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                                 btrfs_debug(root->fs_info,
 193                                            "generation mismatch for (%llu,%d,%llu) %llu != %llu",
 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                 btrfs_err(root->fs_info,
 369                            "readahead: more than %d copies not supported",
 370                            BTRFS_MAX_MIRRORS);
 371                 goto error;
 372         }
 373
 374         for (nzones = 0; nzones < bbio->num_stripes; ++nzones) {
 375                 struct reada_zone *zone;
 376
 377                 dev = bbio->stripes[nzones].dev;
 378                 zone = reada_find_zone(fs_info, dev, logical, bbio);
 379                 if (!zone)
 380                         break;
 381
 382                 re->zones[nzones] = zone;
 383                 spin_lock(&zone->lock);
 384                 if (!zone->elems)
 385                         kref_get(&zone->refcnt);
 386                 ++zone->elems;
 387                 spin_unlock(&zone->lock);
 388                 spin_lock(&fs_info->reada_lock);
 389                 kref_put(&zone->refcnt, reada_zone_release);
 390                 spin_unlock(&fs_info->reada_lock);
 391         }
 392         re->nzones = nzones;
 393         if (nzones == 0) {
 394                 /* not a single zone found, error and out */
 395                 goto error;
 396         }
 397
 398         /* insert extent in reada_tree + all per-device trees, all or nothing */
 399         btrfs_dev_replace_lock(&fs_info->dev_replace);
 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                 BUG_ON(!re_exist);
 405                 re_exist->refcnt++;
 406                 spin_unlock(&fs_info->reada_lock);
 407                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
 408                 goto error;
 409         }
 410         if (ret) {
 411                 spin_unlock(&fs_info->reada_lock);
 412                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
 413                 goto error;
 414         }
 415         prev_dev = NULL;
 416         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 417                         &fs_info->dev_replace);
 418         for (i = 0; i < nzones; ++i) {
 419                 dev = bbio->stripes[i].dev;
 420                 if (dev == prev_dev) {
 421                         /*
 422                          * in case of DUP, just add the first zone. As both
 423                          * are on the same device, there's nothing to gain
 424                          * from adding both.
 425                          * Also, it wouldn't work, as the tree is per device
 426                          * and adding would fail with EEXIST
 427                          */
 428                         continue;
 429                 }
 430                 if (!dev->bdev) {
 431                         /*
 432                          * cannot read ahead on missing device, but for RAID5/6,
 433                          * REQ_GET_READ_MIRRORS return 1. So don't skip missing
 434                          * device for such case.
 435                          */
 436                         if (nzones > 1)
 437                                 continue;
 438                 }
 439                 if (dev_replace_is_ongoing &&
 440                     dev == fs_info->dev_replace.tgtdev) {
 441                         /*
 442                          * as this device is selected for reading only as
 443                          * a last resort, skip it for read ahead.
