mm/mempool.c

   1 /*
   2  *  linux/mm/mempool.c
   3  *
   4  *  memory buffer pool support. Such pools are mostly used
   5  *  for guaranteed, deadlock-free memory allocations during
   6  *  extreme VM load.
   7  *
   8  *  started by Ingo Molnar, Copyright (C) 2001
   9  *  debugging by David Rientjes, Copyright (C) 2015
  10  */
  11
  12 #include <linux/mm.h>
  13 #include <linux/slab.h>
  14 #include <linux/highmem.h>
  15 #include <linux/kasan.h>
  16 #include <linux/kmemleak.h>
  17 #include <linux/export.h>
  18 #include <linux/mempool.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/writeback.h>
  21 #include "slab.h"
  22
  23 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_SLUB_DEBUG_ON)
  24 static void poison_error(mempool_t *pool, void *element, size_t size,
  25                          size_t byte)
  26 {
  27         const int nr = pool->curr_nr;
  28         const int start = max_t(int, byte - (BITS_PER_LONG / 8), 0);
  29         const int end = min_t(int, byte + (BITS_PER_LONG / 8), size);
  30         int i;
  31
  32         pr_err("BUG: mempool element poison mismatch\n");
  33         pr_err("Mempool %p size %zu\n", pool, size);
  34         pr_err(" nr=%d @ %p: %s0x", nr, element, start > 0 ? "... " : "");
  35         for (i = start; i < end; i++)
  36                 pr_cont("%x ", *(u8 *)(element + i));
  37         pr_cont("%s\n", end < size ? "..." : "");
  38         dump_stack();
  39 }
  40
  41 static void __check_element(mempool_t *pool, void *element, size_t size)
  42 {
  43         u8 *obj = element;
  44         size_t i;
  45
  46         for (i = 0; i < size; i++) {
  47                 u8 exp = (i < size - 1) ? POISON_FREE : POISON_END;
  48
  49                 if (obj[i] != exp) {
  50                         poison_error(pool, element, size, i);
  51                         return;
  52                 }
  53         }
  54         memset(obj, POISON_INUSE, size);
  55 }
  56
  57 static void check_element(mempool_t *pool, void *element)
  58 {
  59         /* Mempools backed by slab allocator */
  60         if (pool->free == mempool_free_slab || pool->free == mempool_kfree)
  61                 __check_element(pool, element, ksize(element));
  62
  63         /* Mempools backed by page allocator */
  64         if (pool->free == mempool_free_pages) {
  65                 int order = (int)(long)pool->pool_data;
  66                 void *addr = kmap_atomic((struct page *)element);
  67
  68                 __check_element(pool, addr, 1UL << (PAGE_SHIFT + order));
  69                 kunmap_atomic(addr);
  70         }
  71 }
  72
  73 static void __poison_element(void *element, size_t size)
  74 {
  75         u8 *obj = element;
  76
  77         memset(obj, POISON_FREE, size - 1);
  78         obj[size - 1] = POISON_END;
  79 }
  80
  81 static void poison_element(mempool_t *pool, void *element)
  82 {
  83         /* Mempools backed by slab allocator */
  84         if (pool->alloc == mempool_alloc_slab || pool->alloc == mempool_kmalloc)
  85                 __poison_element(element, ksize(element));
  86
  87         /* Mempools backed by page allocator */
  88         if (pool->alloc == mempool_alloc_pages) {
  89                 int order = (int)(long)pool->pool_data;
  90                 void *addr = kmap_atomic((struct page *)element);
  91
  92                 __poison_element(addr, 1UL << (PAGE_SHIFT + order));
  93                 kunmap_atomic(addr);
  94         }
  95 }
  96 #else /* CONFIG_DEBUG_SLAB || CONFIG_SLUB_DEBUG_ON */
  97 static inline void check_element(mempool_t *pool, void *element)
  98 {
  99 }
 100 static inline void poison_element(mempool_t *pool, void *element)
 101 {
 102 }
 103 #endif /* CONFIG_DEBUG_SLAB || CONFIG_SLUB_DEBUG_ON */
 104
 105 static void kasan_poison_element(mempool_t *pool, void *element)
 106 {
 107         if (pool->alloc == mempool_alloc_slab || pool->alloc == mempool_kmalloc)
