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>
26 #include <linux/export.h>
27 #include <linux/spinlock.h>
28 #include <linux/string.h>
29 #include <linux/swiotlb.h>
30 #include <linux/pfn.h>
31 #include <linux/types.h>
32 #include <linux/ctype.h>
33 #include <linux/highmem.h>
34 #include <linux/gfp.h>
35 #include <linux/scatterlist.h>
36 #include <linux/mem_encrypt.h>
37 #include <linux/set_memory.h>
38 #ifdef CONFIG_DEBUG_FS
39 #include <linux/debugfs.h>
45 #include <linux/init.h>
46 #include <linux/memblock.h>
47 #include <linux/iommu-helper.h>
49 #define CREATE_TRACE_POINTS
50 #include <trace/events/swiotlb.h>
52 #define OFFSET(val,align) ((unsigned long) \
53 ( (val) & ( (align) - 1)))
55 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
58 * Minimum IO TLB size to bother booting with. Systems with mainly
59 * 64bit capable cards will only lightly use the swiotlb. If we can't
60 * allocate a contiguous 1MB, we're probably in trouble anyway.
62 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
64 enum swiotlb_force swiotlb_force;
67 * Used to do a quick range check in swiotlb_tbl_unmap_single and
68 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
71 phys_addr_t io_tlb_start, io_tlb_end;
74 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
75 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
77 static unsigned long io_tlb_nslabs;
80 * The number of used IO TLB block
82 static unsigned long io_tlb_used;
85 * This is a free list describing the number of free entries available from
88 static unsigned int *io_tlb_list;
89 static unsigned int io_tlb_index;
92 * Max segment that we can provide which (if pages are contingous) will
93 * not be bounced (unless SWIOTLB_FORCE is set).
95 unsigned int max_segment;
98 * We need to save away the original address corresponding to a mapped entry
99 * for the sync operations.
101 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
102 static phys_addr_t *io_tlb_orig_addr;
105 * Protect the above data structures in the map and unmap calls
107 static DEFINE_SPINLOCK(io_tlb_lock);
109 static int late_alloc;
112 setup_io_tlb_npages(char *str)
115 io_tlb_nslabs = simple_strtoul(str, &str, 0);
116 /* avoid tail segment of size < IO_TLB_SEGSIZE */
117 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
121 if (!strcmp(str, "force")) {
122 swiotlb_force = SWIOTLB_FORCE;
123 } else if (!strcmp(str, "noforce")) {
124 swiotlb_force = SWIOTLB_NO_FORCE;
130 early_param("swiotlb", setup_io_tlb_npages);
132 static bool no_iotlb_memory;
134 unsigned long swiotlb_nr_tbl(void)
136 return unlikely(no_iotlb_memory) ? 0 : io_tlb_nslabs;
138 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
140 unsigned int swiotlb_max_segment(void)
142 return unlikely(no_iotlb_memory) ? 0 : max_segment;
144 EXPORT_SYMBOL_GPL(swiotlb_max_segment);
146 void swiotlb_set_max_segment(unsigned int val)
148 if (swiotlb_force == SWIOTLB_FORCE)
151 max_segment = rounddown(val, PAGE_SIZE);
154 /* default to 64MB */
155 #define IO_TLB_DEFAULT_SIZE (64UL<<20)
156 unsigned long swiotlb_size_or_default(void)
160 size = io_tlb_nslabs << IO_TLB_SHIFT;
162 return size ? size : (IO_TLB_DEFAULT_SIZE);
165 void swiotlb_print_info(void)
167 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
169 if (no_iotlb_memory) {
170 pr_warn("No low mem\n");
174 pr_info("mapped [mem %#010llx-%#010llx] (%luMB)\n",
175 (unsigned long long)io_tlb_start,
176 (unsigned long long)io_tlb_end,
181 * Early SWIOTLB allocation may be too early to allow an architecture to
182 * perform the desired operations. This function allows the architecture to
183 * call SWIOTLB when the operations are possible. It needs to be called
184 * before the SWIOTLB memory is used.