 444                          */
 445                         continue;
 446                 }
 447                 prev_dev = dev;
 448                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 449                 if (ret) {
 450                         while (--i >= 0) {
 451                                 dev = bbio->stripes[i].dev;
 452                                 BUG_ON(dev == NULL);
 453                                 /* ignore whether the entry was inserted */
 454                                 radix_tree_delete(&dev->reada_extents, index);
 455                         }
 456                         BUG_ON(fs_info == NULL);
 457                         radix_tree_delete(&fs_info->reada_tree, index);
 458                         spin_unlock(&fs_info->reada_lock);
 459                         btrfs_dev_replace_unlock(&fs_info->dev_replace);
 460                         goto error;
 461                 }
 462         }
 463         spin_unlock(&fs_info->reada_lock);
 464         btrfs_dev_replace_unlock(&fs_info->dev_replace);
 465
 466         kfree(bbio);
 467         return re;
 468
 469 error:
 470         while (nzones) {
 471                 struct reada_zone *zone;
 472
 473                 --nzones;
 474                 zone = re->zones[nzones];
 475                 kref_get(&zone->refcnt);
 476                 spin_lock(&zone->lock);
 477                 --zone->elems;
 478                 if (zone->elems == 0) {
 479                         /*
 480                          * no fs_info->reada_lock needed, as this can't be
 481                          * the last ref
 482                          */
 483                         kref_put(&zone->refcnt, reada_zone_release);
 484                 }
 485                 spin_unlock(&zone->lock);
 486
 487                 spin_lock(&fs_info->reada_lock);
 488                 kref_put(&zone->refcnt, reada_zone_release);
 489                 spin_unlock(&fs_info->reada_lock);
 490         }
 491         kfree(bbio);
 492         kfree(re);
 493         return re_exist;
 494 }
 495
 496 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 497                              struct reada_extent *re)
 498 {
 499         int i;
 500         unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
 501
 502         spin_lock(&fs_info->reada_lock);
 503         if (--re->refcnt) {
 504                 spin_unlock(&fs_info->reada_lock);
 505                 return;
 506         }
 507
 508         radix_tree_delete(&fs_info->reada_tree, index);
 509         for (i = 0; i < re->nzones; ++i) {
 510                 struct reada_zone *zone = re->zones[i];
 511
 512                 radix_tree_delete(&zone->device->reada_extents, index);
 513         }
 514
 515         spin_unlock(&fs_info->reada_lock);
 516
 517         for (i = 0; i < re->nzones; ++i) {
 518                 struct reada_zone *zone = re->zones[i];
 519
 520                 kref_get(&zone->refcnt);
 521                 spin_lock(&zone->lock);
 522                 --zone->elems;
 523                 if (zone->elems == 0) {
 524                         /* no fs_info->reada_lock needed, as this can't be
 525                          * the last ref */
 526                         kref_put(&zone->refcnt, reada_zone_release);
 527                 }
 528                 spin_unlock(&zone->lock);
 529
 530                 spin_lock(&fs_info->reada_lock);
 531                 kref_put(&zone->refcnt, reada_zone_release);
 532                 spin_unlock(&fs_info->reada_lock);
 533         }
 534         if (re->scheduled_for)
 535                 atomic_dec(&re->scheduled_for->reada_in_flight);
 536
 537         kfree(re);
 538 }
 539
 540 static void reada_zone_release(struct kref *kref)
 541 {
 542         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 543
 544         radix_tree_delete(&zone->device->reada_zones,
 545                           zone->end >> PAGE_CACHE_SHIFT);
 546
 547         kfree(zone);
 548 }
 549
 550 static void reada_control_release(struct kref *kref)
 551 {
 552         struct reada_control *rc = container_of(kref, struct reada_control,
 553                                                 refcnt);
 554
 555         kfree(rc);
 556 }
 557
 558 static int reada_add_block(struct reada_control *rc, u64 logical,
 559                            struct btrfs_key *top, int level, u64 generation)
 560 {
 561         struct btrfs_root *root = rc->root;
 562         struct reada_extent *re;
 563         struct reada_extctl *rec;
 564
 565         re = reada_find_extent(root, logical, top, level); /* takes one ref */
 566         if (!re)
 567                 return -1;
 568
 569         rec = kzalloc(sizeof(*rec), GFP_NOFS);
 570         if (!rec) {
 571                 reada_extent_put(root->fs_info, re);
 572                 return -1;
 573         }
 574
 575         rec->rc = rc;
 576         rec->generation = generation;
 577         atomic_inc(&rc->elems);
 578
 579         spin_lock(&re->lock);
 580         list_add_tail(&rec->list, &re->extctl);
 581         spin_unlock(&re->lock);
 582
 583         /* leave the ref on the extent */
 584
 585         return 0;
 586 }
 587
 588 /*
