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/rculist.h>
39 #include <linux/scatterlist.h>
40 #include <linux/set_memory.h>
41 #include <linux/spinlock.h>
42 #include <linux/string.h>
43 #include <linux/swiotlb.h>
44 #include <linux/types.h>
45 #ifdef CONFIG_DMA_RESTRICTED_POOL
47 #include <linux/of_fdt.h>
48 #include <linux/of_reserved_mem.h>
49 #include <linux/slab.h>
52 #define CREATE_TRACE_POINTS
53 #include <trace/events/swiotlb.h>
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 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
67 * struct io_tlb_slot - IO TLB slot descriptor
68 * @orig_addr: The original address corresponding to a mapped entry.
69 * @alloc_size: Size of the allocated buffer.
70 * @list: The free list describing the number of free entries available
74 phys_addr_t orig_addr;
79 static bool swiotlb_force_bounce;
80 static bool swiotlb_force_disable;
82 #ifdef CONFIG_SWIOTLB_DYNAMIC
84 static void swiotlb_dyn_alloc(struct work_struct *work);
86 static struct io_tlb_mem io_tlb_default_mem = {
87 .lock = __SPIN_LOCK_UNLOCKED(io_tlb_default_mem.lock),
88 .pools = LIST_HEAD_INIT(io_tlb_default_mem.pools),
89 .dyn_alloc = __WORK_INITIALIZER(io_tlb_default_mem.dyn_alloc,
93 #else /* !CONFIG_SWIOTLB_DYNAMIC */
95 static struct io_tlb_mem io_tlb_default_mem;
97 #endif /* CONFIG_SWIOTLB_DYNAMIC */
99 static unsigned long default_nslabs = IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT;
100 static unsigned long default_nareas;
103 * struct io_tlb_area - IO TLB memory area descriptor
105 * This is a single area with a single lock.
107 * @used: The number of used IO TLB block.
108 * @index: The slot index to start searching in this area for next round.
109 * @lock: The lock to protect the above data structures in the map and
119 * Round up number of slabs to the next power of 2. The last area is going
120 * be smaller than the rest if default_nslabs is not power of two.
121 * The number of slot in an area should be a multiple of IO_TLB_SEGSIZE,
122 * otherwise a segment may span two or more areas. It conflicts with free
123 * contiguous slots tracking: free slots are treated contiguous no matter
124 * whether they cross an area boundary.
126 * Return true if default_nslabs is rounded up.
128 static bool round_up_default_nslabs(void)
133 if (default_nslabs < IO_TLB_SEGSIZE * default_nareas)
134 default_nslabs = IO_TLB_SEGSIZE * default_nareas;
135 else if (is_power_of_2(default_nslabs))
137 default_nslabs = roundup_pow_of_two(default_nslabs);
142 * swiotlb_adjust_nareas() - adjust the number of areas and slots
143 * @nareas: Desired number of areas. Zero is treated as 1.
145 * Adjust the default number of areas in a memory pool.
146 * The default size of the memory pool may also change to meet minimum area
149 static void swiotlb_adjust_nareas(unsigned int nareas)
153 else if (!is_power_of_2(nareas))
154 nareas = roundup_pow_of_two(nareas);
156 default_nareas = nareas;
158 pr_info("area num %d.\n", nareas);
159 if (round_up_default_nslabs())
160 pr_info("SWIOTLB bounce buffer size roundup to %luMB",
161 (default_nslabs << IO_TLB_SHIFT) >> 20);
165 * limit_nareas() - get the maximum number of areas for a given memory pool size
166 * @nareas: Desired number of areas.
167 * @nslots: Total number of slots in the memory pool.
169 * Limit the number of areas to the maximum possible number of areas in
170 * a memory pool of the given size.
172 * Return: Maximum possible number of areas.
174 static unsigned int limit_nareas(unsigned int nareas, unsigned long nslots)
176 if (nslots < nareas * IO_TLB_SEGSIZE)
177 return nslots / IO_TLB_SEGSIZE;
182 setup_io_tlb_npages(char *str)
185 /* avoid tail segment of size < IO_TLB_SEGSIZE */
187 ALIGN(simple_strtoul(str, &str, 0), IO_TLB_SEGSIZE);
192 swiotlb_adjust_nareas(simple_strtoul(str, &str, 0));
195 if (!strcmp(str, "force"))
196 swiotlb_force_bounce = true;
197 else if (!strcmp(str, "noforce"))
198 swiotlb_force_disable = true;
202 early_param("swiotlb", setup_io_tlb_npages);
204 unsigned long swiotlb_size_or_default(void)
206 return default_nslabs << IO_TLB_SHIFT;
209 void __init swiotlb_adjust_size(unsigned long size)
212 * If swiotlb parameter has not been specified, give a chance to
213 * architectures such as those supporting memory encryption to
214 * adjust/expand SWIOTLB size for their use.
216 if (default_nslabs != IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT)
219 size = ALIGN(size, IO_TLB_SIZE);
220 default_nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE);
221 if (round_up_default_nslabs())
222 size = default_nslabs << IO_TLB_SHIFT;
223 pr_info("SWIOTLB bounce buffer size adjusted to %luMB", size >> 20);
226 void swiotlb_print_info(void)
228 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
231 pr_warn("No low mem\n");
235 pr_info("mapped [mem %pa-%pa] (%luMB)\n", &mem->start, &mem->end,
236 (mem->nslabs << IO_TLB_SHIFT) >> 20);
239 static inline unsigned long io_tlb_offset(unsigned long val)
241 return val & (IO_TLB_SEGSIZE - 1);
244 static inline unsigned long nr_slots(u64 val)
246 return DIV_ROUND_UP(val, IO_TLB_SIZE);
250 * Early SWIOTLB allocation may be too early to allow an architecture to
251 * perform the desired operations. This function allows the architecture to
252 * call SWIOTLB when the operations are possible. It needs to be called
253 * before the SWIOTLB memory is used.
