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 extent_buffer *eb, int err)
 213 {
 214         struct btrfs_fs_info *fs_info = eb->fs_info;
 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_device *dev, u64 logical,
 239                                           struct btrfs_bio *bbio)
 240 {
 241         struct btrfs_fs_info *fs_info = dev->fs_info;
 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         ret = radix_tree_preload(GFP_KERNEL);
 274         if (ret) {
 275                 kfree(zone);
 276                 return NULL;
 277         }
 278
 279         zone->start = start;
 280         zone->end = end;
 281         INIT_LIST_HEAD(&zone->list);
 282         spin_lock_init(&zone->lock);
 283         zone->locked = 0;
 284         kref_init(&zone->refcnt);
 285         zone->elems = 0;
 286         zone->device = dev; /* our device always sits at index 0 */
 287         for (i = 0; i < bbio->num_stripes; ++i) {
 288                 /* bounds have already been checked */
 289                 zone->devs[i] = bbio->stripes[i].dev;
 290         }
 291         zone->ndevs = bbio->num_stripes;
 292
 293         spin_lock(&fs_info->reada_lock);
 294         ret = radix_tree_insert(&dev->reada_zones,
 295                                 (unsigned long)(zone->end >> PAGE_SHIFT),
 296                                 zone);
 297
 298         if (ret == -EEXIST) {
 299                 kfree(zone);
 300                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 301                                              logical >> PAGE_SHIFT, 1);
 302                 if (ret == 1 && logical >= zone->start && logical <= zone->end)
 303                         kref_get(&zone->refcnt);
 304                 else
 305                         zone = NULL;
 306         }
 307         spin_unlock(&fs_info->reada_lock);
 308         radix_tree_preload_end();
 309
 310         return zone;
 311 }
 312
 313 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
 314                                               u64 logical,
 315                                               struct btrfs_key *top)
 316 {
 317         int ret;
 318         struct reada_extent *re = NULL;
 319         struct reada_extent *re_exist = NULL;
 320         struct btrfs_bio *bbio = NULL;
 321         struct btrfs_device *dev;
 322         struct btrfs_device *prev_dev;
 323         u64 length;
 324         int real_stripes;
 325         int nzones = 0;
 326         unsigned long index = logical >> PAGE_SHIFT;
 327         int dev_replace_is_ongoing;
 328         int have_zone = 0;
 329
 330         spin_lock(&fs_info->reada_lock);
 331         re = radix_tree_lookup(&fs_info->reada_tree, index);
 332         if (re)
 333                 re->refcnt++;
 334         spin_unlock(&fs_info->reada_lock);
 335
 336         if (re)
 337                 return re;
 338
 339         re = kzalloc(sizeof(*re), GFP_KERNEL);
 340         if (!re)
 341                 return NULL;
 342
 343         re->logical = logical;
 344         re->top = *top;
 345         INIT_LIST_HEAD(&re->extctl);
 346         spin_lock_init(&re->lock);
 347         re->refcnt = 1;
 348
 349         /*
 350          * map block
 351          */
 352         length = fs_info->nodesize;
 353         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
 354                         &length, &bbio, 0);
 355         if (ret || !bbio || length < fs_info->nodesize)
 356                 goto error;
 357
 358         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 359                 btrfs_err(fs_info,
 360                            "readahead: more than %d copies not supported",
 361                            BTRFS_MAX_MIRRORS);
 362                 goto error;
 363         }
 364
 365         real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
 366         for (nzones = 0; nzones < real_stripes; ++nzones) {
 367                 struct reada_zone *zone;
 368
 369                 dev = bbio->stripes[nzones].dev;
 370
 371                 /* cannot read ahead on missing device. */
 372                  if (!dev->bdev)
 373                         continue;
 374
 375                 zone = reada_find_zone(dev, logical, bbio);
 376                 if (!zone)
 377                         continue;
 378
 379                 re->zones[re->nzones++] = zone;
 380                 spin_lock(&zone->lock);
 381                 if (!zone->elems)
 382                         kref_get(&zone->refcnt);
 383                 ++zone->elems;
 384                 spin_unlock(&zone->lock);
 385                 spin_lock(&fs_info->reada_lock);
 386                 kref_put(&zone->refcnt, reada_zone_release);
 387                 spin_unlock(&fs_info->reada_lock);
 388         }
 389         if (re->nzones == 0) {
