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 OFFSET(val,align) ((unsigned long) \
54 ( (val) & ( (align) - 1)))
56 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
59 * Minimum IO TLB size to bother booting with. Systems with mainly
60 * 64bit capable cards will only lightly use the swiotlb. If we can't
61 * allocate a contiguous 1MB, we're probably in trouble anyway.
63 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
65 enum swiotlb_force swiotlb_force;
68 * Used to do a quick range check in swiotlb_tbl_unmap_single and
69 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
72 phys_addr_t io_tlb_start, io_tlb_end;
75 * The number of IO TLB blocks (in groups of 64) between io_tlb_start and
76 * io_tlb_end. This is command line adjustable via setup_io_tlb_npages.
78 static unsigned long io_tlb_nslabs;
81 * The number of used IO TLB block
83 static unsigned long io_tlb_used;
86 * This is a free list describing the number of free entries available from
89 static unsigned int *io_tlb_list;
90 static unsigned int io_tlb_index;
93 * Max segment that we can provide which (if pages are contingous) will
94 * not be bounced (unless SWIOTLB_FORCE is set).
96 static unsigned int max_segment;
99 * We need to save away the original address corresponding to a mapped entry
100 * for the sync operations.
102 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
103 static phys_addr_t *io_tlb_orig_addr;
106 * Protect the above data structures in the map and unmap calls
108 static DEFINE_SPINLOCK(io_tlb_lock);
110 static int late_alloc;
113 setup_io_tlb_npages(char *str)
116 io_tlb_nslabs = simple_strtoul(str, &str, 0);
117 /* avoid tail segment of size < IO_TLB_SEGSIZE */
118 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
122 if (!strcmp(str, "force")) {
123 swiotlb_force = SWIOTLB_FORCE;
124 } else if (!strcmp(str, "noforce")) {
125 swiotlb_force = SWIOTLB_NO_FORCE;
131 early_param("swiotlb", setup_io_tlb_npages);
133 static bool no_iotlb_memory;
135 unsigned long swiotlb_nr_tbl(void)
137 return unlikely(no_iotlb_memory) ? 0 : io_tlb_nslabs;
139 EXPORT_SYMBOL_GPL(swiotlb_nr_tbl);
141 unsigned int swiotlb_max_segment(void)
143 return unlikely(no_iotlb_memory) ? 0 : max_segment;
145 EXPORT_SYMBOL_GPL(swiotlb_max_segment);
147 void swiotlb_set_max_segment(unsigned int val)
149 if (swiotlb_force == SWIOTLB_FORCE)
152 max_segment = rounddown(val, PAGE_SIZE);
155 /* default to 64MB */
156 #define IO_TLB_DEFAULT_SIZE (64UL<<20)
157 unsigned long swiotlb_size_or_default(void)
161 size = io_tlb_nslabs << IO_TLB_SHIFT;
163 return size ? size : (IO_TLB_DEFAULT_SIZE);
166 void swiotlb_print_info(void)
168 unsigned long bytes = io_tlb_nslabs << IO_TLB_SHIFT;
170 if (no_iotlb_memory) {
171 pr_warn("No low mem\n");
175 pr_info("mapped [mem %pa-%pa] (%luMB)\n", &io_tlb_start, &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);
199 int __init swiotlb_init_with_tbl(char *tlb, unsigned long nslabs, int verbose)
201 unsigned long i, bytes;
204 bytes = nslabs << IO_TLB_SHIFT;
206 io_tlb_nslabs = nslabs;
207 io_tlb_start = __pa(tlb);
208 io_tlb_end = io_tlb_start + bytes;
211 * Allocate and initialize the free list array. This array is used
212 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
213 * between io_tlb_start and io_tlb_end.
215 alloc_size = PAGE_ALIGN(io_tlb_nslabs * sizeof(int));
216 io_tlb_list = memblock_alloc(alloc_size, PAGE_SIZE);
218 panic("%s: Failed to allocate %zu bytes align=0x%lx\n",
219 __func__, alloc_size, PAGE_SIZE);
221 alloc_size = PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t));
222 io_tlb_orig_addr = memblock_alloc(alloc_size, PAGE_SIZE);
223 if (!io_tlb_orig_addr)
224 panic("%s: Failed to allocate %zu bytes align=0x%lx\n",
225 __func__, alloc_size, PAGE_SIZE);
227 for (i = 0; i < io_tlb_nslabs; i++) {
228 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
229 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
232 no_iotlb_memory = false;
235 swiotlb_print_info();
237 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
242 * Statically reserve bounce buffer space and initialize bounce buffer data
243 * structures for the software IO TLB used to implement the DMA API.
