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 | * @pad_slots: Number of preceding padding slots. Valid only in the first |
73 | * allocated non-padding slot. |
74 | */ |
75 | struct io_tlb_slot { |
76 | phys_addr_t orig_addr; |
77 | size_t alloc_size; |
78 | unsigned short list; |
79 | unsigned short pad_slots; |
80 | }; |
81 | |
82 | static bool swiotlb_force_bounce; |
83 | static bool swiotlb_force_disable; |
84 | |
85 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
86 | |
87 | static void swiotlb_dyn_alloc(struct work_struct *work); |
88 | |
89 | static struct io_tlb_mem io_tlb_default_mem = { |
90 | .lock = __SPIN_LOCK_UNLOCKED(io_tlb_default_mem.lock), |
91 | .pools = LIST_HEAD_INIT(io_tlb_default_mem.pools), |
92 | .dyn_alloc = __WORK_INITIALIZER(io_tlb_default_mem.dyn_alloc, |
93 | swiotlb_dyn_alloc), |
94 | }; |
95 | |
96 | #else /* !CONFIG_SWIOTLB_DYNAMIC */ |
97 | |
98 | static struct io_tlb_mem io_tlb_default_mem; |
99 | |
100 | #endif /* CONFIG_SWIOTLB_DYNAMIC */ |
101 | |
102 | static unsigned long default_nslabs = IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT; |
103 | static unsigned long default_nareas; |
104 | |
105 | /** |
106 | * struct io_tlb_area - IO TLB memory area descriptor |
107 | * |
108 | * This is a single area with a single lock. |
109 | * |
110 | * @used: The number of used IO TLB block. |
111 | * @index: The slot index to start searching in this area for next round. |
112 | * @lock: The lock to protect the above data structures in the map and |
113 | * unmap calls. |
114 | */ |
115 | struct io_tlb_area { |
116 | unsigned long used; |
117 | unsigned int index; |
118 | spinlock_t lock; |
119 | }; |
120 | |
121 | /* |
122 | * Round up number of slabs to the next power of 2. The last area is going |
123 | * be smaller than the rest if default_nslabs is not power of two. |
124 | * The number of slot in an area should be a multiple of IO_TLB_SEGSIZE, |
125 | * otherwise a segment may span two or more areas. It conflicts with free |
126 | * contiguous slots tracking: free slots are treated contiguous no matter |
127 | * whether they cross an area boundary. |
128 | * |
129 | * Return true if default_nslabs is rounded up. |
130 | */ |
131 | static bool round_up_default_nslabs(void) |
132 | { |
133 | if (!default_nareas) |
134 | return false; |
135 | |
136 | if (default_nslabs < IO_TLB_SEGSIZE * default_nareas) |
137 | default_nslabs = IO_TLB_SEGSIZE * default_nareas; |
138 | else if (is_power_of_2(n: default_nslabs)) |
139 | return false; |
140 | default_nslabs = roundup_pow_of_two(default_nslabs); |
141 | return true; |
142 | } |
143 | |
144 | /** |
145 | * swiotlb_adjust_nareas() - adjust the number of areas and slots |
146 | * @nareas: Desired number of areas. Zero is treated as 1. |
147 | * |
148 | * Adjust the default number of areas in a memory pool. |
149 | * The default size of the memory pool may also change to meet minimum area |
150 | * size requirements. |
151 | */ |
152 | static void swiotlb_adjust_nareas(unsigned int nareas) |
153 | { |
154 | if (!nareas) |
155 | nareas = 1; |
156 | else if (!is_power_of_2(n: nareas)) |
157 | nareas = roundup_pow_of_two(nareas); |
158 | |
159 | default_nareas = nareas; |
160 | |
161 | pr_info("area num %d.\n" , nareas); |
162 | if (round_up_default_nslabs()) |
163 | pr_info("SWIOTLB bounce buffer size roundup to %luMB" , |
164 | (default_nslabs << IO_TLB_SHIFT) >> 20); |
165 | } |
166 | |
167 | /** |
168 | * limit_nareas() - get the maximum number of areas for a given memory pool size |
169 | * @nareas: Desired number of areas. |
170 | * @nslots: Total number of slots in the memory pool. |
171 | * |
172 | * Limit the number of areas to the maximum possible number of areas in |
173 | * a memory pool of the given size. |
174 | * |
175 | * Return: Maximum possible number of areas. |
176 | */ |
177 | static unsigned int limit_nareas(unsigned int nareas, unsigned long nslots) |
178 | { |
179 | if (nslots < nareas * IO_TLB_SEGSIZE) |
180 | return nslots / IO_TLB_SEGSIZE; |
181 | return nareas; |
182 | } |
183 | |
184 | static int __init |
185 | setup_io_tlb_npages(char *str) |
186 | { |
187 | if (isdigit(c: *str)) { |
188 | /* avoid tail segment of size < IO_TLB_SEGSIZE */ |
189 | default_nslabs = |
190 | ALIGN(simple_strtoul(str, &str, 0), IO_TLB_SEGSIZE); |
191 | } |
192 | if (*str == ',') |
193 | ++str; |
194 | if (isdigit(c: *str)) |
195 | swiotlb_adjust_nareas(nareas: simple_strtoul(str, &str, 0)); |
196 | if (*str == ',') |
197 | ++str; |
198 | if (!strcmp(str, "force" )) |
199 | swiotlb_force_bounce = true; |
200 | else if (!strcmp(str, "noforce" )) |
201 | swiotlb_force_disable = true; |
202 | |
203 | return 0; |
204 | } |
205 | early_param("swiotlb" , setup_io_tlb_npages); |
206 | |
207 | unsigned long swiotlb_size_or_default(void) |
208 | { |
209 | return default_nslabs << IO_TLB_SHIFT; |
210 | } |
211 | |
212 | void __init swiotlb_adjust_size(unsigned long size) |
213 | { |
214 | /* |
215 | * If swiotlb parameter has not been specified, give a chance to |
216 | * architectures such as those supporting memory encryption to |
217 | * adjust/expand SWIOTLB size for their use. |
218 | */ |
219 | if (default_nslabs != IO_TLB_DEFAULT_SIZE >> IO_TLB_SHIFT) |
220 | return; |
221 | |
222 | size = ALIGN(size, IO_TLB_SIZE); |
223 | default_nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE); |
224 | if (round_up_default_nslabs()) |
225 | size = default_nslabs << IO_TLB_SHIFT; |
226 | pr_info("SWIOTLB bounce buffer size adjusted to %luMB" , size >> 20); |
227 | } |
228 | |
229 | void swiotlb_print_info(void) |
230 | { |
231 | struct io_tlb_pool *mem = &io_tlb_default_mem.defpool; |
232 | |
233 | if (!mem->nslabs) { |
234 | pr_warn("No low mem\n" ); |
235 | return; |
236 | } |
237 | |
238 | pr_info("mapped [mem %pa-%pa] (%luMB)\n" , &mem->start, &mem->end, |
239 | (mem->nslabs << IO_TLB_SHIFT) >> 20); |
240 | } |
241 | |
242 | static inline unsigned long io_tlb_offset(unsigned long val) |
243 | { |
244 | return val & (IO_TLB_SEGSIZE - 1); |
245 | } |
246 | |
247 | static inline unsigned long nr_slots(u64 val) |
248 | { |
249 | return DIV_ROUND_UP(val, IO_TLB_SIZE); |
250 | } |
251 | |
252 | /* |
253 | * Early SWIOTLB allocation may be too early to allow an architecture to |
254 | * perform the desired operations. This function allows the architecture to |
255 | * call SWIOTLB when the operations are possible. It needs to be called |
256 | * before the SWIOTLB memory is used. |
257 | */ |
258 | void __init swiotlb_update_mem_attributes(void) |
259 | { |
260 | struct io_tlb_pool *mem = &io_tlb_default_mem.defpool; |
261 | unsigned long bytes; |
262 | |
263 | if (!mem->nslabs || mem->late_alloc) |
264 | return; |
265 | bytes = PAGE_ALIGN(mem->nslabs << IO_TLB_SHIFT); |
266 | set_memory_decrypted(addr: (unsigned long)mem->vaddr, numpages: bytes >> PAGE_SHIFT); |
267 | } |
268 | |
269 | static void swiotlb_init_io_tlb_pool(struct io_tlb_pool *mem, phys_addr_t start, |
270 | unsigned long nslabs, bool late_alloc, unsigned int nareas) |
271 | { |
272 | void *vaddr = phys_to_virt(address: start); |
273 | unsigned long bytes = nslabs << IO_TLB_SHIFT, i; |
274 | |
275 | mem->nslabs = nslabs; |
276 | mem->start = start; |
277 | mem->end = mem->start + bytes; |
278 | mem->late_alloc = late_alloc; |
279 | mem->nareas = nareas; |
280 | mem->area_nslabs = nslabs / mem->nareas; |
281 | |
282 | for (i = 0; i < mem->nareas; i++) { |
283 | spin_lock_init(&mem->areas[i].lock); |
284 | mem->areas[i].index = 0; |
285 | mem->areas[i].used = 0; |
286 | } |
287 | |
288 | for (i = 0; i < mem->nslabs; i++) { |
289 | mem->slots[i].list = min(IO_TLB_SEGSIZE - io_tlb_offset(i), |
290 | mem->nslabs - i); |
291 | mem->slots[i].orig_addr = INVALID_PHYS_ADDR; |
292 | mem->slots[i].alloc_size = 0; |
293 | mem->slots[i].pad_slots = 0; |
294 | } |
295 | |
296 | memset(vaddr, 0, bytes); |
297 | mem->vaddr = vaddr; |
298 | return; |
299 | } |
300 | |
301 | /** |
302 | * add_mem_pool() - add a memory pool to the allocator |
303 | * @mem: Software IO TLB allocator. |
