fs/btrfs/reada.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/sched.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/writeback.h>
   9 #include <linux/blkdev.h>
  10 #include <linux/rbtree.h>
  11 #include <linux/slab.h>
  12 #include <linux/workqueue.h>
  13 #include "ctree.h"
  14 #include "volumes.h"
  15 #include "disk-io.h"
  16 #include "transaction.h"
  17 #include "dev-replace.h"
  18
  19 #undef DEBUG
  20
  21 /*
  22  * This is the implementation for the generic read ahead framework.
  23  *
  24  * To trigger a readahead, btrfs_reada_add must be called. It will start
  25  * a read ahead for the given range [start, end) on tree root. The returned
  26  * handle can either be used to wait on the readahead to finish
  27  * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
  28  *
  29  * The read ahead works as follows:
  30  * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
  31  * reada_start_machine will then search for extents to prefetch and trigger
  32  * some reads. When a read finishes for a node, all contained node/leaf
  33  * pointers that lie in the given range will also be enqueued. The reads will
  34  * be triggered in sequential order, thus giving a big win over a naive
  35  * enumeration. It will also make use of multi-device layouts. Each disk
  36  * will have its on read pointer and all disks will by utilized in parallel.
  37  * Also will no two disks read both sides of a mirror simultaneously, as this
  38  * would waste seeking capacity. Instead both disks will read different parts
  39  * of the filesystem.
  40  * Any number of readaheads can be started in parallel. The read order will be
  41  * determined globally, i.e. 2 parallel readaheads will normally finish faster
  42  * than the 2 started one after another.
  43  */
  44
  45 #define MAX_IN_FLIGHT 6
  46
  47 struct reada_extctl {
  48         struct list_head        list;
  49         struct reada_control    *rc;
  50         u64                     generation;
  51 };
  52
  53 struct reada_extent {
  54         u64                     logical;
  55         struct btrfs_key        top;
  56         struct list_head        extctl;
  57         int                     refcnt;
  58         spinlock_t              lock;
  59         struct reada_zone       *zones[BTRFS_MAX_MIRRORS];
  60         int                     nzones;
  61         int                     scheduled;
  62 };
  63
  64 struct reada_zone {
  65         u64                     start;
  66         u64                     end;
  67         u64                     elems;
  68         struct list_head        list;
  69         spinlock_t              lock;
  70         int                     locked;
  71         struct btrfs_device     *device;
  72         struct btrfs_device     *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
  73                                                            * self */
  74         int                     ndevs;
  75         struct kref             refcnt;
  76 };
  77
  78 struct reada_machine_work {
  79         struct btrfs_work       work;
  80         struct btrfs_fs_info    *fs_info;
  81 };
  82
  83 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
  84 static void reada_control_release(struct kref *kref);
  85 static void reada_zone_release(struct kref *kref);
  86 static void reada_start_machine(struct btrfs_fs_info *fs_info);
  87 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
  88
  89 static int reada_add_block(struct reada_control *rc, u64 logical,
  90                            struct btrfs_key *top, u64 generation);
  91
  92 /* recurses */
  93 /* in case of err, eb might be NULL */
  94 static void __readahead_hook(struct btrfs_fs_info *fs_info,
  95                              struct reada_extent *re, struct extent_buffer *eb,
  96                              int err)
  97 {
  98         int nritems;
  99         int i;
 100         u64 bytenr;
 101         u64 generation;
 102         struct list_head list;
 103
 104         spin_lock(&re->lock);
 105         /*
 106          * just take the full list from the extent. afterwards we
 107          * don't need the lock anymore
 108          */
 109         list_replace_init(&re->extctl, &list);
 110         re->scheduled = 0;
 111         spin_unlock(&re->lock);
 112
 113         /*
 114          * this is the error case, the extent buffer has not been
 115          * read correctly. We won't access anything from it and
 116          * just cleanup our data structures. Effectively this will
 117          * cut the branch below this node from read ahead.
