_DEVICE)
657 mark_clean(dma_addr, size);
658}
659
660void
661swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
662 unsigned long offset, size_t size, int dir)
663{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700664 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
665 SYNC_FOR_CPU);
John W. Linville878a97c2005-09-29 14:44:23 -0700666}
667
668void
669swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
670 unsigned long offset, size_t size, int dir)
671{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700672 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
673 SYNC_FOR_DEVICE);
John W. Linville878a97c2005-09-29 14:44:23 -0700674}
675
676/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 * Map a set of buffers described by scatterlist in streaming mode for DMA.
678 * This is the scatter-gather version of the above swiotlb_map_single
679 * interface. Here the scatter gather list elements are each tagged with the
680 * appropriate dma address and length. They are obtained via
681 * sg_dma_{address,length}(SG).
682 *
683 * NOTE: An implementation may be able to use a smaller number of
684 * DMA address/length pairs than there are SG table elements.
685 * (for example via virtual mapping capabilities)
686 * The routine returns the number of addr/length pairs actually
687 * used, at most nents.
688 *
689 * Device ownership issues as mentioned above for swiotlb_map_single are the
690 * same here.
691 */
692int
693swiotlb_map_sg(struct device *hwdev, struct scatterlist *sg, int nelems,
694 int dir)
695{
696 void *addr;
697 unsigned long dev_addr;
698 int i;
699
700 if (dir == DMA_NONE)
701 BUG();
702
703 for (i = 0; i < nelems; i++, sg++) {
704 addr = SG_ENT_VIRT_ADDRESS(sg);
705 dev_addr = virt_to_phys(addr);
706 if (swiotlb_force || address_needs_mapping(hwdev, dev_addr)) {
Andi Kleen7e870232005-12-20 14:45:19 +0100707 void *map = map_single(hwdev, addr, sg->length, dir);
708 sg->dma_address = virt_to_bus(map);
709 if (!map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 /* Don't panic here, we expect map_sg users
711 to do proper error handling. */
712 swiotlb_full(hwdev, sg->length, dir, 0);
713 swiotlb_unmap_sg(hwdev, sg - i, i, dir);
714 sg[0].dma_length = 0;
715 return 0;
716 }
717 } else
718 sg->dma_address = dev_addr;
719 sg->dma_length = sg->length;
720 }
721 return nelems;
722}
723
724/*
725 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
726 * concerning calls here are the same as for swiotlb_unmap_single() above.
727 */
728void
729swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sg, int nelems,
730 int dir)
731{
732 int i;
733
734 if (dir == DMA_NONE)
735 BUG();
736
737 for (i = 0; i < nelems; i++, sg++)
738 if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
739 unmap_single(hwdev, (void *) phys_to_virt(sg->dma_address), sg->dma_length, dir);
740 else if (dir == DMA_FROM_DEVICE)
741 mark_clean(SG_ENT_VIRT_ADDRESS(sg), sg->dma_length);
742}
743
744/*
745 * Make physical memory consistent for a set of streaming mode DMA translations
746 * after a transfer.
747 *
748 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
749 * and usage.
750 */
John W. Linville8270f3f2005-09-29 14:43:32 -0700751static inline void
752swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sg,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700753 int nelems, int dir, int target)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754{
755 int i;
756
757 if (dir == DMA_NONE)
758 BUG();
759
760 for (i = 0; i < nelems; i++, sg++)
761 if (sg->dma_address != SG_ENT_PHYS_ADDRESS(sg))
762 sync_single(hwdev, (void *) sg->dma_address,
John W. Linvillede69e0f2005-09-29 14:44:57 -0700763 sg->dma_length, dir, target);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764}
765
766void
John W. Linville8270f3f2005-09-29 14:43:32 -0700767swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
768 int nelems, int dir)
769{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700770 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
John W. Linville8270f3f2005-09-29 14:43:32 -0700771}
772
773void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
775 int nelems, int dir)
776{
John W. Linvillede69e0f2005-09-29 14:44:57 -0700777 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778}
779
780int
781swiotlb_dma_mapping_error(dma_addr_t dma_addr)
782{
783 return (dma_addr == virt_to_phys(io_tlb_overflow_buffer));
784}
785
786/*
Tony Luck17e5ad62005-09-29 15:52:13 -0700787 * Return whether the given device DMA address mask can be supported
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 * properly. For example, if your device can only drive the low 24-bits
Tony Luck17e5ad62005-09-29 15:52:13 -0700789 * during bus mastering, then you would pass 0x00ffffff as the mask to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 * this function.
791 */
792int
793swiotlb_dma_supported (struct device *hwdev, u64 mask)
794{
795 return (virt_to_phys (io_tlb_end) - 1) <= mask;
796}
797
798EXPORT_SYMBOL(swiotlb_init);
799EXPORT_SYMBOL(swiotlb_map_single);
800EXPORT_SYMBOL(swiotlb_unmap_single);
801EXPORT_SYMBOL(swiotlb_map_sg);
802EXPORT_SYMBOL(swiotlb_unmap_sg);
803EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
804EXPORT_SYMBOL(swiotlb_sync_single_for_device);
John W. Linville878a97c2005-09-29 14:44:23 -0700805EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
806EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
808EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
809EXPORT_SYMBOL(swiotlb_dma_mapping_error);
810EXPORT_SYMBOL(swiotlb_alloc_coherent);
811EXPORT_SYMBOL(swiotlb_free_coherent);
812EXPORT_SYMBOL(swiotlb_dma_supported);