304 | * @pool: Memory pool to be added. |
305 | */ |
306 | static void add_mem_pool(struct io_tlb_mem *mem, struct io_tlb_pool *pool) |
307 | { |
308 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
309 | spin_lock(lock: &mem->lock); |
310 | list_add_rcu(new: &pool->node, head: &mem->pools); |
311 | mem->nslabs += pool->nslabs; |
312 | spin_unlock(lock: &mem->lock); |
313 | #else |
314 | mem->nslabs = pool->nslabs; |
315 | #endif |
316 | } |
317 | |
318 | static void __init *swiotlb_memblock_alloc(unsigned long nslabs, |
319 | unsigned int flags, |
320 | int (*remap)(void *tlb, unsigned long nslabs)) |
321 | { |
322 | size_t bytes = PAGE_ALIGN(nslabs << IO_TLB_SHIFT); |
323 | void *tlb; |
324 | |
325 | /* |
326 | * By default allocate the bounce buffer memory from low memory, but |
327 | * allow to pick a location everywhere for hypervisors with guest |
328 | * memory encryption. |
329 | */ |
330 | if (flags & SWIOTLB_ANY) |
331 | tlb = memblock_alloc(size: bytes, PAGE_SIZE); |
332 | else |
333 | tlb = memblock_alloc_low(size: bytes, PAGE_SIZE); |
334 | |
335 | if (!tlb) { |
336 | pr_warn("%s: Failed to allocate %zu bytes tlb structure\n" , |
337 | __func__, bytes); |
338 | return NULL; |
339 | } |
340 | |
341 | if (remap && remap(tlb, nslabs) < 0) { |
342 | memblock_free(ptr: tlb, PAGE_ALIGN(bytes)); |
343 | pr_warn("%s: Failed to remap %zu bytes\n" , __func__, bytes); |
344 | return NULL; |
345 | } |
346 | |
347 | return tlb; |
348 | } |
349 | |
350 | /* |
351 | * Statically reserve bounce buffer space and initialize bounce buffer data |
352 | * structures for the software IO TLB used to implement the DMA API. |
353 | */ |
354 | void __init swiotlb_init_remap(bool addressing_limit, unsigned int flags, |
355 | int (*remap)(void *tlb, unsigned long nslabs)) |
356 | { |
357 | struct io_tlb_pool *mem = &io_tlb_default_mem.defpool; |
358 | unsigned long nslabs; |
359 | unsigned int nareas; |
360 | size_t alloc_size; |
361 | void *tlb; |
362 | |
363 | if (!addressing_limit && !swiotlb_force_bounce) |
364 | return; |
365 | if (swiotlb_force_disable) |
366 | return; |
367 | |
368 | io_tlb_default_mem.force_bounce = |
369 | swiotlb_force_bounce || (flags & SWIOTLB_FORCE); |
370 | |
371 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
372 | if (!remap) |
373 | io_tlb_default_mem.can_grow = true; |
374 | if (flags & SWIOTLB_ANY) |
375 | io_tlb_default_mem.phys_limit = virt_to_phys(address: high_memory - 1); |
376 | else |
377 | io_tlb_default_mem.phys_limit = ARCH_LOW_ADDRESS_LIMIT; |
378 | #endif |
379 | |
380 | if (!default_nareas) |
381 | swiotlb_adjust_nareas(num_possible_cpus()); |
382 | |
383 | nslabs = default_nslabs; |
384 | nareas = limit_nareas(nareas: default_nareas, nslots: nslabs); |
385 | while ((tlb = swiotlb_memblock_alloc(nslabs, flags, remap)) == NULL) { |
386 | if (nslabs <= IO_TLB_MIN_SLABS) |
387 | return; |
388 | nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE); |
389 | nareas = limit_nareas(nareas, nslots: nslabs); |
390 | } |
391 | |
392 | if (default_nslabs != nslabs) { |
393 | pr_info("SWIOTLB bounce buffer size adjusted %lu -> %lu slabs" , |
394 | default_nslabs, nslabs); |
395 | default_nslabs = nslabs; |
396 | } |
397 | |
398 | alloc_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), nslabs)); |
399 | mem->slots = memblock_alloc(size: alloc_size, PAGE_SIZE); |
400 | if (!mem->slots) { |
401 | pr_warn("%s: Failed to allocate %zu bytes align=0x%lx\n" , |
402 | __func__, alloc_size, PAGE_SIZE); |
403 | return; |
404 | } |
405 | |
406 | mem->areas = memblock_alloc(array_size(sizeof(struct io_tlb_area), |
407 | nareas), SMP_CACHE_BYTES); |
408 | if (!mem->areas) { |
409 | pr_warn("%s: Failed to allocate mem->areas.\n" , __func__); |
410 | return; |
411 | } |
412 | |
413 | swiotlb_init_io_tlb_pool(mem, __pa(tlb), nslabs, late_alloc: false, nareas); |
414 | add_mem_pool(mem: &io_tlb_default_mem, pool: mem); |
415 | |
416 | if (flags & SWIOTLB_VERBOSE) |
417 | swiotlb_print_info(); |
418 | } |
419 | |
420 | void __init swiotlb_init(bool addressing_limit, unsigned int flags) |
421 | { |
422 | swiotlb_init_remap(addressing_limit, flags, NULL); |
423 | } |
424 | |
425 | /* |
426 | * Systems with larger DMA zones (those that don't support ISA) can |
427 | * initialize the swiotlb later using the slab allocator if needed. |
428 | * This should be just like above, but with some error catching. |
429 | */ |
430 | int swiotlb_init_late(size_t size, gfp_t gfp_mask, |
431 | int (*remap)(void *tlb, unsigned long nslabs)) |
432 | { |
433 | struct io_tlb_pool *mem = &io_tlb_default_mem.defpool; |
434 | unsigned long nslabs = ALIGN(size >> IO_TLB_SHIFT, IO_TLB_SEGSIZE); |
435 | unsigned int nareas; |
436 | unsigned char *vstart = NULL; |
437 | unsigned int order, area_order; |
438 | bool retried = false; |
439 | int rc = 0; |
440 | |
441 | if (io_tlb_default_mem.nslabs) |
442 | return 0; |
443 | |
444 | if (swiotlb_force_disable) |
445 | return 0; |
446 | |
447 | io_tlb_default_mem.force_bounce = swiotlb_force_bounce; |
448 | |
449 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
450 | if (!remap) |
451 | io_tlb_default_mem.can_grow = true; |
452 | if (IS_ENABLED(CONFIG_ZONE_DMA) && (gfp_mask & __GFP_DMA)) |
453 | io_tlb_default_mem.phys_limit = DMA_BIT_MASK(zone_dma_bits); |
454 | else if (IS_ENABLED(CONFIG_ZONE_DMA32) && (gfp_mask & __GFP_DMA32)) |
455 | io_tlb_default_mem.phys_limit = DMA_BIT_MASK(32); |
456 | else |
457 | io_tlb_default_mem.phys_limit = virt_to_phys(address: high_memory - 1); |
458 | #endif |
459 | |
460 | if (!default_nareas) |
461 | swiotlb_adjust_nareas(num_possible_cpus()); |
462 | |
463 | retry: |
464 | order = get_order(size: nslabs << IO_TLB_SHIFT); |
465 | nslabs = SLABS_PER_PAGE << order; |
466 | |
467 | while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) { |
468 | vstart = (void *)__get_free_pages(gfp_mask: gfp_mask | __GFP_NOWARN, |
469 | order); |
470 | if (vstart) |
471 | break; |
472 | order--; |
473 | nslabs = SLABS_PER_PAGE << order; |
474 | retried = true; |
475 | } |
476 | |
477 | if (!vstart) |
478 | return -ENOMEM; |
479 | |
480 | if (remap) |
481 | rc = remap(vstart, nslabs); |
482 | if (rc) { |
483 | free_pages(addr: (unsigned long)vstart, order); |
484 | |
485 | nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE); |
486 | if (nslabs < IO_TLB_MIN_SLABS) |
487 | return rc; |
488 | retried = true; |
489 | goto retry; |
490 | } |
491 | |
492 | if (retried) { |
493 | pr_warn("only able to allocate %ld MB\n" , |
494 | (PAGE_SIZE << order) >> 20); |
495 | } |
496 | |
497 | nareas = limit_nareas(nareas: default_nareas, nslots: nslabs); |
498 | area_order = get_order(array_size(sizeof(*mem->areas), nareas)); |
499 | mem->areas = (struct io_tlb_area *) |
500 | __get_free_pages(GFP_KERNEL | __GFP_ZERO, order: area_order); |
501 | if (!mem->areas) |
502 | goto error_area; |
503 | |
504 | mem->slots = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, |
505 | order: get_order(array_size(sizeof(*mem->slots), nslabs))); |
506 | if (!mem->slots) |
507 | goto error_slots; |
508 | |
509 | set_memory_decrypted(addr: (unsigned long)vstart, |
510 | numpages: (nslabs << IO_TLB_SHIFT) >> PAGE_SHIFT); |
511 | swiotlb_init_io_tlb_pool(mem, virt_to_phys(address: vstart), nslabs, late_alloc: true, |
512 | nareas); |
513 | add_mem_pool(mem: &io_tlb_default_mem, pool: mem); |
514 | |
515 | swiotlb_print_info(); |
516 | return 0; |
517 | |
518 | error_slots: |
519 | free_pages(addr: (unsigned long)mem->areas, order: area_order); |
520 | error_area: |
521 | free_pages(addr: (unsigned long)vstart, order); |
522 | return -ENOMEM; |
523 | } |
524 | |
525 | void __init swiotlb_exit(void) |
526 | { |
527 | struct io_tlb_pool *mem = &io_tlb_default_mem.defpool; |
528 | unsigned long tbl_vaddr; |
529 | size_t tbl_size, slots_size; |
530 | unsigned int area_order; |
531 | |
532 | if (swiotlb_force_bounce) |
533 | return; |
534 | |
535 | if (!mem->nslabs) |
536 | return; |
537 | |
538 | pr_info("tearing down default memory pool\n" ); |
539 | tbl_vaddr = (unsigned long)phys_to_virt(address: mem->start); |
