fs/btrfs/reada.c

   1 /*
   2  * Copyright (C) 2011 STRATO.  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/sched.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/writeback.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/rbtree.h>
  24 #include <linux/slab.h>
  25 #include <linux/workqueue.h>
  26 #include "ctree.h"
  27 #include "volumes.h"
  28 #include "disk-io.h"
  29 #include "transaction.h"
  30
  31 #undef DEBUG
  32
  33 /*
  34  * This is the implementation for the generic read ahead framework.
  35  *
  36  * To trigger a readahead, btrfs_reada_add must be called. It will start
  37  * a read ahead for the given range [start, end) on tree root. The returned
  38  * handle can either be used to wait on the readahead to finish
  39  * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
  40  *
  41  * The read ahead works as follows:
  42  * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
  43  * reada_start_machine will then search for extents to prefetch and trigger
  44  * some reads. When a read finishes for a node, all contained node/leaf
  45  * pointers that lie in the given range will also be enqueued. The reads will
  46  * be triggered in sequential order, thus giving a big win over a naive
  47  * enumeration. It will also make use of multi-device layouts. Each disk
  48  * will have its on read pointer and all disks will by utilized in parallel.
  49  * Also will no two disks read both sides of a mirror simultaneously, as this
  50  * would waste seeking capacity. Instead both disks will read different parts
  51  * of the filesystem.
  52  * Any number of readaheads can be started in parallel. The read order will be
  53  * determined globally, i.e. 2 parallel readaheads will normally finish faster
  54  * than the 2 started one after another.
  55  */
  56
  57 #define MAX_IN_FLIGHT 6
  58
  59 struct reada_extctl {
  60         struct list_head        list;
  61         struct reada_control    *rc;
  62         u64                     generation;
  63 };
  64
  65 struct reada_extent {
  66         u64                     logical;
  67         struct btrfs_key        top;
  68         u32                     blocksize;
  69         int                     err;
  70         struct list_head        extctl;
  71         struct kref             refcnt;
  72         spinlock_t              lock;
  73         struct reada_zone       *zones[BTRFS_MAX_MIRRORS];
  74         int                     nzones;
  75         struct btrfs_device     *scheduled_for;
  76 };
  77
  78 struct reada_zone {
  79         u64                     start;
  80         u64                     end;
  81         u64                     elems;
  82         struct list_head        list;
  83         spinlock_t              lock;
  84         int                     locked;
  85         struct btrfs_device     *device;
  86         struct btrfs_device     *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
  87                                                            * self */
  88         int                     ndevs;
  89         struct kref             refcnt;
  90 };
  91
  92 struct reada_machine_work {
  93         struct btrfs_work       work;
  94         struct btrfs_fs_info    *fs_info;
  95 };
  96
  97 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
  98 static void reada_control_release(struct kref *kref);
  99 static void reada_zone_release(struct kref *kref);
 100 static void reada_start_machine(struct btrfs_fs_info *fs_info);
 101 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
 102
 103 static int reada_add_block(struct reada_control *rc, u64 logical,
 104                            struct btrfs_key *top, int level, u64 generation);
 105
 106 /* recurses */
 107 /* in case of err, eb might be NULL */
 108 static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
 109                             u64 start, int err)
 110 {
 111         int level = 0;
 112         int nritems;
 113         int i;
 114         u64 bytenr;
 115         u64 generation;
 116         struct reada_extent *re;
 117         struct btrfs_fs_info *fs_info = root->fs_info;
 118         struct list_head list;
 119         unsigned long index = start >> PAGE_CACHE_SHIFT;
 120         struct btrfs_device *for_dev;
 121
 122         if (eb)
 123                 level = btrfs_header_level(eb);
 124
 125         /* find extent */
 126         spin_lock(&fs_info->reada_lock);
 127         re = radix_tree_lookup(&fs_info->reada_tree, index);
 128         if (re)
 129                 kref_get(&re->refcnt);
 130         spin_unlock(&fs_info->reada_lock);
 131
 132         if (!re)
 133                 return -1;
 134
 135         spin_lock(&re->lock);
 136         /*
 137          * just take the full list from the extent. afterwards we
 138          * don't need the lock anymore
 139          */
 140         list_replace_init(&re->extctl, &list);
 141         for_dev = re->scheduled_for;
 142         re->scheduled_for = NULL;
 143         spin_unlock(&re->lock);
 144
 145         if (err == 0) {
 146                 nritems = level ? btrfs_header_nritems(eb) : 0;
 147                 generation = btrfs_header_generation(eb);
 148                 /*
 149                  * FIXME: currently we just set nritems to 0 if this is a leaf,
 150                  * effectively ignoring the content. In a next step we could
 151                  * trigger more readahead depending from the content, e.g.
