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/slab.h>
  11 #include <linux/workqueue.h>
  12 #include "ctree.h"
  13 #include "volumes.h"
  14 #include "disk-io.h"
  15 #include "transaction.h"
  16 #include "dev-replace.h"
  17
  18 #undef DEBUG
  19
  20 /*
  21  * This is the implementation for the generic read ahead framework.
  22  *
  23  * To trigger a readahead, btrfs_reada_add must be called. It will start
  24  * a read ahead for the given range [start, end) on tree root. The returned
  25  * handle can either be used to wait on the readahead to finish
  26  * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
  27  *
  28  * The read ahead works as follows:
  29  * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
  30  * reada_start_machine will then search for extents to prefetch and trigger
  31  * some reads. When a read finishes for a node, all contained node/leaf
  32  * pointers that lie in the given range will also be enqueued. The reads will
  33  * be triggered in sequential order, thus giving a big win over a naive
  34  * enumeration. It will also make use of multi-device layouts. Each disk
  35  * will have its on read pointer and all disks will by utilized in parallel.
  36  * Also will no two disks read both sides of a mirror simultaneously, as this
  37  * would waste seeking capacity. Instead both disks will read different parts
  38  * of the filesystem.
  39  * Any number of readaheads can be started in parallel. The read order will be
  40  * determined globally, i.e. 2 parallel readaheads will normally finish faster
  41  * than the 2 started one after another.
  42  */
  43
  44 #define MAX_IN_FLIGHT 6
  45
  46 struct reada_extctl {
  47         struct list_head        list;
  48         struct reada_control    *rc;
  49         u64                     generation;
  50 };
  51
  52 struct reada_extent {
  53         u64                     logical;
  54         struct btrfs_key        top;
  55         struct list_head        extctl;
  56         int                     refcnt;
  57         spinlock_t              lock;
  58         struct reada_zone       *zones[BTRFS_MAX_MIRRORS];
  59         int                     nzones;
  60         int                     scheduled;
  61 };
  62
  63 struct reada_zone {
  64         u64                     start;
  65         u64                     end;
  66         u64                     elems;
  67         struct list_head        list;
  68         spinlock_t              lock;
  69         int                     locked;
  70         struct btrfs_device     *device;
  71         struct btrfs_device     *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
  72                                                            * self */
  73         int                     ndevs;
  74         struct kref             refcnt;
  75 };
  76
  77 struct reada_machine_work {
  78         struct btrfs_work       work;
  79         struct btrfs_fs_info    *fs_info;
  80 };
  81
  82 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
  83 static void reada_control_release(struct kref *kref);
  84 static void reada_zone_release(struct kref *kref);
  85 static void reada_start_machine(struct btrfs_fs_info *fs_info);
  86 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
  87
  88 static int reada_add_block(struct reada_control *rc, u64 logical,
  89                            struct btrfs_key *top, u64 generation);
  90
  91 /* recurses */
  92 /* in case of err, eb might be NULL */
  93 static void __readahead_hook(struct btrfs_fs_info *fs_info,
  94                              struct reada_extent *re, struct extent_buffer *eb,
  95                              int err)
  96 {
  97         int nritems;
  98         int i;
  99         u64 bytenr;
 100         u64 generation;
 101         struct list_head list;
 102
 103         spin_lock(&re->lock);
 104         /*
 105          * just take the full list from the extent. afterwards we
 106          * don't need the lock anymore
 107          */
 108         list_replace_init(&re->extctl, &list);
 109         re->scheduled = 0;
 110         spin_unlock(&re->lock);
 111
 112         /*
 113          * this is the error case, the extent buffer has not been
 114          * read correctly. We won't access anything from it and
 115          * just cleanup our data structures. Effectively this will
 116          * cut the branch below this node from read ahead.
 117          */
 118         if (err)
 119                 goto cleanup;
 120
 121         /*
 122          * FIXME: currently we just set nritems to 0 if this is a leaf,
 123          * effectively ignoring the content. In a next step we could
 124          * trigger more readahead depending from the content, e.g.
 125          * fetch the checksums for the extents in the leaf.
