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.
19 #include <linux/cache.h>
20 #include <linux/dma-mapping.h>
22 #include <linux/module.h>
23 #include <linux/spinlock.h>
24 #include <linux/swiotlb.h>
25 #include <linux/string.h>
26 #include <linux/swiotlb.h>
27 #include <linux/types.h>
28 #include <linux/ctype.h>
29 #include <linux/highmem.h>
33 #include <asm/scatterlist.h>
35 #include <linux/init.h>
36 #include <linux/bootmem.h>
37 #include <linux/iommu-helper.h>
39 #define OFFSET(val,align) ((unsigned long) \
40 ( (val) & ( (align) - 1)))
42 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
45 * Minimum IO TLB size to bother booting with. Systems with mainly
46 * 64bit capable cards will only lightly use the swiotlb. If we can't
47 * allocate a contiguous 1MB, we're probably in trouble anyway.
49 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
52 * Enumeration for sync targets
54 enum dma_sync_target {
62 * Used to do a quick range check in swiotlb_unmap_single and
63 * swiotlb_sync_single_*, to see if the memory was in fact allocated by this
66 static char *io_tlb_start, *io_tlb_end;
69 * The number of IO TLB blocks (in groups of 64) betweeen io_tlb_start and
70 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
72 static unsigned long io_tlb_nslabs;
75 * When the IOMMU overflows we return a fallback buffer. This sets the size.
77 static unsigned long io_tlb_overflow = 32*1024;
79 void *io_tlb_overflow_buffer;
82 * This is a free list describing the number of free entries available from
85 static unsigned int *io_tlb_list;
86 static unsigned int io_tlb_index;
89 * We need to save away the original address corresponding to a mapped entry
90 * for the sync operations.
92 static struct swiotlb_phys_addr {
98 * Protect the above data structures in the map and unmap calls
100 static DEFINE_SPINLOCK(io_tlb_lock);
103 setup_io_tlb_npages(char *str)
106 io_tlb_nslabs = simple_strtoul(str, &str, 0);
107 /* avoid tail segment of size < IO_TLB_SEGSIZE */
108 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
112 if (!strcmp(str, "force"))
116 __setup("swiotlb=", setup_io_tlb_npages);
117 /* make io_tlb_overflow tunable too? */
119 void * __weak swiotlb_alloc_boot(size_t size, unsigned long nslabs)
121 return alloc_bootmem_low_pages(size);
124 void * __weak swiotlb_alloc(unsigned order, unsigned long nslabs)
126 return (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN, order);
129 dma_addr_t __weak swiotlb_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
134 phys_addr_t __weak swiotlb_bus_to_phys(dma_addr_t baddr)
139 static dma_addr_t swiotlb_virt_to_bus(struct device *hwdev,
140 volatile void *address)
142 return swiotlb_phys_to_bus(hwdev, virt_to_phys(address));
145 static void *swiotlb_bus_to_virt(dma_addr_t address)
147 return phys_to_virt(swiotlb_bus_to_phys(address));
150 int __weak swiotlb_arch_range_needs_mapping(void *ptr, size_t size)
155 static dma_addr_t swiotlb_sg_to_bus(struct device *hwdev, struct scatterlist *sg)
157 return swiotlb_phys_to_bus(hwdev, page_to_phys(sg_page(sg)) + sg->offset);
160 static void swiotlb_print_info(unsigned long bytes)
162 phys_addr_t pstart, pend;
164 pstart = virt_to_phys(io_tlb_start);
165 pend = virt_to_phys(io_tlb_end);
167 printk(KERN_INFO "Placing %luMB software IO TLB between %p - %p\n",
168 bytes >> 20, io_tlb_start, io_tlb_end);
169 printk(KERN_INFO "software IO TLB at phys %#llx - %#llx\n",
170 (unsigned long long)pstart,
171 (unsigned long long)pend);
175 * Statically reserve bounce buffer space and initialize bounce buffer data
176 * structures for the software IO TLB used to implement the DMA API.
179 swiotlb_init_with_default_size(size_t default_size)
181 unsigned long i, bytes;
183 if (!io_tlb_nslabs) {
184 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
185 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
188 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
191 * Get IO TLB memory from the low pages
193 io_tlb_start = swiotlb_alloc_boot(bytes, io_tlb_nslabs);
195 panic("Cannot allocate SWIOTLB buffer");
196 io_tlb_end = io_tlb_start + bytes;
199 * Allocate and initialize the free list array. This array is used
200 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
201 * between io_tlb_start and io_tlb_end.
