2 * Dynamic DMA mapping support.
4 * This implementation is a fallback for platforms that do not support
5 * I/O TLBs (aka DMA address translation hardware).
6 * Copyright (C) 2000 Asit Mallick <Asit.K.Mallick@intel.com>
7 * Copyright (C) 2000 Goutham Rao <goutham.rao@intel.com>
8 * Copyright (C) 2000, 2003 Hewlett-Packard Co
9 * David Mosberger-Tang <davidm@hpl.hp.com>
11 * 03/05/07 davidm Switch from PCI-DMA to generic device DMA API.
12 * 00/12/13 davidm Rename to swiotlb.c and add mark_clean() to avoid
13 * unnecessary i-cache flushing.
14 * 04/07/.. ak Better overflow handling. Assorted fixes.
15 * 05/09/10 linville Add support for syncing ranges, support syncing for
16 * DMA_BIDIRECTIONAL mappings, miscellaneous cleanup.
17 * 08/12/11 beckyb Add highmem support
20 #include <linux/cache.h>
21 #include <linux/dma-mapping.h>
23 #include <linux/export.h>
24 #include <linux/spinlock.h>
25 #include <linux/string.h>
26 #include <linux/swiotlb.h>
27 #include <linux/pfn.h>
28 #include <linux/types.h>
29 #include <linux/ctype.h>
30 #include <linux/highmem.h>
31 #include <linux/gfp.h>
35 #include <asm/scatterlist.h>
37 #include <linux/init.h>
38 #include <linux/bootmem.h>
39 #include <linux/iommu-helper.h>
41 #define OFFSET(val,align) ((unsigned long) \
42 ( (val) & ( (align) - 1)))
44 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
47 * Minimum IO TLB size to bother booting with. Systems with mainly
48 * 64bit capable cards will only lightly use the swiotlb. If we can't
49 * allocate a contiguous 1MB, we're probably in trouble anyway.
51 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
56 * Used to do a quick range check in swiotlb_tbl_unmap_single and
57 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
60 static char *io_tlb_start, *io_tlb_end;
63 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
64 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
66 static unsigned long io_tlb_nslabs;
69 * When the IOMMU overflows we return a fallback buffer. This sets the size.
71 static unsigned long io_tlb_overflow = 32*1024;
73 static void *io_tlb_overflow_buffer;
76 * This is a free list describing the number of free entries available from
79 static unsigned int *io_tlb_list;
80 static unsigned int io_tlb_index;
83 * We need to save away the original address corresponding to a mapped entry
84 * for the sync operations.
86 static phys_addr_t *io_tlb_orig_addr;
89 * Protect the above data structures in the map and unmap calls
91 static DEFINE_SPINLOCK(io_tlb_lock);
93 static int late_alloc;
96 setup_io_tlb_npages(char *str)
99 io_tlb_nslabs = simple_strtoul(str, &str, 0);
100 /* avoid tail segment of size < IO_TLB_SEGSIZE */
101 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
105 if (!strcmp(str, "force"))
110 __setup("swiotlb=", setup_io_tlb_npages);
111 /* make io_tlb_overflow tunable too? */
113 unsigned long swiotlb_nr_tbl(void)
115 return io_tlb_nslabs;
117 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
118 /* Note that this doesn't work with highmem page */
119 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
120 volatile void *address)
122 return phys_to_dma(hwdev, virt_to_phys(address));
125 void swiotlb_print_info(void)
127 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
128 phys_addr_t pstart, pend;
130 pstart = virt_to_phys(io_tlb_start);
131 pend = virt_to_phys(io_tlb_end);
133 printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
134 (unsigned long long)pstart, (unsigned long long)pend - 1,
135 bytes >> 20, io_tlb_start, io_tlb_end - 1);
138 void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
140 unsigned long i, bytes;
142 bytes = nslabs << IO_TLB_SHIFT;
144 io_tlb_nslabs = nslabs;
146 io_tlb_end = io_tlb_start + bytes;
149 * Allocate and initialize the free list array. This array is used
150 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
151 * between io_tlb_start and io_tlb_end.
