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 #include "dev-replace.h"
  31
  32 #undef DEBUG
  33
  34 /*
  35  * This is the implementation for the generic read ahead framework.
  36  *
  37  * To trigger a readahead, btrfs_reada_add must be called. It will start
  38  * a read ahead for the given range [start, end) on tree root. The returned
  39  * handle can either be used to wait on the readahead to finish
  40  * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
  41  *
  42  * The read ahead works as follows:
  43  * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
  44  * reada_start_machine will then search for extents to prefetch and trigger
  45  * some reads. When a read finishes for a node, all contained node/leaf
  46  * pointers that lie in the given range will also be enqueued. The reads will
  47  * be triggered in sequential order, thus giving a big win over a naive
  48  * enumeration. It will also make use of multi-device layouts. Each disk
  49  * will have its on read pointer and all disks will by utilized in parallel.
  50  * Also will no two disks read both sides of a mirror simultaneously, as this
  51  * would waste seeking capacity. Instead both disks will read different parts
  52  * of the filesystem.
  53  * Any number of readaheads can be started in parallel. The read order will be
  54  * determined globally, i.e. 2 parallel readaheads will normally finish faster
  55  * than the 2 started one after another.
  56  */
  57
  58 #define MAX_IN_FLIGHT 6
  59
  60 struct reada_extctl {
  61         struct list_head        list;
  62         struct reada_control    *rc;
  63         u64                     generation;
  64 };
  65
  66 struct reada_extent {
  67         u64                     logical;
  68         struct btrfs_key        top;
  69         int                     err;
  70         struct list_head        extctl;
  71         int                     refcnt;
  72         spinlock_t              lock;
  73         struct reada_zone       *zones[BTRFS_MAX_MIRRORS];
  74         int                     nzones;
  75         int                     scheduled;
  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, u64 generation);
 105
 106 /* recurses */
 107 /* in case of err, eb might be NULL */
 108 static void __readahead_hook(struct btrfs_fs_info *fs_info,
 109                              struct reada_extent *re, struct extent_buffer *eb,
 110                              int err)
 111 {
 112         int nritems;
 113         int i;
 114         u64 bytenr;
 115         u64 generation;
 116         struct list_head list;
 117
 118         spin_lock(&re->lock);
 119         /*
 120          * just take the full list from the extent. afterwards we
 121          * don't need the lock anymore
 122          */
 123         list_replace_init(&re->extctl, &list);
 124         re->scheduled = 0;
 125         spin_unlock(&re->lock);
 126
 127         /*
 128          * this is the error case, the extent buffer has not been
 129          * read correctly. We won't access anything from it and
 130          * just cleanup our data structures. Effectively this will
 131          * cut the branch below this node from read ahead.
 132          */
 133         if (err)
 134                 goto cleanup;
 135
 136         /*
 137          * FIXME: currently we just set nritems to 0 if this is a leaf,
 138          * effectively ignoring the content. In a next step we could
 139          * trigger more readahead depending from the content, e.g.
 140          * fetch the checksums for the extents in the leaf.
 141          */
 142         if (!btrfs_header_level(eb))
 143                 goto cleanup;
 144
 145         nritems = btrfs_header_nritems(eb);
 146         generation = btrfs_header_generation(eb);
 147         for (i = 0; i < nritems; i++) {
 148                 struct reada_extctl *rec;
 149                 u64 n_gen;
 150                 struct btrfs_key key;
 151                 struct btrfs_key next_key;
 152
 153                 btrfs_node_key_to_cpu(eb, &key, i);
 154                 if (i + 1 < nritems)
 155                         btrfs_node_key_to_cpu(eb, &next_key, i + 1);
 156                 else
 157                         next_key = re->top;
 158                 bytenr = btrfs_node_blockptr(eb, i);
 159                 n_gen = btrfs_node_ptr_generation(eb, i);
 160
 161                 list_for_each_entry(rec, &list, list) {
 162                         struct reada_control *rc = rec->rc;
 163
 164                         /*
 165                          * if the generation doesn't match, just ignore this
 166                          * extctl. This will probably cut off a branch from
 167                          * prefetch. Alternatively one could start a new (sub-)
 168                          * prefetch for this branch, starting again from root.
