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/dma-direct.h>
25 #include <linux/dma-map-ops.h>
27 #include <linux/export.h>
28 #include <linux/spinlock.h>
29 #include <linux/string.h>
30 #include <linux/swiotlb.h>
31 #include <linux/pfn.h>
32 #include <linux/types.h>
33 #include <linux/ctype.h>
34 #include <linux/highmem.h>
35 #include <linux/gfp.h>
36 #include <linux/scatterlist.h>
37 #include <linux/mem_encrypt.h>
38 #include <linux/set_memory.h>
39 #ifdef CONFIG_DEBUG_FS
40 #include <linux/debugfs.h>
46 #include <linux/init.h>
47 #include <linux/memblock.h>
48 #include <linux/iommu-helper.h>
50 #define CREATE_TRACE_POINTS
51 #include <trace/events/swiotlb.h>
53 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
56 * Minimum IO TLB size to bother booting with. Systems with mainly
57 * 64bit capable cards will only lightly use the swiotlb. If we can't
58 * allocate a contiguous 1MB, we're probably in trouble anyway.
60 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
62 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
64 enum swiotlb_force swiotlb_force;
66 struct io_tlb_mem *io_tlb_default_mem;
69 * Max segment that we can provide which (if pages are contingous) will
70 * not be bounced (unless SWIOTLB_FORCE is set).
72 static unsigned int max_segment;
74 static unsigned long default_nslabs = IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT;
77 setup_io_tlb_npages(char *str)
80 /* avoid tail segment of size < IO_TLB_SEGSIZE */
82 ALIGN(simple_strtoul(str, &str, 0), IO_TLB_SEGSIZE);
86 if (!strcmp(str, "force"))
87 swiotlb_force = SWIOTLB_FORCE;
88 else if (!strcmp(str, "noforce"))
89 swiotlb_force = SWIOTLB_NO_FORCE;
93 early_param("swiotlb", setup_io_tlb_npages);
95 unsigned int swiotlb_max_segment(void)
97 return io_tlb_default_mem ? max_segment : 0;
99 EXPORT_SYMBOL_GPL(swiotlb_max_segment);
101 void swiotlb_set_max_segment(unsigned int val)
103 if (swiotlb_force == SWIOTLB_FORCE)
106 max_segment = rounddown(val, PAGE_SIZE);
109 unsigned long swiotlb_size_or_default(void)
111 return default_nslabs << IO_TLB_SHIFT;
114 void __init swiotlb_adjust_size(unsigned long size)
117 * If swiotlb parameter has not been specified, give a chance to
118 * architectures such as those supporting memory encryption to
119 * adjust/expand SWIOTLB size for their use.
121 if (default_nslabs != IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT)
123 size = ALIGN(size, IO_TLB_SIZE);
124 default_nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE);
125 pr_info("SWIOTLB bounce buffer size adjusted to %luMB", size >> 20);
128 void swiotlb_print_info(void)
130 struct io_tlb_mem *mem = io_tlb_default_mem;
133 pr_warn("No low mem\n");
137 pr_info("mapped [mem %pa-%pa] (%luMB)\n", &mem->start, &mem->end,
138 (mem->nslabs << IO_TLB_SHIFT) >> 20);
141 static inline unsigned long io_tlb_offset(unsigned long val)
143 return val & (IO_TLB_SEGSIZE - 1);
146 static inline unsigned long nr_slots(u64 val)
148 return DIV_ROUND_UP(val, IO_TLB_SIZE);
152 * Early SWIOTLB allocation may be too early to allow an architecture to
153 * perform the desired operations. This function allows the architecture to
154 * call SWIOTLB when the operations are possible. It needs to be called
155 * before the SWIOTLB memory is used.
