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         u64                     owner_root;
  56         struct btrfs_key        top;
  57         struct list_head        extctl;
  58         int                     refcnt;
  59         spinlock_t              lock;
  60         struct reada_zone       *zones[BTRFS_MAX_MIRRORS];
  61         int                     nzones;
  62         int                     scheduled;
  63         int                     level;
  64 };
  65
  66 struct reada_zone {
  67         u64                     start;
  68         u64                     end;
  69         u64                     elems;
  70         struct list_head        list;
  71         spinlock_t              lock;
  72         int                     locked;
  73         struct btrfs_device     *device;
  74         struct btrfs_device     *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
  75                                                            * self */
  76         int                     ndevs;
  77         struct kref             refcnt;
  78 };
  79
  80 struct reada_machine_work {
  81         struct btrfs_work       work;
  82         struct btrfs_fs_info    *fs_info;
  83 };
  84
  85 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
  86 static void reada_control_release(struct kref *kref);
  87 static void reada_zone_release(struct kref *kref);
  88 static void reada_start_machine(struct btrfs_fs_info *fs_info);
  89 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
  90
  91 static int reada_add_block(struct reada_control *rc, u64 logical,
  92                            struct btrfs_key *top, u64 owner_root,
  93                            u64 generation, int level);
  94
  95 /* recurses */
  96 /* in case of err, eb might be NULL */
  97 static void __readahead_hook(struct btrfs_fs_info *fs_info,
  98                              struct reada_extent *re, struct extent_buffer *eb,
  99                              int err)
 100 {
 101         int nritems;
 102         int i;
 103         u64 bytenr;
 104         u64 generation;
 105         struct list_head list;
 106
 107         spin_lock(&re->lock);
 108         /*
 109          * just take the full list from the extent. afterwards we
 110          * don't need the lock anymore
 111          */
 112         list_replace_init(&re->extctl, &list);
 113         re->scheduled = 0;
 114         spin_unlock(&re->lock);
 115
 116         /*
 117          * this is the error case, the extent buffer has not been
 118          * read correctly. We won't access anything from it and
 119          * just cleanup our data structures. Effectively this will
 120          * cut the branch below this node from read ahead.
 121          */
 122         if (err)
 123                 goto cleanup;
 124
 125         /*
 126          * FIXME: currently we just set nritems to 0 if this is a leaf,
 127          * effectively ignoring the content. In a next step we could
 128          * trigger more readahead depending from the content, e.g.
 129          * fetch the checksums for the extents in the leaf.
 130          */
 131         if (!btrfs_header_level(eb))
 132                 goto cleanup;
 133
 134         nritems = btrfs_header_nritems(eb);
 135         generation = btrfs_header_generation(eb);
 136         for (i = 0; i < nritems; i++) {
 137                 struct reada_extctl *rec;
 138                 u64 n_gen;
 139                 struct btrfs_key key;
 140                 struct btrfs_key next_key;
 141
 142                 btrfs_node_key_to_cpu(eb, &key, i);
 143                 if (i + 1 < nritems)
 144                         btrfs_node_key_to_cpu(eb, &next_key, i + 1);
 145                 else
 146                         next_key = re->top;
 147                 bytenr = btrfs_node_blockptr(eb, i);
 148                 n_gen = btrfs_node_ptr_generation(eb, i);
 149
 150                 list_for_each_entry(rec, &list, list) {
 151                         struct reada_control *rc = rec->rc;
 152
 153                         /*
 154                          * if the generation doesn't match, just ignore this
 155                          * extctl. This will probably cut off a branch from
 156                          * prefetch. Alternatively one could start a new (sub-)
 157                          * prefetch for this branch, starting again from root.
