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         spin_unlock(&fs_info->reada_lock);
 460         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 461
 462         if (!have_zone)
 463                 goto error;
 464
 465         btrfs_put_bbio(bbio);
 466         return re;
 467
 468 error:
 469         for (nzones = 0; nzones < re->nzones; ++nzones) {
 470                 struct reada_zone *zone;
 471
 472                 zone = re->zones[nzones];
 473                 kref_get(&zone->refcnt);
 474                 spin_lock(&zone->lock);
 475                 --zone->elems;
 476                 if (zone->elems == 0) {
 477                         /*
 478                          * no fs_info->reada_lock needed, as this can't be
 479                          * the last ref
 480                          */
 481                         kref_put(&zone->refcnt, reada_zone_release);
 482                 }
 483                 spin_unlock(&zone->lock);
 484
 485                 spin_lock(&fs_info->reada_lock);
 486                 kref_put(&zone->refcnt, reada_zone_release);
 487                 spin_unlock(&fs_info->reada_lock);
 488         }
 489         btrfs_put_bbio(bbio);
 490         kfree(re);
 491         return re_exist;
 492 }
 493
 494 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 495                              struct reada_extent *re)
 496 {
 497         int i;
 498         unsigned long index = re->logical >> PAGE_SHIFT;
 499
 500         spin_lock(&fs_info->reada_lock);
 501         if (--re->refcnt) {
 502                 spin_unlock(&fs_info->reada_lock);
 503                 return;
 504         }
 505
 506         radix_tree_delete(&fs_info->reada_tree, index);
 507         for (i = 0; i < re->nzones; ++i) {
 508                 struct reada_zone *zone = re->zones[i];
 509
 510                 radix_tree_delete(&zone->device->reada_extents, index);
 511         }
 512
 513         spin_unlock(&fs_info->reada_lock);
 514
 515         for (i = 0; i < re->nzones; ++i) {
 516                 struct reada_zone *zone = re->zones[i];
 517
 518                 kref_get(&zone->refcnt);
 519                 spin_lock(&zone->lock);
 520                 --zone->elems;
 521                 if (zone->elems == 0) {
 522                         /* no fs_info->reada_lock needed, as this can't be
 523                          * the last ref */
 524                         kref_put(&zone->refcnt, reada_zone_release);
 525                 }
 526                 spin_unlock(&zone->lock);
 527
 528                 spin_lock(&fs_info->reada_lock);
 529                 kref_put(&zone->refcnt, reada_zone_release);
 530                 spin_unlock(&fs_info->reada_lock);
 531         }
 532
 533         kfree(re);
 534 }
 535
 536 static void reada_zone_release(struct kref *kref)
 537 {
 538         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 539
 540         radix_tree_delete(&zone->device->reada_zones,
 541                           zone->end >> PAGE_SHIFT);
 542
 543         kfree(zone);
 544 }
 545
 546 static void reada_control_release(struct kref *kref)
 547 {
 548         struct reada_control *rc = container_of(kref, struct reada_control,
 549                                                 refcnt);
 550
 551         kfree(rc);
 552 }
 553
 554 static int reada_add_block(struct reada_control *rc, u64 logical,
 555                            struct btrfs_key *top, u64 generation)
 556 {
 557         struct btrfs_fs_info *fs_info = rc->fs_info;
 558         struct reada_extent *re;
 559         struct reada_extctl *rec;
 560
 561         /* takes one ref */
 562         re = reada_find_extent(fs_info, logical, top);
 563         if (!re)
 564                 return -1;
 565
 566         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 567         if (!rec) {
 568                 reada_extent_put(fs_info, re);
 569                 return -ENOMEM;
 570         }
 571
 572         rec->rc = rc;
 573         rec->generation = generation;
 574         atomic_inc(&rc->elems);
 575
 576         spin_lock(&re->lock);
 577         list_add_tail(&rec->list, &re->extctl);
 578         spin_unlock(&re->lock);
 579
 580         /* leave the ref on the extent */
 581
 582         return 0;
 583 }
 584
 585 /*
 586  * called with fs_info->reada_lock held
 587  */
 588 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 589 {
 590         int i;
 591         unsigned long index = zone->end >> PAGE_SHIFT;
 592
 593         for (i = 0; i < zone->ndevs; ++i) {
 594                 struct reada_zone *peer;
 595                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 596                 if (peer && peer->device != zone->device)
 597                         peer->locked = lock;
 598         }
 599 }
 600
 601 /*
 602  * called with fs_info->reada_lock held
 603  */
 604 static int reada_pick_zone(struct btrfs_device *dev)
 605 {
 606         struct reada_zone *top_zone = NULL;