157 void __init swiotlb_update_mem_attributes(void)
159 struct io_tlb_mem *mem = io_tlb_default_mem;
163 if (!mem || mem->late_alloc)
165 vaddr = phys_to_virt(mem->start);
166 bytes = PAGE_ALIGN(mem->nslabs << IO_TLB_SHIFT);
167 set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT);
168 memset(vaddr, 0, bytes);
171 int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
173 unsigned long bytes = nslabs << IO_TLB_SHIFT, i;
174 struct io_tlb_mem *mem;
177 if (swiotlb_force == SWIOTLB_NO_FORCE)
180 /* protect against double initialization */
181 if (WARN_ON_ONCE(io_tlb_default_mem))
184 alloc_size = PAGE_ALIGN(struct_size(mem, slots, nslabs));
185 mem = memblock_alloc(alloc_size, PAGE_SIZE);
187 panic("%s: Failed to allocate %zu bytes align=0x%lx\n",
188 __func__, alloc_size, PAGE_SIZE);
189 mem->nslabs = nslabs;
190 mem->start = __pa(tlb);
191 mem->end = mem->start + bytes;
193 spin_lock_init(&mem->lock);
194 for (i = 0; i < mem->nslabs; i++) {
195 mem->slots[i].list = IO_TLB_SEGSIZE - io_tlb_offset(i);
196 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
197 mem->slots[i].alloc_size = 0;
200 io_tlb_default_mem = mem;
202 swiotlb_print_info();
203 swiotlb_set_max_segment(mem->nslabs << IO_TLB_SHIFT);
208 * Statically reserve bounce buffer space and initialize bounce buffer data
209 * structures for the software IO TLB used to implement the DMA API.
212 swiotlb_init(int verbose)
214 size_t bytes = PAGE_ALIGN(default_nslabs << IO_TLB_SHIFT);
217 if (swiotlb_force == SWIOTLB_NO_FORCE)
220 /* Get IO TLB memory from the low pages */
221 tlb = memblock_alloc_low(bytes, PAGE_SIZE);
224 if (swiotlb_init_with_tbl(tlb, default_nslabs, verbose))
229 memblock_free_early(__pa(tlb), bytes);
231 pr_warn("Cannot allocate buffer");
235 * Systems with larger DMA zones (those that don't support ISA) can
236 * initialize the swiotlb later using the slab allocator if needed.
237 * This should be just like above, but with some error catching.
240 swiotlb_late_init_with_default_size(size_t default_size)
242 unsigned long nslabs =
243 ALIGN(default_size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE);
245 unsigned char *vstart = NULL;
249 if (swiotlb_force == SWIOTLB_NO_FORCE)
253 * Get IO TLB memory from the low pages
255 order = get_order(nslabs << IO_TLB_SHIFT);
256 nslabs = SLABS_PER_PAGE << order;
257 bytes = nslabs << IO_TLB_SHIFT;
259 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
260 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
270 if (order != get_order(bytes)) {
271 pr_warn("only able to allocate %ld MB\n",
272 (PAGE_SIZE << order) >> 20);
273 nslabs = SLABS_PER_PAGE << order;
275 rc = swiotlb_late_init_with_tbl(vstart, nslabs);
277 free_pages((unsigned long)vstart, order);
283 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
285 unsigned long bytes = nslabs << IO_TLB_SHIFT, i;
286 struct io_tlb_mem *mem;
288 if (swiotlb_force == SWIOTLB_NO_FORCE)
291 /* protect against double initialization */
292 if (WARN_ON_ONCE(io_tlb_default_mem))
295 mem = (void *)__get_free_pages(GFP_KERNEL,
296 get_order(struct_size(mem, slots, nslabs)));
300 mem->nslabs = nslabs;
301 mem->start = virt_to_phys(tlb);
302 mem->end = mem->start + bytes;
305 spin_lock_init(&mem->lock);
306 for (i = 0; i < mem->nslabs; i++) {
307 mem->slots[i].list = IO_TLB_SEGSIZE - io_tlb_offset(i);
