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 #define pr_fmt(fmt) "software IO TLB: " fmt
22 #include <linux/cache.h>
23 #include <linux/dma-mapping.h>
25 #include <linux/export.h>
26 #include <linux/spinlock.h>
27 #include <linux/string.h>
28 #include <linux/swiotlb.h>
29 #include <linux/pfn.h>
30 #include <linux/types.h>
31 #include <linux/ctype.h>
32 #include <linux/highmem.h>
33 #include <linux/gfp.h>
34 #include <linux/scatterlist.h>
39 #include <linux/init.h>
40 #include <linux/bootmem.h>
41 #include <linux/iommu-helper.h>
43 #define CREATE_TRACE_POINTS
44 #include <trace/events/swiotlb.h>
46 #define OFFSET(val,align) ((unsigned long) \
47 ( (val) & ( (align) - 1)))
49 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
52 * Minimum IO TLB size to bother booting with. Systems with mainly
53 * 64bit capable cards will only lightly use the swiotlb. If we can't
54 * allocate a contiguous 1MB, we're probably in trouble anyway.
56 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
58 enum swiotlb_force swiotlb_force;
61 * Used to do a quick range check in swiotlb_tbl_unmap_single and
62 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
65 static phys_addr_t io_tlb_start, io_tlb_end;
68 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
69 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
71 static unsigned long io_tlb_nslabs;
74 * When the IOMMU overflows we return a fallback buffer. This sets the size.
76 static unsigned long io_tlb_overflow = 32*1024;
78 static phys_addr_t io_tlb_overflow_buffer;
81 * This is a free list describing the number of free entries available from
84 static unsigned int *io_tlb_list;
85 static unsigned int io_tlb_index;
88 * We need to save away the original address corresponding to a mapped entry
89 * for the sync operations.
91 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
92 static phys_addr_t *io_tlb_orig_addr;
95 * Protect the above data structures in the map and unmap calls
97 static DEFINE_SPINLOCK(io_tlb_lock);
99 static int late_alloc;
102 setup_io_tlb_npages(char *str)
105 io_tlb_nslabs = simple_strtoul(str, &str, 0);
106 /* avoid tail segment of size < IO_TLB_SEGSIZE */
107 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
111 if (!strcmp(str, "force")) {
112 swiotlb_force = SWIOTLB_FORCE;
113 } else if (!strcmp(str, "noforce")) {
114 swiotlb_force = SWIOTLB_NO_FORCE;
120 early_param("swiotlb", setup_io_tlb_npages);
121 /* make io_tlb_overflow tunable too? */
123 unsigned long swiotlb_nr_tbl(void)
125 return io_tlb_nslabs;
127 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
129 /* default to 64MB */
130 #define IO_TLB_DEFAULT_SIZE (64UL<<20)
131 unsigned long swiotlb_size_or_default(void)
135 size = io_tlb_nslabs << IO_TLB_SHIFT;
137 return size ? size : (IO_TLB_DEFAULT_SIZE);
140 /* Note that this doesn't work with highmem page */
141 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
142 volatile void *address)
144 return phys_to_dma(hwdev, virt_to_phys(address));
147 static bool no_iotlb_memory;
149 void swiotlb_print_info(void)
151 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
153 if (no_iotlb_memory) {
154 pr_warn("No low mem\n");
158 pr_info("mapped [mem %#010llx-%#010llx] (%luMB)\n",
159 (unsigned long long)io_tlb_start,
160 (unsigned long long)io_tlb_end,
164 int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
166 void *v_overflow_buffer;
167 unsigned long i, bytes;
169 bytes = nslabs << IO_TLB_SHIFT;
171 io_tlb_nslabs = nslabs;
172 io_tlb_start = __pa(tlb);
173 io_tlb_end = io_tlb_start + bytes;
176 * Get the overflow emergency buffer
178 v_overflow_buffer = memblock_virt_alloc_low_nopanic(
179 PAGE_ALIGN(io_tlb_overflow),
181 if (!v_overflow_buffer)
184 io_tlb_overflow_buffer = __pa(v_overflow_buffer);
187 * Allocate and initialize the free list array. This array is used
188 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
189 * between io_tlb_start and io_tlb_end.
191 io_tlb_list = memblock_virt_alloc(
192 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
194 io_tlb_orig_addr = memblock_virt_alloc(
195 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
197 for (i = 0; i < io_tlb_nslabs; i++) {
198 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
199 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
202 no_iotlb_memory = false;
205 swiotlb_print_info();
211 * Statically reserve bounce buffer space and initialize bounce buffer data
212 * structures for the software IO TLB used to implement the DMA API.