153 io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
154 for (i = 0; i < io_tlb_nslabs; i++)
155 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
157 io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
160 * Get the overflow emergency buffer
162 io_tlb_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
163 if (!io_tlb_overflow_buffer)
164 panic("Cannot allocate SWIOTLB overflow buffer!\n");
166 swiotlb_print_info();
170 * Statically reserve bounce buffer space and initialize bounce buffer data
171 * structures for the software IO TLB used to implement the DMA API.
174 swiotlb_init_with_default_size(size_t default_size, int verbose)
178 if (!io_tlb_nslabs) {
179 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
180 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
183 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
186 * Get IO TLB memory from the low pages
188 io_tlb_start = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
190 panic("Cannot allocate SWIOTLB buffer");
192 swiotlb_init_with_tbl(io_tlb_start, io_tlb_nslabs, verbose);
196 swiotlb_init(int verbose)
198 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
202 * Systems with larger DMA zones (those that don't support ISA) can
203 * initialize the swiotlb later using the slab allocator if needed.
204 * This should be just like above, but with some error catching.
207 swiotlb_late_init_with_default_size(size_t default_size)
209 unsigned long bytes, req_nslabs = io_tlb_nslabs;
213 if (!io_tlb_nslabs) {
214 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
215 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
219 * Get IO TLB memory from the low pages
221 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
222 io_tlb_nslabs = SLABS_PER_PAGE << order;
223 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
225 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
226 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
234 io_tlb_nslabs = req_nslabs;
237 if (order != get_order(bytes)) {
238 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
239 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
240 io_tlb_nslabs = SLABS_PER_PAGE << order;
242 rc = swiotlb_late_init_with_tbl(io_tlb_start, io_tlb_nslabs);
244 free_pages((unsigned long)io_tlb_start, order);
249 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
251 unsigned long i, bytes;
253 bytes = nslabs << IO_TLB_SHIFT;
255 io_tlb_nslabs = nslabs;
257 io_tlb_end = io_tlb_start + bytes;
259 memset(io_tlb_start, 0, bytes);
262 * Allocate and initialize the free list array. This array is used
263 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
264 * between io_tlb_start and io_tlb_end.
266 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
267 get_order(io_tlb_nslabs * sizeof(int)));
271 for (i = 0; i < io_tlb_nslabs; i++)
272 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
275 io_tlb_orig_addr = (phys_addr_t *)
276 __get_free_pages(GFP_KERNEL,
277 get_order(io_tlb_nslabs *
278 sizeof(phys_addr_t)));
279 if (!io_tlb_orig_addr)
282 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
285 * Get the overflow emergency buffer
287 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
288 get_order(io_tlb_overflow));
289 if (!io_tlb_overflow_buffer)
292 swiotlb_print_info();
299 free_pages((unsigned long)io_tlb_orig_addr,
300 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
301 io_tlb_orig_addr = NULL;
303 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
313 void __init swiotlb_free(void)
315 if (!io_tlb_overflow_buffer)
319 free_pages((unsigned long)io_tlb_overflow_buffer,
320 get_order(io_tlb_overflow));
321 free_pages((unsigned long)io_tlb_orig_addr,
322 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
323 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
325 free_pages((unsigned long)io_tlb_start,
326 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
328 free_bootmem_late(__pa(io_tlb_overflow_buffer),
329 PAGE_ALIGN(io_tlb_overflow));
330 free_bootmem_late(__pa(io_tlb_orig_addr),
331 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
332 free_bootmem_late(__pa(io_tlb_list),
333 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
334 free_bootmem_late(__pa(io_tlb_start),
335 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
340 static int is_swiotlb_buffer(phys_addr_t paddr)
342 return paddr >= virt_to_phys(io_tlb_start) &&
343 paddr < virt_to_phys(io_tlb_end);
347 * Bounce: copy the swiotlb buffer back to the original dma location
349 void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
350 enum dma_data_direction dir)
352 unsigned long pfn = PFN_DOWN(phys);
354 if (PageHighMem(pfn_to_page(pfn))) {
355 /* The buffer does not have a mapping. Map it in and copy */
356 unsigned int offset = phys & ~PAGE_MASK;
362 sz = min_t(size_t, PAGE_SIZE - offset, size);
364 local_irq_save(flags);
365 buffer = kmap_atomic(pfn_to_page(pfn));
366 if (dir == DMA_TO_DEVICE)
367 memcpy(dma_addr, buffer + offset, sz);
369 memcpy(buffer + offset, dma_addr, sz);
370 kunmap_atomic(buffer);
371 local_irq_restore(flags);
379 if (dir == DMA_TO_DEVICE)
380 memcpy(dma_addr, phys_to_virt(phys), size);
382 memcpy(phys_to_virt(phys), dma_addr, size);
385 EXPORT_SYMBOL_GPL(swiotlb_bounce);
387 void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
388 phys_addr_t phys, size_t size,
389 enum dma_data_direction dir)
393 unsigned int nslots, stride, index, wrap;
396 unsigned long offset_slots;
397 unsigned long max_slots;
399 mask = dma_get_seg_boundary(hwdev);
401 tbl_dma_addr &= mask;
403 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
406 * Carefully handle integer overflow which can occur when mask == ~0UL.