 158                          * FIXME: move the generation check out of this loop
 159                          */
 160 #ifdef DEBUG
 161                         if (rec->generation != generation) {
 162                                 btrfs_debug(fs_info,
 163                                             "generation mismatch for (%llu,%d,%llu) %llu != %llu",
 164                                             key.objectid, key.type, key.offset,
 165                                             rec->generation, generation);
 166                         }
 167 #endif
 168                         if (rec->generation == generation &&
 169                             btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
 170                             btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
 171                                 reada_add_block(rc, bytenr, &next_key,
 172                                                 btrfs_header_owner(eb), n_gen,
 173                                                 btrfs_header_level(eb) - 1);
 174                 }
 175         }
 176
 177 cleanup:
 178         /*
 179          * free extctl records
 180          */
 181         while (!list_empty(&list)) {
 182                 struct reada_control *rc;
 183                 struct reada_extctl *rec;
 184
 185                 rec = list_first_entry(&list, struct reada_extctl, list);
 186                 list_del(&rec->list);
 187                 rc = rec->rc;
 188                 kfree(rec);
 189
 190                 kref_get(&rc->refcnt);
 191                 if (atomic_dec_and_test(&rc->elems)) {
 192                         kref_put(&rc->refcnt, reada_control_release);
 193                         wake_up(&rc->wait);
 194                 }
 195                 kref_put(&rc->refcnt, reada_control_release);
 196
 197                 reada_extent_put(fs_info, re);  /* one ref for each entry */
 198         }
 199
 200         return;
 201 }
 202
 203 int btree_readahead_hook(struct extent_buffer *eb, int err)
 204 {
 205         struct btrfs_fs_info *fs_info = eb->fs_info;
 206         int ret = 0;
 207         struct reada_extent *re;
 208
 209         /* find extent */
 210         spin_lock(&fs_info->reada_lock);
 211         re = radix_tree_lookup(&fs_info->reada_tree,
 212                                eb->start >> PAGE_SHIFT);
 213         if (re)
 214                 re->refcnt++;
 215         spin_unlock(&fs_info->reada_lock);
 216         if (!re) {
 217                 ret = -1;
 218                 goto start_machine;
 219         }
 220
 221         __readahead_hook(fs_info, re, eb, err);
 222         reada_extent_put(fs_info, re);  /* our ref */
 223
 224 start_machine:
 225         reada_start_machine(fs_info);
 226         return ret;
 227 }
 228
 229 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
 230                                           struct btrfs_bio *bbio)
 231 {
 232         struct btrfs_fs_info *fs_info = dev->fs_info;
 233         int ret;
 234         struct reada_zone *zone;
 235         struct btrfs_block_group *cache = NULL;
 236         u64 start;
 237         u64 end;
 238         int i;
 239
 240         zone = NULL;
 241         spin_lock(&fs_info->reada_lock);
 242         ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 243                                      logical >> PAGE_SHIFT, 1);
 244         if (ret == 1 && logical >= zone->start && logical <= zone->end) {
 245                 kref_get(&zone->refcnt);
 246                 spin_unlock(&fs_info->reada_lock);
 247                 return zone;
 248         }
 249
 250         spin_unlock(&fs_info->reada_lock);
 251
 252         cache = btrfs_lookup_block_group(fs_info, logical);
 253         if (!cache)
 254                 return NULL;
 255
 256         start = cache->start;
 257         end = start + cache->length - 1;
 258         btrfs_put_block_group(cache);
 259
 260         zone = kzalloc(sizeof(*zone), GFP_KERNEL);
 261         if (!zone)
 262                 return NULL;
 263
 264         ret = radix_tree_preload(GFP_KERNEL);
 265         if (ret) {
 266                 kfree(zone);
 267                 return NULL;
 268         }
 269
 270         zone->start = start;
 271         zone->end = end;
 272         INIT_LIST_HEAD(&zone->list);
 273         spin_lock_init(&zone->lock);
 274         zone->locked = 0;
 275         kref_init(&zone->refcnt);
 276         zone->elems = 0;
 277         zone->device = dev; /* our device always sits at index 0 */
 278         for (i = 0; i < bbio->num_stripes; ++i) {
 279                 /* bounds have already been checked */
 280                 zone->devs[i] = bbio->stripes[i].dev;
 281         }
 282         zone->ndevs = bbio->num_stripes;
 283
 284         spin_lock(&fs_info->reada_lock);
 285         ret = radix_tree_insert(&dev->reada_zones,
 286                                 (unsigned long)(zone->end >> PAGE_SHIFT),
 287                                 zone);
 288
 289         if (ret == -EEXIST) {
 290                 kfree(zone);
 291                 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
 292                                              logical >> PAGE_SHIFT, 1);
 293                 if (ret == 1 && logical >= zone->start && logical <= zone->end)
 294                         kref_get(&zone->refcnt);
 295                 else
 296                         zone = NULL;
 297         }
 298         spin_unlock(&fs_info->reada_lock);
 299         radix_tree_preload_end();
 300
 301         return zone;