 118          */
 119         if (err)
 120                 goto cleanup;
 121
 122         /*
 123          * FIXME: currently we just set nritems to 0 if this is a leaf,
 124          * effectively ignoring the content. In a next step we could
 125          * trigger more readahead depending from the content, e.g.
 126          * fetch the checksums for the extents in the leaf.
 127          */
 128         if (!btrfs_header_level(eb))
 129                 goto cleanup;
 130
 131         nritems = btrfs_header_nritems(eb);
 132         generation = btrfs_header_generation(eb);
 133         for (i = 0; i < nritems; i++) {
 134                 struct reada_extctl *rec;
 135                 u64 n_gen;
 136                 struct btrfs_key key;
 137                 struct btrfs_key next_key;
 138
 139                 btrfs_node_key_to_cpu(eb, &key, i);
 140                 if (i + 1 < nritems)
 141                         btrfs_node_key_to_cpu(eb, &next_key, i + 1);
 142                 else
 143                         next_key = re->top;
 144                 bytenr = btrfs_node_blockptr(eb, i);
 145                 n_gen = btrfs_node_ptr_generation(eb, i);
 146
 147                 list_for_each_entry(rec, &list, list) {
 148                         struct reada_control *rc = rec->rc;
 149
 150                         /*
 151                          * if the generation doesn't match, just ignore this
 152                          * extctl. This will probably cut off a branch from
 153                          * prefetch. Alternatively one could start a new (sub-)
 154                          * prefetch for this branch, starting again from root.
 155                          * FIXME: move the generation check out of this loop
 156                          */
 157 #ifdef DEBUG
 158                         if (rec->generation != generation) {
 159                                 btrfs_debug(fs_info,
 160                                             "generation mismatch for (%llu,%d,%llu) %llu != %llu",
 161                                             key.objectid, key.type, key.offset,
 162                                             rec->generation, generation);
 163                         }
 164 #endif
 165                         if (rec->generation == generation &&
 166                             btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
 167                             btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
 168                                 reada_add_block(rc, bytenr, &next_key, n_gen);
 169                 }
 170         }
 171
 172 cleanup:
 173         /*
 174          * free extctl records
 175          */
 176         while (!list_empty(&list)) {
 177                 struct reada_control *rc;
 178                 struct reada_extctl *rec;
 179
 180                 rec = list_first_entry(&list, struct reada_extctl, list);
 181                 list_del(&rec->list);
 182                 rc = rec->rc;
 183                 kfree(rec);
 184
 185                 kref_get(&rc->refcnt);
 186                 if (atomic_dec_and_test(&rc->elems)) {
 187                         kref_put(&rc->refcnt, reada_control_release);
 188                         wake_up(&rc->wait);
 189                 }
 190                 kref_put(&rc->refcnt, reada_control_release);
 191
 192                 reada_extent_put(fs_info, re);  /* one ref for each entry */
 193         }
 194
 195         return;
 196 }
 197
 198 int btree_readahead_hook(struct extent_buffer *eb, int err)
 199 {
 200         struct btrfs_fs_info *fs_info = eb->fs_info;
 201         int ret = 0;
 202         struct reada_extent *re;
 203
 204         /* find extent */
 205         spin_lock(&fs_info->reada_lock);
 206         re = radix_tree_lookup(&fs_info->reada_tree,
 207                                eb->start >> PAGE_SHIFT);
 208         if (re)
 209                 re->refcnt++;
 210         spin_unlock(&fs_info->reada_lock);
 211         if (!re) {
 212                 ret = -1;
 213                 goto start_machine;
 214         }
 215
 216         __readahead_hook(fs_info, re, eb, err);
 217         reada_extent_put(fs_info, re);  /* our ref */
 218
 219 start_machine:
 220         reada_start_machine(fs_info);
 221         return ret;
 222 }
 223
 224 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
 225                                           struct btrfs_bio *bbio)
 226 {
 227         struct btrfs_fs_info *fs_info = dev->fs_info;
 228         int ret;
 229         struct reada_zone *zone;
 230         struct btrfs_block_group_cache *cache = NULL;
 231         u64 start;
 232         u64 end;
 233         int i;
 234
 235         zone = NULL;
 236         spin_lock(&fs_info->reada_lock);