 152                  * fetch the checksums for the extents in the leaf.
 153                  */
 154         } else {
 155                 /*
 156                  * this is the error case, the extent buffer has not been
 157                  * read correctly. We won't access anything from it and
 158                  * just cleanup our data structures. Effectively this will
 159                  * cut the branch below this node from read ahead.
 160                  */
 161                 nritems = 0;
 162                 generation = 0;
 163         }
 164
 165         for (i = 0; i < nritems; i++) {
 166                 struct reada_extctl *rec;
 167                 u64 n_gen;
 168                 struct btrfs_key key;
 169                 struct btrfs_key next_key;
 170
 171                 btrfs_node_key_to_cpu(eb, &key, i);
 172                 if (i + 1 < nritems)
 173                         btrfs_node_key_to_cpu(eb, &next_key, i + 1);
 174                 else
 175                         next_key = re->top;
 176                 bytenr = btrfs_node_blockptr(eb, i);
 177                 n_gen = btrfs_node_ptr_generation(eb, i);
 178
 179                 list_for_each_entry(rec, &list, list) {
 180                         struct reada_control *rc = rec->rc;
 181
 182                         /*
 183                          * if the generation doesn't match, just ignore this
 184                          * extctl. This will probably cut off a branch from
 185                          * prefetch. Alternatively one could start a new (sub-)
 186                          * prefetch for this branch, starting again from root.
 187                          * FIXME: move the generation check out of this loop
 188                          */
 189 #ifdef DEBUG
 190                         if (rec->generation != generation) {
 191                                 printk(KERN_DEBUG "generation mismatch for "
 192                                                 "(%llu,%d,%llu) %llu != %llu\n",
 193                                        key.objectid, key.type, key.offset,
 194                                        rec->generation, generation);
 195                         }
 196 #endif
 197                         if (rec->generation == generation &&
 198                             btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
 199                             btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
 200                                 reada_add_block(rc, bytenr, &next_key,
 201                                                 level - 1, n_gen);
 202                 }
 203         }
 204         /*
 205          * free extctl records
 206          */
 207         while (!list_empty(&list)) {
 208                 struct reada_control *rc;
 209                 struct reada_extctl *rec;
 210
 211                 rec = list_first_entry(&list, struct reada_extctl, list);
 212                 list_del(&rec->list);
 213                 rc = rec->rc;
 214                 kfree(rec);
 215
 216                 kref_get(&rc->refcnt);
 217                 if (atomic_dec_and_test(&rc->elems)) {
 218                         kref_put(&rc->refcnt, reada_control_release);
 219                         wake_up(&rc->wait);
 220                 }
 221                 kref_put(&rc->refcnt, reada_control_release);
 222
 223                 reada_extent_put(fs_info, re);  /* one ref for each entry */
 224         }
 225         reada_extent_put(fs_info, re);  /* our ref */
 226         if (for_dev)
 227                 atomic_dec(&for_dev->reada_in_flight);
 228
 229         return 0;
 230 }
 231
 232 /*
 233  * start is passed separately in case eb in NULL, which may be the case with
 234  * failed I/O
 235  */
 236 int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
 237                          u64 start, int err)
 238 {
 239         int ret;
 240
 241         ret = __readahead_hook(root, eb, start, err);
 242
 243         reada_start_machine(root->fs_info);
 244
 245         return ret;
 246 }
 247
 248 static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
 249                                           struct btrfs_device *dev, u64 logical,
 250                                           struct btrfs_bio *bbio)
 251 {
 252         int ret;
 253         struct reada_zone *zone;
