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;
61 static phys_addr_t io_tlb_end;
64 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
65 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
67 static unsigned long io_tlb_nslabs;
70 * When the IOMMU overflows we return a fallback buffer. This sets the size.
72 static unsigned long io_tlb_overflow = 32*1024;
74 static void *io_tlb_overflow_buffer;
77 * This is a free list describing the number of free entries available from
80 static unsigned int *io_tlb_list;
81 static unsigned int io_tlb_index;
84 * We need to save away the original address corresponding to a mapped entry
85 * for the sync operations.
87 static phys_addr_t *io_tlb_orig_addr;
90 * Protect the above data structures in the map and unmap calls
92 static DEFINE_SPINLOCK(io_tlb_lock);
94 static int late_alloc;
97 setup_io_tlb_npages(char *str)
100 io_tlb_nslabs = simple_strtoul(str, &str, 0);
101 /* avoid tail segment of size < IO_TLB_SEGSIZE */
102 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
106 if (!strcmp(str, "force"))
111 __setup("swiotlb=", setup_io_tlb_npages);
112 /* make io_tlb_overflow tunable too? */
114 unsigned long swiotlb_nr_tbl(void)
116 return io_tlb_nslabs;
118 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
119 /* Note that this doesn't work with highmem page */
120 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
121 volatile void *address)
123 return phys_to_dma(hwdev, virt_to_phys(address));
126 void swiotlb_print_info(void)
128 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
132 pstart = virt_to_phys(io_tlb_start);
133 vend = phys_to_virt(io_tlb_end);
135 printk(KERN_INFO "software IO TLB [mem %#010llx-%#010llx] (%luMB) mapped at [%p-%p]\n",
136 (unsigned long long)pstart,
137 (unsigned long long)io_tlb_end,
138 bytes >> 20, io_tlb_start, vend - 1);
141 void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
143 unsigned long i, bytes;
145 bytes = nslabs << IO_TLB_SHIFT;
147 io_tlb_nslabs = nslabs;
149 io_tlb_end = __pa(io_tlb_start) + bytes;
152 * Allocate and initialize the free list array. This array is used
153 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
154 * between io_tlb_start and io_tlb_end.
156 io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
157 for (i = 0; i < io_tlb_nslabs; i++)
158 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
160 io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
163 * Get the overflow emergency buffer
165 io_tlb_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
166 if (!io_tlb_overflow_buffer)
167 panic("Cannot allocate SWIOTLB overflow buffer!\n");
169 swiotlb_print_info();
173 * Statically reserve bounce buffer space and initialize bounce buffer data
174 * structures for the software IO TLB used to implement the DMA API.
177 swiotlb_init_with_default_size(size_t default_size, int verbose)
181 if (!io_tlb_nslabs) {
182 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
183 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
186 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
189 * Get IO TLB memory from the low pages
191 io_tlb_start = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
193 panic("Cannot allocate SWIOTLB buffer");
195 swiotlb_init_with_tbl(io_tlb_start, io_tlb_nslabs, verbose);
199 swiotlb_init(int verbose)
201 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
205 * Systems with larger DMA zones (those that don't support ISA) can
206 * initialize the swiotlb later using the slab allocator if needed.
207 * This should be just like above, but with some error catching.
210 swiotlb_late_init_with_default_size(size_t default_size)
212 unsigned long 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,
237 io_tlb_nslabs = req_nslabs;
240 if (order != get_order(bytes)) {
241 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
242 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
243 io_tlb_nslabs = SLABS_PER_PAGE << order;
245 rc = swiotlb_late_init_with_tbl(io_tlb_start, io_tlb_nslabs);
247 free_pages((unsigned long)io_tlb_start, order);
252 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
254 unsigned long i, bytes;
256 bytes = nslabs << IO_TLB_SHIFT;
258 io_tlb_nslabs = nslabs;
260 io_tlb_end = virt_to_phys(io_tlb_start) + bytes;
262 memset(io_tlb_start, 0, bytes);
265 * Allocate and initialize the free list array. This array is used
266 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
267 * between io_tlb_start and io_tlb_end.