540 | tbl_size = PAGE_ALIGN(mem->end - mem->start); |
541 | slots_size = PAGE_ALIGN(array_size(sizeof(*mem->slots), mem->nslabs)); |
542 | |
543 | set_memory_encrypted(addr: tbl_vaddr, numpages: tbl_size >> PAGE_SHIFT); |
544 | if (mem->late_alloc) { |
545 | area_order = get_order(array_size(sizeof(*mem->areas), |
546 | mem->nareas)); |
547 | free_pages(addr: (unsigned long)mem->areas, order: area_order); |
548 | free_pages(addr: tbl_vaddr, order: get_order(size: tbl_size)); |
549 | free_pages(addr: (unsigned long)mem->slots, order: get_order(size: slots_size)); |
550 | } else { |
551 | memblock_free_late(__pa(mem->areas), |
552 | array_size(sizeof(*mem->areas), mem->nareas)); |
553 | memblock_free_late(base: mem->start, size: tbl_size); |
554 | memblock_free_late(__pa(mem->slots), size: slots_size); |
555 | } |
556 | |
557 | memset(mem, 0, sizeof(*mem)); |
558 | } |
559 | |
560 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
561 | |
562 | /** |
563 | * alloc_dma_pages() - allocate pages to be used for DMA |
564 | * @gfp: GFP flags for the allocation. |
565 | * @bytes: Size of the buffer. |
566 | * @phys_limit: Maximum allowed physical address of the buffer. |
567 | * |
568 | * Allocate pages from the buddy allocator. If successful, make the allocated |
569 | * pages decrypted that they can be used for DMA. |
570 | * |
571 | * Return: Decrypted pages, %NULL on allocation failure, or ERR_PTR(-EAGAIN) |
572 | * if the allocated physical address was above @phys_limit. |
573 | */ |
574 | static struct page *alloc_dma_pages(gfp_t gfp, size_t bytes, u64 phys_limit) |
575 | { |
576 | unsigned int order = get_order(size: bytes); |
577 | struct page *page; |
578 | phys_addr_t paddr; |
579 | void *vaddr; |
580 | |
581 | page = alloc_pages(gfp, order); |
582 | if (!page) |
583 | return NULL; |
584 | |
585 | paddr = page_to_phys(page); |
586 | if (paddr + bytes - 1 > phys_limit) { |
587 | __free_pages(page, order); |
588 | return ERR_PTR(error: -EAGAIN); |
589 | } |
590 | |
591 | vaddr = phys_to_virt(address: paddr); |
592 | if (set_memory_decrypted(addr: (unsigned long)vaddr, PFN_UP(bytes))) |
593 | goto error; |
594 | return page; |
595 | |
596 | error: |
597 | /* Intentional leak if pages cannot be encrypted again. */ |
598 | if (!set_memory_encrypted(addr: (unsigned long)vaddr, PFN_UP(bytes))) |
599 | __free_pages(page, order); |
600 | return NULL; |
601 | } |
602 | |
603 | /** |
604 | * swiotlb_alloc_tlb() - allocate a dynamic IO TLB buffer |
605 | * @dev: Device for which a memory pool is allocated. |
606 | * @bytes: Size of the buffer. |
607 | * @phys_limit: Maximum allowed physical address of the buffer. |
608 | * @gfp: GFP flags for the allocation. |
609 | * |
610 | * Return: Allocated pages, or %NULL on allocation failure. |
611 | */ |
612 | static struct page *swiotlb_alloc_tlb(struct device *dev, size_t bytes, |
613 | u64 phys_limit, gfp_t gfp) |
614 | { |
615 | struct page *page; |
616 | |
617 | /* |
618 | * Allocate from the atomic pools if memory is encrypted and |
619 | * the allocation is atomic, because decrypting may block. |
620 | */ |
621 | if (!gfpflags_allow_blocking(gfp_flags: gfp) && dev && force_dma_unencrypted(dev)) { |
622 | void *vaddr; |
623 | |
624 | if (!IS_ENABLED(CONFIG_DMA_COHERENT_POOL)) |
625 | return NULL; |
626 | |
627 | return dma_alloc_from_pool(dev, size: bytes, cpu_addr: &vaddr, flags: gfp, |
628 | phys_addr_ok: dma_coherent_ok); |
629 | } |
630 | |
631 | gfp &= ~GFP_ZONEMASK; |
632 | if (phys_limit <= DMA_BIT_MASK(zone_dma_bits)) |
633 | gfp |= __GFP_DMA; |
634 | else if (phys_limit <= DMA_BIT_MASK(32)) |
635 | gfp |= __GFP_DMA32; |
636 | |
637 | while (IS_ERR(ptr: page = alloc_dma_pages(gfp, bytes, phys_limit))) { |
638 | if (IS_ENABLED(CONFIG_ZONE_DMA32) && |
639 | phys_limit < DMA_BIT_MASK(64) && |
640 | !(gfp & (__GFP_DMA32 | __GFP_DMA))) |
641 | gfp |= __GFP_DMA32; |
642 | else if (IS_ENABLED(CONFIG_ZONE_DMA) && |
643 | !(gfp & __GFP_DMA)) |
644 | gfp = (gfp & ~__GFP_DMA32) | __GFP_DMA; |
645 | else |
646 | return NULL; |
647 | } |
648 | |
649 | return page; |
650 | } |
651 | |
652 | /** |
653 | * swiotlb_free_tlb() - free a dynamically allocated IO TLB buffer |
654 | * @vaddr: Virtual address of the buffer. |
655 | * @bytes: Size of the buffer. |
656 | */ |
657 | static void swiotlb_free_tlb(void *vaddr, size_t bytes) |
658 | { |
659 | if (IS_ENABLED(CONFIG_DMA_COHERENT_POOL) && |
660 | dma_free_from_pool(NULL, start: vaddr, size: bytes)) |
661 | return; |
662 | |
663 | /* Intentional leak if pages cannot be encrypted again. */ |
664 | if (!set_memory_encrypted(addr: (unsigned long)vaddr, PFN_UP(bytes))) |
665 | __free_pages(virt_to_page(vaddr), order: get_order(size: bytes)); |
666 | } |
667 | |
668 | /** |
669 | * swiotlb_alloc_pool() - allocate a new IO TLB memory pool |
670 | * @dev: Device for which a memory pool is allocated. |
671 | * @minslabs: Minimum number of slabs. |
672 | * @nslabs: Desired (maximum) number of slabs. |
673 | * @nareas: Number of areas. |
674 | * @phys_limit: Maximum DMA buffer physical address. |
675 | * @gfp: GFP flags for the allocations. |
676 | * |
677 | * Allocate and initialize a new IO TLB memory pool. The actual number of |
678 | * slabs may be reduced if allocation of @nslabs fails. If even |
679 | * @minslabs cannot be allocated, this function fails. |
680 | * |
681 | * Return: New memory pool, or %NULL on allocation failure. |
682 | */ |
683 | static struct io_tlb_pool *swiotlb_alloc_pool(struct device *dev, |
684 | unsigned long minslabs, unsigned long nslabs, |
685 | unsigned int nareas, u64 phys_limit, gfp_t gfp) |
686 | { |
687 | struct io_tlb_pool *pool; |
688 | unsigned int slot_order; |
689 | struct page *tlb; |
690 | size_t pool_size; |
691 | size_t tlb_size; |
692 | |
693 | if (nslabs > SLABS_PER_PAGE << MAX_PAGE_ORDER) { |
694 | nslabs = SLABS_PER_PAGE << MAX_PAGE_ORDER; |
695 | nareas = limit_nareas(nareas, nslots: nslabs); |
696 | } |
697 | |
698 | pool_size = sizeof(*pool) + array_size(sizeof(*pool->areas), nareas); |
699 | pool = kzalloc(size: pool_size, flags: gfp); |
700 | if (!pool) |
701 | goto error; |
702 | pool->areas = (void *)pool + sizeof(*pool); |
703 | |
704 | tlb_size = nslabs << IO_TLB_SHIFT; |
705 | while (!(tlb = swiotlb_alloc_tlb(dev, bytes: tlb_size, phys_limit, gfp))) { |
706 | if (nslabs <= minslabs) |
707 | goto error_tlb; |
708 | nslabs = ALIGN(nslabs >> 1, IO_TLB_SEGSIZE); |
709 | nareas = limit_nareas(nareas, nslots: nslabs); |
710 | tlb_size = nslabs << IO_TLB_SHIFT; |
711 | } |
712 | |
713 | slot_order = get_order(array_size(sizeof(*pool->slots), nslabs)); |
714 | pool->slots = (struct io_tlb_slot *) |
715 | __get_free_pages(gfp_mask: gfp, order: slot_order); |
716 | if (!pool->slots) |
717 | goto error_slots; |
718 | |
719 | swiotlb_init_io_tlb_pool(mem: pool, page_to_phys(tlb), nslabs, late_alloc: true, nareas); |
720 | return pool; |
721 | |
722 | error_slots: |
723 | swiotlb_free_tlb(page_address(tlb), bytes: tlb_size); |
724 | error_tlb: |
725 | kfree(objp: pool); |
726 | error: |
727 | return NULL; |
728 | } |
729 | |
730 | /** |
731 | * swiotlb_dyn_alloc() - dynamic memory pool allocation worker |
732 | * @work: Pointer to dyn_alloc in struct io_tlb_mem. |
733 | */ |
734 | static void swiotlb_dyn_alloc(struct work_struct *work) |
735 | { |
736 | struct io_tlb_mem *mem = |
737 | container_of(work, struct io_tlb_mem, dyn_alloc); |
738 | struct io_tlb_pool *pool; |
739 | |
740 | pool = swiotlb_alloc_pool(NULL, IO_TLB_MIN_SLABS, nslabs: default_nslabs, |
741 | nareas: default_nareas, phys_limit: mem->phys_limit, GFP_KERNEL); |
742 | if (!pool) { |
743 | pr_warn_ratelimited("Failed to allocate new pool" ); |
744 | return; |
745 | } |
746 | |
747 | add_mem_pool(mem, pool); |
748 | } |
749 | |
750 | /** |
751 | * swiotlb_dyn_free() - RCU callback to free a memory pool |
752 | * @rcu: RCU head in the corresponding struct io_tlb_pool. |
753 | */ |
754 | static void swiotlb_dyn_free(struct rcu_head *rcu) |
755 | { |
756 | struct io_tlb_pool *pool = container_of(rcu, struct io_tlb_pool, rcu); |