 254         struct btrfs_block_group_cache *cache = NULL;
 255         u64 start;
 256         u64 end;
 257         int i;
 258
 259         zone = NULL;
 260         spin_lock(&fs_info->reada_lock);
 261         ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 262                                      logical >> PAGE_CACHE_SHIFT, 1);
 263         if (ret == 1)
 264                 kref_get(&zone->refcnt);
 265         spin_unlock(&fs_info->reada_lock);
 266
 267         if (ret == 1) {
 268                 if (logical >= zone->start && logical < zone->end)
 269                         return zone;
 270                 spin_lock(&fs_info->reada_lock);
 271                 kref_put(&zone->refcnt, reada_zone_release);
 272                 spin_unlock(&fs_info->reada_lock);
 273         }
 274
 275         cache = btrfs_lookup_block_group(fs_info, logical);
 276         if (!cache)
 277                 return NULL;
 278
 279         start = cache->key.objectid;
 280         end = start + cache->key.offset - 1;
 281         btrfs_put_block_group(cache);
 282
 283         zone = kzalloc(sizeof(*zone), GFP_NOFS);
 284         if (!zone)
 285                 return NULL;
 286
 287         zone->start = start;
 288         zone->end = end;
 289         INIT_LIST_HEAD(&zone->list);
 290         spin_lock_init(&zone->lock);
 291         zone->locked = 0;
 292         kref_init(&zone->refcnt);
 293         zone->elems = 0;
 294         zone->device = dev; /* our device always sits at index 0 */
 295         for (i = 0; i < bbio->num_stripes; ++i) {
 296                 /* bounds have already been checked */
 297                 zone->devs[i] = bbio->stripes[i].dev;
 298         }
 299         zone->ndevs = bbio->num_stripes;
 300
 301         spin_lock(&fs_info->reada_lock);
 302         ret = radix_tree_insert(&dev->reada_zones,
 303                                 (unsigned long)(zone->end >> PAGE_CACHE_SHIFT),
 304                                 zone);
 305
 306         if (ret == -EEXIST) {
 307                 kfree(zone);
 308                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 309                                              logical >> PAGE_CACHE_SHIFT, 1);
 310                 if (ret == 1)
 311                         kref_get(&zone->refcnt);
 312         }
 313         spin_unlock(&fs_info->reada_lock);
 314
 315         return zone;
 316 }
 317
 318 static struct reada_extent *reada_find_extent(struct btrfs_root *root,
 319                                               u64 logical,
 320                                               struct btrfs_key *top, int level)
 321 {
 322         int ret;
 323         struct reada_extent *re = NULL;
 324         struct reada_extent *re_exist = NULL;
 325         struct btrfs_fs_info *fs_info = root->fs_info;
 326         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
 327         struct btrfs_bio *bbio = NULL;
 328         struct btrfs_device *dev;
 329         struct btrfs_device *prev_dev;
 330         u32 blocksize;
 331         u64 length;
 332         int nzones = 0;
 333         int i;
 334         unsigned long index = logical >> PAGE_CACHE_SHIFT;
 335
 336         spin_lock(&fs_info->reada_lock);
 337         re = radix_tree_lookup(&fs_info->reada_tree, index);
 338         if (re)
 339                 kref_get(&re->refcnt);
 340         spin_unlock(&fs_info->reada_lock);
 341
 342         if (re)
 343                 return re;
 344
 345         re = kzalloc(sizeof(*re), GFP_NOFS);
 346         if (!re)
 347                 return NULL;
 348
 349         blocksize = btrfs_level_size(root, level);
 350         re->logical = logical;
 351         re->blocksize = blocksize;
 352         re->top = *top;
 353         INIT_LIST_HEAD(&re->extctl);
 354         spin_lock_init(&re->lock);
 355         kref_init(&re->refcnt);
 356
 357         /*
 358          * map block
 359          */
 360         length = blocksize;
 361         ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length, &bbio, 0);
 362         if (ret || !bbio || length < blocksize)
 363                 goto error;
 364
 365         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 366                 printk(KERN_ERR "btrfs readahead: more than %d copies not "
 367                                 "supported", BTRFS_MAX_MIRRORS);
 368                 goto error;
 369         }
 370