186 void __init swiotlb_update_mem_attributes(void)
191 if (no_iotlb_memory || late_alloc)
194 vaddr = phys_to_virt(io_tlb_start);
195 bytes = PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT);
196 set_memory_decrypted((unsigned long)vaddr, bytes >> PAGE_SHIFT);
197 memset(vaddr, 0, bytes);
200 int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
202 unsigned long i, bytes;
205 bytes = nslabs << IO_TLB_SHIFT;
207 io_tlb_nslabs = nslabs;
208 io_tlb_start = __pa(tlb);
209 io_tlb_end = io_tlb_start + bytes;
212 * Allocate and initialize the free list array. This array is used
213 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
214 * between io_tlb_start and io_tlb_end.
216 alloc_size = PAGE_ALIGN(io_tlb_nslabs * sizeof(int));
217 io_tlb_list = memblock_alloc(alloc_size, PAGE_SIZE);
219 panic("%s: Failed to allocate %zu bytes align=0x%lx\n",
220 __func__, alloc_size, PAGE_SIZE);
222 alloc_size = PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t));
223 io_tlb_orig_addr = memblock_alloc(alloc_size, PAGE_SIZE);
224 if (!io_tlb_orig_addr)
225 panic("%s: Failed to allocate %zu bytes align=0x%lx\n",
226 __func__, alloc_size, PAGE_SIZE);
228 for (i = 0; i < io_tlb_nslabs; i++) {
229 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
230 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
233 no_iotlb_memory = false;
236 swiotlb_print_info();
238 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
243 * Statically reserve bounce buffer space and initialize bounce buffer data
244 * structures for the software IO TLB used to implement the DMA API.
247 swiotlb_init(int verbose)
249 size_t default_size = IO_TLB_DEFAULT_SIZE;
250 unsigned char *vstart;
253 if (!io_tlb_nslabs) {
254 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
255 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
258 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
260 /* Get IO TLB memory from the low pages */
261 vstart = memblock_alloc_low(PAGE_ALIGN(bytes), PAGE_SIZE);
262 if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
266 memblock_free_early(io_tlb_start,
267 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
270 pr_warn("Cannot allocate buffer");
271 no_iotlb_memory = true;
275 * Systems with larger DMA zones (those that don't support ISA) can
276 * initialize the swiotlb later using the slab allocator if needed.
277 * This should be just like above, but with some error catching.
280 swiotlb_late_init_with_default_size(size_t default_size)
282 unsigned long bytes, req_nslabs = io_tlb_nslabs;
283 unsigned char *vstart = NULL;
287 if (!io_tlb_nslabs) {
288 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
289 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
293 * Get IO TLB memory from the low pages
295 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
296 io_tlb_nslabs = SLABS_PER_PAGE << order;
297 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
299 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
300 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
308 io_tlb_nslabs = req_nslabs;
311 if (order != get_order(bytes)) {
312 pr_warn("only able to allocate %ld MB\n",
313 (PAGE_SIZE << order) >> 20);
314 io_tlb_nslabs = SLABS_PER_PAGE << order;
316 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
318 free_pages((unsigned long)vstart, order);
323 static void swiotlb_cleanup(void)
332 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
334 unsigned long i, bytes;
336 bytes = nslabs << IO_TLB_SHIFT;
338 io_tlb_nslabs = nslabs;
339 io_tlb_start = virt_to_phys(tlb);
340 io_tlb_end = io_tlb_start + bytes;
342 set_memory_decrypted((unsigned long)tlb, bytes >> PAGE_SHIFT);
343 memset(tlb, 0, bytes);
346 * Allocate and initialize the free list array. This array is used
347 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
348 * between io_tlb_start and io_tlb_end.
