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         struct list_head        extctl;
  70         int                     refcnt;
  71         spinlock_t              lock;
  72         struct reada_zone       *zones[BTRFS_MAX_MIRRORS];
  73         int                     nzones;
  74         int                     scheduled;
  75 };
  76
  77 struct reada_zone {
  78         u64                     start;
  79         u64                     end;
  80         u64                     elems;
  81         struct list_head        list;
  82         spinlock_t              lock;
  83         int                     locked;
  84         struct btrfs_device     *device;
  85         struct btrfs_device     *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
  86                                                            * self */
  87         int                     ndevs;
  88         struct kref             refcnt;
  89 };
  90
  91 struct reada_machine_work {
  92         struct btrfs_work       work;
  93         struct btrfs_fs_info    *fs_info;
  94 };
  95
  96 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
  97 static void reada_control_release(struct kref *kref);
  98 static void reada_zone_release(struct kref *kref);
  99 static void reada_start_machine(struct btrfs_fs_info *fs_info);
 100 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
 101
 102 static int reada_add_block(struct reada_control *rc, u64 logical,
 103                            struct btrfs_key *top, u64 generation);
 104
 105 /* recurses */
 106 /* in case of err, eb might be NULL */
 107 static void __readahead_hook(struct btrfs_fs_info *fs_info,
 108                              struct reada_extent *re, struct extent_buffer *eb,
 109                              int err)
 110 {
 111         int nritems;
 112         int i;
 113         u64 bytenr;
 114         u64 generation;
 115         struct list_head list;
 116
 117         spin_lock(&re->lock);
 118         /*
 119          * just take the full list from the extent. afterwards we
 120          * don't need the lock anymore
 121          */
 122         list_replace_init(&re->extctl, &list);
 123         re->scheduled = 0;
 124         spin_unlock(&re->lock);
 125
 126         /*
 127          * this is the error case, the extent buffer has not been
 128          * read correctly. We won't access anything from it and
 129          * just cleanup our data structures. Effectively this will
 130          * cut the branch below this node from read ahead.
 131          */
 132         if (err)
 133                 goto cleanup;
 134
 135         /*
 136          * FIXME: currently we just set nritems to 0 if this is a leaf,
 137          * effectively ignoring the content. In a next step we could
 138          * trigger more readahead depending from the content, e.g.
 139          * fetch the checksums for the extents in the leaf.
 140          */
 141         if (!btrfs_header_level(eb))
 142                 goto cleanup;
 143
 144         nritems = btrfs_header_nritems(eb);
 145         generation = btrfs_header_generation(eb);
 146         for (i = 0; i < nritems; i++) {
 147                 struct reada_extctl *rec;
 148                 u64 n_gen;
 149                 struct btrfs_key key;
 150                 struct btrfs_key next_key;
 151
 152                 btrfs_node_key_to_cpu(eb, &key, i);
 153                 if (i + 1 < nritems)
 154                         btrfs_node_key_to_cpu(eb, &next_key, i + 1);
 155                 else
 156                         next_key = re->top;
 157                 bytenr = btrfs_node_blockptr(eb, i);
 158                 n_gen = btrfs_node_ptr_generation(eb, i);
 159
 160                 list_for_each_entry(rec, &list, list) {
 161                         struct reada_control *rc = rec->rc;
 162
 163                         /*
 164                          * if the generation doesn't match, just ignore this
 165                          * extctl. This will probably cut off a branch from
 166                          * prefetch. Alternatively one could start a new (sub-)
 167                          * prefetch for this branch, starting again from root.