 589  * called with fs_info->reada_lock held
 590  */
 591 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 592 {
 593         int i;
 594         unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
 595
 596         for (i = 0; i < zone->ndevs; ++i) {
 597                 struct reada_zone *peer;
 598                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 599                 if (peer && peer->device != zone->device)
 600                         peer->locked = lock;
 601         }
 602 }
 603
 604 /*
 605  * called with fs_info->reada_lock held
 606  */
 607 static int reada_pick_zone(struct btrfs_device *dev)
 608 {
 609         struct reada_zone *top_zone = NULL;
 610         struct reada_zone *top_locked_zone = NULL;
 611         u64 top_elems = 0;
 612         u64 top_locked_elems = 0;
 613         unsigned long index = 0;
 614         int ret;
 615
 616         if (dev->reada_curr_zone) {
 617                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 618                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 619                 dev->reada_curr_zone = NULL;
 620         }
 621         /* pick the zone with the most elements */
 622         while (1) {
 623                 struct reada_zone *zone;
 624
 625                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 626                                              (void **)&zone, index, 1);
 627                 if (ret == 0)
 628                         break;
 629                 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
 630                 if (zone->locked) {
 631                         if (zone->elems > top_locked_elems) {
 632                                 top_locked_elems = zone->elems;
 633                                 top_locked_zone = zone;
 634                         }
 635                 } else {
 636                         if (zone->elems > top_elems) {
 637                                 top_elems = zone->elems;
 638                                 top_zone = zone;
 639                         }
 640                 }
 641         }
 642         if (top_zone)
 643                 dev->reada_curr_zone = top_zone;
 644         else if (top_locked_zone)
 645                 dev->reada_curr_zone = top_locked_zone;
 646         else
 647                 return 0;
 648
 649         dev->reada_next = dev->reada_curr_zone->start;
 650         kref_get(&dev->reada_curr_zone->refcnt);
 651         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 652
 653         return 1;
 654 }
 655
 656 static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
 657                                    struct btrfs_device *dev)
 658 {
 659         struct reada_extent *re = NULL;
 660         int mirror_num = 0;
 661         struct extent_buffer *eb = NULL;
 662         u64 logical;
 663         u32 blocksize;
 664         int ret;
 665         int i;
 666         int need_kick = 0;
 667
 668         spin_lock(&fs_info->reada_lock);
 669         if (dev->reada_curr_zone == NULL) {
 670                 ret = reada_pick_zone(dev);
 671                 if (!ret) {
 672                         spin_unlock(&fs_info->reada_lock);
 673                         return 0;
 674                 }
 675         }
 676         /*
 677          * FIXME currently we issue the reads one extent at a time. If we have
 678          * a contiguous block of extents, we could also coagulate them or use
 679          * plugging to speed things up
 680          */
 681         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 682                                      dev->reada_next >> PAGE_CACHE_SHIFT, 1);
 683         if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
 684                 ret = reada_pick_zone(dev);
 685                 if (!ret) {
 686                         spin_unlock(&fs_info->reada_lock);
 687                         return 0;
 688                 }
 689                 re = NULL;
 690                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 691                                         dev->reada_next >> PAGE_CACHE_SHIFT, 1);
 692         }
 693         if (ret == 0) {
 694                 spin_unlock(&fs_info->reada_lock);
 695                 return 0;
 696         }
 697         dev->reada_next = re->logical + re->blocksize;
 698         re->refcnt++;
 699
 700         spin_unlock(&fs_info->reada_lock);
 701
 702         /*
 703          * find mirror num
 704          */
 705         for (i = 0; i < re->nzones; ++i) {
 706                 if (re->zones[i]->device == dev) {
 707                         mirror_num = i + 1;
 708                         break;
 709                 }
 710         }
 711         logical = re->logical;
 712         blocksize = re->blocksize;
 713
 714         spin_lock(&re->lock);
 715         if (re->scheduled_for == NULL) {
 716                 re->scheduled_for = dev;
 717                 need_kick = 1;
 718         }
 719         spin_unlock(&re->lock);
 720
 721         reada_extent_put(fs_info, re);
 722
 723         if (!need_kick)
 724                 return 0;
 725
 726         atomic_inc(&dev->reada_in_flight);
 727         ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