 126          */
 127         if (!btrfs_header_level(eb))
 128                 goto cleanup;
 129
 130         nritems = btrfs_header_nritems(eb);
 131         generation = btrfs_header_generation(eb);
 132         for (i = 0; i < nritems; i++) {
 133                 struct reada_extctl *rec;
 134                 u64 n_gen;
 135                 struct btrfs_key key;
 136                 struct btrfs_key next_key;
 137
 138                 btrfs_node_key_to_cpu(eb, &key, i);
 139                 if (i + 1 < nritems)
 140                         btrfs_node_key_to_cpu(eb, &next_key, i + 1);
 141                 else
 142                         next_key = re->top;
 143                 bytenr = btrfs_node_blockptr(eb, i);
 144                 n_gen = btrfs_node_ptr_generation(eb, i);
 145
 146                 list_for_each_entry(rec, &list, list) {
 147                         struct reada_control *rc = rec->rc;
 148
 149                         /*
 150                          * if the generation doesn't match, just ignore this
 151                          * extctl. This will probably cut off a branch from
 152                          * prefetch. Alternatively one could start a new (sub-)
 153                          * prefetch for this branch, starting again from root.
 154                          * FIXME: move the generation check out of this loop
 155                          */
 156 #ifdef DEBUG
 157                         if (rec->generation != generation) {
 158                                 btrfs_debug(fs_info,
 159                                             "generation mismatch for (%llu,%d,%llu) %llu != %llu",
 160                                             key.objectid, key.type, key.offset,
 161                                             rec->generation, generation);
 162                         }
 163 #endif
 164                         if (rec->generation == generation &&
 165                             btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
 166                             btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
 167                                 reada_add_block(rc, bytenr, &next_key, n_gen);
 168                 }
 169         }
 170
 171 cleanup:
 172         /*
 173          * free extctl records
 174          */
 175         while (!list_empty(&list)) {
 176                 struct reada_control *rc;
 177                 struct reada_extctl *rec;
 178
 179                 rec = list_first_entry(&list, struct reada_extctl, list);
 180                 list_del(&rec->list);
 181                 rc = rec->rc;
 182                 kfree(rec);
 183
 184                 kref_get(&rc->refcnt);
 185                 if (atomic_dec_and_test(&rc->elems)) {
 186                         kref_put(&rc->refcnt, reada_control_release);
 187                         wake_up(&rc->wait);
 188                 }
 189                 kref_put(&rc->refcnt, reada_control_release);
 190
 191                 reada_extent_put(fs_info, re);  /* one ref for each entry */
 192         }
 193
 194         return;
 195 }
 196
 197 int btree_readahead_hook(struct extent_buffer *eb, int err)
 198 {
 199         struct btrfs_fs_info *fs_info = eb->fs_info;
 200         int ret = 0;
 201         struct reada_extent *re;
 202
 203         /* find extent */
 204         spin_lock(&fs_info->reada_lock);
 205         re = radix_tree_lookup(&fs_info->reada_tree,
 206                                eb->start >> PAGE_SHIFT);
 207         if (re)
 208                 re->refcnt++;
 209         spin_unlock(&fs_info->reada_lock);
 210         if (!re) {
 211                 ret = -1;
 212                 goto start_machine;
 213         }
 214
 215         __readahead_hook(fs_info, re, eb, err);
 216         reada_extent_put(fs_info, re);  /* our ref */
 217
 218 start_machine:
 219         reada_start_machine(fs_info);
 220         return ret;
 221 }
 222
 223 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
 224                                           struct btrfs_bio *bbio)
 225 {
 226         struct btrfs_fs_info *fs_info = dev->fs_info;
 227         int ret;
 228         struct reada_zone *zone;
 229         struct btrfs_block_group_cache *cache = NULL;
 230         u64 start;
 231         u64 end;
 232         int i;
 233
 234         zone = NULL;
 235         spin_lock(&fs_info->reada_lock);
 236         ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 237                                      logical >> PAGE_SHIFT, 1);
 238         if (ret == 1 && logical >= zone->start && logical <= zone->end) {
 239                 kref_get(&zone->refcnt);
 240                 spin_unlock(&fs_info->reada_lock);
