Loading...
Note: File does not exist in v4.6.
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Dynamic DMA mapping support.
4 *
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>
11 *
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
19 */
20
21#define pr_fmt(fmt) "software IO TLB: " fmt
22
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>
32#include <linux/io.h>
33#include <linux/iommu-helper.h>
34#include <linux/init.h>
35#include <linux/memblock.h>
36#include <linux/mm.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
46#include <linux/of.h>
47#include <linux/of_fdt.h>
48#include <linux/of_reserved_mem.h>
49#include <linux/slab.h>
50#endif
51
52#define CREATE_TRACE_POINTS
53#include <trace/events/swiotlb.h>
54
55#define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
56
57/*
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.
61 */
62#define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
63
64#define INVALID_PHYS_ADDR (~(phys_addr_t)0)
65
66/**
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
71 * from each index.
72 */
73struct io_tlb_slot {
74 phys_addr_t orig_addr;
75 size_t alloc_size;
76 unsigned int list;
77};
78
79static bool swiotlb_force_bounce;
80static bool swiotlb_force_disable;
81
82#ifdef CONFIG_SWIOTLB_DYNAMIC
83
84static void swiotlb_dyn_alloc(struct work_struct *work);
85
86static 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,
90 swiotlb_dyn_alloc),
91};
92
93#else /* !CONFIG_SWIOTLB_DYNAMIC */
94
95static struct io_tlb_mem io_tlb_default_mem;
96
97#endif /* CONFIG_SWIOTLB_DYNAMIC */
98
99static unsigned long default_nslabs = IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT;
100static unsigned long default_nareas;
101
102/**
103 * struct io_tlb_area - IO TLB memory area descriptor
104 *
105 * This is a single area with a single lock.
106 *
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
110 * unmap calls.
111 */
112struct io_tlb_area {
113 unsigned long used;
114 unsigned int index;
115 spinlock_t lock;
116};
117
118/*
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.
125 *
126 * Return true if default_nslabs is rounded up.
127 */
128static bool round_up_default_nslabs(void)
129{
130 if (!default_nareas)
131 return false;
132
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))
136 return false;
137 default_nslabs = roundup_pow_of_two(default_nslabs);
138 return true;
139}
140
141/**
142 * swiotlb_adjust_nareas() - adjust the number of areas and slots
143 * @nareas: Desired number of areas. Zero is treated as 1.
144 *
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
147 * size requirements.
148 */
149static void swiotlb_adjust_nareas(unsigned int nareas)
150{
151 if (!nareas)
152 nareas = 1;
153 else if (!is_power_of_2(nareas))
154 nareas = roundup_pow_of_two(nareas);
155
156 default_nareas = nareas;
157
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);
162}
163
164/**
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.
168 *
169 * Limit the number of areas to the maximum possible number of areas in
170 * a memory pool of the given size.
171 *
172 * Return: Maximum possible number of areas.
173 */
174static unsigned int limit_nareas(unsigned int nareas, unsigned long nslots)
175{
176 if (nslots < nareas * IO_TLB_SEGSIZE)
177 return nslots / IO_TLB_SEGSIZE;
178 return nareas;
179}
180
181static int __init
182setup_io_tlb_npages(char *str)
183{
184 if (isdigit(*str)) {
185 /* avoid tail segment of size < IO_TLB_SEGSIZE */
186 default_nslabs =
187 ALIGN(simple_strtoul(str, &str, 0), IO_TLB_SEGSIZE);
188 }
189 if (*str == ',')
190 ++str;
191 if (isdigit(*str))
192 swiotlb_adjust_nareas(simple_strtoul(str, &str, 0));
193 if (*str == ',')
194 ++str;
195 if (!strcmp(str, "force"))
196 swiotlb_force_bounce = true;
197 else if (!strcmp(str, "noforce"))
198 swiotlb_force_disable = true;
199
200 return 0;
201}
202early_param("swiotlb", setup_io_tlb_npages);
203
204unsigned long swiotlb_size_or_default(void)
205{
206 return default_nslabs << IO_TLB_SHIFT;
207}
208
209void __init swiotlb_adjust_size(unsigned long size)
210{
211 /*
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.
215 */
216 if (default_nslabs != IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT)
217 return;
218
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);
224}
225
226void swiotlb_print_info(void)
227{
228 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
229
230 if (!mem->nslabs) {
231 pr_warn("No low mem\n");
232 return;
233 }
234
235 pr_info("mapped [mem %pa-%pa] (%luMB)\n", &mem->start, &mem->end,
236 (mem->nslabs << IO_TLB_SHIFT) >> 20);
237}
238
239static inline unsigned long io_tlb_offset(unsigned long val)
240{
241 return val & (IO_TLB_SEGSIZE - 1);
242}
243
244static inline unsigned long nr_slots(u64 val)
245{
246 return DIV_ROUND_UP(val, IO_TLB_SIZE);
247}
248
249/*
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.
254 */
255void __init swiotlb_update_mem_attributes(void)
256{
257 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
258 unsigned long bytes;
259
260 if (!mem->nslabs || mem->late_alloc)
261 return;
262 bytes = PAGE_ALIGN(mem->nslabs << IO_TLB_SHIFT);
263 set_memory_decrypted((unsigned long)mem->vaddr, bytes >> PAGE_SHIFT);
264}
265
266static 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)
268{
269 void *vaddr = phys_to_virt(start);
270 unsigned long bytes = nslabs << IO_TLB_SHIFT, i;
271
272 mem->nslabs = nslabs;
273 mem->start = start;
274 mem->end = mem->start + bytes;
275 mem->late_alloc = late_alloc;
276 mem->nareas = nareas;
277 mem->area_nslabs = nslabs / mem->nareas;
278
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;
283 }
284
285 for (i = 0; i < mem->nslabs; i++) {
286 mem->slots[i].list = min(IO_TLB_SEGSIZE - io_tlb_offset(i),
287 mem->nslabs - i);
288 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
289 mem->slots[i].alloc_size = 0;
290 }
291
292 memset(vaddr, 0, bytes);
293 mem->vaddr = vaddr;
294 return;
295}
296
297/**
298 * add_mem_pool() - add a memory pool to the allocator
299 * @mem: Software IO TLB allocator.
300 * @pool: Memory pool to be added.
301 */
302static void add_mem_pool(struct io_tlb_mem *mem, struct io_tlb_pool *pool)
303{
304#ifdef CONFIG_SWIOTLB_DYNAMIC
305 spin_lock(&mem->lock);
306 list_add_rcu(&pool->node, &mem->pools);
307 mem->nslabs += pool->nslabs;
308 spin_unlock(&mem->lock);
309#else
310 mem->nslabs = pool->nslabs;
311#endif
312}
313
314static void __init *swiotlb_memblock_alloc(unsigned long nslabs,
315 unsigned int flags,
316 int (*remap)(void *tlb, unsigned long nslabs))
317{
318 size_t bytes = PAGE_ALIGN(nslabs << IO_TLB_SHIFT);
319 void *tlb;
320
321 /*
322 * By default allocate the bounce buffer memory from low memory, but
323 * allow to pick a location everywhere for hypervisors with guest
324 * memory encryption.
325 */
326 if (flags & SWIOTLB_ANY)
327 tlb = memblock_alloc(bytes, PAGE_SIZE);
328 else
329 tlb = memblock_alloc_low(bytes, PAGE_SIZE);
330
331 if (!tlb) {
332 pr_warn("%s: Failed to allocate %zu bytes tlb structure\n",
333 __func__, bytes);
334 return NULL;
335 }
336
337 if (remap && remap(tlb, nslabs) < 0) {
338 memblock_free(tlb, PAGE_ALIGN(bytes));
339 pr_warn("%s: Failed to remap %zu bytes\n", __func__, bytes);
340 return NULL;
341 }
342
343 return tlb;
344}
345
346/*
347 * Statically reserve bounce buffer space and initialize bounce buffer data
348 * structures for the software IO TLB used to implement the DMA API.