 390                 /* not a single zone found, error and out */
 391                 goto error;
 392         }
 393
 394         ret = radix_tree_preload(GFP_KERNEL);
 395         if (ret)
 396                 goto error;
 397
 398         /* insert extent in reada_tree + all per-device trees, all or nothing */
 399         btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
 400         spin_lock(&fs_info->reada_lock);
 401         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 402         if (ret == -EEXIST) {
 403                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 404                 re_exist->refcnt++;
 405                 spin_unlock(&fs_info->reada_lock);
 406                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 407                 radix_tree_preload_end();
 408                 goto error;
 409         }
 410         if (ret) {
 411                 spin_unlock(&fs_info->reada_lock);
 412                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 413                 radix_tree_preload_end();
 414                 goto error;
 415         }
 416         radix_tree_preload_end();
 417         prev_dev = NULL;
 418         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 419                         &fs_info->dev_replace);
 420         for (nzones = 0; nzones < re->nzones; ++nzones) {
 421                 dev = re->zones[nzones]->device;
 422
 423                 if (dev == prev_dev) {
 424                         /*
 425                          * in case of DUP, just add the first zone. As both
 426                          * are on the same device, there's nothing to gain
 427                          * from adding both.
 428                          * Also, it wouldn't work, as the tree is per device
 429                          * and adding would fail with EEXIST
 430                          */
 431                         continue;
 432                 }
 433                 if (!dev->bdev)
 434                         continue;
 435
 436                 if (dev_replace_is_ongoing &&
 437                     dev == fs_info->dev_replace.tgtdev) {
 438                         /*
 439                          * as this device is selected for reading only as
 440                          * a last resort, skip it for read ahead.
 441                          */
 442                         continue;
 443                 }
 444                 prev_dev = dev;
 445                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 446                 if (ret) {
 447                         while (--nzones >= 0) {
 448                                 dev = re->zones[nzones]->device;
 449                                 BUG_ON(dev == NULL);
 450                                 /* ignore whether the entry was inserted */
 451                                 radix_tree_delete(&dev->reada_extents, index);
 452                         }
 453                         radix_tree_delete(&fs_info->reada_tree, index);
 454                         spin_unlock(&fs_info->reada_lock);
 455                         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 456                         goto error;
 457                 }
 458                 have_zone = 1;
 459         }
 460         spin_unlock(&fs_info->reada_lock);
 461         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 462
 463         if (!have_zone)
 464                 goto error;
 465
 466         btrfs_put_bbio(bbio);
 467         return re;
 468
 469 error:
 470         for (nzones = 0; nzones < re->nzones; ++nzones) {
 471                 struct reada_zone *zone;
 472
 473                 zone = re->zones[nzones];
 474                 kref_get(&zone->refcnt);
 475                 spin_lock(&zone->lock);
 476                 --zone->elems;
 477                 if (zone->elems == 0) {
 478                         /*
 479                          * no fs_info->reada_lock needed, as this can't be
 480                          * the last ref
 481                          */
 482                         kref_put(&zone->refcnt, reada_zone_release);
 483                 }
 484                 spin_unlock(&zone->lock);
 485
 486                 spin_lock(&fs_info->reada_lock);
 487                 kref_put(&zone->refcnt, reada_zone_release);
 488                 spin_unlock(&fs_info->reada_lock);
 489         }
 490         btrfs_put_bbio(bbio);
 491         kfree(re);
 492         return re_exist;
 493 }
 494
 495 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 496                              struct reada_extent *re)
 497 {
 498         int i;
 499         unsigned long index = re->logical >> PAGE_SHIFT;
 500
 501         spin_lock(&fs_info->reada_lock);
 502         if (--re->refcnt) {