 371         for (nzones = 0; nzones < bbio->num_stripes; ++nzones) {
 372                 struct reada_zone *zone;
 373
 374                 dev = bbio->stripes[nzones].dev;
 375                 zone = reada_find_zone(fs_info, dev, logical, bbio);
 376                 if (!zone)
 377                         break;
 378
 379                 re->zones[nzones] = zone;
 380                 spin_lock(&zone->lock);
 381                 if (!zone->elems)
 382                         kref_get(&zone->refcnt);
 383                 ++zone->elems;
 384                 spin_unlock(&zone->lock);
 385                 spin_lock(&fs_info->reada_lock);
 386                 kref_put(&zone->refcnt, reada_zone_release);
 387                 spin_unlock(&fs_info->reada_lock);
 388         }
 389         re->nzones = nzones;
 390         if (nzones == 0) {
 391                 /* not a single zone found, error and out */
 392                 goto error;
 393         }
 394
 395         /* insert extent in reada_tree + all per-device trees, all or nothing */
 396         spin_lock(&fs_info->reada_lock);
 397         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 398         if (ret == -EEXIST) {
 399                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 400                 BUG_ON(!re_exist);
 401                 kref_get(&re_exist->refcnt);
 402                 spin_unlock(&fs_info->reada_lock);
 403                 goto error;
 404         }
 405         if (ret) {
 406                 spin_unlock(&fs_info->reada_lock);
 407                 goto error;
 408         }
 409         prev_dev = NULL;
 410         for (i = 0; i < nzones; ++i) {
 411                 dev = bbio->stripes[i].dev;
 412                 if (dev == prev_dev) {
 413                         /*
 414                          * in case of DUP, just add the first zone. As both
 415                          * are on the same device, there's nothing to gain
 416                          * from adding both.
 417                          * Also, it wouldn't work, as the tree is per device
 418                          * and adding would fail with EEXIST
 419                          */
 420                         continue;
 421                 }
 422                 prev_dev = dev;
 423                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 424                 if (ret) {
 425                         while (--i >= 0) {
 426                                 dev = bbio->stripes[i].dev;
 427                                 BUG_ON(dev == NULL);
 428                                 radix_tree_delete(&dev->reada_extents, index);
 429                         }
 430                         BUG_ON(fs_info == NULL);
 431                         radix_tree_delete(&fs_info->reada_tree, index);
 432                         spin_unlock(&fs_info->reada_lock);
 433                         goto error;
 434                 }
 435         }
 436         spin_unlock(&fs_info->reada_lock);
 437
 438         kfree(bbio);
 439         return re;
 440
 441 error:
 442         while (nzones) {
 443                 struct reada_zone *zone;
 444
 445                 --nzones;
 446                 zone = re->zones[nzones];
 447                 kref_get(&zone->refcnt);
 448                 spin_lock(&zone->lock);
 449                 --zone->elems;
 450                 if (zone->elems == 0) {
 451                         /*
 452                          * no fs_info->reada_lock needed, as this can't be
 453                          * the last ref
 454                          */
 455                         kref_put(&zone->refcnt, reada_zone_release);
 456                 }
 457                 spin_unlock(&zone->lock);
 458
 459                 spin_lock(&fs_info->reada_lock);
 460                 kref_put(&zone->refcnt, reada_zone_release);
 461                 spin_unlock(&fs_info->reada_lock);
 462         }
 463         kfree(bbio);
 464         kfree(re);
 465         return re_exist;
 466 }
 467
 468 static void reada_kref_dummy(struct kref *kr)
 469 {
 470 }
 471
 472 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 473                              struct reada_extent *re)
 474 {
 475         int i;
 476         unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
 477
 478         spin_lock(&fs_info->reada_lock);
 479         if (!kref_put(&re->refcnt, reada_kref_dummy)) {
