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/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>
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 {
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
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
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);
101 setup_io_tlb_npages(char *str)
104 io_tlb_nslabs = simple_strtoul(str, &str, 0);
105 /* avoid tail segment of size < IO_TLB_SEGSIZE */
106 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
110 if (!strcmp(str, "force"))
114 __setup("swiotlb=", setup_io_tlb_npages);
115 /* make io_tlb_overflow tunable too? */
117 dma_addr_t __weak swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
122 phys_addr_t __weak swiotlb_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
127 /* Note that this doesn't work with highmem page */
128 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
129 volatile void *address)
131 return swiotlb_phys_to_bus(hwdev, virt_to_phys(address));
134 static void swiotlb_print_info(unsigned long bytes)
136 phys_addr_t pstart, pend;
138 pstart = virt_to_phys(io_tlb_start);
139 pend = virt_to_phys(io_tlb_end);
141 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
142 bytes >> 20, io_tlb_start, io_tlb_end);
143 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
144 (unsigned long long)pstart,
145 (unsigned long long)pend);
149 * Statically reserve bounce buffer space and initialize bounce buffer data
150 * structures for the software IO TLB used to implement the DMA API.
153 swiotlb_init_with_default_size(size_t default_size)
155 unsigned long i, bytes;
157 if (!io_tlb_nslabs) {
158 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
159 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
162 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
165 * Get IO TLB memory from the low pages
167 io_tlb_start = alloc_bootmem_low_pages(bytes);
169 panic("Cannot allocate SWIOTLB buffer");
170 io_tlb_end = io_tlb_start + bytes;
173 * Allocate and initialize the free list array. This array is used
174 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
175 * between io_tlb_start and io_tlb_end.
177 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
178 for (i = 0; i < io_tlb_nslabs; i++)
179 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
181 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
184 * Get the overflow emergency buffer
186 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
187 if (!io_tlb_overflow_buffer)
188 panic("Cannot allocate SWIOTLB overflow buffer!\n");
190 swiotlb_print_info(bytes);
196 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
200 * Systems with larger DMA zones (those that don't support ISA) can
201 * initialize the swiotlb later using the slab allocator if needed.
202 * This should be just like above, but with some error catching.
205 swiotlb_late_init_with_default_size(size_t default_size)
207 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
210 if (!io_tlb_nslabs) {
211 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
212 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
216 * Get IO TLB memory from the low pages
218 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
219 io_tlb_nslabs = SLABS_PER_PAGE << order;
220 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
222 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
223 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
233 if (order != get_order(bytes)) {
234 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
235 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
236 io_tlb_nslabs = SLABS_PER_PAGE << order;
237 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
239 io_tlb_end = io_tlb_start + bytes;
240 memset(io_tlb_start, 0, bytes);
243 * Allocate and initialize the free list array. This array is used
244 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
245 * between io_tlb_start and io_tlb_end.
247 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
248 get_order(io_tlb_nslabs * sizeof(int)));
252 for (i = 0; i < io_tlb_nslabs; i++)
253 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
256 io_tlb_orig_addr = (phys_addr_t *)
257 __get_free_pages(GFP_KERNEL,
258 get_order(io_tlb_nslabs *
259 sizeof(phys_addr_t)));
260 if (!io_tlb_orig_addr)
263 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
266 * Get the overflow emergency buffer
268 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
269 get_order(io_tlb_overflow));
270 if (!io_tlb_overflow_buffer)
273 swiotlb_print_info(bytes);
278 free_pages((unsigned long)io_tlb_orig_addr,
279 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
280 io_tlb_orig_addr = NULL;
282 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
287 free_pages((unsigned long)io_tlb_start, order);
290 io_tlb_nslabs = req_nslabs;
294 static int is_swiotlb_buffer(phys_addr_t paddr)
296 return paddr >= virt_to_phys(io_tlb_start) &&
297 paddr < virt_to_phys(io_tlb_end);
301 * Bounce: copy the swiotlb buffer back to the original dma location
303 static void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
304 enum dma_data_direction dir)
306 unsigned long pfn = PFN_DOWN(phys);
308 if (PageHighMem(pfn_to_page(pfn))) {
309 /* The buffer does not have a mapping. Map it in and copy */
310 unsigned int offset = phys & ~PAGE_MASK;
316 sz = min_t(size_t, PAGE_SIZE - offset, size);
318 local_irq_save(flags);
319 buffer = kmap_atomic(pfn_to_page(pfn),
321 if (dir == DMA_TO_DEVICE)
322 memcpy(dma_addr, buffer + offset, sz);
324 memcpy(buffer + offset, dma_addr, sz);
325 kunmap_atomic(buffer, KM_BOUNCE_READ);
326 local_irq_restore(flags);
334 if (dir == DMA_TO_DEVICE)
335 memcpy(dma_addr, phys_to_virt(phys), size);
337 memcpy(phys_to_virt(phys), dma_addr, size);
342 * Allocates bounce buffer and returns its kernel virtual address.