308 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
309 mem->slots[i].alloc_size = 0;
312 set_memory_decrypted((unsigned long)tlb, bytes >> PAGE_SHIFT);
313 memset(tlb, 0, bytes);
315 io_tlb_default_mem = mem;
316 swiotlb_print_info();
317 swiotlb_set_max_segment(mem->nslabs << IO_TLB_SHIFT);
321 void __init swiotlb_exit(void)
323 struct io_tlb_mem *mem = io_tlb_default_mem;
329 size = struct_size(mem, slots, mem->nslabs);
331 free_pages((unsigned long)mem, get_order(size));
333 memblock_free_late(__pa(mem), PAGE_ALIGN(size));
334 io_tlb_default_mem = NULL;
338 * Bounce: copy the swiotlb buffer from or back to the original dma location
340 static void swiotlb_bounce(struct device *dev, phys_addr_t tlb_addr, size_t size,
341 enum dma_data_direction dir)
343 struct io_tlb_mem *mem = io_tlb_default_mem;
344 int index = (tlb_addr - mem->start) >> IO_TLB_SHIFT;
345 phys_addr_t orig_addr = mem->slots[index].orig_addr;
346 size_t alloc_size = mem->slots[index].alloc_size;
347 unsigned long pfn = PFN_DOWN(orig_addr);
348 unsigned char *vaddr = phys_to_virt(tlb_addr);
350 if (orig_addr == INVALID_PHYS_ADDR)
353 if (size > alloc_size) {
354 dev_WARN_ONCE(dev, 1,
355 "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu.\n",
360 if (PageHighMem(pfn_to_page(pfn))) {
361 /* The buffer does not have a mapping. Map it in and copy */
362 unsigned int offset = orig_addr & ~PAGE_MASK;
368 sz = min_t(size_t, PAGE_SIZE - offset, size);
370 local_irq_save(flags);
371 buffer = kmap_atomic(pfn_to_page(pfn));
372 if (dir == DMA_TO_DEVICE)
373 memcpy(vaddr, buffer + offset, sz);
375 memcpy(buffer + offset, vaddr, sz);
376 kunmap_atomic(buffer);
377 local_irq_restore(flags);
384 } else if (dir == DMA_TO_DEVICE) {
385 memcpy(vaddr, phys_to_virt(orig_addr), size);
387 memcpy(phys_to_virt(orig_addr), vaddr, size);
391 #define slot_addr(start, idx) ((start) + ((idx) << IO_TLB_SHIFT))
394 * Return the offset into a iotlb slot required to keep the device happy.
396 static unsigned int swiotlb_align_offset(struct device *dev, u64 addr)
398 return addr & dma_get_min_align_mask(dev) & (IO_TLB_SIZE - 1);
402 * Carefully handle integer overflow which can occur when boundary_mask == ~0UL.
404 static inline unsigned long get_max_slots(unsigned long boundary_mask)
406 if (boundary_mask == ~0UL)
407 return 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
408 return nr_slots(boundary_mask + 1);
411 static unsigned int wrap_index(struct io_tlb_mem *mem, unsigned int index)
413 if (index >= mem->nslabs)
419 * Find a suitable number of IO TLB entries size that will fit this request and
420 * allocate a buffer from that IO TLB pool.
422 static int find_slots(struct device *dev, phys_addr_t orig_addr,
425 struct io_tlb_mem *mem = io_tlb_default_mem;
426 unsigned long boundary_mask = dma_get_seg_boundary(dev);
427 dma_addr_t tbl_dma_addr =
428 phys_to_dma_unencrypted(dev, mem->start) & boundary_mask;
429 unsigned long max_slots = get_max_slots(boundary_mask);
430 unsigned int iotlb_align_mask =
431 dma_get_min_align_mask(dev) & ~(IO_TLB_SIZE - 1);
432 unsigned int nslots = nr_slots(alloc_size), stride;
433 unsigned int index, wrap, count = 0, i;
439 * For mappings with an alignment requirement don't bother looping to
440 * unaligned slots once we found an aligned one. For allocations of
441 * PAGE_SIZE or larger only look for page aligned allocations.
