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-direct.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>
35 #include <linux/mem_encrypt.h>
36 #include <linux/set_memory.h>
41 #include <linux/init.h>
42 #include <linux/bootmem.h>
43 #include <linux/iommu-helper.h>
45 #define CREATE_TRACE_POINTS
46 #include <trace/events/swiotlb.h>
48 #define OFFSET(val,align) ((unsigned long) \
49 ( (val) & ( (align) - 1)))
51 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
54 * Minimum IO TLB size to bother booting with. Systems with mainly
55 * 64bit capable cards will only lightly use the swiotlb. If we can't
56 * allocate a contiguous 1MB, we're probably in trouble anyway.
58 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
60 enum swiotlb_force swiotlb_force;
63 * Used to do a quick range check in swiotlb_tbl_unmap_single and
64 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
67 static phys_addr_t io_tlb_start, io_tlb_end;
70 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
71 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
73 static unsigned long io_tlb_nslabs;
76 * When the IOMMU overflows we return a fallback buffer. This sets the size.
78 static unsigned long io_tlb_overflow = 32*1024;
80 static phys_addr_t io_tlb_overflow_buffer;
83 * This is a free list describing the number of free entries available from
86 static unsigned int *io_tlb_list;
87 static unsigned int io_tlb_index;
90 * Max segment that we can provide which (if pages are contingous) will
91 * not be bounced (unless SWIOTLB_FORCE is set).
93 unsigned int max_segment;
96 * We need to save away the original address corresponding to a mapped entry
97 * for the sync operations.
99 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
100 static phys_addr_t *io_tlb_orig_addr;
103 * Protect the above data structures in the map and unmap calls
105 static DEFINE_SPINLOCK(io_tlb_lock);
107 static int late_alloc;
110 setup_io_tlb_npages(char *str)
113 io_tlb_nslabs = simple_strtoul(str, &str, 0);
114 /* avoid tail segment of size < IO_TLB_SEGSIZE */
115 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
119 if (!strcmp(str, "force")) {
120 swiotlb_force = SWIOTLB_FORCE;
121 } else if (!strcmp(str, "noforce")) {
122 swiotlb_force = SWIOTLB_NO_FORCE;
128 early_param("swiotlb", setup_io_tlb_npages);
129 /* make io_tlb_overflow tunable too? */
131 unsigned long swiotlb_nr_tbl(void)
133 return io_tlb_nslabs;
135 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
137 unsigned int swiotlb_max_segment(void)
141 EXPORT_SYMBOL_GPL(swiotlb_max_segment);
143 void swiotlb_set_max_segment(unsigned int val)
145 if (swiotlb_force == SWIOTLB_FORCE)
148 max_segment = rounddown(val, PAGE_SIZE);
151 /* default to 64MB */
152 #define IO_TLB_DEFAULT_SIZE (64UL<<20)
153 unsigned long swiotlb_size_or_default(void)
157 size = io_tlb_nslabs << IO_TLB_SHIFT;
159 return size ? size : (IO_TLB_DEFAULT_SIZE);
162 static bool no_iotlb_memory;
164 void swiotlb_print_info(void)
166 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
168 if (no_iotlb_memory) {
169 pr_warn("No low mem\n");
173 pr_info("mapped [mem %#010llx-%#010llx] (%luMB)\n",
174 (unsigned long long)io_tlb_start,
175 (unsigned long long)io_tlb_end,
180 * Early SWIOTLB allocation may be too early to allow an architecture to
181 * perform the desired operations. This function allows the architecture to
182 * call SWIOTLB when the operations are possible. It needs to be called
183 * before the SWIOTLB memory is used.
185 void __init swiotlb_update_mem_attributes(void)
190 if (no_iotlb_memory || late_alloc)
193 vaddr = phys_to_virt(io_tlb_start);
194 bytes = PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT);
195 set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT);
196 memset(vaddr, 0, bytes);
198 vaddr = phys_to_virt(io_tlb_overflow_buffer);
199 bytes = PAGE_ALIGN(io_tlb_overflow);
200 set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT);
201 memset(vaddr, 0, bytes);
204 int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
206 void *v_overflow_buffer;
207 unsigned long i, bytes;
209 bytes = nslabs << IO_TLB_SHIFT;
211 io_tlb_nslabs = nslabs;
212 io_tlb_start = __pa(tlb);
213 io_tlb_end = io_tlb_start + bytes;
216 * Get the overflow emergency buffer
218 v_overflow_buffer = memblock_virt_alloc_low_nopanic(
219 PAGE_ALIGN(io_tlb_overflow),
221 if (!v_overflow_buffer)
224 io_tlb_overflow_buffer = __pa(v_overflow_buffer);
227 * Allocate and initialize the free list array. This array is used
228 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
229 * between io_tlb_start and io_tlb_end.
