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         u32                     blocksize;
  70         int                     err;
  71         struct list_head        extctl;
  72         int                     refcnt;
  73         spinlock_t              lock;
  74         struct reada_zone       *zones[BTRFS_MAX_MIRRORS];
  75         int                     nzones;
  76         struct btrfs_device     *scheduled_for;
  77 };
  78
  79 struct reada_zone {
  80         u64                     start;
  81         u64                     end;
  82         u64                     elems;
  83         struct list_head        list;
  84         spinlock_t              lock;
  85         int                     locked;
  86         struct btrfs_device     *device;
  87         struct btrfs_device     *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
  88                                                            * self */
  89         int                     ndevs;
  90         struct kref             refcnt;
  91 };
  92
  93 struct reada_machine_work {
  94         struct btrfs_work       work;
  95         struct btrfs_fs_info    *fs_info;
  96 };
  97
  98 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
  99 static void reada_control_release(struct kref *kref);
 100 static void reada_zone_release(struct kref *kref);
 101 static void reada_start_machine(struct btrfs_fs_info *fs_info);
 102 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
 103
 104 static int reada_add_block(struct reada_control *rc, u64 logical,
 105                            struct btrfs_key *top, int level, u64 generation);
 106
 107 /* recurses */
 108 /* in case of err, eb might be NULL */
 109 static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
 110                             u64 start, int err)
 111 {
 112         int level = 0;
 113         int nritems;
 114         int i;
 115         u64 bytenr;
 116         u64 generation;
 117         struct reada_extent *re;
 118         struct btrfs_fs_info *fs_info = root->fs_info;
 119         struct list_head list;
 120         unsigned long index = start >> PAGE_CACHE_SHIFT;
 121         struct btrfs_device *for_dev;
 122
 123         if (eb)
 124                 level = btrfs_header_level(eb);
 125
 126         /* find extent */
 127         spin_lock(&fs_info->reada_lock);
 128         re = radix_tree_lookup(&fs_info->reada_tree, index);
 129         if (re)
 130                 re->refcnt++;
 131         spin_unlock(&fs_info->reada_lock);
 132
 133         if (!re)
 134                 return -1;
 135
 136         spin_lock(&re->lock);
 137         /*
 138          * just take the full list from the extent. afterwards we
 139          * don't need the lock anymore
 140          */
 141         list_replace_init(&re->extctl, &list);
 142         for_dev = re->scheduled_for;
 143         re->scheduled_for = NULL;
 144         spin_unlock(&re->lock);
 145
 146         if (err == 0) {
 147                 nritems = level ? btrfs_header_nritems(eb) : 0;
 148                 generation = btrfs_header_generation(eb);
 149                 /*
 150                  * FIXME: currently we just set nritems to 0 if this is a leaf,
 151                  * effectively ignoring the content. In a next step we could
 152                  * trigger more readahead depending from the content, e.g.
 153                  * fetch the checksums for the extents in the leaf.
 154                  */
 155         } else {
 156                 /*
 157                  * this is the error case, the extent buffer has not been
 158                  * read correctly. We won't access anything from it and
 159                  * just cleanup our data structures. Effectively this will
 160                  * cut the branch below this node from read ahead.
 161                  */
 162                 nritems = 0;
 163                 generation = 0;
 164         }
 165
 166         for (i = 0; i < nritems; i++) {
 167                 struct reada_extctl *rec;
 168                 u64 n_gen;
 169                 struct btrfs_key key;
 170                 struct btrfs_key next_key;
 171
 172                 btrfs_node_key_to_cpu(eb, &key, i);
 173                 if (i + 1 < nritems)
 174                         btrfs_node_key_to_cpu(eb, &next_key, i + 1);
 175                 else
 176                         next_key = re->top;
 177                 bytenr = btrfs_node_blockptr(eb, i);
 178                 n_gen = btrfs_node_ptr_generation(eb, i);
 179
 180                 list_for_each_entry(rec, &list, list) {
 181                         struct reada_control *rc = rec->rc;
 182
 183                         /*
 184                          * if the generation doesn't match, just ignore this
 185                          * extctl. This will probably cut off a branch from
 186                          * prefetch. Alternatively one could start a new (sub-)
 187                          * prefetch for this branch, starting again from root.