349 */
350void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags,
351 int (*remap)(void *tlb, unsigned long nslabs))
352{
353 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
354 unsigned long nslabs;
355 unsigned int nareas;
356 size_t alloc_size;
357 void *tlb;
358
359 if (!addressing_limit && !swiotlb_force_bounce)
360 return;
361 if (swiotlb_force_disable)
362 return;
363
364 io_tlb_default_mem.force_bounce =
365 swiotlb_force_bounce || (flags & SWIOTLB_FORCE);
366
367#ifdef CONFIG_SWIOTLB_DYNAMIC
368 if (!remap)
369 io_tlb_default_mem.can_grow = true;
370 if (flags & SWIOTLB_ANY)
371 io_tlb_default_mem.phys_limit = virt_to_phys(high_memory - 1);
372 else
373 io_tlb_default_mem.phys_limit = ARCH_LOW_ADDRESS_LIMIT;
374#endif
375
376 if (!default_nareas)
377 swiotlb_adjust_nareas(num_possible_cpus());
378
379 nslabs = default_nslabs;
380 nareas = limit_nareas(default_nareas, nslabs);
381 while ((tlb = swiotlb_memblock_alloc(nslabs, flags, remap)) == NULL) {
382 if (nslabs <= IO_TLB_MIN_SLABS)
383 return;
384 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
385 nareas = limit_nareas(nareas, nslabs);
386 }
387
388 if (default_nslabs != nslabs) {
389 pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs",
390 default_nslabs, nslabs);
391 default_nslabs = nslabs;
392 }
393
394 alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs));
395 mem->slots = memblock_alloc(alloc_size, PAGE_SIZE);
396 if (!mem->slots) {
397 pr_warn("%s: Failed to allocate %zu bytes align=0x%lx\n",
398 __func__, alloc_size, PAGE_SIZE);
399 return;
400 }
401
402 mem->areas = memblock_alloc(array_size(sizeof(struct io_tlb_area),
403 nareas), SMP_CACHE_BYTES);
404 if (!mem->areas) {
405 pr_warn("%s: Failed to allocate mem->areas.\n", __func__);
406 return;
407 }
408
409 swiotlb_init_io_tlb_pool(mem, __pa(tlb), nslabs, false, nareas);
410 add_mem_pool(&io_tlb_default_mem, mem);
411
412 if (flags & SWIOTLB_VERBOSE)
413 swiotlb_print_info();
414}
415
416void __init swiotlb_init(bool addressing_limit, unsigned int flags)
417{
418 swiotlb_init_remap(addressing_limit, flags, NULL);
419}
420
421/*
422 * Systems with larger DMA zones (those that don't support ISA) can
423 * initialize the swiotlb later using the slab allocator if needed.
424 * This should be just like above, but with some error catching.
425 */
426int swiotlb_init_late(size_t size, gfp_t gfp_mask,
427 int (*remap)(void *tlb, unsigned long nslabs))
428{
429 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
430 unsigned long nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE);
431 unsigned int nareas;
432 unsigned char *vstart = NULL;
433 unsigned int order, area_order;
434 bool retried = false;
435 int rc = 0;
436
437 if (io_tlb_default_mem.nslabs)
438 return 0;
439
440 if (swiotlb_force_disable)
441 return 0;
442
443 io_tlb_default_mem.force_bounce = swiotlb_force_bounce;
444
445#ifdef CONFIG_SWIOTLB_DYNAMIC
446 if (!remap)
447 io_tlb_default_mem.can_grow = true;
448 if (IS_ENABLED(CONFIG_ZONE_DMA) && (gfp_mask & __GFP_DMA))
449 io_tlb_default_mem.phys_limit = DMA_BIT_MASK(zone_dma_bits);
450 else if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp_mask & __GFP_DMA32))
451 io_tlb_default_mem.phys_limit = DMA_BIT_MASK(32);
452 else
453 io_tlb_default_mem.phys_limit = virt_to_phys(high_memory - 1);
454#endif
455
456 if (!default_nareas)
457 swiotlb_adjust_nareas(num_possible_cpus());
458
459retry:
460 order = get_order(nslabs << IO_TLB_SHIFT);
461 nslabs = SLABS_PER_PAGE << order;
462
463 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
464 vstart = (void *)__get_free_pages(gfp_mask | __GFP_NOWARN,
465 order);
466 if (vstart)
467 break;
468 order--;
469 nslabs = SLABS_PER_PAGE << order;
470 retried = true;
471 }
472
473 if (!vstart)
474 return -ENOMEM;
475
476 if (remap)
477 rc = remap(vstart, nslabs);
478 if (rc) {
479 free_pages((unsigned long)vstart, order);
480
481 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
482 if (nslabs < IO_TLB_MIN_SLABS)
483 return rc;
484 retried = true;
485 goto retry;
486 }
487
488 if (retried) {
489 pr_warn("only able to allocate %ld MB\n",
490 (PAGE_SIZE << order) >> 20);
491 }
492
493 nareas = limit_nareas(default_nareas, nslabs);
494 area_order = get_order(array_size(sizeof(*mem->areas), nareas));
495 mem->areas = (struct io_tlb_area *)
496 __get_free_pages(GFP_KERNEL | __GFP_ZERO, area_order);
497 if (!mem->areas)
498 goto error_area;
499
500 mem->slots = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
501 get_order(array_size(sizeof(*mem->slots), nslabs)));
502 if (!mem->slots)
503 goto error_slots;
504
505 set_memory_decrypted((unsigned long)vstart,
506 (nslabs << IO_TLB_SHIFT) >> PAGE_SHIFT);
507 swiotlb_init_io_tlb_pool(mem, virt_to_phys(vstart), nslabs, true,
508 nareas);
509 add_mem_pool(&io_tlb_default_mem, mem);
510
511 swiotlb_print_info();
512 return 0;
513
514error_slots:
515 free_pages((unsigned long)mem->areas, area_order);
516error_area:
517 free_pages((unsigned long)vstart, order);
518 return -ENOMEM;
519}
520
521void __init swiotlb_exit(void)
522{
523 struct io_tlb_pool *mem = &io_tlb_default_mem.defpool;
524 unsigned long tbl_vaddr;
525 size_t tbl_size, slots_size;
526 unsigned int area_order;
527
528 if (swiotlb_force_bounce)
529 return;
530
531 if (!mem->nslabs)
532 return;
533
534 pr_info("tearing down default memory pool\n");
535 tbl_vaddr = (unsigned long)phys_to_virt(mem->start);
536 tbl_size = PAGE_ALIGN(mem->end - mem->start);
537 slots_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), mem->nslabs));
538
539 set_memory_encrypted(tbl_vaddr, tbl_size >> PAGE_SHIFT);
540 if (mem->late_alloc) {
541 area_order = get_order(array_size(sizeof(*mem->areas),
542 mem->nareas));
543 free_pages((unsigned long)mem->areas, area_order);
544 free_pages(tbl_vaddr, get_order(tbl_size));
545 free_pages((unsigned long)mem->slots, get_order(slots_size));
546 } else {
547 memblock_free_late(__pa(mem->areas),
548 array_size(sizeof(*mem->areas), mem->nareas));
549 memblock_free_late(mem->start, tbl_size);
550 memblock_free_late(__pa(mem->slots), slots_size);
551 }
552
553 memset(mem, 0, sizeof(*mem));
554}
555
556#ifdef CONFIG_SWIOTLB_DYNAMIC
557
558/**
559 * alloc_dma_pages() - allocate pages to be used for DMA
560 * @gfp: GFP flags for the allocation.