757 | size_t slots_size = array_size(sizeof(*pool->slots), pool->nslabs); |
758 | size_t tlb_size = pool->end - pool->start; |
759 | |
760 | free_pages(addr: (unsigned long)pool->slots, order: get_order(size: slots_size)); |
761 | swiotlb_free_tlb(vaddr: pool->vaddr, bytes: tlb_size); |
762 | kfree(objp: pool); |
763 | } |
764 | |
765 | /** |
766 | * swiotlb_find_pool() - find the IO TLB pool for a physical address |
767 | * @dev: Device which has mapped the DMA buffer. |
768 | * @paddr: Physical address within the DMA buffer. |
769 | * |
770 | * Find the IO TLB memory pool descriptor which contains the given physical |
771 | * address, if any. |
772 | * |
773 | * Return: Memory pool which contains @paddr, or %NULL if none. |
774 | */ |
775 | struct io_tlb_pool *swiotlb_find_pool(struct device *dev, phys_addr_t paddr) |
776 | { |
777 | struct io_tlb_mem *mem = dev->dma_io_tlb_mem; |
778 | struct io_tlb_pool *pool; |
779 | |
780 | rcu_read_lock(); |
781 | list_for_each_entry_rcu(pool, &mem->pools, node) { |
782 | if (paddr >= pool->start && paddr < pool->end) |
783 | goto out; |
784 | } |
785 | |
786 | list_for_each_entry_rcu(pool, &dev->dma_io_tlb_pools, node) { |
787 | if (paddr >= pool->start && paddr < pool->end) |
788 | goto out; |
789 | } |
790 | pool = NULL; |
791 | out: |
792 | rcu_read_unlock(); |
793 | return pool; |
794 | } |
795 | |
796 | /** |
797 | * swiotlb_del_pool() - remove an IO TLB pool from a device |
798 | * @dev: Owning device. |
799 | * @pool: Memory pool to be removed. |
800 | */ |
801 | static void swiotlb_del_pool(struct device *dev, struct io_tlb_pool *pool) |
802 | { |
803 | unsigned long flags; |
804 | |
805 | spin_lock_irqsave(&dev->dma_io_tlb_lock, flags); |
806 | list_del_rcu(entry: &pool->node); |
807 | spin_unlock_irqrestore(lock: &dev->dma_io_tlb_lock, flags); |
808 | |
809 | call_rcu(head: &pool->rcu, func: swiotlb_dyn_free); |
810 | } |
811 | |
812 | #endif /* CONFIG_SWIOTLB_DYNAMIC */ |
813 | |
814 | /** |
815 | * swiotlb_dev_init() - initialize swiotlb fields in &struct device |
816 | * @dev: Device to be initialized. |
817 | */ |
818 | void swiotlb_dev_init(struct device *dev) |
819 | { |
820 | dev->dma_io_tlb_mem = &io_tlb_default_mem; |
821 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
822 | INIT_LIST_HEAD(list: &dev->dma_io_tlb_pools); |
823 | spin_lock_init(&dev->dma_io_tlb_lock); |
824 | dev->dma_uses_io_tlb = false; |
825 | #endif |
826 | } |
827 | |
828 | /** |
829 | * swiotlb_align_offset() - Get required offset into an IO TLB allocation. |
830 | * @dev: Owning device. |
831 | * @align_mask: Allocation alignment mask. |
832 | * @addr: DMA address. |
833 | * |
834 | * Return the minimum offset from the start of an IO TLB allocation which is |
835 | * required for a given buffer address and allocation alignment to keep the |
836 | * device happy. |
837 | * |
838 | * First, the address bits covered by min_align_mask must be identical in the |
839 | * original address and the bounce buffer address. High bits are preserved by |
840 | * choosing a suitable IO TLB slot, but bits below IO_TLB_SHIFT require extra |
841 | * padding bytes before the bounce buffer. |
842 | * |
843 | * Second, @align_mask specifies which bits of the first allocated slot must |
844 | * be zero. This may require allocating additional padding slots, and then the |
845 | * offset (in bytes) from the first such padding slot is returned. |
846 | */ |
847 | static unsigned int swiotlb_align_offset(struct device *dev, |
848 | unsigned int align_mask, u64 addr) |
849 | { |
850 | return addr & dma_get_min_align_mask(dev) & |
851 | (align_mask | (IO_TLB_SIZE - 1)); |
852 | } |
853 | |
854 | /* |
855 | * Bounce: copy the swiotlb buffer from or back to the original dma location |
856 | */ |
857 | static void swiotlb_bounce(struct device *dev, phys_addr_t tlb_addr, size_t size, |
858 | enum dma_data_direction dir) |
859 | { |
860 | struct io_tlb_pool *mem = swiotlb_find_pool(dev, paddr: tlb_addr); |
861 | int index = (tlb_addr - mem->start) >> IO_TLB_SHIFT; |
862 | phys_addr_t orig_addr = mem->slots[index].orig_addr; |
863 | size_t alloc_size = mem->slots[index].alloc_size; |
864 | unsigned long pfn = PFN_DOWN(orig_addr); |
865 | unsigned char *vaddr = mem->vaddr + tlb_addr - mem->start; |
866 | int tlb_offset; |
867 | |
868 | if (orig_addr == INVALID_PHYS_ADDR) |
869 | return; |
870 | |
871 | /* |
872 | * It's valid for tlb_offset to be negative. This can happen when the |
873 | * "offset" returned by swiotlb_align_offset() is non-zero, and the |
874 | * tlb_addr is pointing within the first "offset" bytes of the second |
875 | * or subsequent slots of the allocated swiotlb area. While it's not |
876 | * valid for tlb_addr to be pointing within the first "offset" bytes |
877 | * of the first slot, there's no way to check for such an error since |
878 | * this function can't distinguish the first slot from the second and |
879 | * subsequent slots. |
880 | */ |
881 | tlb_offset = (tlb_addr & (IO_TLB_SIZE - 1)) - |
882 | swiotlb_align_offset(dev, align_mask: 0, addr: orig_addr); |
883 | |
884 | orig_addr += tlb_offset; |
885 | alloc_size -= tlb_offset; |
886 | |
887 | if (size > alloc_size) { |
888 | dev_WARN_ONCE(dev, 1, |
889 | "Buffer overflow detected. Allocation size: %zu. Mapping size: %zu.\n" , |
890 | alloc_size, size); |
891 | size = alloc_size; |
892 | } |
893 | |
894 | if (PageHighMem(pfn_to_page(pfn))) { |
895 | unsigned int offset = orig_addr & ~PAGE_MASK; |
896 | struct page *page; |
897 | unsigned int sz = 0; |
898 | unsigned long flags; |
899 | |
900 | while (size) { |
901 | sz = min_t(size_t, PAGE_SIZE - offset, size); |
902 | |
903 | local_irq_save(flags); |
904 | page = pfn_to_page(pfn); |
905 | if (dir == DMA_TO_DEVICE) |
906 | memcpy_from_page(to: vaddr, page, offset, len: sz); |
907 | else |
908 | memcpy_to_page(page, offset, from: vaddr, len: sz); |
909 | local_irq_restore(flags); |
910 | |
911 | size -= sz; |
912 | pfn++; |
913 | vaddr += sz; |
914 | offset = 0; |
915 | } |
916 | } else if (dir == DMA_TO_DEVICE) { |
917 | memcpy(vaddr, phys_to_virt(orig_addr), size); |
918 | } else { |
919 | memcpy(phys_to_virt(orig_addr), vaddr, size); |
920 | } |
921 | } |
922 | |
923 | static inline phys_addr_t slot_addr(phys_addr_t start, phys_addr_t idx) |
924 | { |
925 | return start + (idx << IO_TLB_SHIFT); |
926 | } |
927 | |
928 | /* |
929 | * Carefully handle integer overflow which can occur when boundary_mask == ~0UL. |
930 | */ |
931 | static inline unsigned long get_max_slots(unsigned long boundary_mask) |
932 | { |
933 | return (boundary_mask >> IO_TLB_SHIFT) + 1; |
934 | } |
935 | |
936 | static unsigned int wrap_area_index(struct io_tlb_pool *mem, unsigned int index) |
937 | { |
938 | if (index >= mem->area_nslabs) |
939 | return 0; |
940 | return index; |
941 | } |
942 | |
943 | /* |
944 | * Track the total used slots with a global atomic value in order to have |
945 | * correct information to determine the high water mark. The mem_used() |
946 | * function gives imprecise results because there's no locking across |
947 | * multiple areas. |
948 | */ |
949 | #ifdef CONFIG_DEBUG_FS |
950 | static void inc_used_and_hiwater(struct io_tlb_mem *mem, unsigned int nslots) |
951 | { |
952 | unsigned long old_hiwater, new_used; |
953 | |
954 | new_used = atomic_long_add_return(i: nslots, v: &mem->total_used); |
955 | old_hiwater = atomic_long_read(v: &mem->used_hiwater); |
956 | do { |
957 | if (new_used <= old_hiwater) |
958 | break; |
959 | } while (!atomic_long_try_cmpxchg(v: &mem->used_hiwater, |
960 | old: &old_hiwater, new: new_used)); |
961 | } |
962 | |
963 | static void dec_used(struct io_tlb_mem *mem, unsigned int nslots) |
964 | { |
965 | atomic_long_sub(i: nslots, v: &mem->total_used); |
966 | } |
967 | |
968 | #else /* !CONFIG_DEBUG_FS */ |
969 | static void inc_used_and_hiwater(struct io_tlb_mem *mem, unsigned int nslots) |
970 | { |
971 | } |
972 | static void dec_used(struct io_tlb_mem *mem, unsigned int nslots) |
973 | { |
974 | } |
975 | #endif /* CONFIG_DEBUG_FS */ |
976 | |