 241                 return zone;
 242         }
 243
 244         spin_unlock(&fs_info->reada_lock);
 245
 246         cache = btrfs_lookup_block_group(fs_info, logical);
 247         if (!cache)
 248                 return NULL;
 249
 250         start = cache->key.objectid;
 251         end = start + cache->key.offset - 1;
 252         btrfs_put_block_group(cache);
 253
 254         zone = kzalloc(sizeof(*zone), GFP_KERNEL);
 255         if (!zone)
 256                 return NULL;
 257
 258         ret = radix_tree_preload(GFP_KERNEL);
 259         if (ret) {
 260                 kfree(zone);
 261                 return NULL;
 262         }
 263
 264         zone->start = start;
 265         zone->end = end;
 266         INIT_LIST_HEAD(&zone->list);
 267         spin_lock_init(&zone->lock);
 268         zone->locked = 0;
 269         kref_init(&zone->refcnt);
 270         zone->elems = 0;
 271         zone->device = dev; /* our device always sits at index 0 */
 272         for (i = 0; i < bbio->num_stripes; ++i) {
 273                 /* bounds have already been checked */
 274                 zone->devs[i] = bbio->stripes[i].dev;
 275         }
 276         zone->ndevs = bbio->num_stripes;
 277
 278         spin_lock(&fs_info->reada_lock);
 279         ret = radix_tree_insert(&dev->reada_zones,
 280                                 (unsigned long)(zone->end >> PAGE_SHIFT),
 281                                 zone);
 282
 283         if (ret == -EEXIST) {
 284                 kfree(zone);
 285                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 286                                              logical >> PAGE_SHIFT, 1);
 287                 if (ret == 1 && logical >= zone->start && logical <= zone->end)
 288                         kref_get(&zone->refcnt);
 289                 else
 290                         zone = NULL;
 291         }
 292         spin_unlock(&fs_info->reada_lock);
 293         radix_tree_preload_end();
 294
 295         return zone;
 296 }
 297
 298 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
 299                                               u64 logical,
 300                                               struct btrfs_key *top)
 301 {
 302         int ret;
 303         struct reada_extent *re = NULL;
 304         struct reada_extent *re_exist = NULL;
 305         struct btrfs_bio *bbio = NULL;
 306         struct btrfs_device *dev;
 307         struct btrfs_device *prev_dev;
 308         u64 length;
 309         int real_stripes;
 310         int nzones = 0;
 311         unsigned long index = logical >> PAGE_SHIFT;
 312         int dev_replace_is_ongoing;
 313         int have_zone = 0;
 314
 315         spin_lock(&fs_info->reada_lock);
 316         re = radix_tree_lookup(&fs_info->reada_tree, index);
 317         if (re)
 318                 re->refcnt++;
 319         spin_unlock(&fs_info->reada_lock);
 320
 321         if (re)
 322                 return re;
 323
 324         re = kzalloc(sizeof(*re), GFP_KERNEL);
 325         if (!re)
 326                 return NULL;
 327
 328         re->logical = logical;
 329         re->top = *top;
 330         INIT_LIST_HEAD(&re->extctl);
 331         spin_lock_init(&re->lock);
 332         re->refcnt = 1;
 333
 334         /*
 335          * map block
 336          */
 337         length = fs_info->nodesize;
 338         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
 339                         &length, &bbio, 0);
 340         if (ret || !bbio || length < fs_info->nodesize)
 341                 goto error;
 342
 343         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 344                 btrfs_err(fs_info,
 345                            "readahead: more than %d copies not supported",
 346                            BTRFS_MAX_MIRRORS);
 347                 goto error;
 348         }
 349
 350         real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
 351         for (nzones = 0; nzones < real_stripes; ++nzones) {
 352                 struct reada_zone *zone;
 353
 354                 dev = bbio->stripes[nzones].dev;
 355
 356                 /* cannot read ahead on missing device. */
 357                 if (!dev->bdev)
 358                         continue;
 359
 360                 zone = reada_find_zone(dev, logical, bbio);
 361                 if (!zone)
 362                         continue;
 363
 364                 re->zones[re->nzones++] = zone;
 365                 spin_lock(&zone->lock);
 366                 if (!zone->elems)
 367                         kref_get(&zone->refcnt);
 368                 ++zone->elems;
 369                 spin_unlock(&zone->lock);
 370                 spin_lock(&fs_info->reada_lock);