246 swiotlb_init(int verbose)
248 size_t default_size = IO_TLB_DEFAULT_SIZE;
249 unsigned char *vstart;
252 if (!io_tlb_nslabs) {
253 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
254 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
257 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
259 /* Get IO TLB memory from the low pages */
260 vstart = memblock_alloc_low(PAGE_ALIGN(bytes), PAGE_SIZE);
261 if (vstart && !swiotlb_init_with_tbl(vstart, io_tlb_nslabs, verbose))
265 memblock_free_early(io_tlb_start,
266 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
269 pr_warn("Cannot allocate buffer");
270 no_iotlb_memory = true;
274 * Systems with larger DMA zones (those that don't support ISA) can
275 * initialize the swiotlb later using the slab allocator if needed.
276 * This should be just like above, but with some error catching.
279 swiotlb_late_init_with_default_size(size_t default_size)
281 unsigned long bytes, req_nslabs = io_tlb_nslabs;
282 unsigned char *vstart = NULL;
286 if (!io_tlb_nslabs) {
287 io_tlb_nslabs = (default_size >> IO_TLB_SHIFT);
288 io_tlb_nslabs = ALIGN(io_tlb_nslabs, IO_TLB_SEGSIZE);
292 * Get IO TLB memory from the low pages
294 order = get_order(io_tlb_nslabs << IO_TLB_SHIFT);
295 io_tlb_nslabs = SLABS_PER_PAGE << order;
296 bytes = io_tlb_nslabs << IO_TLB_SHIFT;
298 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
299 vstart = (void *)__get_free_pages(GFP_DMA | __GFP_NOWARN,
307 io_tlb_nslabs = req_nslabs;
310 if (order != get_order(bytes)) {
311 pr_warn("only able to allocate %ld MB\n",
312 (PAGE_SIZE << order) >> 20);
313 io_tlb_nslabs = SLABS_PER_PAGE << order;
315 rc = swiotlb_late_init_with_tbl(vstart, io_tlb_nslabs);
317 free_pages((unsigned long)vstart, order);
322 static void swiotlb_cleanup(void)
331 swiotlb_late_init_with_tbl(char *tlb, unsigned long nslabs)
333 unsigned long i, bytes;
335 bytes = nslabs << IO_TLB_SHIFT;
337 io_tlb_nslabs = nslabs;
338 io_tlb_start = virt_to_phys(tlb);
339 io_tlb_end = io_tlb_start + bytes;
341 set_memory_decrypted((unsigned long)tlb, bytes >> PAGE_SHIFT);
342 memset(tlb, 0, bytes);
345 * Allocate and initialize the free list array. This array is used
346 * to find contiguous free memory regions of size up to IO_TLB_SEGSIZE
347 * between io_tlb_start and io_tlb_end.
349 io_tlb_list = (unsigned int *)__get_free_pages(GFP_KERNEL,
350 get_order(io_tlb_nslabs * sizeof(int)));
354 io_tlb_orig_addr = (phys_addr_t *)
355 __get_free_pages(GFP_KERNEL,
356 get_order(io_tlb_nslabs *
357 sizeof(phys_addr_t)));
358 if (!io_tlb_orig_addr)
361 for (i = 0; i < io_tlb_nslabs; i++) {
362 io_tlb_list[i] = IO_TLB_SEGSIZE - OFFSET(i, IO_TLB_SEGSIZE);
363 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
366 no_iotlb_memory = false;
368 swiotlb_print_info();
372 swiotlb_set_max_segment(io_tlb_nslabs << IO_TLB_SHIFT);
377 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
385 void __init swiotlb_exit(void)
387 if (!io_tlb_orig_addr)
391 free_pages((unsigned long)io_tlb_orig_addr,
392 get_order(io_tlb_nslabs * sizeof(phys_addr_t)));
393 free_pages((unsigned long)io_tlb_list, get_order(io_tlb_nslabs *
395 free_pages((unsigned long)phys_to_virt(io_tlb_start),
396 get_order(io_tlb_nslabs << IO_TLB_SHIFT));
398 memblock_free_late(__pa(io_tlb_orig_addr),
399 PAGE_ALIGN(io_tlb_nslabs * sizeof(phys_addr_t)));
400 memblock_free_late(__pa(io_tlb_list),
401 PAGE_ALIGN(io_tlb_nslabs * sizeof(int)));
402 memblock_free_late(io_tlb_start,
403 PAGE_ALIGN(io_tlb_nslabs << IO_TLB_SHIFT));
409 * Bounce: copy the swiotlb buffer from or back to the original dma location
411 static void swiotlb_bounce(phys_addr_t orig_addr, phys_addr_t tlb_addr,
412 size_t size, enum dma_data_direction dir)
414 unsigned long pfn = PFN_DOWN(orig_addr);
415 unsigned char *vaddr = phys_to_virt(tlb_addr);
417 if (PageHighMem(pfn_to_page(pfn))) {
418 /* The buffer does not have a mapping. Map it in and copy */
