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(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(volatile void *address)
141 return swiotlb_phys_to_bus(virt_to_phys(address));
144 static void *swiotlb_bus_to_virt(dma_addr_t address)
146 return phys_to_virt(swiotlb_bus_to_phys(address));
149 int __weak swiotlb_arch_range_needs_mapping(void *ptr, size_t size)
154 static dma_addr_t swiotlb_sg_to_bus(struct scatterlist *sg)
156 return swiotlb_phys_to_bus(page_to_phys(sg_page(sg)) + sg->offset);
160 * Statically reserve bounce buffer space and initialize bounce buffer data
161 * structures for the software IO TLB used to implement the DMA API.
164 swiotlb_init_with_default_size(size_t default_size)
166 unsigned long i, bytes;
168 if (!io_tlb_nslabs) {
169 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
170 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
173 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
176 * Get IO TLB memory from the low pages
178 io_tlb_start = swiotlb_alloc_boot(bytes, io_tlb_nslabs);
180 panic("Cannot allocate SWIOTLB buffer");
181 io_tlb_end = io_tlb_start + bytes;
184 * Allocate and initialize the free list array. This array is used
185 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
186 * between io_tlb_start and io_tlb_end.
188 io_tlb_list = alloc_bootmem(io_tlb_nslabs * sizeof(int));
189 for (i = 0; i < io_tlb_nslabs; i++)
190 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
192 io_tlb_orig_addr = alloc_bootmem(io_tlb_nslabs * sizeof(struct swiotlb_phys_addr));
195 * Get the overflow emergency buffer
197 io_tlb_overflow_buffer = alloc_bootmem_low(io_tlb_overflow);
198 if (!io_tlb_overflow_buffer)
199 panic("Cannot allocate SWIOTLB overflow buffer!\n");
201 printk(KERN_INFO "Placing software IO TLB between 0x%lx - 0x%lx\n",
202 swiotlb_virt_to_bus(io_tlb_start), swiotlb_virt_to_bus(io_tlb_end));
208 swiotlb_init_with_default_size(64 * (1<<20)); /* default to 64MB */
212 * Systems with larger DMA zones (those that don't support ISA) can
213 * initialize the swiotlb later using the slab allocator if needed.
214 * This should be just like above, but with some error catching.
217 swiotlb_late_init_with_default_size(size_t default_size)
219 unsigned long i, bytes, req_nslabs = io_tlb_nslabs;
222 if (!io_tlb_nslabs) {
223 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
224 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
228 * Get IO TLB memory from the low pages
230 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
231 io_tlb_nslabs = SLABS_PER_PAGE << order;
232 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
234 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
235 io_tlb_start = swiotlb_alloc(order, io_tlb_nslabs);
244 if (order != get_order(bytes)) {
245 printk(KERN_WARNING "Warning: only able to allocate %ld MB "
246 "for software IO TLB\n", (PAGE_SIZE << order) >> 20);
247 io_tlb_nslabs = SLABS_PER_PAGE << order;
248 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
250 io_tlb_end = io_tlb_start + bytes;
251 memset(io_tlb_start, 0, bytes);
254 * Allocate and initialize the free list array. This array is used
255 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
256 * between io_tlb_start and io_tlb_end.