231 io_tlb_list = memblock_virt_alloc(
232 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)),
234 io_tlb_orig_addr = memblock_virt_alloc(
235 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)),
237 for (i = 0; i < io_tlb_nslabs; i++) {
238 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
239 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
242 no_iotlb_memory = false;
245 swiotlb_print_info();
247 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
252 * Statically reserve bounce buffer space and initialize bounce buffer data
253 * structures for the software IO TLB used to implement the DMA API.
256 swiotlb_init(int verbose)
258 size_t default_size = IO_TLB_DEFAULT_SIZE;
259 unsigned char *vstart;
262 if (!io_tlb_nslabs) {
263 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
264 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
267 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
269 /* Get IO TLB memory from the low pages */
270 vstart = memblock_virt_alloc_low_nopanic(PAGE_ALIGN(bytes), PAGE_SIZE);
271 if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
275 memblock_free_early(io_tlb_start,
276 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
279 pr_warn("Cannot allocate buffer");
280 no_iotlb_memory = true;
284 * Systems with larger DMA zones (those that don't support ISA) can
285 * initialize the swiotlb later using the slab allocator if needed.
286 * This should be just like above, but with some error catching.
289 swiotlb_late_init_with_default_size(size_t default_size)
291 unsigned long bytes, req_nslabs = io_tlb_nslabs;
292 unsigned char *vstart = NULL;
296 if (!io_tlb_nslabs) {
297 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
298 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
302 * Get IO TLB memory from the low pages
304 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
305 io_tlb_nslabs = SLABS_PER_PAGE << order;
306 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
308 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
309 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
317 io_tlb_nslabs = req_nslabs;
320 if (order != get_order(bytes)) {
321 pr_warn("only able to allocate %ld MB\n",
322 (PAGE_SIZE << order) >> 20);
323 io_tlb_nslabs = SLABS_PER_PAGE << order;
325 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
327 free_pages((unsigned long)vstart, order);
333 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
335 unsigned long i, bytes;
336 unsigned char *v_overflow_buffer;
338 bytes = nslabs << IO_TLB_SHIFT;
340 io_tlb_nslabs = nslabs;
341 io_tlb_start = virt_to_phys(tlb);
342 io_tlb_end = io_tlb_start + bytes;
344 set_memory_decrypted((unsigned long)tlb, bytes >> PAGE_SHIFT);
345 memset(tlb, 0, bytes);
348 * Get the overflow emergency buffer
350 v_overflow_buffer = (void *)__get_free_pages(GFP_DMA,
351 get_order(io_tlb_overflow));
352 if (!v_overflow_buffer)
355 set_memory_decrypted((unsigned long)v_overflow_buffer,
356 io_tlb_overflow >> PAGE_SHIFT);
357 memset(v_overflow_buffer, 0, io_tlb_overflow);
358 io_tlb_overflow_buffer = virt_to_phys(v_overflow_buffer);
361 * Allocate and initialize the free list array. This array is used
362 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
363 * between io_tlb_start and io_tlb_end.
