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 int __weak swiotlb_arch_address_needs_mapping(struct device *hwdev,
135 dma_addr_t addr, size_t size)
137 return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size);
140 static void swiotlb_print_info(unsigned long bytes)
142 phys_addr_t pstart, pend;
144 pstart = virt_to_phys(io_tlb_start);
145 pend = virt_to_phys(io_tlb_end);
147 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
148 bytes >> 20, io_tlb_start, io_tlb_end);
149 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
150 (unsigned long long)pstart,
151 (unsigned long long)pend);
155 * Statically reserve bounce buffer space and initialize bounce buffer data
156 * structures for the software IO TLB used to implement the DMA API.
159 swiotlb_init_with_default_size(size_t default_size)
161 unsigned long i, bytes;
163 if (!io_tlb_nslabs) {
164 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
165 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
168 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
171 * Get IO TLB memory from the low pages
173 io_tlb_start = alloc_bootmem_low_pages(bytes);
175 panic("Cannot allocate SWIOTLB buffer");
176 io_tlb_end = io_tlb_start + bytes;
179 * Allocate and initialize the free list array. This array is used
180 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
181 * between io_tlb_start and io_tlb_end.
183 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
184 for (i = 0; i < io_tlb_nslabs; i++)
185 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
187 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
190 * Get the overflow emergency buffer
192 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
193 if (!io_tlb_overflow_buffer)
194 panic("Cannot allocate SWIOTLB overflow buffer!\n");
196 swiotlb_print_info(bytes);
202 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
206 * Systems with larger DMA zones (those that don't support ISA) can
207 * initialize the swiotlb later using the slab allocator if needed.
208 * This should be just like above, but with some error catching.
211 swiotlb_late_init_with_default_size(size_t default_size)
213 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
216 if (!io_tlb_nslabs) {
217 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
218 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
222 * Get IO TLB memory from the low pages
224 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
225 io_tlb_nslabs = SLABS_PER_PAGE << order;
226 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
228 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
229 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
239 if (order != get_order(bytes)) {
240 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
241 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
242 io_tlb_nslabs = SLABS_PER_PAGE << order;
243 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
245 io_tlb_end = io_tlb_start + bytes;
246 memset(io_tlb_start, 0, bytes);
249 * Allocate and initialize the free list array. This array is used
250 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
251 * between io_tlb_start and io_tlb_end.
253 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
254 get_order(io_tlb_nslabs * sizeof(int)));
258 for (i = 0; i < io_tlb_nslabs; i++)
259 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
262 io_tlb_orig_addr = (phys_addr_t *)
263 __get_free_pages(GFP_KERNEL,
264 get_order(io_tlb_nslabs *
265 sizeof(phys_addr_t)));
266 if (!io_tlb_orig_addr)
269 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
272 * Get the overflow emergency buffer
274 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
275 get_order(io_tlb_overflow));
276 if (!io_tlb_overflow_buffer)
279 swiotlb_print_info(bytes);
284 free_pages((unsigned long)io_tlb_orig_addr,
285 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
286 io_tlb_orig_addr = NULL;
288 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
293 free_pages((unsigned long)io_tlb_start, order);
296 io_tlb_nslabs = req_nslabs;
301 address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size)
303 return swiotlb_arch_address_needs_mapping(hwdev, addr, size);
306 static int is_swiotlb_buffer(phys_addr_t paddr)
308 return paddr >= virt_to_phys(io_tlb_start) &&
309 paddr < virt_to_phys(io_tlb_end);
313 * Bounce: copy the swiotlb buffer back to the original dma location
315 static void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
316 enum dma_data_direction dir)
318 unsigned long pfn = PFN_DOWN(phys);
320 if (PageHighMem(pfn_to_page(pfn))) {
321 /* The buffer does not have a mapping. Map it in and copy */
322 unsigned int offset = phys & ~PAGE_MASK;
328 sz = min_t(size_t, PAGE_SIZE - offset, size);
330 local_irq_save(flags);
331 buffer = kmap_atomic(pfn_to_page(pfn),
333 if (dir == DMA_TO_DEVICE)
334 memcpy(dma_addr, buffer + offset, sz);
336 memcpy(buffer + offset, dma_addr, sz);
337 kunmap_atomic(buffer, KM_BOUNCE_READ);
338 local_irq_restore(flags);
346 if (dir == DMA_TO_DEVICE)
347 memcpy(dma_addr, phys_to_virt(phys), size);
349 memcpy(phys_to_virt(phys), dma_addr, size);
354 * Allocates bounce buffer and returns its kernel virtual address.