215 swiotlb_init(int verbose)
217 size_t default_size = IO_TLB_DEFAULT_SIZE;
218 unsigned char *vstart;
221 if (!io_tlb_nslabs) {
222 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
223 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
226 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
228 /* Get IO TLB memory from the low pages */
229 vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
230 if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
234 memblock_free_early(io_tlb_start,
235 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
238 pr_warn("Cannot allocate buffer");
239 no_iotlb_memory = true;
243 * Systems with larger DMA zones (those that don't support ISA) can
244 * initialize the swiotlb later using the slab allocator if needed.
245 * This should be just like above, but with some error catching.
248 swiotlb_late_init_with_default_size(size_t default_size)
250 unsigned long bytes, req_nslabs = io_tlb_nslabs;
251 unsigned char *vstart = NULL;
255 if (!io_tlb_nslabs) {
256 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
257 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
261 * Get IO TLB memory from the low pages
263 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
264 io_tlb_nslabs = SLABS_PER_PAGE << order;
265 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
267 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
268 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
276 io_tlb_nslabs = req_nslabs;
279 if (order != get_order(bytes)) {
280 pr_warn("only able to allocate %ld MB\n",
281 (PAGE_SIZE << order) >> 20);
282 io_tlb_nslabs = SLABS_PER_PAGE << order;
284 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
286 free_pages((unsigned long)vstart, order);
291 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
293 unsigned long i, bytes;
294 unsigned char *v_overflow_buffer;
296 bytes = nslabs << IO_TLB_SHIFT;
298 io_tlb_nslabs = nslabs;
299 io_tlb_start = virt_to_phys(tlb);
300 io_tlb_end = io_tlb_start + bytes;
302 memset(tlb, 0, bytes);
305 * Get the overflow emergency buffer
307 v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
308 get_order(io_tlb_overflow));
309 if (!v_overflow_buffer)
312 io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
315 * Allocate and initialize the free list array. This array is used
316 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
317 * between io_tlb_start and io_tlb_end.
319 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
320 get_order(io_tlb_nslabs * sizeof(int)));
324 io_tlb_orig_addr = (phys_addr_t *)
325 __get_free_pages(GFP_KERNEL,
326 get_order(io_tlb_nslabs *
327 sizeof(phys_addr_t)));
328 if (!io_tlb_orig_addr)
331 for (i = 0; i < io_tlb_nslabs; i++) {
332 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
333 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
336 no_iotlb_memory = false;
338 swiotlb_print_info();
345 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
349 free_pages((unsigned long)v_overflow_buffer,
350 get_order(io_tlb_overflow));
351 io_tlb_overflow_buffer = 0;
359 void __init swiotlb_free(void)
361 if (!io_tlb_orig_addr)
365 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
366 get_order(io_tlb_overflow));
367 free_pages((unsigned long)io_tlb_orig_addr,
368 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
369 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
371 free_pages((unsigned long)phys_to_virt(io_tlb_start),
372 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
374 memblock_free_late(io_tlb_overflow_buffer,
375 PAGE_ALIGN(io_tlb_overflow));
376 memblock_free_late(__pa(io_tlb_orig_addr),
377 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
378 memblock_free_late(__pa(io_tlb_list),
379 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
380 memblock_free_late(io_tlb_start,
381 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
386 int is_swiotlb_buffer(phys_addr_t paddr)
388 return paddr >= io_tlb_start && paddr < io_tlb_end;
392 * Bounce: copy the swiotlb buffer back to the original dma location
394 static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
395 size_t size, enum dma_data_direction dir)
397 unsigned long pfn = PFN_DOWN(orig_addr);
398 unsigned char *vaddr = phys_to_virt(tlb_addr);
400 if (PageHighMem(pfn_to_page(pfn))) {
401 /* The buffer does not have a mapping. Map it in and copy */
402 unsigned int offset = orig_addr & ~PAGE_MASK;
408 sz = min_t(size_t, PAGE_SIZE - offset, size);
410 local_irq_save(flags);
411 buffer = kmap_atomic(pfn_to_page(pfn));
412 if (dir == DMA_TO_DEVICE)
413 memcpy(vaddr, buffer + offset, sz);
415 memcpy(buffer + offset, vaddr, sz);
416 kunmap_atomic(buffer);
417 local_irq_restore(flags);
424 } else if (dir == DMA_TO_DEVICE) {
425 memcpy(vaddr, phys_to_virt(orig_addr), size);
427 memcpy(phys_to_virt(orig_addr), vaddr, size);
431 phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
432 dma_addr_t tbl_dma_addr,
433 phys_addr_t orig_addr, size_t size,
434 enum dma_data_direction dir)
437 phys_addr_t tlb_addr;
438 unsigned int nslots, stride, index, wrap;
441 unsigned long offset_slots;
442 unsigned long max_slots;
445 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
447 mask = dma_get_seg_boundary(hwdev);
449 tbl_dma_addr &= mask;
451 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
454 * Carefully handle integer overflow which can occur when mask == ~0UL.