365 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
366 get_order(io_tlb_nslabs * sizeof(int)));
370 io_tlb_orig_addr = (phys_addr_t *)
371 __get_free_pages(GFP_KERNEL,
372 get_order(io_tlb_nslabs *
373 sizeof(phys_addr_t)));
374 if (!io_tlb_orig_addr)
377 for (i = 0; i < io_tlb_nslabs; i++) {
378 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
379 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
382 no_iotlb_memory = false;
384 swiotlb_print_info();
388 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
393 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
397 free_pages((unsigned long)v_overflow_buffer,
398 get_order(io_tlb_overflow));
399 io_tlb_overflow_buffer = 0;
408 void __init swiotlb_exit(void)
410 if (!io_tlb_orig_addr)
414 free_pages((unsigned long)phys_to_virt(io_tlb_overflow_buffer),
415 get_order(io_tlb_overflow));
416 free_pages((unsigned long)io_tlb_orig_addr,
417 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
418 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
420 free_pages((unsigned long)phys_to_virt(io_tlb_start),
421 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
423 memblock_free_late(io_tlb_overflow_buffer,
424 PAGE_ALIGN(io_tlb_overflow));
425 memblock_free_late(__pa(io_tlb_orig_addr),
426 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
427 memblock_free_late(__pa(io_tlb_list),
428 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
429 memblock_free_late(io_tlb_start,
430 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
436 int is_swiotlb_buffer(phys_addr_t paddr)
438 return paddr >= io_tlb_start && paddr < io_tlb_end;
442 * Bounce: copy the swiotlb buffer back to the original dma location
444 static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
445 size_t size, enum dma_data_direction dir)
447 unsigned long pfn = PFN_DOWN(orig_addr);
448 unsigned char *vaddr = phys_to_virt(tlb_addr);
450 if (PageHighMem(pfn_to_page(pfn))) {
451 /* The buffer does not have a mapping. Map it in and copy */
452 unsigned int offset = orig_addr & ~PAGE_MASK;
458 sz = min_t(size_t, PAGE_SIZE - offset, size);
460 local_irq_save(flags);
461 buffer = kmap_atomic(pfn_to_page(pfn));
462 if (dir == DMA_TO_DEVICE)
463 memcpy(vaddr, buffer + offset, sz);
465 memcpy(buffer + offset, vaddr, sz);
466 kunmap_atomic(buffer);
467 local_irq_restore(flags);
474 } else if (dir == DMA_TO_DEVICE) {
475 memcpy(vaddr, phys_to_virt(orig_addr), size);
477 memcpy(phys_to_virt(orig_addr), vaddr, size);
481 phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
482 dma_addr_t tbl_dma_addr,
483 phys_addr_t orig_addr, size_t size,
484 enum dma_data_direction dir,
488 phys_addr_t tlb_addr;
489 unsigned int nslots, stride, index, wrap;
492 unsigned long offset_slots;
493 unsigned long max_slots;
496 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
498 if (mem_encrypt_active())
499 pr_warn_once("%s is active and system is using DMA bounce buffers\n",
500 sme_active() ? "SME" : "SEV");
502 mask = dma_get_seg_boundary(hwdev);
504 tbl_dma_addr &= mask;
506 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
509 * Carefully handle integer overflow which can occur when mask == ~0UL.
512 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
513 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
516 * For mappings greater than or equal to a page, we limit the stride
517 * (and hence alignment) to a page size.
519 nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
520 if (size >= PAGE_SIZE)
521 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
528 * Find suitable number of IO TLB entries size that will fit this
529 * request and allocate a buffer from that IO TLB pool.
531 spin_lock_irqsave(&io_tlb_lock, flags);
532 index = ALIGN(io_tlb_index, stride);
533 if (index >= io_tlb_nslabs)
538 while (iommu_is_span_boundary(index, nslots, offset_slots,
541 if (index >= io_tlb_nslabs)
548 * If we find a slot that indicates we have 'nslots' number of
549 * contiguous buffers, we allocate the buffers from that slot
550 * and mark the entries as '0' indicating unavailable.
552 if (io_tlb_list[index] >= nslots) {
555 for (i = index; i < (int) (index + nslots); i++)
557 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
558 io_tlb_list[i] = ++count;
559 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
562 * Update the indices to avoid searching in the next
565 io_tlb_index = ((index + nslots) < io_tlb_nslabs
566 ? (index + nslots) : 0);
571 if (index >= io_tlb_nslabs)
573 } while (index != wrap);
576 spin_unlock_irqrestore(&io_tlb_lock, flags);
577 if (!(attrs & DMA_ATTR_NO_WARN) && printk_ratelimit())
578 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes)\n", size);
579 return SWIOTLB_MAP_ERROR;
581 spin_unlock_irqrestore(&io_tlb_lock, flags);
584 * Save away the mapping from the original address to the DMA address.
585 * This is needed when we sync the memory. Then we sync the buffer if
588 for (i = 0; i < nslots; i++)
589 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
590 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
591 (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
592 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_TO_DEVICE);
598 * Allocates bounce buffer and returns its physical address.