357 map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
361 unsigned int nslots, stride, index, wrap;
363 unsigned long start_dma_addr;
365 unsigned long offset_slots;
366 unsigned long max_slots;
368 mask = dma_get_seg_boundary(hwdev);
369 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
371 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
374 * Carefully handle integer overflow which can occur when mask == ~0UL.
377 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
378 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
381 * For mappings greater than a page, we limit the stride (and
382 * hence alignment) to a page size.
384 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
385 if (size > PAGE_SIZE)
386 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
393 * Find suitable number of IO TLB entries size that will fit this
394 * request and allocate a buffer from that IO TLB pool.
396 spin_lock_irqsave(&io_tlb_lock, flags);
397 index = ALIGN(io_tlb_index, stride);
398 if (index >= io_tlb_nslabs)
403 while (iommu_is_span_boundary(index, nslots, offset_slots,
406 if (index >= io_tlb_nslabs)
413 * If we find a slot that indicates we have 'nslots' number of
414 * contiguous buffers, we allocate the buffers from that slot
415 * and mark the entries as '0' indicating unavailable.
417 if (io_tlb_list[index] >= nslots) {
420 for (i = index; i < (int) (index + nslots); i++)
422 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
423 io_tlb_list[i] = ++count;
424 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
427 * Update the indices to avoid searching in the next
430 io_tlb_index = ((index + nslots) < io_tlb_nslabs
431 ? (index + nslots) : 0);
436 if (index >= io_tlb_nslabs)
438 } while (index != wrap);
441 spin_unlock_irqrestore(&io_tlb_lock, flags);
444 spin_unlock_irqrestore(&io_tlb_lock, flags);
447 * Save away the mapping from the original address to the DMA address.
448 * This is needed when we sync the memory. Then we sync the buffer if
451 for (i = 0; i < nslots; i++)
452 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
453 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
454 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
460 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
463 do_unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
466 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
467 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
468 phys_addr_t phys = io_tlb_orig_addr[index];
471 * First, sync the memory before unmapping the entry
473 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
474 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
477 * Return the buffer to the free list by setting the corresponding
478 * entries to indicate the number of contigous entries available.
479 * While returning the entries to the free list, we merge the entries
480 * with slots below and above the pool being returned.
482 spin_lock_irqsave(&io_tlb_lock, flags);
484 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
485 io_tlb_list[index + nslots] : 0);
487 * Step 1: return the slots to the free list, merging the
488 * slots with superceeding slots
490 for (i = index + nslots - 1; i >= index; i--)
491 io_tlb_list[i] = ++count;
493 * Step 2: merge the returned slots with the preceding slots,
494 * if available (non zero)
496 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
497 io_tlb_list[i] = ++count;
499 spin_unlock_irqrestore(&io_tlb_lock, flags);
503 sync_single(struct device *hwdev, char *dma_addr, size_t size,
506 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
507 phys_addr_t phys = io_tlb_orig_addr[index];
509 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
513 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
514 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
516 BUG_ON(dir != DMA_TO_DEVICE);
518 case SYNC_FOR_DEVICE:
519 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
520 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
522 BUG_ON(dir != DMA_FROM_DEVICE);
530 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
531 dma_addr_t *dma_handle, gfp_t flags)
535 int order = get_order(size);
536 u64 dma_mask = DMA_BIT_MASK(32);
538 if (hwdev && hwdev->coherent_dma_mask)
539 dma_mask = hwdev->coherent_dma_mask;
541 ret = (void *)__get_free_pages(flags, order);
543 !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(hwdev, ret),
546 * The allocated memory isn't reachable by the device.