 503                 spin_unlock(&fs_info->reada_lock);
 504                 return;
 505         }
 506
 507         radix_tree_delete(&fs_info->reada_tree, index);
 508         for (i = 0; i < re->nzones; ++i) {
 509                 struct reada_zone *zone = re->zones[i];
 510
 511                 radix_tree_delete(&zone->device->reada_extents, index);
 512         }
 513
 514         spin_unlock(&fs_info->reada_lock);
 515
 516         for (i = 0; i < re->nzones; ++i) {
 517                 struct reada_zone *zone = re->zones[i];
 518
 519                 kref_get(&zone->refcnt);
 520                 spin_lock(&zone->lock);
 521                 --zone->elems;
 522                 if (zone->elems == 0) {
 523                         /* no fs_info->reada_lock needed, as this can't be
 524                          * the last ref */
 525                         kref_put(&zone->refcnt, reada_zone_release);
 526                 }
 527                 spin_unlock(&zone->lock);
 528
 529                 spin_lock(&fs_info->reada_lock);
 530                 kref_put(&zone->refcnt, reada_zone_release);
 531                 spin_unlock(&fs_info->reada_lock);
 532         }
 533
 534         kfree(re);
 535 }
 536
 537 static void reada_zone_release(struct kref *kref)
 538 {
 539         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 540
 541         radix_tree_delete(&zone->device->reada_zones,
 542                           zone->end >> PAGE_SHIFT);
 543
 544         kfree(zone);
 545 }
 546
 547 static void reada_control_release(struct kref *kref)
 548 {
 549         struct reada_control *rc = container_of(kref, struct reada_control,
 550                                                 refcnt);
 551
 552         kfree(rc);
 553 }
 554
 555 static int reada_add_block(struct reada_control *rc, u64 logical,
 556                            struct btrfs_key *top, u64 generation)
 557 {
 558         struct btrfs_fs_info *fs_info = rc->fs_info;
 559         struct reada_extent *re;
 560         struct reada_extctl *rec;
 561
 562         /* takes one ref */
 563         re = reada_find_extent(fs_info, logical, top);
 564         if (!re)
 565                 return -1;
 566
 567         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 568         if (!rec) {
 569                 reada_extent_put(fs_info, re);
 570                 return -ENOMEM;
 571         }
 572
 573         rec->rc = rc;
 574         rec->generation = generation;
 575         atomic_inc(&rc->elems);
 576
 577         spin_lock(&re->lock);
 578         list_add_tail(&rec->list, &re->extctl);
 579         spin_unlock(&re->lock);
 580
 581         /* leave the ref on the extent */
 582
 583         return 0;
 584 }
 585
 586 /*
 587  * called with fs_info->reada_lock held
 588  */
 589 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 590 {
 591         int i;
 592         unsigned long index = zone->end >> PAGE_SHIFT;
 593
 594         for (i = 0; i < zone->ndevs; ++i) {
 595                 struct reada_zone *peer;
 596                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 597                 if (peer && peer->device != zone->device)
 598                         peer->locked = lock;
 599         }
 600 }
 601
 602 /*
 603  * called with fs_info->reada_lock held
 604  */
 605 static int reada_pick_zone(struct btrfs_device *dev)
 606 {
 607         struct reada_zone *top_zone = NULL;
 608         struct reada_zone *top_locked_zone = NULL;
 609         u64 top_elems = 0;
 610         u64 top_locked_elems = 0;
 611         unsigned long index = 0;
 612         int ret;
 613
 614         if (dev->reada_curr_zone) {
 615                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 616                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 617                 dev->reada_curr_zone = NULL;
 618         }
 619         /* pick the zone with the most elements */
 620         while (1) {
 621                 struct reada_zone *zone;
 622
 623                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 624                                              (void **)&zone, index, 1);
 625                 if (ret == 0)
 626                         break;
 627                 index = (zone->end >> PAGE_SHIFT) + 1;
 628                 if (zone->locked) {
 629                         if (zone->elems > top_locked_elems) {
 630                                 top_locked_elems = zone->elems;
 631                                 top_locked_zone = zone;
 632                         }
 633                 } else {
 634                         if (zone->elems > top_elems) {
 635                                 top_elems = zone->elems;
 636                                 top_zone = zone;
 637                         }
 638                 }
 639         }
 640         if (top_zone)