269 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
270 get_order(io_tlb_nslabs * sizeof(int)));
274 for (i = 0; i < io_tlb_nslabs; i++)
275 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
278 io_tlb_orig_addr = (phys_addr_t *)
279 __get_free_pages(GFP_KERNEL,
280 get_order(io_tlb_nslabs *
281 sizeof(phys_addr_t)));
282 if (!io_tlb_orig_addr)
285 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
288 * Get the overflow emergency buffer
290 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
291 get_order(io_tlb_overflow));
292 if (!io_tlb_overflow_buffer)
295 swiotlb_print_info();
302 free_pages((unsigned long)io_tlb_orig_addr,
303 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
304 io_tlb_orig_addr = NULL;
306 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
316 void __init swiotlb_free(void)
318 if (!io_tlb_overflow_buffer)
322 free_pages((unsigned long)io_tlb_overflow_buffer,
323 get_order(io_tlb_overflow));
324 free_pages((unsigned long)io_tlb_orig_addr,
325 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
326 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
328 free_pages((unsigned long)io_tlb_start,
329 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
331 free_bootmem_late(__pa(io_tlb_overflow_buffer),
332 PAGE_ALIGN(io_tlb_overflow));
333 free_bootmem_late(__pa(io_tlb_orig_addr),
334 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
335 free_bootmem_late(__pa(io_tlb_list),
336 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
337 free_bootmem_late(__pa(io_tlb_start),
338 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
343 static int is_swiotlb_buffer(phys_addr_t paddr)
345 return paddr >= virt_to_phys(io_tlb_start) && paddr < io_tlb_end;
349 * Bounce: copy the swiotlb buffer back to the original dma location
351 void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
352 enum dma_data_direction dir)
354 unsigned long pfn = PFN_DOWN(phys);
356 if (PageHighMem(pfn_to_page(pfn))) {
357 /* The buffer does not have a mapping. Map it in and copy */
358 unsigned int offset = phys & ~PAGE_MASK;
364 sz = min_t(size_t, PAGE_SIZE - offset, size);
366 local_irq_save(flags);
367 buffer = kmap_atomic(pfn_to_page(pfn));
368 if (dir == DMA_TO_DEVICE)
369 memcpy(dma_addr, buffer + offset, sz);
371 memcpy(buffer + offset, dma_addr, sz);
372 kunmap_atomic(buffer);
373 local_irq_restore(flags);
381 if (dir == DMA_TO_DEVICE)
382 memcpy(dma_addr, phys_to_virt(phys), size);
384 memcpy(phys_to_virt(phys), dma_addr, size);
387 EXPORT_SYMBOL_GPL(swiotlb_bounce);
389 void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
390 phys_addr_t phys, size_t size,
391 enum dma_data_direction dir)
395 unsigned int nslots, stride, index, wrap;
398 unsigned long offset_slots;
399 unsigned long max_slots;
401 mask = dma_get_seg_boundary(hwdev);
403 tbl_dma_addr &= mask;
405 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
408 * Carefully handle integer overflow which can occur when mask == ~0UL.
411 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
412 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
415 * For mappings greater than a page, we limit the stride (and
416 * hence alignment) to a page size.
418 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
419 if (size > PAGE_SIZE)
420 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
427 * Find suitable number of IO TLB entries size that will fit this
428 * request and allocate a buffer from that IO TLB pool.
430 spin_lock_irqsave(&io_tlb_lock, flags);
431 index = ALIGN(io_tlb_index, stride);
432 if (index >= io_tlb_nslabs)
437 while (iommu_is_span_boundary(index, nslots, offset_slots,
440 if (index >= io_tlb_nslabs)
447 * If we find a slot that indicates we have 'nslots' number of
448 * contiguous buffers, we allocate the buffers from that slot
449 * and mark the entries as '0' indicating unavailable.
451 if (io_tlb_list[index] >= nslots) {
454 for (i = index; i < (int) (index + nslots); i++)
456 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
457 io_tlb_list[i] = ++count;
458 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
461 * Update the indices to avoid searching in the next
464 io_tlb_index = ((index + nslots) < io_tlb_nslabs
465 ? (index + nslots) : 0);
470 if (index >= io_tlb_nslabs)
472 } while (index != wrap);
475 spin_unlock_irqrestore(&io_tlb_lock, flags);
478 spin_unlock_irqrestore(&io_tlb_lock, flags);
481 * Save away the mapping from the original address to the DMA address.
482 * This is needed when we sync the memory. Then we sync the buffer if
485 for (i = 0; i < nslots; i++)
486 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
487 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
488 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
492 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
495 * Allocates bounce buffer and returns its kernel virtual address.
499 map_single(struct device *hwdev, phys_addr_t phys, size_t size,
500 enum dma_data_direction dir)
502 dma_addr_t start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start);
504 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
508 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
511 swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
512 enum dma_data_direction dir)
515 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
516 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
517 phys_addr_t phys = io_tlb_orig_addr[index];
520 * First, sync the memory before unmapping the entry
522 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
523 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
526 * Return the buffer to the free list by setting the corresponding
527 * entries to indicate the number of contiguous entries available.