 188                          * FIXME: move the generation check out of this loop
 189                          */
 190 #ifdef DEBUG
 191                         if (rec->generation != generation) {
 192                                 printk(KERN_DEBUG "generation mismatch for "
 193                                                 "(%llu,%d,%llu) %llu != %llu\n",
 194                                        key.objectid, key.type, key.offset,
 195                                        rec->generation, generation);
 196                         }
 197 #endif
 198                         if (rec->generation == generation &&
 199                             btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
 200                             btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
 201                                 reada_add_block(rc, bytenr, &next_key,
 202                                                 level - 1, n_gen);
 203                 }
 204         }
 205         /*
 206          * free extctl records
 207          */
 208         while (!list_empty(&list)) {
 209                 struct reada_control *rc;
 210                 struct reada_extctl *rec;
 211
 212                 rec = list_first_entry(&list, struct reada_extctl, list);
 213                 list_del(&rec->list);
 214                 rc = rec->rc;
 215                 kfree(rec);
 216
 217                 kref_get(&rc->refcnt);
 218                 if (atomic_dec_and_test(&rc->elems)) {
 219                         kref_put(&rc->refcnt, reada_control_release);
 220                         wake_up(&rc->wait);
 221                 }
 222                 kref_put(&rc->refcnt, reada_control_release);
 223
 224                 reada_extent_put(fs_info, re);  /* one ref for each entry */
 225         }
 226         reada_extent_put(fs_info, re);  /* our ref */
 227         if (for_dev)
 228                 atomic_dec(&for_dev->reada_in_flight);
 229
 230         return 0;
 231 }
 232
 233 /*
 234  * start is passed separately in case eb in NULL, which may be the case with
 235  * failed I/O
 236  */
 237 int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
 238                          u64 start, int err)
 239 {
 240         int ret;
 241
 242         ret = __readahead_hook(root, eb, start, err);
 243
 244         reada_start_machine(root->fs_info);
 245
 246         return ret;
 247 }
 248
 249 static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
 250                                           struct btrfs_device *dev, u64 logical,
 251                                           struct btrfs_bio *bbio)
 252 {
 253         int ret;
 254         struct reada_zone *zone;
 255         struct btrfs_block_group_cache *cache = NULL;
 256         u64 start;
 257         u64 end;
 258         int i;
 259
 260         zone = NULL;
 261         spin_lock(&fs_info->reada_lock);
 262         ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 263                                      logical >> PAGE_CACHE_SHIFT, 1);
 264         if (ret == 1)
 265                 kref_get(&zone->refcnt);
 266         spin_unlock(&fs_info->reada_lock);
 267
 268         if (ret == 1) {
 269                 if (logical >= zone->start && logical < zone->end)
 270                         return zone;
 271                 spin_lock(&fs_info->reada_lock);
 272                 kref_put(&zone->refcnt, reada_zone_release);
 273                 spin_unlock(&fs_info->reada_lock);
 274         }
 275
 276         cache = btrfs_lookup_block_group(fs_info, logical);
 277         if (!cache)
 278                 return NULL;
 279
 280         start = cache->key.objectid;
 281         end = start + cache->key.offset - 1;
 282         btrfs_put_block_group(cache);
 283
 284         zone = kzalloc(sizeof(*zone), GFP_NOFS);
 285         if (!zone)
 286                 return NULL;
 287
 288         zone->start = start;
 289         zone->end = end;
 290         INIT_LIST_HEAD(&zone->list);
 291         spin_lock_init(&zone->lock);
 292         zone->locked = 0;
 293         kref_init(&zone->refcnt);
 294         zone->elems = 0;
 295         zone->device = dev; /* our device always sits at index 0 */
 296         for (i = 0; i < bbio->num_stripes; ++i) {
 297                 /* bounds have already been checked */
 298                 zone->devs[i] = bbio->stripes[i].dev;
 299         }
 300         zone->ndevs = bbio->num_stripes;
 301
 302         spin_lock(&fs_info->reada_lock);
 303         ret = radix_tree_insert(&dev->reada_zones,
 304                                 (unsigned long)(zone->end >> PAGE_CACHE_SHIFT),
