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>
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>
35 #include <asm/scatterlist.h>
37 #include <linux/init.h>
38 #include <linux/bootmem.h>
39 #include <linux/iommu-helper.h>
41 #define OFFSET(val,align) ((unsigned long) \
42 ( (val) & ( (align) - 1)))
44 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
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.
51 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
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
60 static phys_addr_t io_tlb_start, io_tlb_end;
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.
66 static unsigned long io_tlb_nslabs;
69 * When the IOMMU overflows we return a fallback buffer. This sets the size.
71 static unsigned long io_tlb_overflow = 32*1024;
73 static phys_addr_t io_tlb_overflow_buffer;
76 * This is a free list describing the number of free entries available from
79 static unsigned int *io_tlb_list;
80 static unsigned int io_tlb_index;
83 * We need to save away the original address corresponding to a mapped entry
84 * for the sync operations.
86 static phys_addr_t *io_tlb_orig_addr;
89 * Protect the above data structures in the map and unmap calls
91 static DEFINE_SPINLOCK(io_tlb_lock);
93 static int late_alloc;
96 setup_io_tlb_npages(char *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);
105 if (!strcmp(str, "force"))
110 __setup("swiotlb=", setup_io_tlb_npages);
111 /* make io_tlb_overflow tunable too? */
113 unsigned long swiotlb_nr_tbl(void)
115 return io_tlb_nslabs;
117 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
118 /* Note that this doesn't work with highmem page */
119 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
120 volatile void *address)
122 return phys_to_dma(hwdev, virt_to_phys(address));
125 void swiotlb_print_info(void)
127 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
128 unsigned char *vstart, *vend;
130 vstart = phys_to_virt(io_tlb_start);
131 vend = phys_to_virt(io_tlb_end);
133 printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
134 (unsigned long long)io_tlb_start,
135 (unsigned long long)io_tlb_end,
136 bytes >> 20, vstart, vend - 1);
139 void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
141 void *v_overflow_buffer;
142 unsigned long i, bytes;
144 bytes = nslabs << IO_TLB_SHIFT;
146 io_tlb_nslabs = nslabs;
147 io_tlb_start = __pa(tlb);
148 io_tlb_end = io_tlb_start + bytes;
151 * Get the overflow emergency buffer
153 v_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
154 if (!v_overflow_buffer)
155 panic("Cannot allocate SWIOTLB overflow buffer!\n");
157 io_tlb_overflow_buffer = __pa(v_overflow_buffer);
160 * Allocate and initialize the free list array. This array is used
161 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
162 * between io_tlb_start and io_tlb_end.
164 io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
165 for (i = 0; i < io_tlb_nslabs; i++)
166 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
168 io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
171 swiotlb_print_info();
175 * Statically reserve bounce buffer space and initialize bounce buffer data
176 * structures for the software IO TLB used to implement the DMA API.
179 swiotlb_init_with_default_size(size_t default_size, int verbose)
181 unsigned char *vstart;
184 if (!io_tlb_nslabs) {
185 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
186 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
189 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
192 * Get IO TLB memory from the low pages
194 vstart = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
196 panic("Cannot allocate SWIOTLB buffer");
198 swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose);
202 swiotlb_init(int verbose)
204 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
208 * Systems with larger DMA zones (those that don't support ISA) can
209 * initialize the swiotlb later using the slab allocator if needed.
210 * This should be just like above, but with some error catching.