 641                 dev->reada_curr_zone = top_zone;
 642         else if (top_locked_zone)
 643                 dev->reada_curr_zone = top_locked_zone;
 644         else
 645                 return 0;
 646
 647         dev->reada_next = dev->reada_curr_zone->start;
 648         kref_get(&dev->reada_curr_zone->refcnt);
 649         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 650
 651         return 1;
 652 }
 653
 654 static int reada_start_machine_dev(struct btrfs_device *dev)
 655 {
 656         struct btrfs_fs_info *fs_info = dev->fs_info;
 657         struct reada_extent *re = NULL;
 658         int mirror_num = 0;
 659         struct extent_buffer *eb = NULL;
 660         u64 logical;
 661         int ret;
 662         int i;
 663
 664         spin_lock(&fs_info->reada_lock);
 665         if (dev->reada_curr_zone == NULL) {
 666                 ret = reada_pick_zone(dev);
 667                 if (!ret) {
 668                         spin_unlock(&fs_info->reada_lock);
 669                         return 0;
 670                 }
 671         }
 672         /*
 673          * FIXME currently we issue the reads one extent at a time. If we have
 674          * a contiguous block of extents, we could also coagulate them or use
 675          * plugging to speed things up
 676          */
 677         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 678                                      dev->reada_next >> PAGE_SHIFT, 1);
 679         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 680                 ret = reada_pick_zone(dev);
 681                 if (!ret) {
 682                         spin_unlock(&fs_info->reada_lock);
 683                         return 0;
 684                 }
 685                 re = NULL;
 686                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 687                                         dev->reada_next >> PAGE_SHIFT, 1);
 688         }
 689         if (ret == 0) {
 690                 spin_unlock(&fs_info->reada_lock);
 691                 return 0;
 692         }
 693         dev->reada_next = re->logical + fs_info->nodesize;
 694         re->refcnt++;
 695
 696         spin_unlock(&fs_info->reada_lock);
 697
 698         spin_lock(&re->lock);
 699         if (re->scheduled || list_empty(&re->extctl)) {
 700                 spin_unlock(&re->lock);
 701                 reada_extent_put(fs_info, re);
 702                 return 0;
 703         }
 704         re->scheduled = 1;
 705         spin_unlock(&re->lock);
 706
 707         /*
 708          * find mirror num
 709          */
 710         for (i = 0; i < re->nzones; ++i) {
 711                 if (re->zones[i]->device == dev) {
 712                         mirror_num = i + 1;
 713                         break;
 714                 }
 715         }
 716         logical = re->logical;
 717
 718         atomic_inc(&dev->reada_in_flight);
 719         ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
 720         if (ret)
 721                 __readahead_hook(fs_info, re, NULL, ret);
 722         else if (eb)
 723                 __readahead_hook(fs_info, re, eb, ret);
 724
 725         if (eb)
 726                 free_extent_buffer(eb);
 727
 728         atomic_dec(&dev->reada_in_flight);
 729         reada_extent_put(fs_info, re);
 730
 731         return 1;
 732
 733 }
 734
 735 static void reada_start_machine_worker(struct btrfs_work *work)
 736 {
 737         struct reada_machine_work *rmw;
 738         struct btrfs_fs_info *fs_info;
 739         int old_ioprio;
 740
 741         rmw = container_of(work, struct reada_machine_work, work);
 742         fs_info = rmw->fs_info;
 743
 744         kfree(rmw);
 745
 746         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 747                                        task_nice_ioprio(current));
 748         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 749         __reada_start_machine(fs_info);
 750         set_task_ioprio(current, old_ioprio);
 751
 752         atomic_dec(&fs_info->reada_works_cnt);
 753 }
 754
 755 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 756 {
 757         struct btrfs_device *device;
 758         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 759         u64 enqueued;
 760         u64 total = 0;
 761         int i;
 762
 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
 775         if (enqueued == 0)
 776                 return;
 777
 778         /*
 779          * If everything is already in the cache, this is effectively single
 780          * threaded. To a) not hold the caller for too long and b) to utilize
 781          * more cores, we broke the loop above after 10000 iterations and now
 782          * enqueue to workers to finish it. This will distribute the load to
 783          * the cores.