 237         ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 238                                      logical >> PAGE_SHIFT, 1);
 239         if (ret == 1 && logical >= zone->start && logical <= zone->end) {
 240                 kref_get(&zone->refcnt);
 241                 spin_unlock(&fs_info->reada_lock);
 242                 return zone;
 243         }
 244
 245         spin_unlock(&fs_info->reada_lock);
 246
 247         cache = btrfs_lookup_block_group(fs_info, logical);
 248         if (!cache)
 249                 return NULL;
 250
 251         start = cache->key.objectid;
 252         end = start + cache->key.offset - 1;
 253         btrfs_put_block_group(cache);
 254
 255         zone = kzalloc(sizeof(*zone), GFP_KERNEL);
 256         if (!zone)
 257                 return NULL;
 258
 259         ret = radix_tree_preload(GFP_KERNEL);
 260         if (ret) {
 261                 kfree(zone);
 262                 return NULL;
 263         }
 264
 265         zone->start = start;
 266         zone->end = end;
 267         INIT_LIST_HEAD(&zone->list);
 268         spin_lock_init(&zone->lock);
 269         zone->locked = 0;
 270         kref_init(&zone->refcnt);
 271         zone->elems = 0;
 272         zone->device = dev; /* our device always sits at index 0 */
 273         for (i = 0; i < bbio->num_stripes; ++i) {
 274                 /* bounds have already been checked */
 275                 zone->devs[i] = bbio->stripes[i].dev;
 276         }
 277         zone->ndevs = bbio->num_stripes;
 278
 279         spin_lock(&fs_info->reada_lock);
 280         ret = radix_tree_insert(&dev->reada_zones,
 281                                 (unsigned long)(zone->end >> PAGE_SHIFT),
 282                                 zone);
 283
 284         if (ret == -EEXIST) {
 285                 kfree(zone);
 286                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 287                                              logical >> PAGE_SHIFT, 1);
 288                 if (ret == 1 && logical >= zone->start && logical <= zone->end)
 289                         kref_get(&zone->refcnt);
 290                 else
 291                         zone = NULL;
 292         }
 293         spin_unlock(&fs_info->reada_lock);
 294         radix_tree_preload_end();
 295
 296         return zone;
 297 }
 298
 299 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
 300                                               u64 logical,
 301                                               struct btrfs_key *top)
 302 {
 303         int ret;
 304         struct reada_extent *re = NULL;
 305         struct reada_extent *re_exist = NULL;
 306         struct btrfs_bio *bbio = NULL;
 307         struct btrfs_device *dev;
 308         struct btrfs_device *prev_dev;
 309         u64 length;
 310         int real_stripes;
 311         int nzones = 0;
 312         unsigned long index = logical >> PAGE_SHIFT;
 313         int dev_replace_is_ongoing;
 314         int have_zone = 0;
 315
 316         spin_lock(&fs_info->reada_lock);
 317         re = radix_tree_lookup(&fs_info->reada_tree, index);
 318         if (re)
 319                 re->refcnt++;
 320         spin_unlock(&fs_info->reada_lock);
 321
 322         if (re)
 323                 return re;
 324
 325         re = kzalloc(sizeof(*re), GFP_KERNEL);
 326         if (!re)
 327                 return NULL;
 328
 329         re->logical = logical;
 330         re->top = *top;
 331         INIT_LIST_HEAD(&re->extctl);
 332         spin_lock_init(&re->lock);
 333         re->refcnt = 1;
 334
 335         /*
 336          * map block
 337          */
 338         length = fs_info->nodesize;
 339         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
 340                         &length, &bbio, 0);
 341         if (ret || !bbio || length < fs_info->nodesize)
 342                 goto error;
 343
 344         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 345                 btrfs_err(fs_info,
 346                            "readahead: more than %d copies not supported",
 347                            BTRFS_MAX_MIRRORS);
 348                 goto error;
 349         }
 350
 351         real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
 352         for (nzones = 0; nzones < real_stripes; ++nzones) {
 353                 struct reada_zone *zone;
 354
 355                 dev = bbio->stripes[nzones].dev;
 356
 357                 /* cannot read ahead on missing device. */
 358                  if (!dev->bdev)
 359                         continue;
 360
 361                 zone = reada_find_zone(dev, logical, bbio);
 362                 if (!zone)
 363                         continue;
 364
 365                 re->zones[re->nzones++] = zone;