 169                          * FIXME: move the generation check out of this loop
 170                          */
 171 #ifdef DEBUG
 172                         if (rec->generation != generation) {
 173                                 btrfs_debug(fs_info,
 174                                             "generation mismatch for (%llu,%d,%llu) %llu != %llu",
 175                                             key.objectid, key.type, key.offset,
 176                                             rec->generation, generation);
 177                         }
 178 #endif
 179                         if (rec->generation == generation &&
 180                             btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
 181                             btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
 182                                 reada_add_block(rc, bytenr, &next_key, n_gen);
 183                 }
 184         }
 185
 186 cleanup:
 187         /*
 188          * free extctl records
 189          */
 190         while (!list_empty(&list)) {
 191                 struct reada_control *rc;
 192                 struct reada_extctl *rec;
 193
 194                 rec = list_first_entry(&list, struct reada_extctl, list);
 195                 list_del(&rec->list);
 196                 rc = rec->rc;
 197                 kfree(rec);
 198
 199                 kref_get(&rc->refcnt);
 200                 if (atomic_dec_and_test(&rc->elems)) {
 201                         kref_put(&rc->refcnt, reada_control_release);
 202                         wake_up(&rc->wait);
 203                 }
 204                 kref_put(&rc->refcnt, reada_control_release);
 205
 206                 reada_extent_put(fs_info, re);  /* one ref for each entry */
 207         }
 208
 209         return;
 210 }
 211
 212 int btree_readahead_hook(struct btrfs_fs_info *fs_info,
 213                          struct extent_buffer *eb, int err)
 214 {
 215         int ret = 0;
 216         struct reada_extent *re;
 217
 218         /* find extent */
 219         spin_lock(&fs_info->reada_lock);
 220         re = radix_tree_lookup(&fs_info->reada_tree,
 221                                eb->start >> PAGE_SHIFT);
 222         if (re)
 223                 re->refcnt++;
 224         spin_unlock(&fs_info->reada_lock);
 225         if (!re) {
 226                 ret = -1;
 227                 goto start_machine;
 228         }
 229
 230         __readahead_hook(fs_info, re, eb, err);
 231         reada_extent_put(fs_info, re);  /* our ref */
 232
 233 start_machine:
 234         reada_start_machine(fs_info);
 235         return ret;
 236 }
 237
 238 static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
 239                                           struct btrfs_device *dev, u64 logical,
 240                                           struct btrfs_bio *bbio)
 241 {
 242         int ret;
 243         struct reada_zone *zone;
 244         struct btrfs_block_group_cache *cache = NULL;
 245         u64 start;
 246         u64 end;
 247         int i;
 248
 249         zone = NULL;
 250         spin_lock(&fs_info->reada_lock);
 251         ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 252                                      logical >> PAGE_SHIFT, 1);
 253         if (ret == 1 && logical >= zone->start && logical <= zone->end) {
 254                 kref_get(&zone->refcnt);
 255                 spin_unlock(&fs_info->reada_lock);
 256                 return zone;
 257         }
 258
 259         spin_unlock(&fs_info->reada_lock);
 260
 261         cache = btrfs_lookup_block_group(fs_info, logical);
 262         if (!cache)
 263                 return NULL;
 264
 265         start = cache->key.objectid;
 266         end = start + cache->key.offset - 1;
 267         btrfs_put_block_group(cache);
 268
 269         zone = kzalloc(sizeof(*zone), GFP_KERNEL);
 270         if (!zone)
 271                 return NULL;
 272
 273         zone->start = start;
 274         zone->end = end;
 275         INIT_LIST_HEAD(&zone->list);
 276         spin_lock_init(&zone->lock);
 277         zone->locked = 0;
 278         kref_init(&zone->refcnt);
 279         zone->elems = 0;
 280         zone->device = dev; /* our device always sits at index 0 */
 281         for (i = 0; i < bbio->num_stripes; ++i) {
 282                 /* bounds have already been checked */
 283                 zone->devs[i] = bbio->stripes[i].dev;
 284         }
 285         zone->ndevs = bbio->num_stripes;
 286
 287         spin_lock(&fs_info->reada_lock);
 288         ret = radix_tree_insert(&dev->reada_zones,
 289                                 (unsigned long)(zone->end >> PAGE_SHIFT),
 290                                 zone);
 291
 292         if (ret == -EEXIST) {
 293                 kfree(zone);
 294                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 295                                              logical >> PAGE_SHIFT, 1);
 296                 if (ret == 1 && logical >= zone->start && logical <= zone->end)
 297                         kref_get(&zone->refcnt);
 298                 else
 299                         zone = NULL;