419 unsigned int offset = orig_addr & ~PAGE_MASK;
425 sz = min_t(size_t, PAGE_SIZE - offset, size);
427 local_irq_save(flags);
428 buffer = kmap_atomic(pfn_to_page(pfn));
429 if (dir == DMA_TO_DEVICE)
430 memcpy(vaddr, buffer + offset, sz);
432 memcpy(buffer + offset, vaddr, sz);
433 kunmap_atomic(buffer);
434 local_irq_restore(flags);
441 } else if (dir == DMA_TO_DEVICE) {
442 memcpy(vaddr, phys_to_virt(orig_addr), size);
444 memcpy(phys_to_virt(orig_addr), vaddr, size);
448 phys_addr_t swiotlb_tbl_map_single(struct device *hwdev, phys_addr_t orig_addr,
449 size_t mapping_size, size_t alloc_size,
450 enum dma_data_direction dir, unsigned long attrs)
452 dma_addr_t tbl_dma_addr = phys_to_dma_unencrypted(hwdev, io_tlb_start);
454 phys_addr_t tlb_addr;
455 unsigned int nslots, stride, index, wrap;
458 unsigned long offset_slots;
459 unsigned long max_slots;
460 unsigned long tmp_io_tlb_used;
463 panic("Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
465 if (mem_encrypt_active())
466 pr_warn_once("Memory encryption is active and system is using DMA bounce buffers\n");
468 if (mapping_size > alloc_size) {
469 dev_warn_once(hwdev, "Invalid sizes (mapping: %zd bytes, alloc: %zd bytes)",
470 mapping_size, alloc_size);
471 return (phys_addr_t)DMA_MAPPING_ERROR;
474 mask = dma_get_seg_boundary(hwdev);
476 tbl_dma_addr &= mask;
478 offset_slots = ALIGN(tbl_dma_addr, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
481 * Carefully handle integer overflow which can occur when mask == ~0UL.
484 ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
485 : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
488 * For mappings greater than or equal to a page, we limit the stride
489 * (and hence alignment) to a page size.
491 nslots = ALIGN(alloc_size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
492 if (alloc_size >= PAGE_SIZE)
493 stride = (1 << (PAGE_SHIFT - IO_TLB_SHIFT));
500 * Find suitable number of IO TLB entries size that will fit this
501 * request and allocate a buffer from that IO TLB pool.
503 spin_lock_irqsave(&io_tlb_lock, flags);
505 if (unlikely(nslots > io_tlb_nslabs - io_tlb_used))
508 index = ALIGN(io_tlb_index, stride);
509 if (index >= io_tlb_nslabs)
514 while (iommu_is_span_boundary(index, nslots, offset_slots,
517 if (index >= io_tlb_nslabs)
524 * If we find a slot that indicates we have 'nslots' number of
525 * contiguous buffers, we allocate the buffers from that slot
526 * and mark the entries as '0' indicating unavailable.
528 if (io_tlb_list[index] >= nslots) {
531 for (i = index; i < (int) (index + nslots); i++)
533 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE - 1) && io_tlb_list[i]; i--)
534 io_tlb_list[i] = ++count;
535 tlb_addr = io_tlb_start + (index << IO_TLB_SHIFT);
538 * Update the indices to avoid searching in the next
541 io_tlb_index = ((index + nslots) < io_tlb_nslabs
542 ? (index + nslots) : 0);
547 if (index >= io_tlb_nslabs)
549 } while (index != wrap);
552 tmp_io_tlb_used = io_tlb_used;
554 spin_unlock_irqrestore(&io_tlb_lock, flags);
555 if (!(attrs & DMA_ATTR_NO_WARN) && printk_ratelimit())
556 dev_warn(hwdev, "swiotlb buffer is full (sz: %zd bytes), total %lu (slots), used %lu (slots)\n",
557 alloc_size, io_tlb_nslabs, tmp_io_tlb_used);
558 return (phys_addr_t)DMA_MAPPING_ERROR;
560 io_tlb_used += nslots;
561 spin_unlock_irqrestore(&io_tlb_lock, flags);
564 * Save away the mapping from the original address to the DMA address.
565 * This is needed when we sync the memory. Then we sync the buffer if
568 for (i = 0; i < nslots; i++)
569 io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
570 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
571 (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
572 swiotlb_bounce(orig_addr, tlb_addr, mapping_size, DMA_TO_DEVICE);
578 * tlb_addr is the physical address of the bounce buffer to unmap.
