fs/btrfs/inode-map.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/delay.h>
  20 #include <linux/kthread.h>
  21 #include <linux/pagemap.h>
  22
  23 #include "ctree.h"
  24 #include "disk-io.h"
  25 #include "free-space-cache.h"
  26 #include "inode-map.h"
  27 #include "transaction.h"
  28
  29 static int caching_kthread(void *data)
  30 {
  31         struct btrfs_root *root = data;
  32         struct btrfs_fs_info *fs_info = root->fs_info;
  33         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
  34         struct btrfs_key key;
  35         struct btrfs_path *path;
  36         struct extent_buffer *leaf;
  37         u64 last = (u64)-1;
  38         int slot;
  39         int ret;
  40
  41         path = btrfs_alloc_path();
  42         if (!path)
  43                 return -ENOMEM;
  44
  45         /* Since the commit root is read-only, we can safely skip locking. */
  46         path->skip_locking = 1;
  47         path->search_commit_root = 1;
  48         path->reada = 2;
  49
  50         key.objectid = BTRFS_FIRST_FREE_OBJECTID;
  51         key.offset = 0;
  52         key.type = BTRFS_INODE_ITEM_KEY;
  53 again:
  54         /* need to make sure the commit_root doesn't disappear */
  55         mutex_lock(&root->fs_commit_mutex);
  56
  57         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  58         if (ret < 0)
  59                 goto out;
  60
  61         while (1) {
  62                 smp_mb();
  63                 if (fs_info->closing)
  64                         goto out;
  65
  66                 leaf = path->nodes[0];
  67                 slot = path->slots[0];
  68                 if (slot >= btrfs_header_nritems(leaf)) {
  69                         ret = btrfs_next_leaf(root, path);
  70                         if (ret < 0)
  71                                 goto out;
  72                         else if (ret > 0)
  73                                 break;
  74
  75                         if (need_resched() ||
  76                             btrfs_transaction_in_commit(fs_info)) {
  77                                 leaf = path->nodes[0];
  78
  79                                 if (btrfs_header_nritems(leaf) == 0) {
  80                                         WARN_ON(1);
  81                                         break;
  82                                 }
  83
  84                                 /*
  85                                  * Save the key so we can advances forward
  86                                  * in the next search.
  87                                  */
  88                                 btrfs_item_key_to_cpu(leaf, &key, 0);
  89                                 btrfs_release_path(path);
  90                                 root->cache_progress = last;
  91                                 mutex_unlock(&root->fs_commit_mutex);
  92                                 schedule_timeout(1);
  93                                 goto again;
  94                         } else
  95                                 continue;
  96                 }
  97
  98                 btrfs_item_key_to_cpu(leaf, &key, slot);
  99
 100                 if (key.type != BTRFS_INODE_ITEM_KEY)
 101                         goto next;
 102
 103                 if (key.objectid >= root->highest_objectid)
 104                         break;
 105
 106                 if (last != (u64)-1 && last + 1 != key.objectid) {
 107                         __btrfs_add_free_space(ctl, last + 1,
 108                                                key.objectid - last - 1);
 109                         wake_up(&root->cache_wait);
 110                 }
 111
 112                 last = key.objectid;
 113 next:
 114                 path->slots[0]++;
 115         }
 116
 117         if (last < root->highest_objectid - 1) {
 118                 __btrfs_add_free_space(ctl, last + 1,
 119                                        root->highest_objectid - last - 1);
 120         }
 121
 122         spin_lock(&root->cache_lock);
 123         root->cached = BTRFS_CACHE_FINISHED;
 124         spin_unlock(&root->cache_lock);
 125
 126         root->cache_progress = (u64)-1;
 127         btrfs_unpin_free_ino(root);
 128 out:
 129         wake_up(&root->cache_wait);
 130         mutex_unlock(&root->fs_commit_mutex);
 131
 132         btrfs_free_path(path);
 133
 134         return ret;
 135 }
 136
 137 static void start_caching(struct btrfs_root *root)
 138 {
 139         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 140         struct task_struct *tsk;
 141         int ret;
 142         u64 objectid;
 143
 144         spin_lock(&root->cache_lock);
 145         if (root->cached != BTRFS_CACHE_NO) {
 146                 spin_unlock(&root->cache_lock);
 147                 return;
 148         }
 149
 150         root->cached = BTRFS_CACHE_STARTED;
 151         spin_unlock(&root->cache_lock);
 152
 153         ret = load_free_ino_cache(root->fs_info, root);
 154         if (ret == 1) {
 155                 spin_lock(&root->cache_lock);
 156                 root->cached = BTRFS_CACHE_FINISHED;
 157                 spin_unlock(&root->cache_lock);
 158                 return;
 159         }
 160
 161         /*
 162          * It can be quite time-consuming to fill the cache by searching
 163          * through the extent tree, and this can keep ino allocation path
 164          * waiting. Therefore at start we quickly find out the highest
 165          * inode number and we know we can use inode numbers which fall in
 166          * [highest_ino + 1, BTRFS_LAST_FREE_OBJECTID].
