2 * Dynamic DMA mapping support.
4 * This implementation is a fallback for platforms that do not support
5 * I/O TLBs (aka DMA address translation hardware).
6 * Copyright (C) 2000 Asit Mallick <Asit.K.Mallick@intel.com>
7 * Copyright (C) 2000 Goutham Rao <goutham.rao@intel.com>
8 * Copyright (C) 2000, 2003 Hewlett-Packard Co
9 * David Mosberger-Tang <davidm@hpl.hp.com>
11 * 03/05/07 davidm Switch from PCI-DMA to generic device DMA API.
12 * 00/12/13 davidm Rename to swiotlb.c and add mark_clean() to avoid
13 * unnecessary i-cache flushing.
14 * 04/07/.. ak Better overflow handling. Assorted fixes.
15 * 05/09/10 linville Add support for syncing ranges, support syncing for
16 * DMA_BIDIRECTIONAL mappings, miscellaneous cleanup.
17 * 08/12/11 beckyb Add highmem support
20 #include <linux/cache.h>
21 #include <linux/dma-mapping.h>
23 #include <linux/module.h>
24 #include <linux/spinlock.h>
25 #include <linux/string.h>
26 #include <linux/swiotlb.h>
27 #include <linux/pfn.h>
28 #include <linux/types.h>
29 #include <linux/ctype.h>
30 #include <linux/highmem.h>
31 #include <linux/gfp.h>
35 #include <asm/scatterlist.h>
37 #include <linux/init.h>
38 #include <linux/bootmem.h>
39 #include <linux/iommu-helper.h>
41 #define OFFSET(val,align) ((unsigned long) \
42 ( (val) & ( (align) - 1)))
44 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
47 * Minimum IO TLB size to bother booting with. Systems with mainly
48 * 64bit capable cards will only lightly use the swiotlb. If we can't
49 * allocate a contiguous 1MB, we're probably in trouble anyway.
51 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
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 "Placing %luMB software IO TLB between %p - %p\n",
134 bytes >> 20, io_tlb_start, io_tlb_end);
135 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
136 (unsigned long long)pstart,
137 (unsigned long long)pend);
140 void __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
142 unsigned long i, bytes;
144 bytes = nslabs << IO_TLB_SHIFT;
146 io_tlb_nslabs = nslabs;
148 io_tlb_end = io_tlb_start + bytes;
151 * Allocate and initialize the free list array. This array is used
152 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
153 * between io_tlb_start and io_tlb_end.
155 io_tlb_list = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
156 for (i = 0; i < io_tlb_nslabs; i++)
157 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
159 io_tlb_orig_addr = alloc_bootmem_pages(PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
162 * Get the overflow emergency buffer
164 io_tlb_overflow_buffer = alloc_bootmem_low_pages(PAGE_ALIGN(io_tlb_overflow));
165 if (!io_tlb_overflow_buffer)
166 panic("Cannot allocate SWIOTLB overflow buffer!\n");
168 swiotlb_print_info();
172 * Statically reserve bounce buffer space and initialize bounce buffer data
173 * structures for the software IO TLB used to implement the DMA API.
176 swiotlb_init_with_default_size(size_t default_size, int verbose)
180 if (!io_tlb_nslabs) {
181 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
182 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
185 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
188 * Get IO TLB memory from the low pages
190 io_tlb_start = alloc_bootmem_low_pages(PAGE_ALIGN(bytes));
192 panic("Cannot allocate SWIOTLB buffer");
194 swiotlb_init_with_tbl(io_tlb_start, io_tlb_nslabs, verbose);
198 swiotlb_init(int verbose)
200 swiotlb_init_with_default_size(64 * (1<<20), verbose); /* default to 64MB */
204 * Systems with larger DMA zones (those that don't support ISA) can
205 * initialize the swiotlb later using the slab allocator if needed.
206 * This should be just like above, but with some error catching.