561 * @bytes: Size of the buffer.
562 * @phys_limit: Maximum allowed physical address of the buffer.
563 *
564 * Allocate pages from the buddy allocator. If successful, make the allocated
565 * pages decrypted that they can be used for DMA.
566 *
567 * Return: Decrypted pages, %NULL on allocation failure, or ERR_PTR(-EAGAIN)
568 * if the allocated physical address was above @phys_limit.
569 */
570static struct page *alloc_dma_pages(gfp_t gfp, size_t bytes, u64 phys_limit)
571{
572 unsigned int order = get_order(bytes);
573 struct page *page;
574 phys_addr_t paddr;
575 void *vaddr;
576
577 page = alloc_pages(gfp, order);
578 if (!page)
579 return NULL;
580
581 paddr = page_to_phys(page);
582 if (paddr + bytes - 1 > phys_limit) {
583 __free_pages(page, order);
584 return ERR_PTR(-EAGAIN);
585 }
586
587 vaddr = phys_to_virt(paddr);
588 if (set_memory_decrypted((unsigned long)vaddr, PFN_UP(bytes)))
589 goto error;
590 return page;
591
592error:
593 /* Intentional leak if pages cannot be encrypted again. */
594 if (!set_memory_encrypted((unsigned long)vaddr, PFN_UP(bytes)))
595 __free_pages(page, order);
596 return NULL;
597}
598
599/**
600 * swiotlb_alloc_tlb() - allocate a dynamic IO TLB buffer
601 * @dev: Device for which a memory pool is allocated.
602 * @bytes: Size of the buffer.
603 * @phys_limit: Maximum allowed physical address of the buffer.
604 * @gfp: GFP flags for the allocation.
605 *
606 * Return: Allocated pages, or %NULL on allocation failure.
607 */
608static struct page *swiotlb_alloc_tlb(struct device *dev, size_t bytes,
609 u64 phys_limit, gfp_t gfp)
610{
611 struct page *page;
612
613 /*
614 * Allocate from the atomic pools if memory is encrypted and
615 * the allocation is atomic, because decrypting may block.
616 */
617 if (!gfpflags_allow_blocking(gfp) && dev && force_dma_unencrypted(dev)) {
618 void *vaddr;
619
620 if (!IS_ENABLED(CONFIG_DMA_COHERENT_POOL))
621 return NULL;
622
623 return dma_alloc_from_pool(dev, bytes, &vaddr, gfp,
624 dma_coherent_ok);
625 }
626
627 gfp &= ~GFP_ZONEMASK;
628 if (phys_limit <= DMA_BIT_MASK(zone_dma_bits))
629 gfp |= __GFP_DMA;
630 else if (phys_limit <= DMA_BIT_MASK(32))
631 gfp |= __GFP_DMA32;
632
633 while (IS_ERR(page = alloc_dma_pages(gfp, bytes, phys_limit))) {
634 if (IS_ENABLED(CONFIG_ZONE_DMA32) &&
635 phys_limit < DMA_BIT_MASK(64) &&
636 !(gfp & (__GFP_DMA32 | __GFP_DMA)))
637 gfp |= __GFP_DMA32;
638 else if (IS_ENABLED(CONFIG_ZONE_DMA) &&
639 !(gfp & __GFP_DMA))
640 gfp = (gfp & ~__GFP_DMA32) | __GFP_DMA;
641 else
642 return NULL;
643 }
644
645 return page;
646}
647
648/**
649 * swiotlb_free_tlb() - free a dynamically allocated IO TLB buffer
650 * @vaddr: Virtual address of the buffer.
651 * @bytes: Size of the buffer.
652 */
653static void swiotlb_free_tlb(void *vaddr, size_t bytes)
654{
655 if (IS_ENABLED(CONFIG_DMA_COHERENT_POOL) &&
656 dma_free_from_pool(NULL, vaddr, bytes))
657 return;
658
659 /* Intentional leak if pages cannot be encrypted again. */
660 if (!set_memory_encrypted((unsigned long)vaddr, PFN_UP(bytes)))
661 __free_pages(virt_to_page(vaddr), get_order(bytes));
662}
663
664/**
665 * swiotlb_alloc_pool() - allocate a new IO TLB memory pool
666 * @dev: Device for which a memory pool is allocated.
667 * @minslabs: Minimum number of slabs.
668 * @nslabs: Desired (maximum) number of slabs.
669 * @nareas: Number of areas.
670 * @phys_limit: Maximum DMA buffer physical address.
671 * @gfp: GFP flags for the allocations.
672 *
673 * Allocate and initialize a new IO TLB memory pool. The actual number of
674 * slabs may be reduced if allocation of @nslabs fails. If even
675 * @minslabs cannot be allocated, this function fails.
676 *
677 * Return: New memory pool, or %NULL on allocation failure.
678 */
679static struct io_tlb_pool *swiotlb_alloc_pool(struct device *dev,
680 unsigned long minslabs, unsigned long nslabs,
681 unsigned int nareas, u64 phys_limit, gfp_t gfp)
682{
683 struct io_tlb_pool *pool;
684 unsigned int slot_order;
685 struct page *tlb;
686 size_t pool_size;
687 size_t tlb_size;
688
689 if (nslabs > SLABS_PER_PAGE << MAX_PAGE_ORDER) {
690 nslabs = SLABS_PER_PAGE << MAX_PAGE_ORDER;
691 nareas = limit_nareas(nareas, nslabs);
692 }
693
694 pool_size = sizeof(*pool) + array_size(sizeof(*pool->areas), nareas);
695 pool = kzalloc(pool_size, gfp);
696 if (!pool)
697 goto error;
698 pool->areas = (void *)pool + sizeof(*pool);
699
700 tlb_size = nslabs << IO_TLB_SHIFT;
701 while (!(tlb = swiotlb_alloc_tlb(dev, tlb_size, phys_limit, gfp))) {
702 if (nslabs <= minslabs)
703 goto error_tlb;
704 nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE);
705 nareas = limit_nareas(nareas, nslabs);
706 tlb_size = nslabs << IO_TLB_SHIFT;
707 }
708
709 slot_order = get_order(array_size(sizeof(*pool->slots), nslabs));
710 pool->slots = (struct io_tlb_slot *)
711 __get_free_pages(gfp, slot_order);
712 if (!pool->slots)
713 goto error_slots;
714
715 swiotlb_init_io_tlb_pool(pool, page_to_phys(tlb), nslabs, true, nareas);
716 return pool;
717
718error_slots:
719 swiotlb_free_tlb(page_address(tlb), tlb_size);
720error_tlb:
721 kfree(pool);
722error:
723 return NULL;
724}
725
726/**
727 * swiotlb_dyn_alloc() - dynamic memory pool allocation worker
728 * @work: Pointer to dyn_alloc in struct io_tlb_mem.
729 */
730static void swiotlb_dyn_alloc(struct work_struct *work)
731{
732 struct io_tlb_mem *mem =
733 container_of(work, struct io_tlb_mem, dyn_alloc);
734 struct io_tlb_pool *pool;
735
736 pool = swiotlb_alloc_pool(NULL, IO_TLB_MIN_SLABS, default_nslabs,
737 default_nareas, mem->phys_limit, GFP_KERNEL);
738 if (!pool) {
739 pr_warn_ratelimited("Failed to allocate new pool");
740 return;
741 }
742
743 add_mem_pool(mem, pool);
744}
745
746/**
747 * swiotlb_dyn_free() - RCU callback to free a memory pool
748 * @rcu: RCU head in the corresponding struct io_tlb_pool.