345 map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
349 unsigned int nslots, stride, index, wrap;
351 unsigned long start_dma_addr;
353 unsigned long offset_slots;
354 unsigned long max_slots;
356 mask = dma_get_seg_boundary(hwdev);
357 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
359 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
362 * Carefully handle integer overflow which can occur when mask == ~0UL.
365 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
366 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
369 * For mappings greater than a page, we limit the stride (and
370 * hence alignment) to a page size.
372 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
373 if (size > PAGE_SIZE)
374 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
381 * Find suitable number of IO TLB entries size that will fit this
382 * request and allocate a buffer from that IO TLB pool.
384 spin_lock_irqsave(&io_tlb_lock, flags);
385 index = ALIGN(io_tlb_index, stride);
386 if (index >= io_tlb_nslabs)
391 while (iommu_is_span_boundary(index, nslots, offset_slots,
394 if (index >= io_tlb_nslabs)
401 * If we find a slot that indicates we have 'nslots' number of
402 * contiguous buffers, we allocate the buffers from that slot
403 * and mark the entries as '0' indicating unavailable.
405 if (io_tlb_list[index] >= nslots) {
408 for (i = index; i < (int) (index + nslots); i++)
410 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
411 io_tlb_list[i] = ++count;
412 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
415 * Update the indices to avoid searching in the next
418 io_tlb_index = ((index + nslots) < io_tlb_nslabs
419 ? (index + nslots) : 0);
424 if (index >= io_tlb_nslabs)
426 } while (index != wrap);
429 spin_unlock_irqrestore(&io_tlb_lock, flags);
432 spin_unlock_irqrestore(&io_tlb_lock, flags);
435 * Save away the mapping from the original address to the DMA address.
436 * This is needed when we sync the memory. Then we sync the buffer if
439 for (i = 0; i < nslots; i++)
440 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
441 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
442 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
448 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
451 do_unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
454 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
455 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
456 phys_addr_t phys = io_tlb_orig_addr[index];
459 * First, sync the memory before unmapping the entry
461 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
462 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
465 * Return the buffer to the free list by setting the corresponding
466 * entries to indicate the number of contigous entries available.
467 * While returning the entries to the free list, we merge the entries
468 * with slots below and above the pool being returned.
470 spin_lock_irqsave(&io_tlb_lock, flags);
472 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
473 io_tlb_list[index + nslots] : 0);
475 * Step 1: return the slots to the free list, merging the
476 * slots with superceeding slots
478 for (i = index + nslots - 1; i >= index; i--)
479 io_tlb_list[i] = ++count;
481 * Step 2: merge the returned slots with the preceding slots,
482 * if available (non zero)
484 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
485 io_tlb_list[i] = ++count;
487 spin_unlock_irqrestore(&io_tlb_lock, flags);
491 sync_single(struct device *hwdev, char *dma_addr, size_t size,
494 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
495 phys_addr_t phys = io_tlb_orig_addr[index];
497 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
501 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
502 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
504 BUG_ON(dir != DMA_TO_DEVICE);
506 case SYNC_FOR_DEVICE:
507 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
508 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
510 BUG_ON(dir != DMA_FROM_DEVICE);
518 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
519 dma_addr_t *dma_handle, gfp_t flags)
523 int order = get_order(size);
524 u64 dma_mask = DMA_BIT_MASK(32);
526 if (hwdev && hwdev->coherent_dma_mask)
527 dma_mask = hwdev->coherent_dma_mask;
529 ret = (void *)__get_free_pages(flags, order);
530 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size > dma_mask) {
532 * The allocated memory isn't reachable by the device.
