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 #include "block-group.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 = 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->start;
 252         end = start + cache->length - 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         /* Insert extent in reada tree + all per-device trees, all or nothing */
 381         down_read(&fs_info->dev_replace.rwsem);
 382         ret = radix_tree_preload(GFP_KERNEL);
 383         if (ret) {
 384                 up_read(&fs_info->dev_replace.rwsem);
 385                 goto error;
 386         }
 387
 388         spin_lock(&fs_info->reada_lock);
 389         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 390         if (ret == -EEXIST) {
 391                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 392                 re_exist->refcnt++;
 393                 spin_unlock(&fs_info->reada_lock);
 394                 radix_tree_preload_end();
 395                 up_read(&fs_info->dev_replace.rwsem);
 396                 goto error;
 397         }
 398         if (ret) {
 399                 spin_unlock(&fs_info->reada_lock);
 400                 radix_tree_preload_end();
 401                 up_read(&fs_info->dev_replace.rwsem);
 402                 goto error;
 403         }
 404         radix_tree_preload_end();
 405         prev_dev = NULL;
 406         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 407                         &fs_info->dev_replace);
 408         for (nzones = 0; nzones < re->nzones; ++nzones) {
 409                 dev = re->zones[nzones]->device;
 410
 411                 if (dev == prev_dev) {
 412                         /*
 413                          * in case of DUP, just add the first zone. As both
 414                          * are on the same device, there's nothing to gain
 415                          * from adding both.
 416                          * Also, it wouldn't work, as the tree is per device
 417                          * and adding would fail with EEXIST
 418                          */
 419                         continue;
 420                 }
 421                 if (!dev->bdev)
 422                         continue;
 423
 424                 if (dev_replace_is_ongoing &&
 425                     dev == fs_info->dev_replace.tgtdev) {
 426                         /*
 427                          * as this device is selected for reading only as
 428                          * a last resort, skip it for read ahead.
 429                          */
 430                         continue;
 431                 }
 432                 prev_dev = dev;
 433                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 434                 if (ret) {
 435                         while (--nzones >= 0) {
 436                                 dev = re->zones[nzones]->device;
 437                                 BUG_ON(dev == NULL);
 438                                 /* ignore whether the entry was inserted */
 439                                 radix_tree_delete(&dev->reada_extents, index);
 440                         }
 441                         radix_tree_delete(&fs_info->reada_tree, index);
 442                         spin_unlock(&fs_info->reada_lock);
 443                         up_read(&fs_info->dev_replace.rwsem);
 444                         goto error;
 445                 }
 446                 have_zone = 1;
 447         }
 448         if (!have_zone)
 449                 radix_tree_delete(&fs_info->reada_tree, index);
 450         spin_unlock(&fs_info->reada_lock);
 451         up_read(&fs_info->dev_replace.rwsem);
 452
 453         if (!have_zone)
 454                 goto error;
 455
 456         btrfs_put_bbio(bbio);
 457         return re;
 458
 459 error:
 460         for (nzones = 0; nzones < re->nzones; ++nzones) {
 461                 struct reada_zone *zone;
 462
 463                 zone = re->zones[nzones];
 464                 kref_get(&zone->refcnt);
 465                 spin_lock(&zone->lock);
 466                 --zone->elems;
 467                 if (zone->elems == 0) {
 468                         /*
 469                          * no fs_info->reada_lock needed, as this can't be
 470                          * the last ref
 471                          */
 472                         kref_put(&zone->refcnt, reada_zone_release);
 473                 }
 474                 spin_unlock(&zone->lock);
 475
 476                 spin_lock(&fs_info->reada_lock);
 477                 kref_put(&zone->refcnt, reada_zone_release);
 478                 spin_unlock(&fs_info->reada_lock);
 479         }
 480         btrfs_put_bbio(bbio);
 481         kfree(re);
 482         return re_exist;
 483 }
 484
 485 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 486                              struct reada_extent *re)
 487 {
 488         int i;
 489         unsigned long index = re->logical >> PAGE_SHIFT;
 490