 366                 spin_lock(&zone->lock);
 367                 if (!zone->elems)
 368                         kref_get(&zone->refcnt);
 369                 ++zone->elems;
 370                 spin_unlock(&zone->lock);
 371                 spin_lock(&fs_info->reada_lock);
 372                 kref_put(&zone->refcnt, reada_zone_release);
 373                 spin_unlock(&fs_info->reada_lock);
 374         }
 375         if (re->nzones == 0) {
 376                 /* not a single zone found, error and out */
 377                 goto error;
 378         }
 379
 380         ret = radix_tree_preload(GFP_KERNEL);
 381         if (ret)
 382                 goto error;
 383
 384         /* insert extent in reada_tree + all per-device trees, all or nothing */
 385         btrfs_dev_replace_read_lock(&fs_info->dev_replace);
 386         spin_lock(&fs_info->reada_lock);
 387         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 388         if (ret == -EEXIST) {
 389                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 390                 re_exist->refcnt++;
 391                 spin_unlock(&fs_info->reada_lock);
 392                 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
 393                 radix_tree_preload_end();
 394                 goto error;
 395         }
 396         if (ret) {
 397                 spin_unlock(&fs_info->reada_lock);
 398                 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
 399                 radix_tree_preload_end();
 400                 goto error;
 401         }
 402         radix_tree_preload_end();
 403         prev_dev = NULL;
 404         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 405                         &fs_info->dev_replace);
 406         for (nzones = 0; nzones < re->nzones; ++nzones) {
 407                 dev = re->zones[nzones]->device;
 408
 409                 if (dev == prev_dev) {
 410                         /*
 411                          * in case of DUP, just add the first zone. As both
 412                          * are on the same device, there's nothing to gain
 413                          * from adding both.
 414                          * Also, it wouldn't work, as the tree is per device
 415                          * and adding would fail with EEXIST
 416                          */
 417                         continue;
 418                 }
 419                 if (!dev->bdev)
 420                         continue;
 421
 422                 if (dev_replace_is_ongoing &&
 423                     dev == fs_info->dev_replace.tgtdev) {
 424                         /*
 425                          * as this device is selected for reading only as
 426                          * a last resort, skip it for read ahead.
 427                          */
 428                         continue;
 429                 }
 430                 prev_dev = dev;
 431                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 432                 if (ret) {
 433                         while (--nzones >= 0) {
 434                                 dev = re->zones[nzones]->device;
 435                                 BUG_ON(dev == NULL);
 436                                 /* ignore whether the entry was inserted */
 437                                 radix_tree_delete(&dev->reada_extents, index);
 438                         }
 439                         radix_tree_delete(&fs_info->reada_tree, index);
 440                         spin_unlock(&fs_info->reada_lock);
 441                         btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
 442                         goto error;
 443                 }
 444                 have_zone = 1;
 445         }
 446         spin_unlock(&fs_info->reada_lock);
 447         btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
 448
 449         if (!have_zone)
 450                 goto error;
 451
 452         btrfs_put_bbio(bbio);
 453         return re;
 454
 455 error:
 456         for (nzones = 0; nzones < re->nzones; ++nzones) {
 457                 struct reada_zone *zone;
 458
 459                 zone = re->zones[nzones];
 460                 kref_get(&zone->refcnt);
 461                 spin_lock(&zone->lock);
 462                 --zone->elems;
 463                 if (zone->elems == 0) {
 464                         /*
 465                          * no fs_info->reada_lock needed, as this can't be
 466                          * the last ref
 467                          */
 468                         kref_put(&zone->refcnt, reada_zone_release);
 469                 }
 470                 spin_unlock(&zone->lock);
 471
 472                 spin_lock(&fs_info->reada_lock);
