mm/dmapool.c

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