601 map_single(struct device *hwdev, phys_addr_t phys, size_t size,
602 enum dma_data_direction dir, unsigned long attrs)
604 dma_addr_t start_dma_addr;
606 if (swiotlb_force == SWIOTLB_NO_FORCE) {
607 dev_warn_ratelimited(hwdev, "Cannot do DMA to address %pa\n",
609 return SWIOTLB_MAP_ERROR;
612 start_dma_addr = __phys_to_dma(hwdev, io_tlb_start);
613 return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size,
618 * tlb_addr is the physical address of the bounce buffer to unmap.
620 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
621 size_t size, enum dma_data_direction dir,
625 int i, count, nslots = ALIGN(size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
626 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
627 phys_addr_t orig_addr = io_tlb_orig_addr[index];
630 * First, sync the memory before unmapping the entry
632 if (orig_addr != INVALID_PHYS_ADDR &&
633 !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
634 ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
635 swiotlb_bounce(orig_addr, tlb_addr, size, DMA_FROM_DEVICE);
638 * Return the buffer to the free list by setting the corresponding
639 * entries to indicate the number of contiguous entries available.
640 * While returning the entries to the free list, we merge the entries
641 * with slots below and above the pool being returned.
643 spin_lock_irqsave(&io_tlb_lock, flags);
645 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
646 io_tlb_list[index + nslots] : 0);
648 * Step 1: return the slots to the free list, merging the
649 * slots with superceeding slots
651 for (i = index + nslots - 1; i >= index; i--) {
652 io_tlb_list[i] = ++count;
653 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
656 * Step 2: merge the returned slots with the preceding slots,
657 * if available (non zero)
659 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
660 io_tlb_list[i] = ++count;
662 spin_unlock_irqrestore(&io_tlb_lock, flags);
665 void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
666 size_t size, enum dma_data_direction dir,
667 enum dma_sync_target target)
669 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
670 phys_addr_t orig_addr = io_tlb_orig_addr[index];
672 if (orig_addr == INVALID_PHYS_ADDR)
674 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
678 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
679 swiotlb_bounce(orig_addr, tlb_addr,
680 size, DMA_FROM_DEVICE);
682 BUG_ON(dir != DMA_TO_DEVICE);
684 case SYNC_FOR_DEVICE:
685 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
686 swiotlb_bounce(orig_addr, tlb_addr,
687 size, DMA_TO_DEVICE);
689 BUG_ON(dir != DMA_FROM_DEVICE);
696 static inline bool dma_coherent_ok(struct device *dev, dma_addr_t addr,
699 u64 mask = DMA_BIT_MASK(32);
701 if (dev && dev->coherent_dma_mask)
702 mask = dev->coherent_dma_mask;
703 return addr + size - 1 <= mask;
707 swiotlb_alloc_buffer(struct device *dev, size_t size, dma_addr_t *dma_handle,
710 phys_addr_t phys_addr;
712 if (swiotlb_force == SWIOTLB_NO_FORCE)
715 phys_addr = swiotlb_tbl_map_single(dev,
716 __phys_to_dma(dev, io_tlb_start),
717 0, size, DMA_FROM_DEVICE, attrs);
718 if (phys_addr == SWIOTLB_MAP_ERROR)
721 *dma_handle = __phys_to_dma(dev, phys_addr);
722 if (!dma_coherent_ok(dev, *dma_handle, size))
725 memset(phys_to_virt(phys_addr), 0, size);
726 return phys_to_virt(phys_addr);
729 dev_warn(dev, "hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
730 (unsigned long long)dev->coherent_dma_mask,
731 (unsigned long long)*dma_handle);
734 * DMA_TO_DEVICE to avoid memcpy in unmap_single.
735 * DMA_ATTR_SKIP_CPU_SYNC is optional.
737 swiotlb_tbl_unmap_single(dev, phys_addr, size, DMA_TO_DEVICE,
738 DMA_ATTR_SKIP_CPU_SYNC);
740 if (!(attrs & DMA_ATTR_NO_WARN) && printk_ratelimit()) {
742 "swiotlb: coherent allocation failed, size=%zu\n",
749 static bool swiotlb_free_buffer(struct device *dev, size_t size,
752 phys_addr_t phys_addr = dma_to_phys(dev, dma_addr);
754 WARN_ON_ONCE(irqs_disabled());
756 if (!is_swiotlb_buffer(phys_addr))
760 * DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single.