 491         spin_lock(&fs_info->reada_lock);
 492         if (--re->refcnt) {
 493                 spin_unlock(&fs_info->reada_lock);
 494                 return;
 495         }
 496
 497         radix_tree_delete(&fs_info->reada_tree, index);
 498         for (i = 0; i < re->nzones; ++i) {
 499                 struct reada_zone *zone = re->zones[i];
 500
 501                 radix_tree_delete(&zone->device->reada_extents, index);
 502         }
 503
 504         spin_unlock(&fs_info->reada_lock);
 505
 506         for (i = 0; i < re->nzones; ++i) {
 507                 struct reada_zone *zone = re->zones[i];
 508
 509                 kref_get(&zone->refcnt);
 510                 spin_lock(&zone->lock);
 511                 --zone->elems;
 512                 if (zone->elems == 0) {
 513                         /* no fs_info->reada_lock needed, as this can't be
 514                          * the last ref */
 515                         kref_put(&zone->refcnt, reada_zone_release);
 516                 }
 517                 spin_unlock(&zone->lock);
 518
 519                 spin_lock(&fs_info->reada_lock);
 520                 kref_put(&zone->refcnt, reada_zone_release);
 521                 spin_unlock(&fs_info->reada_lock);
 522         }
 523
 524         kfree(re);
 525 }
 526
 527 static void reada_zone_release(struct kref *kref)
 528 {
 529         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 530
 531         radix_tree_delete(&zone->device->reada_zones,
 532                           zone->end >> PAGE_SHIFT);
 533
 534         kfree(zone);
 535 }
 536
 537 static void reada_control_release(struct kref *kref)
 538 {
 539         struct reada_control *rc = container_of(kref, struct reada_control,
 540                                                 refcnt);
 541
 542         kfree(rc);
 543 }
 544
 545 static int reada_add_block(struct reada_control *rc, u64 logical,
 546                            struct btrfs_key *top, u64 generation)
 547 {
 548         struct btrfs_fs_info *fs_info = rc->fs_info;
 549         struct reada_extent *re;
 550         struct reada_extctl *rec;
 551
 552         /* takes one ref */
 553         re = reada_find_extent(fs_info, logical, top);
 554         if (!re)
 555                 return -1;
 556
 557         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 558         if (!rec) {
 559                 reada_extent_put(fs_info, re);
 560                 return -ENOMEM;
 561         }
 562
 563         rec->rc = rc;
 564         rec->generation = generation;
 565         atomic_inc(&rc->elems);
 566
 567         spin_lock(&re->lock);
 568         list_add_tail(&rec->list, &re->extctl);
 569         spin_unlock(&re->lock);
 570
 571         /* leave the ref on the extent */
 572
 573         return 0;
 574 }
 575
 576 /*
 577  * called with fs_info->reada_lock held
 578  */
 579 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 580 {
 581         int i;
 582         unsigned long index = zone->end >> PAGE_SHIFT;
 583
 584         for (i = 0; i < zone->ndevs; ++i) {
 585                 struct reada_zone *peer;
 586                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 587                 if (peer && peer->device != zone->device)
 588                         peer->locked = lock;
 589         }
 590 }
 591
 592 /*
 593  * called with fs_info->reada_lock held
 594  */
 595 static int reada_pick_zone(struct btrfs_device *dev)
 596 {
 597         struct reada_zone *top_zone = NULL;
 598         struct reada_zone *top_locked_zone = NULL;
 599         u64 top_elems = 0;
 600         u64 top_locked_elems = 0;
 601         unsigned long index = 0;
 602         int ret;
 603
 604         if (dev->reada_curr_zone) {
 605                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 606                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 607                 dev->reada_curr_zone = NULL;
 608         }
 609         /* pick the zone with the most elements */
 610         while (1) {
 611                 struct reada_zone *zone;
 612
 613                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 614                                              (void **)&zone, index, 1);
 615                 if (ret == 0)
 616                         break;
 617                 index = (zone->end >> PAGE_SHIFT) + 1;
 618                 if (zone->locked) {
 619                         if (zone->elems > top_locked_elems) {
 620                                 top_locked_elems = zone->elems;
 621                                 top_locked_zone = zone;
 622                         }
 623                 } else {
 624                         if (zone->elems > top_elems) {
 625                                 top_elems = zone->elems;
 626                                 top_zone = zone;
 627                         }
 628                 }
 629         }
 630         if (top_zone)
 631                 dev->reada_curr_zone = top_zone;