 108                 kasan_poison_kfree(element);
 109         if (pool->alloc == mempool_alloc_pages)
 110                 kasan_free_pages(element, (unsigned long)pool->pool_data);
 111 }
 112
 113 static void kasan_unpoison_element(mempool_t *pool, void *element, gfp_t flags)
 114 {
 115         if (pool->alloc == mempool_alloc_slab || pool->alloc == mempool_kmalloc)
 116                 kasan_unpoison_slab(element);
 117         if (pool->alloc == mempool_alloc_pages)
 118                 kasan_alloc_pages(element, (unsigned long)pool->pool_data);
 119 }
 120
 121 static void add_element(mempool_t *pool, void *element)
 122 {
 123         BUG_ON(pool->curr_nr >= pool->min_nr);
 124         poison_element(pool, element);
 125         kasan_poison_element(pool, element);
 126         pool->elements[pool->curr_nr++] = element;
 127 }
 128
 129 static void *remove_element(mempool_t *pool, gfp_t flags)
 130 {
 131         void *element = pool->elements[--pool->curr_nr];
 132
 133         BUG_ON(pool->curr_nr < 0);
 134         kasan_unpoison_element(pool, element, flags);
 135         check_element(pool, element);
 136         return element;
 137 }
 138
 139 /**
 140  * mempool_destroy - deallocate a memory pool
 141  * @pool:      pointer to the memory pool which was allocated via
 142  *             mempool_create().
 143  *
 144  * Free all reserved elements in @pool and @pool itself.  This function
 145  * only sleeps if the free_fn() function sleeps.
 146  */
 147 void mempool_destroy(mempool_t *pool)
 148 {
 149         if (unlikely(!pool))
 150                 return;
 151
 152         while (pool->curr_nr) {
 153                 void *element = remove_element(pool, GFP_KERNEL);
 154                 pool->free(element, pool->pool_data);
 155         }
 156         kfree(pool->elements);
 157         kfree(pool);
 158 }
 159 EXPORT_SYMBOL(mempool_destroy);
 160
 161 /**
 162  * mempool_create - create a memory pool
 163  * @min_nr:    the minimum number of elements guaranteed to be
 164  *             allocated for this pool.
 165  * @alloc_fn:  user-defined element-allocation function.
 166  * @free_fn:   user-defined element-freeing function.
 167  * @pool_data: optional private data available to the user-defined functions.
 168  *
 169  * this function creates and allocates a guaranteed size, preallocated
 170  * memory pool. The pool can be used from the mempool_alloc() and mempool_free()
 171  * functions. This function might sleep. Both the alloc_fn() and the free_fn()
 172  * functions might sleep - as long as the mempool_alloc() function is not called
 173  * from IRQ contexts.
 174  */
 175 mempool_t *mempool_create(int min_nr, mempool_alloc_t *alloc_fn,
 176                                 mempool_free_t *free_fn, void *pool_data)
 177 {
 178         return mempool_create_node(min_nr,alloc_fn,free_fn, pool_data,
 179                                    GFP_KERNEL, NUMA_NO_NODE);
 180 }
 181 EXPORT_SYMBOL(mempool_create);
 182
 183 mempool_t *mempool_create_node(int min_nr, mempool_alloc_t *alloc_fn,
 184                                mempool_free_t *free_fn, void *pool_data,
 185                                gfp_t gfp_mask, int node_id)
 186 {
 187         mempool_t *pool;
 188         pool = kzalloc_node(sizeof(*pool), gfp_mask, node_id);
 189         if (!pool)
 190                 return NULL;
 191         pool->elements = kmalloc_node(min_nr * sizeof(void *),
 192                                       gfp_mask, node_id);
 193         if (!pool->elements) {
 194                 kfree(pool);
 195                 return NULL;
 196         }
 197         spin_lock_init(&pool->lock);
 198         pool->min_nr = min_nr;
 199         pool->pool_data = pool_data;
 200         init_waitqueue_head(&pool->wait);
 201         pool->alloc = alloc_fn;
 202         pool->free = free_fn;
 203
 204         /*
 205          * First pre-allocate the guaranteed number of buffers.
