1 // SPDX-License-Identifier: GPL-2.0-only
3 * Dynamic DMA mapping support.
5 * This implementation is a fallback for platforms that do not support
6 * I/O TLBs (aka DMA address translation hardware).
7 * Copyright (C) 2000 Asit Mallick <Asit.K.Mallick@intel.com>
8 * Copyright (C) 2000 Goutham Rao <goutham.rao@intel.com>
9 * Copyright (C) 2000, 2003 Hewlett-Packard Co
10 * David Mosberger-Tang <davidm@hpl.hp.com>
12 * 03/05/07 davidm Switch from PCI-DMA to generic device DMA API.
13 * 00/12/13 davidm Rename to swiotlb.c and add mark_clean() to avoid
14 * unnecessary i-cache flushing.
15 * 04/07/.. ak Better overflow handling. Assorted fixes.
16 * 05/09/10 linville Add support for syncing ranges, support syncing for
17 * DMA_BIDIRECTIONAL mappings, miscellaneous cleanup.
18 * 08/12/11 beckyb Add highmem support
21 #define pr_fmt(fmt) "software IO TLB: " fmt
23 #include <linux/cache.h>
24 #include <linux/cc_platform.h>
25 #include <linux/ctype.h>
26 #include <linux/debugfs.h>
27 #include <linux/dma-direct.h>
28 #include <linux/dma-map-ops.h>
29 #include <linux/export.h>
30 #include <linux/gfp.h>
31 #include <linux/highmem.h>
33 #include <linux/iommu-helper.h>
34 #include <linux/init.h>
35 #include <linux/memblock.h>
37 #include <linux/pfn.h>
38 #include <linux/scatterlist.h>
39 #include <linux/set_memory.h>
40 #include <linux/spinlock.h>
41 #include <linux/string.h>
42 #include <linux/swiotlb.h>
43 #include <linux/types.h>
44 #ifdef CONFIG_DMA_RESTRICTED_POOL
46 #include <linux/of_fdt.h>
47 #include <linux/of_reserved_mem.h>
48 #include <linux/slab.h>
51 #define CREATE_TRACE_POINTS
52 #include <trace/events/swiotlb.h>
54 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
57 * Minimum IO TLB size to bother booting with. Systems with mainly
58 * 64bit capable cards will only lightly use the swiotlb. If we can't
59 * allocate a contiguous 1MB, we're probably in trouble anyway.
61 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
63 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
66 phys_addr_t orig_addr;
71 static bool swiotlb_force_bounce;
72 static bool swiotlb_force_disable;
74 struct io_tlb_mem io_tlb_default_mem;
76 phys_addr_t swiotlb_unencrypted_base;
78 static unsigned long default_nslabs = IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT;
79 static unsigned long default_nareas;
82 * struct io_tlb_area - IO TLB memory area descriptor
84 * This is a single area with a single lock.
86 * @used: The number of used IO TLB block.
87 * @index: The slot index to start searching in this area for next round.
88 * @lock: The lock to protect the above data structures in the map and
98 * Round up number of slabs to the next power of 2. The last area is going
99 * be smaller than the rest if default_nslabs is not power of two.
100 * The number of slot in an area should be a multiple of IO_TLB_SEGSIZE,
101 * otherwise a segment may span two or more areas. It conflicts with free
102 * contiguous slots tracking: free slots are treated contiguous no matter
103 * whether they cross an area boundary.
105 * Return true if default_nslabs is rounded up.
