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 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
128 volatile void *address)
130 return swiotlb_phys_to_bus(hwdev, virt_to_phys(address));
133 void * __weak swiotlb_bus_to_virt(struct device *hwdev, dma_addr_t address)
135 return phys_to_virt(swiotlb_bus_to_phys(hwdev, address));
138 int __weak swiotlb_arch_address_needs_mapping(struct device *hwdev,
139 dma_addr_t addr, size_t size)
141 return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size);
144 int __weak swiotlb_arch_range_needs_mapping(phys_addr_t paddr, size_t size)
149 static void swiotlb_print_info(unsigned long bytes)
151 phys_addr_t pstart, pend;
153 pstart = virt_to_phys(io_tlb_start);
154 pend = virt_to_phys(io_tlb_end);
156 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
157 bytes >> 20, io_tlb_start, io_tlb_end);
158 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
159 (unsigned long long)pstart,
160 (unsigned long long)pend);
164 * Statically reserve bounce buffer space and initialize bounce buffer data
165 * structures for the software IO TLB used to implement the DMA API.
168 swiotlb_init_with_default_size(size_t default_size)
170 unsigned long i, bytes;
172 if (!io_tlb_nslabs) {
173 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
174 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
177 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
180 * Get IO TLB memory from the low pages
182 io_tlb_start = alloc_bootmem_low_pages(bytes);
184 panic("Cannot allocate SWIOTLB buffer");
185 io_tlb_end = io_tlb_start + bytes;
188 * Allocate and initialize the free list array. This array is used
189 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
190 * between io_tlb_start and io_tlb_end.
192 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
193 for (i = 0; i < io_tlb_nslabs; i++)
194 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
196 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(phys_addr_t));
199 * Get the overflow emergency buffer
201 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
202 if (!io_tlb_overflow_buffer)
203 panic("Cannot allocate SWIOTLB overflow buffer!\n");
205 swiotlb_print_info(bytes);
211 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
215 * Systems with larger DMA zones (those that don't support ISA) can
216 * initialize the swiotlb later using the slab allocator if needed.
217 * This should be just like above, but with some error catching.
220 swiotlb_late_init_with_default_size(size_t default_size)
222 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
225 if (!io_tlb_nslabs) {
226 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
227 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
231 * Get IO TLB memory from the low pages
233 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
234 io_tlb_nslabs = SLABS_PER_PAGE << order;
235 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
237 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
238 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
248 if (order != get_order(bytes)) {
249 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
250 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
251 io_tlb_nslabs = SLABS_PER_PAGE << order;
252 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
254 io_tlb_end = io_tlb_start + bytes;
255 memset(io_tlb_start, 0, bytes);
258 * Allocate and initialize the free list array. This array is used
259 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
260 * between io_tlb_start and io_tlb_end.