 480                 spin_unlock(&fs_info->reada_lock);
 481                 return;
 482         }
 483
 484         radix_tree_delete(&fs_info->reada_tree, index);
 485         for (i = 0; i < re->nzones; ++i) {
 486                 struct reada_zone *zone = re->zones[i];
 487
 488                 radix_tree_delete(&zone->device->reada_extents, index);
 489         }
 490
 491         spin_unlock(&fs_info->reada_lock);
 492
 493         for (i = 0; i < re->nzones; ++i) {
 494                 struct reada_zone *zone = re->zones[i];
 495
 496                 kref_get(&zone->refcnt);
 497                 spin_lock(&zone->lock);
 498                 --zone->elems;
 499                 if (zone->elems == 0) {
 500                         /* no fs_info->reada_lock needed, as this can't be
 501                          * the last ref */
 502                         kref_put(&zone->refcnt, reada_zone_release);
 503                 }
 504                 spin_unlock(&zone->lock);
 505
 506                 spin_lock(&fs_info->reada_lock);
 507                 kref_put(&zone->refcnt, reada_zone_release);
 508                 spin_unlock(&fs_info->reada_lock);
 509         }
 510         if (re->scheduled_for)
 511                 atomic_dec(&re->scheduled_for->reada_in_flight);
 512
 513         kfree(re);
 514 }
 515
 516 static void reada_zone_release(struct kref *kref)
 517 {
 518         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 519
 520         radix_tree_delete(&zone->device->reada_zones,
 521                           zone->end >> PAGE_CACHE_SHIFT);
 522
 523         kfree(zone);
 524 }
 525
 526 static void reada_control_release(struct kref *kref)
 527 {
 528         struct reada_control *rc = container_of(kref, struct reada_control,
 529                                                 refcnt);
 530
 531         kfree(rc);
 532 }
 533
 534 static int reada_add_block(struct reada_control *rc, u64 logical,
 535                            struct btrfs_key *top, int level, u64 generation)
 536 {
 537         struct btrfs_root *root = rc->root;
 538         struct reada_extent *re;
 539         struct reada_extctl *rec;
 540
 541         re = reada_find_extent(root, logical, top, level); /* takes one ref */
 542         if (!re)
 543                 return -1;
 544
 545         rec = kzalloc(sizeof(*rec), GFP_NOFS);
 546         if (!rec) {
 547                 reada_extent_put(root->fs_info, re);
 548                 return -1;
 549         }
 550
 551         rec->rc = rc;
 552         rec->generation = generation;
 553         atomic_inc(&rc->elems);
 554
 555         spin_lock(&re->lock);
 556         list_add_tail(&rec->list, &re->extctl);
 557         spin_unlock(&re->lock);
 558
 559         /* leave the ref on the extent */
 560
 561         return 0;
 562 }
 563
 564 /*
 565  * called with fs_info->reada_lock held
 566  */
 567 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 568 {
 569         int i;
 570         unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
 571
 572         for (i = 0; i < zone->ndevs; ++i) {
 573                 struct reada_zone *peer;
 574                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 575                 if (peer && peer->device != zone->device)
 576                         peer->locked = lock;
 577         }
 578 }
 579
 580 /*
 581  * called with fs_info->reada_lock held
 582  */
 583 static int reada_pick_zone(struct btrfs_device *dev)
 584 {
 585         struct reada_zone *top_zone = NULL;
 586         struct reada_zone *top_locked_zone = NULL;
 587         u64 top_elems = 0;
 588         u64 top_locked_elems = 0;
 589         unsigned long index = 0;
 590         int ret;
 591
 592         if (dev->reada_curr_zone) {
 593                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 594                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 595                 dev->reada_curr_zone = NULL;
 596         }
 597         /* pick the zone with the most elements */
 598         while (1) {
 599                 struct reada_zone *zone;
 600
 601                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 602                                              (void **)&zone, index, 1);
 603                 if (ret == 0)
 604                         break;
 605                 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