 607         struct reada_zone *top_locked_zone = NULL;
 608         u64 top_elems = 0;
 609         u64 top_locked_elems = 0;
 610         unsigned long index = 0;
 611         int ret;
 612
 613         if (dev->reada_curr_zone) {
 614                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 615                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 616                 dev->reada_curr_zone = NULL;
 617         }
 618         /* pick the zone with the most elements */
 619         while (1) {
 620                 struct reada_zone *zone;
 621
 622                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 623                                              (void **)&zone, index, 1);
 624                 if (ret == 0)
 625                         break;
 626                 index = (zone->end >> PAGE_SHIFT) + 1;
 627                 if (zone->locked) {
 628                         if (zone->elems > top_locked_elems) {
 629                                 top_locked_elems = zone->elems;
 630                                 top_locked_zone = zone;
 631                         }
 632                 } else {
 633                         if (zone->elems > top_elems) {
 634                                 top_elems = zone->elems;
 635                                 top_zone = zone;
 636                         }
 637                 }
 638         }
 639         if (top_zone)
 640                 dev->reada_curr_zone = top_zone;
 641         else if (top_locked_zone)
 642                 dev->reada_curr_zone = top_locked_zone;
 643         else
 644                 return 0;
 645
 646         dev->reada_next = dev->reada_curr_zone->start;
 647         kref_get(&dev->reada_curr_zone->refcnt);
 648         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 649
 650         return 1;
 651 }
 652
 653 static int reada_start_machine_dev(struct btrfs_device *dev)
 654 {
 655         struct btrfs_fs_info *fs_info = dev->fs_info;
 656         struct reada_extent *re = NULL;
 657         int mirror_num = 0;
 658         struct extent_buffer *eb = NULL;
 659         u64 logical;
 660         int ret;
 661         int i;
 662
 663         spin_lock(&fs_info->reada_lock);
 664         if (dev->reada_curr_zone == NULL) {
 665                 ret = reada_pick_zone(dev);
 666                 if (!ret) {
 667                         spin_unlock(&fs_info->reada_lock);
 668                         return 0;
 669                 }
 670         }
 671         /*
 672          * FIXME currently we issue the reads one extent at a time. If we have
 673          * a contiguous block of extents, we could also coagulate them or use
 674          * plugging to speed things up
 675          */
 676         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 677                                      dev->reada_next >> PAGE_SHIFT, 1);
 678         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 679                 ret = reada_pick_zone(dev);
 680                 if (!ret) {
 681                         spin_unlock(&fs_info->reada_lock);
 682                         return 0;
 683                 }
 684                 re = NULL;
 685                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 686                                         dev->reada_next >> PAGE_SHIFT, 1);
 687         }
 688         if (ret == 0) {
 689                 spin_unlock(&fs_info->reada_lock);
 690                 return 0;
 691         }
 692         dev->reada_next = re->logical + fs_info->nodesize;
 693         re->refcnt++;
 694
 695         spin_unlock(&fs_info->reada_lock);
 696
 697         spin_lock(&re->lock);
 698         if (re->scheduled || list_empty(&re->extctl)) {
 699                 spin_unlock(&re->lock);
 700                 reada_extent_put(fs_info, re);
 701                 return 0;
 702         }
 703         re->scheduled = 1;
 704         spin_unlock(&re->lock);
 705
 706         /*
 707          * find mirror num
 708          */
 709         for (i = 0; i < re->nzones; ++i) {
 710                 if (re->zones[i]->device == dev) {
 711                         mirror_num = i + 1;
 712                         break;
 713                 }
 714         }
 715         logical = re->logical;
 716
 717         atomic_inc(&dev->reada_in_flight);
 718         ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
 719         if (ret)
 720                 __readahead_hook(fs_info, re, NULL, ret);
 721         else if (eb)
 722                 __readahead_hook(fs_info, re, eb, ret);
 723
 724         if (eb)
 725                 free_extent_buffer(eb);
 726
 727         atomic_dec(&dev->reada_in_flight);
 728         reada_extent_put(fs_info, re);
 729
 730         return 1;
 731
 732 }
 733
 734 static void reada_start_machine_worker(struct btrfs_work *work)
 735 {
 736         struct reada_machine_work *rmw;
 737         struct btrfs_fs_info *fs_info;