107 static bool round_up_default_nslabs(void)
112 if (default_nslabs < IO_TLB_SEGSIZE * default_nareas)
113 default_nslabs = IO_TLB_SEGSIZE * default_nareas;
114 else if (is_power_of_2(default_nslabs))
116 default_nslabs = roundup_pow_of_two(default_nslabs);
120 static void swiotlb_adjust_nareas(unsigned int nareas)
122 /* use a single area when non is specified */
125 else if (!is_power_of_2(nareas))
126 nareas = roundup_pow_of_two(nareas);
128 default_nareas = nareas;
130 pr_info("area num %d.\n", nareas);
131 if (round_up_default_nslabs())
132 pr_info("SWIOTLB bounce buffer size roundup to %luMB",
133 (default_nslabs << IO_TLB_SHIFT) >> 20);
137 setup_io_tlb_npages(char *str)
140 /* avoid tail segment of size < IO_TLB_SEGSIZE */
142 ALIGN(simple_strtoul(str, &str, 0), IO_TLB_SEGSIZE);
147 swiotlb_adjust_nareas(simple_strtoul(str, &str, 0));
150 if (!strcmp(str, "force"))
151 swiotlb_force_bounce = true;
152 else if (!strcmp(str, "noforce"))
153 swiotlb_force_disable = true;
157 early_param("swiotlb", setup_io_tlb_npages);
159 unsigned int swiotlb_max_segment(void)
161 if (!io_tlb_default_mem.nslabs)
163 return rounddown(io_tlb_default_mem.nslabs << IO_TLB_SHIFT, PAGE_SIZE);
165 EXPORT_SYMBOL_GPL(swiotlb_max_segment);
167 unsigned long swiotlb_size_or_default(void)
169 return default_nslabs << IO_TLB_SHIFT;
172 void __init swiotlb_adjust_size(unsigned long size)
175 * If swiotlb parameter has not been specified, give a chance to
176 * architectures such as those supporting memory encryption to
177 * adjust/expand SWIOTLB size for their use.
179 if (default_nslabs != IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT)
182 size = ALIGN(size, IO_TLB_SIZE);
183 default_nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE);
184 if (round_up_default_nslabs())
185 size = default_nslabs << IO_TLB_SHIFT;
186 pr_info("SWIOTLB bounce buffer size adjusted to %luMB", size >> 20);
189 void swiotlb_print_info(void)
191 struct io_tlb_mem *mem = &io_tlb_default_mem;
194 pr_warn("No low mem\n");
198 pr_info("mapped [mem %pa-%pa] (%luMB)\n", &mem->start, &mem->end,
199 (mem->nslabs << IO_TLB_SHIFT) >> 20);
202 static inline unsigned long io_tlb_offset(unsigned long val)
204 return val & (IO_TLB_SEGSIZE - 1);
207 static inline unsigned long nr_slots(u64 val)
209 return DIV_ROUND_UP(val, IO_TLB_SIZE);
213 * Remap swioltb memory in the unencrypted physical address space
214 * when swiotlb_unencrypted_base is set. (e.g. for Hyper-V AMD SEV-SNP
217 #ifdef CONFIG_HAS_IOMEM
218 static void *swiotlb_mem_remap(struct io_tlb_mem *mem, unsigned long bytes)
222 if (swiotlb_unencrypted_base) {
223 phys_addr_t paddr = mem->start + swiotlb_unencrypted_base;
225 vaddr = memremap(paddr, bytes, MEMREMAP_WB);
227 pr_err("Failed to map the unencrypted memory %pa size %lx.\n",
234 static void *swiotlb_mem_remap(struct io_tlb_mem *mem, unsigned long bytes)
241 * Early SWIOTLB allocation may be too early to allow an architecture to
242 * perform the desired operations. This function allows the architecture to
243 * call SWIOTLB when the operations are possible. It needs to be called
244 * before the SWIOTLB memory is used.