 606                 if (zone->locked) {
 607                         if (zone->elems > top_locked_elems) {
 608                                 top_locked_elems = zone->elems;
 609                                 top_locked_zone = zone;
 610                         }
 611                 } else {
 612                         if (zone->elems > top_elems) {
 613                                 top_elems = zone->elems;
 614                                 top_zone = zone;
 615                         }
 616                 }
 617         }
 618         if (top_zone)
 619                 dev->reada_curr_zone = top_zone;
 620         else if (top_locked_zone)
 621                 dev->reada_curr_zone = top_locked_zone;
 622         else
 623                 return 0;
 624
 625         dev->reada_next = dev->reada_curr_zone->start;
 626         kref_get(&dev->reada_curr_zone->refcnt);
 627         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 628
 629         return 1;
 630 }
 631
 632 static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
 633                                    struct btrfs_device *dev)
 634 {
 635         struct reada_extent *re = NULL;
 636         int mirror_num = 0;
 637         struct extent_buffer *eb = NULL;
 638         u64 logical;
 639         u32 blocksize;
 640         int ret;
 641         int i;
 642         int need_kick = 0;
 643
 644         spin_lock(&fs_info->reada_lock);
 645         if (dev->reada_curr_zone == NULL) {
 646                 ret = reada_pick_zone(dev);
 647                 if (!ret) {
 648                         spin_unlock(&fs_info->reada_lock);
 649                         return 0;
 650                 }
 651         }
 652         /*
 653          * FIXME currently we issue the reads one extent at a time. If we have
 654          * a contiguous block of extents, we could also coagulate them or use
 655          * plugging to speed things up
 656          */
 657         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 658                                      dev->reada_next >> PAGE_CACHE_SHIFT, 1);
 659         if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
 660                 ret = reada_pick_zone(dev);
 661                 if (!ret) {
 662                         spin_unlock(&fs_info->reada_lock);
 663                         return 0;
 664                 }
 665                 re = NULL;
 666                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 667                                         dev->reada_next >> PAGE_CACHE_SHIFT, 1);
 668         }
 669         if (ret == 0) {
 670                 spin_unlock(&fs_info->reada_lock);
 671                 return 0;
 672         }
 673         dev->reada_next = re->logical + re->blocksize;
 674         kref_get(&re->refcnt);
 675
 676         spin_unlock(&fs_info->reada_lock);
 677
 678         /*
 679          * find mirror num
 680          */
 681         for (i = 0; i < re->nzones; ++i) {
 682                 if (re->zones[i]->device == dev) {
 683                         mirror_num = i + 1;
 684                         break;
 685                 }
 686         }
 687         logical = re->logical;
 688         blocksize = re->blocksize;
 689
 690         spin_lock(&re->lock);
 691         if (re->scheduled_for == NULL) {
 692                 re->scheduled_for = dev;
 693                 need_kick = 1;
 694         }
 695         spin_unlock(&re->lock);
 696
 697         reada_extent_put(fs_info, re);
 698
 699         if (!need_kick)
 700                 return 0;
 701
 702         atomic_inc(&dev->reada_in_flight);
 703         ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
 704                          mirror_num, &eb);
 705         if (ret)
 706                 __readahead_hook(fs_info->extent_root, NULL, logical, ret);
 707         else if (eb)
 708                 __readahead_hook(fs_info->extent_root, eb, eb->start, ret);
 709
 710         if (eb)
 711                 free_extent_buffer(eb);
 712
 713         return 1;
 714
 715 }
 716
 717 static void reada_start_machine_worker(struct btrfs_work *work)
 718 {
 719         struct reada_machine_work *rmw;
 720         struct btrfs_fs_info *fs_info;
 721
 722         rmw = container_of(work, struct reada_machine_work, work);
 723         fs_info = rmw->fs_info;
 724
 725         kfree(rmw);