749 */
750static void swiotlb_dyn_free(struct rcu_head *rcu)
751{
752 struct io_tlb_pool *pool = container_of(rcu, struct io_tlb_pool, rcu);
753 size_t slots_size = array_size(sizeof(*pool->slots), pool->nslabs);
754 size_t tlb_size = pool->end - pool->start;
755
756 free_pages((unsigned long)pool->slots, get_order(slots_size));
757 swiotlb_free_tlb(pool->vaddr, tlb_size);
758 kfree(pool);
759}
760
761/**
762 * swiotlb_find_pool() - find the IO TLB pool for a physical address
763 * @dev: Device which has mapped the DMA buffer.
764 * @paddr: Physical address within the DMA buffer.
765 *
766 * Find the IO TLB memory pool descriptor which contains the given physical
767 * address, if any.
768 *
769 * Return: Memory pool which contains @paddr, or %NULL if none.
770 */
771struct io_tlb_pool *swiotlb_find_pool(struct device *dev, phys_addr_t paddr)
772{
773 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
774 struct io_tlb_pool *pool;
775
776 rcu_read_lock();
777 list_for_each_entry_rcu(pool, &mem->pools, node) {
778 if (paddr >= pool->start && paddr < pool->end)
779 goto out;
780 }
781
782 list_for_each_entry_rcu(pool, &dev->dma_io_tlb_pools, node) {
783 if (paddr >= pool->start && paddr < pool->end)
784 goto out;
785 }
786 pool = NULL;
787out:
788 rcu_read_unlock();
789 return pool;
790}
791
792/**
793 * swiotlb_del_pool() - remove an IO TLB pool from a device
794 * @dev: Owning device.
795 * @pool: Memory pool to be removed.
796 */
797static void swiotlb_del_pool(struct device *dev, struct io_tlb_pool *pool)
798{
799 unsigned long flags;
800
801 spin_lock_irqsave(&dev->dma_io_tlb_lock, flags);
802 list_del_rcu(&pool->node);
803 spin_unlock_irqrestore(&dev->dma_io_tlb_lock, flags);
804
805 call_rcu(&pool->rcu, swiotlb_dyn_free);
806}
807
808#endif /* CONFIG_SWIOTLB_DYNAMIC */
809
810/**
811 * swiotlb_dev_init() - initialize swiotlb fields in &struct device
812 * @dev: Device to be initialized.
813 */
814void swiotlb_dev_init(struct device *dev)
815{
816 dev->dma_io_tlb_mem = &io_tlb_default_mem;
817#ifdef CONFIG_SWIOTLB_DYNAMIC
818 INIT_LIST_HEAD(&dev->dma_io_tlb_pools);
819 spin_lock_init(&dev->dma_io_tlb_lock);
820 dev->dma_uses_io_tlb = false;
821#endif
822}
823
824/*
825 * Return the offset into a iotlb slot required to keep the device happy.
826 */
827static unsigned int swiotlb_align_offset(struct device *dev, u64 addr)
828{
829 return addr & dma_get_min_align_mask(dev) & (IO_TLB_SIZE - 1);
830}
831
832/*
833 * Bounce: copy the swiotlb buffer from or back to the original dma location
834 */
835static void swiotlb_bounce(struct device *dev, phys_addr_t tlb_addr, size_t size,
836 enum dma_data_direction dir)
837{
838 struct io_tlb_pool *mem = swiotlb_find_pool(dev, tlb_addr);
839 int index = (tlb_addr - mem->start) >> IO_TLB_SHIFT;
840 phys_addr_t orig_addr = mem->slots[index].orig_addr;
841 size_t alloc_size = mem->slots[index].alloc_size;
842 unsigned long pfn = PFN_DOWN(orig_addr);
843 unsigned char *vaddr = mem->vaddr + tlb_addr - mem->start;
844 unsigned int tlb_offset, orig_addr_offset;
845
846 if (orig_addr == INVALID_PHYS_ADDR)
847 return;
848
849 tlb_offset = tlb_addr & (IO_TLB_SIZE - 1);
850 orig_addr_offset = swiotlb_align_offset(dev, orig_addr);
851 if (tlb_offset < orig_addr_offset) {
852 dev_WARN_ONCE(dev, 1,
853 "Access before mapping start detected. orig offset %u, requested offset %u.\n",
854 orig_addr_offset, tlb_offset);
855 return;
856 }
857
858 tlb_offset -= orig_addr_offset;
859 if (tlb_offset > alloc_size) {
860 dev_WARN_ONCE(dev, 1,
861 "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu+%u.\n",
862 alloc_size, size, tlb_offset);
863 return;
864 }
865
866 orig_addr += tlb_offset;
867 alloc_size -= tlb_offset;
868
869 if (size > alloc_size) {
870 dev_WARN_ONCE(dev, 1,
871 "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu.\n",
872 alloc_size, size);
873 size = alloc_size;
874 }
875
876 if (PageHighMem(pfn_to_page(pfn))) {
877 unsigned int offset = orig_addr & ~PAGE_MASK;
878 struct page *page;
879 unsigned int sz = 0;
880 unsigned long flags;
881
882 while (size) {
883 sz = min_t(size_t, PAGE_SIZE - offset, size);
884
885 local_irq_save(flags);
886 page = pfn_to_page(pfn);
887 if (dir == DMA_TO_DEVICE)
888 memcpy_from_page(vaddr, page, offset, sz);
889 else
890 memcpy_to_page(page, offset, vaddr, sz);
891 local_irq_restore(flags);
892
893 size -= sz;
894 pfn++;
895 vaddr += sz;
896 offset = 0;
897 }
898 } else if (dir == DMA_TO_DEVICE) {
899 memcpy(vaddr, phys_to_virt(orig_addr), size);
900 } else {
901 memcpy(phys_to_virt(orig_addr), vaddr, size);
902 }
903}
904
905static inline phys_addr_t slot_addr(phys_addr_t start, phys_addr_t idx)
906{
907 return start + (idx << IO_TLB_SHIFT);
908}
909
910/*
911 * Carefully handle integer overflow which can occur when boundary_mask == ~0UL.
912 */
913static inline unsigned long get_max_slots(unsigned long boundary_mask)
914{
915 return (boundary_mask >> IO_TLB_SHIFT) + 1;
916}
917
918static unsigned int wrap_area_index(struct io_tlb_pool *mem, unsigned int index)
919{
920 if (index >= mem->area_nslabs)
921 return 0;
922 return index;
923}
924
925/*
926 * Track the total used slots with a global atomic value in order to have
927 * correct information to determine the high water mark. The mem_used()
928 * function gives imprecise results because there's no locking across
929 * multiple areas.
930 */
931#ifdef CONFIG_DEBUG_FS
932static void inc_used_and_hiwater(struct io_tlb_mem *mem, unsigned int nslots)
933{
934 unsigned long old_hiwater, new_used;
935
936 new_used = atomic_long_add_return(nslots, &mem->total_used);
937 old_hiwater = atomic_long_read(&mem->used_hiwater);
938 do {
939 if (new_used <= old_hiwater)
940 break;
941 } while (!atomic_long_try_cmpxchg(&mem->used_hiwater,
942 &old_hiwater, new_used));
943}
944
945static void dec_used(struct io_tlb_mem *mem, unsigned int nslots)
946{
947 atomic_long_sub(nslots, &mem->total_used);
948}
949
950#else /* !CONFIG_DEBUG_FS */
951static void inc_used_and_hiwater(struct io_tlb_mem *mem, unsigned int nslots)
952{
953}
954static void dec_used(struct io_tlb_mem *mem, unsigned int nslots)
955{
956}
957#endif /* CONFIG_DEBUG_FS */
958
959/**
960 * swiotlb_search_pool_area() - search one memory area in one pool
961 * @dev: Device which maps the buffer.
