Linux 2.6.33-rc6
[cris-mirror.git] / lib / swiotlb.c
blob437eedb5a53ba1feae04f98b5def86798462f920
1 /*
2 * Dynamic DMA mapping support.
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>
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
20 #include <linux/cache.h>
21 #include <linux/dma-mapping.h>
22 #include <linux/mm.h>
23 #include <linux/module.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>
32 #include <asm/io.h>
33 #include <asm/dma.h>
34 #include <asm/scatterlist.h>
36 #include <linux/init.h>
37 #include <linux/bootmem.h>
38 #include <linux/iommu-helper.h>
40 #define OFFSET(val,align) ((unsigned long) \
41 ( (val) & ( (align) - 1)))
43 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
46 * Minimum IO TLB size to bother booting with. Systems with mainly
47 * 64bit capable cards will only lightly use the swiotlb. If we can't
48 * allocate a contiguous 1MB, we're probably in trouble anyway.
50 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
53 * Enumeration for sync targets
55 enum dma_sync_target {
56 SYNC_FOR_CPU = 0,
57 SYNC_FOR_DEVICE = 1,
60 int swiotlb_force;
63 * Used to do a quick range check in unmap_single and
64 * sync_single_*, to see if the memory was in fact allocated by this
65 * API.
67 static char *io_tlb_start, *io_tlb_end;
70 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
71 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
73 static unsigned long io_tlb_nslabs;
76 * When the IOMMU overflows we return a fallback buffer. This sets the size.
78 static unsigned long io_tlb_overflow = 32*1024;
80 void *io_tlb_overflow_buffer;
83 * This is a free list describing the number of free entries available from
84 * each index
86 static unsigned int *io_tlb_list;
87 static unsigned int io_tlb_index;
90 * We need to save away the original address corresponding to a mapped entry
91 * for the sync operations.
93 static phys_addr_t *io_tlb_orig_addr;
96 * Protect the above data structures in the map and unmap calls
98 static DEFINE_SPINLOCK(io_tlb_lock);
100 static int late_alloc;
102 static int __init
103 setup_io_tlb_npages(char *str)
105 if (isdigit(*str)) {
106 io_tlb_nslabs = simple_strtoul(str, &str, 0);
107 /* avoid tail segment of size < IO_TLB_SEGSIZE */
108 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
110 if (*str == ',')
111 ++str;
112 if (!strcmp(str, "force"))
113 swiotlb_force = 1;
115 return 1;
117 __setup("swiotlb=", setup_io_tlb_npages);
118 /* make io_tlb_overflow tunable too? */
120 /* Note that this doesn't work with highmem page */
121 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
122 volatile void *address)
124 return phys_to_dma(hwdev, virt_to_phys(address));
127 void swiotlb_print_info(void)
129 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
130 phys_addr_t pstart, pend;
132 pstart = virt_to_phys(io_tlb_start);
133 pend = virt_to_phys(io_tlb_end);
135 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
136 bytes >> 20, io_tlb_start, io_tlb_end);
137 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
138 (unsigned long long)pstart,
139 (unsigned long long)pend);
143 * Statically reserve bounce buffer space and initialize bounce buffer data
144 * structures for the software IO TLB used to implement the DMA API.
146 void __init
147 swiotlb_init_with_default_size(size_t default_size, int verbose)
149 unsigned long i, bytes;
151 if (!io_tlb_nslabs) {
152 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
153 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
156 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
159 * Get IO TLB memory from the low pages
161 io_tlb_start = alloc_bootmem_low_pages(bytes);
162 if (!io_tlb_start)
163 panic("Cannot allocate SWIOTLB buffer");
164 io_tlb_end = io_tlb_start + bytes;
167 * Allocate and initialize the free list array. This array is used
168 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
169 * between io_tlb_start and io_tlb_end.
171 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
172 for (i = 0; i < io_tlb_nslabs; i++)
173 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
174 io_tlb_index = 0;
175 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
178 * Get the overflow emergency buffer
180 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
181 if (!io_tlb_overflow_buffer)
182 panic("Cannot allocate SWIOTLB overflow buffer!\n");
183 if (verbose)
184 swiotlb_print_info();
187 void __init
188 swiotlb_init(int verbose)
190 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
194 * Systems with larger DMA zones (those that don't support ISA) can
195 * initialize the swiotlb later using the slab allocator if needed.
196 * This should be just like above, but with some error catching.