977 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
978 | #ifdef CONFIG_DEBUG_FS |
979 | static void inc_transient_used(struct io_tlb_mem *mem, unsigned int nslots) |
980 | { |
981 | atomic_long_add(i: nslots, v: &mem->transient_nslabs); |
982 | } |
983 | |
984 | static void dec_transient_used(struct io_tlb_mem *mem, unsigned int nslots) |
985 | { |
986 | atomic_long_sub(i: nslots, v: &mem->transient_nslabs); |
987 | } |
988 | |
989 | #else /* !CONFIG_DEBUG_FS */ |
990 | static void inc_transient_used(struct io_tlb_mem *mem, unsigned int nslots) |
991 | { |
992 | } |
993 | static void dec_transient_used(struct io_tlb_mem *mem, unsigned int nslots) |
994 | { |
995 | } |
996 | #endif /* CONFIG_DEBUG_FS */ |
997 | #endif /* CONFIG_SWIOTLB_DYNAMIC */ |
998 | |
999 | /** |
1000 | * swiotlb_search_pool_area() - search one memory area in one pool |
1001 | * @dev: Device which maps the buffer. |
1002 | * @pool: Memory pool to be searched. |
1003 | * @area_index: Index of the IO TLB memory area to be searched. |
1004 | * @orig_addr: Original (non-bounced) IO buffer address. |
1005 | * @alloc_size: Total requested size of the bounce buffer, |
1006 | * including initial alignment padding. |
1007 | * @alloc_align_mask: Required alignment of the allocated buffer. |
1008 | * |
1009 | * Find a suitable sequence of IO TLB entries for the request and allocate |
1010 | * a buffer from the given IO TLB memory area. |
1011 | * This function takes care of locking. |
1012 | * |
1013 | * Return: Index of the first allocated slot, or -1 on error. |
1014 | */ |
1015 | static int swiotlb_search_pool_area(struct device *dev, struct io_tlb_pool *pool, |
1016 | int area_index, phys_addr_t orig_addr, size_t alloc_size, |
1017 | unsigned int alloc_align_mask) |
1018 | { |
1019 | struct io_tlb_area *area = pool->areas + area_index; |
1020 | unsigned long boundary_mask = dma_get_seg_boundary(dev); |
1021 | dma_addr_t tbl_dma_addr = |
1022 | phys_to_dma_unencrypted(dev, paddr: pool->start) & boundary_mask; |
1023 | unsigned long max_slots = get_max_slots(boundary_mask); |
1024 | unsigned int iotlb_align_mask = dma_get_min_align_mask(dev); |
1025 | unsigned int nslots = nr_slots(val: alloc_size), stride; |
1026 | unsigned int offset = swiotlb_align_offset(dev, align_mask: 0, addr: orig_addr); |
1027 | unsigned int index, slots_checked, count = 0, i; |
1028 | unsigned long flags; |
1029 | unsigned int slot_base; |
1030 | unsigned int slot_index; |
1031 | |
1032 | BUG_ON(!nslots); |
1033 | BUG_ON(area_index >= pool->nareas); |
1034 | |
1035 | /* |
1036 | * Historically, swiotlb allocations >= PAGE_SIZE were guaranteed to be |
1037 | * page-aligned in the absence of any other alignment requirements. |
1038 | * 'alloc_align_mask' was later introduced to specify the alignment |
1039 | * explicitly, however this is passed as zero for streaming mappings |
1040 | * and so we preserve the old behaviour there in case any drivers are |
1041 | * relying on it. |
1042 | */ |
1043 | if (!alloc_align_mask && !iotlb_align_mask && alloc_size >= PAGE_SIZE) |
1044 | alloc_align_mask = PAGE_SIZE - 1; |
1045 | |
1046 | /* |
1047 | * Ensure that the allocation is at least slot-aligned and update |
1048 | * 'iotlb_align_mask' to ignore bits that will be preserved when |
1049 | * offsetting into the allocation. |
1050 | */ |
1051 | alloc_align_mask |= (IO_TLB_SIZE - 1); |
1052 | iotlb_align_mask &= ~alloc_align_mask; |
1053 | |
1054 | /* |
1055 | * For mappings with an alignment requirement don't bother looping to |
1056 | * unaligned slots once we found an aligned one. |
1057 | */ |
1058 | stride = get_max_slots(max(alloc_align_mask, iotlb_align_mask)); |
1059 | |
1060 | spin_lock_irqsave(&area->lock, flags); |
1061 | if (unlikely(nslots > pool->area_nslabs - area->used)) |
1062 | goto not_found; |
1063 | |
1064 | slot_base = area_index * pool->area_nslabs; |
1065 | index = area->index; |
1066 | |
1067 | for (slots_checked = 0; slots_checked < pool->area_nslabs; ) { |
1068 | phys_addr_t tlb_addr; |
1069 | |
1070 | slot_index = slot_base + index; |
1071 | tlb_addr = slot_addr(start: tbl_dma_addr, idx: slot_index); |
1072 | |
1073 | if ((tlb_addr & alloc_align_mask) || |
1074 | (orig_addr && (tlb_addr & iotlb_align_mask) != |
1075 | (orig_addr & iotlb_align_mask))) { |
1076 | index = wrap_area_index(mem: pool, index: index + 1); |
1077 | slots_checked++; |
1078 | continue; |
1079 | } |
1080 | |
1081 | if (!iommu_is_span_boundary(index: slot_index, nr: nslots, |
1082 | shift: nr_slots(val: tbl_dma_addr), |
1083 | boundary_size: max_slots)) { |
1084 | if (pool->slots[slot_index].list >= nslots) |
1085 | goto found; |
1086 | } |
1087 | index = wrap_area_index(mem: pool, index: index + stride); |
1088 | slots_checked += stride; |
1089 | } |
1090 | |
1091 | not_found: |
1092 | spin_unlock_irqrestore(lock: &area->lock, flags); |
1093 | return -1; |
1094 | |
1095 | found: |
1096 | /* |
1097 | * If we find a slot that indicates we have 'nslots' number of |
1098 | * contiguous buffers, we allocate the buffers from that slot onwards |
1099 | * and set the list of free entries to '0' indicating unavailable. |
1100 | */ |
1101 | for (i = slot_index; i < slot_index + nslots; i++) { |
1102 | pool->slots[i].list = 0; |
1103 | pool->slots[i].alloc_size = alloc_size - (offset + |
1104 | ((i - slot_index) << IO_TLB_SHIFT)); |
1105 | } |
1106 | for (i = slot_index - 1; |
1107 | io_tlb_offset(val: i) != IO_TLB_SEGSIZE - 1 && |
1108 | pool->slots[i].list; i--) |
1109 | pool->slots[i].list = ++count; |
1110 | |
1111 | /* |
1112 | * Update the indices to avoid searching in the next round. |
1113 | */ |
1114 | area->index = wrap_area_index(mem: pool, index: index + nslots); |
1115 | area->used += nslots; |
1116 | spin_unlock_irqrestore(lock: &area->lock, flags); |
1117 | |
1118 | inc_used_and_hiwater(mem: dev->dma_io_tlb_mem, nslots); |
1119 | return slot_index; |
1120 | } |
1121 | |
1122 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1123 | |
1124 | /** |
1125 | * swiotlb_search_area() - search one memory area in all pools |
1126 | * @dev: Device which maps the buffer. |
1127 | * @start_cpu: Start CPU number. |
1128 | * @cpu_offset: Offset from @start_cpu. |
1129 | * @orig_addr: Original (non-bounced) IO buffer address. |
1130 | * @alloc_size: Total requested size of the bounce buffer, |
1131 | * including initial alignment padding. |
1132 | * @alloc_align_mask: Required alignment of the allocated buffer. |
1133 | * @retpool: Used memory pool, updated on return. |
1134 | * |
1135 | * Search one memory area in all pools for a sequence of slots that match the |
1136 | * allocation constraints. |
1137 | * |
1138 | * Return: Index of the first allocated slot, or -1 on error. |
1139 | */ |
1140 | static int swiotlb_search_area(struct device *dev, int start_cpu, |
1141 | int cpu_offset, phys_addr_t orig_addr, size_t alloc_size, |
1142 | unsigned int alloc_align_mask, struct io_tlb_pool **retpool) |
1143 | { |
1144 | struct io_tlb_mem *mem = dev->dma_io_tlb_mem; |
1145 | struct io_tlb_pool *pool; |
1146 | int area_index; |
1147 | int index = -1; |
1148 | |
1149 | rcu_read_lock(); |
1150 | list_for_each_entry_rcu(pool, &mem->pools, node) { |
1151 | if (cpu_offset >= pool->nareas) |
1152 | continue; |
1153 | area_index = (start_cpu + cpu_offset) & (pool->nareas - 1); |
1154 | index = swiotlb_search_pool_area(dev, pool, area_index, |
1155 | orig_addr, alloc_size, |
1156 | alloc_align_mask); |
1157 | if (index >= 0) { |
1158 | *retpool = pool; |
1159 | break; |
1160 | } |
1161 | } |
1162 | rcu_read_unlock(); |
1163 | return index; |
1164 | } |
1165 | |
1166 | /** |
1167 | * swiotlb_find_slots() - search for slots in the whole swiotlb |
1168 | * @dev: Device which maps the buffer. |
1169 | * @orig_addr: Original (non-bounced) IO buffer address. |
1170 | * @alloc_size: Total requested size of the bounce buffer, |
1171 | * including initial alignment padding. |
1172 | * @alloc_align_mask: Required alignment of the allocated buffer. |
1173 | * @retpool: Used memory pool, updated on return. |
1174 | * |
