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>
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)
54 * Enumeration for sync targets
56 enum dma_sync_target {
64 * Used to do a quick range check in unmap_single and
65 * sync_single_*, to see if the memory was in fact allocated by this
68 static char *io_tlb_start, *io_tlb_end;
71 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
72 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
74 static unsigned long io_tlb_nslabs;
77 * When the IOMMU overflows we return a fallback buffer. This sets the size.
79 static unsigned long io_tlb_overflow = 32*1024;
81 void *io_tlb_overflow_buffer;
84 * This is a free list describing the number of free entries available from
87 static unsigned int *io_tlb_list;
88 static unsigned int io_tlb_index;
91 * We need to save away the original address corresponding to a mapped entry
92 * for the sync operations.
94 static phys_addr_t *io_tlb_orig_addr;
97 * Protect the above data structures in the map and unmap calls
99 static DEFINE_SPINLOCK(io_tlb_lock);
101 static int late_alloc;
104 setup_io_tlb_npages(char *str)
107 io_tlb_nslabs = simple_strtoul(str, &str, 0);
108 /* avoid tail segment of size < IO_TLB_SEGSIZE */
109 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
113 if (!strcmp(str, "force"))
118 __setup("swiotlb=", setup_io_tlb_npages);
119 /* make io_tlb_overflow tunable too? */
121 /* Note that this doesn't work with highmem page */
122 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
123 volatile void *address)
125 return phys_to_dma(hwdev, virt_to_phys(address));
128 void swiotlb_print_info(void)
130 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
131 phys_addr_t pstart, pend;
133 pstart = virt_to_phys(io_tlb_start);
134 pend = virt_to_phys(io_tlb_end);
136 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
137 bytes >> 20, io_tlb_start, io_tlb_end);
138 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
139 (unsigned long long)pstart,
140 (unsigned long long)pend);
144 * Statically reserve bounce buffer space and initialize bounce buffer data
145 * structures for the software IO TLB used to implement the DMA API.
148 swiotlb_init_with_default_size(size_t default_size, int verbose)
150 unsigned long i, bytes;
152 if (!io_tlb_nslabs) {
153 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
154 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
157 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
160 * Get IO TLB memory from the low pages
162 io_tlb_start = alloc_bootmem_low_pages(bytes);
164 panic("Cannot allocate SWIOTLB buffer");
165 io_tlb_end = io_tlb_start + bytes;
168 * Allocate and initialize the free list array. This array is used
169 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
170 * between io_tlb_start and io_tlb_end.
172 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
173 for (i = 0; i < io_tlb_nslabs; i++)
174 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
176 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
179 * Get the overflow emergency buffer
181 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
182 if (!io_tlb_overflow_buffer)
183 panic("Cannot allocate SWIOTLB overflow buffer!\n");
185 swiotlb_print_info();
189 swiotlb_init(int verbose)
191 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
195 * Systems with larger DMA zones (those that don't support ISA) can
196 * initialize the swiotlb later using the slab allocator if needed.
197 * This should be just like above, but with some error catching.
200 swiotlb_late_init_with_default_size(size_t default_size)
202 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
205 if (!io_tlb_nslabs) {
206 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
207 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
211 * Get IO TLB memory from the low pages
213 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
214 io_tlb_nslabs = SLABS_PER_PAGE << order;
215 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
217 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
218 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
228 if (order != get_order(bytes)) {
229 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
230 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
231 io_tlb_nslabs = SLABS_PER_PAGE << order;
232 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
234 io_tlb_end = io_tlb_start + bytes;
235 memset(io_tlb_start, 0, bytes);
238 * Allocate and initialize the free list array. This array is used
239 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
240 * between io_tlb_start and io_tlb_end.