962 * @pool: Memory pool to be searched.
963 * @area_index: Index of the IO TLB memory area to be searched.
964 * @orig_addr: Original (non-bounced) IO buffer address.
965 * @alloc_size: Total requested size of the bounce buffer,
966 * including initial alignment padding.
967 * @alloc_align_mask: Required alignment of the allocated buffer.
968 *
969 * Find a suitable sequence of IO TLB entries for the request and allocate
970 * a buffer from the given IO TLB memory area.
971 * This function takes care of locking.
972 *
973 * Return: Index of the first allocated slot, or -1 on error.
974 */
975static int swiotlb_search_pool_area(struct device *dev, struct io_tlb_pool *pool,
976 int area_index, phys_addr_t orig_addr, size_t alloc_size,
977 unsigned int alloc_align_mask)
978{
979 struct io_tlb_area *area = pool->areas + area_index;
980 unsigned long boundary_mask = dma_get_seg_boundary(dev);
981 dma_addr_t tbl_dma_addr =
982 phys_to_dma_unencrypted(dev, pool->start) & boundary_mask;
983 unsigned long max_slots = get_max_slots(boundary_mask);
984 unsigned int iotlb_align_mask =
985 dma_get_min_align_mask(dev) | alloc_align_mask;
986 unsigned int nslots = nr_slots(alloc_size), stride;
987 unsigned int offset = swiotlb_align_offset(dev, orig_addr);
988 unsigned int index, slots_checked, count = 0, i;
989 unsigned long flags;
990 unsigned int slot_base;
991 unsigned int slot_index;
992
993 BUG_ON(!nslots);
994 BUG_ON(area_index >= pool->nareas);
995
996 /*
997 * For allocations of PAGE_SIZE or larger only look for page aligned
998 * allocations.
999 */
1000 if (alloc_size >= PAGE_SIZE)
1001 iotlb_align_mask |= ~PAGE_MASK;
1002 iotlb_align_mask &= ~(IO_TLB_SIZE - 1);
1003
1004 /*
1005 * For mappings with an alignment requirement don't bother looping to
1006 * unaligned slots once we found an aligned one.
1007 */
1008 stride = (iotlb_align_mask >> IO_TLB_SHIFT) + 1;
1009
1010 spin_lock_irqsave(&area->lock, flags);
1011 if (unlikely(nslots > pool->area_nslabs - area->used))
1012 goto not_found;
1013
1014 slot_base = area_index * pool->area_nslabs;
1015 index = area->index;
1016
1017 for (slots_checked = 0; slots_checked < pool->area_nslabs; ) {
1018 slot_index = slot_base + index;
1019
1020 if (orig_addr &&
1021 (slot_addr(tbl_dma_addr, slot_index) &
1022 iotlb_align_mask) != (orig_addr & iotlb_align_mask)) {
1023 index = wrap_area_index(pool, index + 1);
1024 slots_checked++;
1025 continue;
1026 }
1027
1028 if (!iommu_is_span_boundary(slot_index, nslots,
1029 nr_slots(tbl_dma_addr),
1030 max_slots)) {
1031 if (pool->slots[slot_index].list >= nslots)
1032 goto found;
1033 }
1034 index = wrap_area_index(pool, index + stride);
1035 slots_checked += stride;
1036 }
1037
1038not_found:
1039 spin_unlock_irqrestore(&area->lock, flags);
1040 return -1;
1041
1042found:
1043 /*
1044 * If we find a slot that indicates we have 'nslots' number of
1045 * contiguous buffers, we allocate the buffers from that slot onwards
1046 * and set the list of free entries to '0' indicating unavailable.
1047 */
1048 for (i = slot_index; i < slot_index + nslots; i++) {
1049 pool->slots[i].list = 0;
1050 pool->slots[i].alloc_size = alloc_size - (offset +
1051 ((i - slot_index) << IO_TLB_SHIFT));
1052 }
1053 for (i = slot_index - 1;
1054 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 &&
1055 pool->slots[i].list; i--)
1056 pool->slots[i].list = ++count;
1057
1058 /*
1059 * Update the indices to avoid searching in the next round.
1060 */
1061 area->index = wrap_area_index(pool, index + nslots);
1062 area->used += nslots;
1063 spin_unlock_irqrestore(&area->lock, flags);
1064
1065 inc_used_and_hiwater(dev->dma_io_tlb_mem, nslots);
1066 return slot_index;
1067}
1068
1069#ifdef CONFIG_SWIOTLB_DYNAMIC
1070
1071/**
1072 * swiotlb_search_area() - search one memory area in all pools
1073 * @dev: Device which maps the buffer.
1074 * @start_cpu: Start CPU number.
1075 * @cpu_offset: Offset from @start_cpu.
1076 * @orig_addr: Original (non-bounced) IO buffer address.
1077 * @alloc_size: Total requested size of the bounce buffer,
1078 * including initial alignment padding.
1079 * @alloc_align_mask: Required alignment of the allocated buffer.
1080 * @retpool: Used memory pool, updated on return.
1081 *
1082 * Search one memory area in all pools for a sequence of slots that match the
1083 * allocation constraints.
1084 *
1085 * Return: Index of the first allocated slot, or -1 on error.
1086 */
1087static int swiotlb_search_area(struct device *dev, int start_cpu,
1088 int cpu_offset, phys_addr_t orig_addr, size_t alloc_size,
1089 unsigned int alloc_align_mask, struct io_tlb_pool **retpool)
1090{
1091 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1092 struct io_tlb_pool *pool;
1093 int area_index;
1094 int index = -1;
1095
1096 rcu_read_lock();
1097 list_for_each_entry_rcu(pool, &mem->pools, node) {
1098 if (cpu_offset >= pool->nareas)
1099 continue;
1100 area_index = (start_cpu + cpu_offset) & (pool->nareas - 1);
1101 index = swiotlb_search_pool_area(dev, pool, area_index,
1102 orig_addr, alloc_size,
1103 alloc_align_mask);
1104 if (index >= 0) {
1105 *retpool = pool;
1106 break;
1107 }
1108 }
1109 rcu_read_unlock();
1110 return index;
1111}
1112
1113/**
1114 * swiotlb_find_slots() - search for slots in the whole swiotlb
1115 * @dev: Device which maps the buffer.
1116 * @orig_addr: Original (non-bounced) IO buffer address.
1117 * @alloc_size: Total requested size of the bounce buffer,
1118 * including initial alignment padding.
1119 * @alloc_align_mask: Required alignment of the allocated buffer.
1120 * @retpool: Used memory pool, updated on return.
1121 *
1122 * Search through the whole software IO TLB to find a sequence of slots that
1123 * match the allocation constraints.
1124 *
1125 * Return: Index of the first allocated slot, or -1 on error.