1175 | * Search through the whole software IO TLB to find a sequence of slots that |
1176 | * match the allocation constraints. |
1177 | * |
1178 | * Return: Index of the first allocated slot, or -1 on error. |
1179 | */ |
1180 | static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr, |
1181 | size_t alloc_size, unsigned int alloc_align_mask, |
1182 | struct io_tlb_pool **retpool) |
1183 | { |
1184 | struct io_tlb_mem *mem = dev->dma_io_tlb_mem; |
1185 | struct io_tlb_pool *pool; |
1186 | unsigned long nslabs; |
1187 | unsigned long flags; |
1188 | u64 phys_limit; |
1189 | int cpu, i; |
1190 | int index; |
1191 | |
1192 | if (alloc_size > IO_TLB_SEGSIZE * IO_TLB_SIZE) |
1193 | return -1; |
1194 | |
1195 | cpu = raw_smp_processor_id(); |
1196 | for (i = 0; i < default_nareas; ++i) { |
1197 | index = swiotlb_search_area(dev, start_cpu: cpu, cpu_offset: i, orig_addr, alloc_size, |
1198 | alloc_align_mask, retpool: &pool); |
1199 | if (index >= 0) |
1200 | goto found; |
1201 | } |
1202 | |
1203 | if (!mem->can_grow) |
1204 | return -1; |
1205 | |
1206 | schedule_work(work: &mem->dyn_alloc); |
1207 | |
1208 | nslabs = nr_slots(val: alloc_size); |
1209 | phys_limit = min_not_zero(*dev->dma_mask, dev->bus_dma_limit); |
1210 | pool = swiotlb_alloc_pool(dev, minslabs: nslabs, nslabs, nareas: 1, phys_limit, |
1211 | GFP_NOWAIT | __GFP_NOWARN); |
1212 | if (!pool) |
1213 | return -1; |
1214 | |
1215 | index = swiotlb_search_pool_area(dev, pool, area_index: 0, orig_addr, |
1216 | alloc_size, alloc_align_mask); |
1217 | if (index < 0) { |
1218 | swiotlb_dyn_free(rcu: &pool->rcu); |
1219 | return -1; |
1220 | } |
1221 | |
1222 | pool->transient = true; |
1223 | spin_lock_irqsave(&dev->dma_io_tlb_lock, flags); |
1224 | list_add_rcu(new: &pool->node, head: &dev->dma_io_tlb_pools); |
1225 | spin_unlock_irqrestore(lock: &dev->dma_io_tlb_lock, flags); |
1226 | inc_transient_used(mem, nslots: pool->nslabs); |
1227 | |
1228 | found: |
1229 | WRITE_ONCE(dev->dma_uses_io_tlb, true); |
1230 | |
1231 | /* |
1232 | * The general barrier orders reads and writes against a presumed store |
1233 | * of the SWIOTLB buffer address by a device driver (to a driver private |
1234 | * data structure). It serves two purposes. |
1235 | * |
1236 | * First, the store to dev->dma_uses_io_tlb must be ordered before the |
1237 | * presumed store. This guarantees that the returned buffer address |
1238 | * cannot be passed to another CPU before updating dev->dma_uses_io_tlb. |
1239 | * |
1240 | * Second, the load from mem->pools must be ordered before the same |
1241 | * presumed store. This guarantees that the returned buffer address |
1242 | * cannot be observed by another CPU before an update of the RCU list |
1243 | * that was made by swiotlb_dyn_alloc() on a third CPU (cf. multicopy |
1244 | * atomicity). |
1245 | * |
1246 | * See also the comment in is_swiotlb_buffer(). |
1247 | */ |
1248 | smp_mb(); |
1249 | |
1250 | *retpool = pool; |
1251 | return index; |
1252 | } |
1253 | |
1254 | #else /* !CONFIG_SWIOTLB_DYNAMIC */ |
1255 | |
1256 | static int swiotlb_find_slots(struct device *dev, phys_addr_t orig_addr, |
1257 | size_t alloc_size, unsigned int alloc_align_mask, |
1258 | struct io_tlb_pool **retpool) |
1259 | { |
1260 | struct io_tlb_pool *pool; |
1261 | int start, i; |
1262 | int index; |
1263 | |
1264 | *retpool = pool = &dev->dma_io_tlb_mem->defpool; |
1265 | i = start = raw_smp_processor_id() & (pool->nareas - 1); |
1266 | do { |
1267 | index = swiotlb_search_pool_area(dev, pool, i, orig_addr, |
1268 | alloc_size, alloc_align_mask); |
1269 | if (index >= 0) |
1270 | return index; |
1271 | if (++i >= pool->nareas) |
1272 | i = 0; |
1273 | } while (i != start); |
1274 | return -1; |
1275 | } |
1276 | |
1277 | #endif /* CONFIG_SWIOTLB_DYNAMIC */ |
1278 | |
1279 | #ifdef CONFIG_DEBUG_FS |
1280 | |
1281 | /** |
1282 | * mem_used() - get number of used slots in an allocator |
1283 | * @mem: Software IO TLB allocator. |
1284 | * |
1285 | * The result is accurate in this version of the function, because an atomic |
1286 | * counter is available if CONFIG_DEBUG_FS is set. |
1287 | * |
1288 | * Return: Number of used slots. |
1289 | */ |
1290 | static unsigned long mem_used(struct io_tlb_mem *mem) |
1291 | { |
1292 | return atomic_long_read(v: &mem->total_used); |
1293 | } |
1294 | |
1295 | #else /* !CONFIG_DEBUG_FS */ |
1296 | |
1297 | /** |
1298 | * mem_pool_used() - get number of used slots in a memory pool |
1299 | * @pool: Software IO TLB memory pool. |
1300 | * |
1301 | * The result is not accurate, see mem_used(). |
1302 | * |
1303 | * Return: Approximate number of used slots. |
1304 | */ |
1305 | static unsigned long mem_pool_used(struct io_tlb_pool *pool) |
1306 | { |
1307 | int i; |
1308 | unsigned long used = 0; |
1309 | |
1310 | for (i = 0; i < pool->nareas; i++) |
1311 | used += pool->areas[i].used; |
1312 | return used; |
1313 | } |
1314 | |
1315 | /** |
1316 | * mem_used() - get number of used slots in an allocator |
1317 | * @mem: Software IO TLB allocator. |
1318 | * |
1319 | * The result is not accurate, because there is no locking of individual |
1320 | * areas. |
1321 | * |
1322 | * Return: Approximate number of used slots. |
1323 | */ |
1324 | static unsigned long mem_used(struct io_tlb_mem *mem) |
1325 | { |
1326 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1327 | struct io_tlb_pool *pool; |
1328 | unsigned long used = 0; |
1329 | |
1330 | rcu_read_lock(); |
1331 | list_for_each_entry_rcu(pool, &mem->pools, node) |
1332 | used += mem_pool_used(pool); |
1333 | rcu_read_unlock(); |
1334 | |
1335 | return used; |
1336 | #else |
1337 | return mem_pool_used(&mem->defpool); |
1338 | #endif |
1339 | } |
1340 | |
1341 | #endif /* CONFIG_DEBUG_FS */ |
1342 | |
1343 | phys_addr_t swiotlb_tbl_map_single(struct device *dev, phys_addr_t orig_addr, |
1344 | size_t mapping_size, size_t alloc_size, |
1345 | unsigned int alloc_align_mask, enum dma_data_direction dir, |
1346 | unsigned long attrs) |
1347 | { |
1348 | struct io_tlb_mem *mem = dev->dma_io_tlb_mem; |
1349 | unsigned int offset; |
1350 | struct io_tlb_pool *pool; |
1351 | unsigned int i; |
1352 | int index; |
1353 | phys_addr_t tlb_addr; |
1354 | unsigned short pad_slots; |
1355 | |
1356 | if (!mem || !mem->nslabs) { |
1357 | dev_warn_ratelimited(dev, |
1358 | "Can not allocate SWIOTLB buffer earlier and can't now provide you with the DMA bounce buffer" ); |
1359 | return (phys_addr_t)DMA_MAPPING_ERROR; |
1360 | } |
1361 | |
1362 | if (cc_platform_has(attr: CC_ATTR_MEM_ENCRYPT)) |
1363 | pr_warn_once("Memory encryption is active and system is using DMA bounce buffers\n" ); |
1364 | |
1365 | if (mapping_size > alloc_size) { |
1366 | dev_warn_once(dev, "Invalid sizes (mapping: %zd bytes, alloc: %zd bytes)" , |
1367 | mapping_size, alloc_size); |
1368 | return (phys_addr_t)DMA_MAPPING_ERROR; |
1369 | } |
1370 | |
1371 | offset = swiotlb_align_offset(dev, align_mask: alloc_align_mask, addr: orig_addr); |
1372 | index = swiotlb_find_slots(dev, orig_addr, |
1373 | alloc_size: alloc_size + offset, alloc_align_mask, retpool: &pool); |
1374 | if (index == -1) { |
1375 | if (!(attrs & DMA_ATTR_NO_WARN)) |
1376 | dev_warn_ratelimited(dev, |
1377 | "swiotlb buffer is full (sz: %zd bytes), total %lu (slots), used %lu (slots)\n" , |
1378 | alloc_size, mem->nslabs, mem_used(mem)); |
1379 | return (phys_addr_t)DMA_MAPPING_ERROR; |
1380 | } |
1381 | |
1382 | /* |
1383 | * Save away the mapping from the original address to the DMA address. |
1384 | * This is needed when we sync the memory. Then we sync the buffer if |
1385 | * needed. |
1386 | */ |
1387 | pad_slots = offset >> IO_TLB_SHIFT; |
1388 | offset &= (IO_TLB_SIZE - 1); |
1389 | index += pad_slots; |
1390 | pool->slots[index].pad_slots = pad_slots; |
1391 | for (i = 0; i < nr_slots(val: alloc_size + offset); i++) |
1392 | pool->slots[index + i].orig_addr = slot_addr(start: orig_addr, idx: i); |
1393 | tlb_addr = slot_addr(start: pool->start, idx: index) + offset; |
1394 | /* |
1395 | * When the device is writing memory, i.e. dir == DMA_FROM_DEVICE, copy |