 305                                 zone);
 306
 307         if (ret == -EEXIST) {
 308                 kfree(zone);
 309                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 310                                              logical >> PAGE_CACHE_SHIFT, 1);
 311                 if (ret == 1)
 312                         kref_get(&zone->refcnt);
 313         }
 314         spin_unlock(&fs_info->reada_lock);
 315
 316         return zone;
 317 }
 318
 319 static struct reada_extent *reada_find_extent(struct btrfs_root *root,
 320                                               u64 logical,
 321                                               struct btrfs_key *top, int level)
 322 {
 323         int ret;
 324         struct reada_extent *re = NULL;
 325         struct reada_extent *re_exist = NULL;
 326         struct btrfs_fs_info *fs_info = root->fs_info;
 327         struct btrfs_bio *bbio = NULL;
 328         struct btrfs_device *dev;
 329         struct btrfs_device *prev_dev;
 330         u32 blocksize;
 331         u64 length;
 332         int nzones = 0;
 333         int i;
 334         unsigned long index = logical >> PAGE_CACHE_SHIFT;
 335         int dev_replace_is_ongoing;
 336
 337         spin_lock(&fs_info->reada_lock);
 338         re = radix_tree_lookup(&fs_info->reada_tree, index);
 339         if (re)
 340                 re->refcnt++;
 341         spin_unlock(&fs_info->reada_lock);
 342
 343         if (re)
 344                 return re;
 345
 346         re = kzalloc(sizeof(*re), GFP_NOFS);
 347         if (!re)
 348                 return NULL;
 349
 350         blocksize = btrfs_level_size(root, level);
 351         re->logical = logical;
 352         re->blocksize = blocksize;
 353         re->top = *top;
 354         INIT_LIST_HEAD(&re->extctl);
 355         spin_lock_init(&re->lock);
 356         re->refcnt = 1;
 357
 358         /*
 359          * map block
 360          */
 361         length = blocksize;
 362         ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
 363                               &bbio, 0);
 364         if (ret || !bbio || length < blocksize)
 365                 goto error;
 366
 367         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 368                 printk(KERN_ERR "btrfs readahead: more than %d copies not "
 369                                 "supported", BTRFS_MAX_MIRRORS);
 370                 goto error;
 371         }
 372
 373         for (nzones = 0; nzones < bbio->num_stripes; ++nzones) {
 374                 struct reada_zone *zone;
 375
 376                 dev = bbio->stripes[nzones].dev;
 377                 zone = reada_find_zone(fs_info, dev, logical, bbio);
 378                 if (!zone)
 379                         break;
 380
 381                 re->zones[nzones] = zone;
 382                 spin_lock(&zone->lock);
 383                 if (!zone->elems)
 384                         kref_get(&zone->refcnt);
 385                 ++zone->elems;
 386                 spin_unlock(&zone->lock);
 387                 spin_lock(&fs_info->reada_lock);
 388                 kref_put(&zone->refcnt, reada_zone_release);
 389                 spin_unlock(&fs_info->reada_lock);
 390         }
 391         re->nzones = nzones;
 392         if (nzones == 0) {
 393                 /* not a single zone found, error and out */
 394                 goto error;
 395         }
 396
 397         /* insert extent in reada_tree + all per-device trees, all or nothing */
 398         btrfs_dev_replace_lock(&fs_info->dev_replace);
 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                 BUG_ON(!re_exist);
 404                 re_exist->refcnt++;
 405                 spin_unlock(&fs_info->reada_lock);
 406                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
 407                 goto error;
 408         }
 409         if (ret) {
 410                 spin_unlock(&fs_info->reada_lock);
 411                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
 412                 goto error;
 413         }
 414         prev_dev = NULL;
 415         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 416                         &fs_info->dev_replace);
 417         for (i = 0; i < nzones; ++i) {
 418                 dev = bbio->stripes[i].dev;
 419                 if (dev == prev_dev) {
 420                         /*
 421                          * in case of DUP, just add the first zone. As both
 422                          * are on the same device, there's nothing to gain
 423                          * from adding both.