203 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
204 for (i = 0; i < io_tlb_nslabs; i++)
205 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
207 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(struct swiotlb_phys_addr));
210 * Get the overflow emergency buffer
212 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
213 if (!io_tlb_overflow_buffer)
214 panic("Cannot allocate SWIOTLB overflow buffer!\n");
216 swiotlb_print_info(bytes);
222 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
226 * Systems with larger DMA zones (those that don't support ISA) can
227 * initialize the swiotlb later using the slab allocator if needed.
228 * This should be just like above, but with some error catching.
231 swiotlb_late_init_with_default_size(size_t default_size)
233 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
236 if (!io_tlb_nslabs) {
237 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
238 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
242 * Get IO TLB memory from the low pages
244 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
245 io_tlb_nslabs = SLABS_PER_PAGE << order;
246 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
248 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
249 io_tlb_start = swiotlb_alloc(order, io_tlb_nslabs);
258 if (order != get_order(bytes)) {
259 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
260 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
261 io_tlb_nslabs = SLABS_PER_PAGE << order;
262 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
264 io_tlb_end = io_tlb_start + bytes;
265 memset(io_tlb_start, 0, bytes);
268 * Allocate and initialize the free list array. This array is used
269 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
270 * between io_tlb_start and io_tlb_end.
272 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
273 get_order(io_tlb_nslabs * sizeof(int)));
277 for (i = 0; i < io_tlb_nslabs; i++)
278 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
281 io_tlb_orig_addr = (struct swiotlb_phys_addr *)__get_free_pages(GFP_KERNEL,
282 get_order(io_tlb_nslabs * sizeof(struct swiotlb_phys_addr)));
283 if (!io_tlb_orig_addr)
286 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(struct swiotlb_phys_addr));
289 * Get the overflow emergency buffer
291 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
292 get_order(io_tlb_overflow));
293 if (!io_tlb_overflow_buffer)
296 swiotlb_print_info(bytes);
301 free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs *
303 io_tlb_orig_addr = NULL;
305 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
310 free_pages((unsigned long)io_tlb_start, order);
313 io_tlb_nslabs = req_nslabs;
318 address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size)
320 return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size);
323 static inline int range_needs_mapping(void *ptr, size_t size)
325 return swiotlb_force || swiotlb_arch_range_needs_mapping(ptr, size);
328 static int is_swiotlb_buffer(char *addr)
330 return addr >= io_tlb_start && addr < io_tlb_end;
333 static struct swiotlb_phys_addr swiotlb_bus_to_phys_addr(char *dma_addr)
335 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
336 struct swiotlb_phys_addr buffer = io_tlb_orig_addr[index];
337 buffer.offset += (long)dma_addr & ((1 << IO_TLB_SHIFT) - 1);
338 buffer.page += buffer.offset >> PAGE_SHIFT;
339 buffer.offset &= PAGE_SIZE - 1;
344 __sync_single(struct swiotlb_phys_addr buffer, char *dma_addr, size_t size, int dir)
346 if (PageHighMem(buffer.page)) {
348 char *dev, *host, *kmp;
355 if ((bytes + buffer.offset) > PAGE_SIZE)
356 bytes = PAGE_SIZE - buffer.offset;
357 local_irq_save(flags); /* protects KM_BOUNCE_READ */
358 kmp = kmap_atomic(buffer.page, KM_BOUNCE_READ);
359 dev = dma_addr + size - len;
360 host = kmp + buffer.offset;
361 if (dir == DMA_FROM_DEVICE)
362 memcpy(host, dev, bytes);
364 memcpy(dev, host, bytes);
365 kunmap_atomic(kmp, KM_BOUNCE_READ);
366 local_irq_restore(flags);
372 void *v = page_address(buffer.page) + buffer.offset;
374 if (dir == DMA_TO_DEVICE)
375 memcpy(dma_addr, v, size);
377 memcpy(v, dma_addr, size);
382 * Allocates bounce buffer and returns its kernel virtual address.
385 map_single(struct device *hwdev, struct swiotlb_phys_addr buffer, size_t size, int dir)
389 unsigned int nslots, stride, index, wrap;
391 unsigned long start_dma_addr;
393 unsigned long offset_slots;
394 unsigned long max_slots;
395 struct swiotlb_phys_addr slot_buf;
397 mask = dma_get_seg_boundary(hwdev);
398 start_dma_addr = swiotlb_virt_to_bus(hwdev, io_tlb_start) & mask;
400 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
403 * Carefully handle integer overflow which can occur when mask == ~0UL.
406 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
407 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
410 * For mappings greater than a page, we limit the stride (and
411 * hence alignment) to a page size.