1396 | * the original buffer to the TLB buffer before initiating DMA in order |
1397 | * to preserve the original's data if the device does a partial write, |
1398 | * i.e. if the device doesn't overwrite the entire buffer. Preserving |
1399 | * the original data, even if it's garbage, is necessary to match |
1400 | * hardware behavior. Use of swiotlb is supposed to be transparent, |
1401 | * i.e. swiotlb must not corrupt memory by clobbering unwritten bytes. |
1402 | */ |
1403 | swiotlb_bounce(dev, tlb_addr, size: mapping_size, dir: DMA_TO_DEVICE); |
1404 | return tlb_addr; |
1405 | } |
1406 | |
1407 | static void swiotlb_release_slots(struct device *dev, phys_addr_t tlb_addr) |
1408 | { |
1409 | struct io_tlb_pool *mem = swiotlb_find_pool(dev, paddr: tlb_addr); |
1410 | unsigned long flags; |
1411 | unsigned int offset = swiotlb_align_offset(dev, align_mask: 0, addr: tlb_addr); |
1412 | int index, nslots, aindex; |
1413 | struct io_tlb_area *area; |
1414 | int count, i; |
1415 | |
1416 | index = (tlb_addr - offset - mem->start) >> IO_TLB_SHIFT; |
1417 | index -= mem->slots[index].pad_slots; |
1418 | nslots = nr_slots(val: mem->slots[index].alloc_size + offset); |
1419 | aindex = index / mem->area_nslabs; |
1420 | area = &mem->areas[aindex]; |
1421 | |
1422 | /* |
1423 | * Return the buffer to the free list by setting the corresponding |
1424 | * entries to indicate the number of contiguous entries available. |
1425 | * While returning the entries to the free list, we merge the entries |
1426 | * with slots below and above the pool being returned. |
1427 | */ |
1428 | BUG_ON(aindex >= mem->nareas); |
1429 | |
1430 | spin_lock_irqsave(&area->lock, flags); |
1431 | if (index + nslots < ALIGN(index + 1, IO_TLB_SEGSIZE)) |
1432 | count = mem->slots[index + nslots].list; |
1433 | else |
1434 | count = 0; |
1435 | |
1436 | /* |
1437 | * Step 1: return the slots to the free list, merging the slots with |
1438 | * superceeding slots |
1439 | */ |
1440 | for (i = index + nslots - 1; i >= index; i--) { |
1441 | mem->slots[i].list = ++count; |
1442 | mem->slots[i].orig_addr = INVALID_PHYS_ADDR; |
1443 | mem->slots[i].alloc_size = 0; |
1444 | mem->slots[i].pad_slots = 0; |
1445 | } |
1446 | |
1447 | /* |
1448 | * Step 2: merge the returned slots with the preceding slots, if |
1449 | * available (non zero) |
1450 | */ |
1451 | for (i = index - 1; |
1452 | io_tlb_offset(val: i) != IO_TLB_SEGSIZE - 1 && mem->slots[i].list; |
1453 | i--) |
1454 | mem->slots[i].list = ++count; |
1455 | area->used -= nslots; |
1456 | spin_unlock_irqrestore(lock: &area->lock, flags); |
1457 | |
1458 | dec_used(mem: dev->dma_io_tlb_mem, nslots); |
1459 | } |
1460 | |
1461 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1462 | |
1463 | /** |
1464 | * swiotlb_del_transient() - delete a transient memory pool |
1465 | * @dev: Device which mapped the buffer. |
1466 | * @tlb_addr: Physical address within a bounce buffer. |
1467 | * |
1468 | * Check whether the address belongs to a transient SWIOTLB memory pool. |
1469 | * If yes, then delete the pool. |
1470 | * |
1471 | * Return: %true if @tlb_addr belonged to a transient pool that was released. |
1472 | */ |
1473 | static bool swiotlb_del_transient(struct device *dev, phys_addr_t tlb_addr) |
1474 | { |
1475 | struct io_tlb_pool *pool; |
1476 | |
1477 | pool = swiotlb_find_pool(dev, paddr: tlb_addr); |
1478 | if (!pool->transient) |
1479 | return false; |
1480 | |
1481 | dec_used(mem: dev->dma_io_tlb_mem, nslots: pool->nslabs); |
1482 | swiotlb_del_pool(dev, pool); |
1483 | dec_transient_used(mem: dev->dma_io_tlb_mem, nslots: pool->nslabs); |
1484 | return true; |
1485 | } |
1486 | |
1487 | #else /* !CONFIG_SWIOTLB_DYNAMIC */ |
1488 | |
1489 | static inline bool swiotlb_del_transient(struct device *dev, |
1490 | phys_addr_t tlb_addr) |
1491 | { |
1492 | return false; |
1493 | } |
1494 | |
1495 | #endif /* CONFIG_SWIOTLB_DYNAMIC */ |
1496 | |
1497 | /* |
1498 | * tlb_addr is the physical address of the bounce buffer to unmap. |
1499 | */ |
1500 | void swiotlb_tbl_unmap_single(struct device *dev, phys_addr_t tlb_addr, |
1501 | size_t mapping_size, enum dma_data_direction dir, |
1502 | unsigned long attrs) |
1503 | { |
1504 | /* |
1505 | * First, sync the memory before unmapping the entry |
1506 | */ |
1507 | if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) && |
1508 | (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)) |
1509 | swiotlb_bounce(dev, tlb_addr, size: mapping_size, dir: DMA_FROM_DEVICE); |
1510 | |
1511 | if (swiotlb_del_transient(dev, tlb_addr)) |
1512 | return; |
1513 | swiotlb_release_slots(dev, tlb_addr); |
1514 | } |
1515 | |
1516 | void swiotlb_sync_single_for_device(struct device *dev, phys_addr_t tlb_addr, |
1517 | size_t size, enum dma_data_direction dir) |
1518 | { |
1519 | if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL) |
1520 | swiotlb_bounce(dev, tlb_addr, size, dir: DMA_TO_DEVICE); |
1521 | else |
1522 | BUG_ON(dir != DMA_FROM_DEVICE); |
1523 | } |
1524 | |
1525 | void swiotlb_sync_single_for_cpu(struct device *dev, phys_addr_t tlb_addr, |
1526 | size_t size, enum dma_data_direction dir) |
1527 | { |
1528 | if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL) |
1529 | swiotlb_bounce(dev, tlb_addr, size, dir: DMA_FROM_DEVICE); |
1530 | else |
1531 | BUG_ON(dir != DMA_TO_DEVICE); |
1532 | } |
1533 | |
1534 | /* |
1535 | * Create a swiotlb mapping for the buffer at @paddr, and in case of DMAing |
1536 | * to the device copy the data into it as well. |
1537 | */ |
1538 | dma_addr_t swiotlb_map(struct device *dev, phys_addr_t paddr, size_t size, |
1539 | enum dma_data_direction dir, unsigned long attrs) |
1540 | { |
1541 | phys_addr_t swiotlb_addr; |
1542 | dma_addr_t dma_addr; |
1543 | |
1544 | trace_swiotlb_bounced(dev, dev_addr: phys_to_dma(dev, paddr), size); |
1545 | |
1546 | swiotlb_addr = swiotlb_tbl_map_single(dev, orig_addr: paddr, mapping_size: size, alloc_size: size, alloc_align_mask: 0, dir, |
1547 | attrs); |
1548 | if (swiotlb_addr == (phys_addr_t)DMA_MAPPING_ERROR) |
1549 | return DMA_MAPPING_ERROR; |
1550 | |
1551 | /* Ensure that the address returned is DMA'ble */ |
1552 | dma_addr = phys_to_dma_unencrypted(dev, paddr: swiotlb_addr); |
1553 | if (unlikely(!dma_capable(dev, dma_addr, size, true))) { |
1554 | swiotlb_tbl_unmap_single(dev, tlb_addr: swiotlb_addr, mapping_size: size, dir, |
1555 | attrs: attrs | DMA_ATTR_SKIP_CPU_SYNC); |
1556 | dev_WARN_ONCE(dev, 1, |
1557 | "swiotlb addr %pad+%zu overflow (mask %llx, bus limit %llx).\n" , |
1558 | &dma_addr, size, *dev->dma_mask, dev->bus_dma_limit); |
1559 | return DMA_MAPPING_ERROR; |
1560 | } |
1561 | |
1562 | if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) |
1563 | arch_sync_dma_for_device(paddr: swiotlb_addr, size, dir); |
1564 | return dma_addr; |
1565 | } |
1566 | |
1567 | size_t swiotlb_max_mapping_size(struct device *dev) |
1568 | { |
1569 | int min_align_mask = dma_get_min_align_mask(dev); |
1570 | int min_align = 0; |
1571 | |
1572 | /* |
1573 | * swiotlb_find_slots() skips slots according to |
1574 | * min align mask. This affects max mapping size. |
1575 | * Take it into acount here. |
1576 | */ |
1577 | if (min_align_mask) |
1578 | min_align = roundup(min_align_mask, IO_TLB_SIZE); |
1579 | |
1580 | return ((size_t)IO_TLB_SIZE) * IO_TLB_SEGSIZE - min_align; |
1581 | } |
1582 | |
1583 | /** |
1584 | * is_swiotlb_allocated() - check if the default software IO TLB is initialized |
1585 | */ |
1586 | bool is_swiotlb_allocated(void) |
1587 | { |
1588 | return io_tlb_default_mem.nslabs; |
1589 | } |
1590 | |
1591 | bool is_swiotlb_active(struct device *dev) |
1592 | { |
1593 | struct io_tlb_mem *mem = dev->dma_io_tlb_mem; |
1594 | |
1595 | return mem && mem->nslabs; |
1596 | } |
1597 | |
1598 | /** |
1599 | * default_swiotlb_base() - get the base address of the default SWIOTLB |
1600 | * |
1601 | * Get the lowest physical address used by the default software IO TLB pool. |
1602 | */ |
1603 | phys_addr_t default_swiotlb_base(void) |
1604 | { |
1605 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1606 | io_tlb_default_mem.can_grow = false; |