 206          */
 207         while (pool->curr_nr < pool->min_nr) {
 208                 void *element;
 209
 210                 element = pool->alloc(gfp_mask, pool->pool_data);
 211                 if (unlikely(!element)) {
 212                         mempool_destroy(pool);
 213                         return NULL;
 214                 }
 215                 add_element(pool, element);
 216         }
 217         return pool;
 218 }
 219 EXPORT_SYMBOL(mempool_create_node);
 220
 221 /**
 222  * mempool_resize - resize an existing memory pool
 223  * @pool:       pointer to the memory pool which was allocated via
 224  *              mempool_create().
 225  * @new_min_nr: the new minimum number of elements guaranteed to be
 226  *              allocated for this pool.
 227  *
 228  * This function shrinks/grows the pool. In the case of growing,
 229  * it cannot be guaranteed that the pool will be grown to the new
 230  * size immediately, but new mempool_free() calls will refill it.
 231  * This function may sleep.
 232  *
 233  * Note, the caller must guarantee that no mempool_destroy is called
 234  * while this function is running. mempool_alloc() & mempool_free()
 235  * might be called (eg. from IRQ contexts) while this function executes.
 236  */
 237 int mempool_resize(mempool_t *pool, int new_min_nr)
 238 {
 239         void *element;
 240         void **new_elements;
 241         unsigned long flags;
 242
 243         BUG_ON(new_min_nr <= 0);
 244         might_sleep();
 245
 246         spin_lock_irqsave(&pool->lock, flags);
 247         if (new_min_nr <= pool->min_nr) {
 248                 while (new_min_nr < pool->curr_nr) {
 249                         element = remove_element(pool, GFP_KERNEL);
 250                         spin_unlock_irqrestore(&pool->lock, flags);
 251                         pool->free(element, pool->pool_data);
 252                         spin_lock_irqsave(&pool->lock, flags);
 253                 }
 254                 pool->min_nr = new_min_nr;
 255                 goto out_unlock;
 256         }
 257         spin_unlock_irqrestore(&pool->lock, flags);
 258
 259         /* Grow the pool */
 260         new_elements = kmalloc_array(new_min_nr, sizeof(*new_elements),
 261                                      GFP_KERNEL);
 262         if (!new_elements)
 263                 return -ENOMEM;
 264
 265         spin_lock_irqsave(&pool->lock, flags);
 266         if (unlikely(new_min_nr <= pool->min_nr)) {
 267                 /* Raced, other resize will do our work */
 268                 spin_unlock_irqrestore(&pool->lock, flags);
 269                 kfree(new_elements);
 270                 goto out;
 271         }
 272         memcpy(new_elements, pool->elements,
 273                         pool->curr_nr * sizeof(*new_elements));
 274         kfree(pool->elements);
 275         pool->elements = new_elements;
 276         pool->min_nr = new_min_nr;
 277
 278         while (pool->curr_nr < pool->min_nr) {
 279                 spin_unlock_irqrestore(&pool->lock, flags);
 280                 element = pool->alloc(GFP_KERNEL, pool->pool_data);
 281                 if (!element)
 282                         goto out;
 283                 spin_lock_irqsave(&pool->lock, flags);
 284                 if (pool->curr_nr < pool->min_nr) {
 285                         add_element(pool, element);
 286                 } else {
 287                         spin_unlock_irqrestore(&pool->lock, flags);
 288                         pool->free(element, pool->pool_data);   /* Raced */
 289                         goto out;
 290                 }
 291         }
 292 out_unlock:
 293         spin_unlock_irqrestore(&pool->lock, flags);
 294 out:
 295         return 0;
 296 }
 297 EXPORT_SYMBOL(mempool_resize);
 298
 299 /**
 300  * mempool_alloc - allocate an element from a specific memory pool
 301  * @pool:      pointer to the memory pool which was allocated via
 302  *             mempool_create().