258 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
259 get_order(io_tlb_nslabs * sizeof(int)));
263 for (i = 0; i < io_tlb_nslabs; i++)
264 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
267 io_tlb_orig_addr = (struct swiotlb_phys_addr *)__get_free_pages(GFP_KERNEL,
268 get_order(io_tlb_nslabs * sizeof(struct swiotlb_phys_addr)));
269 if (!io_tlb_orig_addr)
272 memset(io_tlb_orig_addr, 0, io_tlb_nslabs * sizeof(struct swiotlb_phys_addr));
275 * Get the overflow emergency buffer
277 io_tlb_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
278 get_order(io_tlb_overflow));
279 if (!io_tlb_overflow_buffer)
282 printk(KERN_INFO "Placing %luMB software IO TLB between 0x%lx - "
283 "0x%lx\n", bytes >> 20,
284 swiotlb_virt_to_bus(io_tlb_start), swiotlb_virt_to_bus(io_tlb_end));
289 free_pages((unsigned long)io_tlb_orig_addr, get_order(io_tlb_nslabs *
291 io_tlb_orig_addr = NULL;
293 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
298 free_pages((unsigned long)io_tlb_start, order);
301 io_tlb_nslabs = req_nslabs;
306 address_needs_mapping(struct device *hwdev, dma_addr_t addr, size_t size)
308 return !is_buffer_dma_capable(dma_get_mask(hwdev), addr, size);
311 static inline int range_needs_mapping(void *ptr, size_t size)
313 return swiotlb_force || swiotlb_arch_range_needs_mapping(ptr, size);
316 static int is_swiotlb_buffer(char *addr)
318 return addr >= io_tlb_start && addr < io_tlb_end;
321 static struct swiotlb_phys_addr swiotlb_bus_to_phys_addr(char *dma_addr)
323 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
324 struct swiotlb_phys_addr buffer = io_tlb_orig_addr[index];
325 buffer.offset += (long)dma_addr & ((1 << IO_TLB_SHIFT) - 1);
326 buffer.page += buffer.offset >> PAGE_SHIFT;
327 buffer.offset &= PAGE_SIZE - 1;
332 __sync_single(struct swiotlb_phys_addr buffer, char *dma_addr, size_t size, int dir)
334 if (PageHighMem(buffer.page)) {
336 char *dev, *host, *kmp;
343 if ((bytes + buffer.offset) > PAGE_SIZE)
344 bytes = PAGE_SIZE - buffer.offset;
345 local_irq_save(flags); /* protects KM_BOUNCE_READ */
346 kmp = kmap_atomic(buffer.page, KM_BOUNCE_READ);
347 dev = dma_addr + size - len;
348 host = kmp + buffer.offset;
349 if (dir == DMA_FROM_DEVICE)
350 memcpy(host, dev, bytes);
352 memcpy(dev, host, bytes);
353 kunmap_atomic(kmp, KM_BOUNCE_READ);
354 local_irq_restore(flags);
360 void *v = page_address(buffer.page) + buffer.offset;
362 if (dir == DMA_TO_DEVICE)
363 memcpy(dma_addr, v, size);
365 memcpy(v, dma_addr, size);
370 * Allocates bounce buffer and returns its kernel virtual address.
373 map_single(struct device *hwdev, struct swiotlb_phys_addr buffer, size_t size, int dir)
377 unsigned int nslots, stride, index, wrap;
379 unsigned long start_dma_addr;
381 unsigned long offset_slots;
382 unsigned long max_slots;
383 struct swiotlb_phys_addr slot_buf;
385 mask = dma_get_seg_boundary(hwdev);
386 start_dma_addr = swiotlb_virt_to_bus(io_tlb_start) & mask;
388 offset_slots = ALIGN(start_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
391 * Carefully handle integer overflow which can occur when mask == ~0UL.
394 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
395 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
398 * For mappings greater than a page, we limit the stride (and
399 * hence alignment) to a page size.
401 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
402 if (size > PAGE_SIZE)
403 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
410 * Find suitable number of IO TLB entries size that will fit this
411 * request and allocate a buffer from that IO TLB pool.
413 spin_lock_irqsave(&io_tlb_lock, flags);
414 index = ALIGN(io_tlb_index, stride);
415 if (index >= io_tlb_nslabs)
420 while (iommu_is_span_boundary(index, nslots, offset_slots,
423 if (index >= io_tlb_nslabs)
430 * If we find a slot that indicates we have 'nslots' number of
431 * contiguous buffers, we allocate the buffers from that slot
432 * and mark the entries as '0' indicating unavailable.
434 if (io_tlb_list[index] >= nslots) {
437 for (i = index; i < (int) (index + nslots); i++)
439 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
440 io_tlb_list[i] = ++count;
441 dma_addr = io_tlb_start + (index << IO_TLB_SHIFT);
444 * Update the indices to avoid searching in the next
447 io_tlb_index = ((index + nslots) < io_tlb_nslabs
448 ? (index + nslots) : 0);
453 if (index >= io_tlb_nslabs)
455 } while (index != wrap);
458 spin_unlock_irqrestore(&io_tlb_lock, flags);
461 spin_unlock_irqrestore(&io_tlb_lock, flags);
464 * Save away the mapping from the original address to the DMA address.
465 * This is needed when we sync the memory. Then we sync the buffer if
469 for (i = 0; i < nslots; i++) {
470 slot_buf.page += slot_buf.offset >> PAGE_SHIFT;
471 slot_buf.offset &= PAGE_SIZE - 1;
472 io_tlb_orig_addr[index+i] = slot_buf;
473 slot_buf.offset += 1 << IO_TLB_SHIFT;
475 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
476 __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
482 * dma_addr is the kernel virtual address of the bounce buffer to unmap.