 784          */
 785         for (i = 0; i < 2; ++i) {
 786                 reada_start_machine(fs_info);
 787                 if (atomic_read(&fs_info->reada_works_cnt) >
 788                     BTRFS_MAX_MIRRORS * 2)
 789                         break;
 790         }
 791 }
 792
 793 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 794 {
 795         struct reada_machine_work *rmw;
 796
 797         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 798         if (!rmw) {
 799                 /* FIXME we cannot handle this properly right now */
 800                 BUG();
 801         }
 802         btrfs_init_work(&rmw->work, btrfs_readahead_helper,
 803                         reada_start_machine_worker, NULL, NULL);
 804         rmw->fs_info = fs_info;
 805
 806         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 807         atomic_inc(&fs_info->reada_works_cnt);
 808 }
 809
 810 #ifdef DEBUG
 811 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 812 {
 813         struct btrfs_device *device;
 814         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 815         unsigned long index;
 816         int ret;
 817         int i;
 818         int j;
 819         int cnt;
 820
 821         spin_lock(&fs_info->reada_lock);
 822         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 823                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 824                         atomic_read(&device->reada_in_flight));
 825                 index = 0;
 826                 while (1) {
 827                         struct reada_zone *zone;
 828                         ret = radix_tree_gang_lookup(&device->reada_zones,
 829                                                      (void **)&zone, index, 1);
 830                         if (ret == 0)
 831                                 break;
 832                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 833                                     zone->start, zone->end, zone->elems,
 834                                     zone->locked);
 835                         for (j = 0; j < zone->ndevs; ++j) {
 836                                 pr_cont(" %lld",
 837                                         zone->devs[j]->devid);
 838                         }
 839                         if (device->reada_curr_zone == zone)
 840                                 pr_cont(" curr off %llu",
 841                                         device->reada_next - zone->start);
 842                         pr_cont("\n");
 843                         index = (zone->end >> PAGE_SHIFT) + 1;
 844                 }
 845                 cnt = 0;
 846                 index = 0;
 847                 while (all) {
 848                         struct reada_extent *re = NULL;
 849
 850                         ret = radix_tree_gang_lookup(&device->reada_extents,
 851                                                      (void **)&re, index, 1);
 852                         if (ret == 0)
 853                                 break;
 854                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 855                                 re->logical, fs_info->nodesize,
 856                                 list_empty(&re->extctl), re->scheduled);
 857
 858                         for (i = 0; i < re->nzones; ++i) {
 859                                 pr_cont(" zone %llu-%llu devs",
 860                                         re->zones[i]->start,
 861                                         re->zones[i]->end);
 862                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 863                                         pr_cont(" %lld",
 864                                                 re->zones[i]->devs[j]->devid);
 865                                 }
 866                         }
 867                         pr_cont("\n");
 868                         index = (re->logical >> PAGE_SHIFT) + 1;
 869                         if (++cnt > 15)
 870                                 break;
 871                 }
 872         }
 873
 874         index = 0;
 875         cnt = 0;
 876         while (all) {
 877                 struct reada_extent *re = NULL;