 473                 kref_put(&zone->refcnt, reada_zone_release);
 474                 spin_unlock(&fs_info->reada_lock);
 475         }
 476         btrfs_put_bbio(bbio);
 477         kfree(re);
 478         return re_exist;
 479 }
 480
 481 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 482                              struct reada_extent *re)
 483 {
 484         int i;
 485         unsigned long index = re->logical >> PAGE_SHIFT;
 486
 487         spin_lock(&fs_info->reada_lock);
 488         if (--re->refcnt) {
 489                 spin_unlock(&fs_info->reada_lock);
 490                 return;
 491         }
 492
 493         radix_tree_delete(&fs_info->reada_tree, index);
 494         for (i = 0; i < re->nzones; ++i) {
 495                 struct reada_zone *zone = re->zones[i];
 496
 497                 radix_tree_delete(&zone->device->reada_extents, index);
 498         }
 499
 500         spin_unlock(&fs_info->reada_lock);
 501
 502         for (i = 0; i < re->nzones; ++i) {
 503                 struct reada_zone *zone = re->zones[i];
 504
 505                 kref_get(&zone->refcnt);
 506                 spin_lock(&zone->lock);
 507                 --zone->elems;
 508                 if (zone->elems == 0) {
 509                         /* no fs_info->reada_lock needed, as this can't be
 510                          * the last ref */
 511                         kref_put(&zone->refcnt, reada_zone_release);
 512                 }
 513                 spin_unlock(&zone->lock);
 514
 515                 spin_lock(&fs_info->reada_lock);
 516                 kref_put(&zone->refcnt, reada_zone_release);
 517                 spin_unlock(&fs_info->reada_lock);
 518         }
 519
 520         kfree(re);
 521 }
 522
 523 static void reada_zone_release(struct kref *kref)
 524 {
 525         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 526
 527         radix_tree_delete(&zone->device->reada_zones,
 528                           zone->end >> PAGE_SHIFT);
 529
 530         kfree(zone);
 531 }
 532
 533 static void reada_control_release(struct kref *kref)
 534 {
 535         struct reada_control *rc = container_of(kref, struct reada_control,
 536                                                 refcnt);
 537
 538         kfree(rc);
 539 }
 540
 541 static int reada_add_block(struct reada_control *rc, u64 logical,
 542                            struct btrfs_key *top, u64 generation)
 543 {
 544         struct btrfs_fs_info *fs_info = rc->fs_info;
 545         struct reada_extent *re;
 546         struct reada_extctl *rec;
 547
 548         /* takes one ref */
 549         re = reada_find_extent(fs_info, logical, top);
 550         if (!re)
 551                 return -1;
 552
 553         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 554         if (!rec) {
 555                 reada_extent_put(fs_info, re);
 556                 return -ENOMEM;
 557         }
 558
 559         rec->rc = rc;
 560         rec->generation = generation;
 561         atomic_inc(&rc->elems);
 562
 563         spin_lock(&re->lock);
 564         list_add_tail(&rec->list, &re->extctl);
 565         spin_unlock(&re->lock);
 566
 567         /* leave the ref on the extent */
 568
 569         return 0;
 570 }
 571
 572 /*
 573  * called with fs_info->reada_lock held
 574  */
 575 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 576 {
 577         int i;
 578         unsigned long index = zone->end >> PAGE_SHIFT;
 579
 580         for (i = 0; i < zone->ndevs; ++i) {
 581                 struct reada_zone *peer;
 582                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 583                 if (peer && peer->device != zone->device)
 584                         peer->locked = lock;
 585         }
 586 }
 587
 588 /*
 589  * called with fs_info->reada_lock held
 590  */
 591 static int reada_pick_zone(struct btrfs_device *dev)
 592 {
 593         struct reada_zone *top_zone = NULL;
 594         struct reada_zone *top_locked_zone = NULL;
 595         u64 top_elems = 0;
 596         u64 top_locked_elems = 0;
 597         unsigned long index = 0;
 598         int ret;
 599
 600         if (dev->reada_curr_zone) {
 601                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 602                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 603                 dev->reada_curr_zone = NULL;
 604         }
 605         /* pick the zone with the most elements */
 606         while (1) {
 607                 struct reada_zone *zone;
 608
 609                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 610                                              (void **)&zone, index, 1);
 611                 if (ret == 0)
 612                         break;
 613                 index = (zone->end >> PAGE_SHIFT) + 1;