 303  * @gfp_mask:  the usual allocation bitmask.
 304  *
 305  * this function only sleeps if the alloc_fn() function sleeps or
 306  * returns NULL. Note that due to preallocation, this function
 307  * *never* fails when called from process contexts. (it might
 308  * fail if called from an IRQ context.)
 309  * Note: using __GFP_ZERO is not supported.
 310  */
 311 void *mempool_alloc(mempool_t *pool, gfp_t gfp_mask)
 312 {
 313         void *element;
 314         unsigned long flags;
 315         wait_queue_entry_t wait;
 316         gfp_t gfp_temp;
 317
 318         VM_WARN_ON_ONCE(gfp_mask & __GFP_ZERO);
 319         might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
 320
 321         gfp_mask |= __GFP_NOMEMALLOC;   /* don't allocate emergency reserves */
 322         gfp_mask |= __GFP_NORETRY;      /* don't loop in __alloc_pages */
 323         gfp_mask |= __GFP_NOWARN;       /* failures are OK */
 324
 325         gfp_temp = gfp_mask & ~(__GFP_DIRECT_RECLAIM|__GFP_IO);
 326
 327 repeat_alloc:
 328
 329         element = pool->alloc(gfp_temp, pool->pool_data);
 330         if (likely(element != NULL))
 331                 return element;
 332
 333         spin_lock_irqsave(&pool->lock, flags);
 334         if (likely(pool->curr_nr)) {
 335                 element = remove_element(pool, gfp_temp);
 336                 spin_unlock_irqrestore(&pool->lock, flags);
 337                 /* paired with rmb in mempool_free(), read comment there */
 338                 smp_wmb();
 339                 /*
 340                  * Update the allocation stack trace as this is more useful
 341                  * for debugging.
 342                  */
 343                 kmemleak_update_trace(element);
 344                 return element;
 345         }
 346
 347         /*
 348          * We use gfp mask w/o direct reclaim or IO for the first round.  If
 349          * alloc failed with that and @pool was empty, retry immediately.
 350          */
 351         if (gfp_temp != gfp_mask) {
 352                 spin_unlock_irqrestore(&pool->lock, flags);
 353                 gfp_temp = gfp_mask;
 354                 goto repeat_alloc;
 355         }
 356
 357         /* We must not sleep if !__GFP_DIRECT_RECLAIM */
 358         if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
 359                 spin_unlock_irqrestore(&pool->lock, flags);
 360                 return NULL;
 361         }
 362
 363         /* Let's wait for someone else to return an element to @pool */
 364         init_wait(&wait);
 365         prepare_to_wait(&pool->wait, &wait, TASK_UNINTERRUPTIBLE);
 366
 367         spin_unlock_irqrestore(&pool->lock, flags);
 368
 369         /*
 370          * FIXME: this should be io_schedule().  The timeout is there as a
 371          * workaround for some DM problems in 2.6.18.
 372          */
 373         io_schedule_timeout(5*HZ);
 374
 375         finish_wait(&pool->wait, &wait);
 376         goto repeat_alloc;
 377 }
 378 EXPORT_SYMBOL(mempool_alloc);
 379
 380 /**
 381  * mempool_free - return an element to the pool.
 382  * @element:   pool element pointer.
 383  * @pool:      pointer to the memory pool which was allocated via
 384  *             mempool_create().
 385  *
 386  * this function only sleeps if the free_fn() function sleeps.
 387  */
 388 void mempool_free(void *element, mempool_t *pool)
 389 {
 390         unsigned long flags;
 391
 392         if (unlikely(element == NULL))
 393                 return;
 394
 395         /*
 396          * Paired with the wmb in mempool_alloc().  The preceding read is
 397          * for @element and the following @pool->curr_nr.  This ensures
 398          * that the visible value of @pool->curr_nr is from after the
 399          * allocation of @element.  This is necessary for fringe cases
 400          * where @element was passed to this task without going through
 401          * barriers.