 424                          * Also, it wouldn't work, as the tree is per device
 425                          * and adding would fail with EEXIST
 426                          */
 427                         continue;
 428                 }
 429                 if (!dev->bdev) {
 430                         /* cannot read ahead on missing device */
 431                         continue;
 432                 }
 433                 if (dev_replace_is_ongoing &&
 434                     dev == fs_info->dev_replace.tgtdev) {
 435                         /*
 436                          * as this device is selected for reading only as
 437                          * a last resort, skip it for read ahead.
 438                          */
 439                         continue;
 440                 }
 441                 prev_dev = dev;
 442                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 443                 if (ret) {
 444                         while (--i >= 0) {
 445                                 dev = bbio->stripes[i].dev;
 446                                 BUG_ON(dev == NULL);
 447                                 /* ignore whether the entry was inserted */
 448                                 radix_tree_delete(&dev->reada_extents, index);
 449                         }
 450                         BUG_ON(fs_info == NULL);
 451                         radix_tree_delete(&fs_info->reada_tree, index);
 452                         spin_unlock(&fs_info->reada_lock);
 453                         btrfs_dev_replace_unlock(&fs_info->dev_replace);
 454                         goto error;
 455                 }
 456         }
 457         spin_unlock(&fs_info->reada_lock);
 458         btrfs_dev_replace_unlock(&fs_info->dev_replace);
 459
 460         kfree(bbio);
 461         return re;
 462
 463 error:
 464         while (nzones) {
 465                 struct reada_zone *zone;
 466
 467                 --nzones;
 468                 zone = re->zones[nzones];
 469                 kref_get(&zone->refcnt);
 470                 spin_lock(&zone->lock);
 471                 --zone->elems;
 472                 if (zone->elems == 0) {
 473                         /*
 474                          * no fs_info->reada_lock needed, as this can't be
 475                          * the last ref
 476                          */
 477                         kref_put(&zone->refcnt, reada_zone_release);
 478                 }
 479                 spin_unlock(&zone->lock);
 480
 481                 spin_lock(&fs_info->reada_lock);
 482                 kref_put(&zone->refcnt, reada_zone_release);
 483                 spin_unlock(&fs_info->reada_lock);
 484         }
 485         kfree(bbio);
 486         kfree(re);
 487         return re_exist;
 488 }
 489
 490 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 491                              struct reada_extent *re)
 492 {
 493         int i;
 494         unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
 495
 496         spin_lock(&fs_info->reada_lock);
 497         if (--re->refcnt) {
 498                 spin_unlock(&fs_info->reada_lock);
 499                 return;
 500         }
 501
 502         radix_tree_delete(&fs_info->reada_tree, index);
 503         for (i = 0; i < re->nzones; ++i) {
 504                 struct reada_zone *zone = re->zones[i];
 505
 506                 radix_tree_delete(&zone->device->reada_extents, index);
 507         }
 508
 509         spin_unlock(&fs_info->reada_lock);
 510
 511         for (i = 0; i < re->nzones; ++i) {
 512                 struct reada_zone *zone = re->zones[i];
 513
 514                 kref_get(&zone->refcnt);
 515                 spin_lock(&zone->lock);
 516                 --zone->elems;
 517                 if (zone->elems == 0) {
 518                         /* no fs_info->reada_lock needed, as this can't be
 519                          * the last ref */
 520                         kref_put(&zone->refcnt, reada_zone_release);
 521                 }
 522                 spin_unlock(&zone->lock);
 523
 524                 spin_lock(&fs_info->reada_lock);
 525                 kref_put(&zone->refcnt, reada_zone_release);
 526                 spin_unlock(&fs_info->reada_lock);
 527         }
 528         if (re->scheduled_for)
 529                 atomic_dec(&re->scheduled_for->reada_in_flight);
 530
 531         kfree(re);
 532 }
 533
 534 static void reada_zone_release(struct kref *kref)
 535 {
 536         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 537
 538         radix_tree_delete(&zone->device->reada_zones,
 539                           zone->end >> PAGE_CACHE_SHIFT);
 540
 541         kfree(zone);
 542 }
 543
 544 static void reada_control_release(struct kref *kref)
 545 {
 546         struct reada_control *rc = container_of(kref, struct reada_control,
 547                                                 refcnt);
 548
 549         kfree(rc);
 550 }
 551
 552 static int reada_add_block(struct reada_control *rc, u64 logical,
 553                            struct btrfs_key *top, int level, u64 generation)
 554 {
 555         struct btrfs_root *root = rc->root;
 556         struct reada_extent *re;
 557         struct reada_extctl *rec;
 558