761 * DMA_ATTR_SKIP_CPU_SYNC is optional.
763 swiotlb_tbl_unmap_single(dev, phys_addr, size, DMA_TO_DEVICE,
764 DMA_ATTR_SKIP_CPU_SYNC);
769 * Map a single buffer of the indicated size for DMA in streaming mode. The
770 * physical address to use is returned.
772 * Once the device is given the dma address, the device owns this memory until
773 * either swiotlb_unmap_page or swiotlb_dma_sync_single is performed.
775 dma_addr_t swiotlb_map_page(struct device *dev, struct page *page,
776 unsigned long offset, size_t size,
777 enum dma_data_direction dir,
780 phys_addr_t map, phys = page_to_phys(page) + offset;
781 dma_addr_t dev_addr = phys_to_dma(dev, phys);
783 BUG_ON(dir == DMA_NONE);
785 * If the address happens to be in the device's DMA window,
786 * we can safely return the device addr and not worry about bounce
789 if (dma_capable(dev, dev_addr, size) && swiotlb_force != SWIOTLB_FORCE)
792 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
794 /* Oh well, have to allocate and map a bounce buffer. */
795 map = map_single(dev, phys, size, dir, attrs);
796 if (map == SWIOTLB_MAP_ERROR)
797 return __phys_to_dma(dev, io_tlb_overflow_buffer);
799 dev_addr = __phys_to_dma(dev, map);
801 /* Ensure that the address returned is DMA'ble */
802 if (dma_capable(dev, dev_addr, size))
805 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
806 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
808 return __phys_to_dma(dev, io_tlb_overflow_buffer);
812 * Unmap a single streaming mode DMA translation. The dma_addr and size must
813 * match what was provided for in a previous swiotlb_map_page call. All
814 * other usages are undefined.
816 * After this call, reads by the cpu to the buffer are guaranteed to see
817 * whatever the device wrote there.
819 static void unmap_single(struct device *hwdev, dma_addr_t dev_addr,
820 size_t size, enum dma_data_direction dir,
823 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
825 BUG_ON(dir == DMA_NONE);
827 if (is_swiotlb_buffer(paddr)) {
828 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
832 if (dir != DMA_FROM_DEVICE)
836 * phys_to_virt doesn't work with hihgmem page but we could
837 * call dma_mark_clean() with hihgmem page here. However, we
838 * are fine since dma_mark_clean() is null on POWERPC. We can
839 * make dma_mark_clean() take a physical address if necessary.
841 dma_mark_clean(phys_to_virt(paddr), size);
844 void swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
845 size_t size, enum dma_data_direction dir,
848 unmap_single(hwdev, dev_addr, size, dir, attrs);
852 * Make physical memory consistent for a single streaming mode DMA translation
855 * If you perform a swiotlb_map_page() but wish to interrogate the buffer
856 * using the cpu, yet do not wish to teardown the dma mapping, you must
857 * call this function before doing so. At the next point you give the dma
858 * address back to the card, you must first perform a
859 * swiotlb_dma_sync_for_device, and then the device again owns the buffer
862 swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
863 size_t size, enum dma_data_direction dir,
864 enum dma_sync_target target)
866 phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
868 BUG_ON(dir == DMA_NONE);
870 if (is_swiotlb_buffer(paddr)) {
871 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
875 if (dir != DMA_FROM_DEVICE)
878 dma_mark_clean(phys_to_virt(paddr), size);
882 swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
883 size_t size, enum dma_data_direction dir)
885 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
889 swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
890 size_t size, enum dma_data_direction dir)
892 swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
896 * Map a set of buffers described by scatterlist in streaming mode for DMA.
897 * This is the scatter-gather version of the above swiotlb_map_page
898 * interface. Here the scatter gather list elements are each tagged with the
899 * appropriate dma address and length. They are obtained via
900 * sg_dma_{address,length}(SG).
902 * NOTE: An implementation may be able to use a smaller number of
903 * DMA address/length pairs than there are SG table elements.
904 * (for example via virtual mapping capabilities)
905 * The routine returns the number of addr/length pairs actually
906 * used, at most nents.