 878
 879                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 880                                              index, 1);
 881                 if (ret == 0)
 882                         break;
 883                 if (!re->scheduled) {
 884                         index = (re->logical >> PAGE_SHIFT) + 1;
 885                         continue;
 886                 }
 887                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 888                         re->logical, fs_info->nodesize,
 889                         list_empty(&re->extctl), re->scheduled);
 890                 for (i = 0; i < re->nzones; ++i) {
 891                         pr_cont(" zone %llu-%llu devs",
 892                                 re->zones[i]->start,
 893                                 re->zones[i]->end);
 894                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 895                                 pr_cont(" %lld",
 896                                        re->zones[i]->devs[j]->devid);
 897                         }
 898                 }
 899                 pr_cont("\n");
 900                 index = (re->logical >> PAGE_SHIFT) + 1;
 901         }
 902         spin_unlock(&fs_info->reada_lock);
 903 }
 904 #endif
 905
 906 /*
 907  * interface
 908  */
 909 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 910                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 911 {
 912         struct reada_control *rc;
 913         u64 start;
 914         u64 generation;
 915         int ret;
 916         struct extent_buffer *node;
 917         static struct btrfs_key max_key = {
 918                 .objectid = (u64)-1,
 919                 .type = (u8)-1,
 920                 .offset = (u64)-1
 921         };
 922
 923         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 924         if (!rc)
 925                 return ERR_PTR(-ENOMEM);
 926
 927         rc->fs_info = root->fs_info;
 928         rc->key_start = *key_start;
 929         rc->key_end = *key_end;
 930         atomic_set(&rc->elems, 0);
 931         init_waitqueue_head(&rc->wait);
 932         kref_init(&rc->refcnt);
 933         kref_get(&rc->refcnt); /* one ref for having elements */
 934
 935         node = btrfs_root_node(root);
 936         start = node->start;
 937         generation = btrfs_header_generation(node);
 938         free_extent_buffer(node);
 939
 940         ret = reada_add_block(rc, start, &max_key, generation);
 941         if (ret) {
 942                 kfree(rc);
 943                 return ERR_PTR(ret);
 944         }
 945
 946         reada_start_machine(root->fs_info);
 947
 948         return rc;
 949 }
 950
 951 #ifdef DEBUG
 952 int btrfs_reada_wait(void *handle)
 953 {
 954         struct reada_control *rc = handle;
 955         struct btrfs_fs_info *fs_info = rc->fs_info;
 956
 957         while (atomic_read(&rc->elems)) {
 958                 if (!atomic_read(&fs_info->reada_works_cnt))
 959                         reada_start_machine(fs_info);
 960                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 961                                    5 * HZ);
 962                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 963         }
 964
 965         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 966
 967         kref_put(&rc->refcnt, reada_control_release);
 968
 969         return 0;
 970 }
 971 #else
 972 int btrfs_reada_wait(void *handle)
 973 {
 974         struct reada_control *rc = handle;
 975         struct btrfs_fs_info *fs_info = rc->fs_info;
 976
 977         while (atomic_read(&rc->elems)) {
 978                 if (!atomic_read(&fs_info->reada_works_cnt))
 979                         reada_start_machine(fs_info);
 980                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 981                                    (HZ + 9) / 10);
 982         }
 983
 984         kref_put(&rc->refcnt, reada_control_release);
 985
 986         return 0;
 987 }
 988 #endif
 989
 990 void btrfs_reada_detach(void *handle)
 991 {
 992         struct reada_control *rc = handle;
 993
 994         kref_put(&rc->refcnt, reada_control_release);
 995 }