209 swiotlb_late_init_with_default_size(size_t default_size)
211 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
214 if (!io_tlb_nslabs) {
215 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
216 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
220 * Get IO TLB memory from the low pages
222 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
223 io_tlb_nslabs = SLABS_PER_PAGE << order;
224 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
226 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
227 io_tlb_start = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
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;
241 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
243 io_tlb_end = io_tlb_start + bytes;
244 memset(io_tlb_start, 0, bytes);
247 * Allocate and initialize the free list array. This array is used
248 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
249 * between io_tlb_start and io_tlb_end.
251 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
252 get_order(io_tlb_nslabs * sizeof(int)));
256 for (i = 0; i < io_tlb_nslabs; i++)
257 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
260 io_tlb_orig_addr = (phys_addr_t *)
261 __get_free_pages(GFP_KERNEL,
262 get_order(io_tlb_nslabs *
263 sizeof(phys_addr_t)));
264 if (!io_tlb_orig_addr)
267 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(phys_addr_t));
270 * Get the overflow emergency buffer
272 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
273 get_order(io_tlb_overflow));
274 if (!io_tlb_overflow_buffer)
277 swiotlb_print_info();
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;
300 void __init swiotlb_free(void)
302 if (!io_tlb_overflow_buffer)
306 free_pages((unsigned long)io_tlb_overflow_buffer,
307 get_order(io_tlb_overflow));
308 free_pages((unsigned long)io_tlb_orig_addr,
309 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
310 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
312 free_pages((unsigned long)io_tlb_start,
313 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
315 free_bootmem_late(__pa(io_tlb_overflow_buffer),
316 PAGE_ALIGN(io_tlb_overflow));
317 free_bootmem_late(__pa(io_tlb_orig_addr),
318 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
319 free_bootmem_late(__pa(io_tlb_list),
320 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
321 free_bootmem_late(__pa(io_tlb_start),
322 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
327 static int is_swiotlb_buffer(phys_addr_t paddr)
329 return paddr >= virt_to_phys(io_tlb_start) &&
330 paddr < virt_to_phys(io_tlb_end);
334 * Bounce: copy the swiotlb buffer back to the original dma location
336 void swiotlb_bounce(phys_addr_t phys, char *dma_addr, size_t size,
337 enum dma_data_direction dir)
339 unsigned long pfn = PFN_DOWN(phys);
341 if (PageHighMem(pfn_to_page(pfn))) {
342 /* The buffer does not have a mapping. Map it in and copy */
343 unsigned int offset = phys & ~PAGE_MASK;
349 sz = min_t(size_t, PAGE_SIZE - offset, size);
351 local_irq_save(flags);
352 buffer = kmap_atomic(pfn_to_page(pfn),
354 if (dir == DMA_TO_DEVICE)
355 memcpy(dma_addr, buffer + offset, sz);
357 memcpy(buffer + offset, dma_addr, sz);
358 kunmap_atomic(buffer, KM_BOUNCE_READ);
359 local_irq_restore(flags);
367 if (dir == DMA_TO_DEVICE)
368 memcpy(dma_addr, phys_to_virt(phys), size);
370 memcpy(phys_to_virt(phys), dma_addr, size);
373 EXPORT_SYMBOL_GPL(swiotlb_bounce);
375 void *swiotlb_tbl_map_single(struct device *hwdev, dma_addr_t tbl_dma_addr,
376 phys_addr_t phys, size_t size,
377 enum dma_data_direction dir)
381 unsigned int nslots, stride, index, wrap;
384 unsigned long offset_slots;
385 unsigned long max_slots;
387 mask = dma_get_seg_boundary(hwdev);
389 tbl_dma_addr &= mask;
391 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
394 * Carefully handle integer overflow which can occur when mask == ~0UL.
397 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
398 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
401 * For mappings greater than a page, we limit the stride (and
402 * hence alignment) to a page size.