 728                          mirror_num, &eb);
 729         if (ret)
 730                 __readahead_hook(fs_info->extent_root, NULL, logical, ret);
 731         else if (eb)
 732                 __readahead_hook(fs_info->extent_root, eb, eb->start, ret);
 733
 734         if (eb)
 735                 free_extent_buffer(eb);
 736
 737         return 1;
 738
 739 }
 740
 741 static void reada_start_machine_worker(struct btrfs_work *work)
 742 {
 743         struct reada_machine_work *rmw;
 744         struct btrfs_fs_info *fs_info;
 745         int old_ioprio;
 746
 747         rmw = container_of(work, struct reada_machine_work, work);
 748         fs_info = rmw->fs_info;
 749
 750         kfree(rmw);
 751
 752         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 753                                        task_nice_ioprio(current));
 754         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 755         __reada_start_machine(fs_info);
 756         set_task_ioprio(current, old_ioprio);
 757 }
 758
 759 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 760 {
 761         struct btrfs_device *device;
 762         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 763         u64 enqueued;
 764         u64 total = 0;
 765         int i;
 766
 767         do {
 768                 enqueued = 0;
 769                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 770                         if (atomic_read(&device->reada_in_flight) <
 771                             MAX_IN_FLIGHT)
 772                                 enqueued += reada_start_machine_dev(fs_info,
 773                                                                     device);
 774                 }
 775                 total += enqueued;
 776         } while (enqueued && total < 10000);
 777
 778         if (enqueued == 0)
 779                 return;
 780
 781         /*
 782          * If everything is already in the cache, this is effectively single
 783          * threaded. To a) not hold the caller for too long and b) to utilize
 784          * more cores, we broke the loop above after 10000 iterations and now
 785          * enqueue to workers to finish it. This will distribute the load to
 786          * the cores.
 787          */
 788         for (i = 0; i < 2; ++i)
 789                 reada_start_machine(fs_info);
 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_NOFS);
 797         if (!rmw) {
 798                 /* FIXME we cannot handle this properly right now */
 799                 BUG();
 800         }
 801         btrfs_init_work(&rmw->work, reada_start_machine_worker, NULL, NULL);
 802         rmw->fs_info = fs_info;
 803
 804         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 805 }
 806
 807 #ifdef DEBUG
 808 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 809 {
 810         struct btrfs_device *device;
 811         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 812         unsigned long index;
 813         int ret;
 814         int i;
 815         int j;
 816         int cnt;
 817
 818         spin_lock(&fs_info->reada_lock);
 819         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 820                 printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
 821                         atomic_read(&device->reada_in_flight));
 822                 index = 0;
 823                 while (1) {
 824                         struct reada_zone *zone;
 825                         ret = radix_tree_gang_lookup(&device->reada_zones,
 826                                                      (void **)&zone, index, 1);
 827                         if (ret == 0)
 828                                 break;
 829                         printk(KERN_DEBUG "  zone %llu-%llu elems %llu locked "
 830                                 "%d devs", zone->start, zone->end, zone->elems,
 831                                 zone->locked);
 832                         for (j = 0; j < zone->ndevs; ++j) {
 833                                 printk(KERN_CONT " %lld",
 834                                         zone->devs[j]->devid);
 835                         }
 836                         if (device->reada_curr_zone == zone)
 837                                 printk(KERN_CONT " curr off %llu",
 838                                         device->reada_next - zone->start);
 839                         printk(KERN_CONT "\n");
 840                         index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
 841                 }
 842                 cnt = 0;
 843                 index = 0;
 844                 while (all) {
 845                         struct reada_extent *re = NULL;
 846
 847                         ret = radix_tree_gang_lookup(&device->reada_extents,
 848                                                      (void **)&re, index, 1);
 849                         if (ret == 0)
 850                                 break;
 851                         printk(KERN_DEBUG
 852                                 "  re: logical %llu size %u empty %d for %lld",
 853                                 re->logical, re->blocksize,
 854                                 list_empty(&re->extctl), re->scheduled_for ?