 559         re = reada_find_extent(root, logical, top, level); /* takes one ref */
 560         if (!re)
 561                 return -1;
 562
 563         rec = kzalloc(sizeof(*rec), GFP_NOFS);
 564         if (!rec) {
 565                 reada_extent_put(root->fs_info, re);
 566                 return -1;
 567         }
 568
 569         rec->rc = rc;
 570         rec->generation = generation;
 571         atomic_inc(&rc->elems);
 572
 573         spin_lock(&re->lock);
 574         list_add_tail(&rec->list, &re->extctl);
 575         spin_unlock(&re->lock);
 576
 577         /* leave the ref on the extent */
 578
 579         return 0;
 580 }
 581
 582 /*
 583  * called with fs_info->reada_lock held
 584  */
 585 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 586 {
 587         int i;
 588         unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
 589
 590         for (i = 0; i < zone->ndevs; ++i) {
 591                 struct reada_zone *peer;
 592                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 593                 if (peer && peer->device != zone->device)
 594                         peer->locked = lock;
 595         }
 596 }
 597
 598 /*
 599  * called with fs_info->reada_lock held
 600  */
 601 static int reada_pick_zone(struct btrfs_device *dev)
 602 {
 603         struct reada_zone *top_zone = NULL;
 604         struct reada_zone *top_locked_zone = NULL;
 605         u64 top_elems = 0;
 606         u64 top_locked_elems = 0;
 607         unsigned long index = 0;
 608         int ret;
 609
 610         if (dev->reada_curr_zone) {
 611                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 612                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 613                 dev->reada_curr_zone = NULL;
 614         }
 615         /* pick the zone with the most elements */
 616         while (1) {
 617                 struct reada_zone *zone;
 618
 619                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 620                                              (void **)&zone, index, 1);
 621                 if (ret == 0)
 622                         break;
 623                 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
 624                 if (zone->locked) {
 625                         if (zone->elems > top_locked_elems) {
 626                                 top_locked_elems = zone->elems;
 627                                 top_locked_zone = zone;
 628                         }
 629                 } else {
 630                         if (zone->elems > top_elems) {
 631                                 top_elems = zone->elems;
 632                                 top_zone = zone;
 633                         }
 634                 }
 635         }
 636         if (top_zone)
 637                 dev->reada_curr_zone = top_zone;
 638         else if (top_locked_zone)
 639                 dev->reada_curr_zone = top_locked_zone;
 640         else
 641                 return 0;
 642
 643         dev->reada_next = dev->reada_curr_zone->start;
 644         kref_get(&dev->reada_curr_zone->refcnt);
 645         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 646
 647         return 1;
 648 }
 649
 650 static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
 651                                    struct btrfs_device *dev)
 652 {
 653         struct reada_extent *re = NULL;
 654         int mirror_num = 0;
 655         struct extent_buffer *eb = NULL;
 656         u64 logical;
 657         u32 blocksize;
 658         int ret;
 659         int i;
 660         int need_kick = 0;
 661
 662         spin_lock(&fs_info->reada_lock);
 663         if (dev->reada_curr_zone == NULL) {
 664                 ret = reada_pick_zone(dev);
 665                 if (!ret) {
 666                         spin_unlock(&fs_info->reada_lock);
 667                         return 0;
 668                 }
 669         }
 670         /*
 671          * FIXME currently we issue the reads one extent at a time. If we have
 672          * a contiguous block of extents, we could also coagulate them or use
 673          * plugging to speed things up
 674          */
 675         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 676                                      dev->reada_next >> PAGE_CACHE_SHIFT, 1);
 677         if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
 678                 ret = reada_pick_zone(dev);
 679                 if (!ret) {
 680                         spin_unlock(&fs_info->reada_lock);
 681                         return 0;
 682                 }
 683                 re = NULL;
 684                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 685                                         dev->reada_next >> PAGE_CACHE_SHIFT, 1);
 686         }
 687         if (ret == 0) {
 688                 spin_unlock(&fs_info->reada_lock);
 689                 return 0;
 690         }
 691         dev->reada_next = re->logical + re->blocksize;
 692         re->refcnt++;
 693
 694         spin_unlock(&fs_info->reada_lock);
 695
 696         /*
 697          * find mirror num
 698          */
 699         for (i = 0; i < re->nzones; ++i) {
 700                 if (re->zones[i]->device == dev) {
 701                         mirror_num = i + 1;