 371                 kref_put(&zone->refcnt, reada_zone_release);
 372                 spin_unlock(&fs_info->reada_lock);
 373         }
 374         if (re->nzones == 0) {
 375                 /* not a single zone found, error and out */
 376                 goto error;
 377         }
 378
 379         /* Insert extent in reada tree + all per-device trees, all or nothing */
 380         down_read(&fs_info->dev_replace.rwsem);
 381         ret = radix_tree_preload(GFP_KERNEL);
 382         if (ret) {
 383                 up_read(&fs_info->dev_replace.rwsem);
 384                 goto error;
 385         }
 386
 387         spin_lock(&fs_info->reada_lock);
 388         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 389         if (ret == -EEXIST) {
 390                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 391                 re_exist->refcnt++;
 392                 spin_unlock(&fs_info->reada_lock);
 393                 radix_tree_preload_end();
 394                 up_read(&fs_info->dev_replace.rwsem);
 395                 goto error;
 396         }
 397         if (ret) {
 398                 spin_unlock(&fs_info->reada_lock);
 399                 radix_tree_preload_end();
 400                 up_read(&fs_info->dev_replace.rwsem);
 401                 goto error;
 402         }
 403         radix_tree_preload_end();
 404         prev_dev = NULL;
 405         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 406                         &fs_info->dev_replace);
 407         for (nzones = 0; nzones < re->nzones; ++nzones) {
 408                 dev = re->zones[nzones]->device;
 409
 410                 if (dev == prev_dev) {
 411                         /*
 412                          * in case of DUP, just add the first zone. As both
 413                          * are on the same device, there's nothing to gain
 414                          * from adding both.
 415                          * Also, it wouldn't work, as the tree is per device
 416                          * and adding would fail with EEXIST
 417                          */
 418                         continue;
 419                 }
 420                 if (!dev->bdev)
 421                         continue;
 422
 423                 if (dev_replace_is_ongoing &&
 424                     dev == fs_info->dev_replace.tgtdev) {
 425                         /*
 426                          * as this device is selected for reading only as
 427                          * a last resort, skip it for read ahead.
 428                          */
 429                         continue;
 430                 }
 431                 prev_dev = dev;
 432                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 433                 if (ret) {
 434                         while (--nzones >= 0) {
 435                                 dev = re->zones[nzones]->device;
 436                                 BUG_ON(dev == NULL);
 437                                 /* ignore whether the entry was inserted */
 438                                 radix_tree_delete(&dev->reada_extents, index);
 439                         }
 440                         radix_tree_delete(&fs_info->reada_tree, index);
 441                         spin_unlock(&fs_info->reada_lock);
 442                         up_read(&fs_info->dev_replace.rwsem);
 443                         goto error;
 444                 }
 445                 have_zone = 1;
 446         }
 447         spin_unlock(&fs_info->reada_lock);
 448         up_read(&fs_info->dev_replace.rwsem);
 449
 450         if (!have_zone)
 451                 goto error;
 452
 453         btrfs_put_bbio(bbio);
 454         return re;
 455
 456 error:
 457         for (nzones = 0; nzones < re->nzones; ++nzones) {
 458                 struct reada_zone *zone;
 459
 460                 zone = re->zones[nzones];
 461                 kref_get(&zone->refcnt);
 462                 spin_lock(&zone->lock);
 463                 --zone->elems;
 464                 if (zone->elems == 0) {
 465                         /*
 466                          * no fs_info->reada_lock needed, as this can't be
 467                          * the last ref
 468                          */
 469                         kref_put(&zone->refcnt, reada_zone_release);
 470                 }
 471                 spin_unlock(&zone->lock);
 472
 473                 spin_lock(&fs_info->reada_lock);
 474                 kref_put(&zone->refcnt, reada_zone_release);
 475                 spin_unlock(&fs_info->reada_lock);
 476         }
 477         btrfs_put_bbio(bbio);
 478         kfree(re);
 479         return re_exist;
 480 }
 481