 168                          * FIXME: move the generation check out of this loop
 169                          */
 170 #ifdef DEBUG
 171                         if (rec->generation != generation) {
 172                                 btrfs_debug(fs_info,
 173                                             "generation mismatch for (%llu,%d,%llu) %llu != %llu",
 174                                             key.objectid, key.type, key.offset,
 175                                             rec->generation, generation);
 176                         }
 177 #endif
 178                         if (rec->generation == generation &&
 179                             btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
 180                             btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
 181                                 reada_add_block(rc, bytenr, &next_key, n_gen);
 182                 }
 183         }
 184
 185 cleanup:
 186         /*
 187          * free extctl records
 188          */
 189         while (!list_empty(&list)) {
 190                 struct reada_control *rc;
 191                 struct reada_extctl *rec;
 192
 193                 rec = list_first_entry(&list, struct reada_extctl, list);
 194                 list_del(&rec->list);
 195                 rc = rec->rc;
 196                 kfree(rec);
 197
 198                 kref_get(&rc->refcnt);
 199                 if (atomic_dec_and_test(&rc->elems)) {
 200                         kref_put(&rc->refcnt, reada_control_release);
 201                         wake_up(&rc->wait);
 202                 }
 203                 kref_put(&rc->refcnt, reada_control_release);
 204
 205                 reada_extent_put(fs_info, re);  /* one ref for each entry */
 206         }
 207
 208         return;
 209 }
 210
 211 int btree_readahead_hook(struct extent_buffer *eb, int err)
 212 {
 213         struct btrfs_fs_info *fs_info = eb->fs_info;
 214         int ret = 0;
 215         struct reada_extent *re;
 216
 217         /* find extent */
 218         spin_lock(&fs_info->reada_lock);
 219         re = radix_tree_lookup(&fs_info->reada_tree,
 220                                eb->start >> PAGE_SHIFT);
 221         if (re)
 222                 re->refcnt++;
 223         spin_unlock(&fs_info->reada_lock);
 224         if (!re) {
 225                 ret = -1;
 226                 goto start_machine;
 227         }
 228
 229         __readahead_hook(fs_info, re, eb, err);
 230         reada_extent_put(fs_info, re);  /* our ref */
 231
 232 start_machine:
 233         reada_start_machine(fs_info);
 234         return ret;
 235 }
 236
 237 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
 238                                           struct btrfs_bio *bbio)
 239 {
 240         struct btrfs_fs_info *fs_info = dev->fs_info;
 241         int ret;
 242         struct reada_zone *zone;
 243         struct btrfs_block_group_cache *cache = NULL;
 244         u64 start;
 245         u64 end;
 246         int i;
 247
 248         zone = NULL;
 249         spin_lock(&fs_info->reada_lock);
 250         ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 251                                      logical >> PAGE_SHIFT, 1);
 252         if (ret == 1 && logical >= zone->start && logical <= zone->end) {
 253                 kref_get(&zone->refcnt);
 254                 spin_unlock(&fs_info->reada_lock);
 255                 return zone;
 256         }
 257
 258         spin_unlock(&fs_info->reada_lock);
 259
 260         cache = btrfs_lookup_block_group(fs_info, logical);
 261         if (!cache)
 262                 return NULL;
 263
 264         start = cache->key.objectid;
 265         end = start + cache->key.offset - 1;
 266         btrfs_put_block_group(cache);
 267
 268         zone = kzalloc(sizeof(*zone), GFP_KERNEL);
 269         if (!zone)
 270                 return NULL;
 271
 272         ret = radix_tree_preload(GFP_KERNEL);
 273         if (ret) {
 274                 kfree(zone);
 275                 return NULL;
 276         }
 277
 278         zone->start = start;
 279         zone->end = end;
 280         INIT_LIST_HEAD(&zone->list);
 281         spin_lock_init(&zone->lock);
 282         zone->locked = 0;
 283         kref_init(&zone->refcnt);
 284         zone->elems = 0;
 285         zone->device = dev; /* our device always sits at index 0 */
 286         for (i = 0; i < bbio->num_stripes; ++i) {
 287                 /* bounds have already been checked */
 288                 zone->devs[i] = bbio->stripes[i].dev;
 289         }
 290         zone->ndevs = bbio->num_stripes;
 291
 292         spin_lock(&fs_info->reada_lock);
 293         ret = radix_tree_insert(&dev->reada_zones,
 294                                 (unsigned long)(zone->end >> PAGE_SHIFT),
 295                                 zone);
 296
 297         if (ret == -EEXIST) {
 298                 kfree(zone);
 299                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 300                                              logical >> PAGE_SHIFT, 1);
 301                 if (ret == 1 && logical >= zone->start && logical <= zone->end)
 302                         kref_get(&zone->refcnt);
 303                 else
 304                         zone = NULL;
 305         }