 302 }
 303
 304 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
 305                                               u64 logical,
 306                                               struct btrfs_key *top,
 307                                               u64 owner_root, int level)
 308 {
 309         int ret;
 310         struct reada_extent *re = NULL;
 311         struct reada_extent *re_exist = NULL;
 312         struct btrfs_bio *bbio = NULL;
 313         struct btrfs_device *dev;
 314         struct btrfs_device *prev_dev;
 315         u64 length;
 316         int real_stripes;
 317         int nzones = 0;
 318         unsigned long index = logical >> PAGE_SHIFT;
 319         int dev_replace_is_ongoing;
 320         int have_zone = 0;
 321
 322         spin_lock(&fs_info->reada_lock);
 323         re = radix_tree_lookup(&fs_info->reada_tree, index);
 324         if (re)
 325                 re->refcnt++;
 326         spin_unlock(&fs_info->reada_lock);
 327
 328         if (re)
 329                 return re;
 330
 331         re = kzalloc(sizeof(*re), GFP_KERNEL);
 332         if (!re)
 333                 return NULL;
 334
 335         re->logical = logical;
 336         re->top = *top;
 337         INIT_LIST_HEAD(&re->extctl);
 338         spin_lock_init(&re->lock);
 339         re->refcnt = 1;
 340         re->owner_root = owner_root;
 341         re->level = level;
 342
 343         /*
 344          * map block
 345          */
 346         length = fs_info->nodesize;
 347         ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
 348                         &length, &bbio, 0);
 349         if (ret || !bbio || length < fs_info->nodesize)
 350                 goto error;
 351
 352         if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
 353                 btrfs_err(fs_info,
 354                            "readahead: more than %d copies not supported",
 355                            BTRFS_MAX_MIRRORS);
 356                 goto error;
 357         }
 358
 359         real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
 360         for (nzones = 0; nzones < real_stripes; ++nzones) {
 361                 struct reada_zone *zone;
 362
 363                 dev = bbio->stripes[nzones].dev;
 364
 365                 /* cannot read ahead on missing device. */
 366                 if (!dev->bdev)
 367                         continue;
 368
 369                 zone = reada_find_zone(dev, logical, bbio);
 370                 if (!zone)
 371                         continue;
 372
 373                 re->zones[re->nzones++] = zone;
 374                 spin_lock(&zone->lock);
 375                 if (!zone->elems)
 376                         kref_get(&zone->refcnt);
 377                 ++zone->elems;
 378                 spin_unlock(&zone->lock);
 379                 spin_lock(&fs_info->reada_lock);
 380                 kref_put(&zone->refcnt, reada_zone_release);
 381                 spin_unlock(&fs_info->reada_lock);
 382         }
 383         if (re->nzones == 0) {
 384                 /* not a single zone found, error and out */
 385                 goto error;
 386         }
 387
 388         /* Insert extent in reada tree + all per-device trees, all or nothing */
 389         down_read(&fs_info->dev_replace.rwsem);
 390         ret = radix_tree_preload(GFP_KERNEL);
 391         if (ret) {
 392                 up_read(&fs_info->dev_replace.rwsem);
 393                 goto error;
 394         }
 395
 396         spin_lock(&fs_info->reada_lock);
 397         ret = radix_tree_insert(&fs_info->reada_tree, index, re);
 398         if (ret == -EEXIST) {
 399                 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
 400                 re_exist->refcnt++;
 401                 spin_unlock(&fs_info->reada_lock);
 402                 radix_tree_preload_end();
 403                 up_read(&fs_info->dev_replace.rwsem);
 404                 goto error;
 405         }
 406         if (ret) {
 407                 spin_unlock(&fs_info->reada_lock);
 408                 radix_tree_preload_end();
 409                 up_read(&fs_info->dev_replace.rwsem);
 410                 goto error;
 411         }
 412         radix_tree_preload_end();
 413         prev_dev = NULL;
 414         dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
 415                         &fs_info->dev_replace);
 416         for (nzones = 0; nzones < re->nzones; ++nzones) {
 417                 dev = re->zones[nzones]->device;
 418
 419                 if (dev == prev_dev) {
 420                         /*
 421                          * in case of DUP, just add the first zone. As both
 422                          * are on the same device, there's nothing to gain
 423                          * from adding both.
 424                          * Also, it wouldn't work, as the tree is per device
 425                          * and adding would fail with EEXIST
 426                          */
 427                         continue;
 428                 }
 429                 if (!dev->bdev)
 430                         continue;
 431
 432                 if (test_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state))
 433                         continue;
 434
 435                 if (dev_replace_is_ongoing &&
 436                     dev == fs_info->dev_replace.tgtdev) {
 437                         /*
 438                          * as this device is selected for reading only as
 439                          * a last resort, skip it for read ahead.