 167          */
 168         ret = btrfs_find_free_objectid(root, &objectid);
 169         if (!ret && objectid <= BTRFS_LAST_FREE_OBJECTID) {
 170                 __btrfs_add_free_space(ctl, objectid,
 171                                        BTRFS_LAST_FREE_OBJECTID - objectid + 1);
 172         }
 173
 174         tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu\n",
 175                           root->root_key.objectid);
 176         BUG_ON(IS_ERR(tsk));
 177 }
 178
 179 int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
 180 {
 181 again:
 182         *objectid = btrfs_find_ino_for_alloc(root);
 183
 184         if (*objectid != 0)
 185                 return 0;
 186
 187         start_caching(root);
 188
 189         wait_event(root->cache_wait,
 190                    root->cached == BTRFS_CACHE_FINISHED ||
 191                    root->free_ino_ctl->free_space > 0);
 192
 193         if (root->cached == BTRFS_CACHE_FINISHED &&
 194             root->free_ino_ctl->free_space == 0)
 195                 return -ENOSPC;
 196         else
 197                 goto again;
 198 }
 199
 200 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
 201 {
 202         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 203         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 204 again:
 205         if (root->cached == BTRFS_CACHE_FINISHED) {
 206                 __btrfs_add_free_space(ctl, objectid, 1);
 207         } else {
 208                 /*
 209                  * If we are in the process of caching free ino chunks,
 210                  * to avoid adding the same inode number to the free_ino
 211                  * tree twice due to cross transaction, we'll leave it
 212                  * in the pinned tree until a transaction is committed
 213                  * or the caching work is done.
 214                  */
 215
 216                 mutex_lock(&root->fs_commit_mutex);
 217                 spin_lock(&root->cache_lock);
 218                 if (root->cached == BTRFS_CACHE_FINISHED) {
 219                         spin_unlock(&root->cache_lock);
 220                         mutex_unlock(&root->fs_commit_mutex);
 221                         goto again;
 222                 }
 223                 spin_unlock(&root->cache_lock);
 224
 225                 start_caching(root);
 226
 227                 if (objectid <= root->cache_progress ||
 228                     objectid > root->highest_objectid)
 229                         __btrfs_add_free_space(ctl, objectid, 1);
 230                 else
 231                         __btrfs_add_free_space(pinned, objectid, 1);
 232
 233                 mutex_unlock(&root->fs_commit_mutex);
 234         }
 235 }
 236
 237 /*
 238  * When a transaction is committed, we'll move those inode numbers which
 239  * are smaller than root->cache_progress from pinned tree to free_ino tree,
 240  * and others will just be dropped, because the commit root we were
 241  * searching has changed.