199 swiotlb_late_init_with_default_size(size_t default_size)
201 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
202 unsigned int order;
204 if (!io_tlb_nslabs) {
205 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
206 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
210 * Get IO TLB memory from the low pages
212 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
213 io_tlb_nslabs = SLABS_PER_PAGE << order;
214 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
216 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
217 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
218 order);
219 if (io_tlb_start)
220 break;
221 order--;
224 if (!io_tlb_start)
225 goto cleanup1;
227 if (order != get_order(bytes)) {
228 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
229 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
230 io_tlb_nslabs = SLABS_PER_PAGE << order;
231 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
233 io_tlb_end = io_tlb_start + bytes;
234 memset(io_tlb_start, 0, bytes);
237 * Allocate and initialize the free list array. This array is used
238 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
239 * between io_tlb_start and io_tlb_end.
241 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
242 get_order(io_tlb_nslabs * sizeof(int)));
243 if (!io_tlb_list)
244 goto cleanup2;
246 for (i = 0; i < io_tlb_nslabs; i++)
247 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
248 io_tlb_index = 0;
250 io_tlb_orig_addr = (phys_addr_t *)
251 __get_free_pages(GFP_KERNEL,
252 get_order(io_tlb_nslabs *
253 sizeof(phys_addr_t)));
254 if (!io_tlb_orig_addr)
255 goto cleanup3;
257 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
260 * Get the overflow emergency buffer
262 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
263 get_order(io_tlb_overflow));
264 if (!io_tlb_overflow_buffer)
265 goto cleanup4;
267 swiotlb_print_info();
269 late_alloc = 1;
271 return 0;
273 cleanup4:
274 free_pages((unsigned long)io_tlb_orig_addr,
275 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
276 io_tlb_orig_addr = NULL;
277 cleanup3:
278 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
279 sizeof(int)));
280 io_tlb_list = NULL;
281 cleanup2:
282 io_tlb_end = NULL;
283 free_pages((unsigned long)io_tlb_start, order);
284 io_tlb_start = NULL;
285 cleanup1:
286 io_tlb_nslabs = req_nslabs;
287 return -ENOMEM;
290 void __init swiotlb_free(void)
292 if (!io_tlb_overflow_buffer)
293 return;
295 if (late_alloc) {
296 free_pages((unsigned long)io_tlb_overflow_buffer,
297 get_order(io_tlb_overflow));
298 free_pages((unsigned long)io_tlb_orig_addr,
299 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
300 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
301 sizeof(int)));
302 free_pages((unsigned long)io_tlb_start,
303 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
304 } else {
305 free_bootmem_late(__pa(io_tlb_overflow_buffer),
306 io_tlb_overflow);
307 free_bootmem_late(__pa(io_tlb_orig_addr),
308 io_tlb_nslabs * sizeof(phys_addr_t));
309 free_bootmem_late(__pa(io_tlb_list),
310 io_tlb_nslabs * sizeof(int));
311 free_bootmem_late(__pa(io_tlb_start),
312 io_tlb_nslabs << IO_TLB_SHIFT);
316 static int is_swiotlb_buffer(phys_addr_t paddr)
318 return paddr >= virt_to_phys(io_tlb_start) &&
319 paddr < virt_to_phys(io_tlb_end);
323 * Bounce: copy the swiotlb buffer back to the original dma location
325 static void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
326 enum dma_data_direction dir)
328 unsigned long pfn = PFN_DOWN(phys);
330 if (PageHighMem(pfn_to_page(pfn))) {
331 /* The buffer does not have a mapping. Map it in and copy */
332 unsigned int offset = phys & ~PAGE_MASK;
333 char *buffer;
334 unsigned int sz = 0;
335 unsigned long flags;
337 while (size) {
338 sz = min_t(size_t, PAGE_SIZE - offset, size);
340 local_irq_save(flags);
341 buffer = kmap_atomic(pfn_to_page(pfn),
342 KM_BOUNCE_READ);
343 if (dir == DMA_TO_DEVICE)
344 memcpy(dma_addr, buffer + offset, sz);
345 else
346 memcpy(buffer + offset, dma_addr, sz);
347 kunmap_atomic(buffer, KM_BOUNCE_READ);
348 local_irq_restore(flags);
350 size -= sz;
351 pfn++;
352 dma_addr += sz;
353 offset = 0;
355 } else {
356 if (dir == DMA_TO_DEVICE)
357 memcpy(dma_addr, phys_to_virt(phys), size);
358 else
359 memcpy(phys_to_virt(phys), dma_addr, size);
364 * Allocates bounce buffer and returns its kernel virtual address.