350 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
351 get_order(io_tlb_nslabs * sizeof(int)));
355 io_tlb_orig_addr = (phys_addr_t *)
356 __get_free_pages(GFP_KERNEL,
357 get_order(io_tlb_nslabs *
358 sizeof(phys_addr_t)));
359 if (!io_tlb_orig_addr)
362 for (i = 0; i < io_tlb_nslabs; i++) {
363 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
364 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
367 no_iotlb_memory = false;
369 swiotlb_print_info();
373 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
378 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
386 void __init swiotlb_exit(void)
388 if (!io_tlb_orig_addr)
392 free_pages((unsigned long)io_tlb_orig_addr,
393 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
394 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
396 free_pages((unsigned long)phys_to_virt(io_tlb_start),
397 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
399 memblock_free_late(__pa(io_tlb_orig_addr),
400 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
401 memblock_free_late(__pa(io_tlb_list),
402 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
403 memblock_free_late(io_tlb_start,
404 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
410 * Bounce: copy the swiotlb buffer from or back to the original dma location
412 static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
413 size_t size, enum dma_data_direction dir)
415 unsigned long pfn = PFN_DOWN(orig_addr);
416 unsigned char *vaddr = phys_to_virt(tlb_addr);
418 if (PageHighMem(pfn_to_page(pfn))) {
419 /* The buffer does not have a mapping. Map it in and copy */
420 unsigned int offset = orig_addr & ~PAGE_MASK;
426 sz = min_t(size_t, PAGE_SIZE - offset, size);
428 local_irq_save(flags);
429 buffer = kmap_atomic(pfn_to_page(pfn));
430 if (dir == DMA_TO_DEVICE)
431 memcpy(vaddr, buffer + offset, sz);
433 memcpy(buffer + offset, vaddr, sz);
434 kunmap_atomic(buffer);
435 local_irq_restore(flags);
442 } else if (dir == DMA_TO_DEVICE) {
443 memcpy(vaddr, phys_to_virt(orig_addr), size);
445 memcpy(phys_to_virt(orig_addr), vaddr, size);
449 phys_addr_t swiotlb_tbl_map_single(struct device *hwdev,
450 dma_addr_t tbl_dma_addr,
451 phys_addr_t orig_addr,
454 enum dma_data_direction dir,
458 phys_addr_t tlb_addr;
459 unsigned int nslots, stride, index, wrap;
462 unsigned long offset_slots;
463 unsigned long max_slots;
464 unsigned long tmp_io_tlb_used;
467 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
469 if (mem_encrypt_active())
470 pr_warn_once("Memory encryption is active and system is using DMA bounce buffers\n");
472 if (mapping_size > alloc_size) {
473 dev_warn_once(hwdev, "Invalid sizes (mapping: %zd bytes, alloc: %zd bytes)",
474 mapping_size, alloc_size);
475 return (phys_addr_t)DMA_MAPPING_ERROR;
478 mask = dma_get_seg_boundary(hwdev);
480 tbl_dma_addr &= mask;
482 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
485 * Carefully handle integer overflow which can occur when mask == ~0UL.
488 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
489 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
492 * For mappings greater than or equal to a page, we limit the stride
493 * (and hence alignment) to a page size.
495 nslots = ALIGN(alloc_size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
496 if (alloc_size >= PAGE_SIZE)
497 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
504 * Find suitable number of IO TLB entries size that will fit this
505 * request and allocate a buffer from that IO TLB pool.
507 spin_lock_irqsave(&io_tlb_lock, flags);
509 if (unlikely(nslots > io_tlb_nslabs - io_tlb_used))
512 index = ALIGN(io_tlb_index, stride);
513 if (index >= io_tlb_nslabs)
518 while (iommu_is_span_boundary(index, nslots, offset_slots,
521 if (index >= io_tlb_nslabs)
528 * If we find a slot that indicates we have 'nslots' number of
529 * contiguous buffers, we allocate the buffers from that slot
530 * and mark the entries as '0' indicating unavailable.
532 if (io_tlb_list[index] >= nslots) {
535 for (i = index; i < (int) (index + nslots); i++)
537 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
538 io_tlb_list[i] = ++count;
539 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
542 * Update the indices to avoid searching in the next
545 io_tlb_index = ((index + nslots) < io_tlb_nslabs
546 ? (index + nslots) : 0);
551 if (index >= io_tlb_nslabs)
553 } while (index != wrap);
556 tmp_io_tlb_used = io_tlb_used;
558 spin_unlock_irqrestore(&io_tlb_lock, flags);
559 if (!(attrs & DMA_ATTR_NO_WARN) && printk_ratelimit())
560 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes), total %lu (slots), used %lu (slots)\n",
561 alloc_size, io_tlb_nslabs, tmp_io_tlb_used);
562 return (phys_addr_t)DMA_MAPPING_ERROR;
564 io_tlb_used += nslots;
565 spin_unlock_irqrestore(&io_tlb_lock, flags);
568 * Save away the mapping from the original address to the DMA address.
569 * This is needed when we sync the memory. Then we sync the buffer if
572 for (i = 0; i < nslots; i++)
573 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
575 * When dir == DMA_FROM_DEVICE we could omit the copy from the orig
576 * to the tlb buffer, if we knew for sure the device will
577 * overwirte the entire current content. But we don't. Thus
578 * unconditional bounce may prevent leaking swiotlb content (i.e.