534 free_pages((unsigned long) ret, order);
539 * We are either out of memory or the device can't DMA
540 * to GFP_DMA memory; fall back on map_single(), which
541 * will grab memory from the lowest available address range.
543 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
548 memset(ret, 0, size);
549 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
551 /* Confirm address can be DMA'd by device */
552 if (dev_addr + size > dma_mask) {
553 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
554 (unsigned long long)dma_mask,
555 (unsigned long long)dev_addr);
557 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
558 do_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
561 *dma_handle = dev_addr;
564 EXPORT_SYMBOL(swiotlb_alloc_coherent);
567 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
570 phys_addr_t paddr = swiotlb_bus_to_phys(hwdev, dev_addr);
572 WARN_ON(irqs_disabled());
573 if (!is_swiotlb_buffer(paddr))
574 free_pages((unsigned long)vaddr, get_order(size));
576 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
577 do_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
579 EXPORT_SYMBOL(swiotlb_free_coherent);
582 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
585 * Ran out of IOMMU space for this operation. This is very bad.
586 * Unfortunately the drivers cannot handle this operation properly.
587 * unless they check for dma_mapping_error (most don't)
588 * When the mapping is small enough return a static buffer to limit
589 * the damage, or panic when the transfer is too big.
591 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
592 "device %s\n", size, dev ? dev_name(dev) : "?");
594 if (size > io_tlb_overflow && do_panic) {
595 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
596 panic("DMA: Memory would be corrupted\n");
597 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
598 panic("DMA: Random memory would be DMAed\n");
603 * Map a single buffer of the indicated size for DMA in streaming mode. The
604 * physical address to use is returned.
606 * Once the device is given the dma address, the device owns this memory until
607 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
609 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
610 unsigned long offset, size_t size,
611 enum dma_data_direction dir,
612 struct dma_attrs *attrs)
614 phys_addr_t phys = page_to_phys(page) + offset;
615 dma_addr_t dev_addr = swiotlb_phys_to_bus(dev, phys);
618 BUG_ON(dir == DMA_NONE);
620 * If the address happens to be in the device's DMA window,
621 * we can safely return the device addr and not worry about bounce
624 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
628 * Oh well, have to allocate and map a bounce buffer.
630 map = map_single(dev, phys, size, dir);
632 swiotlb_full(dev, size, dir, 1);
633 map = io_tlb_overflow_buffer;
636 dev_addr = swiotlb_virt_to_bus(dev, map);
639 * Ensure that the address returned is DMA'ble
641 if (!dma_capable(dev, dev_addr, size))
642 panic("map_single: bounce buffer is not DMA'ble");
646 EXPORT_SYMBOL_GPL(swiotlb_map_page);
649 * Unmap a single streaming mode DMA translation. The dma_addr and size must
650 * match what was provided for in a previous swiotlb_map_page call. All
651 * other usages are undefined.
653 * After this call, reads by the cpu to the buffer are guaranteed to see
654 * whatever the device wrote there.
656 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
657 size_t size, int dir)
659 phys_addr_t paddr = swiotlb_bus_to_phys(hwdev, dev_addr);
661 BUG_ON(dir == DMA_NONE);
663 if (is_swiotlb_buffer(paddr)) {
664 do_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
668 if (dir != DMA_FROM_DEVICE)
672 * phys_to_virt doesn't work with hihgmem page but we could
673 * call dma_mark_clean() with hihgmem page here. However, we
674 * are fine since dma_mark_clean() is null on POWERPC. We can
675 * make dma_mark_clean() take a physical address if necessary.