528 * While returning the entries to the free list, we merge the entries
529 * with slots below and above the pool being returned.
531 spin_lock_irqsave(&io_tlb_lock, flags);
533 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
534 io_tlb_list[index + nslots] : 0);
536 * Step 1: return the slots to the free list, merging the
537 * slots with superceeding slots
539 for (i = index + nslots - 1; i >= index; i--)
540 io_tlb_list[i] = ++count;
542 * Step 2: merge the returned slots with the preceding slots,
543 * if available (non zero)
545 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
546 io_tlb_list[i] = ++count;
548 spin_unlock_irqrestore(&io_tlb_lock, flags);
550 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
553 swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
554 enum dma_data_direction dir,
555 enum dma_sync_target target)
557 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
558 phys_addr_t phys = io_tlb_orig_addr[index];
560 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
564 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
565 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
567 BUG_ON(dir != DMA_TO_DEVICE);
569 case SYNC_FOR_DEVICE:
570 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
571 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
573 BUG_ON(dir != DMA_FROM_DEVICE);
579 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
582 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
583 dma_addr_t *dma_handle, gfp_t flags)
587 int order = get_order(size);
588 u64 dma_mask = DMA_BIT_MASK(32);
590 if (hwdev && hwdev->coherent_dma_mask)
591 dma_mask = hwdev->coherent_dma_mask;
593 ret = (void *)__get_free_pages(flags, order);
594 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
596 * The allocated memory isn't reachable by the device.
598 free_pages((unsigned long) ret, order);
603 * We are either out of memory or the device can't DMA to
604 * GFP_DMA memory; fall back on map_single(), which
605 * will grab memory from the lowest available address range.
607 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
612 memset(ret, 0, size);
613 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
615 /* Confirm address can be DMA'd by device */
616 if (dev_addr + size - 1 > dma_mask) {
617 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
618 (unsigned long long)dma_mask,
619 (unsigned long long)dev_addr);
621 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
622 swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
625 *dma_handle = dev_addr;
628 EXPORT_SYMBOL(swiotlb_alloc_coherent);
631 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
634 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
636 WARN_ON(irqs_disabled());
637 if (!is_swiotlb_buffer(paddr))
638 free_pages((unsigned long)vaddr, get_order(size));
640 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
641 swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
643 EXPORT_SYMBOL(swiotlb_free_coherent);
646 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
650 * Ran out of IOMMU space for this operation. This is very bad.
651 * Unfortunately the drivers cannot handle this operation properly.
652 * unless they check for dma_mapping_error (most don't)
653 * When the mapping is small enough return a static buffer to limit
654 * the damage, or panic when the transfer is too big.
656 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
657 "device %s\n", size, dev ? dev_name(dev) : "?");
659 if (size <= io_tlb_overflow || !do_panic)
662 if (dir == DMA_BIDIRECTIONAL)
663 panic("DMA: Random memory could be DMA accessed\n");
664 if (dir == DMA_FROM_DEVICE)
665 panic("DMA: Random memory could be DMA written\n");
666 if (dir == DMA_TO_DEVICE)
667 panic("DMA: Random memory could be DMA read\n");
671 * Map a single buffer of the indicated size for DMA in streaming mode. The
672 * physical address to use is returned.
674 * Once the device is given the dma address, the device owns this memory until
675 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
677 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
678 unsigned long offset, size_t size,
679 enum dma_data_direction dir,
680 struct dma_attrs *attrs)
682 phys_addr_t phys = page_to_phys(page) + offset;
683 dma_addr_t dev_addr = phys_to_dma(dev, phys);
686 BUG_ON(dir == DMA_NONE);
688 * If the address happens to be in the device's DMA window,
689 * we can safely return the device addr and not worry about bounce
692 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
696 * Oh well, have to allocate and map a bounce buffer.
698 map = map_single(dev, phys, size, dir);
700 swiotlb_full(dev, size, dir, 1);
701 map = io_tlb_overflow_buffer;
704 dev_addr = swiotlb_virt_to_bus(dev, map);
707 * Ensure that the address returned is DMA'ble
709 if (!dma_capable(dev, dev_addr, size)) {
710 swiotlb_tbl_unmap_single(dev, map, size, dir);
711 dev_addr = swiotlb_virt_to_bus(dev, io_tlb_overflow_buffer);
716 EXPORT_SYMBOL_GPL(swiotlb_map_page);
719 * Unmap a single streaming mode DMA translation. The dma_addr and size must
720 * match what was provided for in a previous swiotlb_map_page call. All
721 * other usages are undefined.