 242  *
 243  * Must be called with root->fs_commit_mutex held
 244  */
 245 void btrfs_unpin_free_ino(struct btrfs_root *root)
 246 {
 247         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 248         struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
 249         struct btrfs_free_space *info;
 250         struct rb_node *n;
 251         u64 count;
 252
 253         while (1) {
 254                 n = rb_first(rbroot);
 255                 if (!n)
 256                         break;
 257
 258                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 259                 BUG_ON(info->bitmap);
 260
 261                 if (info->offset > root->cache_progress)
 262                         goto free;
 263                 else if (info->offset + info->bytes > root->cache_progress)
 264                         count = root->cache_progress - info->offset + 1;
 265                 else
 266                         count = info->bytes;
 267
 268                 __btrfs_add_free_space(ctl, info->offset, count);
 269 free:
 270                 rb_erase(&info->offset_index, rbroot);
 271                 kfree(info);
 272         }
 273 }
 274
 275 #define INIT_THRESHOLD  (((1024 * 32) / 2) / sizeof(struct btrfs_free_space))
 276 #define INODES_PER_BITMAP (PAGE_CACHE_SIZE * 8)
 277
 278 /*
 279  * The goal is to keep the memory used by the free_ino tree won't
 280  * exceed the memory if we use bitmaps only.
 281  */
 282 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 283 {
 284         struct btrfs_free_space *info;
 285         struct rb_node *n;
 286         int max_ino;
 287         int max_bitmaps;
 288
 289         n = rb_last(&ctl->free_space_offset);
 290         if (!n) {
 291                 ctl->extents_thresh = INIT_THRESHOLD;
 292                 return;
 293         }
 294         info = rb_entry(n, struct btrfs_free_space, offset_index);
 295
 296         /*
 297          * Find the maximum inode number in the filesystem. Note we
 298          * ignore the fact that this can be a bitmap, because we are
 299          * not doing precise calculation.
 300          */
 301         max_ino = info->bytes - 1;
 302
 303         max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
 304         if (max_bitmaps <= ctl->total_bitmaps) {
 305                 ctl->extents_thresh = 0;
 306                 return;
 307         }
 308
 309         ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
 310                                 PAGE_CACHE_SIZE / sizeof(*info);
 311 }
 312
 313 /*
 314  * We don't fall back to bitmap, if we are below the extents threshold
 315  * or this chunk of inode numbers is a big one.
 316  */
 317 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
 318                        struct btrfs_free_space *info)
 319 {
 320         if (ctl->free_extents < ctl->extents_thresh ||
 321             info->bytes > INODES_PER_BITMAP / 10)
 322                 return false;
 323
 324         return true;
 325 }
 326
 327 static struct btrfs_free_space_op free_ino_op = {
 328         .recalc_thresholds      = recalculate_thresholds,
 329         .use_bitmap             = use_bitmap,
 330 };
 331
 332 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
 333 {
 334 }
 335
 336 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
 337                               struct btrfs_free_space *info)
 338 {
 339         /*
 340          * We always use extents for two reasons:
 341          *
 342          * - The pinned tree is only used during the process of caching
 343          *   work.
 344          * - Make code simpler. See btrfs_unpin_free_ino().
 345          */
 346         return false;
 347 }
 348
 349 static struct btrfs_free_space_op pinned_free_ino_op = {
 350         .recalc_thresholds      = pinned_recalc_thresholds,
 351         .use_bitmap             = pinned_use_bitmap,
 352 };
 353
 354 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
 355 {
 356         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 357         struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
 358
 359         spin_lock_init(&ctl->tree_lock);
 360         ctl->unit = 1;
 361         ctl->start = 0;
 362         ctl->private = NULL;
 363         ctl->op = &free_ino_op;
 364
 365         /*
 366          * Initially we allow to use 16K of ram to cache chunks of
 367          * inode numbers before we resort to bitmaps. This is somewhat
 368          * arbitrary, but it will be adjusted in runtime.