677 dma_mark_clean(phys_to_virt(paddr), size);
680 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
681 size_t size, enum dma_data_direction dir,
682 struct dma_attrs *attrs)
684 unmap_single(hwdev, dev_addr, size, dir);
686 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
689 * Make physical memory consistent for a single streaming mode DMA translation
692 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
693 * using the cpu, yet do not wish to teardown the dma mapping, you must
694 * call this function before doing so. At the next point you give the dma
695 * address back to the card, you must first perform a
696 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
699 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
700 size_t size, int dir, int target)
702 phys_addr_t paddr = swiotlb_bus_to_phys(hwdev, dev_addr);
704 BUG_ON(dir == DMA_NONE);
706 if (is_swiotlb_buffer(paddr)) {
707 sync_single(hwdev, phys_to_virt(paddr), size, dir, target);
711 if (dir != DMA_FROM_DEVICE)
714 dma_mark_clean(phys_to_virt(paddr), size);
718 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
719 size_t size, enum dma_data_direction dir)
721 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
723 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
726 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
727 size_t size, enum dma_data_direction dir)
729 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
731 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
734 * Same as above, but for a sub-range of the mapping.
737 swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
738 unsigned long offset, size_t size,
741 swiotlb_sync_single(hwdev, dev_addr + offset, size, dir, target);
745 swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
746 unsigned long offset, size_t size,
747 enum dma_data_direction dir)
749 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
752 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
755 swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
756 unsigned long offset, size_t size,
757 enum dma_data_direction dir)
759 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
762 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
765 * Map a set of buffers described by scatterlist in streaming mode for DMA.
766 * This is the scatter-gather version of the above swiotlb_map_page
767 * interface. Here the scatter gather list elements are each tagged with the
768 * appropriate dma address and length. They are obtained via
769 * sg_dma_{address,length}(SG).
771 * NOTE: An implementation may be able to use a smaller number of
772 * DMA address/length pairs than there are SG table elements.
773 * (for example via virtual mapping capabilities)
774 * The routine returns the number of addr/length pairs actually
775 * used, at most nents.
777 * Device ownership issues as mentioned above for swiotlb_map_page are the
781 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
782 enum dma_data_direction dir, struct dma_attrs *attrs)
784 struct scatterlist *sg;
787 BUG_ON(dir == DMA_NONE);
789 for_each_sg(sgl, sg, nelems, i) {
790 phys_addr_t paddr = sg_phys(sg);
791 dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
794 !dma_capable(hwdev, dev_addr, sg->length)) {
795 void *map = map_single(hwdev, sg_phys(sg),
798 /* Don't panic here, we expect map_sg users
799 to do proper error handling. */
800 swiotlb_full(hwdev, sg->length, dir, 0);
801 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
803 sgl[0].dma_length = 0;
806 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
808 sg->dma_address = dev_addr;
809 sg->dma_length = sg->length;
813 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
816 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
819 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
821 EXPORT_SYMBOL(swiotlb_map_sg);
824 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
825 * concerning calls here are the same as for swiotlb_unmap_page() above.
828 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
829 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
831 struct scatterlist *sg;
834 BUG_ON(dir == DMA_NONE);
836 for_each_sg(sgl, sg, nelems, i)
837 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
840 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
843 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
846 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
848 EXPORT_SYMBOL(swiotlb_unmap_sg);
851 * Make physical memory consistent for a set of streaming mode DMA translations
854 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
858 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
859 int nelems, int dir, int target)
861 struct scatterlist *sg;
864 for_each_sg(sgl, sg, nelems, i)
865 swiotlb_sync_single(hwdev, sg->dma_address,
866 sg->dma_length, dir, target);
870 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
871 int nelems, enum dma_data_direction dir)
873 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
875 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
878 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
879 int nelems, enum dma_data_direction dir)
881 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
883 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
886 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
888 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
890 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
893 * Return whether the given device DMA address mask can be supported
894 * properly. For example, if your device can only drive the low 24-bits
895 * during bus mastering, then you would pass 0x00ffffff as the mask to
899 swiotlb_dma_supported(struct device *hwdev, u64 mask)
901 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
903 EXPORT_SYMBOL(swiotlb_dma_supported);