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 btrfs_fs_info *fs_info,
 213                          struct extent_buffer *eb, int err)
 214 {
 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_fs_info *fs_info,
 239                                           struct btrfs_device *dev, u64 logical,
 240                                           struct btrfs_bio *bbio)
 241 {
 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         zone->start = start;
 274         zone->end = end;
 275         INIT_LIST_HEAD(&zone->list);
 276         spin_lock_init(&zone->lock);
 277         zone->locked = 0;
 278         kref_init(&zone->refcnt);
 279         zone->elems = 0;
 280         zone->device = dev; /* our device always sits at index 0 */
 281         for (i = 0; i < bbio->num_stripes; ++i) {
 282                 /* bounds have already been checked */
 283                 zone->devs[i] = bbio->stripes[i].dev;
 284         }
 285         zone->ndevs = bbio->num_stripes;
 286
 287         spin_lock(&fs_info->reada_lock);
 288         ret = radix_tree_insert(&dev->reada_zones,
 289                                 (unsigned long)(zone->end >> PAGE_SHIFT),
 290                                 zone);
 291
 292         if (ret == -EEXIST) {
 293                 kfree(zone);
 294                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 295                                              logical >> PAGE_SHIFT, 1);
 296                 if (ret == 1 && logical >= zone->start && logical <= zone->end)
 297                         kref_get(&zone->refcnt);
 298                 else
 299                         zone = NULL;
 300         }
 301         spin_unlock(&fs_info->reada_lock);
 302
 303         return zone;
 304 }
 305
 306 static struct reada_extent *reada_find_extent(struct btrfs_root *root,
 307                                               u64 logical,
 308                                               struct btrfs_key *top)
 309 {
 310         int ret;
 311         struct reada_extent *re = NULL;
 312         struct reada_extent *re_exist = NULL;
 313         struct btrfs_fs_info *fs_info = root->fs_info;
 314         struct btrfs_bio *bbio = NULL;
 315         struct btrfs_device *dev;
 316         struct btrfs_device *prev_dev;
 317         u32 blocksize;
 318         u64 length;
 319         int real_stripes;
 320         int nzones = 0;
 321         unsigned long index = logical >> PAGE_SHIFT;
 322         int dev_replace_is_ongoing;
 323         int have_zone = 0;
 324
 325         spin_lock(&fs_info->reada_lock);
 326         re = radix_tree_lookup(&fs_info->reada_tree, index);
 327         if (re)
 328                 re->refcnt++;
 329         spin_unlock(&fs_info->reada_lock);
 330
 331         if (re)
 332                 return re;
 333
 334         re = kzalloc(sizeof(*re), GFP_KERNEL);
 335         if (!re)
 336                 return NULL;
 337
 338         blocksize = fs_info->nodesize;
 339         re->logical = logical;
 340         re->top = *top;
 341         INIT_LIST_HEAD(&re->extctl);
 342         spin_lock_init(&re->lock);
 343         re->refcnt = 1;
 344
 345         /*
 346          * map block
 347          */
 348         length = blocksize;
 349         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
 350                         &length, &bbio, 0);
 351         if (ret || !bbio || length < blocksize)
 352                 goto error;
 353
 354         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 355                 btrfs_err(fs_info,
 356                            "readahead: more than %d copies not supported",
 357                            BTRFS_MAX_MIRRORS);
 358                 goto error;
 359         }
 360
 361         real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
 362         for (nzones = 0; nzones < real_stripes; ++nzones) {
 363                 struct reada_zone *zone;
 364
 365                 dev = bbio->stripes[nzones].dev;
 366
 367                 /* cannot read ahead on missing device. */
 368                  if (!dev->bdev)
 369                         continue;
 370
 371                 zone = reada_find_zone(fs_info, dev, logical, bbio);
 372                 if (!zone)
 373                         continue;
 374
 375                 re->zones[re->nzones++] = zone;
 376                 spin_lock(&zone->lock);
 377                 if (!zone->elems)
 378                         kref_get(&zone->refcnt);
 379                 ++zone->elems;
 380                 spin_unlock(&zone->lock);
 381                 spin_lock(&fs_info->reada_lock);
 382                 kref_put(&zone->refcnt, reada_zone_release);
 383                 spin_unlock(&fs_info->reada_lock);
 384         }
 385         if (re->nzones == 0) {
 386                 /* not a single zone found, error and out */
 387                 goto error;
 388         }
 389
 390         /* insert extent in reada_tree + all per-device trees, all or nothing */
 391         btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
 392         spin_lock(&fs_info->reada_lock);
 393         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 394         if (ret == -EEXIST) {
 395                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 396                 re_exist->refcnt++;
 397                 spin_unlock(&fs_info->reada_lock);
 398                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 399                 goto error;
 400         }
 401         if (ret) {
 402                 spin_unlock(&fs_info->reada_lock);
 403                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 404                 goto error;
 405         }
 406         prev_dev = NULL;
 407         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 408                         &fs_info->dev_replace);
 409         for (nzones = 0; nzones < re->nzones; ++nzones) {
 410                 dev = re->zones[nzones]->device;
 411
 412                 if (dev == prev_dev) {
 413                         /*
 414                          * in case of DUP, just add the first zone. As both
 415                          * are on the same device, there's nothing to gain
 416                          * from adding both.