 306         spin_unlock(&fs_info->reada_lock);
 307         radix_tree_preload_end();
 308
 309         return zone;
 310 }
 311
 312 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
 313                                               u64 logical,
 314                                               struct btrfs_key *top)
 315 {
 316         int ret;
 317         struct reada_extent *re = NULL;
 318         struct reada_extent *re_exist = NULL;
 319         struct btrfs_bio *bbio = NULL;
 320         struct btrfs_device *dev;
 321         struct btrfs_device *prev_dev;
 322         u64 length;
 323         int real_stripes;
 324         int nzones = 0;
 325         unsigned long index = logical >> PAGE_SHIFT;
 326         int dev_replace_is_ongoing;
 327         int have_zone = 0;
 328
 329         spin_lock(&fs_info->reada_lock);
 330         re = radix_tree_lookup(&fs_info->reada_tree, index);
 331         if (re)
 332                 re->refcnt++;
 333         spin_unlock(&fs_info->reada_lock);
 334
 335         if (re)
 336                 return re;
 337
 338         re = kzalloc(sizeof(*re), GFP_KERNEL);
 339         if (!re)
 340                 return NULL;
 341
 342         re->logical = logical;
 343         re->top = *top;
 344         INIT_LIST_HEAD(&re->extctl);
 345         spin_lock_init(&re->lock);
 346         re->refcnt = 1;
 347
 348         /*
 349          * map block
 350          */
 351         length = fs_info->nodesize;
 352         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
 353                         &length, &bbio, 0);
 354         if (ret || !bbio || length < fs_info->nodesize)
 355                 goto error;
 356
 357         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 358                 btrfs_err(fs_info,
 359                            "readahead: more than %d copies not supported",
 360                            BTRFS_MAX_MIRRORS);
 361                 goto error;
 362         }
 363
 364         real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
 365         for (nzones = 0; nzones < real_stripes; ++nzones) {
 366                 struct reada_zone *zone;
 367
 368                 dev = bbio->stripes[nzones].dev;
 369
 370                 /* cannot read ahead on missing device. */
 371                  if (!dev->bdev)
 372                         continue;
 373
 374                 zone = reada_find_zone(dev, logical, bbio);
 375                 if (!zone)
 376                         continue;
 377
 378                 re->zones[re->nzones++] = zone;
 379                 spin_lock(&zone->lock);
 380                 if (!zone->elems)
 381                         kref_get(&zone->refcnt);
 382                 ++zone->elems;
 383                 spin_unlock(&zone->lock);
 384                 spin_lock(&fs_info->reada_lock);
 385                 kref_put(&zone->refcnt, reada_zone_release);
 386                 spin_unlock(&fs_info->reada_lock);
 387         }
 388         if (re->nzones == 0) {
 389                 /* not a single zone found, error and out */
 390                 goto error;
 391         }
 392
 393         ret = radix_tree_preload(GFP_KERNEL);
 394         if (ret)
 395                 goto error;
 396
 397         /* insert extent in reada_tree + all per-device trees, all or nothing */
 398         btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
 399         spin_lock(&fs_info->reada_lock);
 400         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 401         if (ret == -EEXIST) {
 402                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 403                 re_exist->refcnt++;
 404                 spin_unlock(&fs_info->reada_lock);
 405                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 406                 radix_tree_preload_end();
 407                 goto error;
 408         }
 409         if (ret) {
 410                 spin_unlock(&fs_info->reada_lock);
 411                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 412                 radix_tree_preload_end();
 413                 goto error;
 414         }
 415         radix_tree_preload_end();
 416         prev_dev = NULL;
 417         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 418                         &fs_info->dev_replace);
 419         for (nzones = 0; nzones < re->nzones; ++nzones) {
 420                 dev = re->zones[nzones]->device;
 421
 422                 if (dev == prev_dev) {
 423                         /*
 424                          * in case of DUP, just add the first zone. As both
 425                          * are on the same device, there's nothing to gain
 426                          * from adding both.
 427                          * Also, it wouldn't work, as the tree is per device
 428                          * and adding would fail with EEXIST
 429                          */
 430                         continue;
 431                 }
 432                 if (!dev->bdev)
 433                         continue;
 434
 435                 if (dev_replace_is_ongoing &&
 436                     dev == fs_info->dev_replace.tgtdev) {
 437                         /*
 438                          * as this device is selected for reading only as
 439                          * a last resort, skip it for read ahead.