 632         else if (top_locked_zone)
 633                 dev->reada_curr_zone = top_locked_zone;
 634         else
 635                 return 0;
 636
 637         dev->reada_next = dev->reada_curr_zone->start;
 638         kref_get(&dev->reada_curr_zone->refcnt);
 639         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 640
 641         return 1;
 642 }
 643
 644 static int reada_tree_block_flagged(struct btrfs_fs_info *fs_info, u64 bytenr,
 645                                     int mirror_num, struct extent_buffer **eb)
 646 {
 647         struct extent_buffer *buf = NULL;
 648         int ret;
 649
 650         buf = btrfs_find_create_tree_block(fs_info, bytenr);
 651         if (IS_ERR(buf))
 652                 return 0;
 653
 654         set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags);
 655
 656         ret = read_extent_buffer_pages(buf, WAIT_PAGE_LOCK, mirror_num);
 657         if (ret) {
 658                 free_extent_buffer_stale(buf);
 659                 return ret;
 660         }
 661
 662         if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) {
 663                 free_extent_buffer_stale(buf);
 664                 return -EIO;
 665         } else if (extent_buffer_uptodate(buf)) {
 666                 *eb = buf;
 667         } else {
 668                 free_extent_buffer(buf);
 669         }
 670         return 0;
 671 }
 672
 673 static int reada_start_machine_dev(struct btrfs_device *dev)
 674 {
 675         struct btrfs_fs_info *fs_info = dev->fs_info;
 676         struct reada_extent *re = NULL;
 677         int mirror_num = 0;
 678         struct extent_buffer *eb = NULL;
 679         u64 logical;
 680         int ret;
 681         int i;
 682
 683         spin_lock(&fs_info->reada_lock);
 684         if (dev->reada_curr_zone == NULL) {
 685                 ret = reada_pick_zone(dev);
 686                 if (!ret) {
 687                         spin_unlock(&fs_info->reada_lock);
 688                         return 0;
 689                 }
 690         }
 691         /*
 692          * FIXME currently we issue the reads one extent at a time. If we have
 693          * a contiguous block of extents, we could also coagulate them or use
 694          * plugging to speed things up
 695          */
 696         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 697                                      dev->reada_next >> PAGE_SHIFT, 1);
 698         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 699                 ret = reada_pick_zone(dev);
 700                 if (!ret) {
 701                         spin_unlock(&fs_info->reada_lock);
 702                         return 0;
 703                 }
 704                 re = NULL;
 705                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 706                                         dev->reada_next >> PAGE_SHIFT, 1);
 707         }
 708         if (ret == 0) {
 709                 spin_unlock(&fs_info->reada_lock);
 710                 return 0;
 711         }
 712         dev->reada_next = re->logical + fs_info->nodesize;
 713         re->refcnt++;
 714
 715         spin_unlock(&fs_info->reada_lock);
 716
 717         spin_lock(&re->lock);
 718         if (re->scheduled || list_empty(&re->extctl)) {
 719                 spin_unlock(&re->lock);
 720                 reada_extent_put(fs_info, re);
 721                 return 0;
 722         }
 723         re->scheduled = 1;
 724         spin_unlock(&re->lock);
 725
 726         /*
 727          * find mirror num
 728          */
 729         for (i = 0; i < re->nzones; ++i) {
 730                 if (re->zones[i]->device == dev) {
 731                         mirror_num = i + 1;
 732                         break;
 733                 }
 734         }
 735         logical = re->logical;
 736
 737         atomic_inc(&dev->reada_in_flight);
 738         ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
 739         if (ret)
 740                 __readahead_hook(fs_info, re, NULL, ret);
 741         else if (eb)
 742                 __readahead_hook(fs_info, re, eb, ret);
 743
 744         if (eb)
 745                 free_extent_buffer(eb);
 746
 747         atomic_dec(&dev->reada_in_flight);
 748         reada_extent_put(fs_info, re);
 749
 750         return 1;
 751
 752 }
 753
 754 static void reada_start_machine_worker(struct btrfs_work *work)
 755 {
 756         struct reada_machine_work *rmw;
 757         int old_ioprio;
 758
 759         rmw = container_of(work, struct reada_machine_work, work);
 760
 761         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 762                                        task_nice_ioprio(current));
 763         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 764         __reada_start_machine(rmw->fs_info);
 765         set_task_ioprio(current, old_ioprio);
 766
 767         atomic_dec(&rmw->fs_info->reada_works_cnt);
 768
 769         kfree(rmw);
 770 }
 771
 772 /* Try to start up to 10k READA requests for a group of devices */