242 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
243 get_order(io_tlb_nslabs * sizeof(int)));
247 for (i = 0; i < io_tlb_nslabs; i++)
248 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
251 io_tlb_orig_addr = (phys_addr_t *)
252 __get_free_pages(GFP_KERNEL,
253 get_order(io_tlb_nslabs *
254 sizeof(phys_addr_t)));
255 if (!io_tlb_orig_addr)
258 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
261 * Get the overflow emergency buffer
263 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
264 get_order(io_tlb_overflow));
265 if (!io_tlb_overflow_buffer)
268 swiotlb_print_info();
275 free_pages((unsigned long)io_tlb_orig_addr,
276 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
277 io_tlb_orig_addr = NULL;
279 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
284 free_pages((unsigned long)io_tlb_start, order);
287 io_tlb_nslabs = req_nslabs;
291 void __init swiotlb_free(void)
293 if (!io_tlb_overflow_buffer)
297 free_pages((unsigned long)io_tlb_overflow_buffer,
298 get_order(io_tlb_overflow));
299 free_pages((unsigned long)io_tlb_orig_addr,
300 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
301 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
303 free_pages((unsigned long)io_tlb_start,
304 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
306 free_bootmem_late(__pa(io_tlb_overflow_buffer),
308 free_bootmem_late(__pa(io_tlb_orig_addr),
309 io_tlb_nslabs * sizeof(phys_addr_t));
310 free_bootmem_late(__pa(io_tlb_list),
311 io_tlb_nslabs * sizeof(int));
312 free_bootmem_late(__pa(io_tlb_start),
313 io_tlb_nslabs << IO_TLB_SHIFT);
317 static int is_swiotlb_buffer(phys_addr_t paddr)
319 return paddr >= virt_to_phys(io_tlb_start) &&
320 paddr < virt_to_phys(io_tlb_end);
324 * Bounce: copy the swiotlb buffer back to the original dma location
326 static void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
327 enum dma_data_direction dir)
329 unsigned long pfn = PFN_DOWN(phys);
331 if (PageHighMem(pfn_to_page(pfn))) {
332 /* The buffer does not have a mapping. Map it in and copy */
333 unsigned int offset = phys & ~PAGE_MASK;
339 sz = min_t(size_t, PAGE_SIZE - offset, size);
341 local_irq_save(flags);
342 buffer = kmap_atomic(pfn_to_page(pfn),
344 if (dir == DMA_TO_DEVICE)
345 memcpy(dma_addr, buffer + offset, sz);
347 memcpy(buffer + offset, dma_addr, sz);
348 kunmap_atomic(buffer, KM_BOUNCE_READ);
349 local_irq_restore(flags);
357 if (dir == DMA_TO_DEVICE)
358 memcpy(dma_addr, phys_to_virt(phys), size);
360 memcpy(phys_to_virt(phys), dma_addr, size);
365 * Allocates bounce buffer and returns its kernel virtual address.
368 map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
372 unsigned int nslots, stride, index, wrap;
374 unsigned long start_dma_addr;
376 unsigned long offset_slots;
377 unsigned long max_slots;
379 mask = dma_get_seg_boundary(hwdev);
380 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
382 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
385 * Carefully handle integer overflow which can occur when mask == ~0UL.
388 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
389 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
392 * For mappings greater than a page, we limit the stride (and
393 * hence alignment) to a page size.
395 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
396 if (size > PAGE_SIZE)
397 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
404 * Find suitable number of IO TLB entries size that will fit this
405 * request and allocate a buffer from that IO TLB pool.
407 spin_lock_irqsave(&io_tlb_lock, flags);
408 index = ALIGN(io_tlb_index, stride);
409 if (index >= io_tlb_nslabs)
414 while (iommu_is_span_boundary(index, nslots, offset_slots,
417 if (index >= io_tlb_nslabs)
424 * If we find a slot that indicates we have 'nslots' number of
425 * contiguous buffers, we allocate the buffers from that slot
426 * and mark the entries as '0' indicating unavailable.
428 if (io_tlb_list[index] >= nslots) {
431 for (i = index; i < (int) (index + nslots); i++)
433 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
434 io_tlb_list[i] = ++count;
435 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
438 * Update the indices to avoid searching in the next
441 io_tlb_index = ((index + nslots) < io_tlb_nslabs
442 ? (index + nslots) : 0);
447 if (index >= io_tlb_nslabs)
449 } while (index != wrap);
452 spin_unlock_irqrestore(&io_tlb_lock, flags);
455 spin_unlock_irqrestore(&io_tlb_lock, flags);
458 * Save away the mapping from the original address to the DMA address.
459 * This is needed when we sync the memory. Then we sync the buffer if
462 for (i = 0; i < nslots; i++)
463 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
464 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
465 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
471 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
474 do_unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
477 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
478 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
479 phys_addr_t phys = io_tlb_orig_addr[index];
482 * First, sync the memory before unmapping the entry
484 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
485 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
488 * Return the buffer to the free list by setting the corresponding
489 * entries to indicate the number of contiguous entries available.