 482 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 483                              struct reada_extent *re)
 484 {
 485         int i;
 486         unsigned long index = re->logical >> PAGE_SHIFT;
 487
 488         spin_lock(&fs_info->reada_lock);
 489         if (--re->refcnt) {
 490                 spin_unlock(&fs_info->reada_lock);
 491                 return;
 492         }
 493
 494         radix_tree_delete(&fs_info->reada_tree, index);
 495         for (i = 0; i < re->nzones; ++i) {
 496                 struct reada_zone *zone = re->zones[i];
 497
 498                 radix_tree_delete(&zone->device->reada_extents, index);
 499         }
 500
 501         spin_unlock(&fs_info->reada_lock);
 502
 503         for (i = 0; i < re->nzones; ++i) {
 504                 struct reada_zone *zone = re->zones[i];
 505
 506                 kref_get(&zone->refcnt);
 507                 spin_lock(&zone->lock);
 508                 --zone->elems;
 509                 if (zone->elems == 0) {
 510                         /* no fs_info->reada_lock needed, as this can't be
 511                          * the last ref */
 512                         kref_put(&zone->refcnt, reada_zone_release);
 513                 }
 514                 spin_unlock(&zone->lock);
 515
 516                 spin_lock(&fs_info->reada_lock);
 517                 kref_put(&zone->refcnt, reada_zone_release);
 518                 spin_unlock(&fs_info->reada_lock);
 519         }
 520
 521         kfree(re);
 522 }
 523
 524 static void reada_zone_release(struct kref *kref)
 525 {
 526         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 527
 528         radix_tree_delete(&zone->device->reada_zones,
 529                           zone->end >> PAGE_SHIFT);
 530
 531         kfree(zone);
 532 }
 533
 534 static void reada_control_release(struct kref *kref)
 535 {
 536         struct reada_control *rc = container_of(kref, struct reada_control,
 537                                                 refcnt);
 538
 539         kfree(rc);
 540 }
 541
 542 static int reada_add_block(struct reada_control *rc, u64 logical,
 543                            struct btrfs_key *top, u64 generation)
 544 {
 545         struct btrfs_fs_info *fs_info = rc->fs_info;
 546         struct reada_extent *re;
 547         struct reada_extctl *rec;
 548
 549         /* takes one ref */
 550         re = reada_find_extent(fs_info, logical, top);
 551         if (!re)
 552                 return -1;
 553
 554         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 555         if (!rec) {
 556                 reada_extent_put(fs_info, re);
 557                 return -ENOMEM;
 558         }
 559
 560         rec->rc = rc;
 561         rec->generation = generation;
 562         atomic_inc(&rc->elems);
 563
 564         spin_lock(&re->lock);
 565         list_add_tail(&rec->list, &re->extctl);
 566         spin_unlock(&re->lock);
 567
 568         /* leave the ref on the extent */
 569
 570         return 0;
 571 }
 572
 573 /*
 574  * called with fs_info->reada_lock held
 575  */
 576 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 577 {
 578         int i;
 579         unsigned long index = zone->end >> PAGE_SHIFT;
 580
 581         for (i = 0; i < zone->ndevs; ++i) {
 582                 struct reada_zone *peer;
 583                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 584                 if (peer && peer->device != zone->device)
 585                         peer->locked = lock;
 586         }
 587 }
 588
 589 /*
 590  * called with fs_info->reada_lock held
 591  */
 592 static int reada_pick_zone(struct btrfs_device *dev)
 593 {
 594         struct reada_zone *top_zone = NULL;
 595         struct reada_zone *top_locked_zone = NULL;
 596         u64 top_elems = 0;
 597         u64 top_locked_elems = 0;
 598         unsigned long index = 0;
 599         int ret;
 600
 601         if (dev->reada_curr_zone) {
 602                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 603                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 604                 dev->reada_curr_zone = NULL;
 605         }
 606         /* pick the zone with the most elements */
 607         while (1) {
 608                 struct reada_zone *zone;
 609
 610                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 611                                              (void **)&zone, index, 1);
 612                 if (ret == 0)
 613                         break;
 614                 index = (zone->end >> PAGE_SHIFT) + 1;
 615                 if (zone->locked) {
 616                         if (zone->elems > top_locked_elems) {
 617                                 top_locked_elems = zone->elems;