 417                          * Also, it wouldn't work, as the tree is per device
 418                          * and adding would fail with EEXIST
 419                          */
 420                         continue;
 421                 }
 422                 if (!dev->bdev)
 423                         continue;
 424
 425                 if (dev_replace_is_ongoing &&
 426                     dev == fs_info->dev_replace.tgtdev) {
 427                         /*
 428                          * as this device is selected for reading only as
 429                          * a last resort, skip it for read ahead.
 430                          */
 431                         continue;
 432                 }
 433                 prev_dev = dev;
 434                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 435                 if (ret) {
 436                         while (--nzones >= 0) {
 437                                 dev = re->zones[nzones]->device;
 438                                 BUG_ON(dev == NULL);
 439                                 /* ignore whether the entry was inserted */
 440                                 radix_tree_delete(&dev->reada_extents, index);
 441                         }
 442                         radix_tree_delete(&fs_info->reada_tree, index);
 443                         spin_unlock(&fs_info->reada_lock);
 444                         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 445                         goto error;
 446                 }
 447                 have_zone = 1;
 448         }
 449         spin_unlock(&fs_info->reada_lock);
 450         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
 451
 452         if (!have_zone)
 453                 goto error;
 454
 455         btrfs_put_bbio(bbio);
 456         return re;
 457
 458 error:
 459         for (nzones = 0; nzones < re->nzones; ++nzones) {
 460                 struct reada_zone *zone;
 461
 462                 zone = re->zones[nzones];
 463                 kref_get(&zone->refcnt);
 464                 spin_lock(&zone->lock);
 465                 --zone->elems;
 466                 if (zone->elems == 0) {
 467                         /*
 468                          * no fs_info->reada_lock needed, as this can't be
 469                          * the last ref
 470                          */
 471                         kref_put(&zone->refcnt, reada_zone_release);
 472                 }
 473                 spin_unlock(&zone->lock);
 474
 475                 spin_lock(&fs_info->reada_lock);
 476                 kref_put(&zone->refcnt, reada_zone_release);
 477                 spin_unlock(&fs_info->reada_lock);
 478         }
 479         btrfs_put_bbio(bbio);
 480         kfree(re);
 481         return re_exist;
 482 }
 483
 484 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 485                              struct reada_extent *re)
 486 {
 487         int i;
 488         unsigned long index = re->logical >> PAGE_SHIFT;
 489
 490         spin_lock(&fs_info->reada_lock);
 491         if (--re->refcnt) {
 492                 spin_unlock(&fs_info->reada_lock);
 493                 return;
 494         }
 495
 496         radix_tree_delete(&fs_info->reada_tree, index);
 497         for (i = 0; i < re->nzones; ++i) {
 498                 struct reada_zone *zone = re->zones[i];
 499
 500                 radix_tree_delete(&zone->device->reada_extents, index);
 501         }
 502
 503         spin_unlock(&fs_info->reada_lock);
 504
 505         for (i = 0; i < re->nzones; ++i) {
 506                 struct reada_zone *zone = re->zones[i];
 507
 508                 kref_get(&zone->refcnt);
 509                 spin_lock(&zone->lock);
 510                 --zone->elems;
 511                 if (zone->elems == 0) {
 512                         /* no fs_info->reada_lock needed, as this can't be
 513                          * the last ref */
 514                         kref_put(&zone->refcnt, reada_zone_release);
 515                 }
 516                 spin_unlock(&zone->lock);
 517
 518                 spin_lock(&fs_info->reada_lock);
 519                 kref_put(&zone->refcnt, reada_zone_release);
 520                 spin_unlock(&fs_info->reada_lock);
 521         }
 522
 523         kfree(re);
 524 }
 525
 526 static void reada_zone_release(struct kref *kref)
 527 {