413 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
414 if (size > PAGE_SIZE)
415 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
422 * Find suitable number of IO TLB entries size that will fit this
423 * request and allocate a buffer from that IO TLB pool.
425 spin_lock_irqsave(&io_tlb_lock, flags);
426 index = ALIGN(io_tlb_index, stride);
427 if (index >= io_tlb_nslabs)
432 while (iommu_is_span_boundary(index, nslots, offset_slots,
435 if (index >= io_tlb_nslabs)
442 * If we find a slot that indicates we have 'nslots' number of
443 * contiguous buffers, we allocate the buffers from that slot
444 * and mark the entries as '0' indicating unavailable.
446 if (io_tlb_list[index] >= nslots) {
449 for (i = index; i < (int) (index + nslots); i++)
451 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
452 io_tlb_list[i] = ++count;
453 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
456 * Update the indices to avoid searching in the next
459 io_tlb_index = ((index + nslots) < io_tlb_nslabs
460 ? (index + nslots) : 0);
465 if (index >= io_tlb_nslabs)
467 } while (index != wrap);
470 spin_unlock_irqrestore(&io_tlb_lock, flags);
473 spin_unlock_irqrestore(&io_tlb_lock, flags);
476 * Save away the mapping from the original address to the DMA address.
477 * This is needed when we sync the memory. Then we sync the buffer if
481 for (i = 0; i < nslots; i++) {
482 slot_buf.page += slot_buf.offset >> PAGE_SHIFT;
483 slot_buf.offset &= PAGE_SIZE - 1;
484 io_tlb_orig_addr[index+i] = slot_buf;
485 slot_buf.offset += 1 << IO_TLB_SHIFT;
487 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
488 __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
494 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
497 unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
500 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
501 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
502 struct swiotlb_phys_addr buffer = swiotlb_bus_to_phys_addr(dma_addr);
505 * First, sync the memory before unmapping the entry
507 if ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL))
509 * bounce... copy the data back into the original buffer * and
510 * delete the bounce buffer.
512 __sync_single(buffer, dma_addr, size, DMA_FROM_DEVICE);
515 * Return the buffer to the free list by setting the corresponding
516 * entries to indicate the number of contigous entries available.
517 * While returning the entries to the free list, we merge the entries
518 * with slots below and above the pool being returned.
520 spin_lock_irqsave(&io_tlb_lock, flags);
522 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
523 io_tlb_list[index + nslots] : 0);
525 * Step 1: return the slots to the free list, merging the
526 * slots with superceeding slots
528 for (i = index + nslots - 1; i >= index; i--)
529 io_tlb_list[i] = ++count;
531 * Step 2: merge the returned slots with the preceding slots,
532 * if available (non zero)
534 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
535 io_tlb_list[i] = ++count;
537 spin_unlock_irqrestore(&io_tlb_lock, flags);
541 sync_single(struct device *hwdev, char *dma_addr, size_t size,
544 struct swiotlb_phys_addr buffer = swiotlb_bus_to_phys_addr(dma_addr);
548 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
549 __sync_single(buffer, dma_addr, size, DMA_FROM_DEVICE);
551 BUG_ON(dir != DMA_TO_DEVICE);
553 case SYNC_FOR_DEVICE:
554 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
555 __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
557 BUG_ON(dir != DMA_FROM_DEVICE);
565 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
566 dma_addr_t *dma_handle, gfp_t flags)
570 int order = get_order(size);
571 u64 dma_mask = DMA_32BIT_MASK;
573 if (hwdev && hwdev->coherent_dma_mask)
574 dma_mask = hwdev->coherent_dma_mask;
576 ret = (void *)__get_free_pages(flags, order);
578 !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(hwdev, ret),
581 * The allocated memory isn't reachable by the device.
582 * Fall back on swiotlb_map_single().
584 free_pages((unsigned long) ret, order);
589 * We are either out of memory or the device can't DMA
590 * to GFP_DMA memory; fall back on
591 * swiotlb_map_single(), which will grab memory from
592 * the lowest available address range.
594 struct swiotlb_phys_addr buffer;
595 buffer.page = virt_to_page(NULL);
597 ret = map_single(hwdev, buffer, size, DMA_FROM_DEVICE);
602 memset(ret, 0, size);
603 dev_addr = swiotlb_virt_to_bus(hwdev, ret);
605 /* Confirm address can be DMA'd by device */
606 if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) {
607 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
608 (unsigned long long)dma_mask,
609 (unsigned long long)dev_addr);
611 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
612 unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
615 *dma_handle = dev_addr;
620 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
621 dma_addr_t dma_handle)
623 WARN_ON(irqs_disabled());
624 if (!is_swiotlb_buffer(vaddr))
625 free_pages((unsigned long) vaddr, get_order(size));
627 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
628 unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
632 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
635 * Ran out of IOMMU space for this operation. This is very bad.