490 * While returning the entries to the free list, we merge the entries
491 * with slots below and above the pool being returned.
493 spin_lock_irqsave(&io_tlb_lock, flags);
495 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
496 io_tlb_list[index + nslots] : 0);
498 * Step 1: return the slots to the free list, merging the
499 * slots with superceeding slots
501 for (i = index + nslots - 1; i >= index; i--)
502 io_tlb_list[i] = ++count;
504 * Step 2: merge the returned slots with the preceding slots,
505 * if available (non zero)
507 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
508 io_tlb_list[i] = ++count;
510 spin_unlock_irqrestore(&io_tlb_lock, flags);
514 sync_single(struct device *hwdev, char *dma_addr, size_t size,
517 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
518 phys_addr_t phys = io_tlb_orig_addr[index];
520 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
524 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
525 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
527 BUG_ON(dir != DMA_TO_DEVICE);
529 case SYNC_FOR_DEVICE:
530 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
531 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
533 BUG_ON(dir != DMA_FROM_DEVICE);
541 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
542 dma_addr_t *dma_handle, gfp_t flags)
546 int order = get_order(size);
547 u64 dma_mask = DMA_BIT_MASK(32);
549 if (hwdev && hwdev->coherent_dma_mask)
550 dma_mask = hwdev->coherent_dma_mask;
552 ret = (void *)__get_free_pages(flags, order);
553 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
555 * The allocated memory isn't reachable by the device.
557 free_pages((unsigned long) ret, order);
562 * We are either out of memory or the device can't DMA
563 * to GFP_DMA memory; fall back on map_single(), which
564 * will grab memory from the lowest available address range.
566 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
571 memset(ret, 0, size);
572 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
574 /* Confirm address can be DMA'd by device */
575 if (dev_addr + size - 1 > dma_mask) {
576 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
577 (unsigned long long)dma_mask,
578 (unsigned long long)dev_addr);
580 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
581 do_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
584 *dma_handle = dev_addr;
587 EXPORT_SYMBOL(swiotlb_alloc_coherent);
590 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
593 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
595 WARN_ON(irqs_disabled());
596 if (!is_swiotlb_buffer(paddr))
597 free_pages((unsigned long)vaddr, get_order(size));
599 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
600 do_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
602 EXPORT_SYMBOL(swiotlb_free_coherent);
605 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
608 * Ran out of IOMMU space for this operation. This is very bad.
609 * Unfortunately the drivers cannot handle this operation properly.
610 * unless they check for dma_mapping_error (most don't)
611 * When the mapping is small enough return a static buffer to limit
612 * the damage, or panic when the transfer is too big.
614 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
615 "device %s\n", size, dev ? dev_name(dev) : "?");
617 if (size <= io_tlb_overflow || !do_panic)
620 if (dir == DMA_BIDIRECTIONAL)
621 panic("DMA: Random memory could be DMA accessed\n");
622 if (dir == DMA_FROM_DEVICE)
623 panic("DMA: Random memory could be DMA written\n");
624 if (dir == DMA_TO_DEVICE)
625 panic("DMA: Random memory could be DMA read\n");
629 * Map a single buffer of the indicated size for DMA in streaming mode. The
630 * physical address to use is returned.
632 * Once the device is given the dma address, the device owns this memory until
633 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
635 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
636 unsigned long offset, size_t size,
637 enum dma_data_direction dir,
638 struct dma_attrs *attrs)
640 phys_addr_t phys = page_to_phys(page) + offset;
641 dma_addr_t dev_addr = phys_to_dma(dev, phys);
644 BUG_ON(dir == DMA_NONE);
646 * If the address happens to be in the device's DMA window,
647 * we can safely return the device addr and not worry about bounce
650 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
654 * Oh well, have to allocate and map a bounce buffer.
656 map = map_single(dev, phys, size, dir);
658 swiotlb_full(dev, size, dir, 1);
659 map = io_tlb_overflow_buffer;
662 dev_addr = swiotlb_virt_to_bus(dev, map);
665 * Ensure that the address returned is DMA'ble
667 if (!dma_capable(dev, dev_addr, size))
668 panic("map_single: bounce buffer is not DMA'ble");
672 EXPORT_SYMBOL_GPL(swiotlb_map_page);
675 * Unmap a single streaming mode DMA translation. The dma_addr and size must
676 * match what was provided for in a previous swiotlb_map_page call. All
677 * other usages are undefined.