246 void __init swiotlb_update_mem_attributes(void)
248 struct io_tlb_mem *mem = &io_tlb_default_mem;
252 if (!mem->nslabs || mem->late_alloc)
254 vaddr = phys_to_virt(mem->start);
255 bytes = PAGE_ALIGN(mem->nslabs << IO_TLB_SHIFT);
256 set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT);
258 mem->vaddr = swiotlb_mem_remap(mem, bytes);
263 static void swiotlb_init_io_tlb_mem(struct io_tlb_mem *mem, phys_addr_t start,
264 unsigned long nslabs, unsigned int flags,
265 bool late_alloc, unsigned int nareas)
267 void *vaddr = phys_to_virt(start);
268 unsigned long bytes = nslabs << IO_TLB_SHIFT, i;
270 mem->nslabs = nslabs;
272 mem->end = mem->start + bytes;
273 mem->late_alloc = late_alloc;
274 mem->nareas = nareas;
275 mem->area_nslabs = nslabs / mem->nareas;
277 mem->force_bounce = swiotlb_force_bounce || (flags & SWIOTLB_FORCE);
279 for (i = 0; i < mem->nareas; i++) {
280 spin_lock_init(&mem->areas[i].lock);
281 mem->areas[i].index = 0;
282 mem->areas[i].used = 0;
285 for (i = 0; i < mem->nslabs; i++) {
286 mem->slots[i].list = IO_TLB_SEGSIZE - io_tlb_offset(i);
287 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
288 mem->slots[i].alloc_size = 0;
292 * If swiotlb_unencrypted_base is set, the bounce buffer memory will
293 * be remapped and cleared in swiotlb_update_mem_attributes.
295 if (swiotlb_unencrypted_base)
298 memset(vaddr, 0, bytes);
303 static void *swiotlb_memblock_alloc(unsigned long nslabs, unsigned int flags,
304 int (*remap)(void *tlb, unsigned long nslabs))
306 size_t bytes = PAGE_ALIGN(nslabs << IO_TLB_SHIFT);
310 * By default allocate the bounce buffer memory from low memory, but
311 * allow to pick a location everywhere for hypervisors with guest
314 if (flags & SWIOTLB_ANY)
315 tlb = memblock_alloc(bytes, PAGE_SIZE);
317 tlb = memblock_alloc_low(bytes, PAGE_SIZE);
320 pr_warn("%s: Failed to allocate %zu bytes tlb structure\n",
325 if (remap && remap(tlb, nslabs) < 0) {
326 memblock_free(tlb, PAGE_ALIGN(bytes));
327 pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
335 * Statically reserve bounce buffer space and initialize bounce buffer data
336 * structures for the software IO TLB used to implement the DMA API.
338 void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags,
339 int (*remap)(void *tlb, unsigned long nslabs))
341 struct io_tlb_mem *mem = &io_tlb_default_mem;
342 unsigned long nslabs;
346 if (!addressing_limit && !swiotlb_force_bounce)
348 if (swiotlb_force_disable)
352 * default_nslabs maybe changed when adjust area number.
353 * So allocate bounce buffer after adjusting area number.
356 swiotlb_adjust_nareas(num_possible_cpus());
358 nslabs = default_nslabs;
359 while ((tlb = swiotlb_memblock_alloc(nslabs, flags, remap)) == NULL) {
360 if (nslabs <= IO_TLB_MIN_SLABS)
362 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
365 if (default_nslabs != nslabs) {
366 pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs",
367 default_nslabs, nslabs);
368 default_nslabs = nslabs;
371 alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs));
372 mem->slots = memblock_alloc(alloc_size, PAGE_SIZE);
374 pr_warn("%s: Failed to allocate %zu bytes align=0x%lx\n",
375 __func__, alloc_size, PAGE_SIZE);
379 mem->areas = memblock_alloc(array_size(sizeof(struct io_tlb_area),
380 default_nareas), SMP_CACHE_BYTES);
382 pr_warn("%s: Failed to allocate mem->areas.\n", __func__);
386 swiotlb_init_io_tlb_mem(mem, __pa(tlb), nslabs, flags, false,
389 if (flags & SWIOTLB_VERBOSE)
390 swiotlb_print_info();
393 void __init swiotlb_init(bool addressing_limit, unsigned int flags)
395 swiotlb_init_remap(addressing_limit, flags, NULL);
399 * Systems with larger DMA zones (those that don't support ISA) can
400 * initialize the swiotlb later using the slab allocator if needed.
401 * This should be just like above, but with some error catching.