443 stride = (iotlb_align_mask >> IO_TLB_SHIFT) + 1;
444 if (alloc_size >= PAGE_SIZE)
445 stride = max(stride, stride << (PAGE_SHIFT - IO_TLB_SHIFT));
447 spin_lock_irqsave(&mem->lock, flags);
448 if (unlikely(nslots > mem->nslabs - mem->used))
451 index = wrap = wrap_index(mem, ALIGN(mem->index, stride));
453 if ((slot_addr(tbl_dma_addr, index) & iotlb_align_mask) !=
454 (orig_addr & iotlb_align_mask)) {
455 index = wrap_index(mem, index + 1);
460 * If we find a slot that indicates we have 'nslots' number of
461 * contiguous buffers, we allocate the buffers from that slot
462 * and mark the entries as '0' indicating unavailable.
464 if (!iommu_is_span_boundary(index, nslots,
465 nr_slots(tbl_dma_addr),
467 if (mem->slots[index].list >= nslots)
470 index = wrap_index(mem, index + stride);
471 } while (index != wrap);
474 spin_unlock_irqrestore(&mem->lock, flags);
478 for (i = index; i < index + nslots; i++)
479 mem->slots[i].list = 0;
481 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 &&
482 mem->slots[i].list; i--)
483 mem->slots[i].list = ++count;
486 * Update the indices to avoid searching in the next round.
488 if (index + nslots < mem->nslabs)
489 mem->index = index + nslots;
494 spin_unlock_irqrestore(&mem->lock, flags);
498 phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
499 size_t mapping_size, size_t alloc_size,
500 enum dma_data_direction dir, unsigned long attrs)
502 struct io_tlb_mem *mem = io_tlb_default_mem;
503 unsigned int offset = swiotlb_align_offset(dev, orig_addr);
506 phys_addr_t tlb_addr;
509 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
511 if (mem_encrypt_active())
512 pr_warn_once("Memory encryption is active and system is using DMA bounce buffers\n");
514 if (mapping_size > alloc_size) {
515 dev_warn_once(dev, "Invalid sizes (mapping: %zd bytes, alloc: %zd bytes)",
516 mapping_size, alloc_size);
517 return (phys_addr_t)DMA_MAPPING_ERROR;
520 index = find_slots(dev, orig_addr, alloc_size + offset);
522 if (!(attrs & DMA_ATTR_NO_WARN))
523 dev_warn_ratelimited(dev,
524 "swiotlb buffer is full (sz: %zd bytes), total %lu (slots), used %lu (slots)\n",
525 alloc_size, mem->nslabs, mem->used);
526 return (phys_addr_t)DMA_MAPPING_ERROR;
530 * Save away the mapping from the original address to the DMA address.
531 * This is needed when we sync the memory. Then we sync the buffer if
534 for (i = 0; i < nr_slots(alloc_size + offset); i++) {
535 mem->slots[index + i].orig_addr = slot_addr(orig_addr, i);
536 mem->slots[index + i].alloc_size =
537 alloc_size - (i << IO_TLB_SHIFT);
539 tlb_addr = slot_addr(mem->start, index) + offset;
540 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
541 (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
542 swiotlb_bounce(dev, tlb_addr, mapping_size, DMA_TO_DEVICE);
547 * tlb_addr is the physical address of the bounce buffer to unmap.
549 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
550 size_t mapping_size, enum dma_data_direction dir,
553 struct io_tlb_mem *mem = io_tlb_default_mem;
555 unsigned int offset = swiotlb_align_offset(hwdev, tlb_addr);
556 int index = (tlb_addr - offset - mem->start) >> IO_TLB_SHIFT;
557 int nslots = nr_slots(mem->slots[index].alloc_size + offset);
561 * First, sync the memory before unmapping the entry
563 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
564 (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
565 swiotlb_bounce(hwdev, tlb_addr, mapping_size, DMA_FROM_DEVICE);
568 * Return the buffer to the free list by setting the corresponding
569 * entries to indicate the number of contiguous entries available.