409 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
410 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
413 * For mappings greater than a page, we limit the stride (and
414 * hence alignment) to a page size.
416 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
417 if (size > PAGE_SIZE)
418 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
425 * Find suitable number of IO TLB entries size that will fit this
426 * request and allocate a buffer from that IO TLB pool.
428 spin_lock_irqsave(&io_tlb_lock, flags);
429 index = ALIGN(io_tlb_index, stride);
430 if (index >= io_tlb_nslabs)
435 while (iommu_is_span_boundary(index, nslots, offset_slots,
438 if (index >= io_tlb_nslabs)
445 * If we find a slot that indicates we have 'nslots' number of
446 * contiguous buffers, we allocate the buffers from that slot
447 * and mark the entries as '0' indicating unavailable.
449 if (io_tlb_list[index] >= nslots) {
452 for (i = index; i < (int) (index + nslots); i++)
454 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
455 io_tlb_list[i] = ++count;
456 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
459 * Update the indices to avoid searching in the next
462 io_tlb_index = ((index + nslots) < io_tlb_nslabs
463 ? (index + nslots) : 0);
468 if (index >= io_tlb_nslabs)
470 } while (index != wrap);
473 spin_unlock_irqrestore(&io_tlb_lock, flags);
476 spin_unlock_irqrestore(&io_tlb_lock, flags);
479 * Save away the mapping from the original address to the DMA address.
480 * This is needed when we sync the memory. Then we sync the buffer if
483 for (i = 0; i < nslots; i++)
484 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
485 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
486 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
490 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
493 * Allocates bounce buffer and returns its kernel virtual address.
497 map_single(struct device *hwdev, phys_addr_t phys, size_t size,
498 enum dma_data_direction dir)
500 dma_addr_t start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start);
502 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
506 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
509 swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
510 enum dma_data_direction dir)
513 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
514 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
515 phys_addr_t phys = io_tlb_orig_addr[index];
518 * First, sync the memory before unmapping the entry
520 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
521 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
524 * Return the buffer to the free list by setting the corresponding
525 * entries to indicate the number of contiguous entries available.
526 * While returning the entries to the free list, we merge the entries
527 * with slots below and above the pool being returned.
529 spin_lock_irqsave(&io_tlb_lock, flags);
531 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
532 io_tlb_list[index + nslots] : 0);
534 * Step 1: return the slots to the free list, merging the
535 * slots with superceeding slots
537 for (i = index + nslots - 1; i >= index; i--)
538 io_tlb_list[i] = ++count;
540 * Step 2: merge the returned slots with the preceding slots,
541 * if available (non zero)
543 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
544 io_tlb_list[i] = ++count;
546 spin_unlock_irqrestore(&io_tlb_lock, flags);
548 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
551 swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
552 enum dma_data_direction dir,
553 enum dma_sync_target target)
555 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
556 phys_addr_t phys = io_tlb_orig_addr[index];
558 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
562 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
563 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
565 BUG_ON(dir != DMA_TO_DEVICE);
567 case SYNC_FOR_DEVICE:
568 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
569 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
571 BUG_ON(dir != DMA_FROM_DEVICE);
577 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
580 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
581 dma_addr_t *dma_handle, gfp_t flags)
585 int order = get_order(size);
586 u64 dma_mask = DMA_BIT_MASK(32);
588 if (hwdev && hwdev->coherent_dma_mask)
589 dma_mask = hwdev->coherent_dma_mask;
591 ret = (void *)__get_free_pages(flags, order);
592 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
594 * The allocated memory isn't reachable by the device.