255 void __init swiotlb_update_mem_attributes(void)
257 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
260 if (!mem->nslabs || mem->late_alloc)
262 bytes = PAGE_ALIGN(mem->nslabs << IO_TLB_SHIFT);
263 set_memory_decrypted((unsigned long)mem->vaddr, bytes >> PAGE_SHIFT);
266 static void swiotlb_init_io_tlb_pool(struct io_tlb_pool *mem, phys_addr_t start,
267 unsigned long nslabs, bool late_alloc, unsigned int nareas)
269 void *vaddr = phys_to_virt(start);
270 unsigned long bytes = nslabs << IO_TLB_SHIFT, i;
272 mem->nslabs = nslabs;
274 mem->end = mem->start + bytes;
275 mem->late_alloc = late_alloc;
276 mem->nareas = nareas;
277 mem->area_nslabs = nslabs / mem->nareas;
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;
291 memset(vaddr, 0, bytes);
297 * add_mem_pool() - add a memory pool to the allocator
298 * @mem: Software IO TLB allocator.
299 * @pool: Memory pool to be added.
301 static void add_mem_pool(struct io_tlb_mem *mem, struct io_tlb_pool *pool)
303 #ifdef CONFIG_SWIOTLB_DYNAMIC
304 spin_lock(&mem->lock);
305 list_add_rcu(&pool->node, &mem->pools);
306 mem->nslabs += pool->nslabs;
307 spin_unlock(&mem->lock);
309 mem->nslabs = pool->nslabs;
313 static void __init *swiotlb_memblock_alloc(unsigned long nslabs,
315 int (*remap)(void *tlb, unsigned long nslabs))
317 size_t bytes = PAGE_ALIGN(nslabs << IO_TLB_SHIFT);
321 * By default allocate the bounce buffer memory from low memory, but
322 * allow to pick a location everywhere for hypervisors with guest
325 if (flags & SWIOTLB_ANY)
326 tlb = memblock_alloc(bytes, PAGE_SIZE);
328 tlb = memblock_alloc_low(bytes, PAGE_SIZE);
331 pr_warn("%s: Failed to allocate %zu bytes tlb structure\n",
336 if (remap && remap(tlb, nslabs) < 0) {
337 memblock_free(tlb, PAGE_ALIGN(bytes));
338 pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
346 * Statically reserve bounce buffer space and initialize bounce buffer data
347 * structures for the software IO TLB used to implement the DMA API.
349 void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags,
350 int (*remap)(void *tlb, unsigned long nslabs))
352 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
353 unsigned long nslabs;
358 if (!addressing_limit && !swiotlb_force_bounce)
360 if (swiotlb_force_disable)
363 io_tlb_default_mem.force_bounce =
364 swiotlb_force_bounce || (flags & SWIOTLB_FORCE);
366 #ifdef CONFIG_SWIOTLB_DYNAMIC
368 io_tlb_default_mem.can_grow = true;
369 if (flags & SWIOTLB_ANY)
370 io_tlb_default_mem.phys_limit = virt_to_phys(high_memory - 1);
372 io_tlb_default_mem.phys_limit = ARCH_LOW_ADDRESS_LIMIT;
376 swiotlb_adjust_nareas(num_possible_cpus());
378 nslabs = default_nslabs;
379 nareas = limit_nareas(default_nareas, nslabs);
380 while ((tlb = swiotlb_memblock_alloc(nslabs, flags, remap)) == NULL) {
381 if (nslabs <= IO_TLB_MIN_SLABS)
383 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
384 nareas = limit_nareas(nareas, nslabs);
387 if (default_nslabs != nslabs) {
388 pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs",
389 default_nslabs, nslabs);
390 default_nslabs = nslabs;
393 alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs));
394 mem->slots = memblock_alloc(alloc_size, PAGE_SIZE);
396 pr_warn("%s: Failed to allocate %zu bytes align=0x%lx\n",
397 __func__, alloc_size, PAGE_SIZE);
401 mem->areas = memblock_alloc(array_size(sizeof(struct io_tlb_area),
402 nareas), SMP_CACHE_BYTES);
404 pr_warn("%s: Failed to allocate mem->areas.\n", __func__);
408 swiotlb_init_io_tlb_pool(mem, __pa(tlb), nslabs, false, nareas);
409 add_mem_pool(&io_tlb_default_mem, mem);
411 if (flags & SWIOTLB_VERBOSE)
412 swiotlb_print_info();
415 void __init swiotlb_init(bool addressing_limit, unsigned int flags)
417 swiotlb_init_remap(addressing_limit, flags, NULL);
421 * Systems with larger DMA zones (those that don't support ISA) can
422 * initialize the swiotlb later using the slab allocator if needed.
423 * This should be just like above, but with some error catching.
425 int swiotlb_init_late(size_t size, gfp_t gfp_mask,
426 int (*remap)(void *tlb, unsigned long nslabs))
428 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
429 unsigned long nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE);
431 unsigned char *vstart = NULL;
432 unsigned int order, area_order;
433 bool retried = false;
436 if (io_tlb_default_mem.nslabs)
439 if (swiotlb_force_disable)
442 io_tlb_default_mem.force_bounce = swiotlb_force_bounce;
444 #ifdef CONFIG_SWIOTLB_DYNAMIC
446 io_tlb_default_mem.can_grow = true;
447 if (IS_ENABLED(CONFIG_ZONE_DMA) && (gfp_mask & __GFP_DMA))
448 io_tlb_default_mem.phys_limit = DMA_BIT_MASK(zone_dma_bits);
449 else if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp_mask & __GFP_DMA32))
450 io_tlb_default_mem.phys_limit = DMA_BIT_MASK(32);
452 io_tlb_default_mem.phys_limit = virt_to_phys(high_memory - 1);
456 swiotlb_adjust_nareas(num_possible_cpus());
459 order = get_order(nslabs << IO_TLB_SHIFT);
460 nslabs = SLABS_PER_PAGE << order;
462 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
463 vstart = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
468 nslabs = SLABS_PER_PAGE << order;
476 rc = remap(vstart, nslabs);
478 free_pages((unsigned long)vstart, order);
480 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
481 if (nslabs < IO_TLB_MIN_SLABS)
488 pr_warn("only able to allocate %ld MB\n",
489 (PAGE_SIZE << order) >> 20);
492 nareas = limit_nareas(default_nareas, nslabs);
493 area_order = get_order(array_size(sizeof(*mem->areas), nareas));
494 mem->areas = (struct io_tlb_area *)
495 __get_free_pages(GFP_KERNEL | __GFP_ZERO, area_order);
499 mem->slots = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
500 get_order(array_size(sizeof(*mem->slots), nslabs)));
504 set_memory_decrypted((unsigned long)vstart,
505 (nslabs << IO_TLB_SHIFT) >> PAGE_SHIFT);
506 swiotlb_init_io_tlb_pool(mem, virt_to_phys(vstart), nslabs, true,
508 add_mem_pool(&io_tlb_default_mem, mem);
510 swiotlb_print_info();
514 free_pages((unsigned long)mem->areas, area_order);
516 free_pages((unsigned long)vstart, order);
520 void __init swiotlb_exit(void)
522 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
523 unsigned long tbl_vaddr;
524 size_t tbl_size, slots_size;
525 unsigned int area_order;
527 if (swiotlb_force_bounce)
533 pr_info("tearing down default memory pool\n");
534 tbl_vaddr = (unsigned long)phys_to_virt(mem->start);
535 tbl_size = PAGE_ALIGN(mem->end - mem->start);
536 slots_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), mem->nslabs));
538 set_memory_encrypted(tbl_vaddr, tbl_size >> PAGE_SHIFT);
539 if (mem->late_alloc) {
540 area_order = get_order(array_size(sizeof(*mem->areas),
542 free_pages((unsigned long)mem->areas, area_order);
543 free_pages(tbl_vaddr, get_order(tbl_size));
544 free_pages((unsigned long)mem->slots, get_order(slots_size));
546 memblock_free_late(__pa(mem->areas),
547 array_size(sizeof(*mem->areas), mem->nareas));
548 memblock_free_late(mem->start, tbl_size);
549 memblock_free_late(__pa(mem->slots), slots_size);
552 memset(mem, 0, sizeof(*mem));
555 #ifdef CONFIG_SWIOTLB_DYNAMIC
558 * alloc_dma_pages() - allocate pages to be used for DMA
559 * @gfp: GFP flags for the allocation.