485 unmap_single(struct device *hwdev, char *dma_addr, size_t size, int dir)
488 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
489 int index = (dma_addr - io_tlb_start) >> IO_TLB_SHIFT;
490 struct swiotlb_phys_addr buffer = swiotlb_bus_to_phys_addr(dma_addr);
493 * First, sync the memory before unmapping the entry
495 if ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL))
497 * bounce... copy the data back into the original buffer * and
498 * delete the bounce buffer.
500 __sync_single(buffer, dma_addr, size, DMA_FROM_DEVICE);
503 * Return the buffer to the free list by setting the corresponding
504 * entries to indicate the number of contigous entries available.
505 * While returning the entries to the free list, we merge the entries
506 * with slots below and above the pool being returned.
508 spin_lock_irqsave(&io_tlb_lock, flags);
510 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
511 io_tlb_list[index + nslots] : 0);
513 * Step 1: return the slots to the free list, merging the
514 * slots with superceeding slots
516 for (i = index + nslots - 1; i >= index; i--)
517 io_tlb_list[i] = ++count;
519 * Step 2: merge the returned slots with the preceding slots,
520 * if available (non zero)
522 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
523 io_tlb_list[i] = ++count;
525 spin_unlock_irqrestore(&io_tlb_lock, flags);
529 sync_single(struct device *hwdev, char *dma_addr, size_t size,
532 struct swiotlb_phys_addr buffer = swiotlb_bus_to_phys_addr(dma_addr);
536 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
537 __sync_single(buffer, dma_addr, size, DMA_FROM_DEVICE);
539 BUG_ON(dir != DMA_TO_DEVICE);
541 case SYNC_FOR_DEVICE:
542 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
543 __sync_single(buffer, dma_addr, size, DMA_TO_DEVICE);
545 BUG_ON(dir != DMA_FROM_DEVICE);
553 swiotlb_alloc_coherent(struct device *hwdev, size_t size,
554 dma_addr_t *dma_handle, gfp_t flags)
558 int order = get_order(size);
559 u64 dma_mask = DMA_32BIT_MASK;
561 if (hwdev && hwdev->coherent_dma_mask)
562 dma_mask = hwdev->coherent_dma_mask;
564 ret = (void *)__get_free_pages(flags, order);
565 if (ret && !is_buffer_dma_capable(dma_mask, swiotlb_virt_to_bus(ret), size)) {
567 * The allocated memory isn't reachable by the device.
568 * Fall back on swiotlb_map_single().
570 free_pages((unsigned long) ret, order);
575 * We are either out of memory or the device can't DMA
576 * to GFP_DMA memory; fall back on
577 * swiotlb_map_single(), which will grab memory from
578 * the lowest available address range.
580 struct swiotlb_phys_addr buffer;
581 buffer.page = virt_to_page(NULL);
583 ret = map_single(hwdev, buffer, size, DMA_FROM_DEVICE);
588 memset(ret, 0, size);
589 dev_addr = swiotlb_virt_to_bus(ret);
591 /* Confirm address can be DMA'd by device */
592 if (!is_buffer_dma_capable(dma_mask, dev_addr, size)) {
593 printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
594 (unsigned long long)dma_mask,
595 (unsigned long long)dev_addr);
597 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
598 unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
601 *dma_handle = dev_addr;
606 swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
607 dma_addr_t dma_handle)
609 WARN_ON(irqs_disabled());
610 if (!is_swiotlb_buffer(vaddr))
611 free_pages((unsigned long) vaddr, get_order(size));
613 /* DMA_TO_DEVICE to avoid memcpy in unmap_single */
614 unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
618 swiotlb_full(struct device *dev, size_t size, int dir, int do_panic)
621 * Ran out of IOMMU space for this operation. This is very bad.
622 * Unfortunately the drivers cannot handle this operation properly.
623 * unless they check for dma_mapping_error (most don't)
624 * When the mapping is small enough return a static buffer to limit
625 * the damage, or panic when the transfer is too big.
627 printk(KERN_ERR "DMA: Out of SW-IOMMU space for %zu bytes at "
628 "device %s\n", size, dev ? dev->bus_id : "?");
630 if (size > io_tlb_overflow && do_panic) {
631 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
632 panic("DMA: Memory would be corrupted\n");
633 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
634 panic("DMA: Random memory would be DMAed\n");
639 * Map a single buffer of the indicated size for DMA in streaming mode. The
640 * physical address to use is returned.
642 * Once the device is given the dma address, the device owns this memory until
643 * either swiotlb_unmap_single or swiotlb_dma_sync_single is performed.
