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/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         size = ALIGN(size, align);
 148         allocation = max_t(size_t, size, PAGE_SIZE);
 149
 150         if (!boundary)
 151                 boundary = allocation;
 152         else if ((boundary < size) || (boundary & (boundary - 1)))
 153                 return NULL;
 154
 155         retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
 156         if (!retval)
 157                 return retval;
 158
 159         strlcpy(retval->name, name, sizeof(retval->name));
 160
 161         retval->dev = dev;
 162
 163         INIT_LIST_HEAD(&retval->page_list);
 164         spin_lock_init(&retval->lock);
 165         retval->size = size;
 166         retval->boundary = boundary;
 167         retval->allocation = allocation;
 168
 169         INIT_LIST_HEAD(&retval->pools);
 170
 171         /*
 172          * pools_lock ensures that the ->dma_pools list does not get corrupted.
 173          * pools_reg_lock ensures that there is not a race between
 174          * dma_pool_create() and dma_pool_destroy() or within dma_pool_create()
 175          * when the first invocation of dma_pool_create() failed on
 176          * device_create_file() and the second assumes that it has been done (I
 177          * know it is a short window).
 178          */
 179         mutex_lock(&pools_reg_lock);
 180         mutex_lock(&pools_lock);
 181         if (list_empty(&dev->dma_pools))
 182                 empty = true;
 183         list_add(&retval->pools, &dev->dma_pools);
 184         mutex_unlock(&pools_lock);
 185         if (empty) {
 186                 int err;
 187
 188                 err = device_create_file(dev, &dev_attr_pools);
 189                 if (err) {
 190                         mutex_lock(&pools_lock);
 191                         list_del(&retval->pools);
 192                         mutex_unlock(&pools_lock);
 193                         mutex_unlock(&pools_reg_lock);
 194                         kfree(retval);
 195                         return NULL;
 196                 }
 197         }
 198         mutex_unlock(&pools_reg_lock);
 199         return retval;
 200 }
 201 EXPORT_SYMBOL(dma_pool_create);
 202
 203 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
 204 {
 205         unsigned int offset = 0;
 206         unsigned int next_boundary = pool->boundary;
 207
 208         do {
 209                 unsigned int next = offset + pool->size;
 210                 if (unlikely((next + pool->size) >= next_boundary)) {
 211                         next = next_boundary;
 212                         next_boundary += pool->boundary;
 213                 }
 214                 *(int *)(page->vaddr + offset) = next;
 215                 offset = next;
 216         } while (offset < pool->allocation);
 217 }
 218
 219 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
 220 {
 221         struct dma_page *page;
 222
 223         page = kmalloc(sizeof(*page), mem_flags);
 224         if (!page)
 225                 return NULL;
 226         page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
 227                                          &page->dma, mem_flags);
 228         if (page->vaddr) {
 229 #ifdef  DMAPOOL_DEBUG
 230                 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
 231 #endif
 232                 pool_initialise_page(pool, page);
 233                 page->in_use = 0;
 234                 page->offset = 0;
 235         } else {
 236                 kfree(page);
 237                 page = NULL;
 238         }
 239         return page;
 240 }
 241
 242 static inline bool is_page_busy(struct dma_page *page)
 243 {
 244         return page->in_use != 0;
 245 }
 246
 247 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
 248 {
 249         dma_addr_t dma = page->dma;
 250
 251 #ifdef  DMAPOOL_DEBUG
 252         memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
 253 #endif
 254         dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma);
 255         list_del(&page->page_list);
 256         kfree(page);
 257 }
 258
 259 /**
 260  * dma_pool_destroy - destroys a pool of dma memory blocks.
 261  * @pool: dma pool that will be destroyed
 262  * Context: !in_interrupt()
 263  *
 264  * Caller guarantees that no more memory from the pool is in use,
 265  * and that nothing will try to use the pool after this call.
