| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * DMA Pool allocator |
| 4 | * |
| 5 | * Copyright 2001 David Brownell |
| 6 | * Copyright 2007 Intel Corporation |
| 7 | * Author: Matthew Wilcox <willy@linux.intel.com> |
| 8 | * |
| 9 | * This allocator returns small blocks of a given size which are DMA-able by |
| 10 | * the given device. It uses the dma_alloc_coherent page allocator to get |
| 11 | * new pages, then splits them up into blocks of the required size. |
| 12 | * Many older drivers still have their own code to do this. |
| 13 | * |
| 14 | * The current design of this allocator is fairly simple. The pool is |
| 15 | * represented by the 'struct dma_pool' which keeps a doubly-linked list of |
| 16 | * allocated pages. Each page in the page_list is split into blocks of at |
| 17 | * least 'size' bytes. Free blocks are tracked in an unsorted singly-linked |
| 18 | * list of free blocks within the page. Used blocks aren't tracked, but we |
| 19 | * keep a count of how many are currently allocated from each page. |
| 20 | */ |
| 21 | |
| 22 | #include <linux/device.h> |
| 23 | #include <linux/dma-mapping.h> |
| 24 | #include <linux/dmapool.h> |
| 25 | #include <linux/kernel.h> |
| 26 | #include <linux/list.h> |
| 27 | #include <linux/export.h> |
| 28 | #include <linux/mutex.h> |
| 29 | #include <linux/poison.h> |
| 30 | #include <linux/sched.h> |
| 31 | #include <linux/slab.h> |
| 32 | #include <linux/stat.h> |
| 33 | #include <linux/spinlock.h> |
| 34 | #include <linux/string.h> |
| 35 | #include <linux/types.h> |
| 36 | #include <linux/wait.h> |
| 37 | |
| 38 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_SLUB_DEBUG_ON) |
| 39 | #define DMAPOOL_DEBUG 1 |
| 40 | #endif |
| 41 | |
| 42 | struct dma_pool { /* the pool */ |
| 43 | struct list_head page_list; |
| 44 | spinlock_t lock; |
| 45 | size_t size; |
| 46 | struct device *dev; |
| 47 | size_t allocation; |
| 48 | size_t boundary; |
| 49 | char name[32]; |
| 50 | struct list_head pools; |
| 51 | }; |
| 52 | |
| 53 | struct dma_page { /* cacheable header for 'allocation' bytes */ |
| 54 | struct list_head page_list; |
| 55 | void *vaddr; |
| 56 | dma_addr_t dma; |
| 57 | unsigned int in_use; |
| 58 | unsigned int offset; |
| 59 | }; |
| 60 | |
| 61 | static DEFINE_MUTEX(pools_lock); |
| 62 | static DEFINE_MUTEX(pools_reg_lock); |
| 63 | |
| 64 | static ssize_t |
| 65 | show_pools(struct device *dev, struct device_attribute *attr, char *buf) |
| 66 | { |
| 67 | unsigned temp; |
| 68 | unsigned size; |
| 69 | char *next; |
| 70 | struct dma_page *page; |
| 71 | struct dma_pool *pool; |
| 72 | |
| 73 | next = buf; |
| 74 | size = PAGE_SIZE; |
| 75 | |
| 76 | temp = scnprintf(next, size, "poolinfo - 0.1\n"); |
| 77 | size -= temp; |
| 78 | next += temp; |
| 79 | |
| 80 | mutex_lock(&pools_lock); |
| 81 | list_for_each_entry(pool, &dev->dma_pools, pools) { |
| 82 | unsigned pages = 0; |
| 83 | unsigned blocks = 0; |
| 84 | |
| 85 | spin_lock_irq(&pool->lock); |
| 86 | list_for_each_entry(page, &pool->page_list, page_list) { |
| 87 | pages++; |