 440                          */
 441                         continue;
 442                 }
 443                 prev_dev = dev;
 444                 ret = radix_tree_insert(&dev->reada_extents, index, re);
 445                 if (ret) {
 446                         while (--nzones >= 0) {
 447                                 dev = re->zones[nzones]->device;
 448                                 BUG_ON(dev == NULL);
 449                                 /* ignore whether the entry was inserted */
 450                                 radix_tree_delete(&dev->reada_extents, index);
 451                         }
 452                         radix_tree_delete(&fs_info->reada_tree, index);
 453                         spin_unlock(&fs_info->reada_lock);
 454                         up_read(&fs_info->dev_replace.rwsem);
 455                         goto error;
 456                 }
 457                 have_zone = 1;
 458         }
 459         if (!have_zone)
 460                 radix_tree_delete(&fs_info->reada_tree, index);
 461         spin_unlock(&fs_info->reada_lock);
 462         up_read(&fs_info->dev_replace.rwsem);
 463
 464         if (!have_zone)
 465                 goto error;
 466
 467         btrfs_put_bbio(bbio);
 468         return re;
 469
 470 error:
 471         for (nzones = 0; nzones < re->nzones; ++nzones) {
 472                 struct reada_zone *zone;
 473
 474                 zone = re->zones[nzones];
 475                 kref_get(&zone->refcnt);
 476                 spin_lock(&zone->lock);
 477                 --zone->elems;
 478                 if (zone->elems == 0) {
 479                         /*
 480                          * no fs_info->reada_lock needed, as this can't be
 481                          * the last ref
 482                          */
 483                         kref_put(&zone->refcnt, reada_zone_release);
 484                 }
 485                 spin_unlock(&zone->lock);
 486
 487                 spin_lock(&fs_info->reada_lock);
 488                 kref_put(&zone->refcnt, reada_zone_release);
 489                 spin_unlock(&fs_info->reada_lock);
 490         }
 491         btrfs_put_bbio(bbio);
 492         kfree(re);
 493         return re_exist;
 494 }
 495
 496 static void reada_extent_put(struct btrfs_fs_info *fs_info,
 497                              struct reada_extent *re)
 498 {
 499         int i;
 500         unsigned long index = re->logical >> PAGE_SHIFT;
 501
 502         spin_lock(&fs_info->reada_lock);
 503         if (--re->refcnt) {
 504                 spin_unlock(&fs_info->reada_lock);
 505                 return;
 506         }
 507
 508         radix_tree_delete(&fs_info->reada_tree, index);
 509         for (i = 0; i < re->nzones; ++i) {
 510                 struct reada_zone *zone = re->zones[i];
 511
 512                 radix_tree_delete(&zone->device->reada_extents, index);
 513         }
 514
 515         spin_unlock(&fs_info->reada_lock);
 516
 517         for (i = 0; i < re->nzones; ++i) {
 518                 struct reada_zone *zone = re->zones[i];
 519
 520                 kref_get(&zone->refcnt);
 521                 spin_lock(&zone->lock);
 522                 --zone->elems;
 523                 if (zone->elems == 0) {
 524                         /* no fs_info->reada_lock needed, as this can't be
 525                          * the last ref */
 526                         kref_put(&zone->refcnt, reada_zone_release);
 527                 }
 528                 spin_unlock(&zone->lock);
 529
 530                 spin_lock(&fs_info->reada_lock);
 531                 kref_put(&zone->refcnt, reada_zone_release);
 532                 spin_unlock(&fs_info->reada_lock);
 533         }
 534
 535         kfree(re);
 536 }
 537
 538 static void reada_zone_release(struct kref *kref)
 539 {
 540         struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
 541
 542         lockdep_assert_held(&zone->device->fs_info->reada_lock);
 543
 544         radix_tree_delete(&zone->device->reada_zones,
 545                           zone->end >> PAGE_SHIFT);
 546
 547         kfree(zone);
 548 }
 549
 550 static void reada_control_release(struct kref *kref)
 551 {
 552         struct reada_control *rc = container_of(kref, struct reada_control,
 553                                                 refcnt);
 554
 555         kfree(rc);
 556 }
 557
 558 static int reada_add_block(struct reada_control *rc, u64 logical,
 559                            struct btrfs_key *top, u64 owner_root,
 560                            u64 generation, int level)
 561 {
 562         struct btrfs_fs_info *fs_info = rc->fs_info;
 563         struct reada_extent *re;
 564         struct reada_extctl *rec;
 565
 566         /* takes one ref */
 567         re = reada_find_extent(fs_info, logical, top, owner_root, level);