213 swiotlb_late_init_with_default_size(size_t default_size)
215 unsigned long bytes, req_nslabs = io_tlb_nslabs;
216 unsigned char *vstart = NULL;
220 if (!io_tlb_nslabs) {
221 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
222 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
226 * Get IO TLB memory from the low pages
228 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
229 io_tlb_nslabs = SLABS_PER_PAGE << order;
230 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
232 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
233 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
241 io_tlb_nslabs = req_nslabs;
244 if (order != get_order(bytes)) {
245 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
246 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
247 io_tlb_nslabs = SLABS_PER_PAGE << order;
249 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
251 free_pages((unsigned long)vstart, order);
256 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
258 unsigned long i, bytes;
259 unsigned char *v_overflow_buffer;
261 bytes = nslabs << IO_TLB_SHIFT;
263 io_tlb_nslabs = nslabs;
264 io_tlb_start = virt_to_phys(tlb);
265 io_tlb_end = io_tlb_start + bytes;
267 memset(tlb, 0, bytes);
270 * Get the overflow emergency buffer
272 v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
273 get_order(io_tlb_overflow));
274 if (!v_overflow_buffer)
277 io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
280 * Allocate and initialize the free list array. This array is used
281 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
282 * between io_tlb_start and io_tlb_end.
284 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
285 get_order(io_tlb_nslabs * sizeof(int)));
289 for (i = 0; i < io_tlb_nslabs; i++)
290 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
293 io_tlb_orig_addr = (phys_addr_t *)
294 __get_free_pages(GFP_KERNEL,
295 get_order(io_tlb_nslabs *
296 sizeof(phys_addr_t)));
297 if (!io_tlb_orig_addr)
300 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
302 swiotlb_print_info();
309 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
313 free_pages((unsigned long)v_overflow_buffer,
314 get_order(io_tlb_overflow));
315 io_tlb_overflow_buffer = 0;
323 void __init swiotlb_free(void)
325 if (!io_tlb_orig_addr)
329 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
330 get_order(io_tlb_overflow));
331 free_pages((unsigned long)io_tlb_orig_addr,
332 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
333 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
335 free_pages((unsigned long)phys_to_virt(io_tlb_start),
336 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
338 free_bootmem_late(io_tlb_overflow_buffer,
339 PAGE_ALIGN(io_tlb_overflow));
340 free_bootmem_late(__pa(io_tlb_orig_addr),
341 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
342 free_bootmem_late(__pa(io_tlb_list),
343 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
344 free_bootmem_late(io_tlb_start,
345 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
350 static int is_swiotlb_buffer(phys_addr_t paddr)
352 return paddr >= io_tlb_start && paddr < io_tlb_end;
356 * Bounce: copy the swiotlb buffer back to the original dma location
358 void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
359 enum dma_data_direction dir)
361 unsigned long pfn = PFN_DOWN(phys);
363 if (PageHighMem(pfn_to_page(pfn))) {
364 /* The buffer does not have a mapping. Map it in and copy */
365 unsigned int offset = phys & ~PAGE_MASK;
371 sz = min_t(size_t, PAGE_SIZE - offset, size);
373 local_irq_save(flags);
374 buffer = kmap_atomic(pfn_to_page(pfn));
375 if (dir == DMA_TO_DEVICE)
376 memcpy(dma_addr, buffer + offset, sz);
378 memcpy(buffer + offset, dma_addr, sz);
379 kunmap_atomic(buffer);
380 local_irq_restore(flags);
388 if (dir == DMA_TO_DEVICE)
389 memcpy(dma_addr, phys_to_virt(phys), size);
391 memcpy(phys_to_virt(phys), dma_addr, size);
394 EXPORT_SYMBOL_GPL(swiotlb_bounce);
396 phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
397 dma_addr_t tbl_dma_addr,
398 phys_addr_t orig_addr, size_t size,
399 enum dma_data_direction dir)
402 phys_addr_t tlb_addr;
403 unsigned int nslots, stride, index, wrap;
406 unsigned long offset_slots;
407 unsigned long max_slots;
409 mask = dma_get_seg_boundary(hwdev);
411 tbl_dma_addr &= mask;
413 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
416 * Carefully handle integer overflow which can occur when mask == ~0UL.
419 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
420 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
423 * For mappings greater than a page, we limit the stride (and
424 * hence alignment) to a page size.
426 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
427 if (size > PAGE_SIZE)
428 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
435 * Find suitable number of IO TLB entries size that will fit this
436 * request and allocate a buffer from that IO TLB pool.