366 static void *
367 map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
369 unsigned long flags;
370 char *dma_addr;
371 unsigned int nslots, stride, index, wrap;
372 int i;
373 unsigned long start_dma_addr;
374 unsigned long mask;
375 unsigned long offset_slots;
376 unsigned long max_slots;
378 mask = dma_get_seg_boundary(hwdev);
379 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
381 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
384 * Carefully handle integer overflow which can occur when mask == ~0UL.
386 max_slots = mask + 1
387 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
388 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
391 * For mappings greater than a page, we limit the stride (and
392 * hence alignment) to a page size.
394 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
395 if (size > PAGE_SIZE)
396 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
397 else
398 stride = 1;
400 BUG_ON(!nslots);
403 * Find suitable number of IO TLB entries size that will fit this
404 * request and allocate a buffer from that IO TLB pool.
406 spin_lock_irqsave(&io_tlb_lock, flags);
407 index = ALIGN(io_tlb_index, stride);
408 if (index >= io_tlb_nslabs)
409 index = 0;
410 wrap = index;
412 do {
413 while (iommu_is_span_boundary(index, nslots, offset_slots,
414 max_slots)) {
415 index += stride;
416 if (index >= io_tlb_nslabs)
417 index = 0;
418 if (index == wrap)
419 goto not_found;
423 * If we find a slot that indicates we have 'nslots' number of
424 * contiguous buffers, we allocate the buffers from that slot
425 * and mark the entries as '0' indicating unavailable.
427 if (io_tlb_list[index] >= nslots) {
428 int count = 0;
430 for (i = index; i < (int) (index + nslots); i++)
431 io_tlb_list[i] = 0;
432 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
433 io_tlb_list[i] = ++count;
434 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
437 * Update the indices to avoid searching in the next
438 * round.
440 io_tlb_index = ((index + nslots) < io_tlb_nslabs
441 ? (index + nslots) : 0);
443 goto found;
445 index += stride;
446 if (index >= io_tlb_nslabs)
447 index = 0;
448 } while (index != wrap);
450 not_found:
451 spin_unlock_irqrestore(&io_tlb_lock, flags);
452 return NULL;
453 found:
454 spin_unlock_irqrestore(&io_tlb_lock, flags);
457 * Save away the mapping from the original address to the DMA address.
458 * This is needed when we sync the memory. Then we sync the buffer if
459 * needed.
461 for (i = 0; i < nslots; i++)
462 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
463 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
464 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
466 return dma_addr;
470 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
472 static void
473 do_unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
475 unsigned long flags;
476 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
477 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
478 phys_addr_t phys = io_tlb_orig_addr[index];
481 * First, sync the memory before unmapping the entry
483 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
484 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
487 * Return the buffer to the free list by setting the corresponding
488 * entries to indicate the number of contiguous entries available.
489 * While returning the entries to the free list, we merge the entries
490 * with slots below and above the pool being returned.
492 spin_lock_irqsave(&io_tlb_lock, flags);
494 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
495 io_tlb_list[index + nslots] : 0);
497 * Step 1: return the slots to the free list, merging the
498 * slots with superceeding slots
500 for (i = index + nslots - 1; i >= index; i--)
501 io_tlb_list[i] = ++count;
503 * Step 2: merge the returned slots with the preceding slots,
504 * if available (non zero)
506 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
507 io_tlb_list[i] = ++count;
509 spin_unlock_irqrestore(&io_tlb_lock, flags);
512 static void
513 sync_single(struct device *hwdev, char *dma_addr, size_t size,
514 int dir, int target)
516 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
517 phys_addr_t phys = io_tlb_orig_addr[index];
519 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
521 switch (target) {
522 case SYNC_FOR_CPU:
523 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
524 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
525 else
526 BUG_ON(dir != DMA_TO_DEVICE);
527 break;
528 case SYNC_FOR_DEVICE:
529 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
530 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
531 else
532 BUG_ON(dir != DMA_FROM_DEVICE);
533 break;
534 default:
535 BUG();
539 void *
540 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
541 dma_addr_t *dma_handle, gfp_t flags)
543 dma_addr_t dev_addr;
544 void *ret;
545 int order = get_order(size);
546 u64 dma_mask = DMA_BIT_MASK(32);
548 if (hwdev && hwdev->coherent_dma_mask)
549 dma_mask = hwdev->coherent_dma_mask;
551 ret = (void *)__get_free_pages(flags, order);
552 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
554 * The allocated memory isn't reachable by the device.