 855                                 re->scheduled_for->devid : -1);
 856
 857                         for (i = 0; i < re->nzones; ++i) {
 858                                 printk(KERN_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                                         printk(KERN_CONT " %lld",
 863                                                 re->zones[i]->devs[j]->devid);
 864                                 }
 865                         }
 866                         printk(KERN_CONT "\n");
 867                         index = (re->logical >> PAGE_CACHE_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_for) {
 883                         index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
 884                         continue;
 885                 }
 886                 printk(KERN_DEBUG
 887                         "re: logical %llu size %u list empty %d for %lld",
 888                         re->logical, re->blocksize, list_empty(&re->extctl),
 889                         re->scheduled_for ? re->scheduled_for->devid : -1);
 890                 for (i = 0; i < re->nzones; ++i) {
 891                         printk(KERN_CONT " zone %llu-%llu devs",
 892                                 re->zones[i]->start,
 893                                 re->zones[i]->end);
 894                         for (i = 0; i < re->nzones; ++i) {
 895                                 printk(KERN_CONT " zone %llu-%llu devs",
 896                                         re->zones[i]->start,
 897                                         re->zones[i]->end);
 898                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 899                                         printk(KERN_CONT " %lld",
 900                                                 re->zones[i]->devs[j]->devid);
 901                                 }
 902                         }
 903                 }
 904                 printk(KERN_CONT "\n");
 905                 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
 906         }
 907         spin_unlock(&fs_info->reada_lock);
 908 }
 909 #endif
 910
 911 /*
 912  * interface
 913  */
 914 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 915                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 916 {
 917         struct reada_control *rc;
 918         u64 start;
 919         u64 generation;
 920         int level;
 921         struct extent_buffer *node;
 922         static struct btrfs_key max_key = {
 923                 .objectid = (u64)-1,
 924                 .type = (u8)-1,
 925                 .offset = (u64)-1
 926         };
 927
 928         rc = kzalloc(sizeof(*rc), GFP_NOFS);
 929         if (!rc)
 930                 return ERR_PTR(-ENOMEM);
 931
 932         rc->root = root;
 933         rc->key_start = *key_start;
 934         rc->key_end = *key_end;
 935         atomic_set(&rc->elems, 0);
 936         init_waitqueue_head(&rc->wait);
 937         kref_init(&rc->refcnt);
 938         kref_get(&rc->refcnt); /* one ref for having elements */
 939
 940         node = btrfs_root_node(root);
 941         start = node->start;
 942         level = btrfs_header_level(node);
 943         generation = btrfs_header_generation(node);
 944         free_extent_buffer(node);
 945
 946         if (reada_add_block(rc, start, &max_key, level, generation)) {
 947                 kfree(rc);
 948                 return ERR_PTR(-ENOMEM);
 949         }
 950
 951         reada_start_machine(root->fs_info);
 952
 953         return rc;
 954 }
 955
 956 #ifdef DEBUG
 957 int btrfs_reada_wait(void *handle)
 958 {
 959         struct reada_control *rc = handle;
 960
 961         while (atomic_read(&rc->elems)) {
 962                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 963                                    5 * HZ);
 964                 dump_devs(rc->root->fs_info,
 965                           atomic_read(&rc->elems) < 10 ? 1 : 0);
 966         }
 967
 968         dump_devs(rc->root->fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 969
 970         kref_put(&rc->refcnt, reada_control_release);
 971
 972         return 0;
 973 }
 974 #else
 975 int btrfs_reada_wait(void *handle)
 976 {
 977         struct reada_control *rc = handle;
 978
 979         while (atomic_read(&rc->elems)) {
 980                 wait_event(rc->wait, atomic_read(&rc->elems) == 0);
 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 }