548 free_pages((unsigned long) ret, order);
553 * We are either out of memory or the device can't DMA
554 * to GFP_DMA memory; fall back on map_single(), which
555 * will grab memory from the lowest available address range.
557 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
562 memset(ret, 0, size);
563 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
565 /* Confirm address can be DMA'd by device */
566 if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) {
567 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
568 (unsigned long long)dma_mask,
569 (unsigned long long)dev_addr);
571 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
572 do_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
575 *dma_handle = dev_addr;
578 EXPORT_SYMBOL(swiotlb_alloc_coherent);
581 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
584 phys_addr_t paddr = swiotlb_bus_to_phys(hwdev, dev_addr);
586 WARN_ON(irqs_disabled());
587 if (!is_swiotlb_buffer(paddr))
588 free_pages((unsigned long)vaddr, get_order(size));
590 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
591 do_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
593 EXPORT_SYMBOL(swiotlb_free_coherent);
596 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
599 * Ran out of IOMMU space for this operation. This is very bad.
600 * Unfortunately the drivers cannot handle this operation properly.
601 * unless they check for dma_mapping_error (most don't)
602 * When the mapping is small enough return a static buffer to limit
603 * the damage, or panic when the transfer is too big.
605 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
606 "device %s\n", size, dev ? dev_name(dev) : "?");
608 if (size > io_tlb_overflow && do_panic) {
609 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
610 panic("DMA: Memory would be corrupted\n");
611 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
612 panic("DMA: Random memory would be DMAed\n");
617 * Map a single buffer of the indicated size for DMA in streaming mode. The
618 * physical address to use is returned.
620 * Once the device is given the dma address, the device owns this memory until
621 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
623 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
624 unsigned long offset, size_t size,
625 enum dma_data_direction dir,
626 struct dma_attrs *attrs)
628 phys_addr_t phys = page_to_phys(page) + offset;
629 dma_addr_t dev_addr = swiotlb_phys_to_bus(dev, phys);
632 BUG_ON(dir == DMA_NONE);
634 * If the address happens to be in the device's DMA window,
635 * we can safely return the device addr and not worry about bounce
638 if (!address_needs_mapping(dev, dev_addr, size) && !swiotlb_force)
642 * Oh well, have to allocate and map a bounce buffer.
644 map = map_single(dev, phys, size, dir);
646 swiotlb_full(dev, size, dir, 1);
647 map = io_tlb_overflow_buffer;
650 dev_addr = swiotlb_virt_to_bus(dev, map);
653 * Ensure that the address returned is DMA'ble
655 if (address_needs_mapping(dev, dev_addr, size))
656 panic("map_single: bounce buffer is not DMA'ble");
660 EXPORT_SYMBOL_GPL(swiotlb_map_page);
663 * Unmap a single streaming mode DMA translation. The dma_addr and size must
664 * match what was provided for in a previous swiotlb_map_page call. All
665 * other usages are undefined.
667 * After this call, reads by the cpu to the buffer are guaranteed to see
668 * whatever the device wrote there.
670 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
671 size_t size, int dir)
673 phys_addr_t paddr = swiotlb_bus_to_phys(hwdev, dev_addr);
675 BUG_ON(dir == DMA_NONE);
677 if (is_swiotlb_buffer(paddr)) {
678 do_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
682 if (dir != DMA_FROM_DEVICE)
686 * phys_to_virt doesn't work with hihgmem page but we could
687 * call dma_mark_clean() with hihgmem page here. However, we
688 * are fine since dma_mark_clean() is null on POWERPC. We can
689 * make dma_mark_clean() take a physical address if necessary.