 618                                 top_locked_zone = zone;
 619                         }
 620                 } else {
 621                         if (zone->elems > top_elems) {
 622                                 top_elems = zone->elems;
 623                                 top_zone = zone;
 624                         }
 625                 }
 626         }
 627         if (top_zone)
 628                 dev->reada_curr_zone = top_zone;
 629         else if (top_locked_zone)
 630                 dev->reada_curr_zone = top_locked_zone;
 631         else
 632                 return 0;
 633
 634         dev->reada_next = dev->reada_curr_zone->start;
 635         kref_get(&dev->reada_curr_zone->refcnt);
 636         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 637
 638         return 1;
 639 }
 640
 641 static int reada_start_machine_dev(struct btrfs_device *dev)
 642 {
 643         struct btrfs_fs_info *fs_info = dev->fs_info;
 644         struct reada_extent *re = NULL;
 645         int mirror_num = 0;
 646         struct extent_buffer *eb = NULL;
 647         u64 logical;
 648         int ret;
 649         int i;
 650
 651         spin_lock(&fs_info->reada_lock);
 652         if (dev->reada_curr_zone == NULL) {
 653                 ret = reada_pick_zone(dev);
 654                 if (!ret) {
 655                         spin_unlock(&fs_info->reada_lock);
 656                         return 0;
 657                 }
 658         }
 659         /*
 660          * FIXME currently we issue the reads one extent at a time. If we have
 661          * a contiguous block of extents, we could also coagulate them or use
 662          * plugging to speed things up
 663          */
 664         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 665                                      dev->reada_next >> PAGE_SHIFT, 1);
 666         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 667                 ret = reada_pick_zone(dev);
 668                 if (!ret) {
 669                         spin_unlock(&fs_info->reada_lock);
 670                         return 0;
 671                 }
 672                 re = NULL;
 673                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 674                                         dev->reada_next >> PAGE_SHIFT, 1);
 675         }
 676         if (ret == 0) {
 677                 spin_unlock(&fs_info->reada_lock);
 678                 return 0;
 679         }
 680         dev->reada_next = re->logical + fs_info->nodesize;
 681         re->refcnt++;
 682
 683         spin_unlock(&fs_info->reada_lock);
 684
 685         spin_lock(&re->lock);
 686         if (re->scheduled || list_empty(&re->extctl)) {
 687                 spin_unlock(&re->lock);
 688                 reada_extent_put(fs_info, re);
 689                 return 0;
 690         }
 691         re->scheduled = 1;
 692         spin_unlock(&re->lock);
 693
 694         /*
 695          * find mirror num
 696          */
 697         for (i = 0; i < re->nzones; ++i) {
 698                 if (re->zones[i]->device == dev) {
 699                         mirror_num = i + 1;
 700                         break;
 701                 }
 702         }
 703         logical = re->logical;
 704
 705         atomic_inc(&dev->reada_in_flight);
 706         ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
 707         if (ret)
 708                 __readahead_hook(fs_info, re, NULL, ret);
 709         else if (eb)
 710                 __readahead_hook(fs_info, re, eb, ret);
 711
 712         if (eb)
 713                 free_extent_buffer(eb);
 714
 715         atomic_dec(&dev->reada_in_flight);
 716         reada_extent_put(fs_info, re);
 717
 718         return 1;
 719
 720 }
 721
 722 static void reada_start_machine_worker(struct btrfs_work *work)
 723 {
 724         struct reada_machine_work *rmw;
 725         struct btrfs_fs_info *fs_info;
 726         int old_ioprio;
 727
 728         rmw = container_of(work, struct reada_machine_work, work);
 729         fs_info = rmw->fs_info;
 730
 731         kfree(rmw);
 732
 733         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 734                                        task_nice_ioprio(current));
 735         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 736         __reada_start_machine(fs_info);
 737         set_task_ioprio(current, old_ioprio);
 738
 739         atomic_dec(&fs_info->reada_works_cnt);
 740 }
 741
 742 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 743 {
 744         struct btrfs_device *device;
 745         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 746         u64 enqueued;
 747         u64 total = 0;
