drivers/md/dm-bio-prison.c

   1 /*
   2  * Copyright (C) 2012 Red Hat, Inc.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm.h"
   8 #include "dm-bio-prison.h"
   9
  10 #include <linux/spinlock.h>
  11 #include <linux/mempool.h>
  12 #include <linux/module.h>
  13 #include <linux/slab.h>
  14
  15 /*----------------------------------------------------------------*/
  16
  17 struct dm_bio_prison_cell {
  18         struct hlist_node list;
  19         struct dm_bio_prison *prison;
  20         struct dm_cell_key key;
  21         struct bio *holder;
  22         struct bio_list bios;
  23 };
  24
  25 struct dm_bio_prison {
  26         spinlock_t lock;
  27         mempool_t *cell_pool;
  28
  29         unsigned nr_buckets;
  30         unsigned hash_mask;
  31         struct hlist_head *cells;
  32 };
  33
  34 /*----------------------------------------------------------------*/
  35
  36 static uint32_t calc_nr_buckets(unsigned nr_cells)
  37 {
  38         uint32_t n = 128;
  39
  40         nr_cells /= 4;
  41         nr_cells = min(nr_cells, 8192u);
  42
  43         while (n < nr_cells)
  44                 n <<= 1;
  45
  46         return n;
  47 }
  48
  49 static struct kmem_cache *_cell_cache;
  50
  51 /*
  52  * @nr_cells should be the number of cells you want in use _concurrently_.
  53  * Don't confuse it with the number of distinct keys.
  54  */
  55 struct dm_bio_prison *dm_bio_prison_create(unsigned nr_cells)
  56 {
  57         unsigned i;
  58         uint32_t nr_buckets = calc_nr_buckets(nr_cells);
  59         size_t len = sizeof(struct dm_bio_prison) +
  60                 (sizeof(struct hlist_head) * nr_buckets);
  61         struct dm_bio_prison *prison = kmalloc(len, GFP_KERNEL);
  62
  63         if (!prison)
  64                 return NULL;
  65
  66         spin_lock_init(&prison->lock);
  67         prison->cell_pool = mempool_create_slab_pool(nr_cells, _cell_cache);
  68         if (!prison->cell_pool) {
  69                 kfree(prison);
  70                 return NULL;
  71         }
  72
  73         prison->nr_buckets = nr_buckets;
  74         prison->hash_mask = nr_buckets - 1;
  75         prison->cells = (struct hlist_head *) (prison + 1);
  76         for (i = 0; i < nr_buckets; i++)
  77                 INIT_HLIST_HEAD(prison->cells + i);
  78
  79         return prison;
  80 }
  81 EXPORT_SYMBOL_GPL(dm_bio_prison_create);
  82
  83 void dm_bio_prison_destroy(struct dm_bio_prison *prison)
  84 {
  85         mempool_destroy(prison->cell_pool);
  86         kfree(prison);
  87 }
  88 EXPORT_SYMBOL_GPL(dm_bio_prison_destroy);
  89
  90 static uint32_t hash_key(struct dm_bio_prison *prison, struct dm_cell_key *key)
  91 {
  92         const unsigned long BIG_PRIME = 4294967291UL;
  93         uint64_t hash = key->block * BIG_PRIME;
  94
  95         return (uint32_t) (hash & prison->hash_mask);
  96 }
  97
  98 static int keys_equal(struct dm_cell_key *lhs, struct dm_cell_key *rhs)
  99 {
 100                return (lhs->virtual == rhs->virtual) &&
 101                        (lhs->dev == rhs->dev) &&
 102                        (lhs->block == rhs->block);
 103 }
 104
 105 static struct dm_bio_prison_cell *__search_bucket(struct hlist_head *bucket,
 106                                                   struct dm_cell_key *key)
 107 {
 108         struct dm_bio_prison_cell *cell;
 109         struct hlist_node *tmp;
 110
 111         hlist_for_each_entry(cell, tmp, bucket, list)
 112                 if (keys_equal(&cell->key, key))
 113                         return cell;
 114
 115         return NULL;
 116 }
 117
 118 /*
 119  * This may block if a new cell needs allocating.  You must ensure that
 120  * cells will be unlocked even if the calling thread is blocked.
 121  *
 122  * Returns 1 if the cell was already held, 0 if @inmate is the new holder.
 123  */
 124 int dm_bio_detain(struct dm_bio_prison *prison, struct dm_cell_key *key,
 125                   struct bio *inmate, struct dm_bio_prison_cell **ref)
 126 {
 127         int r = 1;
 128         unsigned long flags;
 129         uint32_t hash = hash_key(prison, key);
 130         struct dm_bio_prison_cell *cell, *cell2;
 131
 132         BUG_ON(hash > prison->nr_buckets);
 133
 134         spin_lock_irqsave(&prison->lock, flags);
 135
 136         cell = __search_bucket(prison->cells + hash, key);
 137         if (cell) {
 138                 bio_list_add(&cell->bios, inmate);
 139                 goto out;
 140         }
 141
 142         /*
 143          * Allocate a new cell
 144          */
 145         spin_unlock_irqrestore(&prison->lock, flags);
 146         cell2 = mempool_alloc(prison->cell_pool, GFP_NOIO);
 147         spin_lock_irqsave(&prison->lock, flags);
 148
 149         /*
 150          * We've been unlocked, so we have to double check that
 151          * nobody else has inserted this cell in the meantime.