403 int swiotlb_init_late(size_t size, gfp_t gfp_mask,
404 int (*remap)(void *tlb, unsigned long nslabs))
406 struct io_tlb_mem *mem = &io_tlb_default_mem;
407 unsigned long nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE);
408 unsigned char *vstart = NULL;
409 unsigned int order, area_order;
410 bool retried = false;
413 if (swiotlb_force_disable)
417 order = get_order(nslabs << IO_TLB_SHIFT);
418 nslabs = SLABS_PER_PAGE << order;
420 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
421 vstart = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
426 nslabs = SLABS_PER_PAGE << order;
434 rc = remap(vstart, nslabs);
436 free_pages((unsigned long)vstart, order);
438 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
439 if (nslabs < IO_TLB_MIN_SLABS)
446 pr_warn("only able to allocate %ld MB\n",
447 (PAGE_SIZE << order) >> 20);
451 swiotlb_adjust_nareas(num_possible_cpus());
453 area_order = get_order(array_size(sizeof(*mem->areas),
455 mem->areas = (struct io_tlb_area *)
456 __get_free_pages(GFP_KERNEL | __GFP_ZERO, area_order);
460 mem->slots = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
461 get_order(array_size(sizeof(*mem->slots), nslabs)));
465 set_memory_decrypted((unsigned long)vstart,
466 (nslabs << IO_TLB_SHIFT) >> PAGE_SHIFT);
467 swiotlb_init_io_tlb_mem(mem, virt_to_phys(vstart), nslabs, 0, true,
470 swiotlb_print_info();
474 free_pages((unsigned long)mem->areas, area_order);
476 free_pages((unsigned long)vstart, order);
480 void __init swiotlb_exit(void)
482 struct io_tlb_mem *mem = &io_tlb_default_mem;
483 unsigned long tbl_vaddr;
484 size_t tbl_size, slots_size;
485 unsigned int area_order;
487 if (swiotlb_force_bounce)
493 pr_info("tearing down default memory pool\n");
494 tbl_vaddr = (unsigned long)phys_to_virt(mem->start);
495 tbl_size = PAGE_ALIGN(mem->end - mem->start);
496 slots_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), mem->nslabs));
498 set_memory_encrypted(tbl_vaddr, tbl_size >> PAGE_SHIFT);
499 if (mem->late_alloc) {
500 area_order = get_order(array_size(sizeof(*mem->areas),
502 free_pages((unsigned long)mem->areas, area_order);
503 free_pages(tbl_vaddr, get_order(tbl_size));
504 free_pages((unsigned long)mem->slots, get_order(slots_size));
506 memblock_free_late(__pa(mem->areas),
507 array_size(sizeof(*mem->areas), mem->nareas));
508 memblock_free_late(mem->start, tbl_size);
509 memblock_free_late(__pa(mem->slots), slots_size);
512 memset(mem, 0, sizeof(*mem));
516 * Return the offset into a iotlb slot required to keep the device happy.
518 static unsigned int swiotlb_align_offset(struct device *dev, u64 addr)
520 return addr & dma_get_min_align_mask(dev) & (IO_TLB_SIZE - 1);
524 * Bounce: copy the swiotlb buffer from or back to the original dma location
526 static void swiotlb_bounce(struct device *dev, phys_addr_t tlb_addr, size_t size,
527 enum dma_data_direction dir)
529 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
530 int index = (tlb_addr - mem->start) >> IO_TLB_SHIFT;
531 phys_addr_t orig_addr = mem->slots[index].orig_addr;
532 size_t alloc_size = mem->slots[index].alloc_size;
533 unsigned long pfn = PFN_DOWN(orig_addr);
534 unsigned char *vaddr = mem->vaddr + tlb_addr - mem->start;
535 unsigned int tlb_offset, orig_addr_offset;
537 if (orig_addr == INVALID_PHYS_ADDR)
540 tlb_offset = tlb_addr & (IO_TLB_SIZE - 1);
541 orig_addr_offset = swiotlb_align_offset(dev, orig_addr);
542 if (tlb_offset < orig_addr_offset) {
543 dev_WARN_ONCE(dev, 1,