 440                          */
 441                         continue;
 442                 }
 443                 prev_dev = dev;
 444                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 445                 if (ret) {
 446                         while (--nzones >= 0) {
 447                                 dev = re->zones[nzones]->device;
 448                                 BUG_ON(dev == NULL);
 449                                 /* ignore whether the entry was inserted */
 450                                 radix_tree_delete(&dev->reada_extents, index);
 451                         }
 452                         radix_tree_delete(&fs_info->reada_tree, index);
 453                         spin_unlock(&fs_info->reada_lock);
 454                         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 455                         goto error;
 456                 }
 457                 have_zone = 1;
 458         }
 459         if (!have_zone)
 460                 radix_tree_delete(&fs_info->reada_tree, index);
 461         spin_unlock(&fs_info->reada_lock);
 462         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 463
 464         if (!have_zone)
 465                 goto error;
 466
 467         btrfs_put_bbio(bbio);
 468         return re;
 469
 470 error:
 471         for (nzones = 0; nzones < re->nzones; ++nzones) {
 472                 struct reada_zone *zone;
 473
 474                 zone = re->zones[nzones];
 475                 kref_get(&zone->refcnt);
 476                 spin_lock(&zone->lock);
 477                 --zone->elems;
 478                 if (zone->elems == 0) {
 479                         /*
 480                          * no fs_info->reada_lock needed, as this can't be
 481                          * the last ref
 482                          */
 483                         kref_put(&zone->refcnt, reada_zone_release);
 484                 }
 485                 spin_unlock(&zone->lock);
 486
 487                 spin_lock(&fs_info->reada_lock);
 488                 kref_put(&zone->refcnt, reada_zone_release);
 489                 spin_unlock(&fs_info->reada_lock);
 490         }
 491         btrfs_put_bbio(bbio);
 492         kfree(re);
 493         return re_exist;
 494 }
 495
 496 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 497                              struct reada_extent *re)
 498 {
 499         int i;
 500         unsigned long index = re->logical >> PAGE_SHIFT;
 501
 502         spin_lock(&fs_info->reada_lock);
 503         if (--re->refcnt) {
 504                 spin_unlock(&fs_info->reada_lock);
 505                 return;
 506         }
 507
 508         radix_tree_delete(&fs_info->reada_tree, index);
 509         for (i = 0; i < re->nzones; ++i) {
 510                 struct reada_zone *zone = re->zones[i];
 511
 512                 radix_tree_delete(&zone->device->reada_extents, index);
 513         }
 514
 515         spin_unlock(&fs_info->reada_lock);
 516
 517         for (i = 0; i < re->nzones; ++i) {
 518                 struct reada_zone *zone = re->zones[i];
 519
 520                 kref_get(&zone->refcnt);
 521                 spin_lock(&zone->lock);
 522                 --zone->elems;
 523                 if (zone->elems == 0) {
 524                         /* no fs_info->reada_lock needed, as this can't be
 525                          * the last ref */
 526                         kref_put(&zone->refcnt, reada_zone_release);
 527                 }
 528                 spin_unlock(&zone->lock);
 529
 530                 spin_lock(&fs_info->reada_lock);
 531                 kref_put(&zone->refcnt, reada_zone_release);
 532                 spin_unlock(&fs_info->reada_lock);
 533         }
 534
 535         kfree(re);
 536 }
 537
 538 static void reada_zone_release(struct kref *kref)
 539 {
 540         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 541
 542         radix_tree_delete(&zone->device->reada_zones,
 543                           zone->end >> PAGE_SHIFT);
 544
 545         kfree(zone);
 546 }
 547
 548 static void reada_control_release(struct kref *kref)
 549 {
 550         struct reada_control *rc = container_of(kref, struct reada_control,
 551                                                 refcnt);
 552
 553         kfree(rc);
 554 }
 555
 556 static int reada_add_block(struct reada_control *rc, u64 logical,
 557                            struct btrfs_key *top, u64 generation)
 558 {
 559         struct btrfs_fs_info *fs_info = rc->fs_info;
 560         struct reada_extent *re;
 561         struct reada_extctl *rec;
 562
 563         /* takes one ref */
 564         re = reada_find_extent(fs_info, logical, top);
 565         if (!re)
 566                 return -1;
 567
 568         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 569         if (!rec) {
 570                 reada_extent_put(fs_info, re);
 571                 return -ENOMEM;
 572         }
 573
 574         rec->rc = rc;
 575         rec->generation = generation;
 576         atomic_inc(&rc->elems);
 577
 578         spin_lock(&re->lock);
 579         list_add_tail(&rec->list, &re->extctl);
 580         spin_unlock(&re->lock);
 581
 582         /* leave the ref on the extent */
 583
 584         return 0;
 585 }
 586