636 * Unfortunately the drivers cannot handle this operation properly.
637 * unless they check for dma_mapping_error (most don't)
638 * When the mapping is small enough return a static buffer to limit
639 * the damage, or panic when the transfer is too big.
641 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
642 "device %s\n", size, dev ? dev->bus_id : "?");
644 if (size > io_tlb_overflow && do_panic) {
645 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
646 panic("DMA: Memory would be corrupted\n");
647 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
648 panic("DMA: Random memory would be DMAed\n");
653 * Map a single buffer of the indicated size for DMA in streaming mode. The
654 * physical address to use is returned.
656 * Once the device is given the dma address, the device owns this memory until
657 * either swiotlb_unmap_single or swiotlb_dma_sync_single is performed.
660 swiotlb_map_single_attrs(struct device *hwdev, void *ptr, size_t size,
661 int dir, struct dma_attrs *attrs)
663 dma_addr_t dev_addr = swiotlb_virt_to_bus(hwdev, ptr);
665 struct swiotlb_phys_addr buffer;
667 BUG_ON(dir == DMA_NONE);
669 * If the pointer passed in happens to be in the device's DMA window,
670 * we can safely return the device addr and not worry about bounce
673 if (!address_needs_mapping(hwdev, dev_addr, size) &&
674 !range_needs_mapping(ptr, size))
678 * Oh well, have to allocate and map a bounce buffer.
680 buffer.page = virt_to_page(ptr);
681 buffer.offset = (unsigned long)ptr & ~PAGE_MASK;
682 map = map_single(hwdev, buffer, size, dir);
684 swiotlb_full(hwdev, size, dir, 1);
685 map = io_tlb_overflow_buffer;
688 dev_addr = swiotlb_virt_to_bus(hwdev, map);
691 * Ensure that the address returned is DMA'ble
693 if (address_needs_mapping(hwdev, dev_addr, size))
694 panic("map_single: bounce buffer is not DMA'ble");
698 EXPORT_SYMBOL(swiotlb_map_single_attrs);
701 swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir)
703 return swiotlb_map_single_attrs(hwdev, ptr, size, dir, NULL);
707 * Unmap a single streaming mode DMA translation. The dma_addr and size must
708 * match what was provided for in a previous swiotlb_map_single call. All
709 * other usages are undefined.
711 * After this call, reads by the cpu to the buffer are guaranteed to see
712 * whatever the device wrote there.
715 swiotlb_unmap_single_attrs(struct device *hwdev, dma_addr_t dev_addr,
716 size_t size, int dir, struct dma_attrs *attrs)
718 char *dma_addr = swiotlb_bus_to_virt(dev_addr);
720 BUG_ON(dir == DMA_NONE);
721 if (is_swiotlb_buffer(dma_addr))
722 unmap_single(hwdev, dma_addr, size, dir);
723 else if (dir == DMA_FROM_DEVICE)
724 dma_mark_clean(dma_addr, size);
726 EXPORT_SYMBOL(swiotlb_unmap_single_attrs);
729 swiotlb_unmap_single(struct device *hwdev, dma_addr_t dev_addr, size_t size,
732 return swiotlb_unmap_single_attrs(hwdev, dev_addr, size, dir, NULL);
735 * Make physical memory consistent for a single streaming mode DMA translation
738 * If you perform a swiotlb_map_single() but wish to interrogate the buffer
739 * using the cpu, yet do not wish to teardown the dma mapping, you must
740 * call this function before doing so. At the next point you give the dma
741 * address back to the card, you must first perform a
742 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
745 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
746 size_t size, int dir, int target)
748 char *dma_addr = swiotlb_bus_to_virt(dev_addr);
750 BUG_ON(dir == DMA_NONE);
751 if (is_swiotlb_buffer(dma_addr))
752 sync_single(hwdev, dma_addr, size, dir, target);
753 else if (dir == DMA_FROM_DEVICE)
754 dma_mark_clean(dma_addr, size);
758 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
759 size_t size, int dir)
761 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
765 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
766 size_t size, int dir)
768 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
772 * Same as above, but for a sub-range of the mapping.