 369          */
 370         ctl->extents_thresh = INIT_THRESHOLD;
 371
 372         spin_lock_init(&pinned->tree_lock);
 373         pinned->unit = 1;
 374         pinned->start = 0;
 375         pinned->private = NULL;
 376         pinned->extents_thresh = 0;
 377         pinned->op = &pinned_free_ino_op;
 378 }
 379
 380 int btrfs_save_ino_cache(struct btrfs_root *root,
 381                          struct btrfs_trans_handle *trans)
 382 {
 383         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
 384         struct btrfs_path *path;
 385         struct inode *inode;
 386         u64 alloc_hint = 0;
 387         int ret;
 388         int prealloc;
 389         bool retry = false;
 390
 391         path = btrfs_alloc_path();
 392         if (!path)
 393                 return -ENOMEM;
 394 again:
 395         inode = lookup_free_ino_inode(root, path);
 396         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
 397                 ret = PTR_ERR(inode);
 398                 goto out;
 399         }
 400
 401         if (IS_ERR(inode)) {
 402                 BUG_ON(retry);
 403                 retry = true;
 404
 405                 ret = create_free_ino_inode(root, trans, path);
 406                 if (ret)
 407                         goto out;
 408                 goto again;
 409         }
 410
 411         BTRFS_I(inode)->generation = 0;
 412         ret = btrfs_update_inode(trans, root, inode);
 413         WARN_ON(ret);
 414
 415         if (i_size_read(inode) > 0) {
 416                 ret = btrfs_truncate_free_space_cache(root, trans, path, inode);
 417                 if (ret)
 418                         goto out_put;
 419         }
 420
 421         spin_lock(&root->cache_lock);
 422         if (root->cached != BTRFS_CACHE_FINISHED) {
 423                 ret = -1;
 424                 spin_unlock(&root->cache_lock);
 425                 goto out_put;
 426         }
 427         spin_unlock(&root->cache_lock);
 428
 429         spin_lock(&ctl->tree_lock);
 430         prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
 431         prealloc = ALIGN(prealloc, PAGE_CACHE_SIZE);
 432         prealloc += ctl->total_bitmaps * PAGE_CACHE_SIZE;
 433         spin_unlock(&ctl->tree_lock);
 434
 435         /* Just to make sure we have enough space */
 436         prealloc += 8 * PAGE_CACHE_SIZE;
 437
 438         ret = btrfs_check_data_free_space(inode, prealloc);
 439         if (ret)
 440                 goto out_put;
 441
 442         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
 443                                               prealloc, prealloc, &alloc_hint);
 444         if (ret)
 445                 goto out_put;
 446         btrfs_free_reserved_data_space(inode, prealloc);
 447
 448 out_put:
 449         iput(inode);
 450 out:
 451         if (ret == 0)
 452                 ret = btrfs_write_out_ino_cache(root, trans, path);
 453
 454         btrfs_free_path(path);
 455         return ret;
 456 }
 457
 458 static int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
 459 {
 460         struct btrfs_path *path;
 461         int ret;
 462         struct extent_buffer *l;
 463         struct btrfs_key search_key;
 464         struct btrfs_key found_key;
 465         int slot;
 466
 467         path = btrfs_alloc_path();
 468         if (!path)
 469                 return -ENOMEM;
 470
 471         search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
 472         search_key.type = -1;
 473         search_key.offset = (u64)-1;
 474         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 475         if (ret < 0)
 476                 goto error;
 477         BUG_ON(ret == 0);
 478         if (path->slots[0] > 0) {
 479                 slot = path->slots[0] - 1;
 480                 l = path->nodes[0];
 481                 btrfs_item_key_to_cpu(l, &found_key, slot);
 482                 *objectid = max_t(u64, found_key.objectid,
 483                                   BTRFS_FIRST_FREE_OBJECTID - 1);
 484         } else {
 485                 *objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
 486         }
 487         ret = 0;
 488 error:
 489         btrfs_free_path(path);
 490         return ret;
 491 }
 492
 493 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
 494 {
 495         int ret;
 496         mutex_lock(&root->objectid_mutex);
 497
 498         if (unlikely(root->highest_objectid < BTRFS_FIRST_FREE_OBJECTID)) {
 499                 ret = btrfs_find_highest_objectid(root,
 500                                                   &root->highest_objectid);
 501                 if (ret)
 502                         goto out;
 503         }
 504
 505         if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
 506                 ret = -ENOSPC;
 507                 goto out;
 508         }
 509
 510         *objectid = ++root->highest_objectid;
 511         ret = 0;
 512 out:
 513         mutex_unlock(&root->objectid_mutex);
 514         return ret;
 515 }