 702                         break;
 703                 }
 704         }
 705         logical = re->logical;
 706         blocksize = re->blocksize;
 707
 708         spin_lock(&re->lock);
 709         if (re->scheduled_for == NULL) {
 710                 re->scheduled_for = dev;
 711                 need_kick = 1;
 712         }
 713         spin_unlock(&re->lock);
 714
 715         reada_extent_put(fs_info, re);
 716
 717         if (!need_kick)
 718                 return 0;
 719
 720         atomic_inc(&dev->reada_in_flight);
 721         ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
 722                          mirror_num, &eb);
 723         if (ret)
 724                 __readahead_hook(fs_info->extent_root, NULL, logical, ret);
 725         else if (eb)
 726                 __readahead_hook(fs_info->extent_root, eb, eb->start, ret);
 727
 728         if (eb)
 729                 free_extent_buffer(eb);
 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
 753 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 754 {
 755         struct btrfs_device *device;
 756         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 757         u64 enqueued;
 758         u64 total = 0;
 759         int i;
 760
 761         do {
 762                 enqueued = 0;
 763                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 764                         if (atomic_read(&device->reada_in_flight) <
 765                             MAX_IN_FLIGHT)
 766                                 enqueued += reada_start_machine_dev(fs_info,
 767                                                                     device);
 768                 }
 769                 total += enqueued;
 770         } while (enqueued && total < 10000);
 771
 772         if (enqueued == 0)
 773                 return;
 774
 775         /*
 776          * If everything is already in the cache, this is effectively single
 777          * threaded. To a) not hold the caller for too long and b) to utilize
 778          * more cores, we broke the loop above after 10000 iterations and now
 779          * enqueue to workers to finish it. This will distribute the load to
 780          * the cores.
 781          */
 782         for (i = 0; i < 2; ++i)
 783                 reada_start_machine(fs_info);
 784 }
 785
 786 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 787 {
 788         struct reada_machine_work *rmw;
 789
 790         rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
 791         if (!rmw) {
 792                 /* FIXME we cannot handle this properly right now */
 793                 BUG();
 794         }
 795         rmw->work.func = reada_start_machine_worker;
 796         rmw->fs_info = fs_info;
 797
 798         btrfs_queue_worker(&fs_info->readahead_workers, &rmw->work);
 799 }
 800
 801 #ifdef DEBUG
 802 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 803 {
 804         struct btrfs_device *device;
 805         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 806         unsigned long index;
 807         int ret;
 808         int i;
 809         int j;
 810         int cnt;
 811
 812         spin_lock(&fs_info->reada_lock);
 813         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 814                 printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
 815                         atomic_read(&device->reada_in_flight));
 816                 index = 0;
 817                 while (1) {
 818                         struct reada_zone *zone;
 819                         ret = radix_tree_gang_lookup(&device->reada_zones,
 820                                                      (void **)&zone, index, 1);
 821                         if (ret == 0)
 822                                 break;
 823                         printk(KERN_DEBUG "  zone %llu-%llu elems %llu locked "
 824                                 "%d devs", zone->start, zone->end, zone->elems,
 825                                 zone->locked);
 826                         for (j = 0; j < zone->ndevs; ++j) {
 827                                 printk(KERN_CONT " %lld",
 828                                         zone->devs[j]->devid);
 829                         }
 830                         if (device->reada_curr_zone == zone)
 831                                 printk(KERN_CONT " curr off %llu",
 832                                         device->reada_next - zone->start);
 833                         printk(KERN_CONT "\n");
 834                         index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
 835                 }
 836                 cnt = 0;
 837                 index = 0;
 838                 while (all) {
 839                         struct reada_extent *re = NULL;
 840
 841                         ret = radix_tree_gang_lookup(&device->reada_extents,
 842                                                      (void **)&re, index, 1);
 843                         if (ret == 0)
 844                                 break;
 845                         printk(KERN_DEBUG
 846                                 "  re: logical %llu size %u empty %d for %lld",
 847                                 re->logical, re->blocksize,
 848                                 list_empty(&re->extctl), re->scheduled_for ?