 568         if (!re)
 569                 return -1;
 570
 571         rec = kzalloc(sizeof(*rec), GFP_KERNEL);
 572         if (!rec) {
 573                 reada_extent_put(fs_info, re);
 574                 return -ENOMEM;
 575         }
 576
 577         rec->rc = rc;
 578         rec->generation = generation;
 579         atomic_inc(&rc->elems);
 580
 581         spin_lock(&re->lock);
 582         list_add_tail(&rec->list, &re->extctl);
 583         spin_unlock(&re->lock);
 584
 585         /* leave the ref on the extent */
 586
 587         return 0;
 588 }
 589
 590 /*
 591  * called with fs_info->reada_lock held
 592  */
 593 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
 594 {
 595         int i;
 596         unsigned long index = zone->end >> PAGE_SHIFT;
 597
 598         for (i = 0; i < zone->ndevs; ++i) {
 599                 struct reada_zone *peer;
 600                 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
 601                 if (peer && peer->device != zone->device)
 602                         peer->locked = lock;
 603         }
 604 }
 605
 606 /*
 607  * called with fs_info->reada_lock held
 608  */
 609 static int reada_pick_zone(struct btrfs_device *dev)
 610 {
 611         struct reada_zone *top_zone = NULL;
 612         struct reada_zone *top_locked_zone = NULL;
 613         u64 top_elems = 0;
 614         u64 top_locked_elems = 0;
 615         unsigned long index = 0;
 616         int ret;
 617
 618         if (dev->reada_curr_zone) {
 619                 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
 620                 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
 621                 dev->reada_curr_zone = NULL;
 622         }
 623         /* pick the zone with the most elements */
 624         while (1) {
 625                 struct reada_zone *zone;
 626
 627                 ret = radix_tree_gang_lookup(&dev->reada_zones,
 628                                              (void **)&zone, index, 1);
 629                 if (ret == 0)
 630                         break;
 631                 index = (zone->end >> PAGE_SHIFT) + 1;
 632                 if (zone->locked) {
 633                         if (zone->elems > top_locked_elems) {
 634                                 top_locked_elems = zone->elems;
 635                                 top_locked_zone = zone;
 636                         }
 637                 } else {
 638                         if (zone->elems > top_elems) {
 639                                 top_elems = zone->elems;
 640                                 top_zone = zone;
 641                         }
 642                 }
 643         }
 644         if (top_zone)
 645                 dev->reada_curr_zone = top_zone;
 646         else if (top_locked_zone)
 647                 dev->reada_curr_zone = top_locked_zone;
 648         else
 649                 return 0;
 650
 651         dev->reada_next = dev->reada_curr_zone->start;
 652         kref_get(&dev->reada_curr_zone->refcnt);
 653         reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
 654
 655         return 1;
 656 }
 657
 658 static int reada_tree_block_flagged(struct btrfs_fs_info *fs_info, u64 bytenr,
 659                                     u64 owner_root, int level, int mirror_num,
 660                                     struct extent_buffer **eb)
 661 {
 662         struct extent_buffer *buf = NULL;
 663         int ret;
 664
 665         buf = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
 666         if (IS_ERR(buf))
 667                 return 0;
 668
 669         set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags);
 670
 671         ret = read_extent_buffer_pages(buf, WAIT_PAGE_LOCK, mirror_num);
 672         if (ret) {
 673                 free_extent_buffer_stale(buf);
 674                 return ret;
 675         }
 676
 677         if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) {
 678                 free_extent_buffer_stale(buf);
 679                 return -EIO;
 680         } else if (extent_buffer_uptodate(buf)) {
 681                 *eb = buf;
 682         } else {
 683                 free_extent_buffer(buf);
 684         }
 685         return 0;
 686 }
 687
 688 static int reada_start_machine_dev(struct btrfs_device *dev)
 689 {
 690         struct btrfs_fs_info *fs_info = dev->fs_info;
 691         struct reada_extent *re = NULL;
 692         int mirror_num = 0;
 693         struct extent_buffer *eb = NULL;
 694         u64 logical;
 695         int ret;
 696         int i;
 697
 698         spin_lock(&fs_info->reada_lock);
 699         if (dev->reada_curr_zone == NULL) {
 700                 ret = reada_pick_zone(dev);
 701                 if (!ret) {
 702                         spin_unlock(&fs_info->reada_lock);