 748         int i;
 749
 750         do {
 751                 enqueued = 0;
 752                 mutex_lock(&fs_devices->device_list_mutex);
 753                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 754                         if (atomic_read(&device->reada_in_flight) <
 755                             MAX_IN_FLIGHT)
 756                                 enqueued += reada_start_machine_dev(device);
 757                 }
 758                 mutex_unlock(&fs_devices->device_list_mutex);
 759                 total += enqueued;
 760         } while (enqueued && total < 10000);
 761
 762         if (enqueued == 0)
 763                 return;
 764
 765         /*
 766          * If everything is already in the cache, this is effectively single
 767          * threaded. To a) not hold the caller for too long and b) to utilize
 768          * more cores, we broke the loop above after 10000 iterations and now
 769          * enqueue to workers to finish it. This will distribute the load to
 770          * the cores.
 771          */
 772         for (i = 0; i < 2; ++i) {
 773                 reada_start_machine(fs_info);
 774                 if (atomic_read(&fs_info->reada_works_cnt) >
 775                     BTRFS_MAX_MIRRORS * 2)
 776                         break;
 777         }
 778 }
 779
 780 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 781 {
 782         struct reada_machine_work *rmw;
 783
 784         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 785         if (!rmw) {
 786                 /* FIXME we cannot handle this properly right now */
 787                 BUG();
 788         }
 789         btrfs_init_work(&rmw->work, btrfs_readahead_helper,
 790                         reada_start_machine_worker, NULL, NULL);
 791         rmw->fs_info = fs_info;
 792
 793         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 794         atomic_inc(&fs_info->reada_works_cnt);
 795 }
 796
 797 #ifdef DEBUG
 798 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 799 {
 800         struct btrfs_device *device;
 801         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 802         unsigned long index;
 803         int ret;
 804         int i;
 805         int j;
 806         int cnt;
 807
 808         spin_lock(&fs_info->reada_lock);
 809         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 810                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 811                         atomic_read(&device->reada_in_flight));
 812                 index = 0;
 813                 while (1) {
 814                         struct reada_zone *zone;
 815                         ret = radix_tree_gang_lookup(&device->reada_zones,
 816                                                      (void **)&zone, index, 1);
 817                         if (ret == 0)
 818                                 break;
 819                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 820                                     zone->start, zone->end, zone->elems,
 821                                     zone->locked);
 822                         for (j = 0; j < zone->ndevs; ++j) {
 823                                 pr_cont(" %lld",
 824                                         zone->devs[j]->devid);
 825                         }
 826                         if (device->reada_curr_zone == zone)
 827                                 pr_cont(" curr off %llu",
 828                                         device->reada_next - zone->start);
 829                         pr_cont("\n");
 830                         index = (zone->end >> PAGE_SHIFT) + 1;
 831                 }
 832                 cnt = 0;
 833                 index = 0;
 834                 while (all) {
 835                         struct reada_extent *re = NULL;
 836
 837                         ret = radix_tree_gang_lookup(&device->reada_extents,
 838                                                      (void **)&re, index, 1);
 839                         if (ret == 0)
 840                                 break;
 841                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 842                                 re->logical, fs_info->nodesize,
 843                                 list_empty(&re->extctl), re->scheduled);
 844
 845                         for (i = 0; i < re->nzones; ++i) {
 846                                 pr_cont(" zone %llu-%llu devs",
 847                                         re->zones[i]->start,
 848                                         re->zones[i]->end);
 849                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 850                                         pr_cont(" %lld",