 849                                 re->scheduled_for->devid : -1);
 850
 851                         for (i = 0; i < re->nzones; ++i) {
 852                                 printk(KERN_CONT " zone %llu-%llu devs",
 853                                         re->zones[i]->start,
 854                                         re->zones[i]->end);
 855                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 856                                         printk(KERN_CONT " %lld",
 857                                                 re->zones[i]->devs[j]->devid);
 858                                 }
 859                         }
 860                         printk(KERN_CONT "\n");
 861                         index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
 862                         if (++cnt > 15)
 863                                 break;
 864                 }
 865         }
 866
 867         index = 0;
 868         cnt = 0;
 869         while (all) {
 870                 struct reada_extent *re = NULL;
 871
 872                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 873                                              index, 1);
 874                 if (ret == 0)
 875                         break;
 876                 if (!re->scheduled_for) {
 877                         index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
 878                         continue;
 879                 }
 880                 printk(KERN_DEBUG
 881                         "re: logical %llu size %u list empty %d for %lld",
 882                         re->logical, re->blocksize, list_empty(&re->extctl),
 883                         re->scheduled_for ? re->scheduled_for->devid : -1);
 884                 for (i = 0; i < re->nzones; ++i) {
 885                         printk(KERN_CONT " zone %llu-%llu devs",
 886                                 re->zones[i]->start,
 887                                 re->zones[i]->end);
 888                         for (i = 0; i < re->nzones; ++i) {
 889                                 printk(KERN_CONT " zone %llu-%llu devs",
 890                                         re->zones[i]->start,
 891                                         re->zones[i]->end);
 892                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 893                                         printk(KERN_CONT " %lld",
 894                                                 re->zones[i]->devs[j]->devid);
 895                                 }
 896                         }
 897                 }
 898                 printk(KERN_CONT "\n");
 899                 index = (re->logical >> PAGE_CACHE_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 level;
 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_NOFS);
 923         if (!rc)
 924                 return ERR_PTR(-ENOMEM);
 925
 926         rc->root = root;
 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         level = btrfs_header_level(node);
 937         generation = btrfs_header_generation(node);
 938         free_extent_buffer(node);
 939
 940         if (reada_add_block(rc, start, &max_key, level, generation)) {
 941                 kfree(rc);
 942                 return ERR_PTR(-ENOMEM);
 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
 955         while (atomic_read(&rc->elems)) {
 956                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 957                                    5 * HZ);
 958                 dump_devs(rc->root->fs_info,
 959                           atomic_read(&rc->elems) < 10 ? 1 : 0);
 960         }
 961
 962         dump_devs(rc->root->fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 963
 964         kref_put(&rc->refcnt, reada_control_release);
 965
 966         return 0;
 967 }
 968 #else
 969 int btrfs_reada_wait(void *handle)
 970 {
 971         struct reada_control *rc = handle;
 972
 973         while (atomic_read(&rc->elems)) {
 974                 wait_event(rc->wait, atomic_read(&rc->elems) == 0);
 975         }
 976
 977         kref_put(&rc->refcnt, reada_control_release);
 978
 979         return 0;
 980 }
 981 #endif
 982
 983 void btrfs_reada_detach(void *handle)
 984 {
 985         struct reada_control *rc = handle;
 986
 987         kref_put(&rc->refcnt, reada_control_release);
 988 }