 402          *
 403          * For example, assume @p is %NULL at the beginning and one task
 404          * performs "p = mempool_alloc(...);" while another task is doing
 405          * "while (!p) cpu_relax(); mempool_free(p, ...);".  This function
 406          * may end up using curr_nr value which is from before allocation
 407          * of @p without the following rmb.
 408          */
 409         smp_rmb();
 410
 411         /*
 412          * For correctness, we need a test which is guaranteed to trigger
 413          * if curr_nr + #allocated == min_nr.  Testing curr_nr < min_nr
 414          * without locking achieves that and refilling as soon as possible
 415          * is desirable.
 416          *
 417          * Because curr_nr visible here is always a value after the
 418          * allocation of @element, any task which decremented curr_nr below
 419          * min_nr is guaranteed to see curr_nr < min_nr unless curr_nr gets
 420          * incremented to min_nr afterwards.  If curr_nr gets incremented
 421          * to min_nr after the allocation of @element, the elements
 422          * allocated after that are subject to the same guarantee.
 423          *
 424          * Waiters happen iff curr_nr is 0 and the above guarantee also
 425          * ensures that there will be frees which return elements to the
 426          * pool waking up the waiters.
 427          */
 428         if (unlikely(pool->curr_nr < pool->min_nr)) {
 429                 spin_lock_irqsave(&pool->lock, flags);
 430                 if (likely(pool->curr_nr < pool->min_nr)) {
 431                         add_element(pool, element);
 432                         spin_unlock_irqrestore(&pool->lock, flags);
 433                         wake_up(&pool->wait);
 434                         return;
 435                 }
 436                 spin_unlock_irqrestore(&pool->lock, flags);
 437         }
 438         pool->free(element, pool->pool_data);
 439 }
 440 EXPORT_SYMBOL(mempool_free);
 441
 442 /*
 443  * A commonly used alloc and free fn.
 444  */
 445 void *mempool_alloc_slab(gfp_t gfp_mask, void *pool_data)
 446 {
 447         struct kmem_cache *mem = pool_data;
 448         VM_BUG_ON(mem->ctor);
 449         return kmem_cache_alloc(mem, gfp_mask);
 450 }
 451 EXPORT_SYMBOL(mempool_alloc_slab);
 452
 453 void mempool_free_slab(void *element, void *pool_data)
 454 {
 455         struct kmem_cache *mem = pool_data;
 456         kmem_cache_free(mem, element);
 457 }
 458 EXPORT_SYMBOL(mempool_free_slab);
 459
 460 /*
 461  * A commonly used alloc and free fn that kmalloc/kfrees the amount of memory
 462  * specified by pool_data
 463  */
 464 void *mempool_kmalloc(gfp_t gfp_mask, void *pool_data)
 465 {
 466         size_t size = (size_t)pool_data;
 467         return kmalloc(size, gfp_mask);
 468 }
 469 EXPORT_SYMBOL(mempool_kmalloc);
 470
 471 void mempool_kfree(void *element, void *pool_data)
 472 {
 473         kfree(element);
 474 }
 475 EXPORT_SYMBOL(mempool_kfree);
 476
 477 /*
 478  * A simple mempool-backed page allocator that allocates pages
 479  * of the order specified by pool_data.
 480  */
 481 void *mempool_alloc_pages(gfp_t gfp_mask, void *pool_data)
 482 {
 483         int order = (int)(long)pool_data;
 484         return alloc_pages(gfp_mask, order);
 485 }
 486 EXPORT_SYMBOL(mempool_alloc_pages);
 487
 488 void mempool_free_pages(void *element, void *pool_data)
 489 {
 490         int order = (int)(long)pool_data;
 491         __free_pages(element, order);
 492 }
 493 EXPORT_SYMBOL(mempool_free_pages);