457 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
458 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
461 * For mappings greater than or equal to a page, we limit the stride
462 * (and hence alignment) to a page size.
464 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
465 if (size >= PAGE_SIZE)
466 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
473 * Find suitable number of IO TLB entries size that will fit this
474 * request and allocate a buffer from that IO TLB pool.
476 spin_lock_irqsave(&io_tlb_lock, flags);
477 index = ALIGN(io_tlb_index, stride);
478 if (index >= io_tlb_nslabs)
483 while (iommu_is_span_boundary(index, nslots, offset_slots,
486 if (index >= io_tlb_nslabs)
493 * If we find a slot that indicates we have 'nslots' number of
494 * contiguous buffers, we allocate the buffers from that slot
495 * and mark the entries as '0' indicating unavailable.
497 if (io_tlb_list[index] >= nslots) {
500 for (i = index; i < (int) (index + nslots); i++)
502 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
503 io_tlb_list[i] = ++count;
504 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
507 * Update the indices to avoid searching in the next
510 io_tlb_index = ((index + nslots) < io_tlb_nslabs
511 ? (index + nslots) : 0);
516 if (index >= io_tlb_nslabs)
518 } while (index != wrap);
521 spin_unlock_irqrestore(&io_tlb_lock, flags);
522 if (printk_ratelimit())
523 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
524 return SWIOTLB_MAP_ERROR;
526 spin_unlock_irqrestore(&io_tlb_lock, flags);
529 * Save away the mapping from the original address to the DMA address.
530 * This is needed when we sync the memory. Then we sync the buffer if
533 for (i = 0; i < nslots; i++)
534 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
535 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
536 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
540 EXPORT_SYMBOL_GPL(swiotlb_tbl_map_single);
543 * Allocates bounce buffer and returns its kernel virtual address.
547 map_single(struct device *hwdev, phys_addr_t phys, size_t size,
548 enum dma_data_direction dir)
550 dma_addr_t start_dma_addr;
552 if (swiotlb_force == SWIOTLB_NO_FORCE) {
553 dev_warn_ratelimited(hwdev, "Cannot do DMA to address %pa\n",
555 return SWIOTLB_MAP_ERROR;
558 start_dma_addr = phys_to_dma(hwdev, io_tlb_start);
559 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
563 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
565 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
566 size_t size, enum dma_data_direction dir)
569 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
570 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
571 phys_addr_t orig_addr = io_tlb_orig_addr[index];
574 * First, sync the memory before unmapping the entry
576 if (orig_addr != INVALID_PHYS_ADDR &&
577 ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
578 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
581 * Return the buffer to the free list by setting the corresponding
582 * entries to indicate the number of contiguous entries available.
583 * While returning the entries to the free list, we merge the entries
584 * with slots below and above the pool being returned.
586 spin_lock_irqsave(&io_tlb_lock, flags);
588 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
589 io_tlb_list[index + nslots] : 0);
591 * Step 1: return the slots to the free list, merging the
592 * slots with superceeding slots
594 for (i = index + nslots - 1; i >= index; i--) {
595 io_tlb_list[i] = ++count;
596 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
599 * Step 2: merge the returned slots with the preceding slots,
600 * if available (non zero)
602 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
603 io_tlb_list[i] = ++count;
605 spin_unlock_irqrestore(&io_tlb_lock, flags);
607 EXPORT_SYMBOL_GPL(swiotlb_tbl_unmap_single);
609 void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
610 size_t size, enum dma_data_direction dir,
611 enum dma_sync_target target)
613 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
614 phys_addr_t orig_addr = io_tlb_orig_addr[index];
616 if (orig_addr == INVALID_PHYS_ADDR)
618 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
622 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
623 swiotlb_bounce(orig_addr, tlb_addr,
624 size, DMA_FROM_DEVICE);
626 BUG_ON(dir != DMA_TO_DEVICE);
628 case SYNC_FOR_DEVICE:
629 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
630 swiotlb_bounce(orig_addr, tlb_addr,
631 size, DMA_TO_DEVICE);
633 BUG_ON(dir != DMA_FROM_DEVICE);
639 EXPORT_SYMBOL_GPL(swiotlb_tbl_sync_single);
642 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
643 dma_addr_t *dma_handle, gfp_t flags)
647 int order = get_order(size);
648 u64 dma_mask = DMA_BIT_MASK(32);
650 if (hwdev && hwdev->coherent_dma_mask)
651 dma_mask = hwdev->coherent_dma_mask;
653 ret = (void *)__get_free_pages(flags, order);
655 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
656 if (dev_addr + size - 1 > dma_mask) {
658 * The allocated memory isn't reachable by the device.