1607 | #endif |
1608 | return io_tlb_default_mem.defpool.start; |
1609 | } |
1610 | |
1611 | /** |
1612 | * default_swiotlb_limit() - get the address limit of the default SWIOTLB |
1613 | * |
1614 | * Get the highest physical address used by the default software IO TLB pool. |
1615 | */ |
1616 | phys_addr_t default_swiotlb_limit(void) |
1617 | { |
1618 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1619 | return io_tlb_default_mem.phys_limit; |
1620 | #else |
1621 | return io_tlb_default_mem.defpool.end - 1; |
1622 | #endif |
1623 | } |
1624 | |
1625 | #ifdef CONFIG_DEBUG_FS |
1626 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1627 | static unsigned long mem_transient_used(struct io_tlb_mem *mem) |
1628 | { |
1629 | return atomic_long_read(v: &mem->transient_nslabs); |
1630 | } |
1631 | |
1632 | static int io_tlb_transient_used_get(void *data, u64 *val) |
1633 | { |
1634 | struct io_tlb_mem *mem = data; |
1635 | |
1636 | *val = mem_transient_used(mem); |
1637 | return 0; |
1638 | } |
1639 | |
1640 | DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_transient_used, io_tlb_transient_used_get, |
1641 | NULL, "%llu\n" ); |
1642 | #endif /* CONFIG_SWIOTLB_DYNAMIC */ |
1643 | |
1644 | static int io_tlb_used_get(void *data, u64 *val) |
1645 | { |
1646 | struct io_tlb_mem *mem = data; |
1647 | |
1648 | *val = mem_used(mem); |
1649 | return 0; |
1650 | } |
1651 | |
1652 | static int io_tlb_hiwater_get(void *data, u64 *val) |
1653 | { |
1654 | struct io_tlb_mem *mem = data; |
1655 | |
1656 | *val = atomic_long_read(v: &mem->used_hiwater); |
1657 | return 0; |
1658 | } |
1659 | |
1660 | static int io_tlb_hiwater_set(void *data, u64 val) |
1661 | { |
1662 | struct io_tlb_mem *mem = data; |
1663 | |
1664 | /* Only allow setting to zero */ |
1665 | if (val != 0) |
1666 | return -EINVAL; |
1667 | |
1668 | atomic_long_set(v: &mem->used_hiwater, i: val); |
1669 | return 0; |
1670 | } |
1671 | |
1672 | DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_used, io_tlb_used_get, NULL, "%llu\n" ); |
1673 | DEFINE_DEBUGFS_ATTRIBUTE(fops_io_tlb_hiwater, io_tlb_hiwater_get, |
1674 | io_tlb_hiwater_set, "%llu\n" ); |
1675 | |
1676 | static void swiotlb_create_debugfs_files(struct io_tlb_mem *mem, |
1677 | const char *dirname) |
1678 | { |
1679 | mem->debugfs = debugfs_create_dir(name: dirname, parent: io_tlb_default_mem.debugfs); |
1680 | if (!mem->nslabs) |
1681 | return; |
1682 | |
1683 | debugfs_create_ulong(name: "io_tlb_nslabs" , mode: 0400, parent: mem->debugfs, value: &mem->nslabs); |
1684 | debugfs_create_file(name: "io_tlb_used" , mode: 0400, parent: mem->debugfs, data: mem, |
1685 | fops: &fops_io_tlb_used); |
1686 | debugfs_create_file(name: "io_tlb_used_hiwater" , mode: 0600, parent: mem->debugfs, data: mem, |
1687 | fops: &fops_io_tlb_hiwater); |
1688 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1689 | debugfs_create_file(name: "io_tlb_transient_nslabs" , mode: 0400, parent: mem->debugfs, |
1690 | data: mem, fops: &fops_io_tlb_transient_used); |
1691 | #endif |
1692 | } |
1693 | |
1694 | static int __init swiotlb_create_default_debugfs(void) |
1695 | { |
1696 | swiotlb_create_debugfs_files(mem: &io_tlb_default_mem, dirname: "swiotlb" ); |
1697 | return 0; |
1698 | } |
1699 | |
1700 | late_initcall(swiotlb_create_default_debugfs); |
1701 | |
1702 | #else /* !CONFIG_DEBUG_FS */ |
1703 | |
1704 | static inline void swiotlb_create_debugfs_files(struct io_tlb_mem *mem, |
1705 | const char *dirname) |
1706 | { |
1707 | } |
1708 | |
1709 | #endif /* CONFIG_DEBUG_FS */ |
1710 | |
1711 | #ifdef CONFIG_DMA_RESTRICTED_POOL |
1712 | |
1713 | struct page *swiotlb_alloc(struct device *dev, size_t size) |
1714 | { |
1715 | struct io_tlb_mem *mem = dev->dma_io_tlb_mem; |
1716 | struct io_tlb_pool *pool; |
1717 | phys_addr_t tlb_addr; |
1718 | unsigned int align; |
1719 | int index; |
1720 | |
1721 | if (!mem) |
1722 | return NULL; |
1723 | |
1724 | align = (1 << (get_order(size) + PAGE_SHIFT)) - 1; |
1725 | index = swiotlb_find_slots(dev, orig_addr: 0, alloc_size: size, alloc_align_mask: align, retpool: &pool); |
1726 | if (index == -1) |
1727 | return NULL; |
1728 | |
1729 | tlb_addr = slot_addr(start: pool->start, idx: index); |
1730 | if (unlikely(!PAGE_ALIGNED(tlb_addr))) { |
1731 | dev_WARN_ONCE(dev, 1, "Cannot allocate pages from non page-aligned swiotlb addr 0x%pa.\n" , |
1732 | &tlb_addr); |
1733 | swiotlb_release_slots(dev, tlb_addr); |
1734 | return NULL; |
1735 | } |
1736 | |
1737 | return pfn_to_page(PFN_DOWN(tlb_addr)); |
1738 | } |
1739 | |
1740 | bool swiotlb_free(struct device *dev, struct page *page, size_t size) |
1741 | { |
1742 | phys_addr_t tlb_addr = page_to_phys(page); |
1743 | |
1744 | if (!is_swiotlb_buffer(dev, paddr: tlb_addr)) |
1745 | return false; |
1746 | |
1747 | swiotlb_release_slots(dev, tlb_addr); |
1748 | |
1749 | return true; |
1750 | } |
1751 | |
1752 | static int rmem_swiotlb_device_init(struct reserved_mem *rmem, |
1753 | struct device *dev) |
1754 | { |
1755 | struct io_tlb_mem *mem = rmem->priv; |
1756 | unsigned long nslabs = rmem->size >> IO_TLB_SHIFT; |
1757 | |
1758 | /* Set Per-device io tlb area to one */ |
1759 | unsigned int nareas = 1; |
1760 | |
1761 | if (PageHighMem(pfn_to_page(PHYS_PFN(rmem->base)))) { |
1762 | dev_err(dev, "Restricted DMA pool must be accessible within the linear mapping." ); |
1763 | return -EINVAL; |
1764 | } |
1765 | |
1766 | /* |
1767 | * Since multiple devices can share the same pool, the private data, |
1768 | * io_tlb_mem struct, will be initialized by the first device attached |
1769 | * to it. |
1770 | */ |
1771 | if (!mem) { |
1772 | struct io_tlb_pool *pool; |
1773 | |
1774 | mem = kzalloc(size: sizeof(*mem), GFP_KERNEL); |
1775 | if (!mem) |
1776 | return -ENOMEM; |
1777 | pool = &mem->defpool; |
1778 | |
1779 | pool->slots = kcalloc(n: nslabs, size: sizeof(*pool->slots), GFP_KERNEL); |
1780 | if (!pool->slots) { |
1781 | kfree(objp: mem); |
1782 | return -ENOMEM; |
1783 | } |
1784 | |
1785 | pool->areas = kcalloc(n: nareas, size: sizeof(*pool->areas), |
1786 | GFP_KERNEL); |
1787 | if (!pool->areas) { |
1788 | kfree(objp: pool->slots); |
1789 | kfree(objp: mem); |
1790 | return -ENOMEM; |
1791 | } |
1792 | |
1793 | set_memory_decrypted(addr: (unsigned long)phys_to_virt(address: rmem->base), |
1794 | numpages: rmem->size >> PAGE_SHIFT); |
1795 | swiotlb_init_io_tlb_pool(mem: pool, start: rmem->base, nslabs, |
1796 | late_alloc: false, nareas); |
1797 | mem->force_bounce = true; |
1798 | mem->for_alloc = true; |
1799 | #ifdef CONFIG_SWIOTLB_DYNAMIC |
1800 | spin_lock_init(&mem->lock); |
1801 | #endif |
1802 | add_mem_pool(mem, pool); |
1803 | |
1804 | rmem->priv = mem; |
1805 | |
1806 | swiotlb_create_debugfs_files(mem, dirname: rmem->name); |
1807 | } |
1808 | |
1809 | dev->dma_io_tlb_mem = mem; |
1810 | |
1811 | return 0; |
1812 | } |
1813 | |
1814 | static void rmem_swiotlb_device_release(struct reserved_mem *rmem, |
1815 | struct device *dev) |
1816 | { |
1817 | dev->dma_io_tlb_mem = &io_tlb_default_mem; |
1818 | } |
1819 | |
1820 | static const struct reserved_mem_ops rmem_swiotlb_ops = { |
1821 | .device_init = rmem_swiotlb_device_init, |
1822 | .device_release = rmem_swiotlb_device_release, |
1823 | }; |
1824 | |
1825 | static int __init rmem_swiotlb_setup(struct reserved_mem *rmem) |
1826 | { |
1827 | unsigned long node = rmem->fdt_node; |
1828 | |
1829 | if (of_get_flat_dt_prop(node, name: "reusable" , NULL) || |
1830 | of_get_flat_dt_prop(node, name: "linux,cma-default" , NULL) || |
1831 | of_get_flat_dt_prop(node, name: "linux,dma-default" , NULL) || |
1832 | of_get_flat_dt_prop(node, name: "no-map" , NULL)) |
1833 | return -EINVAL; |
1834 | |
1835 | rmem->ops = &rmem_swiotlb_ops; |
1836 | pr_info("Reserved memory: created restricted DMA pool at %pa, size %ld MiB\n" , |
1837 | &rmem->base, (unsigned long)rmem->size / SZ_1M); |
1838 | return 0; |
1839 | } |
1840 | |
1841 | RESERVEDMEM_OF_DECLARE(dma, "restricted-dma-pool" , rmem_swiotlb_setup); |
1842 | #endif /* CONFIG_DMA_RESTRICTED_POOL */ |
1843 | |