570 * While returning the entries to the free list, we merge the entries
571 * with slots below and above the pool being returned.
573 spin_lock_irqsave(&mem->lock, flags);
574 if (index + nslots < ALIGN(index + 1, IO_TLB_SEGSIZE))
575 count = mem->slots[index + nslots].list;
580 * Step 1: return the slots to the free list, merging the slots with
583 for (i = index + nslots - 1; i >= index; i--) {
584 mem->slots[i].list = ++count;
585 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
586 mem->slots[i].alloc_size = 0;
590 * Step 2: merge the returned slots with the preceding slots, if
591 * available (non zero)
594 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 && mem->slots[i].list;
596 mem->slots[i].list = ++count;
598 spin_unlock_irqrestore(&mem->lock, flags);
601 void swiotlb_sync_single_for_device(struct device *dev, phys_addr_t tlb_addr,
602 size_t size, enum dma_data_direction dir)
604 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
605 swiotlb_bounce(dev, tlb_addr, size, DMA_TO_DEVICE);
607 BUG_ON(dir != DMA_FROM_DEVICE);
610 void swiotlb_sync_single_for_cpu(struct device *dev, phys_addr_t tlb_addr,
611 size_t size, enum dma_data_direction dir)
613 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
614 swiotlb_bounce(dev, tlb_addr, size, DMA_FROM_DEVICE);
616 BUG_ON(dir != DMA_TO_DEVICE);
620 * Create a swiotlb mapping for the buffer at @paddr, and in case of DMAing
621 * to the device copy the data into it as well.
623 dma_addr_t swiotlb_map(struct device *dev, phys_addr_t paddr, size_t size,
624 enum dma_data_direction dir, unsigned long attrs)
626 phys_addr_t swiotlb_addr;
629 trace_swiotlb_bounced(dev, phys_to_dma(dev, paddr), size,
632 swiotlb_addr = swiotlb_tbl_map_single(dev, paddr, size, size, dir,
634 if (swiotlb_addr == (phys_addr_t)DMA_MAPPING_ERROR)
635 return DMA_MAPPING_ERROR;
637 /* Ensure that the address returned is DMA'ble */
638 dma_addr = phys_to_dma_unencrypted(dev, swiotlb_addr);
639 if (unlikely(!dma_capable(dev, dma_addr, size, true))) {
640 swiotlb_tbl_unmap_single(dev, swiotlb_addr, size, dir,
641 attrs | DMA_ATTR_SKIP_CPU_SYNC);
642 dev_WARN_ONCE(dev, 1,
643 "swiotlb addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
644 &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
645 return DMA_MAPPING_ERROR;
648 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
649 arch_sync_dma_for_device(swiotlb_addr, size, dir);
653 size_t swiotlb_max_mapping_size(struct device *dev)
655 return ((size_t)IO_TLB_SIZE) * IO_TLB_SEGSIZE;
658 bool is_swiotlb_active(void)
660 return io_tlb_default_mem != NULL;
662 EXPORT_SYMBOL_GPL(is_swiotlb_active);
664 #ifdef CONFIG_DEBUG_FS
666 static int __init swiotlb_create_debugfs(void)
668 struct io_tlb_mem *mem = io_tlb_default_mem;
672 mem->debugfs = debugfs_create_dir("swiotlb", NULL);
673 debugfs_create_ulong("io_tlb_nslabs", 0400, mem->debugfs, &mem->nslabs);
674 debugfs_create_ulong("io_tlb_used", 0400, mem->debugfs, &mem->used);
678 late_initcall(swiotlb_create_debugfs);