596 free_pages((unsigned long) ret, order);
601 * We are either out of memory or the device can't DMA to
602 * GFP_DMA memory; fall back on map_single(), which
603 * will grab memory from the lowest available address range.
605 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
610 memset(ret, 0, size);
611 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
613 /* Confirm address can be DMA'd by device */
614 if (dev_addr + size - 1 > dma_mask) {
615 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
616 (unsigned long long)dma_mask,
617 (unsigned long long)dev_addr);
619 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
620 swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
623 *dma_handle = dev_addr;
626 EXPORT_SYMBOL(swiotlb_alloc_coherent);
629 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
632 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
634 WARN_ON(irqs_disabled());
635 if (!is_swiotlb_buffer(paddr))
636 free_pages((unsigned long)vaddr, get_order(size));
638 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
639 swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
641 EXPORT_SYMBOL(swiotlb_free_coherent);
644 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
648 * Ran out of IOMMU space for this operation. This is very bad.
649 * Unfortunately the drivers cannot handle this operation properly.
650 * unless they check for dma_mapping_error (most don't)
651 * When the mapping is small enough return a static buffer to limit
652 * the damage, or panic when the transfer is too big.
654 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
655 "device %s\n", size, dev ? dev_name(dev) : "?");
657 if (size <= io_tlb_overflow || !do_panic)
660 if (dir == DMA_BIDIRECTIONAL)
661 panic("DMA: Random memory could be DMA accessed\n");
662 if (dir == DMA_FROM_DEVICE)
663 panic("DMA: Random memory could be DMA written\n");
664 if (dir == DMA_TO_DEVICE)
665 panic("DMA: Random memory could be DMA read\n");
669 * Map a single buffer of the indicated size for DMA in streaming mode. The
670 * physical address to use is returned.
672 * Once the device is given the dma address, the device owns this memory until
673 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
675 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
676 unsigned long offset, size_t size,
677 enum dma_data_direction dir,
678 struct dma_attrs *attrs)
680 phys_addr_t phys = page_to_phys(page) + offset;
681 dma_addr_t dev_addr = phys_to_dma(dev, phys);
684 BUG_ON(dir == DMA_NONE);
686 * If the address happens to be in the device's DMA window,
687 * we can safely return the device addr and not worry about bounce
690 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
694 * Oh well, have to allocate and map a bounce buffer.
696 map = map_single(dev, phys, size, dir);
698 swiotlb_full(dev, size, dir, 1);
699 map = io_tlb_overflow_buffer;
702 dev_addr = swiotlb_virt_to_bus(dev, map);
705 * Ensure that the address returned is DMA'ble
707 if (!dma_capable(dev, dev_addr, size)) {
708 swiotlb_tbl_unmap_single(dev, map, size, dir);
709 dev_addr = swiotlb_virt_to_bus(dev, io_tlb_overflow_buffer);
714 EXPORT_SYMBOL_GPL(swiotlb_map_page);
717 * Unmap a single streaming mode DMA translation. The dma_addr and size must
718 * match what was provided for in a previous swiotlb_map_page call. All
719 * other usages are undefined.
721 * After this call, reads by the cpu to the buffer are guaranteed to see
722 * whatever the device wrote there.