 266  */
 267 void dma_pool_destroy(struct dma_pool *pool)
 268 {
 269         struct dma_page *page, *tmp;
 270         bool empty = false;
 271
 272         if (unlikely(!pool))
 273                 return;
 274
 275         mutex_lock(&pools_reg_lock);
 276         mutex_lock(&pools_lock);
 277         list_del(&pool->pools);
 278         if (pool->dev && list_empty(&pool->dev->dma_pools))
 279                 empty = true;
 280         mutex_unlock(&pools_lock);
 281         if (empty)
 282                 device_remove_file(pool->dev, &dev_attr_pools);
 283         mutex_unlock(&pools_reg_lock);
 284
 285         list_for_each_entry_safe(page, tmp, &pool->page_list, page_list) {
 286                 if (is_page_busy(page)) {
 287                         if (pool->dev)
 288                                 dev_err(pool->dev, "%s %s, %p busy\n", __func__,
 289                                         pool->name, page->vaddr);
 290                         else
 291                                 pr_err("%s %s, %p busy\n", __func__,
 292                                        pool->name, page->vaddr);
 293                         /* leak the still-in-use consistent memory */
 294                         list_del(&page->page_list);
 295                         kfree(page);
 296                 } else
 297                         pool_free_page(pool, page);
 298         }
 299
 300         kfree(pool);
 301 }
 302 EXPORT_SYMBOL(dma_pool_destroy);
 303
 304 /**
 305  * dma_pool_alloc - get a block of consistent memory
 306  * @pool: dma pool that will produce the block
 307  * @mem_flags: GFP_* bitmask
 308  * @handle: pointer to dma address of block
 309  *
 310  * Return: the kernel virtual address of a currently unused block,
 311  * and reports its dma address through the handle.
 312  * If such a memory block can't be allocated, %NULL is returned.
 313  */
 314 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
 315                      dma_addr_t *handle)
 316 {
 317         unsigned long flags;
 318         struct dma_page *page;
 319         size_t offset;
 320         void *retval;
 321
 322         might_sleep_if(gfpflags_allow_blocking(mem_flags));
 323
 324         spin_lock_irqsave(&pool->lock, flags);
 325         list_for_each_entry(page, &pool->page_list, page_list) {
 326                 if (page->offset < pool->allocation)
 327                         goto ready;
 328         }
 329
 330         /* pool_alloc_page() might sleep, so temporarily drop &pool->lock */
 331         spin_unlock_irqrestore(&pool->lock, flags);
 332
 333         page = pool_alloc_page(pool, mem_flags & (~__GFP_ZERO));
 334         if (!page)
 335                 return NULL;
 336
 337         spin_lock_irqsave(&pool->lock, flags);
 338
 339         list_add(&page->page_list, &pool->page_list);
 340  ready:
 341         page->in_use++;
 342         offset = page->offset;
 343         page->offset = *(int *)(page->vaddr + offset);
 344         retval = offset + page->vaddr;
 345         *handle = offset + page->dma;
 346 #ifdef  DMAPOOL_DEBUG
 347         {
 348                 int i;
 349                 u8 *data = retval;
 350                 /* page->offset is stored in first 4 bytes */
 351                 for (i = sizeof(page->offset); i < pool->size; i++) {
 352                         if (data[i] == POOL_POISON_FREED)
 353                                 continue;
 354                         if (pool->dev)
 355                                 dev_err(pool->dev, "%s %s, %p (corrupted)\n",
 356                                         __func__, pool->name, retval);
 357                         else
 358                                 pr_err("%s %s, %p (corrupted)\n",
 359                                         __func__, pool->name, retval);
 360
 361                         /*
 362                          * Dump the first 4 bytes even if they are not
 363                          * POOL_POISON_FREED
 364                          */
 365                         print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1,
 366                                         data, pool->size, 1);
 367                         break;
 368                 }
 369         }
 370         if (!(mem_flags & __GFP_ZERO))
 371                 memset(retval, POOL_POISON_ALLOCATED, pool->size);
 372 #endif
 373         spin_unlock_irqrestore(&pool->lock, flags);
 374
 375         if (want_init_on_alloc(mem_flags))
 376                 memset(retval, 0, pool->size);
 377
 378         return retval;
 379 }
 380 EXPORT_SYMBOL(dma_pool_alloc);
 381
 382 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
 383 {
 384         struct dma_page *page;
 385
 386         list_for_each_entry(page, &pool->page_list, page_list) {
 387                 if (dma < page->dma)
 388                         continue;
 389                 if ((dma - page->dma) < pool->allocation)
 390                         return page;
 391         }
 392         return NULL;
 393 }
 394
 395 /**
 396  * dma_pool_free - put block back into dma pool
 397  * @pool: the dma pool holding the block
 398  * @vaddr: virtual address of block
 399  * @dma: dma address of block
 400  *
 401  * Caller promises neither device nor driver will again touch this block
 402  * unless it is first re-allocated.