723 * After this call, reads by the cpu to the buffer are guaranteed to see
724 * whatever the device wrote there.
726 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
727 size_t size, enum dma_data_direction dir)
729 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
731 BUG_ON(dir == DMA_NONE);
733 if (is_swiotlb_buffer(paddr)) {
734 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
738 if (dir != DMA_FROM_DEVICE)
742 * phys_to_virt doesn't work with hihgmem page but we could
743 * call dma_mark_clean() with hihgmem page here. However, we
744 * are fine since dma_mark_clean() is null on POWERPC. We can
745 * make dma_mark_clean() take a physical address if necessary.
747 dma_mark_clean(phys_to_virt(paddr), size);
750 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
751 size_t size, enum dma_data_direction dir,
752 struct dma_attrs *attrs)
754 unmap_single(hwdev, dev_addr, size, dir);
756 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
759 * Make physical memory consistent for a single streaming mode DMA translation
762 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
763 * using the cpu, yet do not wish to teardown the dma mapping, you must
764 * call this function before doing so. At the next point you give the dma
765 * address back to the card, you must first perform a
766 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
769 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
770 size_t size, enum dma_data_direction dir,
771 enum dma_sync_target target)
773 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
775 BUG_ON(dir == DMA_NONE);
777 if (is_swiotlb_buffer(paddr)) {
778 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
783 if (dir != DMA_FROM_DEVICE)
786 dma_mark_clean(phys_to_virt(paddr), size);
790 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
791 size_t size, enum dma_data_direction dir)
793 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
795 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
798 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
799 size_t size, enum dma_data_direction dir)
801 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
803 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
806 * Map a set of buffers described by scatterlist in streaming mode for DMA.
807 * This is the scatter-gather version of the above swiotlb_map_page
808 * interface. Here the scatter gather list elements are each tagged with the
809 * appropriate dma address and length. They are obtained via
810 * sg_dma_{address,length}(SG).
812 * NOTE: An implementation may be able to use a smaller number of
813 * DMA address/length pairs than there are SG table elements.
814 * (for example via virtual mapping capabilities)
815 * The routine returns the number of addr/length pairs actually
816 * used, at most nents.
818 * Device ownership issues as mentioned above for swiotlb_map_page are the
822 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
823 enum dma_data_direction dir, struct dma_attrs *attrs)
825 struct scatterlist *sg;
828 BUG_ON(dir == DMA_NONE);
830 for_each_sg(sgl, sg, nelems, i) {
831 phys_addr_t paddr = sg_phys(sg);
832 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
835 !dma_capable(hwdev, dev_addr, sg->length)) {
836 void *map = map_single(hwdev, sg_phys(sg),
839 /* Don't panic here, we expect map_sg users
840 to do proper error handling. */
841 swiotlb_full(hwdev, sg->length, dir, 0);
842 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
844 sgl[0].dma_length = 0;
847 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
849 sg->dma_address = dev_addr;
850 sg->dma_length = sg->length;
854 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
857 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
858 enum dma_data_direction dir)
860 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
862 EXPORT_SYMBOL(swiotlb_map_sg);
865 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
866 * concerning calls here are the same as for swiotlb_unmap_page() above.
869 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
870 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
872 struct scatterlist *sg;
875 BUG_ON(dir == DMA_NONE);
877 for_each_sg(sgl, sg, nelems, i)
878 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
881 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
884 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
885 enum dma_data_direction dir)
887 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
889 EXPORT_SYMBOL(swiotlb_unmap_sg);
892 * Make physical memory consistent for a set of streaming mode DMA translations
895 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
899 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
900 int nelems, enum dma_data_direction dir,
901 enum dma_sync_target target)
903 struct scatterlist *sg;
906 for_each_sg(sgl, sg, nelems, i)
907 swiotlb_sync_single(hwdev, sg->dma_address,
908 sg->dma_length, dir, target);
912 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
913 int nelems, enum dma_data_direction dir)
915 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
917 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
920 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
921 int nelems, enum dma_data_direction dir)
923 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
925 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
928 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
930 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
932 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
935 * Return whether the given device DMA address mask can be supported
936 * properly. For example, if your device can only drive the low 24-bits
937 * during bus mastering, then you would pass 0x00ffffff as the mask to
941 swiotlb_dma_supported(struct device *hwdev, u64 mask)
943 return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
945 EXPORT_SYMBOL(swiotlb_dma_supported);