| 88 | blocks += page->in_use; |
| 89 | } |
| 90 | spin_unlock_irq(&pool->lock); |
| 91 | |
| 92 | /* per-pool info, no real statistics yet */ |
| 93 | temp = scnprintf(next, size, "%-16s %4u %4zu %4zu %2u\n", |
| 94 | pool->name, blocks, |
| 95 | pages * (pool->allocation / pool->size), |
| 96 | pool->size, pages); |
| 97 | size -= temp; |
| 98 | next += temp; |
| 99 | } |
| 100 | mutex_unlock(&pools_lock); |
| 101 | |
| 102 | return PAGE_SIZE - size; |
| 103 | } |
| 104 | |
| 105 | static DEVICE_ATTR(pools, 0444, show_pools, NULL); |
| 106 | |
| 107 | /** |
| 108 | * dma_pool_create - Creates a pool of consistent memory blocks, for dma. |
| 109 | * @name: name of pool, for diagnostics |
| 110 | * @dev: device that will be doing the DMA |
| 111 | * @size: size of the blocks in this pool. |
| 112 | * @align: alignment requirement for blocks; must be a power of two |
| 113 | * @boundary: returned blocks won't cross this power of two boundary |
| 114 | * Context: not in_interrupt() |
| 115 | * |
| 116 | * Given one of these pools, dma_pool_alloc() |
| 117 | * may be used to allocate memory. Such memory will all have "consistent" |
| 118 | * DMA mappings, accessible by the device and its driver without using |
| 119 | * cache flushing primitives. The actual size of blocks allocated may be |
| 120 | * larger than requested because of alignment. |
| 121 | * |
| 122 | * If @boundary is nonzero, objects returned from dma_pool_alloc() won't |
| 123 | * cross that size boundary. This is useful for devices which have |
| 124 | * addressing restrictions on individual DMA transfers, such as not crossing |
| 125 | * boundaries of 4KBytes. |
| 126 | * |
| 127 | * Return: a dma allocation pool with the requested characteristics, or |
| 128 | * %NULL if one can't be created. |
| 129 | */ |
| 130 | struct dma_pool *dma_pool_create(const char *name, struct device *dev, |
| 131 | size_t size, size_t align, size_t boundary) |
| 132 | { |
| 133 | struct dma_pool *retval; |
| 134 | size_t allocation; |
| 135 | bool empty = false; |
| 136 | |
| 137 | if (align == 0) |
| 138 | align = 1; |
| 139 | else if (align & (align - 1)) |
| 140 | return NULL; |
| 141 | |
| 142 | if (size == 0) |
| 143 | return NULL; |
| 144 | else if (size < 4) |
| 145 | size = 4; |
| 146 | |
| 147 | if ((size % align) != 0) |
| 148 | size = ALIGN(size, align); |
| 149 | |
| 150 | allocation = max_t(size_t, size, PAGE_SIZE); |
| 151 | |
| 152 | if (!boundary) |
| 153 | boundary = allocation; |
| 154 | else if ((boundary < size) || (boundary & (boundary - 1))) |
| 155 | return NULL; |
| 156 | |
| 157 | retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev)); |
| 158 | if (!retval) |
| 159 | return retval; |
| 160 | |
| 161 | strlcpy(retval->name, name, sizeof(retval->name)); |
| 162 | |
| 163 | retval->dev = dev; |
| 164 | |
| 165 | INIT_LIST_HEAD(&retval->page_list); |
| 166 | spin_lock_init(&retval->lock); |
| 167 | retval->size = size; |
| 168 | retval->boundary = boundary; |
| 169 | retval->allocation = allocation; |
| 170 | |
| 171 | INIT_LIST_HEAD(&retval->pools); |