 300         }
 301         spin_unlock(&fs_info->reada_lock);
 302
 303         return zone;
 304 }
 305
 306 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
 307                                               u64 logical,
 308                                               struct btrfs_key *top)
 309 {
 310         int ret;
 311         struct reada_extent *re = NULL;
 312         struct reada_extent *re_exist = NULL;
 313         struct btrfs_bio *bbio = NULL;
 314         struct btrfs_device *dev;
 315         struct btrfs_device *prev_dev;
 316         u32 blocksize;
 317         u64 length;
 318         int real_stripes;
 319         int nzones = 0;
 320         unsigned long index = logical >> PAGE_SHIFT;
 321         int dev_replace_is_ongoing;
 322         int have_zone = 0;
 323
 324         spin_lock(&fs_info->reada_lock);
 325         re = radix_tree_lookup(&fs_info->reada_tree, index);
 326         if (re)
 327                 re->refcnt++;
 328         spin_unlock(&fs_info->reada_lock);
 329
 330         if (re)
 331                 return re;
 332
 333         re = kzalloc(sizeof(*re), GFP_KERNEL);
 334         if (!re)
 335                 return NULL;
 336
 337         blocksize = fs_info->nodesize;
 338         re->logical = logical;
 339         re->top = *top;
 340         INIT_LIST_HEAD(&re->extctl);
 341         spin_lock_init(&re->lock);
 342         re->refcnt = 1;
 343
 344         /*
 345          * map block
 346          */
 347         length = blocksize;
 348         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
 349                         &length, &bbio, 0);
 350         if (ret || !bbio || length < blocksize)
 351                 goto error;
 352
 353         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 354                 btrfs_err(fs_info,
 355                            "readahead: more than %d copies not supported",
 356                            BTRFS_MAX_MIRRORS);
 357                 goto error;
 358         }
 359
 360         real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
 361         for (nzones = 0; nzones < real_stripes; ++nzones) {
 362                 struct reada_zone *zone;
 363
 364                 dev = bbio->stripes[nzones].dev;
 365
 366                 /* cannot read ahead on missing device. */
 367                  if (!dev->bdev)
 368                         continue;
 369
 370                 zone = reada_find_zone(fs_info, dev, logical, bbio);
 371                 if (!zone)
 372                         continue;
 373
 374                 re->zones[re->nzones++] = zone;
 375                 spin_lock(&zone->lock);
 376                 if (!zone->elems)
 377                         kref_get(&zone->refcnt);
 378                 ++zone->elems;
 379                 spin_unlock(&zone->lock);
 380                 spin_lock(&fs_info->reada_lock);
 381                 kref_put(&zone->refcnt, reada_zone_release);
 382                 spin_unlock(&fs_info->reada_lock);
 383         }
 384         if (re->nzones == 0) {
 385                 /* not a single zone found, error and out */
 386                 goto error;
 387         }
 388
 389         /* insert extent in reada_tree + all per-device trees, all or nothing */
 390         btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
 391         spin_lock(&fs_info->reada_lock);
 392         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 393         if (ret == -EEXIST) {
 394                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 395                 re_exist->refcnt++;
 396                 spin_unlock(&fs_info->reada_lock);
 397                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 398                 goto error;
 399         }
 400         if (ret) {
 401                 spin_unlock(&fs_info->reada_lock);
 402                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 403                 goto error;
 404         }
 405         prev_dev = NULL;
 406         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 407                         &fs_info->dev_replace);
 408         for (nzones = 0; nzones < re->nzones; ++nzones) {
 409                 dev = re->zones[nzones]->device;
 410
 411                 if (dev == prev_dev) {
 412                         /*
 413                          * in case of DUP, just add the first zone. As both
 414                          * are on the same device, there's nothing to gain
 415                          * from adding both.
 416                          * Also, it wouldn't work, as the tree is per device
 417                          * and adding would fail with EEXIST
 418                          */
 419                         continue;
 420                 }
 421                 if (!dev->bdev)
 422                         continue;
 423
 424                 if (dev_replace_is_ongoing &&
 425                     dev == fs_info->dev_replace.tgtdev) {
 426                         /*
 427                          * as this device is selected for reading only as
 428                          * a last resort, skip it for read ahead.