646 swiotlb_map_single_attrs(struct device *hwdev, void *ptr, size_t size,
647 int dir, struct dma_attrs *attrs)
649 dma_addr_t dev_addr = swiotlb_virt_to_bus(ptr);
651 struct swiotlb_phys_addr buffer;
653 BUG_ON(dir == DMA_NONE);
655 * If the pointer passed in happens to be in the device's DMA window,
656 * we can safely return the device addr and not worry about bounce
659 if (!address_needs_mapping(hwdev, dev_addr, size) &&
660 !range_needs_mapping(ptr, size))
664 * Oh well, have to allocate and map a bounce buffer.
666 buffer.page = virt_to_page(ptr);
667 buffer.offset = (unsigned long)ptr & ~PAGE_MASK;
668 map = map_single(hwdev, buffer, size, dir);
670 swiotlb_full(hwdev, size, dir, 1);
671 map = io_tlb_overflow_buffer;
674 dev_addr = swiotlb_virt_to_bus(map);
677 * Ensure that the address returned is DMA'ble
679 if (address_needs_mapping(hwdev, dev_addr, size))
680 panic("map_single: bounce buffer is not DMA'ble");
684 EXPORT_SYMBOL(swiotlb_map_single_attrs);
687 swiotlb_map_single(struct device *hwdev, void *ptr, size_t size, int dir)
689 return swiotlb_map_single_attrs(hwdev, ptr, size, dir, NULL);
693 * Unmap a single streaming mode DMA translation. The dma_addr and size must
694 * match what was provided for in a previous swiotlb_map_single call. All
695 * other usages are undefined.
697 * After this call, reads by the cpu to the buffer are guaranteed to see
698 * whatever the device wrote there.
701 swiotlb_unmap_single_attrs(struct device *hwdev, dma_addr_t dev_addr,
702 size_t size, int dir, struct dma_attrs *attrs)
704 char *dma_addr = swiotlb_bus_to_virt(dev_addr);
706 BUG_ON(dir == DMA_NONE);
707 if (is_swiotlb_buffer(dma_addr))
708 unmap_single(hwdev, dma_addr, size, dir);
709 else if (dir == DMA_FROM_DEVICE)
710 dma_mark_clean(dma_addr, size);
712 EXPORT_SYMBOL(swiotlb_unmap_single_attrs);
715 swiotlb_unmap_single(struct device *hwdev, dma_addr_t dev_addr, size_t size,
718 return swiotlb_unmap_single_attrs(hwdev, dev_addr, size, dir, NULL);
721 * Make physical memory consistent for a single streaming mode DMA translation
724 * If you perform a swiotlb_map_single() but wish to interrogate the buffer
725 * using the cpu, yet do not wish to teardown the dma mapping, you must
726 * call this function before doing so. At the next point you give the dma
727 * address back to the card, you must first perform a
728 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
731 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
732 size_t size, int dir, int target)
734 char *dma_addr = swiotlb_bus_to_virt(dev_addr);
736 BUG_ON(dir == DMA_NONE);
737 if (is_swiotlb_buffer(dma_addr))
738 sync_single(hwdev, dma_addr, size, dir, target);
739 else if (dir == DMA_FROM_DEVICE)
740 dma_mark_clean(dma_addr, size);
744 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
745 size_t size, int dir)
747 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
751 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
752 size_t size, int dir)
754 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
758 * Same as above, but for a sub-range of the mapping.
761 swiotlb_sync_single_range(struct device *hwdev, dma_addr_t dev_addr,
762 unsigned long offset, size_t size,
765 char *dma_addr = swiotlb_bus_to_virt(dev_addr) + offset;
767 BUG_ON(dir == DMA_NONE);
768 if (is_swiotlb_buffer(dma_addr))
769 sync_single(hwdev, dma_addr, size, dir, target);
770 else if (dir == DMA_FROM_DEVICE)
771 dma_mark_clean(dma_addr, size);
775 swiotlb_sync_single_range_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
776 unsigned long offset, size_t size, int dir)
778 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
783 swiotlb_sync_single_range_for_device(struct device *hwdev, dma_addr_t dev_addr,
784 unsigned long offset, size_t size, int dir)
786 swiotlb_sync_single_range(hwdev, dev_addr, offset, size, dir,
790 void swiotlb_unmap_sg_attrs(struct device *, struct scatterlist *, int, int,
793 * Map a set of buffers described by scatterlist in streaming mode for DMA.