 528         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 529
 530         radix_tree_delete(&zone->device->reada_zones,
 531                           zone->end >> PAGE_SHIFT);
 532
 533         kfree(zone);
 534 }
 535
 536 static void reada_control_release(struct kref *kref)
 537 {
 538         struct reada_control *rc = container_of(kref, struct reada_control,
 539                                                 refcnt);
 540
 541         kfree(rc);
 542 }
 543
 544 static int reada_add_block(struct reada_control *rc, u64 logical,
 545                            struct btrfs_key *top, u64 generation)
 546 {
 547         struct btrfs_fs_info *fs_info = rc->fs_info;
 548         struct reada_extent *re;
 549         struct reada_extctl *rec;
 550
 551         /* takes one ref */
 552         re = reada_find_extent(fs_info->tree_root, logical, top);
 553         if (!re)
 554                 return -1;
 555
 556         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 557         if (!rec) {
 558                 reada_extent_put(fs_info, re);
 559                 return -ENOMEM;
 560         }
 561
 562         rec->rc = rc;
 563         rec->generation = generation;
 564         atomic_inc(&rc->elems);
 565
 566         spin_lock(&re->lock);
 567         list_add_tail(&rec->list, &re->extctl);
 568         spin_unlock(&re->lock);
 569
 570         /* leave the ref on the extent */
 571
 572         return 0;
 573 }
 574
 575 /*
 576  * called with fs_info->reada_lock held
 577  */
 578 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 579 {
 580         int i;
 581         unsigned long index = zone->end >> PAGE_SHIFT;
 582
 583         for (i = 0; i < zone->ndevs; ++i) {
 584                 struct reada_zone *peer;
 585                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 586                 if (peer && peer->device != zone->device)
 587                         peer->locked = lock;
 588         }
 589 }
 590
 591 /*
 592  * called with fs_info->reada_lock held
 593  */
 594 static int reada_pick_zone(struct btrfs_device *dev)
 595 {
 596         struct reada_zone *top_zone = NULL;
 597         struct reada_zone *top_locked_zone = NULL;
 598         u64 top_elems = 0;
 599         u64 top_locked_elems = 0;
 600         unsigned long index = 0;
 601         int ret;
 602
 603         if (dev->reada_curr_zone) {
 604                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 605                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 606                 dev->reada_curr_zone = NULL;
 607         }
 608         /* pick the zone with the most elements */
 609         while (1) {
 610                 struct reada_zone *zone;
 611
 612                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 613                                              (void **)&zone, index, 1);
 614                 if (ret == 0)
 615                         break;
 616                 index = (zone->end >> PAGE_SHIFT) + 1;
 617                 if (zone->locked) {
 618                         if (zone->elems > top_locked_elems) {
 619                                 top_locked_elems = zone->elems;
 620                                 top_locked_zone = zone;
 621                         }
 622                 } else {
 623                         if (zone->elems > top_elems) {
 624                                 top_elems = zone->elems;
 625                                 top_zone = zone;
 626                         }
 627                 }
 628         }
 629         if (top_zone)
 630                 dev->reada_curr_zone = top_zone;
 631         else if (top_locked_zone)
 632                 dev->reada_curr_zone = top_locked_zone;
 633         else
 634                 return 0;
 635
 636         dev->reada_next = dev->reada_curr_zone->start;
 637         kref_get(&dev->reada_curr_zone->refcnt);
 638         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 639
 640         return 1;
 641 }
 642
 643 static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
 644                                    struct btrfs_device *dev)
 645 {
 646         struct reada_extent *re = NULL;
 647         int mirror_num = 0;
 648         struct extent_buffer *eb = NULL;
 649         u64 logical;
 650         int ret;
 651         int i;
 652
 653         spin_lock(&fs_info->reada_lock);