556 free_pages((unsigned long) ret, order);
557 ret = NULL;
559 if (!ret) {
561 * We are either out of memory or the device can't DMA
562 * to GFP_DMA memory; fall back on map_single(), which
563 * will grab memory from the lowest available address range.
565 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
566 if (!ret)
567 return NULL;
570 memset(ret, 0, size);
571 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
573 /* Confirm address can be DMA'd by device */
574 if (dev_addr + size - 1 > dma_mask) {
575 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
576 (unsigned long long)dma_mask,
577 (unsigned long long)dev_addr);
579 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
580 do_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
581 return NULL;
583 *dma_handle = dev_addr;
584 return ret;
586 EXPORT_SYMBOL(swiotlb_alloc_coherent);
588 void
589 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
590 dma_addr_t dev_addr)
592 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
594 WARN_ON(irqs_disabled());
595 if (!is_swiotlb_buffer(paddr))
596 free_pages((unsigned long)vaddr, get_order(size));
597 else
598 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
599 do_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
601 EXPORT_SYMBOL(swiotlb_free_coherent);
603 static void
604 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
607 * Ran out of IOMMU space for this operation. This is very bad.
608 * Unfortunately the drivers cannot handle this operation properly.
609 * unless they check for dma_mapping_error (most don't)
610 * When the mapping is small enough return a static buffer to limit
611 * the damage, or panic when the transfer is too big.
613 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
614 "device %s\n", size, dev ? dev_name(dev) : "?");
616 if (size <= io_tlb_overflow || !do_panic)
617 return;
619 if (dir == DMA_BIDIRECTIONAL)
620 panic("DMA: Random memory could be DMA accessed\n");
621 if (dir == DMA_FROM_DEVICE)
622 panic("DMA: Random memory could be DMA written\n");
623 if (dir == DMA_TO_DEVICE)
624 panic("DMA: Random memory could be DMA read\n");
628 * Map a single buffer of the indicated size for DMA in streaming mode. The
629 * physical address to use is returned.
631 * Once the device is given the dma address, the device owns this memory until
632 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
634 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
635 unsigned long offset, size_t size,
636 enum dma_data_direction dir,
637 struct dma_attrs *attrs)
639 phys_addr_t phys = page_to_phys(page) + offset;
640 dma_addr_t dev_addr = phys_to_dma(dev, phys);
641 void *map;
643 BUG_ON(dir == DMA_NONE);
645 * If the address happens to be in the device's DMA window,
646 * we can safely return the device addr and not worry about bounce
647 * buffering it.
649 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
650 return dev_addr;
653 * Oh well, have to allocate and map a bounce buffer.
655 map = map_single(dev, phys, size, dir);
656 if (!map) {
657 swiotlb_full(dev, size, dir, 1);
658 map = io_tlb_overflow_buffer;
661 dev_addr = swiotlb_virt_to_bus(dev, map);
664 * Ensure that the address returned is DMA'ble
666 if (!dma_capable(dev, dev_addr, size))
667 panic("map_single: bounce buffer is not DMA'ble");
669 return dev_addr;
671 EXPORT_SYMBOL_GPL(swiotlb_map_page);
674 * Unmap a single streaming mode DMA translation. The dma_addr and size must
675 * match what was provided for in a previous swiotlb_map_page call. All
676 * other usages are undefined.
678 * After this call, reads by the cpu to the buffer are guaranteed to see
679 * whatever the device wrote there.
681 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
682 size_t size, int dir)
684 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
686 BUG_ON(dir == DMA_NONE);
688 if (is_swiotlb_buffer(paddr)) {
689 do_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
690 return;
693 if (dir != DMA_FROM_DEVICE)
694 return;
697 * phys_to_virt doesn't work with hihgmem page but we could
698 * call dma_mark_clean() with hihgmem page here. However, we
699 * are fine since dma_mark_clean() is null on POWERPC. We can
700 * make dma_mark_clean() take a physical address if necessary.