775 swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
776 unsigned long offset, size_t size,
779 char *dma_addr = swiotlb_bus_to_virt(dev_addr) + offset;
781 BUG_ON(dir == DMA_NONE);
782 if (is_swiotlb_buffer(dma_addr))
783 sync_single(hwdev, dma_addr, size, dir, target);
784 else if (dir == DMA_FROM_DEVICE)
785 dma_mark_clean(dma_addr, size);
789 swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
790 unsigned long offset, size_t size, int dir)
792 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
797 swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
798 unsigned long offset, size_t size, int dir)
800 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
804 void swiotlb_unmap_sg_attrs(struct device *, struct scatterlist *, int, int,
807 * Map a set of buffers described by scatterlist in streaming mode for DMA.
808 * This is the scatter-gather version of the above swiotlb_map_single
809 * interface. Here the scatter gather list elements are each tagged with the
810 * appropriate dma address and length. They are obtained via
811 * sg_dma_{address,length}(SG).
813 * NOTE: An implementation may be able to use a smaller number of
814 * DMA address/length pairs than there are SG table elements.
815 * (for example via virtual mapping capabilities)
816 * The routine returns the number of addr/length pairs actually
817 * used, at most nents.
819 * Device ownership issues as mentioned above for swiotlb_map_single are the
823 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
824 int dir, struct dma_attrs *attrs)
826 struct scatterlist *sg;
827 struct swiotlb_phys_addr buffer;
831 BUG_ON(dir == DMA_NONE);
833 for_each_sg(sgl, sg, nelems, i) {
834 dev_addr = swiotlb_sg_to_bus(hwdev, sg);
835 if (range_needs_mapping(sg_virt(sg), sg->length) ||
836 address_needs_mapping(hwdev, dev_addr, sg->length)) {
838 buffer.page = sg_page(sg);
839 buffer.offset = sg->offset;
840 map = map_single(hwdev, buffer, sg->length, dir);
842 /* Don't panic here, we expect map_sg users
843 to do proper error handling. */
844 swiotlb_full(hwdev, sg->length, dir, 0);
845 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
847 sgl[0].dma_length = 0;
850 sg->dma_address = swiotlb_virt_to_bus(hwdev, map);
852 sg->dma_address = dev_addr;
853 sg->dma_length = sg->length;
857 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
860 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
863 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
867 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
868 * concerning calls here are the same as for swiotlb_unmap_single() above.
871 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
872 int nelems, int dir, struct dma_attrs *attrs)
874 struct scatterlist *sg;
877 BUG_ON(dir == DMA_NONE);
879 for_each_sg(sgl, sg, nelems, i) {
880 if (sg->dma_address != swiotlb_sg_to_bus(hwdev, sg))
881 unmap_single(hwdev, swiotlb_bus_to_virt(sg->dma_address),
882 sg->dma_length, dir);
883 else if (dir == DMA_FROM_DEVICE)
884 dma_mark_clean(swiotlb_bus_to_virt(sg->dma_address), sg->dma_length);
887 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
890 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
893 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
897 * Make physical memory consistent for a set of streaming mode DMA translations
900 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
904 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
905 int nelems, int dir, int target)
907 struct scatterlist *sg;
910 BUG_ON(dir == DMA_NONE);
912 for_each_sg(sgl, sg, nelems, i) {
913 if (sg->dma_address != swiotlb_sg_to_bus(hwdev, sg))
914 sync_single(hwdev, swiotlb_bus_to_virt(sg->dma_address),
915 sg->dma_length, dir, target);
916 else if (dir == DMA_FROM_DEVICE)
917 dma_mark_clean(swiotlb_bus_to_virt(sg->dma_address), sg->dma_length);
922 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
925 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
929 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
932 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
936 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
938 return (dma_addr == swiotlb_virt_to_bus(hwdev, io_tlb_overflow_buffer));
942 * Return whether the given device DMA address mask can be supported
943 * properly. For example, if your device can only drive the low 24-bits
944 * during bus mastering, then you would pass 0x00ffffff as the mask to
948 swiotlb_dma_supported(struct device *hwdev, u64 mask)
950 return swiotlb_virt_to_bus(hwdev, io_tlb_end - 1) <= mask;
953 EXPORT_SYMBOL(swiotlb_map_single);
954 EXPORT_SYMBOL(swiotlb_unmap_single);
955 EXPORT_SYMBOL(swiotlb_map_sg);
956 EXPORT_SYMBOL(swiotlb_unmap_sg);
957 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
958 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
959 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
960 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
961 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
962 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
963 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
964 EXPORT_SYMBOL(swiotlb_alloc_coherent);
965 EXPORT_SYMBOL(swiotlb_free_coherent);
966 EXPORT_SYMBOL(swiotlb_dma_supported);