 738         int old_ioprio;
 739
 740         rmw = container_of(work, struct reada_machine_work, work);
 741         fs_info = rmw->fs_info;
 742
 743         kfree(rmw);
 744
 745         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 746                                        task_nice_ioprio(current));
 747         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 748         __reada_start_machine(fs_info);
 749         set_task_ioprio(current, old_ioprio);
 750
 751         atomic_dec(&fs_info->reada_works_cnt);
 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         do {
 763                 enqueued = 0;
 764                 mutex_lock(&fs_devices->device_list_mutex);
 765                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 766                         if (atomic_read(&device->reada_in_flight) <
 767                             MAX_IN_FLIGHT)
 768                                 enqueued += reada_start_machine_dev(device);
 769                 }
 770                 mutex_unlock(&fs_devices->device_list_mutex);
 771                 total += enqueued;
 772         } while (enqueued && total < 10000);
 773
 774         if (enqueued == 0)
 775                 return;
 776
 777         /*
 778          * If everything is already in the cache, this is effectively single
 779          * threaded. To a) not hold the caller for too long and b) to utilize
 780          * more cores, we broke the loop above after 10000 iterations and now
 781          * enqueue to workers to finish it. This will distribute the load to
 782          * the cores.
 783          */
 784         for (i = 0; i < 2; ++i) {
 785                 reada_start_machine(fs_info);
 786                 if (atomic_read(&fs_info->reada_works_cnt) >
 787                     BTRFS_MAX_MIRRORS * 2)
 788                         break;
 789         }
 790 }
 791
 792 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 793 {
 794         struct reada_machine_work *rmw;
 795
 796         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 797         if (!rmw) {
 798                 /* FIXME we cannot handle this properly right now */
 799                 BUG();
 800         }
 801         btrfs_init_work(&rmw->work, btrfs_readahead_helper,
 802                         reada_start_machine_worker, NULL, NULL);
 803         rmw->fs_info = fs_info;
 804
 805         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 806         atomic_inc(&fs_info->reada_works_cnt);
 807 }
 808
 809 #ifdef DEBUG
 810 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 811 {
 812         struct btrfs_device *device;
 813         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 814         unsigned long index;
 815         int ret;
 816         int i;
 817         int j;
 818         int cnt;
 819
 820         spin_lock(&fs_info->reada_lock);
 821         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 822                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 823                         atomic_read(&device->reada_in_flight));
 824                 index = 0;
 825                 while (1) {
 826                         struct reada_zone *zone;
 827                         ret = radix_tree_gang_lookup(&device->reada_zones,
 828                                                      (void **)&zone, index, 1);
 829                         if (ret == 0)
 830                                 break;
 831                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 832                                     zone->start, zone->end, zone->elems,
 833                                     zone->locked);
 834                         for (j = 0; j < zone->ndevs; ++j) {
 835                                 pr_cont(" %lld",
 836                                         zone->devs[j]->devid);
 837                         }
 838                         if (device->reada_curr_zone == zone)
 839                                 pr_cont(" curr off %llu",
 840                                         device->reada_next - zone->start);
 841                         pr_cont("\n");
 842                         index = (zone->end >> PAGE_SHIFT) + 1;
 843                 }
 844                 cnt = 0;
 845                 index = 0;
 846                 while (all) {
 847                         struct reada_extent *re = NULL;
 848
 849                         ret = radix_tree_gang_lookup(&device->reada_extents,
 850                                                      (void **)&re, index, 1);
 851                         if (ret == 0)
 852                                 break;
 853                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 854                                 re->logical, fs_info->nodesize,
 855                                 list_empty(&re->extctl), re->scheduled);
 856
 857                         for (i = 0; i < re->nzones; ++i) {
 858                                 pr_cont(" zone %llu-%llu devs",