560 * @bytes: Size of the buffer.
562 * Allocate pages from the buddy allocator. If successful, make the allocated
563 * pages decrypted that they can be used for DMA.
565 * Return: Decrypted pages, or %NULL on failure.
567 static struct page *alloc_dma_pages(gfp_t gfp, size_t bytes)
569 unsigned int order = get_order(bytes);
573 page = alloc_pages(gfp, order);
577 vaddr = page_address(page);
578 if (set_memory_decrypted((unsigned long)vaddr, PFN_UP(bytes)))
583 __free_pages(page, order);
588 * swiotlb_alloc_tlb() - allocate a dynamic IO TLB buffer
589 * @dev: Device for which a memory pool is allocated.
590 * @bytes: Size of the buffer.
591 * @phys_limit: Maximum allowed physical address of the buffer.
592 * @gfp: GFP flags for the allocation.
594 * Return: Allocated pages, or %NULL on allocation failure.
596 static struct page *swiotlb_alloc_tlb(struct device *dev, size_t bytes,
597 u64 phys_limit, gfp_t gfp)
602 * Allocate from the atomic pools if memory is encrypted and
603 * the allocation is atomic, because decrypting may block.
605 if (!gfpflags_allow_blocking(gfp) && dev && force_dma_unencrypted(dev)) {
608 if (!IS_ENABLED(CONFIG_DMA_COHERENT_POOL))
611 return dma_alloc_from_pool(dev, bytes, &vaddr, gfp,
615 gfp &= ~GFP_ZONEMASK;
616 if (phys_limit <= DMA_BIT_MASK(zone_dma_bits))
618 else if (phys_limit <= DMA_BIT_MASK(32))
621 while ((page = alloc_dma_pages(gfp, bytes)) &&
622 page_to_phys(page) + bytes - 1 > phys_limit) {
623 /* allocated, but too high */
624 __free_pages(page, get_order(bytes));
626 if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
627 phys_limit < DMA_BIT_MASK(64) &&
628 !(gfp & (__GFP_DMA32 | __GFP_DMA)))
630 else if (IS_ENABLED(CONFIG_ZONE_DMA) &&
632 gfp = (gfp & ~__GFP_DMA32) | __GFP_DMA;
641 * swiotlb_free_tlb() - free a dynamically allocated IO TLB buffer
642 * @vaddr: Virtual address of the buffer.
643 * @bytes: Size of the buffer.
645 static void swiotlb_free_tlb(void *vaddr, size_t bytes)
647 if (IS_ENABLED(CONFIG_DMA_COHERENT_POOL) &&
648 dma_free_from_pool(NULL, vaddr, bytes))
651 /* Intentional leak if pages cannot be encrypted again. */
652 if (!set_memory_encrypted((unsigned long)vaddr, PFN_UP(bytes)))
653 __free_pages(virt_to_page(vaddr), get_order(bytes));
657 * swiotlb_alloc_pool() - allocate a new IO TLB memory pool
658 * @dev: Device for which a memory pool is allocated.
659 * @minslabs: Minimum number of slabs.
660 * @nslabs: Desired (maximum) number of slabs.
661 * @nareas: Number of areas.
662 * @phys_limit: Maximum DMA buffer physical address.
663 * @gfp: GFP flags for the allocations.
665 * Allocate and initialize a new IO TLB memory pool. The actual number of
666 * slabs may be reduced if allocation of @nslabs fails. If even
667 * @minslabs cannot be allocated, this function fails.
669 * Return: New memory pool, or %NULL on allocation failure.
671 static struct io_tlb_pool *swiotlb_alloc_pool(struct device *dev,
672 unsigned long minslabs, unsigned long nslabs,
673 unsigned int nareas, u64 phys_limit, gfp_t gfp)
675 struct io_tlb_pool *pool;
676 unsigned int slot_order;
681 if (nslabs > SLABS_PER_PAGE << MAX_ORDER) {
682 nslabs = SLABS_PER_PAGE << MAX_ORDER;
683 nareas = limit_nareas(nareas, nslabs);
686 pool_size = sizeof(*pool) + array_size(sizeof(*pool->areas), nareas);
687 pool = kzalloc(pool_size, gfp);
690 pool->areas = (void *)pool + sizeof(*pool);
692 tlb_size = nslabs << IO_TLB_SHIFT;
693 while (!(tlb = swiotlb_alloc_tlb(dev, tlb_size, phys_limit, gfp))) {
694 if (nslabs <= minslabs)
696 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
697 nareas = limit_nareas(nareas, nslabs);
698 tlb_size = nslabs << IO_TLB_SHIFT;
701 slot_order = get_order(array_size(sizeof(*pool->slots), nslabs));
702 pool->slots = (struct io_tlb_slot *)
703 __get_free_pages(gfp, slot_order);
707 swiotlb_init_io_tlb_pool(pool, page_to_phys(tlb), nslabs, true, nareas);
711 swiotlb_free_tlb(page_address(tlb), tlb_size);
719 * swiotlb_dyn_alloc() - dynamic memory pool allocation worker
720 * @work: Pointer to dyn_alloc in struct io_tlb_mem.