544 "Access before mapping start detected. orig offset %u, requested offset %u.\n",
545 orig_addr_offset, tlb_offset);
549 tlb_offset -= orig_addr_offset;
550 if (tlb_offset > alloc_size) {
551 dev_WARN_ONCE(dev, 1,
552 "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu+%u.\n",
553 alloc_size, size, tlb_offset);
557 orig_addr += tlb_offset;
558 alloc_size -= tlb_offset;
560 if (size > alloc_size) {
561 dev_WARN_ONCE(dev, 1,
562 "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu.\n",
567 if (PageHighMem(pfn_to_page(pfn))) {
568 unsigned int offset = orig_addr & ~PAGE_MASK;
574 sz = min_t(size_t, PAGE_SIZE - offset, size);
576 local_irq_save(flags);
577 page = pfn_to_page(pfn);
578 if (dir == DMA_TO_DEVICE)
579 memcpy_from_page(vaddr, page, offset, sz);
581 memcpy_to_page(page, offset, vaddr, sz);
582 local_irq_restore(flags);
589 } else if (dir == DMA_TO_DEVICE) {
590 memcpy(vaddr, phys_to_virt(orig_addr), size);
592 memcpy(phys_to_virt(orig_addr), vaddr, size);
596 static inline phys_addr_t slot_addr(phys_addr_t start, phys_addr_t idx)
598 return start + (idx << IO_TLB_SHIFT);
602 * Carefully handle integer overflow which can occur when boundary_mask == ~0UL.
604 static inline unsigned long get_max_slots(unsigned long boundary_mask)
606 if (boundary_mask == ~0UL)
607 return 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
608 return nr_slots(boundary_mask + 1);
611 static unsigned int wrap_area_index(struct io_tlb_mem *mem, unsigned int index)
613 if (index >= mem->area_nslabs)
619 * Find a suitable number of IO TLB entries size that will fit this request and
620 * allocate a buffer from that IO TLB pool.
622 static int swiotlb_do_find_slots(struct device *dev, int area_index,
623 phys_addr_t orig_addr, size_t alloc_size,
624 unsigned int alloc_align_mask)
626 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
627 struct io_tlb_area *area = mem->areas + area_index;
628 unsigned long boundary_mask = dma_get_seg_boundary(dev);
629 dma_addr_t tbl_dma_addr =
630 phys_to_dma_unencrypted(dev, mem->start) & boundary_mask;
631 unsigned long max_slots = get_max_slots(boundary_mask);
632 unsigned int iotlb_align_mask =
633 dma_get_min_align_mask(dev) & ~(IO_TLB_SIZE - 1);
634 unsigned int nslots = nr_slots(alloc_size), stride;
635 unsigned int index, wrap, count = 0, i;
636 unsigned int offset = swiotlb_align_offset(dev, orig_addr);
638 unsigned int slot_base;
639 unsigned int slot_index;
642 BUG_ON(area_index >= mem->nareas);
645 * For mappings with an alignment requirement don't bother looping to
646 * unaligned slots once we found an aligned one. For allocations of
647 * PAGE_SIZE or larger only look for page aligned allocations.
649 stride = (iotlb_align_mask >> IO_TLB_SHIFT) + 1;
650 if (alloc_size >= PAGE_SIZE)
651 stride = max(stride, stride << (PAGE_SHIFT - IO_TLB_SHIFT));
652 stride = max(stride, (alloc_align_mask >> IO_TLB_SHIFT) + 1);
654 spin_lock_irqsave(&area->lock, flags);
655 if (unlikely(nslots > mem->area_nslabs - area->used))
658 slot_base = area_index * mem->area_nslabs;
659 index = wrap = wrap_area_index(mem, ALIGN(area->index, stride));
662 slot_index = slot_base + index;
665 (slot_addr(tbl_dma_addr, slot_index) &
666 iotlb_align_mask) != (orig_addr & iotlb_align_mask)) {
667 index = wrap_area_index(mem, index + 1);
672 * If we find a slot that indicates we have 'nslots' number of
673 * contiguous buffers, we allocate the buffers from that slot
674 * and mark the entries as '0' indicating unavailable.