 614                 if (zone->locked) {
 615                         if (zone->elems > top_locked_elems) {
 616                                 top_locked_elems = zone->elems;
 617                                 top_locked_zone = zone;
 618                         }
 619                 } else {
 620                         if (zone->elems > top_elems) {
 621                                 top_elems = zone->elems;
 622                                 top_zone = zone;
 623                         }
 624                 }
 625         }
 626         if (top_zone)
 627                 dev->reada_curr_zone = top_zone;
 628         else if (top_locked_zone)
 629                 dev->reada_curr_zone = top_locked_zone;
 630         else
 631                 return 0;
 632
 633         dev->reada_next = dev->reada_curr_zone->start;
 634         kref_get(&dev->reada_curr_zone->refcnt);
 635         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 636
 637         return 1;
 638 }
 639
 640 static int reada_start_machine_dev(struct btrfs_device *dev)
 641 {
 642         struct btrfs_fs_info *fs_info = dev->fs_info;
 643         struct reada_extent *re = NULL;
 644         int mirror_num = 0;
 645         struct extent_buffer *eb = NULL;
 646         u64 logical;
 647         int ret;
 648         int i;
 649
 650         spin_lock(&fs_info->reada_lock);
 651         if (dev->reada_curr_zone == NULL) {
 652                 ret = reada_pick_zone(dev);
 653                 if (!ret) {
 654                         spin_unlock(&fs_info->reada_lock);
 655                         return 0;
 656                 }
 657         }
 658         /*
 659          * FIXME currently we issue the reads one extent at a time. If we have
 660          * a contiguous block of extents, we could also coagulate them or use
 661          * plugging to speed things up
 662          */
 663         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 664                                      dev->reada_next >> PAGE_SHIFT, 1);
 665         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 666                 ret = reada_pick_zone(dev);
 667                 if (!ret) {
 668                         spin_unlock(&fs_info->reada_lock);
 669                         return 0;
 670                 }
 671                 re = NULL;
 672                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 673                                         dev->reada_next >> PAGE_SHIFT, 1);
 674         }
 675         if (ret == 0) {
 676                 spin_unlock(&fs_info->reada_lock);
 677                 return 0;
 678         }
 679         dev->reada_next = re->logical + fs_info->nodesize;
 680         re->refcnt++;
 681
 682         spin_unlock(&fs_info->reada_lock);
 683
 684         spin_lock(&re->lock);
 685         if (re->scheduled || list_empty(&re->extctl)) {
 686                 spin_unlock(&re->lock);
 687                 reada_extent_put(fs_info, re);
 688                 return 0;
 689         }
 690         re->scheduled = 1;
 691         spin_unlock(&re->lock);
 692
 693         /*
 694          * find mirror num
 695          */
 696         for (i = 0; i < re->nzones; ++i) {
 697                 if (re->zones[i]->device == dev) {
 698                         mirror_num = i + 1;
 699                         break;
 700                 }
 701         }
 702         logical = re->logical;
 703
 704         atomic_inc(&dev->reada_in_flight);
 705         ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
 706         if (ret)
 707                 __readahead_hook(fs_info, re, NULL, ret);
 708         else if (eb)
 709                 __readahead_hook(fs_info, re, eb, ret);
 710
 711         if (eb)
 712                 free_extent_buffer(eb);
 713
 714         atomic_dec(&dev->reada_in_flight);
 715         reada_extent_put(fs_info, re);
 716
 717         return 1;
 718
 719 }
 720
 721 static void reada_start_machine_worker(struct btrfs_work *work)
 722 {
 723         struct reada_machine_work *rmw;
 724         struct btrfs_fs_info *fs_info;
 725         int old_ioprio;
 726
 727         rmw = container_of(work, struct reada_machine_work, work);
 728         fs_info = rmw->fs_info;
 729
 730         kfree(rmw);
 731
 732         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 733                                        task_nice_ioprio(current));
 734         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 735         __reada_start_machine(fs_info);
 736         set_task_ioprio(current, old_ioprio);
 737
 738         atomic_dec(&fs_info->reada_works_cnt);
 739 }
 740
 741 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 742 {
 743         struct btrfs_device *device;