722 static void swiotlb_dyn_alloc(struct work_struct *work)
724 struct io_tlb_mem *mem =
725 container_of(work, struct io_tlb_mem, dyn_alloc);
726 struct io_tlb_pool *pool;
728 pool = swiotlb_alloc_pool(NULL, IO_TLB_MIN_SLABS, default_nslabs,
729 default_nareas, mem->phys_limit, GFP_KERNEL);
731 pr_warn_ratelimited("Failed to allocate new pool");
735 add_mem_pool(mem, pool);
739 * swiotlb_dyn_free() - RCU callback to free a memory pool
740 * @rcu: RCU head in the corresponding struct io_tlb_pool.
742 static void swiotlb_dyn_free(struct rcu_head *rcu)
744 struct io_tlb_pool *pool = container_of(rcu, struct io_tlb_pool, rcu);
745 size_t slots_size = array_size(sizeof(*pool->slots), pool->nslabs);
746 size_t tlb_size = pool->end - pool->start;
748 free_pages((unsigned long)pool->slots, get_order(slots_size));
749 swiotlb_free_tlb(pool->vaddr, tlb_size);
754 * swiotlb_find_pool() - find the IO TLB pool for a physical address
755 * @dev: Device which has mapped the DMA buffer.
756 * @paddr: Physical address within the DMA buffer.
758 * Find the IO TLB memory pool descriptor which contains the given physical
761 * Return: Memory pool which contains @paddr, or %NULL if none.
763 struct io_tlb_pool *swiotlb_find_pool(struct device *dev, phys_addr_t paddr)
765 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
766 struct io_tlb_pool *pool;
769 list_for_each_entry_rcu(pool, &mem->pools, node) {
770 if (paddr >= pool->start && paddr < pool->end)
774 list_for_each_entry_rcu(pool, &dev->dma_io_tlb_pools, node) {
775 if (paddr >= pool->start && paddr < pool->end)
785 * swiotlb_del_pool() - remove an IO TLB pool from a device
786 * @dev: Owning device.
787 * @pool: Memory pool to be removed.
789 static void swiotlb_del_pool(struct device *dev, struct io_tlb_pool *pool)
793 spin_lock_irqsave(&dev->dma_io_tlb_lock, flags);
794 list_del_rcu(&pool->node);
795 spin_unlock_irqrestore(&dev->dma_io_tlb_lock, flags);
797 call_rcu(&pool->rcu, swiotlb_dyn_free);
800 #endif /* CONFIG_SWIOTLB_DYNAMIC */
803 * swiotlb_dev_init() - initialize swiotlb fields in &struct device
804 * @dev: Device to be initialized.
806 void swiotlb_dev_init(struct device *dev)
808 dev->dma_io_tlb_mem = &io_tlb_default_mem;
809 #ifdef CONFIG_SWIOTLB_DYNAMIC
810 INIT_LIST_HEAD(&dev->dma_io_tlb_pools);
811 spin_lock_init(&dev->dma_io_tlb_lock);
812 dev->dma_uses_io_tlb = false;
817 * Return the offset into a iotlb slot required to keep the device happy.
819 static unsigned int swiotlb_align_offset(struct device *dev, u64 addr)
821 return addr & dma_get_min_align_mask(dev) & (IO_TLB_SIZE - 1);
825 * Bounce: copy the swiotlb buffer from or back to the original dma location
827 static void swiotlb_bounce(struct device *dev, phys_addr_t tlb_addr, size_t size,
828 enum dma_data_direction dir)
830 struct io_tlb_pool *mem = swiotlb_find_pool(dev, tlb_addr);
831 int index = (tlb_addr - mem->start) >> IO_TLB_SHIFT;
832 phys_addr_t orig_addr = mem->slots[index].orig_addr;
833 size_t alloc_size = mem->slots[index].alloc_size;
834 unsigned long pfn = PFN_DOWN(orig_addr);
835 unsigned char *vaddr = mem->vaddr + tlb_addr - mem->start;
836 unsigned int tlb_offset, orig_addr_offset;
838 if (orig_addr == INVALID_PHYS_ADDR)
841 tlb_offset = tlb_addr & (IO_TLB_SIZE - 1);
842 orig_addr_offset = swiotlb_align_offset(dev, orig_addr);
843 if (tlb_offset < orig_addr_offset) {
844 dev_WARN_ONCE(dev, 1,
845 "Access before mapping start detected. orig offset %u, requested offset %u.\n",
846 orig_addr_offset, tlb_offset);
850 tlb_offset -= orig_addr_offset;
851 if (tlb_offset > alloc_size) {
852 dev_WARN_ONCE(dev, 1,
853 "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu+%u.\n",
854 alloc_size, size, tlb_offset);
858 orig_addr += tlb_offset;
859 alloc_size -= tlb_offset;
861 if (size > alloc_size) {
862 dev_WARN_ONCE(dev, 1,
863 "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu.\n",
868 if (PageHighMem(pfn_to_page(pfn))) {
869 unsigned int offset = orig_addr & ~PAGE_MASK;
875 sz = min_t(size_t, PAGE_SIZE - offset, size);
877 local_irq_save(flags);
878 page = pfn_to_page(pfn);
879 if (dir == DMA_TO_DEVICE)
880 memcpy_from_page(vaddr, page, offset, sz);
882 memcpy_to_page(page, offset, vaddr, sz);
883 local_irq_restore(flags);
890 } else if (dir == DMA_TO_DEVICE) {
891 memcpy(vaddr, phys_to_virt(orig_addr), size);
893 memcpy(phys_to_virt(orig_addr), vaddr, size);
897 static inline phys_addr_t slot_addr(phys_addr_t start, phys_addr_t idx)
899 return start + (idx << IO_TLB_SHIFT);
903 * Carefully handle integer overflow which can occur when boundary_mask == ~0UL.