 587 /*
 588  * called with fs_info->reada_lock held
 589  */
 590 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 591 {
 592         int i;
 593         unsigned long index = zone->end >> PAGE_SHIFT;
 594
 595         for (i = 0; i < zone->ndevs; ++i) {
 596                 struct reada_zone *peer;
 597                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 598                 if (peer && peer->device != zone->device)
 599                         peer->locked = lock;
 600         }
 601 }
 602
 603 /*
 604  * called with fs_info->reada_lock held
 605  */
 606 static int reada_pick_zone(struct btrfs_device *dev)
 607 {
 608         struct reada_zone *top_zone = NULL;
 609         struct reada_zone *top_locked_zone = NULL;
 610         u64 top_elems = 0;
 611         u64 top_locked_elems = 0;
 612         unsigned long index = 0;
 613         int ret;
 614
 615         if (dev->reada_curr_zone) {
 616                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 617                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 618                 dev->reada_curr_zone = NULL;
 619         }
 620         /* pick the zone with the most elements */
 621         while (1) {
 622                 struct reada_zone *zone;
 623
 624                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 625                                              (void **)&zone, index, 1);
 626                 if (ret == 0)
 627                         break;
 628                 index = (zone->end >> PAGE_SHIFT) + 1;
 629                 if (zone->locked) {
 630                         if (zone->elems > top_locked_elems) {
 631                                 top_locked_elems = zone->elems;
 632                                 top_locked_zone = zone;
 633                         }
 634                 } else {
 635                         if (zone->elems > top_elems) {
 636                                 top_elems = zone->elems;
 637                                 top_zone = zone;
 638                         }
 639                 }
 640         }
 641         if (top_zone)
 642                 dev->reada_curr_zone = top_zone;
 643         else if (top_locked_zone)
 644                 dev->reada_curr_zone = top_locked_zone;
 645         else
 646                 return 0;
 647
 648         dev->reada_next = dev->reada_curr_zone->start;
 649         kref_get(&dev->reada_curr_zone->refcnt);
 650         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 651
 652         return 1;
 653 }
 654
 655 static int reada_start_machine_dev(struct btrfs_device *dev)
 656 {
 657         struct btrfs_fs_info *fs_info = dev->fs_info;
 658         struct reada_extent *re = NULL;
 659         int mirror_num = 0;
 660         struct extent_buffer *eb = NULL;
 661         u64 logical;
 662         int ret;
 663         int i;
 664
 665         spin_lock(&fs_info->reada_lock);
 666         if (dev->reada_curr_zone == NULL) {
 667                 ret = reada_pick_zone(dev);
 668                 if (!ret) {
 669                         spin_unlock(&fs_info->reada_lock);
 670                         return 0;
 671                 }
 672         }
 673         /*
 674          * FIXME currently we issue the reads one extent at a time. If we have
 675          * a contiguous block of extents, we could also coagulate them or use
 676          * plugging to speed things up
 677          */
 678         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 679                                      dev->reada_next >> PAGE_SHIFT, 1);
 680         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 681                 ret = reada_pick_zone(dev);
 682                 if (!ret) {
 683                         spin_unlock(&fs_info->reada_lock);
 684                         return 0;
 685                 }
 686                 re = NULL;
 687                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 688                                         dev->reada_next >> PAGE_SHIFT, 1);
 689         }
 690         if (ret == 0) {
 691                 spin_unlock(&fs_info->reada_lock);
 692                 return 0;
 693         }
 694         dev->reada_next = re->logical + fs_info->nodesize;
 695         re->refcnt++;
 696
 697         spin_unlock(&fs_info->reada_lock);
 698
 699         spin_lock(&re->lock);
 700         if (re->scheduled || list_empty(&re->extctl)) {
 701                 spin_unlock(&re->lock);
 702                 reada_extent_put(fs_info, re);
 703                 return 0;
 704         }
 705         re->scheduled = 1;
 706         spin_unlock(&re->lock);
 707
 708         /*
 709          * find mirror num
 710          */
 711         for (i = 0; i < re->nzones; ++i) {
 712                 if (re->zones[i]->device == dev) {
 713                         mirror_num = i + 1;
 714                         break;
 715                 }
 716         }
 717         logical = re->logical;
 718
 719         atomic_inc(&dev->reada_in_flight);
 720         ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
 721         if (ret)
 722                 __readahead_hook(fs_info, re, NULL, ret);
 723         else if (eb)
 724                 __readahead_hook(fs_info, re, eb, ret);
 725
 726         if (eb)