 744         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 745         u64 enqueued;
 746         u64 total = 0;
 747         int i;
 748
 749         do {
 750                 enqueued = 0;
 751                 mutex_lock(&fs_devices->device_list_mutex);
 752                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 753                         if (atomic_read(&device->reada_in_flight) <
 754                             MAX_IN_FLIGHT)
 755                                 enqueued += reada_start_machine_dev(device);
 756                 }
 757                 mutex_unlock(&fs_devices->device_list_mutex);
 758                 total += enqueued;
 759         } while (enqueued && total < 10000);
 760
 761         if (enqueued == 0)
 762                 return;
 763
 764         /*
 765          * If everything is already in the cache, this is effectively single
 766          * threaded. To a) not hold the caller for too long and b) to utilize
 767          * more cores, we broke the loop above after 10000 iterations and now
 768          * enqueue to workers to finish it. This will distribute the load to
 769          * the cores.
 770          */
 771         for (i = 0; i < 2; ++i) {
 772                 reada_start_machine(fs_info);
 773                 if (atomic_read(&fs_info->reada_works_cnt) >
 774                     BTRFS_MAX_MIRRORS * 2)
 775                         break;
 776         }
 777 }
 778
 779 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 780 {
 781         struct reada_machine_work *rmw;
 782
 783         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 784         if (!rmw) {
 785                 /* FIXME we cannot handle this properly right now */
 786                 BUG();
 787         }
 788         btrfs_init_work(&rmw->work, btrfs_readahead_helper,
 789                         reada_start_machine_worker, NULL, NULL);
 790         rmw->fs_info = fs_info;
 791
 792         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 793         atomic_inc(&fs_info->reada_works_cnt);
 794 }
 795
 796 #ifdef DEBUG
 797 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 798 {
 799         struct btrfs_device *device;
 800         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 801         unsigned long index;
 802         int ret;
 803         int i;
 804         int j;
 805         int cnt;
 806
 807         spin_lock(&fs_info->reada_lock);
 808         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 809                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 810                         atomic_read(&device->reada_in_flight));
 811                 index = 0;
 812                 while (1) {
 813                         struct reada_zone *zone;
 814                         ret = radix_tree_gang_lookup(&device->reada_zones,
 815                                                      (void **)&zone, index, 1);
 816                         if (ret == 0)
 817                                 break;
 818                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 819                                     zone->start, zone->end, zone->elems,
 820                                     zone->locked);
 821                         for (j = 0; j < zone->ndevs; ++j) {
 822                                 pr_cont(" %lld",
 823                                         zone->devs[j]->devid);
 824                         }
 825                         if (device->reada_curr_zone == zone)
 826                                 pr_cont(" curr off %llu",
 827                                         device->reada_next - zone->start);
 828                         pr_cont("\n");
 829                         index = (zone->end >> PAGE_SHIFT) + 1;
 830                 }
 831                 cnt = 0;
 832                 index = 0;
 833                 while (all) {
 834                         struct reada_extent *re = NULL;
 835
 836                         ret = radix_tree_gang_lookup(&device->reada_extents,
 837                                                      (void **)&re, index, 1);
 838                         if (ret == 0)
 839                                 break;
 840                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 841                                 re->logical, fs_info->nodesize,
 842                                 list_empty(&re->extctl), re->scheduled);
 843
 844                         for (i = 0; i < re->nzones; ++i) {
 845                                 pr_cont(" zone %llu-%llu devs",
 846                                         re->zones[i]->start,
 847                                         re->zones[i]->end);
 848                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 849                                         pr_cont(" %lld",