 726
 727         __reada_start_machine(fs_info);
 728 }
 729
 730 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 731 {
 732         struct btrfs_device *device;
 733         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 734         u64 enqueued;
 735         u64 total = 0;
 736         int i;
 737
 738         do {
 739                 enqueued = 0;
 740                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 741                         if (atomic_read(&device->reada_in_flight) <
 742                             MAX_IN_FLIGHT)
 743                                 enqueued += reada_start_machine_dev(fs_info,
 744                                                                     device);
 745                 }
 746                 total += enqueued;
 747         } while (enqueued && total < 10000);
 748
 749         if (enqueued == 0)
 750                 return;
 751
 752         /*
 753          * If everything is already in the cache, this is effectively single
 754          * threaded. To a) not hold the caller for too long and b) to utilize
 755          * more cores, we broke the loop above after 10000 iterations and now
 756          * enqueue to workers to finish it. This will distribute the load to
 757          * the cores.
 758          */
 759         for (i = 0; i < 2; ++i)
 760                 reada_start_machine(fs_info);
 761 }
 762
 763 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 764 {
 765         struct reada_machine_work *rmw;
 766
 767         rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
 768         if (!rmw) {
 769                 /* FIXME we cannot handle this properly right now */
 770                 BUG();
 771         }
 772         rmw->work.func = reada_start_machine_worker;
 773         rmw->fs_info = fs_info;
 774
 775         btrfs_queue_worker(&fs_info->readahead_workers, &rmw->work);
 776 }
 777
 778 #ifdef DEBUG
 779 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 780 {
 781         struct btrfs_device *device;
 782         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 783         unsigned long index;
 784         int ret;
 785         int i;
 786         int j;
 787         int cnt;
 788
 789         spin_lock(&fs_info->reada_lock);
 790         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 791                 printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
 792                         atomic_read(&device->reada_in_flight));
 793                 index = 0;
 794                 while (1) {
 795                         struct reada_zone *zone;
 796                         ret = radix_tree_gang_lookup(&device->reada_zones,
 797                                                      (void **)&zone, index, 1);
 798                         if (ret == 0)
 799                                 break;
 800                         printk(KERN_DEBUG "  zone %llu-%llu elems %llu locked "
 801                                 "%d devs", zone->start, zone->end, zone->elems,
 802                                 zone->locked);
 803                         for (j = 0; j < zone->ndevs; ++j) {
 804                                 printk(KERN_CONT " %lld",
 805                                         zone->devs[j]->devid);
 806                         }
 807                         if (device->reada_curr_zone == zone)
 808                                 printk(KERN_CONT " curr off %llu",
 809                                         device->reada_next - zone->start);
 810                         printk(KERN_CONT "\n");
 811                         index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
 812                 }
 813                 cnt = 0;
 814                 index = 0;
 815                 while (all) {
 816                         struct reada_extent *re = NULL;
 817
 818                         ret = radix_tree_gang_lookup(&device->reada_extents,
 819                                                      (void **)&re, index, 1);
 820                         if (ret == 0)
 821                                 break;
 822                         printk(KERN_DEBUG
 823                                 "  re: logical %llu size %u empty %d for %lld",
 824                                 re->logical, re->blocksize,
 825                                 list_empty(&re->extctl), re->scheduled_for ?