 727                 free_extent_buffer(eb);
 728
 729         atomic_dec(&dev->reada_in_flight);
 730         reada_extent_put(fs_info, re);
 731
 732         return 1;
 733
 734 }
 735
 736 static void reada_start_machine_worker(struct btrfs_work *work)
 737 {
 738         struct reada_machine_work *rmw;
 739         int old_ioprio;
 740
 741         rmw = container_of(work, struct reada_machine_work, work);
 742
 743         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 744                                        task_nice_ioprio(current));
 745         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 746         __reada_start_machine(rmw->fs_info);
 747         set_task_ioprio(current, old_ioprio);
 748
 749         atomic_dec(&rmw->fs_info->reada_works_cnt);
 750
 751         kfree(rmw);
 752 }
 753
 754 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 755 {
 756         struct btrfs_device *device;
 757         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 758         u64 enqueued;
 759         u64 total = 0;
 760         int i;
 761
 762 again:
 763         do {
 764                 enqueued = 0;
 765                 mutex_lock(&fs_devices->device_list_mutex);
 766                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 767                         if (atomic_read(&device->reada_in_flight) <
 768                             MAX_IN_FLIGHT)
 769                                 enqueued += reada_start_machine_dev(device);
 770                 }
 771                 mutex_unlock(&fs_devices->device_list_mutex);
 772                 total += enqueued;
 773         } while (enqueued && total < 10000);
 774         if (fs_devices->seed) {
 775                 fs_devices = fs_devices->seed;
 776                 goto again;
 777         }
 778
 779         if (enqueued == 0)
 780                 return;
 781
 782         /*
 783          * If everything is already in the cache, this is effectively single
 784          * threaded. To a) not hold the caller for too long and b) to utilize
 785          * more cores, we broke the loop above after 10000 iterations and now
 786          * enqueue to workers to finish it. This will distribute the load to
 787          * the cores.
 788          */
 789         for (i = 0; i < 2; ++i) {
 790                 reada_start_machine(fs_info);
 791                 if (atomic_read(&fs_info->reada_works_cnt) >
 792                     BTRFS_MAX_MIRRORS * 2)
 793                         break;
 794         }
 795 }
 796
 797 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 798 {
 799         struct reada_machine_work *rmw;
 800
 801         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 802         if (!rmw) {
 803                 /* FIXME we cannot handle this properly right now */
 804                 BUG();
 805         }
 806         btrfs_init_work(&rmw->work, btrfs_readahead_helper,
 807                         reada_start_machine_worker, NULL, NULL);
 808         rmw->fs_info = fs_info;
 809
 810         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 811         atomic_inc(&fs_info->reada_works_cnt);
 812 }
 813
 814 #ifdef DEBUG
 815 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 816 {
 817         struct btrfs_device *device;
 818         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 819         unsigned long index;
 820         int ret;
 821         int i;
 822         int j;
 823         int cnt;
 824
 825         spin_lock(&fs_info->reada_lock);
 826         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 827                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 828                         atomic_read(&device->reada_in_flight));
 829                 index = 0;
 830                 while (1) {
 831                         struct reada_zone *zone;
 832                         ret = radix_tree_gang_lookup(&device->reada_zones,
 833                                                      (void **)&zone, index, 1);
 834                         if (ret == 0)
 835                                 break;
 836                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 837                                     zone->start, zone->end, zone->elems,
 838                                     zone->locked);
 839                         for (j = 0; j < zone->ndevs; ++j) {
 840                                 pr_cont(" %lld",
 841                                         zone->devs[j]->devid);
 842                         }
 843                         if (device->reada_curr_zone == zone)
 844                                 pr_cont(" curr off %llu",
 845                                         device->reada_next - zone->start);
 846                         pr_cont("\n");
 847                         index = (zone->end >> PAGE_SHIFT) + 1;
 848                 }
 849                 cnt = 0;
 850                 index = 0;
 851                 while (all) {
 852                         struct reada_extent *re = NULL;
 853
 854                         ret = radix_tree_gang_lookup(&device->reada_extents,
 855                                                      (void **)&re, index, 1);
 856                         if (ret == 0)
 857                                 break;