702 dma_mark_clean(phys_to_virt(paddr), size);
705 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
706 size_t size, enum dma_data_direction dir,
707 struct dma_attrs *attrs)
709 unmap_single(hwdev, dev_addr, size, dir);
711 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
714 * Make physical memory consistent for a single streaming mode DMA translation
715 * after a transfer.
717 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
718 * using the cpu, yet do not wish to teardown the dma mapping, you must
719 * call this function before doing so. At the next point you give the dma
720 * address back to the card, you must first perform a
721 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
723 static void
724 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
725 size_t size, int dir, int target)
727 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
729 BUG_ON(dir == DMA_NONE);
731 if (is_swiotlb_buffer(paddr)) {
732 sync_single(hwdev, phys_to_virt(paddr), size, dir, target);
733 return;
736 if (dir != DMA_FROM_DEVICE)
737 return;
739 dma_mark_clean(phys_to_virt(paddr), size);
742 void
743 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
744 size_t size, enum dma_data_direction dir)
746 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
748 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
750 void
751 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
752 size_t size, enum dma_data_direction dir)
754 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
756 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
759 * Same as above, but for a sub-range of the mapping.
761 static void
762 swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
763 unsigned long offset, size_t size,
764 int dir, int target)
766 swiotlb_sync_single(hwdev, dev_addr + offset, size, dir, target);
769 void
770 swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
771 unsigned long offset, size_t size,
772 enum dma_data_direction dir)
774 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
775 SYNC_FOR_CPU);
777 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
779 void
780 swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
781 unsigned long offset, size_t size,
782 enum dma_data_direction dir)
784 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
785 SYNC_FOR_DEVICE);
787 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
790 * Map a set of buffers described by scatterlist in streaming mode for DMA.
791 * This is the scatter-gather version of the above swiotlb_map_page
792 * interface. Here the scatter gather list elements are each tagged with the
793 * appropriate dma address and length. They are obtained via
794 * sg_dma_{address,length}(SG).
796 * NOTE: An implementation may be able to use a smaller number of
797 * DMA address/length pairs than there are SG table elements.
798 * (for example via virtual mapping capabilities)
799 * The routine returns the number of addr/length pairs actually
800 * used, at most nents.
802 * Device ownership issues as mentioned above for swiotlb_map_page are the
803 * same here.
806 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
807 enum dma_data_direction dir, struct dma_attrs *attrs)
809 struct scatterlist *sg;
810 int i;
812 BUG_ON(dir == DMA_NONE);
814 for_each_sg(sgl, sg, nelems, i) {
815 phys_addr_t paddr = sg_phys(sg);
816 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
818 if (swiotlb_force ||
819 !dma_capable(hwdev, dev_addr, sg->length)) {
820 void *map = map_single(hwdev, sg_phys(sg),
821 sg->length, dir);
822 if (!map) {
823 /* Don't panic here, we expect map_sg users
824 to do proper error handling. */
825 swiotlb_full(hwdev, sg->length, dir, 0);
826 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
827 attrs);
828 sgl[0].dma_length = 0;
829 return 0;
831 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
832 } else
833 sg->dma_address = dev_addr;
834 sg->dma_length = sg->length;
836 return nelems;
838 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
841 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
842 int dir)
844 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
846 EXPORT_SYMBOL(swiotlb_map_sg);
849 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
850 * concerning calls here are the same as for swiotlb_unmap_page() above.
852 void
853 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
854 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
856 struct scatterlist *sg;
857 int i;
859 BUG_ON(dir == DMA_NONE);
861 for_each_sg(sgl, sg, nelems, i)
862 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
865 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
867 void
868 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
869 int dir)
871 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
873 EXPORT_SYMBOL(swiotlb_unmap_sg);
876 * Make physical memory consistent for a set of streaming mode DMA translations
877 * after a transfer.
879 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
880 * and usage.
882 static void
883 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
884 int nelems, int dir, int target)
886 struct scatterlist *sg;
887 int i;
889 for_each_sg(sgl, sg, nelems, i)
890 swiotlb_sync_single(hwdev, sg->dma_address,
891 sg->dma_length, dir, target);
894 void
895 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
896 int nelems, enum dma_data_direction dir)
898 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
900 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
902 void
903 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
904 int nelems, enum dma_data_direction dir)
906 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
908 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
911 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
913 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
915 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
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.
924 swiotlb_dma_supported(struct device *hwdev, u64 mask)
926 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
928 EXPORT_SYMBOL(swiotlb_dma_supported);