 429                          */
 430                         continue;
 431                 }
 432                 prev_dev = dev;
 433                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 434                 if (ret) {
 435                         while (--nzones >= 0) {
 436                                 dev = re->zones[nzones]->device;
 437                                 BUG_ON(dev == NULL);
 438                                 /* ignore whether the entry was inserted */
 439                                 radix_tree_delete(&dev->reada_extents, index);
 440                         }
 441                         radix_tree_delete(&fs_info->reada_tree, index);
 442                         spin_unlock(&fs_info->reada_lock);
 443                         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 444                         goto error;
 445                 }
 446                 have_zone = 1;
 447         }
 448         spin_unlock(&fs_info->reada_lock);
 449         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 450
 451         if (!have_zone)
 452                 goto error;
 453
 454         btrfs_put_bbio(bbio);
 455         return re;
 456
 457 error:
 458         for (nzones = 0; nzones < re->nzones; ++nzones) {
 459                 struct reada_zone *zone;
 460
 461                 zone = re->zones[nzones];
 462                 kref_get(&zone->refcnt);
 463                 spin_lock(&zone->lock);
 464                 --zone->elems;
 465                 if (zone->elems == 0) {
 466                         /*
 467                          * no fs_info->reada_lock needed, as this can't be
 468                          * the last ref
 469                          */
 470                         kref_put(&zone->refcnt, reada_zone_release);
 471                 }
 472                 spin_unlock(&zone->lock);
 473
 474                 spin_lock(&fs_info->reada_lock);
 475                 kref_put(&zone->refcnt, reada_zone_release);
 476                 spin_unlock(&fs_info->reada_lock);
 477         }
 478         btrfs_put_bbio(bbio);
 479         kfree(re);
 480         return re_exist;
 481 }
 482
 483 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 484                              struct reada_extent *re)
 485 {
 486         int i;
 487         unsigned long index = re->logical >> PAGE_SHIFT;
 488
 489         spin_lock(&fs_info->reada_lock);
 490         if (--re->refcnt) {
 491                 spin_unlock(&fs_info->reada_lock);
 492                 return;
 493         }
 494
 495         radix_tree_delete(&fs_info->reada_tree, index);
 496         for (i = 0; i < re->nzones; ++i) {
 497                 struct reada_zone *zone = re->zones[i];
 498
 499                 radix_tree_delete(&zone->device->reada_extents, index);
 500         }
 501
 502         spin_unlock(&fs_info->reada_lock);
 503
 504         for (i = 0; i < re->nzones; ++i) {
 505                 struct reada_zone *zone = re->zones[i];
 506
 507                 kref_get(&zone->refcnt);
 508                 spin_lock(&zone->lock);
 509                 --zone->elems;
 510                 if (zone->elems == 0) {
 511                         /* no fs_info->reada_lock needed, as this can't be
 512                          * the last ref */
 513                         kref_put(&zone->refcnt, reada_zone_release);
 514                 }
 515                 spin_unlock(&zone->lock);
 516
 517                 spin_lock(&fs_info->reada_lock);
 518                 kref_put(&zone->refcnt, reada_zone_release);
 519                 spin_unlock(&fs_info->reada_lock);
 520         }
 521
 522         kfree(re);
 523 }
 524
 525 static void reada_zone_release(struct kref *kref)
 526 {
 527         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 528
 529         radix_tree_delete(&zone->device->reada_zones,
 530                           zone->end >> PAGE_SHIFT);
 531
 532         kfree(zone);
 533 }
 534
 535 static void reada_control_release(struct kref *kref)
 536 {
 537         struct reada_control *rc = container_of(kref, struct reada_control,
 538                                                 refcnt);
 539
 540         kfree(rc);
 541 }
 542
 543 static int reada_add_block(struct reada_control *rc, u64 logical,
 544                            struct btrfs_key *top, u64 generation)
 545 {
 546         struct btrfs_fs_info *fs_info = rc->fs_info;
 547         struct reada_extent *re;
 548         struct reada_extctl *rec;
 549
 550         /* takes one ref */
 551         re = reada_find_extent(fs_info, logical, top);
 552         if (!re)
 553                 return -1;
 554
 555         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 556         if (!rec) {
 557                 reada_extent_put(fs_info, re);
 558                 return -ENOMEM;
 559         }
 560
 561         rec->rc = rc;
 562         rec->generation = generation;
 563         atomic_inc(&rc->elems);
 564
 565         spin_lock(&re->lock);
 566         list_add_tail(&rec->list, &re->extctl);
 567         spin_unlock(&re->lock);
 568
 569         /* leave the ref on the extent */
 570
 571         return 0;
 572 }
 573
 574 /*
 575  * called with fs_info->reada_lock held
 576  */