 773 static int reada_start_for_fsdevs(struct btrfs_fs_devices *fs_devices)
 774 {
 775         u64 enqueued;
 776         u64 total = 0;
 777         struct btrfs_device *device;
 778
 779         do {
 780                 enqueued = 0;
 781                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 782                         if (atomic_read(&device->reada_in_flight) <
 783                             MAX_IN_FLIGHT)
 784                                 enqueued += reada_start_machine_dev(device);
 785                 }
 786                 total += enqueued;
 787         } while (enqueued && total < 10000);
 788
 789         return total;
 790 }
 791
 792 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 793 {
 794         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices, *seed_devs;
 795         int i;
 796         u64 enqueued = 0;
 797
 798         mutex_lock(&fs_devices->device_list_mutex);
 799
 800         enqueued += reada_start_for_fsdevs(fs_devices);
 801         list_for_each_entry(seed_devs, &fs_devices->seed_list, seed_list)
 802                 enqueued += reada_start_for_fsdevs(seed_devs);
 803
 804         mutex_unlock(&fs_devices->device_list_mutex);
 805         if (enqueued == 0)
 806                 return;
 807
 808         /*
 809          * If everything is already in the cache, this is effectively single
 810          * threaded. To a) not hold the caller for too long and b) to utilize
 811          * more cores, we broke the loop above after 10000 iterations and now
 812          * enqueue to workers to finish it. This will distribute the load to
 813          * the cores.
 814          */
 815         for (i = 0; i < 2; ++i) {
 816                 reada_start_machine(fs_info);
 817                 if (atomic_read(&fs_info->reada_works_cnt) >
 818                     BTRFS_MAX_MIRRORS * 2)
 819                         break;
 820         }
 821 }
 822
 823 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 824 {
 825         struct reada_machine_work *rmw;
 826
 827         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 828         if (!rmw) {
 829                 /* FIXME we cannot handle this properly right now */
 830                 BUG();
 831         }
 832         btrfs_init_work(&rmw->work, reada_start_machine_worker, NULL, NULL);
 833         rmw->fs_info = fs_info;
 834
 835         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 836         atomic_inc(&fs_info->reada_works_cnt);
 837 }
 838
 839 #ifdef DEBUG
 840 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 841 {
 842         struct btrfs_device *device;
 843         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 844         unsigned long index;
 845         int ret;
 846         int i;
 847         int j;
 848         int cnt;
 849
 850         spin_lock(&fs_info->reada_lock);
 851         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 852                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 853                         atomic_read(&device->reada_in_flight));
 854                 index = 0;
 855                 while (1) {
 856                         struct reada_zone *zone;
 857                         ret = radix_tree_gang_lookup(&device->reada_zones,
 858                                                      (void **)&zone, index, 1);
 859                         if (ret == 0)
 860                                 break;
 861                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 862                                     zone->start, zone->end, zone->elems,
 863                                     zone->locked);
 864                         for (j = 0; j < zone->ndevs; ++j) {
 865                                 pr_cont(" %lld",
 866                                         zone->devs[j]->devid);
 867                         }
 868                         if (device->reada_curr_zone == zone)
 869                                 pr_cont(" curr off %llu",
 870                                         device->reada_next - zone->start);
 871                         pr_cont("\n");
 872                         index = (zone->end >> PAGE_SHIFT) + 1;
 873                 }
 874                 cnt = 0;
 875                 index = 0;
 876                 while (all) {
 877                         struct reada_extent *re = NULL;
 878
 879                         ret = radix_tree_gang_lookup(&device->reada_extents,
 880                                                      (void **)&re, index, 1);
 881                         if (ret == 0)
 882                                 break;
 883                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 884                                 re->logical, fs_info->nodesize,
 885                                 list_empty(&re->extctl), re->scheduled);
 886
 887                         for (i = 0; i < re->nzones; ++i) {
 888                                 pr_cont(" zone %llu-%llu devs",