 826                                 re->scheduled_for->devid : -1);
 827
 828                         for (i = 0; i < re->nzones; ++i) {
 829                                 printk(KERN_CONT " zone %llu-%llu devs",
 830                                         re->zones[i]->start,
 831                                         re->zones[i]->end);
 832                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 833                                         printk(KERN_CONT " %lld",
 834                                                 re->zones[i]->devs[j]->devid);
 835                                 }
 836                         }
 837                         printk(KERN_CONT "\n");
 838                         index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
 839                         if (++cnt > 15)
 840                                 break;
 841                 }
 842         }
 843
 844         index = 0;
 845         cnt = 0;
 846         while (all) {
 847                 struct reada_extent *re = NULL;
 848
 849                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 850                                              index, 1);
 851                 if (ret == 0)
 852                         break;
 853                 if (!re->scheduled_for) {
 854                         index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
 855                         continue;
 856                 }
 857                 printk(KERN_DEBUG
 858                         "re: logical %llu size %u list empty %d for %lld",
 859                         re->logical, re->blocksize, list_empty(&re->extctl),
 860                         re->scheduled_for ? re->scheduled_for->devid : -1);
 861                 for (i = 0; i < re->nzones; ++i) {
 862                         printk(KERN_CONT " zone %llu-%llu devs",
 863                                 re->zones[i]->start,
 864                                 re->zones[i]->end);
 865                         for (i = 0; i < re->nzones; ++i) {
 866                                 printk(KERN_CONT " zone %llu-%llu devs",
 867                                         re->zones[i]->start,
 868                                         re->zones[i]->end);
 869                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 870                                         printk(KERN_CONT " %lld",
 871                                                 re->zones[i]->devs[j]->devid);
 872                                 }
 873                         }
 874                 }
 875                 printk(KERN_CONT "\n");
 876                 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
 877         }
 878         spin_unlock(&fs_info->reada_lock);
 879 }
 880 #endif
 881
 882 /*
 883  * interface
 884  */
 885 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 886                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 887 {
 888         struct reada_control *rc;
 889         u64 start;
 890         u64 generation;
 891         int level;
 892         struct extent_buffer *node;
 893         static struct btrfs_key max_key = {
 894                 .objectid = (u64)-1,
 895                 .type = (u8)-1,
 896                 .offset = (u64)-1
 897         };
 898
 899         rc = kzalloc(sizeof(*rc), GFP_NOFS);
 900         if (!rc)
 901                 return ERR_PTR(-ENOMEM);
 902
 903         rc->root = root;
 904         rc->key_start = *key_start;
 905         rc->key_end = *key_end;
 906         atomic_set(&rc->elems, 0);
 907         init_waitqueue_head(&rc->wait);
 908         kref_init(&rc->refcnt);
 909         kref_get(&rc->refcnt); /* one ref for having elements */
 910
 911         node = btrfs_root_node(root);
 912         start = node->start;
 913         level = btrfs_header_level(node);
 914         generation = btrfs_header_generation(node);
 915         free_extent_buffer(node);
 916
 917         reada_add_block(rc, start, &max_key, level, generation);
 918
 919         reada_start_machine(root->fs_info);
 920
 921         return rc;
 922 }
 923
 924 #ifdef DEBUG
 925 int btrfs_reada_wait(void *handle)
 926 {
 927         struct reada_control *rc = handle;
 928
 929         while (atomic_read(&rc->elems)) {
 930                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 931                                    5 * HZ);
 932                 dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
 933         }
 934
 935         dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
 936
 937         kref_put(&rc->refcnt, reada_control_release);
 938
 939         return 0;
 940 }
 941 #else
 942 int btrfs_reada_wait(void *handle)
 943 {
 944         struct reada_control *rc = handle;
 945
 946         while (atomic_read(&rc->elems)) {
 947                 wait_event(rc->wait, atomic_read(&rc->elems) == 0);
 948         }
 949
 950         kref_put(&rc->refcnt, reada_control_release);
 951
 952         return 0;
 953 }
 954 #endif
 955
 956 void btrfs_reada_detach(void *handle)
 957 {
 958         struct reada_control *rc = handle;
 959
 960         kref_put(&rc->refcnt, reada_control_release);
 961 }