 858                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 859                                 re->logical, fs_info->nodesize,
 860                                 list_empty(&re->extctl), re->scheduled);
 861
 862                         for (i = 0; i < re->nzones; ++i) {
 863                                 pr_cont(" zone %llu-%llu devs",
 864                                         re->zones[i]->start,
 865                                         re->zones[i]->end);
 866                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 867                                         pr_cont(" %lld",
 868                                                 re->zones[i]->devs[j]->devid);
 869                                 }
 870                         }
 871                         pr_cont("\n");
 872                         index = (re->logical >> PAGE_SHIFT) + 1;
 873                         if (++cnt > 15)
 874                                 break;
 875                 }
 876         }
 877
 878         index = 0;
 879         cnt = 0;
 880         while (all) {
 881                 struct reada_extent *re = NULL;
 882
 883                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 884                                              index, 1);
 885                 if (ret == 0)
 886                         break;
 887                 if (!re->scheduled) {
 888                         index = (re->logical >> PAGE_SHIFT) + 1;
 889                         continue;
 890                 }
 891                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 892                         re->logical, fs_info->nodesize,
 893                         list_empty(&re->extctl), re->scheduled);
 894                 for (i = 0; i < re->nzones; ++i) {
 895                         pr_cont(" zone %llu-%llu devs",
 896                                 re->zones[i]->start,
 897                                 re->zones[i]->end);
 898                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 899                                 pr_cont(" %lld",
 900                                        re->zones[i]->devs[j]->devid);
 901                         }
 902                 }
 903                 pr_cont("\n");
 904                 index = (re->logical >> PAGE_SHIFT) + 1;
 905         }
 906         spin_unlock(&fs_info->reada_lock);
 907 }
 908 #endif
 909
 910 /*
 911  * interface
 912  */
 913 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 914                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 915 {
 916         struct reada_control *rc;
 917         u64 start;
 918         u64 generation;
 919         int ret;
 920         struct extent_buffer *node;
 921         static struct btrfs_key max_key = {
 922                 .objectid = (u64)-1,
 923                 .type = (u8)-1,
 924                 .offset = (u64)-1
 925         };
 926
 927         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 928         if (!rc)
 929                 return ERR_PTR(-ENOMEM);
 930
 931         rc->fs_info = root->fs_info;
 932         rc->key_start = *key_start;
 933         rc->key_end = *key_end;
 934         atomic_set(&rc->elems, 0);
 935         init_waitqueue_head(&rc->wait);
 936         kref_init(&rc->refcnt);
 937         kref_get(&rc->refcnt); /* one ref for having elements */
 938
 939         node = btrfs_root_node(root);
 940         start = node->start;
 941         generation = btrfs_header_generation(node);
 942         free_extent_buffer(node);
 943
 944         ret = reada_add_block(rc, start, &max_key, generation);
 945         if (ret) {
 946                 kfree(rc);
 947                 return ERR_PTR(ret);
 948         }
 949
 950         reada_start_machine(root->fs_info);
 951
 952         return rc;
 953 }
 954
 955 #ifdef DEBUG
 956 int btrfs_reada_wait(void *handle)
 957 {
 958         struct reada_control *rc = handle;
 959         struct btrfs_fs_info *fs_info = rc->fs_info;
 960
 961         while (atomic_read(&rc->elems)) {
 962                 if (!atomic_read(&fs_info->reada_works_cnt))
 963                         reada_start_machine(fs_info);
 964                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 965                                    5 * HZ);
 966                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 967         }
 968
 969         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 970
 971         kref_put(&rc->refcnt, reada_control_release);
 972
 973         return 0;
 974 }
 975 #else
 976 int btrfs_reada_wait(void *handle)
 977 {
 978         struct reada_control *rc = handle;
 979         struct btrfs_fs_info *fs_info = rc->fs_info;
 980
 981         while (atomic_read(&rc->elems)) {
 982                 if (!atomic_read(&fs_info->reada_works_cnt))
 983                         reada_start_machine(fs_info);
 984                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 985                                    (HZ + 9) / 10);
 986         }
 987
 988         kref_put(&rc->refcnt, reada_control_release);
 989
 990         return 0;
 991 }
 992 #endif
 993
 994 void btrfs_reada_detach(void *handle)
 995 {
 996         struct reada_control *rc = handle;
 997
 998         kref_put(&rc->refcnt, reada_control_release);
 999 }