 577 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 578 {
 579         int i;
 580         unsigned long index = zone->end >> PAGE_SHIFT;
 581
 582         for (i = 0; i < zone->ndevs; ++i) {
 583                 struct reada_zone *peer;
 584                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 585                 if (peer && peer->device != zone->device)
 586                         peer->locked = lock;
 587         }
 588 }
 589
 590 /*
 591  * called with fs_info->reada_lock held
 592  */
 593 static int reada_pick_zone(struct btrfs_device *dev)
 594 {
 595         struct reada_zone *top_zone = NULL;
 596         struct reada_zone *top_locked_zone = NULL;
 597         u64 top_elems = 0;
 598         u64 top_locked_elems = 0;
 599         unsigned long index = 0;
 600         int ret;
 601
 602         if (dev->reada_curr_zone) {
 603                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 604                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 605                 dev->reada_curr_zone = NULL;
 606         }
 607         /* pick the zone with the most elements */
 608         while (1) {
 609                 struct reada_zone *zone;
 610
 611                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 612                                              (void **)&zone, index, 1);
 613                 if (ret == 0)
 614                         break;
 615                 index = (zone->end >> PAGE_SHIFT) + 1;
 616                 if (zone->locked) {
 617                         if (zone->elems > top_locked_elems) {
 618                                 top_locked_elems = zone->elems;
 619                                 top_locked_zone = zone;
 620                         }
 621                 } else {
 622                         if (zone->elems > top_elems) {
 623                                 top_elems = zone->elems;
 624                                 top_zone = zone;
 625                         }
 626                 }
 627         }
 628         if (top_zone)
 629                 dev->reada_curr_zone = top_zone;
 630         else if (top_locked_zone)
 631                 dev->reada_curr_zone = top_locked_zone;
 632         else
 633                 return 0;
 634
 635         dev->reada_next = dev->reada_curr_zone->start;
 636         kref_get(&dev->reada_curr_zone->refcnt);
 637         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 638
 639         return 1;
 640 }
 641
 642 static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
 643                                    struct btrfs_device *dev)
 644 {
 645         struct reada_extent *re = NULL;
 646         int mirror_num = 0;
 647         struct extent_buffer *eb = NULL;
 648         u64 logical;
 649         int ret;
 650         int i;
 651
 652         spin_lock(&fs_info->reada_lock);
 653         if (dev->reada_curr_zone == NULL) {
 654                 ret = reada_pick_zone(dev);
 655                 if (!ret) {
 656                         spin_unlock(&fs_info->reada_lock);
 657                         return 0;
 658                 }
 659         }
 660         /*
 661          * FIXME currently we issue the reads one extent at a time. If we have
 662          * a contiguous block of extents, we could also coagulate them or use
 663          * plugging to speed things up
 664          */
 665         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 666                                      dev->reada_next >> PAGE_SHIFT, 1);
 667         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 668                 ret = reada_pick_zone(dev);
 669                 if (!ret) {
 670                         spin_unlock(&fs_info->reada_lock);
 671                         return 0;
 672                 }
 673                 re = NULL;
 674                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 675                                         dev->reada_next >> PAGE_SHIFT, 1);
 676         }
 677         if (ret == 0) {
 678                 spin_unlock(&fs_info->reada_lock);
 679                 return 0;
 680         }
 681         dev->reada_next = re->logical + fs_info->nodesize;
 682         re->refcnt++;
 683
 684         spin_unlock(&fs_info->reada_lock);
 685
 686         spin_lock(&re->lock);
 687         if (re->scheduled || list_empty(&re->extctl)) {
 688                 spin_unlock(&re->lock);
 689                 reada_extent_put(fs_info, re);
 690                 return 0;
 691         }
 692         re->scheduled = 1;
 693         spin_unlock(&re->lock);
 694
 695         /*
 696          * find mirror num
 697          */
 698         for (i = 0; i < re->nzones; ++i) {
 699                 if (re->zones[i]->device == dev) {
 700                         mirror_num = i + 1;
 701                         break;
 702                 }
 703         }
 704         logical = re->logical;
 705
 706         atomic_inc(&dev->reada_in_flight);
 707         ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
 708         if (ret)
 709                 __readahead_hook(fs_info, re, NULL, ret);
 710         else if (eb)
 711                 __readahead_hook(fs_info, re, eb, ret);
 712
 713         if (eb)
 714                 free_extent_buffer(eb);
 715