 152          */
 153         cell = __search_bucket(prison->cells + hash, key);
 154         if (cell) {
 155                 mempool_free(cell2, prison->cell_pool);
 156                 bio_list_add(&cell->bios, inmate);
 157                 goto out;
 158         }
 159
 160         /*
 161          * Use new cell.
 162          */
 163         cell = cell2;
 164
 165         cell->prison = prison;
 166         memcpy(&cell->key, key, sizeof(cell->key));
 167         cell->holder = inmate;
 168         bio_list_init(&cell->bios);
 169         hlist_add_head(&cell->list, prison->cells + hash);
 170
 171         r = 0;
 172
 173 out:
 174         spin_unlock_irqrestore(&prison->lock, flags);
 175
 176         *ref = cell;
 177
 178         return r;
 179 }
 180 EXPORT_SYMBOL_GPL(dm_bio_detain);
 181
 182 /*
 183  * @inmates must have been initialised prior to this call
 184  */
 185 static void __cell_release(struct dm_bio_prison_cell *cell, struct bio_list *inmates)
 186 {
 187         struct dm_bio_prison *prison = cell->prison;
 188
 189         hlist_del(&cell->list);
 190
 191         if (inmates) {
 192                 bio_list_add(inmates, cell->holder);
 193                 bio_list_merge(inmates, &cell->bios);
 194         }
 195
 196         mempool_free(cell, prison->cell_pool);
 197 }
 198
 199 void dm_cell_release(struct dm_bio_prison_cell *cell, struct bio_list *bios)
 200 {
 201         unsigned long flags;
 202         struct dm_bio_prison *prison = cell->prison;
 203
 204         spin_lock_irqsave(&prison->lock, flags);
 205         __cell_release(cell, bios);
 206         spin_unlock_irqrestore(&prison->lock, flags);
 207 }
 208 EXPORT_SYMBOL_GPL(dm_cell_release);
 209
 210 /*
 211  * There are a couple of places where we put a bio into a cell briefly
 212  * before taking it out again.  In these situations we know that no other
 213  * bio may be in the cell.  This function releases the cell, and also does
 214  * a sanity check.
 215  */
 216 static void __cell_release_singleton(struct dm_bio_prison_cell *cell, struct bio *bio)
 217 {
 218         BUG_ON(cell->holder != bio);
 219         BUG_ON(!bio_list_empty(&cell->bios));
 220
 221         __cell_release(cell, NULL);
 222 }
 223
 224 void dm_cell_release_singleton(struct dm_bio_prison_cell *cell, struct bio *bio)
 225 {
 226         unsigned long flags;
 227         struct dm_bio_prison *prison = cell->prison;
 228
 229         spin_lock_irqsave(&prison->lock, flags);
 230         __cell_release_singleton(cell, bio);
 231         spin_unlock_irqrestore(&prison->lock, flags);
 232 }
 233 EXPORT_SYMBOL_GPL(dm_cell_release_singleton);
 234
 235 /*
 236  * Sometimes we don't want the holder, just the additional bios.
 237  */
 238 static void __cell_release_no_holder(struct dm_bio_prison_cell *cell, struct bio_list *inmates)
 239 {
 240         struct dm_bio_prison *prison = cell->prison;
 241
 242         hlist_del(&cell->list);
 243         bio_list_merge(inmates, &cell->bios);
 244
 245         mempool_free(cell, prison->cell_pool);
 246 }
 247
 248 void dm_cell_release_no_holder(struct dm_bio_prison_cell *cell, struct bio_list *inmates)
 249 {
 250         unsigned long flags;
 251         struct dm_bio_prison *prison = cell->prison;
 252
 253         spin_lock_irqsave(&prison->lock, flags);
 254         __cell_release_no_holder(cell, inmates);
 255         spin_unlock_irqrestore(&prison->lock, flags);
 256 }
 257 EXPORT_SYMBOL_GPL(dm_cell_release_no_holder);
 258
 259 void dm_cell_error(struct dm_bio_prison_cell *cell)
 260 {
 261         struct dm_bio_prison *prison = cell->prison;
 262         struct bio_list bios;
 263         struct bio *bio;
 264         unsigned long flags;
 265
 266         bio_list_init(&bios);
 267
 268         spin_lock_irqsave(&prison->lock, flags);
 269         __cell_release(cell, &bios);
 270         spin_unlock_irqrestore(&prison->lock, flags);
 271
 272         while ((bio = bio_list_pop(&bios)))
 273                 bio_io_error(bio);
 274 }
 275 EXPORT_SYMBOL_GPL(dm_cell_error);
 276
 277 /*----------------------------------------------------------------*/
 278
 279 #define DEFERRED_SET_SIZE 64
 280
 281 struct dm_deferred_entry {
 282         struct dm_deferred_set *ds;