660 free_pages((unsigned long) ret, order);
666 * We are either out of memory or the device can't DMA to
667 * GFP_DMA memory; fall back on map_single(), which
668 * will grab memory from the lowest available address range.
670 phys_addr_t paddr = map_single(hwdev, 0, size, DMA_FROM_DEVICE);
671 if (paddr == SWIOTLB_MAP_ERROR)
674 ret = phys_to_virt(paddr);
675 dev_addr = phys_to_dma(hwdev, paddr);
677 /* Confirm address can be DMA'd by device */
678 if (dev_addr + size - 1 > dma_mask) {
679 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
680 (unsigned long long)dma_mask,
681 (unsigned long long)dev_addr);
683 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
684 swiotlb_tbl_unmap_single(hwdev, paddr,
685 size, DMA_TO_DEVICE);
690 *dma_handle = dev_addr;
691 memset(ret, 0, size);
696 pr_warn("coherent allocation failed for device %s size=%zu\n",
697 dev_name(hwdev), size);
702 EXPORT_SYMBOL(swiotlb_alloc_coherent);
705 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
708 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
710 WARN_ON(irqs_disabled());
711 if (!is_swiotlb_buffer(paddr))
712 free_pages((unsigned long)vaddr, get_order(size));
714 /* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
715 swiotlb_tbl_unmap_single(hwdev, paddr, size, DMA_TO_DEVICE);
717 EXPORT_SYMBOL(swiotlb_free_coherent);
720 swiotlb_full(struct device *dev, size_t size, enum dma_data_direction dir,
723 if (swiotlb_force == SWIOTLB_NO_FORCE)
727 * Ran out of IOMMU space for this operation. This is very bad.
728 * Unfortunately the drivers cannot handle this operation properly.
729 * unless they check for dma_mapping_error (most don't)
730 * When the mapping is small enough return a static buffer to limit
731 * the damage, or panic when the transfer is too big.
733 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
734 "device %s\n", size, dev ? dev_name(dev) : "?");
736 if (size <= io_tlb_overflow || !do_panic)
739 if (dir == DMA_BIDIRECTIONAL)
740 panic("DMA: Random memory could be DMA accessed\n");
741 if (dir == DMA_FROM_DEVICE)
742 panic("DMA: Random memory could be DMA written\n");
743 if (dir == DMA_TO_DEVICE)
744 panic("DMA: Random memory could be DMA read\n");
748 * Map a single buffer of the indicated size for DMA in streaming mode. The
749 * physical address to use is returned.
751 * Once the device is given the dma address, the device owns this memory until
752 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
754 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
755 unsigned long offset, size_t size,
756 enum dma_data_direction dir,
759 phys_addr_t map, phys = page_to_phys(page) + offset;
760 dma_addr_t dev_addr = phys_to_dma(dev, phys);
762 BUG_ON(dir == DMA_NONE);
764 * If the address happens to be in the device's DMA window,
765 * we can safely return the device addr and not worry about bounce
768 if (dma_capable(dev, dev_addr, size) && swiotlb_force != SWIOTLB_FORCE)
771 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
773 /* Oh well, have to allocate and map a bounce buffer. */
774 map = map_single(dev, phys, size, dir);
775 if (map == SWIOTLB_MAP_ERROR) {
776 swiotlb_full(dev, size, dir, 1);
777 return phys_to_dma(dev, io_tlb_overflow_buffer);
780 dev_addr = phys_to_dma(dev, map);
782 /* Ensure that the address returned is DMA'ble */
783 if (!dma_capable(dev, dev_addr, size)) {
784 swiotlb_tbl_unmap_single(dev, map, size, dir);
785 return phys_to_dma(dev, io_tlb_overflow_buffer);
790 EXPORT_SYMBOL_GPL(swiotlb_map_page);
793 * Unmap a single streaming mode DMA translation. The dma_addr and size must
794 * match what was provided for in a previous swiotlb_map_page call. All
795 * other usages are undefined.
797 * After this call, reads by the cpu to the buffer are guaranteed to see
798 * whatever the device wrote there.