262 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
263 get_order(io_tlb_nslabs * sizeof(int)));
267 for (i = 0; i < io_tlb_nslabs; i++)
268 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
271 io_tlb_orig_addr = (phys_addr_t *)
272 __get_free_pages(GFP_KERNEL,
273 get_order(io_tlb_nslabs *
274 sizeof(phys_addr_t)));
275 if (!io_tlb_orig_addr)
278 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
281 * Get the overflow emergency buffer
283 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
284 get_order(io_tlb_overflow));
285 if (!io_tlb_overflow_buffer)
288 swiotlb_print_info(bytes);
293 free_pages((unsigned long)io_tlb_orig_addr,
294 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
295 io_tlb_orig_addr = NULL;
297 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
302 free_pages((unsigned long)io_tlb_start, order);
305 io_tlb_nslabs = req_nslabs;
310 address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size)
312 return swiotlb_arch_address_needs_mapping(hwdev, addr, size);
315 static inline int range_needs_mapping(phys_addr_t paddr, size_t size)
317 return swiotlb_force || swiotlb_arch_range_needs_mapping(paddr, size);
320 static int is_swiotlb_buffer(char *addr)
322 return addr >= io_tlb_start && addr < io_tlb_end;
326 * Bounce: copy the swiotlb buffer back to the original dma location
328 static void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
329 enum dma_data_direction dir)
331 unsigned long pfn = PFN_DOWN(phys);
333 if (PageHighMem(pfn_to_page(pfn))) {
334 /* The buffer does not have a mapping. Map it in and copy */
335 unsigned int offset = phys & ~PAGE_MASK;
341 sz = min_t(size_t, PAGE_SIZE - offset, size);
343 local_irq_save(flags);
344 buffer = kmap_atomic(pfn_to_page(pfn),
346 if (dir == DMA_TO_DEVICE)
347 memcpy(dma_addr, buffer + offset, sz);
349 memcpy(buffer + offset, dma_addr, sz);
350 kunmap_atomic(buffer, KM_BOUNCE_READ);
351 local_irq_restore(flags);
359 if (dir == DMA_TO_DEVICE)
360 memcpy(dma_addr, phys_to_virt(phys), size);
362 memcpy(phys_to_virt(phys), dma_addr, size);
367 * Allocates bounce buffer and returns its kernel virtual address.
370 map_single(struct device *hwdev, phys_addr_t phys, size_t size, int dir)
374 unsigned int nslots, stride, index, wrap;
376 unsigned long start_dma_addr;
378 unsigned long offset_slots;
379 unsigned long max_slots;
381 mask = dma_get_seg_boundary(hwdev);
382 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
384 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
387 * Carefully handle integer overflow which can occur when mask == ~0UL.
390 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
391 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
394 * For mappings greater than a page, we limit the stride (and
395 * hence alignment) to a page size.
397 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
398 if (size > PAGE_SIZE)
399 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
406 * Find suitable number of IO TLB entries size that will fit this
407 * request and allocate a buffer from that IO TLB pool.
409 spin_lock_irqsave(&io_tlb_lock, flags);
410 index = ALIGN(io_tlb_index, stride);
411 if (index >= io_tlb_nslabs)
416 while (iommu_is_span_boundary(index, nslots, offset_slots,
419 if (index >= io_tlb_nslabs)
426 * If we find a slot that indicates we have 'nslots' number of
427 * contiguous buffers, we allocate the buffers from that slot
428 * and mark the entries as '0' indicating unavailable.
430 if (io_tlb_list[index] >= nslots) {
433 for (i = index; i < (int) (index + nslots); i++)
435 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
436 io_tlb_list[i] = ++count;
437 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
440 * Update the indices to avoid searching in the next
443 io_tlb_index = ((index + nslots) < io_tlb_nslabs
444 ? (index + nslots) : 0);
449 if (index >= io_tlb_nslabs)
451 } while (index != wrap);
454 spin_unlock_irqrestore(&io_tlb_lock, flags);
457 spin_unlock_irqrestore(&io_tlb_lock, flags);
460 * Save away the mapping from the original address to the DMA address.
461 * This is needed when we sync the memory. Then we sync the buffer if
464 for (i = 0; i < nslots; i++)
465 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
466 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
467 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
473 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
476 do_unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
479 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
480 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
481 phys_addr_t phys = io_tlb_orig_addr[index];
484 * First, sync the memory before unmapping the entry
486 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
487 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
490 * Return the buffer to the free list by setting the corresponding
491 * entries to indicate the number of contigous entries available.
492 * While returning the entries to the free list, we merge the entries
493 * with slots below and above the pool being returned.