 716         atomic_dec(&dev->reada_in_flight);
 717         reada_extent_put(fs_info, re);
 718
 719         return 1;
 720
 721 }
 722
 723 static void reada_start_machine_worker(struct btrfs_work *work)
 724 {
 725         struct reada_machine_work *rmw;
 726         struct btrfs_fs_info *fs_info;
 727         int old_ioprio;
 728
 729         rmw = container_of(work, struct reada_machine_work, work);
 730         fs_info = rmw->fs_info;
 731
 732         kfree(rmw);
 733
 734         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 735                                        task_nice_ioprio(current));
 736         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 737         __reada_start_machine(fs_info);
 738         set_task_ioprio(current, old_ioprio);
 739
 740         atomic_dec(&fs_info->reada_works_cnt);
 741 }
 742
 743 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 744 {
 745         struct btrfs_device *device;
 746         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 747         u64 enqueued;
 748         u64 total = 0;
 749         int i;
 750
 751         do {
 752                 enqueued = 0;
 753                 mutex_lock(&fs_devices->device_list_mutex);
 754                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 755                         if (atomic_read(&device->reada_in_flight) <
 756                             MAX_IN_FLIGHT)
 757                                 enqueued += reada_start_machine_dev(fs_info,
 758                                                                     device);
 759                 }
 760                 mutex_unlock(&fs_devices->device_list_mutex);
 761                 total += enqueued;
 762         } while (enqueued && total < 10000);
 763
 764         if (enqueued == 0)
 765                 return;
 766
 767         /*
 768          * If everything is already in the cache, this is effectively single
 769          * threaded. To a) not hold the caller for too long and b) to utilize
 770          * more cores, we broke the loop above after 10000 iterations and now
 771          * enqueue to workers to finish it. This will distribute the load to
 772          * the cores.
 773          */
 774         for (i = 0; i < 2; ++i) {
 775                 reada_start_machine(fs_info);
 776                 if (atomic_read(&fs_info->reada_works_cnt) >
 777                     BTRFS_MAX_MIRRORS * 2)
 778                         break;
 779         }
 780 }
 781
 782 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 783 {
 784         struct reada_machine_work *rmw;
 785
 786         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 787         if (!rmw) {
 788                 /* FIXME we cannot handle this properly right now */
 789                 BUG();
 790         }
 791         btrfs_init_work(&rmw->work, btrfs_readahead_helper,
 792                         reada_start_machine_worker, NULL, NULL);
 793         rmw->fs_info = fs_info;
 794
 795         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 796         atomic_inc(&fs_info->reada_works_cnt);
 797 }
 798
 799 #ifdef DEBUG
 800 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 801 {
 802         struct btrfs_device *device;
 803         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 804         unsigned long index;
 805         int ret;
 806         int i;
 807         int j;
 808         int cnt;
 809
 810         spin_lock(&fs_info->reada_lock);
 811         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 812                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 813                         atomic_read(&device->reada_in_flight));
 814                 index = 0;
 815                 while (1) {
 816                         struct reada_zone *zone;
 817                         ret = radix_tree_gang_lookup(&device->reada_zones,
 818                                                      (void **)&zone, index, 1);
 819                         if (ret == 0)
 820                                 break;
 821                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 822                                     zone->start, zone->end, zone->elems,
 823                                     zone->locked);
 824                         for (j = 0; j < zone->ndevs; ++j) {
 825                                 pr_cont(" %lld",
 826                                         zone->devs[j]->devid);
 827                         }
 828                         if (device->reada_curr_zone == zone)
 829                                 pr_cont(" curr off %llu",
 830                                         device->reada_next - zone->start);
 831                         pr_cont("\n");
 832                         index = (zone->end >> PAGE_SHIFT) + 1;
 833                 }
 834                 cnt = 0;
 835                 index = 0;
 836                 while (all) {
 837                         struct reada_extent *re = NULL;
 838
 839                         ret = radix_tree_gang_lookup(&device->reada_extents,
 840                                                      (void **)&re, index, 1);
 841                         if (ret == 0)
 842                                 break;