 403  */
 404 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
 405 {
 406         struct dma_page *page;
 407         unsigned long flags;
 408         unsigned int offset;
 409
 410         spin_lock_irqsave(&pool->lock, flags);
 411         page = pool_find_page(pool, dma);
 412         if (!page) {
 413                 spin_unlock_irqrestore(&pool->lock, flags);
 414                 if (pool->dev)
 415                         dev_err(pool->dev, "%s %s, %p/%pad (bad dma)\n",
 416                                 __func__, pool->name, vaddr, &dma);
 417                 else
 418                         pr_err("%s %s, %p/%pad (bad dma)\n",
 419                                __func__, pool->name, vaddr, &dma);
 420                 return;
 421         }
 422
 423         offset = vaddr - page->vaddr;
 424         if (want_init_on_free())
 425                 memset(vaddr, 0, pool->size);
 426 #ifdef  DMAPOOL_DEBUG
 427         if ((dma - page->dma) != offset) {
 428                 spin_unlock_irqrestore(&pool->lock, flags);
 429                 if (pool->dev)
 430                         dev_err(pool->dev, "%s %s, %p (bad vaddr)/%pad\n",
 431                                 __func__, pool->name, vaddr, &dma);
 432                 else
 433                         pr_err("%s %s, %p (bad vaddr)/%pad\n",
 434                                __func__, pool->name, vaddr, &dma);
 435                 return;
 436         }
 437         {
 438                 unsigned int chain = page->offset;
 439                 while (chain < pool->allocation) {
 440                         if (chain != offset) {
 441                                 chain = *(int *)(page->vaddr + chain);
 442                                 continue;
 443                         }
 444                         spin_unlock_irqrestore(&pool->lock, flags);
 445                         if (pool->dev)
 446                                 dev_err(pool->dev, "%s %s, dma %pad already free\n",
 447                                         __func__, pool->name, &dma);
 448                         else
 449                                 pr_err("%s %s, dma %pad already free\n",
 450                                        __func__, pool->name, &dma);
 451                         return;
 452                 }
 453         }
 454         memset(vaddr, POOL_POISON_FREED, pool->size);
 455 #endif
 456
 457         page->in_use--;
 458         *(int *)vaddr = page->offset;
 459         page->offset = offset;
 460         /*
 461          * Resist a temptation to do
 462          *    if (!is_page_busy(page)) pool_free_page(pool, page);
 463          * Better have a few empty pages hang around.
 464          */
 465         spin_unlock_irqrestore(&pool->lock, flags);
 466 }
 467 EXPORT_SYMBOL(dma_pool_free);
 468
 469 /*
 470  * Managed DMA pool
 471  */
 472 static void dmam_pool_release(struct device *dev, void *res)
 473 {
 474         struct dma_pool *pool = *(struct dma_pool **)res;
 475
 476         dma_pool_destroy(pool);
 477 }
 478
 479 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
 480 {
 481         return *(struct dma_pool **)res == match_data;
 482 }
 483
 484 /**
 485  * dmam_pool_create - Managed dma_pool_create()
 486  * @name: name of pool, for diagnostics
 487  * @dev: device that will be doing the DMA
 488  * @size: size of the blocks in this pool.
 489  * @align: alignment requirement for blocks; must be a power of two
 490  * @allocation: returned blocks won't cross this boundary (or zero)
 491  *
 492  * Managed dma_pool_create().  DMA pool created with this function is
 493  * automatically destroyed on driver detach.
 494  *
 495  * Return: a managed dma allocation pool with the requested
 496  * characteristics, or %NULL if one can't be created.
 497  */
 498 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
 499                                   size_t size, size_t align, size_t allocation)
 500 {
 501         struct dma_pool **ptr, *pool;
 502
 503         ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
 504         if (!ptr)
 505                 return NULL;
 506
 507         pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
 508         if (pool)
 509                 devres_add(dev, ptr);
 510         else
 511                 devres_free(ptr);
 512
 513         return pool;
 514 }
 515 EXPORT_SYMBOL(dmam_pool_create);
 516
 517 /**
 518  * dmam_pool_destroy - Managed dma_pool_destroy()
 519  * @pool: dma pool that will be destroyed
 520  *
 521  * Managed dma_pool_destroy().
 522  */
 523 void dmam_pool_destroy(struct dma_pool *pool)
 524 {
 525         struct device *dev = pool->dev;
 526
 527         WARN_ON(devres_release(dev, dmam_pool_release, dmam_pool_match, pool));
 528 }
 529 EXPORT_SYMBOL(dmam_pool_destroy);