495 spin_lock_irqsave(&io_tlb_lock, flags);
497 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
498 io_tlb_list[index + nslots] : 0);
500 * Step 1: return the slots to the free list, merging the
501 * slots with superceeding slots
503 for (i = index + nslots - 1; i >= index; i--)
504 io_tlb_list[i] = ++count;
506 * Step 2: merge the returned slots with the preceding slots,
507 * if available (non zero)
509 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
510 io_tlb_list[i] = ++count;
512 spin_unlock_irqrestore(&io_tlb_lock, flags);
516 sync_single(struct device *hwdev, char *dma_addr, size_t size,
519 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
520 phys_addr_t phys = io_tlb_orig_addr[index];
522 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
526 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
527 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
529 BUG_ON(dir != DMA_TO_DEVICE);
531 case SYNC_FOR_DEVICE:
532 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
533 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
535 BUG_ON(dir != DMA_FROM_DEVICE);
543 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
544 dma_addr_t *dma_handle, gfp_t flags)
548 int order = get_order(size);
549 u64 dma_mask = DMA_BIT_MASK(32);
551 if (hwdev && hwdev->coherent_dma_mask)
552 dma_mask = hwdev->coherent_dma_mask;
554 ret = (void *)__get_free_pages(flags, order);
556 !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(hwdev, ret),
559 * The allocated memory isn't reachable by the device.
561 free_pages((unsigned long) ret, order);
566 * We are either out of memory or the device can't DMA
567 * to GFP_DMA memory; fall back on map_single(), which
568 * will grab memory from the lowest available address range.
570 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
575 memset(ret, 0, size);
576 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
578 /* Confirm address can be DMA'd by device */
579 if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) {
580 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
581 (unsigned long long)dma_mask,
582 (unsigned long long)dev_addr);
584 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
585 do_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
588 *dma_handle = dev_addr;
591 EXPORT_SYMBOL(swiotlb_alloc_coherent);
594 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
595 dma_addr_t dma_handle)
597 WARN_ON(irqs_disabled());
598 if (!is_swiotlb_buffer(vaddr))
599 free_pages((unsigned long) vaddr, get_order(size));
601 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
602 do_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
604 EXPORT_SYMBOL(swiotlb_free_coherent);
607 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
610 * Ran out of IOMMU space for this operation. This is very bad.
611 * Unfortunately the drivers cannot handle this operation properly.
612 * unless they check for dma_mapping_error (most don't)
613 * When the mapping is small enough return a static buffer to limit
614 * the damage, or panic when the transfer is too big.
616 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
617 "device %s\n", size, dev ? dev_name(dev) : "?");
619 if (size > io_tlb_overflow && do_panic) {
620 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
621 panic("DMA: Memory would be corrupted\n");
622 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
623 panic("DMA: Random memory would be DMAed\n");
628 * Map a single buffer of the indicated size for DMA in streaming mode. The
629 * physical address to use is returned.
631 * Once the device is given the dma address, the device owns this memory until
632 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
634 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
635 unsigned long offset, size_t size,
636 enum dma_data_direction dir,
637 struct dma_attrs *attrs)
639 phys_addr_t phys = page_to_phys(page) + offset;
640 dma_addr_t dev_addr = swiotlb_phys_to_bus(dev, phys);
643 BUG_ON(dir == DMA_NONE);
645 * If the address happens to be in the device's DMA window,
646 * we can safely return the device addr and not worry about bounce
649 if (!address_needs_mapping(dev, dev_addr, size) &&
650 !range_needs_mapping(phys, size))
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 (address_needs_mapping(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 char *dma_addr = swiotlb_bus_to_virt(hwdev, dev_addr);
687 BUG_ON(dir == DMA_NONE);
689 if (is_swiotlb_buffer(dma_addr)) {
690 do_unmap_single(hwdev, dma_addr, size, dir);
694 if (dir != DMA_FROM_DEVICE)
697 dma_mark_clean(dma_addr, size);
700 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
701 size_t size, enum dma_data_direction dir,
702 struct dma_attrs *attrs)
704 unmap_single(hwdev, dev_addr, size, dir);
706 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
709 * Make physical memory consistent for a single streaming mode DMA translation
712 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
713 * using the cpu, yet do not wish to teardown the dma mapping, you must
714 * call this function before doing so. At the next point you give the dma
715 * address back to the card, you must first perform a
716 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
719 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
720 size_t size, int dir, int target)
722 char *dma_addr = swiotlb_bus_to_virt(hwdev, dev_addr);
724 BUG_ON(dir == DMA_NONE);
726 if (is_swiotlb_buffer(dma_addr)) {
727 sync_single(hwdev, dma_addr, size, dir, target);
731 if (dir != DMA_FROM_DEVICE)
734 dma_mark_clean(dma_addr, size);
738 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
739 size_t size, enum dma_data_direction dir)
741 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
743 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
746 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
747 size_t size, enum dma_data_direction dir)
749 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
751 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
754 * Same as above, but for a sub-range of the mapping.