905 static inline unsigned long get_max_slots(unsigned long boundary_mask)
907 return (boundary_mask >> IO_TLB_SHIFT) + 1;
910 static unsigned int wrap_area_index(struct io_tlb_pool *mem, unsigned int index)
912 if (index >= mem->area_nslabs)
918 * Track the total used slots with a global atomic value in order to have
919 * correct information to determine the high water mark. The mem_used()
920 * function gives imprecise results because there's no locking across
923 #ifdef CONFIG_DEBUG_FS
924 static void inc_used_and_hiwater(struct io_tlb_mem *mem, unsigned int nslots)
926 unsigned long old_hiwater, new_used;
928 new_used = atomic_long_add_return(nslots, &mem->total_used);
929 old_hiwater = atomic_long_read(&mem->used_hiwater);
931 if (new_used <= old_hiwater)
933 } while (!atomic_long_try_cmpxchg(&mem->used_hiwater,
934 &old_hiwater, new_used));
937 static void dec_used(struct io_tlb_mem *mem, unsigned int nslots)
939 atomic_long_sub(nslots, &mem->total_used);
942 #else /* !CONFIG_DEBUG_FS */
943 static void inc_used_and_hiwater(struct io_tlb_mem *mem, unsigned int nslots)
946 static void dec_used(struct io_tlb_mem *mem, unsigned int nslots)
949 #endif /* CONFIG_DEBUG_FS */
952 * swiotlb_area_find_slots() - search for slots in one IO TLB memory area
953 * @dev: Device which maps the buffer.
954 * @pool: Memory pool to be searched.
955 * @area_index: Index of the IO TLB memory area to be searched.
956 * @orig_addr: Original (non-bounced) IO buffer address.
957 * @alloc_size: Total requested size of the bounce buffer,
958 * including initial alignment padding.
959 * @alloc_align_mask: Required alignment of the allocated buffer.
961 * Find a suitable sequence of IO TLB entries for the request and allocate
962 * a buffer from the given IO TLB memory area.
963 * This function takes care of locking.
965 * Return: Index of the first allocated slot, or -1 on error.
967 static int swiotlb_area_find_slots(struct device *dev, struct io_tlb_pool *pool,
968 int area_index, phys_addr_t orig_addr, size_t alloc_size,
969 unsigned int alloc_align_mask)
971 struct io_tlb_area *area = pool->areas + area_index;
972 unsigned long boundary_mask = dma_get_seg_boundary(dev);
973 dma_addr_t tbl_dma_addr =
974 phys_to_dma_unencrypted(dev, pool->start) & boundary_mask;
975 unsigned long max_slots = get_max_slots(boundary_mask);
976 unsigned int iotlb_align_mask =
977 dma_get_min_align_mask(dev) | alloc_align_mask;
978 unsigned int nslots = nr_slots(alloc_size), stride;
979 unsigned int offset = swiotlb_align_offset(dev, orig_addr);
980 unsigned int index, slots_checked, count = 0, i;
982 unsigned int slot_base;
983 unsigned int slot_index;
986 BUG_ON(area_index >= pool->nareas);
989 * For allocations of PAGE_SIZE or larger only look for page aligned
992 if (alloc_size >= PAGE_SIZE)
993 iotlb_align_mask |= ~PAGE_MASK;
994 iotlb_align_mask &= ~(IO_TLB_SIZE - 1);
997 * For mappings with an alignment requirement don't bother looping to
998 * unaligned slots once we found an aligned one.
1000 stride = (iotlb_align_mask >> IO_TLB_SHIFT) + 1;
1002 spin_lock_irqsave(&area->lock, flags);
1003 if (unlikely(nslots > pool->area_nslabs - area->used))
1006 slot_base = area_index * pool->area_nslabs;
1007 index = area->index;
1009 for (slots_checked = 0; slots_checked < pool->area_nslabs; ) {
1010 slot_index = slot_base + index;
1013 (slot_addr(tbl_dma_addr, slot_index) &
1014 iotlb_align_mask) != (orig_addr & iotlb_align_mask)) {
1015 index = wrap_area_index(pool, index + 1);
1020 if (!iommu_is_span_boundary(slot_index, nslots,
1021 nr_slots(tbl_dma_addr),
1023 if (pool->slots[slot_index].list >= nslots)
1026 index = wrap_area_index(pool, index + stride);
1027 slots_checked += stride;
1031 spin_unlock_irqrestore(&area->lock, flags);
1036 * If we find a slot that indicates we have 'nslots' number of
1037 * contiguous buffers, we allocate the buffers from that slot onwards
1038 * and set the list of free entries to '0' indicating unavailable.
1040 for (i = slot_index; i < slot_index + nslots; i++) {
1041 pool->slots[i].list = 0;
1042 pool->slots[i].alloc_size = alloc_size - (offset +
1043 ((i - slot_index) << IO_TLB_SHIFT));
1045 for (i = slot_index - 1;
1046 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 &&
1047 pool->slots[i].list; i--)
1048 pool->slots[i].list = ++count;
1051 * Update the indices to avoid searching in the next round.
1053 area->index = wrap_area_index(pool, index + nslots);
1054 area->used += nslots;
1055 spin_unlock_irqrestore(&area->lock, flags);
1057 inc_used_and_hiwater(dev->dma_io_tlb_mem, nslots);
1062 * swiotlb_pool_find_slots() - search for slots in one memory pool
1063 * @dev: Device which maps the buffer.
1064 * @pool: Memory pool to be searched.
1065 * @orig_addr: Original (non-bounced) IO buffer address.
1066 * @alloc_size: Total requested size of the bounce buffer,
1067 * including initial alignment padding.
1068 * @alloc_align_mask: Required alignment of the allocated buffer.
1070 * Search through one memory pool to find a sequence of slots that match the
1071 * allocation constraints.
1073 * Return: Index of the first allocated slot, or -1 on error.
1075 static int swiotlb_pool_find_slots(struct device *dev, struct io_tlb_pool *pool,
1076 phys_addr_t orig_addr, size_t alloc_size,
1077 unsigned int alloc_align_mask)
1079 int start = raw_smp_processor_id() & (pool->nareas - 1);
1080 int i = start, index;
1083 index = swiotlb_area_find_slots(dev, pool, i, orig_addr,
1084 alloc_size, alloc_align_mask);
1087 if (++i >= pool->nareas)
1089 } while (i != start);
1094 #ifdef CONFIG_SWIOTLB_DYNAMIC
1097 * swiotlb_find_slots() - search for slots in the whole swiotlb
1098 * @dev: Device which maps the buffer.
1099 * @orig_addr: Original (non-bounced) IO buffer address.