 654         if (dev->reada_curr_zone == NULL) {
 655                 ret = reada_pick_zone(dev);
 656                 if (!ret) {
 657                         spin_unlock(&fs_info->reada_lock);
 658                         return 0;
 659                 }
 660         }
 661         /*
 662          * FIXME currently we issue the reads one extent at a time. If we have
 663          * a contiguous block of extents, we could also coagulate them or use
 664          * plugging to speed things up
 665          */
 666         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 667                                      dev->reada_next >> PAGE_SHIFT, 1);
 668         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 669                 ret = reada_pick_zone(dev);
 670                 if (!ret) {
 671                         spin_unlock(&fs_info->reada_lock);
 672                         return 0;
 673                 }
 674                 re = NULL;
 675                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 676                                         dev->reada_next >> PAGE_SHIFT, 1);
 677         }
 678         if (ret == 0) {
 679                 spin_unlock(&fs_info->reada_lock);
 680                 return 0;
 681         }
 682         dev->reada_next = re->logical + fs_info->nodesize;
 683         re->refcnt++;
 684
 685         spin_unlock(&fs_info->reada_lock);
 686
 687         spin_lock(&re->lock);
 688         if (re->scheduled || list_empty(&re->extctl)) {
 689                 spin_unlock(&re->lock);
 690                 reada_extent_put(fs_info, re);
 691                 return 0;
 692         }
 693         re->scheduled = 1;
 694         spin_unlock(&re->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
 707         atomic_inc(&dev->reada_in_flight);
 708         ret = reada_tree_block_flagged(fs_info->extent_root, logical,
 709                         mirror_num, &eb);
 710         if (ret)
 711                 __readahead_hook(fs_info, re, NULL, ret);
 712         else if (eb)
 713                 __readahead_hook(fs_info, re, eb, ret);
 714
 715         if (eb)
 716                 free_extent_buffer(eb);
 717
 718         atomic_dec(&dev->reada_in_flight);
 719         reada_extent_put(fs_info, re);
 720
 721         return 1;
 722
 723 }
 724
 725 static void reada_start_machine_worker(struct btrfs_work *work)
 726 {
 727         struct reada_machine_work *rmw;
 728         struct btrfs_fs_info *fs_info;
 729         int old_ioprio;
 730
 731         rmw = container_of(work, struct reada_machine_work, work);
 732         fs_info = rmw->fs_info;
 733
 734         kfree(rmw);
 735
 736         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 737                                        task_nice_ioprio(current));
 738         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 739         __reada_start_machine(fs_info);
 740         set_task_ioprio(current, old_ioprio);
 741
 742         atomic_dec(&fs_info->reada_works_cnt);
 743 }
 744
 745 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 746 {
 747         struct btrfs_device *device;
 748         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 749         u64 enqueued;
 750         u64 total = 0;
 751         int i;
 752
 753         do {
 754                 enqueued = 0;
 755                 mutex_lock(&fs_devices->device_list_mutex);
 756                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 757                         if (atomic_read(&device->reada_in_flight) <
 758                             MAX_IN_FLIGHT)
 759                                 enqueued += reada_start_machine_dev(fs_info,
 760                                                                     device);
 761                 }
 762                 mutex_unlock(&fs_devices->device_list_mutex);
 763                 total += enqueued;
 764         } while (enqueued && total < 10000);
 765
 766         if (enqueued == 0)
 767                 return;
 768
 769         /*
 770          * If everything is already in the cache, this is effectively single
 771          * threaded. To a) not hold the caller for too long and b) to utilize
 772          * more cores, we broke the loop above after 10000 iterations and now
 773          * enqueue to workers to finish it. This will distribute the load to
 774          * the cores.