404 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
405 if (size > PAGE_SIZE)
406 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
413 * Find suitable number of IO TLB entries size that will fit this
414 * request and allocate a buffer from that IO TLB pool.
416 spin_lock_irqsave(&io_tlb_lock, flags);
417 index = ALIGN(io_tlb_index, stride);
418 if (index >= io_tlb_nslabs)
423 while (iommu_is_span_boundary(index, nslots, offset_slots,
426 if (index >= io_tlb_nslabs)
433 * If we find a slot that indicates we have 'nslots' number of
434 * contiguous buffers, we allocate the buffers from that slot
435 * and mark the entries as '0' indicating unavailable.
437 if (io_tlb_list[index] >= nslots) {
440 for (i = index; i < (int) (index + nslots); i++)
442 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
443 io_tlb_list[i] = ++count;
444 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
447 * Update the indices to avoid searching in the next
450 io_tlb_index = ((index + nslots) < io_tlb_nslabs
451 ? (index + nslots) : 0);
456 if (index >= io_tlb_nslabs)
458 } while (index != wrap);
461 spin_unlock_irqrestore(&io_tlb_lock, flags);
464 spin_unlock_irqrestore(&io_tlb_lock, flags);
467 * Save away the mapping from the original address to the DMA address.
468 * This is needed when we sync the memory. Then we sync the buffer if
471 for (i = 0; i < nslots; i++)
472 io_tlb_orig_addr[index+i] = phys + (i << IO_TLB_SHIFT);
473 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
474 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
478 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
481 * Allocates bounce buffer and returns its kernel virtual address.
485 map_single(struct device *hwdev, phys_addr_t phys, size_t size,
486 enum dma_data_direction dir)
488 dma_addr_t start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start);
490 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
494 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
497 swiotlb_tbl_unmap_single(struct device *hwdev, char *dma_addr, size_t size,
498 enum dma_data_direction dir)
501 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
502 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
503 phys_addr_t phys = io_tlb_orig_addr[index];
506 * First, sync the memory before unmapping the entry
508 if (phys && ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
509 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
512 * Return the buffer to the free list by setting the corresponding
513 * entries to indicate the number of contiguous entries available.
514 * While returning the entries to the free list, we merge the entries
515 * with slots below and above the pool being returned.
517 spin_lock_irqsave(&io_tlb_lock, flags);
519 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
520 io_tlb_list[index + nslots] : 0);
522 * Step 1: return the slots to the free list, merging the
523 * slots with superceeding slots
525 for (i = index + nslots - 1; i >= index; i--)
526 io_tlb_list[i] = ++count;
528 * Step 2: merge the returned slots with the preceding slots,
529 * if available (non zero)
531 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
532 io_tlb_list[i] = ++count;
534 spin_unlock_irqrestore(&io_tlb_lock, flags);
536 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
539 swiotlb_tbl_sync_single(struct device *hwdev, char *dma_addr, size_t size,
540 enum dma_data_direction dir,
541 enum dma_sync_target target)
543 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
544 phys_addr_t phys = io_tlb_orig_addr[index];
546 phys += ((unsigned long)dma_addr & ((1 << IO_TLB_SHIFT) - 1));
550 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
551 swiotlb_bounce(phys, dma_addr, size, DMA_FROM_DEVICE);
553 BUG_ON(dir != DMA_TO_DEVICE);
555 case SYNC_FOR_DEVICE:
556 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
557 swiotlb_bounce(phys, dma_addr, size, DMA_TO_DEVICE);
559 BUG_ON(dir != DMA_FROM_DEVICE);
565 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
568 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
569 dma_addr_t *dma_handle, gfp_t flags)
573 int order = get_order(size);
574 u64 dma_mask = DMA_BIT_MASK(32);
576 if (hwdev && hwdev->coherent_dma_mask)
577 dma_mask = hwdev->coherent_dma_mask;
579 ret = (void *)__get_free_pages(flags, order);
580 if (ret && swiotlb_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
582 * The allocated memory isn't reachable by the device.