1126 */
1127static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr,
1128 size_t alloc_size, unsigned int alloc_align_mask,
1129 struct io_tlb_pool **retpool)
1130{
1131 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1132 struct io_tlb_pool *pool;
1133 unsigned long nslabs;
1134 unsigned long flags;
1135 u64 phys_limit;
1136 int cpu, i;
1137 int index;
1138
1139 if (alloc_size > IO_TLB_SEGSIZE * IO_TLB_SIZE)
1140 return -1;
1141
1142 cpu = raw_smp_processor_id();
1143 for (i = 0; i < default_nareas; ++i) {
1144 index = swiotlb_search_area(dev, cpu, i, orig_addr, alloc_size,
1145 alloc_align_mask, &pool);
1146 if (index >= 0)
1147 goto found;
1148 }
1149
1150 if (!mem->can_grow)
1151 return -1;
1152
1153 schedule_work(&mem->dyn_alloc);
1154
1155 nslabs = nr_slots(alloc_size);
1156 phys_limit = min_not_zero(*dev->dma_mask, dev->bus_dma_limit);
1157 pool = swiotlb_alloc_pool(dev, nslabs, nslabs, 1, phys_limit,
1158 GFP_NOWAIT | __GFP_NOWARN);
1159 if (!pool)
1160 return -1;
1161
1162 index = swiotlb_search_pool_area(dev, pool, 0, orig_addr,
1163 alloc_size, alloc_align_mask);
1164 if (index < 0) {
1165 swiotlb_dyn_free(&pool->rcu);
1166 return -1;
1167 }
1168
1169 pool->transient = true;
1170 spin_lock_irqsave(&dev->dma_io_tlb_lock, flags);
1171 list_add_rcu(&pool->node, &dev->dma_io_tlb_pools);
1172 spin_unlock_irqrestore(&dev->dma_io_tlb_lock, flags);
1173
1174found:
1175 WRITE_ONCE(dev->dma_uses_io_tlb, true);
1176
1177 /*
1178 * The general barrier orders reads and writes against a presumed store
1179 * of the SWIOTLB buffer address by a device driver (to a driver private
1180 * data structure). It serves two purposes.
1181 *
1182 * First, the store to dev->dma_uses_io_tlb must be ordered before the
1183 * presumed store. This guarantees that the returned buffer address
1184 * cannot be passed to another CPU before updating dev->dma_uses_io_tlb.
1185 *
1186 * Second, the load from mem->pools must be ordered before the same
1187 * presumed store. This guarantees that the returned buffer address
1188 * cannot be observed by another CPU before an update of the RCU list
1189 * that was made by swiotlb_dyn_alloc() on a third CPU (cf. multicopy
1190 * atomicity).
1191 *
1192 * See also the comment in is_swiotlb_buffer().
1193 */
1194 smp_mb();
1195
1196 *retpool = pool;
1197 return index;
1198}
1199
1200#else /* !CONFIG_SWIOTLB_DYNAMIC */
1201
1202static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr,
1203 size_t alloc_size, unsigned int alloc_align_mask,
1204 struct io_tlb_pool **retpool)
1205{
1206 struct io_tlb_pool *pool;
1207 int start, i;
1208 int index;
1209
1210 *retpool = pool = &dev->dma_io_tlb_mem->defpool;
1211 i = start = raw_smp_processor_id() & (pool->nareas - 1);
1212 do {
1213 index = swiotlb_search_pool_area(dev, pool, i, orig_addr,
1214 alloc_size, alloc_align_mask);
1215 if (index >= 0)
1216 return index;
1217 if (++i >= pool->nareas)
1218 i = 0;
1219 } while (i != start);
1220 return -1;
1221}
1222
1223#endif /* CONFIG_SWIOTLB_DYNAMIC */
1224
1225#ifdef CONFIG_DEBUG_FS
1226
1227/**
1228 * mem_used() - get number of used slots in an allocator
1229 * @mem: Software IO TLB allocator.
1230 *
1231 * The result is accurate in this version of the function, because an atomic
1232 * counter is available if CONFIG_DEBUG_FS is set.
1233 *
1234 * Return: Number of used slots.
1235 */
1236static unsigned long mem_used(struct io_tlb_mem *mem)
1237{
1238 return atomic_long_read(&mem->total_used);
1239}
1240
1241#else /* !CONFIG_DEBUG_FS */
1242
1243/**
1244 * mem_pool_used() - get number of used slots in a memory pool
1245 * @pool: Software IO TLB memory pool.
1246 *
1247 * The result is not accurate, see mem_used().
1248 *
1249 * Return: Approximate number of used slots.
1250 */
1251static unsigned long mem_pool_used(struct io_tlb_pool *pool)
1252{
1253 int i;
1254 unsigned long used = 0;
1255
1256 for (i = 0; i < pool->nareas; i++)
1257 used += pool->areas[i].used;
1258 return used;
1259}
1260
1261/**
1262 * mem_used() - get number of used slots in an allocator
1263 * @mem: Software IO TLB allocator.
1264 *
1265 * The result is not accurate, because there is no locking of individual
1266 * areas.
1267 *
1268 * Return: Approximate number of used slots.
1269 */
1270static unsigned long mem_used(struct io_tlb_mem *mem)
1271{
1272#ifdef CONFIG_SWIOTLB_DYNAMIC
1273 struct io_tlb_pool *pool;
1274 unsigned long used = 0;
1275
1276 rcu_read_lock();
1277 list_for_each_entry_rcu(pool, &mem->pools, node)
1278 used += mem_pool_used(pool);
1279 rcu_read_unlock();
1280
1281 return used;
1282#else
1283 return mem_pool_used(&mem->defpool);
1284#endif
1285}
1286
1287#endif /* CONFIG_DEBUG_FS */
1288
1289phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr,
1290 size_t mapping_size, size_t alloc_size,
1291 unsigned int alloc_align_mask, enum dma_data_direction dir,
1292 unsigned long attrs)
1293{
1294 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1295 unsigned int offset = swiotlb_align_offset(dev, orig_addr);
1296 struct io_tlb_pool *pool;
1297 unsigned int i;
1298 int index;
1299 phys_addr_t tlb_addr;
1300
1301 if (!mem || !mem->nslabs) {
1302 dev_warn_ratelimited(dev,
1303 "Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer");
1304 return (phys_addr_t)DMA_MAPPING_ERROR;
1305 }
1306
1307 if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
1308 pr_warn_once("Memory encryption is active and system is using DMA bounce buffers\n");
1309
1310 if (mapping_size > alloc_size) {
1311 dev_warn_once(dev, "Invalid sizes (mapping: %zd bytes, alloc: %zd bytes)",
1312 mapping_size, alloc_size);
1313 return (phys_addr_t)DMA_MAPPING_ERROR;
1314 }
1315
1316 index = swiotlb_find_slots(dev, orig_addr,
1317 alloc_size + offset, alloc_align_mask, &pool);
1318 if (index == -1) {
1319 if (!(attrs & DMA_ATTR_NO_WARN))
1320 dev_warn_ratelimited(dev,
1321 "swiotlb buffer is full (sz: %zd bytes), total %lu (slots), used %lu (slots)\n",
1322 alloc_size, mem->nslabs, mem_used(mem));
1323 return (phys_addr_t)DMA_MAPPING_ERROR;
1324 }
1325
1326 /*
1327 * Save away the mapping from the original address to the DMA address.
1328 * This is needed when we sync the memory. Then we sync the buffer if
1329 * needed.
1330 */
1331 for (i = 0; i < nr_slots(alloc_size + offset); i++)
1332 pool->slots[index + i].orig_addr = slot_addr(orig_addr, i);
1333 tlb_addr = slot_addr(pool->start, index) + offset;
1334 /*
1335 * When the device is writing memory, i.e. dir == DMA_FROM_DEVICE, copy
1336 * the original buffer to the TLB buffer before initiating DMA in order
1337 * to preserve the original's data if the device does a partial write,
1338 * i.e. if the device doesn't overwrite the entire buffer. Preserving
1339 * the original data, even if it's garbage, is necessary to match
1340 * hardware behavior. Use of swiotlb is supposed to be transparent,
1341 * i.e. swiotlb must not corrupt memory by clobbering unwritten bytes.