438 spin_lock_irqsave(&io_tlb_lock, flags);
439 index = ALIGN(io_tlb_index, stride);
440 if (index >= io_tlb_nslabs)
445 while (iommu_is_span_boundary(index, nslots, offset_slots,
448 if (index >= io_tlb_nslabs)
455 * If we find a slot that indicates we have 'nslots' number of
456 * contiguous buffers, we allocate the buffers from that slot
457 * and mark the entries as '0' indicating unavailable.
459 if (io_tlb_list[index] >= nslots) {
462 for (i = index; i < (int) (index + nslots); i++)
464 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
465 io_tlb_list[i] = ++count;
466 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
469 * Update the indices to avoid searching in the next
472 io_tlb_index = ((index + nslots) < io_tlb_nslabs
473 ? (index + nslots) : 0);
478 if (index >= io_tlb_nslabs)
480 } while (index != wrap);
483 spin_unlock_irqrestore(&io_tlb_lock, flags);
484 return SWIOTLB_MAP_ERROR;
486 spin_unlock_irqrestore(&io_tlb_lock, flags);
489 * Save away the mapping from the original address to the DMA address.
490 * This is needed when we sync the memory. Then we sync the buffer if
493 for (i = 0; i < nslots; i++)
494 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
495 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
496 swiotlb_bounce(orig_addr, phys_to_virt(tlb_addr), size,
501 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
504 * Allocates bounce buffer and returns its kernel virtual address.
507 phys_addr_t map_single(struct device *hwdev, phys_addr_t phys, size_t size,
508 enum dma_data_direction dir)
510 dma_addr_t start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
512 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
516 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
518 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
519 size_t size, enum dma_data_direction dir)
522 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
523 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
524 phys_addr_t orig_addr = io_tlb_orig_addr[index];
527 * First, sync the memory before unmapping the entry
529 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
530 swiotlb_bounce(orig_addr, phys_to_virt(tlb_addr),
531 size, DMA_FROM_DEVICE);
534 * Return the buffer to the free list by setting the corresponding
535 * entries to indicate the number of contiguous entries available.
536 * While returning the entries to the free list, we merge the entries
537 * with slots below and above the pool being returned.
539 spin_lock_irqsave(&io_tlb_lock, flags);
541 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
542 io_tlb_list[index + nslots] : 0);
544 * Step 1: return the slots to the free list, merging the
545 * slots with superceeding slots
547 for (i = index + nslots - 1; i >= index; i--)
548 io_tlb_list[i] = ++count;
550 * Step 2: merge the returned slots with the preceding slots,
551 * if available (non zero)
553 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
554 io_tlb_list[i] = ++count;
556 spin_unlock_irqrestore(&io_tlb_lock, flags);
558 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
561 swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
562 enum dma_data_direction dir,
563 enum dma_sync_target target)
565 int index = (dma_addr - (char *)phys_to_virt(io_tlb_start)) >> IO_TLB_SHIFT;
566 phys_addr_t phys = io_tlb_orig_addr[index];
568 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
572 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
573 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
575 BUG_ON(dir != DMA_TO_DEVICE);
577 case SYNC_FOR_DEVICE:
578 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
579 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
581 BUG_ON(dir != DMA_FROM_DEVICE);
587 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
590 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
591 dma_addr_t *dma_handle, gfp_t flags)
595 int order = get_order(size);
596 u64 dma_mask = DMA_BIT_MASK(32);
598 if (hwdev && hwdev->coherent_dma_mask)
599 dma_mask = hwdev->coherent_dma_mask;
601 ret = (void *)__get_free_pages(flags, order);
603 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
604 if (dev_addr + size - 1 > dma_mask) {
606 * The allocated memory isn't reachable by the device.
608 free_pages((unsigned long) ret, order);
614 * We are either out of memory or the device can't DMA to
615 * GFP_DMA memory; fall back on map_single(), which
616 * will grab memory from the lowest available address range.