 889                                         re->zones[i]->start,
 890                                         re->zones[i]->end);
 891                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 892                                         pr_cont(" %lld",
 893                                                 re->zones[i]->devs[j]->devid);
 894                                 }
 895                         }
 896                         pr_cont("\n");
 897                         index = (re->logical >> PAGE_SHIFT) + 1;
 898                         if (++cnt > 15)
 899                                 break;
 900                 }
 901         }
 902
 903         index = 0;
 904         cnt = 0;
 905         while (all) {
 906                 struct reada_extent *re = NULL;
 907
 908                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 909                                              index, 1);
 910                 if (ret == 0)
 911                         break;
 912                 if (!re->scheduled) {
 913                         index = (re->logical >> PAGE_SHIFT) + 1;
 914                         continue;
 915                 }
 916                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 917                         re->logical, fs_info->nodesize,
 918                         list_empty(&re->extctl), re->scheduled);
 919                 for (i = 0; i < re->nzones; ++i) {
 920                         pr_cont(" zone %llu-%llu devs",
 921                                 re->zones[i]->start,
 922                                 re->zones[i]->end);
 923                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 924                                 pr_cont(" %lld",
 925                                        re->zones[i]->devs[j]->devid);
 926                         }
 927                 }
 928                 pr_cont("\n");
 929                 index = (re->logical >> PAGE_SHIFT) + 1;
 930         }
 931         spin_unlock(&fs_info->reada_lock);
 932 }
 933 #endif
 934
 935 /*
 936  * interface
 937  */
 938 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 939                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 940 {
 941         struct reada_control *rc;
 942         u64 start;
 943         u64 generation;
 944         int ret;
 945         struct extent_buffer *node;
 946         static struct btrfs_key max_key = {
 947                 .objectid = (u64)-1,
 948                 .type = (u8)-1,
 949                 .offset = (u64)-1
 950         };
 951
 952         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 953         if (!rc)
 954                 return ERR_PTR(-ENOMEM);
 955
 956         rc->fs_info = root->fs_info;
 957         rc->key_start = *key_start;
 958         rc->key_end = *key_end;
 959         atomic_set(&rc->elems, 0);
 960         init_waitqueue_head(&rc->wait);
 961         kref_init(&rc->refcnt);
 962         kref_get(&rc->refcnt); /* one ref for having elements */
 963
 964         node = btrfs_root_node(root);
 965         start = node->start;
 966         generation = btrfs_header_generation(node);
 967         free_extent_buffer(node);
 968
 969         ret = reada_add_block(rc, start, &max_key, generation);
 970         if (ret) {
 971                 kfree(rc);
 972                 return ERR_PTR(ret);
 973         }
 974
 975         reada_start_machine(root->fs_info);
 976
 977         return rc;
 978 }
 979
 980 #ifdef DEBUG
 981 int btrfs_reada_wait(void *handle)
 982 {
 983         struct reada_control *rc = handle;
 984         struct btrfs_fs_info *fs_info = rc->fs_info;
 985
 986         while (atomic_read(&rc->elems)) {
 987                 if (!atomic_read(&fs_info->reada_works_cnt))
 988                         reada_start_machine(fs_info);
 989                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
 990                                    5 * HZ);
 991                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 992         }
 993
 994         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
 995
 996         kref_put(&rc->refcnt, reada_control_release);
 997
 998         return 0;
 999 }
1000 #else
1001 int btrfs_reada_wait(void *handle)
1002 {
1003         struct reada_control *rc = handle;
1004         struct btrfs_fs_info *fs_info = rc->fs_info;
1005
1006         while (atomic_read(&rc->elems)) {
1007                 if (!atomic_read(&fs_info->reada_works_cnt))
1008                         reada_start_machine(fs_info);
1009                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1010                                    (HZ + 9) / 10);
1011         }
1012
1013         kref_put(&rc->refcnt, reada_control_release);
1014
1015         return 0;
1016 }
1017 #endif
1018
1019 void btrfs_reada_detach(void *handle)
1020 {
1021         struct reada_control *rc = handle;
1022
1023         kref_put(&rc->refcnt, reada_control_release);
1024 }