579 * kernel memory) to user-space.
581 swiotlb_bounce(orig_addr, tlb_addr, mapping_size, DMA_TO_DEVICE);
586 * tlb_addr is the physical address of the bounce buffer to unmap.
588 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
589 size_t mapping_size, size_t alloc_size,
590 enum dma_data_direction dir, unsigned long attrs)
593 int i, count, nslots = ALIGN(alloc_size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
594 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
595 phys_addr_t orig_addr = io_tlb_orig_addr[index];
598 * First, sync the memory before unmapping the entry
600 if (orig_addr != INVALID_PHYS_ADDR &&
601 !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
602 ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
603 swiotlb_bounce(orig_addr, tlb_addr, mapping_size, DMA_FROM_DEVICE);
606 * Return the buffer to the free list by setting the corresponding
607 * entries to indicate the number of contiguous entries available.
608 * While returning the entries to the free list, we merge the entries
609 * with slots below and above the pool being returned.
611 spin_lock_irqsave(&io_tlb_lock, flags);
613 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
614 io_tlb_list[index + nslots] : 0);
616 * Step 1: return the slots to the free list, merging the
617 * slots with superceeding slots
619 for (i = index + nslots - 1; i >= index; i--) {
620 io_tlb_list[i] = ++count;
621 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
624 * Step 2: merge the returned slots with the preceding slots,
625 * if available (non zero)
627 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
628 io_tlb_list[i] = ++count;
630 io_tlb_used -= nslots;
632 spin_unlock_irqrestore(&io_tlb_lock, flags);
635 void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
636 size_t size, enum dma_data_direction dir,
637 enum dma_sync_target target)
639 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
640 phys_addr_t orig_addr = io_tlb_orig_addr[index];
642 if (orig_addr == INVALID_PHYS_ADDR)
644 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
648 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
649 swiotlb_bounce(orig_addr, tlb_addr,
650 size, DMA_FROM_DEVICE);
652 BUG_ON(dir != DMA_TO_DEVICE);
654 case SYNC_FOR_DEVICE:
655 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
656 swiotlb_bounce(orig_addr, tlb_addr,
657 size, DMA_TO_DEVICE);
659 BUG_ON(dir != DMA_FROM_DEVICE);
667 * Create a swiotlb mapping for the buffer at @phys, and in case of DMAing
668 * to the device copy the data into it as well.
670 bool swiotlb_map(struct device *dev, phys_addr_t *phys, dma_addr_t *dma_addr,
671 size_t size, enum dma_data_direction dir, unsigned long attrs)
673 trace_swiotlb_bounced(dev, *dma_addr, size, swiotlb_force);
675 if (unlikely(swiotlb_force == SWIOTLB_NO_FORCE)) {
676 dev_warn_ratelimited(dev,
677 "Cannot do DMA to address %pa\n", phys);
681 /* Oh well, have to allocate and map a bounce buffer. */
682 *phys = swiotlb_tbl_map_single(dev, __phys_to_dma(dev, io_tlb_start),
683 *phys, size, size, dir, attrs);
684 if (*phys == (phys_addr_t)DMA_MAPPING_ERROR)
687 /* Ensure that the address returned is DMA'ble */
688 *dma_addr = __phys_to_dma(dev, *phys);
689 if (unlikely(!dma_capable(dev, *dma_addr, size))) {
690 swiotlb_tbl_unmap_single(dev, *phys, size, size, dir,
691 attrs | DMA_ATTR_SKIP_CPU_SYNC);
698 size_t swiotlb_max_mapping_size(struct device *dev)
700 return ((size_t)1 << IO_TLB_SHIFT) * IO_TLB_SEGSIZE;
703 bool is_swiotlb_active(void)
706 * When SWIOTLB is initialized, even if io_tlb_start points to physical
707 * address zero, io_tlb_end surely doesn't.
709 return io_tlb_end != 0;
712 #ifdef CONFIG_DEBUG_FS
714 static int __init swiotlb_create_debugfs(void)
718 root = debugfs_create_dir("swiotlb", NULL);
719 debugfs_create_ulong("io_tlb_nslabs", 0400, root, &io_tlb_nslabs);
720 debugfs_create_ulong("io_tlb_used", 0400, root, &io_tlb_used);
724 late_initcall(swiotlb_create_debugfs);