1342 */
1343 swiotlb_bounce(dev, tlb_addr, mapping_size, DMA_TO_DEVICE);
1344 return tlb_addr;
1345}
1346
1347static void swiotlb_release_slots(struct device *dev, phys_addr_t tlb_addr)
1348{
1349 struct io_tlb_pool *mem = swiotlb_find_pool(dev, tlb_addr);
1350 unsigned long flags;
1351 unsigned int offset = swiotlb_align_offset(dev, tlb_addr);
1352 int index = (tlb_addr - offset - mem->start) >> IO_TLB_SHIFT;
1353 int nslots = nr_slots(mem->slots[index].alloc_size + offset);
1354 int aindex = index / mem->area_nslabs;
1355 struct io_tlb_area *area = &mem->areas[aindex];
1356 int count, i;
1357
1358 /*
1359 * Return the buffer to the free list by setting the corresponding
1360 * entries to indicate the number of contiguous entries available.
1361 * While returning the entries to the free list, we merge the entries
1362 * with slots below and above the pool being returned.
1363 */
1364 BUG_ON(aindex >= mem->nareas);
1365
1366 spin_lock_irqsave(&area->lock, flags);
1367 if (index + nslots < ALIGN(index + 1, IO_TLB_SEGSIZE))
1368 count = mem->slots[index + nslots].list;
1369 else
1370 count = 0;
1371
1372 /*
1373 * Step 1: return the slots to the free list, merging the slots with
1374 * superceeding slots
1375 */
1376 for (i = index + nslots - 1; i >= index; i--) {
1377 mem->slots[i].list = ++count;
1378 mem->slots[i].orig_addr = INVALID_PHYS_ADDR;
1379 mem->slots[i].alloc_size = 0;
1380 }
1381
1382 /*
1383 * Step 2: merge the returned slots with the preceding slots, if
1384 * available (non zero)
1385 */
1386 for (i = index - 1;
1387 io_tlb_offset(i) != IO_TLB_SEGSIZE - 1 && mem->slots[i].list;
1388 i--)
1389 mem->slots[i].list = ++count;
1390 area->used -= nslots;
1391 spin_unlock_irqrestore(&area->lock, flags);
1392
1393 dec_used(dev->dma_io_tlb_mem, nslots);
1394}
1395
1396#ifdef CONFIG_SWIOTLB_DYNAMIC
1397
1398/**
1399 * swiotlb_del_transient() - delete a transient memory pool
1400 * @dev: Device which mapped the buffer.
1401 * @tlb_addr: Physical address within a bounce buffer.
1402 *
1403 * Check whether the address belongs to a transient SWIOTLB memory pool.
1404 * If yes, then delete the pool.
1405 *
1406 * Return: %true if @tlb_addr belonged to a transient pool that was released.
1407 */
1408static bool swiotlb_del_transient(struct device *dev, phys_addr_t tlb_addr)
1409{
1410 struct io_tlb_pool *pool;
1411
1412 pool = swiotlb_find_pool(dev, tlb_addr);
1413 if (!pool->transient)
1414 return false;
1415
1416 dec_used(dev->dma_io_tlb_mem, pool->nslabs);
1417 swiotlb_del_pool(dev, pool);
1418 return true;
1419}
1420
1421#else /* !CONFIG_SWIOTLB_DYNAMIC */
1422
1423static inline bool swiotlb_del_transient(struct device *dev,
1424 phys_addr_t tlb_addr)
1425{
1426 return false;
1427}
1428
1429#endif /* CONFIG_SWIOTLB_DYNAMIC */
1430
1431/*
1432 * tlb_addr is the physical address of the bounce buffer to unmap.
1433 */
1434void swiotlb_tbl_unmap_single(struct device *dev, phys_addr_t tlb_addr,
1435 size_t mapping_size, enum dma_data_direction dir,
1436 unsigned long attrs)
1437{
1438 /*
1439 * First, sync the memory before unmapping the entry
1440 */
1441 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
1442 (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL))
1443 swiotlb_bounce(dev, tlb_addr, mapping_size, DMA_FROM_DEVICE);
1444
1445 if (swiotlb_del_transient(dev, tlb_addr))
1446 return;
1447 swiotlb_release_slots(dev, tlb_addr);
1448}
1449
1450void swiotlb_sync_single_for_device(struct device *dev, phys_addr_t tlb_addr,
1451 size_t size, enum dma_data_direction dir)
1452{
1453 if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)
1454 swiotlb_bounce(dev, tlb_addr, size, DMA_TO_DEVICE);
1455 else
1456 BUG_ON(dir != DMA_FROM_DEVICE);
1457}
1458
1459void swiotlb_sync_single_for_cpu(struct device *dev, phys_addr_t tlb_addr,
1460 size_t size, enum dma_data_direction dir)
1461{
1462 if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
1463 swiotlb_bounce(dev, tlb_addr, size, DMA_FROM_DEVICE);
1464 else
1465 BUG_ON(dir != DMA_TO_DEVICE);
1466}
1467
1468/*
1469 * Create a swiotlb mapping for the buffer at @paddr, and in case of DMAing
1470 * to the device copy the data into it as well.
1471 */
1472dma_addr_t swiotlb_map(struct device *dev, phys_addr_t paddr, size_t size,
1473 enum dma_data_direction dir, unsigned long attrs)
1474{
1475 phys_addr_t swiotlb_addr;
1476 dma_addr_t dma_addr;
1477
1478 trace_swiotlb_bounced(dev, phys_to_dma(dev, paddr), size);
1479
1480 swiotlb_addr = swiotlb_tbl_map_single(dev, paddr, size, size, 0, dir,
1481 attrs);
1482 if (swiotlb_addr == (phys_addr_t)DMA_MAPPING_ERROR)
1483 return DMA_MAPPING_ERROR;
1484
1485 /* Ensure that the address returned is DMA'ble */
1486 dma_addr = phys_to_dma_unencrypted(dev, swiotlb_addr);
1487 if (unlikely(!dma_capable(dev, dma_addr, size, true))) {
1488 swiotlb_tbl_unmap_single(dev, swiotlb_addr, size, dir,
1489 attrs | DMA_ATTR_SKIP_CPU_SYNC);
1490 dev_WARN_ONCE(dev, 1,
1491 "swiotlb addr %pad+%zu overflow (mask %llx, bus limit %llx).\n",
1492 &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit);
1493 return DMA_MAPPING_ERROR;
1494 }
1495
1496 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1497 arch_sync_dma_for_device(swiotlb_addr, size, dir);
1498 return dma_addr;
1499}
1500
1501size_t swiotlb_max_mapping_size(struct device *dev)
1502{
1503 int min_align_mask = dma_get_min_align_mask(dev);
1504 int min_align = 0;
1505
1506 /*
1507 * swiotlb_find_slots() skips slots according to
1508 * min align mask. This affects max mapping size.
1509 * Take it into acount here.
1510 */
1511 if (min_align_mask)
1512 min_align = roundup(min_align_mask, IO_TLB_SIZE);
1513
1514 return ((size_t)IO_TLB_SIZE) * IO_TLB_SEGSIZE - min_align;
1515}
1516
1517/**
1518 * is_swiotlb_allocated() - check if the default software IO TLB is initialized
1519 */
1520bool is_swiotlb_allocated(void)
1521{
1522 return io_tlb_default_mem.nslabs;
1523}
1524
1525bool is_swiotlb_active(struct device *dev)
1526{
1527 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1528
1529 return mem && mem->nslabs;
1530}
1531
1532/**
1533 * default_swiotlb_base() - get the base address of the default SWIOTLB
1534 *
1535 * Get the lowest physical address used by the default software IO TLB pool.
1536 */
1537phys_addr_t default_swiotlb_base(void)
1538{
1539#ifdef CONFIG_SWIOTLB_DYNAMIC
1540 io_tlb_default_mem.can_grow = false;
1541#endif
1542 return io_tlb_default_mem.defpool.start;
1543}
1544
1545/**
1546 * default_swiotlb_limit() - get the address limit of the default SWIOTLB
1547 *
1548 * Get the highest physical address used by the default software IO TLB pool.