1100 * @alloc_size: Total requested size of the bounce buffer,
1101 * including initial alignment padding.
1102 * @alloc_align_mask: Required alignment of the allocated buffer.
1103 * @retpool: Used memory pool, updated on return.
1105 * Search through the whole software IO TLB to find a sequence of slots that
1106 * match the allocation constraints.
1108 * Return: Index of the first allocated slot, or -1 on error.
1110 static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr,
1111 size_t alloc_size, unsigned int alloc_align_mask,
1112 struct io_tlb_pool **retpool)
1114 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1115 struct io_tlb_pool *pool;
1116 unsigned long nslabs;
1117 unsigned long flags;
1122 list_for_each_entry_rcu(pool, &mem->pools, node) {
1123 index = swiotlb_pool_find_slots(dev, pool, orig_addr,
1124 alloc_size, alloc_align_mask);
1134 schedule_work(&mem->dyn_alloc);
1136 nslabs = nr_slots(alloc_size);
1137 phys_limit = min_not_zero(*dev->dma_mask, dev->bus_dma_limit);
1138 pool = swiotlb_alloc_pool(dev, nslabs, nslabs, 1, phys_limit,
1139 GFP_NOWAIT | __GFP_NOWARN);
1143 index = swiotlb_pool_find_slots(dev, pool, orig_addr,
1144 alloc_size, alloc_align_mask);
1146 swiotlb_dyn_free(&pool->rcu);
1150 pool->transient = true;
1151 spin_lock_irqsave(&dev->dma_io_tlb_lock, flags);
1152 list_add_rcu(&pool->node, &dev->dma_io_tlb_pools);
1153 spin_unlock_irqrestore(&dev->dma_io_tlb_lock, flags);
1156 WRITE_ONCE(dev->dma_uses_io_tlb, true);
1159 * The general barrier orders reads and writes against a presumed store
1160 * of the SWIOTLB buffer address by a device driver (to a driver private
1161 * data structure). It serves two purposes.
1163 * First, the store to dev->dma_uses_io_tlb must be ordered before the
1164 * presumed store. This guarantees that the returned buffer address
1165 * cannot be passed to another CPU before updating dev->dma_uses_io_tlb.
1167 * Second, the load from mem->pools must be ordered before the same
1168 * presumed store. This guarantees that the returned buffer address
1169 * cannot be observed by another CPU before an update of the RCU list
1170 * that was made by swiotlb_dyn_alloc() on a third CPU (cf. multicopy
1173 * See also the comment in is_swiotlb_buffer().
1181 #else /* !CONFIG_SWIOTLB_DYNAMIC */
1183 static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr,
1184 size_t alloc_size, unsigned int alloc_align_mask,
1185 struct io_tlb_pool **retpool)
1187 *retpool = &dev->dma_io_tlb_mem->defpool;
1188 return swiotlb_pool_find_slots(dev, *retpool,
1189 orig_addr, alloc_size, alloc_align_mask);
1192 #endif /* CONFIG_SWIOTLB_DYNAMIC */
1194 #ifdef CONFIG_DEBUG_FS
1197 * mem_used() - get number of used slots in an allocator
1198 * @mem: Software IO TLB allocator.
1200 * The result is accurate in this version of the function, because an atomic
1201 * counter is available if CONFIG_DEBUG_FS is set.
1203 * Return: Number of used slots.
1205 static unsigned long mem_used(struct io_tlb_mem *mem)
1207 return atomic_long_read(&mem->total_used);
1210 #else /* !CONFIG_DEBUG_FS */
1213 * mem_pool_used() - get number of used slots in a memory pool
1214 * @pool: Software IO TLB memory pool.
1216 * The result is not accurate, see mem_used().
1218 * Return: Approximate number of used slots.
1220 static unsigned long mem_pool_used(struct io_tlb_pool *pool)
1223 unsigned long used = 0;
1225 for (i = 0; i < pool->nareas; i++)
1226 used += pool->areas[i].used;
1231 * mem_used() - get number of used slots in an allocator
1232 * @mem: Software IO TLB allocator.
1234 * The result is not accurate, because there is no locking of individual
1237 * Return: Approximate number of used slots.
1239 static unsigned long mem_used(struct io_tlb_mem *mem)
1241 #ifdef CONFIG_SWIOTLB_DYNAMIC
1242 struct io_tlb_pool *pool;
1243 unsigned long used = 0;
1246 list_for_each_entry_rcu(pool, &mem->pools, node)
1247 used += mem_pool_used(pool);
1252 return mem_pool_used(&mem->defpool);
1256 #endif /* CONFIG_DEBUG_FS */
1258 phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
1259 size_t mapping_size, size_t alloc_size,
1260 unsigned int alloc_align_mask, enum dma_data_direction dir,
1261 unsigned long attrs)
1263 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1264 unsigned int offset = swiotlb_align_offset(dev, orig_addr);
1265 struct io_tlb_pool *pool;
1268 phys_addr_t tlb_addr;
1270 if (!mem || !mem->nslabs) {
1271 dev_warn_ratelimited(dev,
1272 "Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
1273 return (phys_addr_t)DMA_MAPPING_ERROR;
1276 if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
1277 pr_warn_once("Memory encryption is active and system is using DMA bounce buffers\n");
1279 if (mapping_size > alloc_size) {
1280 dev_warn_once(dev, "Invalid sizes (mapping: %zd bytes, alloc: %zd bytes)",
1281 mapping_size, alloc_size);
1282 return (phys_addr_t)DMA_MAPPING_ERROR;
1285 index = swiotlb_find_slots(dev, orig_addr,
1286 alloc_size + offset, alloc_align_mask, &pool);
1288 if (!(attrs & DMA_ATTR_NO_WARN))
1289 dev_warn_ratelimited(dev,
1290 "swiotlb buffer is full (sz: %zd bytes), total %lu (slots), used %lu (slots)\n",
1291 alloc_size, mem->nslabs, mem_used(mem));
1292 return (phys_addr_t)DMA_MAPPING_ERROR;
1296 * Save away the mapping from the original address to the DMA address.
1297 * This is needed when we sync the memory. Then we sync the buffer if
1300 for (i = 0; i < nr_slots(alloc_size + offset); i++)
1301 pool->slots[index + i].orig_addr = slot_addr(orig_addr, i);
1302 tlb_addr = slot_addr(pool->start, index) + offset;
1304 * When dir == DMA_FROM_DEVICE we could omit the copy from the orig
1305 * to the tlb buffer, if we knew for sure the device will
1306 * overwrite the entire current content. But we don't. Thus
1307 * unconditional bounce may prevent leaking swiotlb content (i.e.