 859                                         re->zones[i]->start,
 860                                         re->zones[i]->end);
 861                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 862                                         pr_cont(" %lld",
 863                                                 re->zones[i]->devs[j]->devid);
 864                                 }
 865                         }
 866                         pr_cont("\n");
 867                         index = (re->logical >> PAGE_SHIFT) + 1;
 868                         if (++cnt > 15)
 869                                 break;
 870                 }
 871         }
 872
 873         index = 0;
 874         cnt = 0;
 875         while (all) {
 876                 struct reada_extent *re = NULL;
 877
 878                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 879                                              index, 1);
 880                 if (ret == 0)
 881                         break;
 882                 if (!re->scheduled) {
 883                         index = (re->logical >> PAGE_SHIFT) + 1;
 884                         continue;
 885                 }
 886                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 887                         re->logical, fs_info->nodesize,
 888                         list_empty(&re->extctl), re->scheduled);
 889                 for (i = 0; i < re->nzones; ++i) {
 890                         pr_cont(" zone %llu-%llu devs",
 891                                 re->zones[i]->start,
 892                                 re->zones[i]->end);
 893                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 894                                 pr_cont(" %lld",
 895                                        re->zones[i]->devs[j]->devid);
 896                         }
 897                 }
 898                 pr_cont("\n");
 899                 index = (re->logical >> PAGE_SHIFT) + 1;
 900         }
 901         spin_unlock(&fs_info->reada_lock);
 902 }
 903 #endif
 904
 905 /*
 906  * interface
 907  */
 908 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 909                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 910 {
 911         struct reada_control *rc;
 912         u64 start;
 913         u64 generation;
 914         int ret;
 915         struct extent_buffer *node;
 916         static struct btrfs_key max_key = {
 917                 .objectid = (u64)-1,
 918                 .type = (u8)-1,
 919                 .offset = (u64)-1
 920         };
 921
 922         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 923         if (!rc)
 924                 return ERR_PTR(-ENOMEM);
 925
 926         rc->fs_info = root->fs_info;
 927         rc->key_start = *key_start;
 928         rc->key_end = *key_end;
 929         atomic_set(&rc->elems, 0);
 930         init_waitqueue_head(&rc->wait);
 931         kref_init(&rc->refcnt);
 932         kref_get(&rc->refcnt); /* one ref for having elements */
 933
 934         node = btrfs_root_node(root);
 935         start = node->start;
 936         generation = btrfs_header_generation(node);
 937         free_extent_buffer(node);
 938
 939         ret = reada_add_block(rc, start, &max_key, generation);
 940         if (ret) {
 941                 kfree(rc);
 942                 return ERR_PTR(ret);
 943         }
 944
 945         reada_start_machine(root->fs_info);
 946
 947         return rc;
 948 }
 949
 950 #ifdef DEBUG
 951 int btrfs_reada_wait(void *handle)
 952 {
 953         struct reada_control *rc = handle;
 954         struct btrfs_fs_info *fs_info = rc->fs_info;
 955
 956         while (atomic_read(&rc->elems)) {
 957                 if (!atomic_read(&fs_info->reada_works_cnt))
 958                         reada_start_machine(fs_info);
 959                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 960                                    5 * HZ);
 961                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 962         }
 963
 964         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 965
 966         kref_put(&rc->refcnt, reada_control_release);
 967
 968         return 0;
 969 }
 970 #else
 971 int btrfs_reada_wait(void *handle)
 972 {
 973         struct reada_control *rc = handle;
 974         struct btrfs_fs_info *fs_info = rc->fs_info;
 975
 976         while (atomic_read(&rc->elems)) {
 977                 if (!atomic_read(&fs_info->reada_works_cnt))
 978                         reada_start_machine(fs_info);
 979                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 980                                    (HZ + 9) / 10);
 981         }
 982
 983         kref_put(&rc->refcnt, reada_control_release);
 984
 985         return 0;
 986 }
 987 #endif
 988
 989 void btrfs_reada_detach(void *handle)
 990 {
 991         struct reada_control *rc = handle;
 992
 993         kref_put(&rc->refcnt, reada_control_release);
 994 }