800 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
801 size_t size, enum dma_data_direction dir)
803 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
805 BUG_ON(dir == DMA_NONE);
807 if (is_swiotlb_buffer(paddr)) {
808 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
812 if (dir != DMA_FROM_DEVICE)
816 * phys_to_virt doesn't work with hihgmem page but we could
817 * call dma_mark_clean() with hihgmem page here. However, we
818 * are fine since dma_mark_clean() is null on POWERPC. We can
819 * make dma_mark_clean() take a physical address if necessary.
821 dma_mark_clean(phys_to_virt(paddr), size);
824 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
825 size_t size, enum dma_data_direction dir,
828 unmap_single(hwdev, dev_addr, size, dir);
830 EXPORT_SYMBOL_GPL(swiotlb_unmap_page);
833 * Make physical memory consistent for a single streaming mode DMA translation
836 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
837 * using the cpu, yet do not wish to teardown the dma mapping, you must
838 * call this function before doing so. At the next point you give the dma
839 * address back to the card, you must first perform a
840 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
843 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
844 size_t size, enum dma_data_direction dir,
845 enum dma_sync_target target)
847 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
849 BUG_ON(dir == DMA_NONE);
851 if (is_swiotlb_buffer(paddr)) {
852 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
856 if (dir != DMA_FROM_DEVICE)
859 dma_mark_clean(phys_to_virt(paddr), size);
863 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
864 size_t size, enum dma_data_direction dir)
866 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
868 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
871 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
872 size_t size, enum dma_data_direction dir)
874 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
876 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
879 * Map a set of buffers described by scatterlist in streaming mode for DMA.
880 * This is the scatter-gather version of the above swiotlb_map_page
881 * interface. Here the scatter gather list elements are each tagged with the
882 * appropriate dma address and length. They are obtained via
883 * sg_dma_{address,length}(SG).
885 * NOTE: An implementation may be able to use a smaller number of
886 * DMA address/length pairs than there are SG table elements.
887 * (for example via virtual mapping capabilities)
888 * The routine returns the number of addr/length pairs actually
889 * used, at most nents.
891 * Device ownership issues as mentioned above for swiotlb_map_page are the
895 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
896 enum dma_data_direction dir, unsigned long attrs)
898 struct scatterlist *sg;
901 BUG_ON(dir == DMA_NONE);
903 for_each_sg(sgl, sg, nelems, i) {
904 phys_addr_t paddr = sg_phys(sg);
905 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
907 if (swiotlb_force == SWIOTLB_FORCE ||
908 !dma_capable(hwdev, dev_addr, sg->length)) {
909 phys_addr_t map = map_single(hwdev, sg_phys(sg),
911 if (map == SWIOTLB_MAP_ERROR) {
912 /* Don't panic here, we expect map_sg users
913 to do proper error handling. */
914 swiotlb_full(hwdev, sg->length, dir, 0);
915 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
920 sg->dma_address = phys_to_dma(hwdev, map);
922 sg->dma_address = dev_addr;
923 sg_dma_len(sg) = sg->length;
927 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
930 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
931 enum dma_data_direction dir)
933 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, 0);
935 EXPORT_SYMBOL(swiotlb_map_sg);
938 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
939 * concerning calls here are the same as for swiotlb_unmap_page() above.
942 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
943 int nelems, enum dma_data_direction dir,
946 struct scatterlist *sg;
949 BUG_ON(dir == DMA_NONE);
951 for_each_sg(sgl, sg, nelems, i)
952 unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
955 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
958 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
959 enum dma_data_direction dir)
961 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, 0);
963 EXPORT_SYMBOL(swiotlb_unmap_sg);
966 * Make physical memory consistent for a set of streaming mode DMA translations
969 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
973 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
974 int nelems, enum dma_data_direction dir,
975 enum dma_sync_target target)
977 struct scatterlist *sg;
980 for_each_sg(sgl, sg, nelems, i)
981 swiotlb_sync_single(hwdev, sg->dma_address,
982 sg_dma_len(sg), dir, target);
986 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
987 int nelems, enum dma_data_direction dir)
989 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
991 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
994 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
995 int nelems, enum dma_data_direction dir)
997 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
999 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
1002 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
1004 return (dma_addr == phys_to_dma(hwdev, io_tlb_overflow_buffer));
1006 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
1009 * Return whether the given device DMA address mask can be supported
1010 * properly. For example, if your device can only drive the low 24-bits
1011 * during bus mastering, then you would pass 0x00ffffff as the mask to
1015 swiotlb_dma_supported(struct device *hwdev, u64 mask)
1017 return phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
1019 EXPORT_SYMBOL(swiotlb_dma_supported);