| 172 | |
| 173 | /* |
| 174 | * pools_lock ensures that the ->dma_pools list does not get corrupted. |
| 175 | * pools_reg_lock ensures that there is not a race between |
| 176 | * dma_pool_create() and dma_pool_destroy() or within dma_pool_create() |
| 177 | * when the first invocation of dma_pool_create() failed on |
| 178 | * device_create_file() and the second assumes that it has been done (I |
| 179 | * know it is a short window). |
| 180 | */ |
| 181 | mutex_lock(&pools_reg_lock); |
| 182 | mutex_lock(&pools_lock); |
| 183 | if (list_empty(&dev->dma_pools)) |
| 184 | empty = true; |
| 185 | list_add(&retval->pools, &dev->dma_pools); |
| 186 | mutex_unlock(&pools_lock); |
| 187 | if (empty) { |
| 188 | int err; |
| 189 | |
| 190 | err = device_create_file(dev, &dev_attr_pools); |
| 191 | if (err) { |
| 192 | mutex_lock(&pools_lock); |
| 193 | list_del(&retval->pools); |
| 194 | mutex_unlock(&pools_lock); |
| 195 | mutex_unlock(&pools_reg_lock); |
| 196 | kfree(retval); |
| 197 | return NULL; |
| 198 | } |
| 199 | } |
| 200 | mutex_unlock(&pools_reg_lock); |
| 201 | return retval; |
| 202 | } |
| 203 | EXPORT_SYMBOL(dma_pool_create); |
| 204 | |
| 205 | static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page) |
| 206 | { |
| 207 | unsigned int offset = 0; |
| 208 | unsigned int next_boundary = pool->boundary; |
| 209 | |
| 210 | do { |
| 211 | unsigned int next = offset + pool->size; |
| 212 | if (unlikely((next + pool->size) >= next_boundary)) { |
| 213 | next = next_boundary; |
| 214 | next_boundary += pool->boundary; |
| 215 | } |
| 216 | *(int *)(page->vaddr + offset) = next; |
| 217 | offset = next; |
| 218 | } while (offset < pool->allocation); |
| 219 | } |
| 220 | |
| 221 | static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags) |
| 222 | { |
| 223 | struct dma_page *page; |
| 224 | |
| 225 | page = kmalloc(sizeof(*page), mem_flags); |
| 226 | if (!page) |
| 227 | return NULL; |
| 228 | page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation, |
| 229 | &page->dma, mem_flags); |
| 230 | if (page->vaddr) { |
| 231 | #ifdef DMAPOOL_DEBUG |
| 232 | memset(page->vaddr, POOL_POISON_FREED, pool->allocation); |
| 233 | #endif |
| 234 | pool_initialise_page(pool, page); |
| 235 | page->in_use = 0; |
| 236 | page->offset = 0; |
| 237 | } else { |
| 238 | kfree(page); |
| 239 | page = NULL; |
| 240 | } |
| 241 | return page; |
| 242 | } |
| 243 | |
| 244 | static inline bool is_page_busy(struct dma_page *page) |
| 245 | { |
| 246 | return page->in_use != 0; |
| 247 | } |
| 248 | |
| 249 | static void pool_free_page(struct dma_pool *pool, struct dma_page *page) |
| 250 | { |
| 251 | dma_addr_t dma = page->dma; |
| 252 | |
| 253 | #ifdef DMAPOOL_DEBUG |
| 254 | memset(page->vaddr, POOL_POISON_FREED, pool->allocation); |
| 255 | #endif |
| 256 | dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma); |
| 257 | list_del(&page->page_list); |
| 258 | kfree(page); |
| 259 | } |
| 260 | |
| 261 | /** |
| 262 | * dma_pool_destroy - destroys a pool of dma memory blocks. |
| 263 | * @pool: dma pool that will be destroyed |