1308 * kernel memory) to user-space.
1310 swiotlb_bounce(dev, tlb_addr, mapping_size, DMA_TO_DEVICE);
1314 static void swiotlb_release_slots(struct device *dev, phys_addr_t tlb_addr)
1316 struct io_tlb_pool *mem = swiotlb_find_pool(dev, tlb_addr);
1317 unsigned long flags;
1318 unsigned int offset = swiotlb_align_offset(dev, tlb_addr);
1319 int index = (tlb_addr - offset - mem->start) >> IO_TLB_SHIFT;
1320 int nslots = nr_slots(mem->slots[index].alloc_size + offset);
1321 int aindex = index / mem->area_nslabs;
1322 struct io_tlb_area *area = &mem->areas[aindex];
1326 * Return the buffer to the free list by setting the corresponding
1327 * entries to indicate the number of contiguous entries available.
1328 * While returning the entries to the free list, we merge the entries
1329 * with slots below and above the pool being returned.
1331 BUG_ON(aindex >= mem->nareas);
1333 spin_lock_irqsave(&area->lock, flags);
1334 if (index + nslots < ALIGN(index + 1, IO_TLB_SEGSIZE))
1335 count = mem->slots[index + nslots].list;
1340 * Step 1: return the slots to the free list, merging the slots with
1341 * superceeding slots
1343 for (i = index + nslots - 1; i >= index; i--) {
1344 mem->slots[i].list = ++count;
1345 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
1346 mem->slots[i].alloc_size = 0;
1350 * Step 2: merge the returned slots with the preceding slots, if
1351 * available (non zero)
1354 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 && mem->slots[i].list;
1356 mem->slots[i].list = ++count;
1357 area->used -= nslots;
1358 spin_unlock_irqrestore(&area->lock, flags);
1360 dec_used(dev->dma_io_tlb_mem, nslots);
1363 #ifdef CONFIG_SWIOTLB_DYNAMIC
1366 * swiotlb_del_transient() - delete a transient memory pool
1367 * @dev: Device which mapped the buffer.
1368 * @tlb_addr: Physical address within a bounce buffer.
1370 * Check whether the address belongs to a transient SWIOTLB memory pool.
1371 * If yes, then delete the pool.
1373 * Return: %true if @tlb_addr belonged to a transient pool that was released.
1375 static bool swiotlb_del_transient(struct device *dev, phys_addr_t tlb_addr)
1377 struct io_tlb_pool *pool;
1379 pool = swiotlb_find_pool(dev, tlb_addr);
1380 if (!pool->transient)
1383 dec_used(dev->dma_io_tlb_mem, pool->nslabs);
1384 swiotlb_del_pool(dev, pool);
1388 #else /* !CONFIG_SWIOTLB_DYNAMIC */
1390 static inline bool swiotlb_del_transient(struct device *dev,
1391 phys_addr_t tlb_addr)
1396 #endif /* CONFIG_SWIOTLB_DYNAMIC */
1399 * tlb_addr is the physical address of the bounce buffer to unmap.
1401 void swiotlb_tbl_unmap_single(struct device *dev, phys_addr_t tlb_addr,
1402 size_t mapping_size, enum dma_data_direction dir,
1403 unsigned long attrs)
1406 * First, sync the memory before unmapping the entry
1408 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
1409 (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
1410 swiotlb_bounce(dev, tlb_addr, mapping_size, DMA_FROM_DEVICE);
1412 if (swiotlb_del_transient(dev, tlb_addr))
1414 swiotlb_release_slots(dev, tlb_addr);
1417 void swiotlb_sync_single_for_device(struct device *dev, phys_addr_t tlb_addr,
1418 size_t size, enum dma_data_direction dir)
1420 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
1421 swiotlb_bounce(dev, tlb_addr, size, DMA_TO_DEVICE);
1423 BUG_ON(dir != DMA_FROM_DEVICE);
1426 void swiotlb_sync_single_for_cpu(struct device *dev, phys_addr_t tlb_addr,
1427 size_t size, enum dma_data_direction dir)
1429 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
1430 swiotlb_bounce(dev, tlb_addr, size, DMA_FROM_DEVICE);
1432 BUG_ON(dir != DMA_TO_DEVICE);
1436 * Create a swiotlb mapping for the buffer at @paddr, and in case of DMAing
1437 * to the device copy the data into it as well.
1439 dma_addr_t swiotlb_map(struct device *dev, phys_addr_t paddr, size_t size,
1440 enum dma_data_direction dir, unsigned long attrs)
1442 phys_addr_t swiotlb_addr;
1443 dma_addr_t dma_addr;
1445 trace_swiotlb_bounced(dev, phys_to_dma(dev, paddr), size);
1447 swiotlb_addr = swiotlb_tbl_map_single(dev, paddr, size, size, 0, dir,
1449 if (swiotlb_addr == (phys_addr_t)DMA_MAPPING_ERROR)
1450 return DMA_MAPPING_ERROR;
1452 /* Ensure that the address returned is DMA'ble */
1453 dma_addr = phys_to_dma_unencrypted(dev, swiotlb_addr);
1454 if (unlikely(!dma_capable(dev, dma_addr, size, true))) {
1455 swiotlb_tbl_unmap_single(dev, swiotlb_addr, size, dir,
1456 attrs | DMA_ATTR_SKIP_CPU_SYNC);
1457 dev_WARN_ONCE(dev, 1,
1458 "swiotlb addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
1459 &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
1460 return DMA_MAPPING_ERROR;
1463 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1464 arch_sync_dma_for_device(swiotlb_addr, size, dir);
1468 size_t swiotlb_max_mapping_size(struct device *dev)
1470 int min_align_mask = dma_get_min_align_mask(dev);
1474 * swiotlb_find_slots() skips slots according to
1475 * min align mask. This affects max mapping size.
1476 * Take it into acount here.
1479 min_align = roundup(min_align_mask, IO_TLB_SIZE);
1481 return ((size_t)IO_TLB_SIZE) * IO_TLB_SEGSIZE - min_align;
1485 * is_swiotlb_allocated() - check if the default software IO TLB is initialized
1487 bool is_swiotlb_allocated(void)
1489 return io_tlb_default_mem.nslabs;
1492 bool is_swiotlb_active(struct device *dev)
1494 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1496 return mem && mem->nslabs;
1500 * default_swiotlb_base() - get the base address of the default SWIOTLB
1502 * Get the lowest physical address used by the default software IO TLB pool.