794 * This is the scatter-gather version of the above swiotlb_map_single
795 * interface. Here the scatter gather list elements are each tagged with the
796 * appropriate dma address and length. They are obtained via
797 * sg_dma_{address,length}(SG).
799 * NOTE: An implementation may be able to use a smaller number of
800 * DMA address/length pairs than there are SG table elements.
801 * (for example via virtual mapping capabilities)
802 * The routine returns the number of addr/length pairs actually
803 * used, at most nents.
805 * Device ownership issues as mentioned above for swiotlb_map_single are the
809 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
810 int dir, struct dma_attrs *attrs)
812 struct scatterlist *sg;
813 struct swiotlb_phys_addr buffer;
817 BUG_ON(dir == DMA_NONE);
819 for_each_sg(sgl, sg, nelems, i) {
820 dev_addr = swiotlb_sg_to_bus(sg);
821 if (range_needs_mapping(sg_virt(sg), sg->length) ||
822 address_needs_mapping(hwdev, dev_addr, sg->length)) {
824 buffer.page = sg_page(sg);
825 buffer.offset = sg->offset;
826 map = map_single(hwdev, buffer, sg->length, dir);
828 /* Don't panic here, we expect map_sg users
829 to do proper error handling. */
830 swiotlb_full(hwdev, sg->length, dir, 0);
831 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
833 sgl[0].dma_length = 0;
836 sg->dma_address = swiotlb_virt_to_bus(map);
838 sg->dma_address = dev_addr;
839 sg->dma_length = sg->length;
843 EXPORT_SYMBOL(swiotlb_map_sg_attrs);
846 swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
849 return swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
853 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
854 * concerning calls here are the same as for swiotlb_unmap_single() above.
857 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
858 int nelems, int dir, struct dma_attrs *attrs)
860 struct scatterlist *sg;
863 BUG_ON(dir == DMA_NONE);
865 for_each_sg(sgl, sg, nelems, i) {
866 if (sg->dma_address != swiotlb_sg_to_bus(sg))
867 unmap_single(hwdev, swiotlb_bus_to_virt(sg->dma_address),
868 sg->dma_length, dir);
869 else if (dir == DMA_FROM_DEVICE)
870 dma_mark_clean(swiotlb_bus_to_virt(sg->dma_address), sg->dma_length);
873 EXPORT_SYMBOL(swiotlb_unmap_sg_attrs);
876 swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
879 return swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
883 * Make physical memory consistent for a set of streaming mode DMA translations
886 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
890 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
891 int nelems, int dir, int target)
893 struct scatterlist *sg;
896 BUG_ON(dir == DMA_NONE);
898 for_each_sg(sgl, sg, nelems, i) {
899 if (sg->dma_address != swiotlb_sg_to_bus(sg))
900 sync_single(hwdev, swiotlb_bus_to_virt(sg->dma_address),
901 sg->dma_length, dir, target);
902 else if (dir == DMA_FROM_DEVICE)
903 dma_mark_clean(swiotlb_bus_to_virt(sg->dma_address), sg->dma_length);
908 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
911 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
915 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
918 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
922 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
924 return (dma_addr == swiotlb_virt_to_bus(io_tlb_overflow_buffer));
928 * Return whether the given device DMA address mask can be supported
929 * properly. For example, if your device can only drive the low 24-bits
930 * during bus mastering, then you would pass 0x00ffffff as the mask to
934 swiotlb_dma_supported(struct device *hwdev, u64 mask)
936 return swiotlb_virt_to_bus(io_tlb_end - 1) <= mask;
939 EXPORT_SYMBOL(swiotlb_map_single);
940 EXPORT_SYMBOL(swiotlb_unmap_single);
941 EXPORT_SYMBOL(swiotlb_map_sg);
942 EXPORT_SYMBOL(swiotlb_unmap_sg);
943 EXPORT_SYMBOL(swiotlb_sync_single_for_cpu);
944 EXPORT_SYMBOL(swiotlb_sync_single_for_device);
945 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_cpu);
946 EXPORT_SYMBOL_GPL(swiotlb_sync_single_range_for_device);
947 EXPORT_SYMBOL(swiotlb_sync_sg_for_cpu);
948 EXPORT_SYMBOL(swiotlb_sync_sg_for_device);
949 EXPORT_SYMBOL(swiotlb_dma_mapping_error);
950 EXPORT_SYMBOL(swiotlb_alloc_coherent);
951 EXPORT_SYMBOL(swiotlb_free_coherent);
952 EXPORT_SYMBOL(swiotlb_dma_supported);