618 phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
619 if (paddr == SWIOTLB_MAP_ERROR)
622 ret = phys_to_virt(paddr);
623 dev_addr = phys_to_dma(hwdev, paddr);
625 /* Confirm address can be DMA'd by device */
626 if (dev_addr + size - 1 > dma_mask) {
627 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
628 (unsigned long long)dma_mask,
629 (unsigned long long)dev_addr);
631 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
632 swiotlb_tbl_unmap_single(hwdev, paddr,
633 size, DMA_TO_DEVICE);
638 *dma_handle = dev_addr;
639 memset(ret, 0, size);
643 EXPORT_SYMBOL(swiotlb_alloc_coherent);
646 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
649 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
651 WARN_ON(irqs_disabled());
652 if (!is_swiotlb_buffer(paddr))
653 free_pages((unsigned long)vaddr, get_order(size));
655 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
656 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
658 EXPORT_SYMBOL(swiotlb_free_coherent);
661 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
665 * Ran out of IOMMU space for this operation. This is very bad.
666 * Unfortunately the drivers cannot handle this operation properly.
667 * unless they check for dma_mapping_error (most don't)
668 * When the mapping is small enough return a static buffer to limit
669 * the damage, or panic when the transfer is too big.
671 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
672 "device %s\n", size, dev ? dev_name(dev) : "?");
674 if (size <= io_tlb_overflow || !do_panic)
677 if (dir == DMA_BIDIRECTIONAL)
678 panic("DMA: Random memory could be DMA accessed\n");
679 if (dir == DMA_FROM_DEVICE)
680 panic("DMA: Random memory could be DMA written\n");
681 if (dir == DMA_TO_DEVICE)
682 panic("DMA: Random memory could be DMA read\n");
686 * Map a single buffer of the indicated size for DMA in streaming mode. The
687 * physical address to use is returned.
689 * Once the device is given the dma address, the device owns this memory until
690 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
692 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
693 unsigned long offset, size_t size,
694 enum dma_data_direction dir,
695 struct dma_attrs *attrs)
697 phys_addr_t map, phys = page_to_phys(page) + offset;
698 dma_addr_t dev_addr = phys_to_dma(dev, phys);
700 BUG_ON(dir == DMA_NONE);
702 * If the address happens to be in the device's DMA window,
703 * we can safely return the device addr and not worry about bounce
706 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
709 /* Oh well, have to allocate and map a bounce buffer. */
710 map = map_single(dev, phys, size, dir);
711 if (map == SWIOTLB_MAP_ERROR) {
712 swiotlb_full(dev, size, dir, 1);
713 return phys_to_dma(dev, io_tlb_overflow_buffer);
716 dev_addr = phys_to_dma(dev, map);
718 /* Ensure that the address returned is DMA'ble */
719 if (!dma_capable(dev, dev_addr, size)) {
720 swiotlb_tbl_unmap_single(dev, map, size, dir);
721 return phys_to_dma(dev, io_tlb_overflow_buffer);
726 EXPORT_SYMBOL_GPL(swiotlb_map_page);
729 * Unmap a single streaming mode DMA translation. The dma_addr and size must
730 * match what was provided for in a previous swiotlb_map_page call. All
731 * other usages are undefined.
733 * After this call, reads by the cpu to the buffer are guaranteed to see
734 * whatever the device wrote there.
736 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
737 size_t size, enum dma_data_direction dir)
739 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
741 BUG_ON(dir == DMA_NONE);
743 if (is_swiotlb_buffer(paddr)) {
744 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
748 if (dir != DMA_FROM_DEVICE)
752 * phys_to_virt doesn't work with hihgmem page but we could
753 * call dma_mark_clean() with hihgmem page here. However, we
754 * are fine since dma_mark_clean() is null on POWERPC. We can
755 * make dma_mark_clean() take a physical address if necessary.