676 if (!iommu_is_span_boundary(slot_index, nslots,
677 nr_slots(tbl_dma_addr),
679 if (mem->slots[slot_index].list >= nslots)
682 index = wrap_area_index(mem, index + stride);
683 } while (index != wrap);
686 spin_unlock_irqrestore(&area->lock, flags);
690 for (i = slot_index; i < slot_index + nslots; i++) {
691 mem->slots[i].list = 0;
692 mem->slots[i].alloc_size = alloc_size - (offset +
693 ((i - slot_index) << IO_TLB_SHIFT));
695 for (i = slot_index - 1;
696 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 &&
697 mem->slots[i].list; i--)
698 mem->slots[i].list = ++count;
701 * Update the indices to avoid searching in the next round.
703 if (index + nslots < mem->area_nslabs)
704 area->index = index + nslots;
707 area->used += nslots;
708 spin_unlock_irqrestore(&area->lock, flags);
712 static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr,
713 size_t alloc_size, unsigned int alloc_align_mask)
715 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
716 int start = raw_smp_processor_id() & (mem->nareas - 1);
717 int i = start, index;
720 index = swiotlb_do_find_slots(dev, i, orig_addr, alloc_size,
724 if (++i >= mem->nareas)
726 } while (i != start);
731 static unsigned long mem_used(struct io_tlb_mem *mem)
734 unsigned long used = 0;
736 for (i = 0; i < mem->nareas; i++)
737 used += mem->areas[i].used;
741 phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
742 size_t mapping_size, size_t alloc_size,
743 unsigned int alloc_align_mask, enum dma_data_direction dir,
746 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
747 unsigned int offset = swiotlb_align_offset(dev, orig_addr);
750 phys_addr_t tlb_addr;
752 if (!mem || !mem->nslabs) {
753 dev_warn_ratelimited(dev,
754 "Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
755 return (phys_addr_t)DMA_MAPPING_ERROR;
758 if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
759 pr_warn_once("Memory encryption is active and system is using DMA bounce buffers\n");
761 if (mapping_size > alloc_size) {
762 dev_warn_once(dev, "Invalid sizes (mapping: %zd bytes, alloc: %zd bytes)",
763 mapping_size, alloc_size);
764 return (phys_addr_t)DMA_MAPPING_ERROR;
767 index = swiotlb_find_slots(dev, orig_addr,
768 alloc_size + offset, alloc_align_mask);
770 if (!(attrs & DMA_ATTR_NO_WARN))
771 dev_warn_ratelimited(dev,
772 "swiotlb buffer is full (sz: %zd bytes), total %lu (slots), used %lu (slots)\n",
773 alloc_size, mem->nslabs, mem_used(mem));
774 return (phys_addr_t)DMA_MAPPING_ERROR;
778 * Save away the mapping from the original address to the DMA address.
779 * This is needed when we sync the memory. Then we sync the buffer if
782 for (i = 0; i < nr_slots(alloc_size + offset); i++)
783 mem->slots[index + i].orig_addr = slot_addr(orig_addr, i);
784 tlb_addr = slot_addr(mem->start, index) + offset;
786 * When dir == DMA_FROM_DEVICE we could omit the copy from the orig
787 * to the tlb buffer, if we knew for sure the device will
788 * overwrite the entire current content. But we don't. Thus
789 * unconditional bounce may prevent leaking swiotlb content (i.e.
790 * kernel memory) to user-space.
792 swiotlb_bounce(dev, tlb_addr, mapping_size, DMA_TO_DEVICE);
796 static void swiotlb_release_slots(struct device *dev, phys_addr_t tlb_addr)
798 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
800 unsigned int offset = swiotlb_align_offset(dev, tlb_addr);
801 int index = (tlb_addr - offset - mem->start) >> IO_TLB_SHIFT;
802 int nslots = nr_slots(mem->slots[index].alloc_size + offset);
803 int aindex = index / mem->area_nslabs;
804 struct io_tlb_area *area = &mem->areas[aindex];
808 * Return the buffer to the free list by setting the corresponding
809 * entries to indicate the number of contiguous entries available.