908 * Device ownership issues as mentioned above for swiotlb_map_page are the
912 swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, int nelems,
913 enum dma_data_direction dir, unsigned long attrs)
915 struct scatterlist *sg;
918 BUG_ON(dir == DMA_NONE);
920 for_each_sg(sgl, sg, nelems, i) {
921 phys_addr_t paddr = sg_phys(sg);
922 dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
924 if (swiotlb_force == SWIOTLB_FORCE ||
925 !dma_capable(hwdev, dev_addr, sg->length)) {
926 phys_addr_t map = map_single(hwdev, sg_phys(sg),
927 sg->length, dir, attrs);
928 if (map == SWIOTLB_MAP_ERROR) {
929 /* Don't panic here, we expect map_sg users
930 to do proper error handling. */
931 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
932 swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
937 sg->dma_address = __phys_to_dma(hwdev, map);
939 sg->dma_address = dev_addr;
940 sg_dma_len(sg) = sg->length;
946 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
947 * concerning calls here are the same as for swiotlb_unmap_page() above.
950 swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
951 int nelems, enum dma_data_direction dir,
954 struct scatterlist *sg;
957 BUG_ON(dir == DMA_NONE);
959 for_each_sg(sgl, sg, nelems, i)
960 unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir,
965 * Make physical memory consistent for a set of streaming mode DMA translations
968 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
972 swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
973 int nelems, enum dma_data_direction dir,
974 enum dma_sync_target target)
976 struct scatterlist *sg;
979 for_each_sg(sgl, sg, nelems, i)
980 swiotlb_sync_single(hwdev, sg->dma_address,
981 sg_dma_len(sg), dir, target);
985 swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
986 int nelems, enum dma_data_direction dir)
988 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
992 swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
993 int nelems, enum dma_data_direction dir)
995 swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
999 swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
1001 return (dma_addr == __phys_to_dma(hwdev, io_tlb_overflow_buffer));
1005 * Return whether the given device DMA address mask can be supported
1006 * properly. For example, if your device can only drive the low 24-bits
1007 * during bus mastering, then you would pass 0x00ffffff as the mask to
1011 swiotlb_dma_supported(struct device *hwdev, u64 mask)
1013 return __phys_to_dma(hwdev, io_tlb_end - 1) <= mask;
1016 void *swiotlb_alloc(struct device *dev, size_t size, dma_addr_t *dma_handle,
1017 gfp_t gfp, unsigned long attrs)
1021 /* temporary workaround: */
1022 if (gfp & __GFP_NOWARN)
1023 attrs |= DMA_ATTR_NO_WARN;
1026 * Don't print a warning when the first allocation attempt fails.
1027 * swiotlb_alloc_coherent() will print a warning when the DMA memory
1028 * allocation ultimately failed.
1030 gfp |= __GFP_NOWARN;
1032 vaddr = dma_direct_alloc(dev, size, dma_handle, gfp, attrs);
1034 vaddr = swiotlb_alloc_buffer(dev, size, dma_handle, attrs);
1038 void swiotlb_free(struct device *dev, size_t size, void *vaddr,
1039 dma_addr_t dma_addr, unsigned long attrs)
1041 if (!swiotlb_free_buffer(dev, size, dma_addr))
1042 dma_direct_free(dev, size, vaddr, dma_addr, attrs);
1045 const struct dma_map_ops swiotlb_dma_ops = {
1046 .mapping_error = swiotlb_dma_mapping_error,
1047 .alloc = swiotlb_alloc,
1048 .free = swiotlb_free,
1049 .sync_single_for_cpu = swiotlb_sync_single_for_cpu,
1050 .sync_single_for_device = swiotlb_sync_single_for_device,
1051 .sync_sg_for_cpu = swiotlb_sync_sg_for_cpu,
1052 .sync_sg_for_device = swiotlb_sync_sg_for_device,
1053 .map_sg = swiotlb_map_sg_attrs,
1054 .unmap_sg = swiotlb_unmap_sg_attrs,
1055 .map_page = swiotlb_map_page,
1056 .unmap_page = swiotlb_unmap_page,
1057 .dma_supported = dma_direct_supported,
1059 EXPORT_SYMBOL(swiotlb_dma_ops);