| 264 | * Context: !in_interrupt() |
| 265 | * |
| 266 | * Caller guarantees that no more memory from the pool is in use, |
| 267 | * and that nothing will try to use the pool after this call. |
| 268 | */ |
| 269 | void dma_pool_destroy(struct dma_pool *pool) |
| 270 | { |
| 271 | bool empty = false; |
| 272 | |
| 273 | if (unlikely(!pool)) |
| 274 | return; |
| 275 | |
| 276 | mutex_lock(&pools_reg_lock); |
| 277 | mutex_lock(&pools_lock); |
| 278 | list_del(&pool->pools); |
| 279 | if (pool->dev && list_empty(&pool->dev->dma_pools)) |
| 280 | empty = true; |
| 281 | mutex_unlock(&pools_lock); |
| 282 | if (empty) |
| 283 | device_remove_file(pool->dev, &dev_attr_pools); |
| 284 | mutex_unlock(&pools_reg_lock); |
| 285 | |
| 286 | while (!list_empty(&pool->page_list)) { |
| 287 | struct dma_page *page; |
| 288 | page = list_entry(pool->page_list.next, |
| 289 | struct dma_page, page_list); |
| 290 | if (is_page_busy(page)) { |
| 291 | if (pool->dev) |
| 292 | dev_err(pool->dev, |
| 293 | "dma_pool_destroy %s, %p busy\n", |
| 294 | pool->name, page->vaddr); |
| 295 | else |
| 296 | pr_err("dma_pool_destroy %s, %p busy\n", |
| 297 | pool->name, page->vaddr); |
| 298 | /* leak the still-in-use consistent memory */ |
| 299 | list_del(&page->page_list); |
| 300 | kfree(page); |
| 301 | } else |
| 302 | pool_free_page(pool, page); |
| 303 | } |
| 304 | |
| 305 | kfree(pool); |
| 306 | } |
| 307 | EXPORT_SYMBOL(dma_pool_destroy); |
| 308 | |
| 309 | /** |
| 310 | * dma_pool_alloc - get a block of consistent memory |
| 311 | * @pool: dma pool that will produce the block |
| 312 | * @mem_flags: GFP_* bitmask |
| 313 | * @handle: pointer to dma address of block |
| 314 | * |
| 315 | * Return: the kernel virtual address of a currently unused block, |
| 316 | * and reports its dma address through the handle. |
| 317 | * If such a memory block can't be allocated, %NULL is returned. |
| 318 | */ |
| 319 | void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags, |
| 320 | dma_addr_t *handle) |
| 321 | { |
| 322 | unsigned long flags; |
| 323 | struct dma_page *page; |
| 324 | size_t offset; |
| 325 | void *retval; |
| 326 | |
| 327 | might_sleep_if(gfpflags_allow_blocking(mem_flags)); |
| 328 | |
| 329 | spin_lock_irqsave(&pool->lock, flags); |
| 330 | list_for_each_entry(page, &pool->page_list, page_list) { |
| 331 | if (page->offset < pool->allocation) |
| 332 | goto ready; |
| 333 | } |
| 334 | |
| 335 | /* pool_alloc_page() might sleep, so temporarily drop &pool->lock */ |
| 336 | spin_unlock_irqrestore(&pool->lock, flags); |
| 337 | |
| 338 | page = pool_alloc_page(pool, mem_flags & (~__GFP_ZERO)); |
| 339 | if (!page) |
| 340 | return NULL; |
| 341 | |
| 342 | spin_lock_irqsave(&pool->lock, flags); |
| 343 | |
| 344 | list_add(&page->page_list, &pool->page_list); |
| 345 | ready: |
| 346 | page->in_use++; |
| 347 | offset = page->offset; |
| 348 | page->offset = *(int *)(page->vaddr + offset); |
| 349 | retval = offset + page->vaddr; |
| 350 | *handle = offset + page->dma; |
| 351 | #ifdef DMAPOOL_DEBUG |