 851                                                 re->zones[i]->devs[j]->devid);
 852                                 }
 853                         }
 854                         pr_cont("\n");
 855                         index = (re->logical >> PAGE_SHIFT) + 1;
 856                         if (++cnt > 15)
 857                                 break;
 858                 }
 859         }
 860
 861         index = 0;
 862         cnt = 0;
 863         while (all) {
 864                 struct reada_extent *re = NULL;
 865
 866                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 867                                              index, 1);
 868                 if (ret == 0)
 869                         break;
 870                 if (!re->scheduled) {
 871                         index = (re->logical >> PAGE_SHIFT) + 1;
 872                         continue;
 873                 }
 874                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 875                         re->logical, fs_info->nodesize,
 876                         list_empty(&re->extctl), re->scheduled);
 877                 for (i = 0; i < re->nzones; ++i) {
 878                         pr_cont(" zone %llu-%llu devs",
 879                                 re->zones[i]->start,
 880                                 re->zones[i]->end);
 881                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 882                                 pr_cont(" %lld",
 883                                        re->zones[i]->devs[j]->devid);
 884                         }
 885                 }
 886                 pr_cont("\n");
 887                 index = (re->logical >> PAGE_SHIFT) + 1;
 888         }
 889         spin_unlock(&fs_info->reada_lock);
 890 }
 891 #endif
 892
 893 /*
 894  * interface
 895  */
 896 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 897                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 898 {
 899         struct reada_control *rc;
 900         u64 start;
 901         u64 generation;
 902         int ret;
 903         struct extent_buffer *node;
 904         static struct btrfs_key max_key = {
 905                 .objectid = (u64)-1,
 906                 .type = (u8)-1,
 907                 .offset = (u64)-1
 908         };
 909
 910         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 911         if (!rc)
 912                 return ERR_PTR(-ENOMEM);
 913
 914         rc->fs_info = root->fs_info;
 915         rc->key_start = *key_start;
 916         rc->key_end = *key_end;
 917         atomic_set(&rc->elems, 0);
 918         init_waitqueue_head(&rc->wait);
 919         kref_init(&rc->refcnt);
 920         kref_get(&rc->refcnt); /* one ref for having elements */
 921
 922         node = btrfs_root_node(root);
 923         start = node->start;
 924         generation = btrfs_header_generation(node);
 925         free_extent_buffer(node);
 926
 927         ret = reada_add_block(rc, start, &max_key, generation);
 928         if (ret) {
 929                 kfree(rc);
 930                 return ERR_PTR(ret);
 931         }
 932
 933         reada_start_machine(root->fs_info);
 934
 935         return rc;
 936 }
 937
 938 #ifdef DEBUG
 939 int btrfs_reada_wait(void *handle)
 940 {
 941         struct reada_control *rc = handle;
 942         struct btrfs_fs_info *fs_info = rc->fs_info;
 943
 944         while (atomic_read(&rc->elems)) {
 945                 if (!atomic_read(&fs_info->reada_works_cnt))
 946                         reada_start_machine(fs_info);
 947                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 948                                    5 * HZ);
 949                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 950         }
 951
 952         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 953
 954         kref_put(&rc->refcnt, reada_control_release);
 955
 956         return 0;
 957 }
 958 #else
 959 int btrfs_reada_wait(void *handle)
 960 {
 961         struct reada_control *rc = handle;
 962         struct btrfs_fs_info *fs_info = rc->fs_info;
 963
 964         while (atomic_read(&rc->elems)) {
 965                 if (!atomic_read(&fs_info->reada_works_cnt))
 966                         reada_start_machine(fs_info);
 967                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 968                                    (HZ + 9) / 10);
 969         }
 970
 971         kref_put(&rc->refcnt, reada_control_release);
 972
 973         return 0;
 974 }
 975 #endif
 976
 977 void btrfs_reada_detach(void *handle)
 978 {
 979         struct reada_control *rc = handle;
 980
 981         kref_put(&rc->refcnt, reada_control_release);
 982 }