 703                         return 0;
 704                 }
 705         }
 706         /*
 707          * FIXME currently we issue the reads one extent at a time. If we have
 708          * a contiguous block of extents, we could also coagulate them or use
 709          * plugging to speed things up
 710          */
 711         ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 712                                      dev->reada_next >> PAGE_SHIFT, 1);
 713         if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
 714                 ret = reada_pick_zone(dev);
 715                 if (!ret) {
 716                         spin_unlock(&fs_info->reada_lock);
 717                         return 0;
 718                 }
 719                 re = NULL;
 720                 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
 721                                         dev->reada_next >> PAGE_SHIFT, 1);
 722         }
 723         if (ret == 0) {
 724                 spin_unlock(&fs_info->reada_lock);
 725                 return 0;
 726         }
 727         dev->reada_next = re->logical + fs_info->nodesize;
 728         re->refcnt++;
 729
 730         spin_unlock(&fs_info->reada_lock);
 731
 732         spin_lock(&re->lock);
 733         if (re->scheduled || list_empty(&re->extctl)) {
 734                 spin_unlock(&re->lock);
 735                 reada_extent_put(fs_info, re);
 736                 return 0;
 737         }
 738         re->scheduled = 1;
 739         spin_unlock(&re->lock);
 740
 741         /*
 742          * find mirror num
 743          */
 744         for (i = 0; i < re->nzones; ++i) {
 745                 if (re->zones[i]->device == dev) {
 746                         mirror_num = i + 1;
 747                         break;
 748                 }
 749         }
 750         logical = re->logical;
 751
 752         atomic_inc(&dev->reada_in_flight);
 753         ret = reada_tree_block_flagged(fs_info, logical, re->owner_root,
 754                                        re->level, mirror_num, &eb);
 755         if (ret)
 756                 __readahead_hook(fs_info, re, NULL, ret);
 757         else if (eb)
 758                 __readahead_hook(fs_info, re, eb, ret);
 759
 760         if (eb)
 761                 free_extent_buffer(eb);
 762
 763         atomic_dec(&dev->reada_in_flight);
 764         reada_extent_put(fs_info, re);
 765
 766         return 1;
 767
 768 }
 769
 770 static void reada_start_machine_worker(struct btrfs_work *work)
 771 {
 772         struct reada_machine_work *rmw;
 773         int old_ioprio;
 774
 775         rmw = container_of(work, struct reada_machine_work, work);
 776
 777         old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
 778                                        task_nice_ioprio(current));
 779         set_task_ioprio(current, BTRFS_IOPRIO_READA);
 780         __reada_start_machine(rmw->fs_info);
 781         set_task_ioprio(current, old_ioprio);
 782
 783         atomic_dec(&rmw->fs_info->reada_works_cnt);
 784
 785         kfree(rmw);
 786 }
 787
 788 /* Try to start up to 10k READA requests for a group of devices */
 789 static int reada_start_for_fsdevs(struct btrfs_fs_devices *fs_devices)
 790 {
 791         u64 enqueued;
 792         u64 total = 0;
 793         struct btrfs_device *device;
 794
 795         do {
 796                 enqueued = 0;
 797                 list_for_each_entry(device, &fs_devices->devices, dev_list) {
 798                         if (atomic_read(&device->reada_in_flight) <
 799                             MAX_IN_FLIGHT)
 800                                 enqueued += reada_start_machine_dev(device);
 801                 }
 802                 total += enqueued;
 803         } while (enqueued && total < 10000);
 804
 805         return total;
 806 }
 807
 808 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
 809 {
 810         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices, *seed_devs;
 811         int i;
 812         u64 enqueued = 0;
 813
 814         mutex_lock(&fs_devices->device_list_mutex);
 815
 816         enqueued += reada_start_for_fsdevs(fs_devices);
 817         list_for_each_entry(seed_devs, &fs_devices->seed_list, seed_list)
 818                 enqueued += reada_start_for_fsdevs(seed_devs);
 819
 820         mutex_unlock(&fs_devices->device_list_mutex);
 821         if (enqueued == 0)
 822                 return;
 823
 824         /*
 825          * If everything is already in the cache, this is effectively single
 826          * threaded. To a) not hold the caller for too long and b) to utilize
 827          * more cores, we broke the loop above after 10000 iterations and now
 828          * enqueue to workers to finish it. This will distribute the load to
 829          * the cores.