584 free_pages((unsigned long) ret, order);
589 * We are either out of memory or the device can't DMA to
590 * GFP_DMA memory; fall back on map_single(), which
591 * will grab memory from the lowest available address range.
593 ret = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
598 memset(ret, 0, size);
599 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
601 /* Confirm address can be DMA'd by device */
602 if (dev_addr + size - 1 > dma_mask) {
603 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
604 (unsigned long long)dma_mask,
605 (unsigned long long)dev_addr);
607 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
608 swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
611 *dma_handle = dev_addr;
614 EXPORT_SYMBOL(swiotlb_alloc_coherent);
617 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
620 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
622 WARN_ON(irqs_disabled());
623 if (!is_swiotlb_buffer(paddr))
624 free_pages((unsigned long)vaddr, get_order(size));
626 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
627 swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
629 EXPORT_SYMBOL(swiotlb_free_coherent);
632 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
636 * Ran out of IOMMU space for this operation. This is very bad.
637 * Unfortunately the drivers cannot handle this operation properly.
638 * unless they check for dma_mapping_error (most don't)
639 * When the mapping is small enough return a static buffer to limit
640 * the damage, or panic when the transfer is too big.
642 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
643 "device %s\n", size, dev ? dev_name(dev) : "?");
645 if (size <= io_tlb_overflow || !do_panic)
648 if (dir == DMA_BIDIRECTIONAL)
649 panic("DMA: Random memory could be DMA accessed\n");
650 if (dir == DMA_FROM_DEVICE)
651 panic("DMA: Random memory could be DMA written\n");
652 if (dir == DMA_TO_DEVICE)
653 panic("DMA: Random memory could be DMA read\n");
657 * Map a single buffer of the indicated size for DMA in streaming mode. The
658 * physical address to use is returned.
660 * Once the device is given the dma address, the device owns this memory until
661 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
663 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
664 unsigned long offset, size_t size,
665 enum dma_data_direction dir,
666 struct dma_attrs *attrs)
668 phys_addr_t phys = page_to_phys(page) + offset;
669 dma_addr_t dev_addr = phys_to_dma(dev, phys);
672 BUG_ON(dir == DMA_NONE);
674 * If the address happens to be in the device's DMA window,
675 * we can safely return the device addr and not worry about bounce
678 if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
682 * Oh well, have to allocate and map a bounce buffer.
684 map = map_single(dev, phys, size, dir);
686 swiotlb_full(dev, size, dir, 1);
687 map = io_tlb_overflow_buffer;
690 dev_addr = swiotlb_virt_to_bus(dev, map);
693 * Ensure that the address returned is DMA'ble
695 if (!dma_capable(dev, dev_addr, size)) {
696 swiotlb_tbl_unmap_single(dev, map, size, dir);
697 dev_addr = swiotlb_virt_to_bus(dev, io_tlb_overflow_buffer);
702 EXPORT_SYMBOL_GPL(swiotlb_map_page);
705 * Unmap a single streaming mode DMA translation. The dma_addr and size must
706 * match what was provided for in a previous swiotlb_map_page call. All
707 * other usages are undefined.
709 * After this call, reads by the cpu to the buffer are guaranteed to see
710 * whatever the device wrote there.
712 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
713 size_t size, enum dma_data_direction dir)
715 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
717 BUG_ON(dir == DMA_NONE);
719 if (is_swiotlb_buffer(paddr)) {
720 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
724 if (dir != DMA_FROM_DEVICE)
728 * phys_to_virt doesn't work with hihgmem page but we could
729 * call dma_mark_clean() with hihgmem page here. However, we
730 * are fine since dma_mark_clean() is null on POWERPC. We can
731 * make dma_mark_clean() take a physical address if necessary.