 843                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 844                                 re->logical, fs_info->nodesize,
 845                                 list_empty(&re->extctl), re->scheduled);
 846
 847                         for (i = 0; i < re->nzones; ++i) {
 848                                 pr_cont(" zone %llu-%llu devs",
 849                                         re->zones[i]->start,
 850                                         re->zones[i]->end);
 851                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 852                                         pr_cont(" %lld",
 853                                                 re->zones[i]->devs[j]->devid);
 854                                 }
 855                         }
 856                         pr_cont("\n");
 857                         index = (re->logical >> PAGE_SHIFT) + 1;
 858                         if (++cnt > 15)
 859                                 break;
 860                 }
 861         }
 862
 863         index = 0;
 864         cnt = 0;
 865         while (all) {
 866                 struct reada_extent *re = NULL;
 867
 868                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 869                                              index, 1);
 870                 if (ret == 0)
 871                         break;
 872                 if (!re->scheduled) {
 873                         index = (re->logical >> PAGE_SHIFT) + 1;
 874                         continue;
 875                 }
 876                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 877                         re->logical, fs_info->nodesize,
 878                         list_empty(&re->extctl), re->scheduled);
 879                 for (i = 0; i < re->nzones; ++i) {
 880                         pr_cont(" zone %llu-%llu devs",
 881                                 re->zones[i]->start,
 882                                 re->zones[i]->end);
 883                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 884                                 pr_cont(" %lld",
 885                                        re->zones[i]->devs[j]->devid);
 886                         }
 887                 }
 888                 pr_cont("\n");
 889                 index = (re->logical >> PAGE_SHIFT) + 1;
 890         }
 891         spin_unlock(&fs_info->reada_lock);
 892 }
 893 #endif
 894
 895 /*
 896  * interface
 897  */
 898 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 899                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 900 {
 901         struct reada_control *rc;
 902         u64 start;
 903         u64 generation;
 904         int ret;
 905         struct extent_buffer *node;
 906         static struct btrfs_key max_key = {
 907                 .objectid = (u64)-1,
 908                 .type = (u8)-1,
 909                 .offset = (u64)-1
 910         };
 911
 912         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 913         if (!rc)
 914                 return ERR_PTR(-ENOMEM);
 915
 916         rc->fs_info = root->fs_info;
 917         rc->key_start = *key_start;
 918         rc->key_end = *key_end;
 919         atomic_set(&rc->elems, 0);
 920         init_waitqueue_head(&rc->wait);
 921         kref_init(&rc->refcnt);
 922         kref_get(&rc->refcnt); /* one ref for having elements */
 923
 924         node = btrfs_root_node(root);
 925         start = node->start;
 926         generation = btrfs_header_generation(node);
 927         free_extent_buffer(node);
 928
 929         ret = reada_add_block(rc, start, &max_key, generation);
 930         if (ret) {
 931                 kfree(rc);
 932                 return ERR_PTR(ret);
 933         }
 934
 935         reada_start_machine(root->fs_info);
 936
 937         return rc;
 938 }
 939
 940 #ifdef DEBUG
 941 int btrfs_reada_wait(void *handle)
 942 {
 943         struct reada_control *rc = handle;
 944         struct btrfs_fs_info *fs_info = rc->fs_info;
 945
 946         while (atomic_read(&rc->elems)) {
 947                 if (!atomic_read(&fs_info->reada_works_cnt))
 948                         reada_start_machine(fs_info);
 949                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 950                                    5 * HZ);
 951                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 952         }
 953
 954         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 955
 956         kref_put(&rc->refcnt, reada_control_release);
 957
 958         return 0;
 959 }
 960 #else
 961 int btrfs_reada_wait(void *handle)
 962 {
 963         struct reada_control *rc = handle;
 964         struct btrfs_fs_info *fs_info = rc->fs_info;
 965
 966         while (atomic_read(&rc->elems)) {
 967                 if (!atomic_read(&fs_info->reada_works_cnt))
 968                         reada_start_machine(fs_info);
 969                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 970                                    (HZ + 9) / 10);
 971         }
 972
 973         kref_put(&rc->refcnt, reada_control_release);
 974
 975         return 0;
 976 }
 977 #endif
 978
 979 void btrfs_reada_detach(void *handle)
 980 {
 981         struct reada_control *rc = handle;
 982
 983         kref_put(&rc->refcnt, reada_control_release);
 984 }