 830          */
 831         for (i = 0; i < 2; ++i) {
 832                 reada_start_machine(fs_info);
 833                 if (atomic_read(&fs_info->reada_works_cnt) >
 834                     BTRFS_MAX_MIRRORS * 2)
 835                         break;
 836         }
 837 }
 838
 839 static void reada_start_machine(struct btrfs_fs_info *fs_info)
 840 {
 841         struct reada_machine_work *rmw;
 842
 843         rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
 844         if (!rmw) {
 845                 /* FIXME we cannot handle this properly right now */
 846                 BUG();
 847         }
 848         btrfs_init_work(&rmw->work, reada_start_machine_worker, NULL, NULL);
 849         rmw->fs_info = fs_info;
 850
 851         btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
 852         atomic_inc(&fs_info->reada_works_cnt);
 853 }
 854
 855 #ifdef DEBUG
 856 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
 857 {
 858         struct btrfs_device *device;
 859         struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 860         unsigned long index;
 861         int ret;
 862         int i;
 863         int j;
 864         int cnt;
 865
 866         spin_lock(&fs_info->reada_lock);
 867         list_for_each_entry(device, &fs_devices->devices, dev_list) {
 868                 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
 869                         atomic_read(&device->reada_in_flight));
 870                 index = 0;
 871                 while (1) {
 872                         struct reada_zone *zone;
 873                         ret = radix_tree_gang_lookup(&device->reada_zones,
 874                                                      (void **)&zone, index, 1);
 875                         if (ret == 0)
 876                                 break;
 877                         pr_debug("  zone %llu-%llu elems %llu locked %d devs",
 878                                     zone->start, zone->end, zone->elems,
 879                                     zone->locked);
 880                         for (j = 0; j < zone->ndevs; ++j) {
 881                                 pr_cont(" %lld",
 882                                         zone->devs[j]->devid);
 883                         }
 884                         if (device->reada_curr_zone == zone)
 885                                 pr_cont(" curr off %llu",
 886                                         device->reada_next - zone->start);
 887                         pr_cont("\n");
 888                         index = (zone->end >> PAGE_SHIFT) + 1;
 889                 }
 890                 cnt = 0;
 891                 index = 0;
 892                 while (all) {
 893                         struct reada_extent *re = NULL;
 894
 895                         ret = radix_tree_gang_lookup(&device->reada_extents,
 896                                                      (void **)&re, index, 1);
 897                         if (ret == 0)
 898                                 break;
 899                         pr_debug("  re: logical %llu size %u empty %d scheduled %d",
 900                                 re->logical, fs_info->nodesize,
 901                                 list_empty(&re->extctl), re->scheduled);
 902
 903                         for (i = 0; i < re->nzones; ++i) {
 904                                 pr_cont(" zone %llu-%llu devs",
 905                                         re->zones[i]->start,
 906                                         re->zones[i]->end);
 907                                 for (j = 0; j < re->zones[i]->ndevs; ++j) {
 908                                         pr_cont(" %lld",
 909                                                 re->zones[i]->devs[j]->devid);
 910                                 }
 911                         }
 912                         pr_cont("\n");
 913                         index = (re->logical >> PAGE_SHIFT) + 1;
 914                         if (++cnt > 15)
 915                                 break;
 916                 }
 917         }
 918
 919         index = 0;
 920         cnt = 0;
 921         while (all) {
 922                 struct reada_extent *re = NULL;
 923
 924                 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
 925                                              index, 1);
 926                 if (ret == 0)
 927                         break;
 928                 if (!re->scheduled) {
 929                         index = (re->logical >> PAGE_SHIFT) + 1;
 930                         continue;
 931                 }
 932                 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
 933                         re->logical, fs_info->nodesize,
 934                         list_empty(&re->extctl), re->scheduled);
 935                 for (i = 0; i < re->nzones; ++i) {
 936                         pr_cont(" zone %llu-%llu devs",
 937                                 re->zones[i]->start,
 938                                 re->zones[i]->end);
 939                         for (j = 0; j < re->zones[i]->ndevs; ++j) {
 940                                 pr_cont(" %lld",