733 dma_mark_clean(phys_to_virt(paddr), size);
736 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
737 size_t size, enum dma_data_direction dir,
738 struct dma_attrs *attrs)
740 unmap_single(hwdev, dev_addr, size, dir);
742 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
745 * Make physical memory consistent for a single streaming mode DMA translation
748 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
749 * using the cpu, yet do not wish to teardown the dma mapping, you must
750 * call this function before doing so. At the next point you give the dma
751 * address back to the card, you must first perform a
752 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
755 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
756 size_t size, enum dma_data_direction dir,
757 enum dma_sync_target target)
759 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
761 BUG_ON(dir == DMA_NONE);
763 if (is_swiotlb_buffer(paddr)) {
764 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
769 if (dir != DMA_FROM_DEVICE)
772 dma_mark_clean(phys_to_virt(paddr), size);
776 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
777 size_t size, enum dma_data_direction dir)
779 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
781 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
784 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
785 size_t size, enum dma_data_direction dir)
787 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
789 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
792 * Map a set of buffers described by scatterlist in streaming mode for DMA.
793 * This is the scatter-gather version of the above swiotlb_map_page
794 * interface. Here the scatter gather list elements are each tagged with the
795 * appropriate dma address and length. They are obtained via
796 * sg_dma_{address,length}(SG).
798 * NOTE: An implementation may be able to use a smaller number of
799 * DMA address/length pairs than there are SG table elements.
800 * (for example via virtual mapping capabilities)
801 * The routine returns the number of addr/length pairs actually
802 * used, at most nents.
804 * Device ownership issues as mentioned above for swiotlb_map_page are the
808 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
809 enum dma_data_direction dir, struct dma_attrs *attrs)
811 struct scatterlist *sg;
814 BUG_ON(dir == DMA_NONE);
816 for_each_sg(sgl, sg, nelems, i) {
817 phys_addr_t paddr = sg_phys(sg);
818 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
821 !dma_capable(hwdev, dev_addr, sg->length)) {
822 void *map = map_single(hwdev, sg_phys(sg),
825 /* Don't panic here, we expect map_sg users
826 to do proper error handling. */
827 swiotlb_full(hwdev, sg->length, dir, 0);
828 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
830 sgl[0].dma_length = 0;
833 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
835 sg->dma_address = dev_addr;
836 sg->dma_length = sg->length;
840 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
843 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
844 enum dma_data_direction dir)
846 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
848 EXPORT_SYMBOL(swiotlb_map_sg);
851 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
852 * concerning calls here are the same as for swiotlb_unmap_page() above.
855 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
856 int nelems, enum dma_data_direction dir, struct dma_attrs *attrs)
858 struct scatterlist *sg;
861 BUG_ON(dir == DMA_NONE);
863 for_each_sg(sgl, sg, nelems, i)
864 unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
867 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
870 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
871 enum dma_data_direction dir)
873 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
875 EXPORT_SYMBOL(swiotlb_unmap_sg);
878 * Make physical memory consistent for a set of streaming mode DMA translations
881 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
885 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
886 int nelems, enum dma_data_direction dir,
887 enum dma_sync_target target)
889 struct scatterlist *sg;
892 for_each_sg(sgl, sg, nelems, i)
893 swiotlb_sync_single(hwdev, sg->dma_address,
894 sg->dma_length, dir, target);
898 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
899 int nelems, enum dma_data_direction dir)
901 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
903 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
906 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
907 int nelems, enum dma_data_direction dir)
909 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
911 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
914 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
916 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
918 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
921 * Return whether the given device DMA address mask can be supported
922 * properly. For example, if your device can only drive the low 24-bits
923 * during bus mastering, then you would pass 0x00ffffff as the mask to
927 swiotlb_dma_supported(struct device *hwdev, u64 mask)
929 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
931 EXPORT_SYMBOL(swiotlb_dma_supported);