691 dma_mark_clean(phys_to_virt(paddr), size);
694 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
695 size_t size, enum dma_data_direction dir,
696 struct dma_attrs *attrs)
698 unmap_single(hwdev, dev_addr, size, dir);
700 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
703 * Make physical memory consistent for a single streaming mode DMA translation
706 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
707 * using the cpu, yet do not wish to teardown the dma mapping, you must
708 * call this function before doing so. At the next point you give the dma
709 * address back to the card, you must first perform a
710 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
713 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
714 size_t size, int dir, int target)
716 phys_addr_t paddr = swiotlb_bus_to_phys(hwdev, dev_addr);
718 BUG_ON(dir == DMA_NONE);
720 if (is_swiotlb_buffer(paddr)) {
721 sync_single(hwdev, phys_to_virt(paddr), size, dir, target);
725 if (dir != DMA_FROM_DEVICE)
728 dma_mark_clean(phys_to_virt(paddr), size);
732 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
733 size_t size, enum dma_data_direction dir)
735 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
737 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
740 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
741 size_t size, enum dma_data_direction dir)
743 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
745 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
748 * Same as above, but for a sub-range of the mapping.
751 swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
752 unsigned long offset, size_t size,
755 swiotlb_sync_single(hwdev, dev_addr + offset, size, dir, target);
759 swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
760 unsigned long offset, size_t size,
761 enum dma_data_direction dir)
763 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
766 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
769 swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
770 unsigned long offset, size_t size,
771 enum dma_data_direction dir)
773 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
776 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
779 * Map a set of buffers described by scatterlist in streaming mode for DMA.
780 * This is the scatter-gather version of the above swiotlb_map_page
781 * interface. Here the scatter gather list elements are each tagged with the
782 * appropriate dma address and length. They are obtained via
783 * sg_dma_{address,length}(SG).
785 * NOTE: An implementation may be able to use a smaller number of
786 * DMA address/length pairs than there are SG table elements.
787 * (for example via virtual mapping capabilities)
788 * The routine returns the number of addr/length pairs actually
789 * used, at most nents.
791 * Device ownership issues as mentioned above for swiotlb_map_page are the
795 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
796 enum dma_data_direction dir, struct dma_attrs *attrs)
798 struct scatterlist *sg;
801 BUG_ON(dir == DMA_NONE);
803 for_each_sg(sgl, sg, nelems, i) {
804 phys_addr_t paddr = sg_phys(sg);
805 dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
808 address_needs_mapping(hwdev, dev_addr, sg->length)) {
809 void *map = map_single(hwdev, sg_phys(sg),
812 /* Don't panic here, we expect map_sg users
813 to do proper error handling. */
814 swiotlb_full(hwdev, sg->length, dir, 0);
815 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
817 sgl[0].dma_length = 0;
820 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
822 sg->dma_address = dev_addr;
823 sg->dma_length = sg->length;
827 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
830 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
833 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
835 EXPORT_SYMBOL(swiotlb_map_sg);
838 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
839 * concerning calls here are the same as for swiotlb_unmap_page() above.
842 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
843 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
845 struct scatterlist *sg;
848 BUG_ON(dir == DMA_NONE);
850 for_each_sg(sgl, sg, nelems, i)
851 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
854 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
857 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
860 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
862 EXPORT_SYMBOL(swiotlb_unmap_sg);
865 * Make physical memory consistent for a set of streaming mode DMA translations
868 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
872 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
873 int nelems, int dir, int target)
875 struct scatterlist *sg;
878 for_each_sg(sgl, sg, nelems, i)
879 swiotlb_sync_single(hwdev, sg->dma_address,
880 sg->dma_length, dir, target);
884 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
885 int nelems, enum dma_data_direction dir)
887 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
889 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
892 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
893 int nelems, enum dma_data_direction dir)
895 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
897 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
900 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
902 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
904 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
907 * Return whether the given device DMA address mask can be supported
908 * properly. For example, if your device can only drive the low 24-bits
909 * during bus mastering, then you would pass 0x00ffffff as the mask to
913 swiotlb_dma_supported(struct device *hwdev, u64 mask)
915 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
917 EXPORT_SYMBOL(swiotlb_dma_supported);