 283         unsigned count;
 284         struct list_head work_items;
 285 };
 286
 287 struct dm_deferred_set {
 288         spinlock_t lock;
 289         unsigned current_entry;
 290         unsigned sweeper;
 291         struct dm_deferred_entry entries[DEFERRED_SET_SIZE];
 292 };
 293
 294 struct dm_deferred_set *dm_deferred_set_create(void)
 295 {
 296         int i;
 297         struct dm_deferred_set *ds;
 298
 299         ds = kmalloc(sizeof(*ds), GFP_KERNEL);
 300         if (!ds)
 301                 return NULL;
 302
 303         spin_lock_init(&ds->lock);
 304         ds->current_entry = 0;
 305         ds->sweeper = 0;
 306         for (i = 0; i < DEFERRED_SET_SIZE; i++) {
 307                 ds->entries[i].ds = ds;
 308                 ds->entries[i].count = 0;
 309                 INIT_LIST_HEAD(&ds->entries[i].work_items);
 310         }
 311
 312         return ds;
 313 }
 314 EXPORT_SYMBOL_GPL(dm_deferred_set_create);
 315
 316 void dm_deferred_set_destroy(struct dm_deferred_set *ds)
 317 {
 318         kfree(ds);
 319 }
 320 EXPORT_SYMBOL_GPL(dm_deferred_set_destroy);
 321
 322 struct dm_deferred_entry *dm_deferred_entry_inc(struct dm_deferred_set *ds)
 323 {
 324         unsigned long flags;
 325         struct dm_deferred_entry *entry;
 326
 327         spin_lock_irqsave(&ds->lock, flags);
 328         entry = ds->entries + ds->current_entry;
 329         entry->count++;
 330         spin_unlock_irqrestore(&ds->lock, flags);
 331
 332         return entry;
 333 }
 334 EXPORT_SYMBOL_GPL(dm_deferred_entry_inc);
 335
 336 static unsigned ds_next(unsigned index)
 337 {
 338         return (index + 1) % DEFERRED_SET_SIZE;
 339 }
 340
 341 static void __sweep(struct dm_deferred_set *ds, struct list_head *head)
 342 {
 343         while ((ds->sweeper != ds->current_entry) &&
 344                !ds->entries[ds->sweeper].count) {
 345                 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
 346                 ds->sweeper = ds_next(ds->sweeper);
 347         }
 348
 349         if ((ds->sweeper == ds->current_entry) && !ds->entries[ds->sweeper].count)
 350                 list_splice_init(&ds->entries[ds->sweeper].work_items, head);
 351 }
 352
 353 void dm_deferred_entry_dec(struct dm_deferred_entry *entry, struct list_head *head)
 354 {
 355         unsigned long flags;
 356
 357         spin_lock_irqsave(&entry->ds->lock, flags);
 358         BUG_ON(!entry->count);
 359         --entry->count;
 360         __sweep(entry->ds, head);
 361         spin_unlock_irqrestore(&entry->ds->lock, flags);
 362 }
 363 EXPORT_SYMBOL_GPL(dm_deferred_entry_dec);
 364
 365 /*
 366  * Returns 1 if deferred or 0 if no pending items to delay job.
 367  */
 368 int dm_deferred_set_add_work(struct dm_deferred_set *ds, struct list_head *work)
 369 {
 370         int r = 1;
 371         unsigned long flags;
 372         unsigned next_entry;
 373
 374         spin_lock_irqsave(&ds->lock, flags);
 375         if ((ds->sweeper == ds->current_entry) &&
 376             !ds->entries[ds->current_entry].count)
 377                 r = 0;
 378         else {
 379                 list_add(work, &ds->entries[ds->current_entry].work_items);
 380                 next_entry = ds_next(ds->current_entry);
 381                 if (!ds->entries[next_entry].count)
 382                         ds->current_entry = next_entry;
 383         }
 384         spin_unlock_irqrestore(&ds->lock, flags);
 385
 386         return r;
 387 }
 388 EXPORT_SYMBOL_GPL(dm_deferred_set_add_work);
 389
 390 /*----------------------------------------------------------------*/
 391
 392 static int __init dm_bio_prison_init(void)
 393 {
 394         _cell_cache = KMEM_CACHE(dm_bio_prison_cell, 0);
 395         if (!_cell_cache)
 396                 return -ENOMEM;
 397
 398         return 0;
 399 }
 400
 401 static void __exit dm_bio_prison_exit(void)
 402 {
 403         kmem_cache_destroy(_cell_cache);
 404         _cell_cache = NULL;
 405 }
 406
 407 /*
 408  * module hooks
 409  */
 410 module_init(dm_bio_prison_init);
 411 module_exit(dm_bio_prison_exit);
 412
 413 MODULE_DESCRIPTION(DM_NAME " bio prison");
 414 MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
 415 MODULE_LICENSE("GPL");