810 * While returning the entries to the free list, we merge the entries
811 * with slots below and above the pool being returned.
813 BUG_ON(aindex >= mem->nareas);
815 spin_lock_irqsave(&area->lock, flags);
816 if (index + nslots < ALIGN(index + 1, IO_TLB_SEGSIZE))
817 count = mem->slots[index + nslots].list;
822 * Step 1: return the slots to the free list, merging the slots with
825 for (i = index + nslots - 1; i >= index; i--) {
826 mem->slots[i].list = ++count;
827 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
828 mem->slots[i].alloc_size = 0;
832 * Step 2: merge the returned slots with the preceding slots, if
833 * available (non zero)
836 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 && mem->slots[i].list;
838 mem->slots[i].list = ++count;
839 area->used -= nslots;
840 spin_unlock_irqrestore(&area->lock, flags);
844 * tlb_addr is the physical address of the bounce buffer to unmap.
846 void swiotlb_tbl_unmap_single(struct device *dev, phys_addr_t tlb_addr,
847 size_t mapping_size, enum dma_data_direction dir,
851 * First, sync the memory before unmapping the entry
853 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
854 (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
855 swiotlb_bounce(dev, tlb_addr, mapping_size, DMA_FROM_DEVICE);
857 swiotlb_release_slots(dev, tlb_addr);
860 void swiotlb_sync_single_for_device(struct device *dev, phys_addr_t tlb_addr,
861 size_t size, enum dma_data_direction dir)
863 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
864 swiotlb_bounce(dev, tlb_addr, size, DMA_TO_DEVICE);
866 BUG_ON(dir != DMA_FROM_DEVICE);
869 void swiotlb_sync_single_for_cpu(struct device *dev, phys_addr_t tlb_addr,
870 size_t size, enum dma_data_direction dir)
872 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
873 swiotlb_bounce(dev, tlb_addr, size, DMA_FROM_DEVICE);
875 BUG_ON(dir != DMA_TO_DEVICE);
879 * Create a swiotlb mapping for the buffer at @paddr, and in case of DMAing
880 * to the device copy the data into it as well.
882 dma_addr_t swiotlb_map(struct device *dev, phys_addr_t paddr, size_t size,
883 enum dma_data_direction dir, unsigned long attrs)
885 phys_addr_t swiotlb_addr;
888 trace_swiotlb_bounced(dev, phys_to_dma(dev, paddr), size);
890 swiotlb_addr = swiotlb_tbl_map_single(dev, paddr, size, size, 0, dir,
892 if (swiotlb_addr == (phys_addr_t)DMA_MAPPING_ERROR)
893 return DMA_MAPPING_ERROR;
895 /* Ensure that the address returned is DMA'ble */
896 dma_addr = phys_to_dma_unencrypted(dev, swiotlb_addr);
897 if (unlikely(!dma_capable(dev, dma_addr, size, true))) {
898 swiotlb_tbl_unmap_single(dev, swiotlb_addr, size, dir,
899 attrs | DMA_ATTR_SKIP_CPU_SYNC);
900 dev_WARN_ONCE(dev, 1,
901 "swiotlb addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
902 &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
903 return DMA_MAPPING_ERROR;
906 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
907 arch_sync_dma_for_device(swiotlb_addr, size, dir);
911 size_t swiotlb_max_mapping_size(struct device *dev)
913 int min_align_mask = dma_get_min_align_mask(dev);
917 * swiotlb_find_slots() skips slots according to
918 * min align mask. This affects max mapping size.
919 * Take it into acount here.