 775          */
 776         for (i = 0; i < 2; ++i) {
 777                 reada_start_machine(fs_info);
 778                 if (atomic_read(&fs_info->reada_works_cnt) >
 779                     BTRFS_MAX_MIRRORS * 2)
 780                         break;
 781         }
 782 }
 783
 784 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 785 {
 786         struct reada_machine_work *rmw;
 787
 788         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 789         if (!rmw) {
 790                 /* FIXME we cannot handle this properly right now */
 791                 BUG();
 792         }
 793         btrfs_init_work(&rmw->work, btrfs_readahead_helper,
 794                         reada_start_machine_worker, NULL, NULL);
 795         rmw->fs_info = fs_info;
 796
 797         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 798         atomic_inc(&fs_info->reada_works_cnt);
 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                 btrfs_debug(fs_info, "dev %lld has %d in flight", 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                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 824                                     zone->start, zone->end, zone->elems,
 825                                     zone->locked);
 826                         for (j = 0; j < zone->ndevs; ++j) {
 827                                 pr_cont(" %lld",
 828                                         zone->devs[j]->devid);
 829                         }
 830                         if (device->reada_curr_zone == zone)
 831                                 pr_cont(" curr off %llu",
 832                                         device->reada_next - zone->start);
 833                         pr_cont("\n");
 834                         index = (zone->end >> PAGE_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                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 846                                 re->logical, fs_info->nodesize,
 847                                 list_empty(&re->extctl), re->scheduled);
 848
 849                         for (i = 0; i < re->nzones; ++i) {
 850                                 pr_cont(" zone %llu-%llu devs",
 851                                         re->zones[i]->start,
 852                                         re->zones[i]->end);
 853                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 854                                         pr_cont(" %lld",
 855                                                 re->zones[i]->devs[j]->devid);
 856                                 }
 857                         }
 858                         pr_cont("\n");
 859                         index = (re->logical >> PAGE_SHIFT) + 1;
 860                         if (++cnt > 15)
 861                                 break;
 862                 }
 863         }
 864
 865         index = 0;
 866         cnt = 0;
 867         while (all) {
 868                 struct reada_extent *re = NULL;
 869
 870                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 871                                              index, 1);
 872                 if (ret == 0)
 873                         break;
 874                 if (!re->scheduled) {
 875                         index = (re->logical >> PAGE_SHIFT) + 1;
 876                         continue;
 877                 }
 878                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 879                         re->logical, fs_info->nodesize,
 880                         list_empty(&re->extctl), re->scheduled);
 881                 for (i = 0; i < re->nzones; ++i) {
 882                         pr_cont(" zone %llu-%llu devs",
 883                                 re->zones[i]->start,
 884                                 re->zones[i]->end);
 885                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 886                                 pr_cont(" %lld",
 887                                        re->zones[i]->devs[j]->devid);
 888                         }
 889                 }
 890                 pr_cont("\n");
 891                 index = (re->logical >> PAGE_SHIFT) + 1;
 892         }
 893         spin_unlock(&fs_info->reada_lock);
 894 }
 895 #endif
 896
 897 /*
 898  * interface
 899  */
 900 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 901                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 902 {
 903         struct reada_control *rc;
 904         u64 start;
 905         u64 generation;
 906         int ret;
 907         struct extent_buffer *node;
 908         static struct btrfs_key max_key = {
 909                 .objectid = (u64)-1,
 910                 .type = (u8)-1,
 911                 .offset = (u64)-1
 912         };
 913
 914         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 915         if (!rc)
 916                 return ERR_PTR(-ENOMEM);
 917
 918         rc->fs_info = root->fs_info;
 919         rc->key_start = *key_start;
 920         rc->key_end = *key_end;
 921         atomic_set(&rc->elems, 0);
 922         init_waitqueue_head(&rc->wait);
 923         kref_init(&rc->refcnt);
 924         kref_get(&rc->refcnt); /* one ref for having elements */
 925
 926         node = btrfs_root_node(root);
 927         start = node->start;
 928         generation = btrfs_header_generation(node);
 929         free_extent_buffer(node);
 930
 931         ret = reada_add_block(rc, start, &max_key, generation);
 932         if (ret) {
 933                 kfree(rc);
 934                 return ERR_PTR(ret);
 935         }
 936
 937         reada_start_machine(root->fs_info);
 938
 939         return rc;
 940 }
 941
 942 #ifdef DEBUG
 943 int btrfs_reada_wait(void *handle)
 944 {
 945         struct reada_control *rc = handle;
 946         struct btrfs_fs_info *fs_info = rc->fs_info;
 947
 948         while (atomic_read(&rc->elems)) {
 949                 if (!atomic_read(&fs_info->reada_works_cnt))
 950                         reada_start_machine(fs_info);
 951                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 952                                    5 * HZ);
 953                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 954         }
 955
 956         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 957
 958         kref_put(&rc->refcnt, reada_control_release);
 959
 960         return 0;
 961 }
 962 #else
 963 int btrfs_reada_wait(void *handle)
 964 {
 965         struct reada_control *rc = handle;
 966         struct btrfs_fs_info *fs_info = rc->fs_info;
 967
 968         while (atomic_read(&rc->elems)) {
 969                 if (!atomic_read(&fs_info->reada_works_cnt))
 970                         reada_start_machine(fs_info);
 971                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 972                                    (HZ + 9) / 10);
 973         }
 974
 975         kref_put(&rc->refcnt, reada_control_release);
 976
 977         return 0;
 978 }
 979 #endif
 980
 981 void btrfs_reada_detach(void *handle)
 982 {
 983         struct reada_control *rc = handle;
 984
 985         kref_put(&rc->refcnt, reada_control_release);
 986 }