757 dma_mark_clean(phys_to_virt(paddr), size);
760 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
761 size_t size, enum dma_data_direction dir,
762 struct dma_attrs *attrs)
764 unmap_single(hwdev, dev_addr, size, dir);
766 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
769 * Make physical memory consistent for a single streaming mode DMA translation
772 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
773 * using the cpu, yet do not wish to teardown the dma mapping, you must
774 * call this function before doing so. At the next point you give the dma
775 * address back to the card, you must first perform a
776 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
779 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
780 size_t size, enum dma_data_direction dir,
781 enum dma_sync_target target)
783 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
785 BUG_ON(dir == DMA_NONE);
787 if (is_swiotlb_buffer(paddr)) {
788 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
793 if (dir != DMA_FROM_DEVICE)
796 dma_mark_clean(phys_to_virt(paddr), size);
800 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
801 size_t size, enum dma_data_direction dir)
803 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
805 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
808 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
809 size_t size, enum dma_data_direction dir)
811 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
813 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
816 * Map a set of buffers described by scatterlist in streaming mode for DMA.
817 * This is the scatter-gather version of the above swiotlb_map_page
818 * interface. Here the scatter gather list elements are each tagged with the
819 * appropriate dma address and length. They are obtained via
820 * sg_dma_{address,length}(SG).
822 * NOTE: An implementation may be able to use a smaller number of
823 * DMA address/length pairs than there are SG table elements.
824 * (for example via virtual mapping capabilities)
825 * The routine returns the number of addr/length pairs actually
826 * used, at most nents.
828 * Device ownership issues as mentioned above for swiotlb_map_page are the
832 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
833 enum dma_data_direction dir, struct dma_attrs *attrs)
835 struct scatterlist *sg;
838 BUG_ON(dir == DMA_NONE);
840 for_each_sg(sgl, sg, nelems, i) {
841 phys_addr_t paddr = sg_phys(sg);
842 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
845 !dma_capable(hwdev, dev_addr, sg->length)) {
846 phys_addr_t map = map_single(hwdev, sg_phys(sg),
848 if (map == SWIOTLB_MAP_ERROR) {
849 /* Don't panic here, we expect map_sg users
850 to do proper error handling. */
851 swiotlb_full(hwdev, sg->length, dir, 0);
852 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
854 sgl[0].dma_length = 0;
857 sg->dma_address = phys_to_dma(hwdev, map);
859 sg->dma_address = dev_addr;
860 sg->dma_length = sg->length;
864 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
867 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
868 enum dma_data_direction dir)
870 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
872 EXPORT_SYMBOL(swiotlb_map_sg);
875 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
876 * concerning calls here are the same as for swiotlb_unmap_page() above.
879 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
880 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
882 struct scatterlist *sg;
885 BUG_ON(dir == DMA_NONE);
887 for_each_sg(sgl, sg, nelems, i)
888 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
891 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
894 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
895 enum dma_data_direction dir)
897 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
899 EXPORT_SYMBOL(swiotlb_unmap_sg);
902 * Make physical memory consistent for a set of streaming mode DMA translations
905 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
909 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
910 int nelems, enum dma_data_direction dir,
911 enum dma_sync_target target)
913 struct scatterlist *sg;
916 for_each_sg(sgl, sg, nelems, i)
917 swiotlb_sync_single(hwdev, sg->dma_address,
918 sg->dma_length, dir, target);
922 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
923 int nelems, enum dma_data_direction dir)
925 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
927 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
930 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
931 int nelems, enum dma_data_direction dir)
933 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
935 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
938 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
940 return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
942 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
945 * Return whether the given device DMA address mask can be supported
946 * properly. For example, if your device can only drive the low 24-bits
947 * during bus mastering, then you would pass 0x00ffffff as the mask to
951 swiotlb_dma_supported(struct device *hwdev, u64 mask)
953 return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
955 EXPORT_SYMBOL(swiotlb_dma_supported);