679 * After this call, reads by the cpu to the buffer are guaranteed to see
680 * whatever the device wrote there.
682 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
683 size_t size, int dir)
685 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
687 BUG_ON(dir == DMA_NONE);
689 if (is_swiotlb_buffer(paddr)) {
690 do_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
694 if (dir != DMA_FROM_DEVICE)
698 * phys_to_virt doesn't work with hihgmem page but we could
699 * call dma_mark_clean() with hihgmem page here. However, we
700 * are fine since dma_mark_clean() is null on POWERPC. We can
701 * make dma_mark_clean() take a physical address if necessary.
703 dma_mark_clean(phys_to_virt(paddr), size);
706 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
707 size_t size, enum dma_data_direction dir,
708 struct dma_attrs *attrs)
710 unmap_single(hwdev, dev_addr, size, dir);
712 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
715 * Make physical memory consistent for a single streaming mode DMA translation
718 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
719 * using the cpu, yet do not wish to teardown the dma mapping, you must
720 * call this function before doing so. At the next point you give the dma
721 * address back to the card, you must first perform a
722 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
725 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
726 size_t size, int dir, int target)
728 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
730 BUG_ON(dir == DMA_NONE);
732 if (is_swiotlb_buffer(paddr)) {
733 sync_single(hwdev, phys_to_virt(paddr), size, dir, target);
737 if (dir != DMA_FROM_DEVICE)
740 dma_mark_clean(phys_to_virt(paddr), size);
744 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
745 size_t size, enum dma_data_direction dir)
747 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
749 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
752 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
753 size_t size, enum dma_data_direction dir)
755 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
757 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
760 * Map a set of buffers described by scatterlist in streaming mode for DMA.
761 * This is the scatter-gather version of the above swiotlb_map_page
762 * interface. Here the scatter gather list elements are each tagged with the
763 * appropriate dma address and length. They are obtained via
764 * sg_dma_{address,length}(SG).
766 * NOTE: An implementation may be able to use a smaller number of
767 * DMA address/length pairs than there are SG table elements.
768 * (for example via virtual mapping capabilities)
769 * The routine returns the number of addr/length pairs actually
770 * used, at most nents.
772 * Device ownership issues as mentioned above for swiotlb_map_page are the
776 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
777 enum dma_data_direction dir, struct dma_attrs *attrs)
779 struct scatterlist *sg;
782 BUG_ON(dir == DMA_NONE);
784 for_each_sg(sgl, sg, nelems, i) {
785 phys_addr_t paddr = sg_phys(sg);
786 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
789 !dma_capable(hwdev, dev_addr, sg->length)) {
790 void *map = map_single(hwdev, sg_phys(sg),
793 /* Don't panic here, we expect map_sg users
794 to do proper error handling. */
795 swiotlb_full(hwdev, sg->length, dir, 0);
796 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
798 sgl[0].dma_length = 0;
801 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
803 sg->dma_address = dev_addr;
804 sg->dma_length = sg->length;
808 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
811 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
814 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
816 EXPORT_SYMBOL(swiotlb_map_sg);
819 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
820 * concerning calls here are the same as for swiotlb_unmap_page() above.
823 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
824 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
826 struct scatterlist *sg;
829 BUG_ON(dir == DMA_NONE);
831 for_each_sg(sgl, sg, nelems, i)
832 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
835 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
838 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
841 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
843 EXPORT_SYMBOL(swiotlb_unmap_sg);
846 * Make physical memory consistent for a set of streaming mode DMA translations
849 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
853 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
854 int nelems, int dir, int target)
856 struct scatterlist *sg;
859 for_each_sg(sgl, sg, nelems, i)
860 swiotlb_sync_single(hwdev, sg->dma_address,
861 sg->dma_length, dir, target);
865 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
866 int nelems, enum dma_data_direction dir)
868 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
870 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
873 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
874 int nelems, enum dma_data_direction dir)
876 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
878 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
881 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
883 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
885 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
888 * Return whether the given device DMA address mask can be supported
889 * properly. For example, if your device can only drive the low 24-bits
890 * during bus mastering, then you would pass 0x00ffffff as the mask to
894 swiotlb_dma_supported(struct device *hwdev, u64 mask)
896 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
898 EXPORT_SYMBOL(swiotlb_dma_supported);