724 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
725 size_t size, enum dma_data_direction dir)
727 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
729 BUG_ON(dir == DMA_NONE);
731 if (is_swiotlb_buffer(paddr)) {
732 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
736 if (dir != DMA_FROM_DEVICE)
740 * phys_to_virt doesn't work with hihgmem page but we could
741 * call dma_mark_clean() with hihgmem page here. However, we
742 * are fine since dma_mark_clean() is null on POWERPC. We can
743 * make dma_mark_clean() take a physical address if necessary.
745 dma_mark_clean(phys_to_virt(paddr), size);
748 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
749 size_t size, enum dma_data_direction dir,
750 struct dma_attrs *attrs)
752 unmap_single(hwdev, dev_addr, size, dir);
754 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
757 * Make physical memory consistent for a single streaming mode DMA translation
760 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
761 * using the cpu, yet do not wish to teardown the dma mapping, you must
762 * call this function before doing so. At the next point you give the dma
763 * address back to the card, you must first perform a
764 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
767 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
768 size_t size, enum dma_data_direction dir,
769 enum dma_sync_target target)
771 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
773 BUG_ON(dir == DMA_NONE);
775 if (is_swiotlb_buffer(paddr)) {
776 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
781 if (dir != DMA_FROM_DEVICE)
784 dma_mark_clean(phys_to_virt(paddr), size);
788 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
789 size_t size, enum dma_data_direction dir)
791 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
793 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
796 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
797 size_t size, enum dma_data_direction dir)
799 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
801 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
804 * Map a set of buffers described by scatterlist in streaming mode for DMA.
805 * This is the scatter-gather version of the above swiotlb_map_page
806 * interface. Here the scatter gather list elements are each tagged with the
807 * appropriate dma address and length. They are obtained via
808 * sg_dma_{address,length}(SG).
810 * NOTE: An implementation may be able to use a smaller number of
811 * DMA address/length pairs than there are SG table elements.
812 * (for example via virtual mapping capabilities)
813 * The routine returns the number of addr/length pairs actually
814 * used, at most nents.
816 * Device ownership issues as mentioned above for swiotlb_map_page are the
820 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
821 enum dma_data_direction dir, struct dma_attrs *attrs)
823 struct scatterlist *sg;
826 BUG_ON(dir == DMA_NONE);
828 for_each_sg(sgl, sg, nelems, i) {
829 phys_addr_t paddr = sg_phys(sg);
830 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
833 !dma_capable(hwdev, dev_addr, sg->length)) {
834 void *map = map_single(hwdev, sg_phys(sg),
837 /* Don't panic here, we expect map_sg users
838 to do proper error handling. */
839 swiotlb_full(hwdev, sg->length, dir, 0);
840 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
842 sgl[0].dma_length = 0;
845 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
847 sg->dma_address = dev_addr;
848 sg->dma_length = sg->length;
852 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
855 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
856 enum dma_data_direction dir)
858 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
860 EXPORT_SYMBOL(swiotlb_map_sg);
863 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
864 * concerning calls here are the same as for swiotlb_unmap_page() above.
867 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
868 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
870 struct scatterlist *sg;
873 BUG_ON(dir == DMA_NONE);
875 for_each_sg(sgl, sg, nelems, i)
876 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
879 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
882 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
883 enum dma_data_direction dir)
885 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
887 EXPORT_SYMBOL(swiotlb_unmap_sg);
890 * Make physical memory consistent for a set of streaming mode DMA translations
893 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
897 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
898 int nelems, enum dma_data_direction dir,
899 enum dma_sync_target target)
901 struct scatterlist *sg;
904 for_each_sg(sgl, sg, nelems, i)
905 swiotlb_sync_single(hwdev, sg->dma_address,
906 sg->dma_length, dir, target);
910 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
911 int nelems, enum dma_data_direction dir)
913 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
915 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
918 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
919 int nelems, enum dma_data_direction dir)
921 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
923 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
926 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
928 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
930 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
933 * Return whether the given device DMA address mask can be supported
934 * properly. For example, if your device can only drive the low 24-bits
935 * during bus mastering, then you would pass 0x00ffffff as the mask to
939 swiotlb_dma_supported(struct device *hwdev, u64 mask)
941 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
943 EXPORT_SYMBOL(swiotlb_dma_supported);