 850                                                 re->zones[i]->devs[j]->devid);
 851                                 }
 852                         }
 853                         pr_cont("\n");
 854                         index = (re->logical >> PAGE_SHIFT) + 1;
 855                         if (++cnt > 15)
 856                                 break;
 857                 }
 858         }
 859
 860         index = 0;
 861         cnt = 0;
 862         while (all) {
 863                 struct reada_extent *re = NULL;
 864
 865                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 866                                              index, 1);
 867                 if (ret == 0)
 868                         break;
 869                 if (!re->scheduled) {
 870                         index = (re->logical >> PAGE_SHIFT) + 1;
 871                         continue;
 872                 }
 873                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 874                         re->logical, fs_info->nodesize,
 875                         list_empty(&re->extctl), re->scheduled);
 876                 for (i = 0; i < re->nzones; ++i) {
 877                         pr_cont(" zone %llu-%llu devs",
 878                                 re->zones[i]->start,
 879                                 re->zones[i]->end);
 880                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 881                                 pr_cont(" %lld",
 882                                        re->zones[i]->devs[j]->devid);
 883                         }
 884                 }
 885                 pr_cont("\n");
 886                 index = (re->logical >> PAGE_SHIFT) + 1;
 887         }
 888         spin_unlock(&fs_info->reada_lock);
 889 }
 890 #endif
 891
 892 /*
 893  * interface
 894  */
 895 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 896                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 897 {
 898         struct reada_control *rc;
 899         u64 start;
 900         u64 generation;
 901         int ret;
 902         struct extent_buffer *node;
 903         static struct btrfs_key max_key = {
 904                 .objectid = (u64)-1,
 905                 .type = (u8)-1,
 906                 .offset = (u64)-1
 907         };
 908
 909         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 910         if (!rc)
 911                 return ERR_PTR(-ENOMEM);
 912
 913         rc->fs_info = root->fs_info;
 914         rc->key_start = *key_start;
 915         rc->key_end = *key_end;
 916         atomic_set(&rc->elems, 0);
 917         init_waitqueue_head(&rc->wait);
 918         kref_init(&rc->refcnt);
 919         kref_get(&rc->refcnt); /* one ref for having elements */
 920
 921         node = btrfs_root_node(root);
 922         start = node->start;
 923         generation = btrfs_header_generation(node);
 924         free_extent_buffer(node);
 925
 926         ret = reada_add_block(rc, start, &max_key, generation);
 927         if (ret) {
 928                 kfree(rc);
 929                 return ERR_PTR(ret);
 930         }
 931
 932         reada_start_machine(root->fs_info);
 933
 934         return rc;
 935 }
 936
 937 #ifdef DEBUG
 938 int btrfs_reada_wait(void *handle)
 939 {
 940         struct reada_control *rc = handle;
 941         struct btrfs_fs_info *fs_info = rc->fs_info;
 942
 943         while (atomic_read(&rc->elems)) {
 944                 if (!atomic_read(&fs_info->reada_works_cnt))
 945                         reada_start_machine(fs_info);
 946                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 947                                    5 * HZ);
 948                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 949         }
 950
 951         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 952
 953         kref_put(&rc->refcnt, reada_control_release);
 954
 955         return 0;
 956 }
 957 #else
 958 int btrfs_reada_wait(void *handle)
 959 {
 960         struct reada_control *rc = handle;
 961         struct btrfs_fs_info *fs_info = rc->fs_info;
 962
 963         while (atomic_read(&rc->elems)) {
 964                 if (!atomic_read(&fs_info->reada_works_cnt))
 965                         reada_start_machine(fs_info);
 966                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 967                                    (HZ + 9) / 10);
 968         }
 969
 970         kref_put(&rc->refcnt, reada_control_release);
 971
 972         return 0;
 973 }
 974 #endif
 975
 976 void btrfs_reada_detach(void *handle)
 977 {
 978         struct reada_control *rc = handle;
 979
 980         kref_put(&rc->refcnt, reada_control_release);
 981 }