922 min_align = roundup(min_align_mask, IO_TLB_SIZE);
924 return ((size_t)IO_TLB_SIZE) * IO_TLB_SEGSIZE - min_align;
927 bool is_swiotlb_active(struct device *dev)
929 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
931 return mem && mem->nslabs;
933 EXPORT_SYMBOL_GPL(is_swiotlb_active);
935 static int io_tlb_used_get(void *data, u64 *val)
937 *val = mem_used(&io_tlb_default_mem);
940 DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_used, io_tlb_used_get, NULL, "%llu\n");
942 static void swiotlb_create_debugfs_files(struct io_tlb_mem *mem,
945 mem->debugfs = debugfs_create_dir(dirname, io_tlb_default_mem.debugfs);
949 debugfs_create_ulong("io_tlb_nslabs", 0400, mem->debugfs, &mem->nslabs);
950 debugfs_create_file("io_tlb_used", 0400, mem->debugfs, NULL,
954 static int __init __maybe_unused swiotlb_create_default_debugfs(void)
956 swiotlb_create_debugfs_files(&io_tlb_default_mem, "swiotlb");
960 #ifdef CONFIG_DEBUG_FS
961 late_initcall(swiotlb_create_default_debugfs);
964 #ifdef CONFIG_DMA_RESTRICTED_POOL
966 struct page *swiotlb_alloc(struct device *dev, size_t size)
968 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
969 phys_addr_t tlb_addr;
975 index = swiotlb_find_slots(dev, 0, size, 0);
979 tlb_addr = slot_addr(mem->start, index);
981 return pfn_to_page(PFN_DOWN(tlb_addr));
984 bool swiotlb_free(struct device *dev, struct page *page, size_t size)
986 phys_addr_t tlb_addr = page_to_phys(page);
988 if (!is_swiotlb_buffer(dev, tlb_addr))
991 swiotlb_release_slots(dev, tlb_addr);
996 static int rmem_swiotlb_device_init(struct reserved_mem *rmem,
999 struct io_tlb_mem *mem = rmem->priv;
1000 unsigned long nslabs = rmem->size >> IO_TLB_SHIFT;
1002 /* Set Per-device io tlb area to one */
1003 unsigned int nareas = 1;
1006 * Since multiple devices can share the same pool, the private data,
1007 * io_tlb_mem struct, will be initialized by the first device attached
1011 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
1015 mem->slots = kcalloc(nslabs, sizeof(*mem->slots), GFP_KERNEL);
1021 mem->areas = kcalloc(nareas, sizeof(*mem->areas),
1029 set_memory_decrypted((unsigned long)phys_to_virt(rmem->base),
1030 rmem->size >> PAGE_SHIFT);
1031 swiotlb_init_io_tlb_mem(mem, rmem->base, nslabs, SWIOTLB_FORCE,
1033 mem->for_alloc = true;
1037 swiotlb_create_debugfs_files(mem, rmem->name);
1040 dev->dma_io_tlb_mem = mem;
1045 static void rmem_swiotlb_device_release(struct reserved_mem *rmem,
1048 dev->dma_io_tlb_mem = &io_tlb_default_mem;
1051 static const struct reserved_mem_ops rmem_swiotlb_ops = {
1052 .device_init = rmem_swiotlb_device_init,
1053 .device_release = rmem_swiotlb_device_release,
1056 static int __init rmem_swiotlb_setup(struct reserved_mem *rmem)
1058 unsigned long node = rmem->fdt_node;
1060 if (of_get_flat_dt_prop(node, "reusable", NULL) ||
1061 of_get_flat_dt_prop(node, "linux,cma-default", NULL) ||
1062 of_get_flat_dt_prop(node, "linux,dma-default", NULL) ||
1063 of_get_flat_dt_prop(node, "no-map", NULL))
1066 if (PageHighMem(pfn_to_page(PHYS_PFN(rmem->base)))) {
1067 pr_err("Restricted DMA pool must be accessible within the linear mapping.");
1071 rmem->ops = &rmem_swiotlb_ops;
1072 pr_info("Reserved memory: created restricted DMA pool at %pa, size %ld MiB\n",
1073 &rmem->base, (unsigned long)rmem->size / SZ_1M);
1077 RESERVEDMEM_OF_DECLARE(dma, "restricted-dma-pool", rmem_swiotlb_setup);
1078 #endif /* CONFIG_DMA_RESTRICTED_POOL */