580 void swiotlb_tbl_unmap_single(struct device *hwdev, phys_addr_t tlb_addr,
581 size_t mapping_size, size_t alloc_size,
582 enum dma_data_direction dir, unsigned long attrs)
585 int i, count, nslots = ALIGN(alloc_size, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT;
586 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
587 phys_addr_t orig_addr = io_tlb_orig_addr[index];
590 * First, sync the memory before unmapping the entry
592 if (orig_addr != INVALID_PHYS_ADDR &&
593 !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
594 ((dir == DMA_FROM_DEVICE) || (dir == DMA_BIDIRECTIONAL)))
595 swiotlb_bounce(orig_addr, tlb_addr, mapping_size, DMA_FROM_DEVICE);
598 * Return the buffer to the free list by setting the corresponding
599 * entries to indicate the number of contiguous entries available.
600 * While returning the entries to the free list, we merge the entries
601 * with slots below and above the pool being returned.
603 spin_lock_irqsave(&io_tlb_lock, flags);
605 count = ((index + nslots) < ALIGN(index + 1, IO_TLB_SEGSIZE) ?
606 io_tlb_list[index + nslots] : 0);
608 * Step 1: return the slots to the free list, merging the
609 * slots with superceeding slots
611 for (i = index + nslots - 1; i >= index; i--) {
612 io_tlb_list[i] = ++count;
613 io_tlb_orig_addr[i] = INVALID_PHYS_ADDR;
616 * Step 2: merge the returned slots with the preceding slots,
617 * if available (non zero)
619 for (i = index - 1; (OFFSET(i, IO_TLB_SEGSIZE) != IO_TLB_SEGSIZE -1) && io_tlb_list[i]; i--)
620 io_tlb_list[i] = ++count;
622 io_tlb_used -= nslots;
624 spin_unlock_irqrestore(&io_tlb_lock, flags);
627 void swiotlb_tbl_sync_single(struct device *hwdev, phys_addr_t tlb_addr,
628 size_t size, enum dma_data_direction dir,
629 enum dma_sync_target target)
631 int index = (tlb_addr - io_tlb_start) >> IO_TLB_SHIFT;
632 phys_addr_t orig_addr = io_tlb_orig_addr[index];
634 if (orig_addr == INVALID_PHYS_ADDR)
636 orig_addr += (unsigned long)tlb_addr & ((1 << IO_TLB_SHIFT) - 1);
640 if (likely(dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
641 swiotlb_bounce(orig_addr, tlb_addr,
642 size, DMA_FROM_DEVICE);
644 BUG_ON(dir != DMA_TO_DEVICE);
646 case SYNC_FOR_DEVICE:
647 if (likely(dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
648 swiotlb_bounce(orig_addr, tlb_addr,
649 size, DMA_TO_DEVICE);
651 BUG_ON(dir != DMA_FROM_DEVICE);
659 * Create a swiotlb mapping for the buffer at @paddr, and in case of DMAing
660 * to the device copy the data into it as well.
662 dma_addr_t swiotlb_map(struct device *dev, phys_addr_t paddr, size_t size,
663 enum dma_data_direction dir, unsigned long attrs)
665 phys_addr_t swiotlb_addr;
668 trace_swiotlb_bounced(dev, phys_to_dma(dev, paddr), size,
671 swiotlb_addr = swiotlb_tbl_map_single(dev, paddr, size, size, dir,
673 if (swiotlb_addr == (phys_addr_t)DMA_MAPPING_ERROR)
674 return DMA_MAPPING_ERROR;
676 /* Ensure that the address returned is DMA'ble */
677 dma_addr = phys_to_dma_unencrypted(dev, swiotlb_addr);
678 if (unlikely(!dma_capable(dev, dma_addr, size, true))) {
679 swiotlb_tbl_unmap_single(dev, swiotlb_addr, size, size, dir,
680 attrs | DMA_ATTR_SKIP_CPU_SYNC);
681 dev_WARN_ONCE(dev, 1,
682 "swiotlb addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
683 &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
684 return DMA_MAPPING_ERROR;
687 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
688 arch_sync_dma_for_device(swiotlb_addr, size, dir);
692 size_t swiotlb_max_mapping_size(struct device *dev)
694 return ((size_t)1 << IO_TLB_SHIFT) * IO_TLB_SEGSIZE;
697 bool is_swiotlb_active(void)
700 * When SWIOTLB is initialized, even if io_tlb_start points to physical
701 * address zero, io_tlb_end surely doesn't.
703 return io_tlb_end != 0;
706 #ifdef CONFIG_DEBUG_FS
708 static int __init swiotlb_create_debugfs(void)
712 root = debugfs_create_dir("swiotlb", NULL);
713 debugfs_create_ulong("io_tlb_nslabs", 0400, root, &io_tlb_nslabs);
714 debugfs_create_ulong("io_tlb_used", 0400, root, &io_tlb_used);
718 late_initcall(swiotlb_create_debugfs);