1504 phys_addr_t default_swiotlb_base(void)
1506 #ifdef CONFIG_SWIOTLB_DYNAMIC
1507 io_tlb_default_mem.can_grow = false;
1509 return io_tlb_default_mem.defpool.start;
1513 * default_swiotlb_limit() - get the address limit of the default SWIOTLB
1515 * Get the highest physical address used by the default software IO TLB pool.
1517 phys_addr_t default_swiotlb_limit(void)
1519 #ifdef CONFIG_SWIOTLB_DYNAMIC
1520 return io_tlb_default_mem.phys_limit;
1522 return io_tlb_default_mem.defpool.end - 1;
1526 #ifdef CONFIG_DEBUG_FS
1528 static int io_tlb_used_get(void *data, u64 *val)
1530 struct io_tlb_mem *mem = data;
1532 *val = mem_used(mem);
1536 static int io_tlb_hiwater_get(void *data, u64 *val)
1538 struct io_tlb_mem *mem = data;
1540 *val = atomic_long_read(&mem->used_hiwater);
1544 static int io_tlb_hiwater_set(void *data, u64 val)
1546 struct io_tlb_mem *mem = data;
1548 /* Only allow setting to zero */
1552 atomic_long_set(&mem->used_hiwater, val);
1556 DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_used, io_tlb_used_get, NULL, "%llu\n");
1557 DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_hiwater, io_tlb_hiwater_get,
1558 io_tlb_hiwater_set, "%llu\n");
1560 static void swiotlb_create_debugfs_files(struct io_tlb_mem *mem,
1561 const char *dirname)
1563 atomic_long_set(&mem->total_used, 0);
1564 atomic_long_set(&mem->used_hiwater, 0);
1566 mem->debugfs = debugfs_create_dir(dirname, io_tlb_default_mem.debugfs);
1570 debugfs_create_ulong("io_tlb_nslabs", 0400, mem->debugfs, &mem->nslabs);
1571 debugfs_create_file("io_tlb_used", 0400, mem->debugfs, mem,
1573 debugfs_create_file("io_tlb_used_hiwater", 0600, mem->debugfs, mem,
1574 &fops_io_tlb_hiwater);
1577 static int __init swiotlb_create_default_debugfs(void)
1579 swiotlb_create_debugfs_files(&io_tlb_default_mem, "swiotlb");
1583 late_initcall(swiotlb_create_default_debugfs);
1585 #else /* !CONFIG_DEBUG_FS */
1587 static inline void swiotlb_create_debugfs_files(struct io_tlb_mem *mem,
1588 const char *dirname)
1592 #endif /* CONFIG_DEBUG_FS */
1594 #ifdef CONFIG_DMA_RESTRICTED_POOL
1596 struct page *swiotlb_alloc(struct device *dev, size_t size)
1598 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1599 struct io_tlb_pool *pool;
1600 phys_addr_t tlb_addr;
1606 index = swiotlb_find_slots(dev, 0, size, 0, &pool);
1610 tlb_addr = slot_addr(pool->start, index);
1612 return pfn_to_page(PFN_DOWN(tlb_addr));
1615 bool swiotlb_free(struct device *dev, struct page *page, size_t size)
1617 phys_addr_t tlb_addr = page_to_phys(page);
1619 if (!is_swiotlb_buffer(dev, tlb_addr))
1622 swiotlb_release_slots(dev, tlb_addr);
1627 static int rmem_swiotlb_device_init(struct reserved_mem *rmem,
1630 struct io_tlb_mem *mem = rmem->priv;
1631 unsigned long nslabs = rmem->size >> IO_TLB_SHIFT;
1633 /* Set Per-device io tlb area to one */
1634 unsigned int nareas = 1;
1636 if (PageHighMem(pfn_to_page(PHYS_PFN(rmem->base)))) {
1637 dev_err(dev, "Restricted DMA pool must be accessible within the linear mapping.");
1642 * Since multiple devices can share the same pool, the private data,
1643 * io_tlb_mem struct, will be initialized by the first device attached
1647 struct io_tlb_pool *pool;
1649 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
1652 pool = &mem->defpool;
1654 pool->slots = kcalloc(nslabs, sizeof(*pool->slots), GFP_KERNEL);
1660 pool->areas = kcalloc(nareas, sizeof(*pool->areas),
1668 set_memory_decrypted((unsigned long)phys_to_virt(rmem->base),
1669 rmem->size >> PAGE_SHIFT);
1670 swiotlb_init_io_tlb_pool(pool, rmem->base, nslabs,
1672 mem->force_bounce = true;
1673 mem->for_alloc = true;
1674 #ifdef CONFIG_SWIOTLB_DYNAMIC
1675 spin_lock_init(&mem->lock);
1677 add_mem_pool(mem, pool);
1681 swiotlb_create_debugfs_files(mem, rmem->name);
1684 dev->dma_io_tlb_mem = mem;
1689 static void rmem_swiotlb_device_release(struct reserved_mem *rmem,
1692 dev->dma_io_tlb_mem = &io_tlb_default_mem;
1695 static const struct reserved_mem_ops rmem_swiotlb_ops = {
1696 .device_init = rmem_swiotlb_device_init,
1697 .device_release = rmem_swiotlb_device_release,
1700 static int __init rmem_swiotlb_setup(struct reserved_mem *rmem)
1702 unsigned long node = rmem->fdt_node;
1704 if (of_get_flat_dt_prop(node, "reusable", NULL) ||
1705 of_get_flat_dt_prop(node, "linux,cma-default", NULL) ||
1706 of_get_flat_dt_prop(node, "linux,dma-default", NULL) ||
1707 of_get_flat_dt_prop(node, "no-map", NULL))
1710 rmem->ops = &rmem_swiotlb_ops;
1711 pr_info("Reserved memory: created restricted DMA pool at %pa, size %ld MiB\n",
1712 &rmem->base, (unsigned long)rmem->size / SZ_1M);
1716 RESERVEDMEM_OF_DECLARE(dma, "restricted-dma-pool", rmem_swiotlb_setup);
1717 #endif /* CONFIG_DMA_RESTRICTED_POOL */