1549 */
1550phys_addr_t default_swiotlb_limit(void)
1551{
1552#ifdef CONFIG_SWIOTLB_DYNAMIC
1553 return io_tlb_default_mem.phys_limit;
1554#else
1555 return io_tlb_default_mem.defpool.end - 1;
1556#endif
1557}
1558
1559#ifdef CONFIG_DEBUG_FS
1560
1561static int io_tlb_used_get(void *data, u64 *val)
1562{
1563 struct io_tlb_mem *mem = data;
1564
1565 *val = mem_used(mem);
1566 return 0;
1567}
1568
1569static int io_tlb_hiwater_get(void *data, u64 *val)
1570{
1571 struct io_tlb_mem *mem = data;
1572
1573 *val = atomic_long_read(&mem->used_hiwater);
1574 return 0;
1575}
1576
1577static int io_tlb_hiwater_set(void *data, u64 val)
1578{
1579 struct io_tlb_mem *mem = data;
1580
1581 /* Only allow setting to zero */
1582 if (val != 0)
1583 return -EINVAL;
1584
1585 atomic_long_set(&mem->used_hiwater, val);
1586 return 0;
1587}
1588
1589DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_used, io_tlb_used_get, NULL, "%llu\n");
1590DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_hiwater, io_tlb_hiwater_get,
1591 io_tlb_hiwater_set, "%llu\n");
1592
1593static void swiotlb_create_debugfs_files(struct io_tlb_mem *mem,
1594 const char *dirname)
1595{
1596 atomic_long_set(&mem->total_used, 0);
1597 atomic_long_set(&mem->used_hiwater, 0);
1598
1599 mem->debugfs = debugfs_create_dir(dirname, io_tlb_default_mem.debugfs);
1600 if (!mem->nslabs)
1601 return;
1602
1603 debugfs_create_ulong("io_tlb_nslabs", 0400, mem->debugfs, &mem->nslabs);
1604 debugfs_create_file("io_tlb_used", 0400, mem->debugfs, mem,
1605 &fops_io_tlb_used);
1606 debugfs_create_file("io_tlb_used_hiwater", 0600, mem->debugfs, mem,
1607 &fops_io_tlb_hiwater);
1608}
1609
1610static int __init swiotlb_create_default_debugfs(void)
1611{
1612 swiotlb_create_debugfs_files(&io_tlb_default_mem, "swiotlb");
1613 return 0;
1614}
1615
1616late_initcall(swiotlb_create_default_debugfs);
1617
1618#else /* !CONFIG_DEBUG_FS */
1619
1620static inline void swiotlb_create_debugfs_files(struct io_tlb_mem *mem,
1621 const char *dirname)
1622{
1623}
1624
1625#endif /* CONFIG_DEBUG_FS */
1626
1627#ifdef CONFIG_DMA_RESTRICTED_POOL
1628
1629struct page *swiotlb_alloc(struct device *dev, size_t size)
1630{
1631 struct io_tlb_mem *mem = dev->dma_io_tlb_mem;
1632 struct io_tlb_pool *pool;
1633 phys_addr_t tlb_addr;
1634 int index;
1635
1636 if (!mem)
1637 return NULL;
1638
1639 index = swiotlb_find_slots(dev, 0, size, 0, &pool);
1640 if (index == -1)
1641 return NULL;
1642
1643 tlb_addr = slot_addr(pool->start, index);
1644
1645 return pfn_to_page(PFN_DOWN(tlb_addr));
1646}
1647
1648bool swiotlb_free(struct device *dev, struct page *page, size_t size)
1649{
1650 phys_addr_t tlb_addr = page_to_phys(page);
1651
1652 if (!is_swiotlb_buffer(dev, tlb_addr))
1653 return false;
1654
1655 swiotlb_release_slots(dev, tlb_addr);
1656
1657 return true;
1658}
1659
1660static int rmem_swiotlb_device_init(struct reserved_mem *rmem,
1661 struct device *dev)
1662{
1663 struct io_tlb_mem *mem = rmem->priv;
1664 unsigned long nslabs = rmem->size >> IO_TLB_SHIFT;
1665
1666 /* Set Per-device io tlb area to one */
1667 unsigned int nareas = 1;
1668
1669 if (PageHighMem(pfn_to_page(PHYS_PFN(rmem->base)))) {
1670 dev_err(dev, "Restricted DMA pool must be accessible within the linear mapping.");
1671 return -EINVAL;
1672 }
1673
1674 /*
1675 * Since multiple devices can share the same pool, the private data,
1676 * io_tlb_mem struct, will be initialized by the first device attached
1677 * to it.
1678 */
1679 if (!mem) {
1680 struct io_tlb_pool *pool;
1681
1682 mem = kzalloc(sizeof(*mem), GFP_KERNEL);
1683 if (!mem)
1684 return -ENOMEM;
1685 pool = &mem->defpool;
1686
1687 pool->slots = kcalloc(nslabs, sizeof(*pool->slots), GFP_KERNEL);
1688 if (!pool->slots) {
1689 kfree(mem);
1690 return -ENOMEM;
1691 }
1692
1693 pool->areas = kcalloc(nareas, sizeof(*pool->areas),
1694 GFP_KERNEL);
1695 if (!pool->areas) {
1696 kfree(pool->slots);
1697 kfree(mem);
1698 return -ENOMEM;
1699 }
1700
1701 set_memory_decrypted((unsigned long)phys_to_virt(rmem->base),
1702 rmem->size >> PAGE_SHIFT);
1703 swiotlb_init_io_tlb_pool(pool, rmem->base, nslabs,
1704 false, nareas);
1705 mem->force_bounce = true;
1706 mem->for_alloc = true;
1707#ifdef CONFIG_SWIOTLB_DYNAMIC
1708 spin_lock_init(&mem->lock);
1709#endif
1710 add_mem_pool(mem, pool);
1711
1712 rmem->priv = mem;
1713
1714 swiotlb_create_debugfs_files(mem, rmem->name);
1715 }
1716
1717 dev->dma_io_tlb_mem = mem;
1718
1719 return 0;
1720}
1721
1722static void rmem_swiotlb_device_release(struct reserved_mem *rmem,
1723 struct device *dev)
1724{
1725 dev->dma_io_tlb_mem = &io_tlb_default_mem;
1726}
1727
1728static const struct reserved_mem_ops rmem_swiotlb_ops = {
1729 .device_init = rmem_swiotlb_device_init,
1730 .device_release = rmem_swiotlb_device_release,
1731};
1732
1733static int __init rmem_swiotlb_setup(struct reserved_mem *rmem)
1734{
1735 unsigned long node = rmem->fdt_node;
1736
1737 if (of_get_flat_dt_prop(node, "reusable", NULL) ||
1738 of_get_flat_dt_prop(node, "linux,cma-default", NULL) ||
1739 of_get_flat_dt_prop(node, "linux,dma-default", NULL) ||
1740 of_get_flat_dt_prop(node, "no-map", NULL))
1741 return -EINVAL;
1742
1743 rmem->ops = &rmem_swiotlb_ops;
1744 pr_info("Reserved memory: created restricted DMA pool at %pa, size %ld MiB\n",
1745 &rmem->base, (unsigned long)rmem->size / SZ_1M);
1746 return 0;
1747}
1748
1749RESERVEDMEM_OF_DECLARE(dma, "restricted-dma-pool", rmem_swiotlb_setup);
1750#endif /* CONFIG_DMA_RESTRICTED_POOL */