| 352 | { |
| 353 | int i; |
| 354 | u8 *data = retval; |
| 355 | /* page->offset is stored in first 4 bytes */ |
| 356 | for (i = sizeof(page->offset); i < pool->size; i++) { |
| 357 | if (data[i] == POOL_POISON_FREED) |
| 358 | continue; |
| 359 | if (pool->dev) |
| 360 | dev_err(pool->dev, |
| 361 | "dma_pool_alloc %s, %p (corrupted)\n", |
| 362 | pool->name, retval); |
| 363 | else |
| 364 | pr_err("dma_pool_alloc %s, %p (corrupted)\n", |
| 365 | pool->name, retval); |
| 366 | |
| 367 | /* |
| 368 | * Dump the first 4 bytes even if they are not |
| 369 | * POOL_POISON_FREED |
| 370 | */ |
| 371 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, |
| 372 | data, pool->size, 1); |
| 373 | break; |
| 374 | } |
| 375 | } |
| 376 | if (!(mem_flags & __GFP_ZERO)) |
| 377 | memset(retval, POOL_POISON_ALLOCATED, pool->size); |
| 378 | #endif |
| 379 | spin_unlock_irqrestore(&pool->lock, flags); |
| 380 | |
| 381 | if (mem_flags & __GFP_ZERO) |
| 382 | memset(retval, 0, pool->size); |
| 383 | |
| 384 | return retval; |
| 385 | } |
| 386 | EXPORT_SYMBOL(dma_pool_alloc); |
| 387 | |
| 388 | static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma) |
| 389 | { |
| 390 | struct dma_page *page; |
| 391 | |
| 392 | list_for_each_entry(page, &pool->page_list, page_list) { |
| 393 | if (dma < page->dma) |
| 394 | continue; |
| 395 | if ((dma - page->dma) < pool->allocation) |
| 396 | return page; |
| 397 | } |
| 398 | return NULL; |
| 399 | } |
| 400 | |
| 401 | /** |
| 402 | * dma_pool_free - put block back into dma pool |
| 403 | * @pool: the dma pool holding the block |
| 404 | * @vaddr: virtual address of block |
| 405 | * @dma: dma address of block |
| 406 | * |
| 407 | * Caller promises neither device nor driver will again touch this block |
| 408 | * unless it is first re-allocated. |
| 409 | */ |
| 410 | void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma) |
| 411 | { |
| 412 | struct dma_page *page; |
| 413 | unsigned long flags; |
| 414 | unsigned int offset; |
| 415 | |
| 416 | spin_lock_irqsave(&pool->lock, flags); |
| 417 | page = pool_find_page(pool, dma); |
| 418 | if (!page) { |
| 419 | spin_unlock_irqrestore(&pool->lock, flags); |
| 420 | if (pool->dev) |
| 421 | dev_err(pool->dev, |
| 422 | "dma_pool_free %s, %p/%lx (bad dma)\n", |
| 423 | pool->name, vaddr, (unsigned long)dma); |
| 424 | else |
| 425 | pr_err("dma_pool_free %s, %p/%lx (bad dma)\n", |
| 426 | pool->name, vaddr, (unsigned long)dma); |
| 427 | return; |
| 428 | } |
| 429 | |
| 430 | offset = vaddr - page->vaddr; |
| 431 | #ifdef DMAPOOL_DEBUG |
| 432 | if ((dma - page->dma) != offset) { |
| 433 | spin_unlock_irqrestore(&pool->lock, flags); |
| 434 | if (pool->dev) |
| 435 | dev_err(pool->dev, |
| 436 | "dma_pool_free %s, %p (bad vaddr)/%pad\n", |
| 437 | pool->name, vaddr, &dma); |
| 438 | else |
| 439 | pr_err("dma_pool_free %s, %p (bad vaddr)/%pad\n", |
| 440 | pool->name, vaddr, &dma); |
| 441 | return; |
| 442 | } |
| 443 | { |
| 444 | unsigned int chain = page->offset; |
| 445 | while (chain < pool->allocation) { |