757 swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
758 unsigned long offset, size_t size,
761 swiotlb_sync_single(hwdev, dev_addr + offset, size, dir, target);
765 swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
766 unsigned long offset, size_t size,
767 enum dma_data_direction dir)
769 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
772 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
775 swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
776 unsigned long offset, size_t size,
777 enum dma_data_direction dir)
779 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
782 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
785 * Map a set of buffers described by scatterlist in streaming mode for DMA.
786 * This is the scatter-gather version of the above swiotlb_map_page
787 * interface. Here the scatter gather list elements are each tagged with the
788 * appropriate dma address and length. They are obtained via
789 * sg_dma_{address,length}(SG).
791 * NOTE: An implementation may be able to use a smaller number of
792 * DMA address/length pairs than there are SG table elements.
793 * (for example via virtual mapping capabilities)
794 * The routine returns the number of addr/length pairs actually
795 * used, at most nents.
797 * Device ownership issues as mentioned above for swiotlb_map_page are the
801 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
802 enum dma_data_direction dir, struct dma_attrs *attrs)
804 struct scatterlist *sg;
807 BUG_ON(dir == DMA_NONE);
809 for_each_sg(sgl, sg, nelems, i) {
810 phys_addr_t paddr = sg_phys(sg);
811 dma_addr_t dev_addr = swiotlb_phys_to_bus(hwdev, paddr);
813 if (range_needs_mapping(paddr, sg->length) ||
814 address_needs_mapping(hwdev, dev_addr, sg->length)) {
815 void *map = map_single(hwdev, sg_phys(sg),
818 /* Don't panic here, we expect map_sg users
819 to do proper error handling. */
820 swiotlb_full(hwdev, sg->length, dir, 0);
821 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
823 sgl[0].dma_length = 0;
826 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
828 sg->dma_address = dev_addr;
829 sg->dma_length = sg->length;
833 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
836 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
839 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
841 EXPORT_SYMBOL(swiotlb_map_sg);
844 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
845 * concerning calls here are the same as for swiotlb_unmap_page() above.
848 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
849 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
851 struct scatterlist *sg;
854 BUG_ON(dir == DMA_NONE);
856 for_each_sg(sgl, sg, nelems, i)
857 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
860 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
863 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
866 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
868 EXPORT_SYMBOL(swiotlb_unmap_sg);
871 * Make physical memory consistent for a set of streaming mode DMA translations
874 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
878 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
879 int nelems, int dir, int target)
881 struct scatterlist *sg;
884 for_each_sg(sgl, sg, nelems, i)
885 swiotlb_sync_single(hwdev, sg->dma_address,
886 sg->dma_length, dir, target);
890 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
891 int nelems, enum dma_data_direction dir)
893 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
895 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
898 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
899 int nelems, enum dma_data_direction dir)
901 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
903 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
906 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
908 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
910 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
913 * Return whether the given device DMA address mask can be supported
914 * properly. For example, if your device can only drive the low 24-bits
915 * during bus mastering, then you would pass 0x00ffffff as the mask to
919 swiotlb_dma_supported(struct device *hwdev, u64 mask)
921 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
923 EXPORT_SYMBOL(swiotlb_dma_supported);