| 446 | if (chain != offset) { |
| 447 | chain = *(int *)(page->vaddr + chain); |
| 448 | continue; |
| 449 | } |
| 450 | spin_unlock_irqrestore(&pool->lock, flags); |
| 451 | if (pool->dev) |
| 452 | dev_err(pool->dev, "dma_pool_free %s, dma %pad already free\n", |
| 453 | pool->name, &dma); |
| 454 | else |
| 455 | pr_err("dma_pool_free %s, dma %pad already free\n", |
| 456 | pool->name, &dma); |
| 457 | return; |
| 458 | } |
| 459 | } |
| 460 | memset(vaddr, POOL_POISON_FREED, pool->size); |
| 461 | #endif |
| 462 | |
| 463 | page->in_use--; |
| 464 | *(int *)vaddr = page->offset; |
| 465 | page->offset = offset; |
| 466 | /* |
| 467 | * Resist a temptation to do |
| 468 | * if (!is_page_busy(page)) pool_free_page(pool, page); |
| 469 | * Better have a few empty pages hang around. |
| 470 | */ |
| 471 | spin_unlock_irqrestore(&pool->lock, flags); |
| 472 | } |
| 473 | EXPORT_SYMBOL(dma_pool_free); |
| 474 | |
| 475 | /* |
| 476 | * Managed DMA pool |
| 477 | */ |
| 478 | static void dmam_pool_release(struct device *dev, void *res) |
| 479 | { |
| 480 | struct dma_pool *pool = *(struct dma_pool **)res; |
| 481 | |
| 482 | dma_pool_destroy(pool); |
| 483 | } |
| 484 | |
| 485 | static int dmam_pool_match(struct device *dev, void *res, void *match_data) |
| 486 | { |
| 487 | return *(struct dma_pool **)res == match_data; |
| 488 | } |
| 489 | |
| 490 | /** |
| 491 | * dmam_pool_create - Managed dma_pool_create() |
| 492 | * @name: name of pool, for diagnostics |
| 493 | * @dev: device that will be doing the DMA |
| 494 | * @size: size of the blocks in this pool. |
| 495 | * @align: alignment requirement for blocks; must be a power of two |
| 496 | * @allocation: returned blocks won't cross this boundary (or zero) |
| 497 | * |
| 498 | * Managed dma_pool_create(). DMA pool created with this function is |
| 499 | * automatically destroyed on driver detach. |
| 500 | * |
| 501 | * Return: a managed dma allocation pool with the requested |
| 502 | * characteristics, or %NULL if one can't be created. |
| 503 | */ |
| 504 | struct dma_pool *dmam_pool_create(const char *name, struct device *dev, |
| 505 | size_t size, size_t align, size_t allocation) |
| 506 | { |
| 507 | struct dma_pool **ptr, *pool; |
| 508 | |
| 509 | ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL); |
| 510 | if (!ptr) |
| 511 | return NULL; |
| 512 | |
| 513 | pool = *ptr = dma_pool_create(name, dev, size, align, allocation); |
| 514 | if (pool) |
| 515 | devres_add(dev, ptr); |
| 516 | else |
| 517 | devres_free(ptr); |
| 518 | |
| 519 | return pool; |
| 520 | } |
| 521 | EXPORT_SYMBOL(dmam_pool_create); |
| 522 | |
| 523 | /** |
| 524 | * dmam_pool_destroy - Managed dma_pool_destroy() |
| 525 | * @pool: dma pool that will be destroyed |
| 526 | * |
| 527 | * Managed dma_pool_destroy(). |
| 528 | */ |
| 529 | void dmam_pool_destroy(struct dma_pool *pool) |
| 530 | { |
| 531 | struct device *dev = pool->dev; |
| 532 | |
| 533 | WARN_ON(devres_release(dev, dmam_pool_release, dmam_pool_match, pool)); |
| 534 | } |
| 535 | EXPORT_SYMBOL(dmam_pool_destroy); |