 941                                        re->zones[i]->devs[j]->devid);
 942                         }
 943                 }
 944                 pr_cont("\n");
 945                 index = (re->logical >> PAGE_SHIFT) + 1;
 946         }
 947         spin_unlock(&fs_info->reada_lock);
 948 }
 949 #endif
 950
 951 /*
 952  * interface
 953  */
 954 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
 955                         struct btrfs_key *key_start, struct btrfs_key *key_end)
 956 {
 957         struct reada_control *rc;
 958         u64 start;
 959         u64 generation;
 960         int ret;
 961         int level;
 962         struct extent_buffer *node;
 963         static struct btrfs_key max_key = {
 964                 .objectid = (u64)-1,
 965                 .type = (u8)-1,
 966                 .offset = (u64)-1
 967         };
 968
 969         rc = kzalloc(sizeof(*rc), GFP_KERNEL);
 970         if (!rc)
 971                 return ERR_PTR(-ENOMEM);
 972
 973         rc->fs_info = root->fs_info;
 974         rc->key_start = *key_start;
 975         rc->key_end = *key_end;
 976         atomic_set(&rc->elems, 0);
 977         init_waitqueue_head(&rc->wait);
 978         kref_init(&rc->refcnt);
 979         kref_get(&rc->refcnt); /* one ref for having elements */
 980
 981         node = btrfs_root_node(root);
 982         start = node->start;
 983         generation = btrfs_header_generation(node);
 984         level = btrfs_header_level(node);
 985         free_extent_buffer(node);
 986
 987         ret = reada_add_block(rc, start, &max_key, root->root_key.objectid,
 988                               generation, level);
 989         if (ret) {
 990                 kfree(rc);
 991                 return ERR_PTR(ret);
 992         }
 993
 994         reada_start_machine(root->fs_info);
 995
 996         return rc;
 997 }
 998
 999 #ifdef DEBUG
1000 int btrfs_reada_wait(void *handle)
1001 {
1002         struct reada_control *rc = handle;
1003         struct btrfs_fs_info *fs_info = rc->fs_info;
1004
1005         while (atomic_read(&rc->elems)) {
1006                 if (!atomic_read(&fs_info->reada_works_cnt))
1007                         reada_start_machine(fs_info);
1008                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1009                                    5 * HZ);
1010                 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
1011         }
1012
1013         dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
1014
1015         kref_put(&rc->refcnt, reada_control_release);
1016
1017         return 0;
1018 }
1019 #else
1020 int btrfs_reada_wait(void *handle)
1021 {
1022         struct reada_control *rc = handle;
1023         struct btrfs_fs_info *fs_info = rc->fs_info;
1024
1025         while (atomic_read(&rc->elems)) {
1026                 if (!atomic_read(&fs_info->reada_works_cnt))
1027                         reada_start_machine(fs_info);
1028                 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1029                                    (HZ + 9) / 10);
1030         }
1031
1032         kref_put(&rc->refcnt, reada_control_release);
1033
1034         return 0;
1035 }
1036 #endif
1037
1038 void btrfs_reada_detach(void *handle)
1039 {
1040         struct reada_control *rc = handle;
1041
1042         kref_put(&rc->refcnt, reada_control_release);
1043 }
1044
1045 /*
1046  * Before removing a device (device replace or device remove ioctls), call this
1047  * function to wait for all existing readahead requests on the device and to
1048  * make sure no one queues more readahead requests for the device.
1049  *
1050  * Must be called without holding neither the device list mutex nor the device
1051  * replace semaphore, otherwise it will deadlock.
1052  */
1053 void btrfs_reada_remove_dev(struct btrfs_device *dev)
1054 {
1055         struct btrfs_fs_info *fs_info = dev->fs_info;
1056
1057         /* Serialize with readahead extent creation at reada_find_extent(). */
1058         spin_lock(&fs_info->reada_lock);
1059         set_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state);
1060         spin_unlock(&fs_info->reada_lock);
1061
1062         /*
1063          * There might be readahead requests added to the radix trees which
1064          * were not yet added to the readahead work queue. We need to start
1065          * them and wait for their completion, otherwise we can end up with
1066          * use-after-free problems when dropping the last reference on the
1067          * readahead extents and their zones, as they need to access the
1068          * device structure.
1069          */
1070         reada_start_machine(fs_info);
1071         btrfs_flush_workqueue(fs_info->readahead_workers);
1072 }
1073
1074 /*
1075  * If when removing a device (device replace or device remove ioctls) an error
1076  * happens after calling btrfs_reada_remove_dev(), call this to undo what that
1077  * function did. This is safe to call even if btrfs_reada_remove_dev() was not
1078  * called before.
1079  */
1080 void btrfs_reada_undo_remove_dev(struct btrfs_device *dev)
1081 {
1082         spin_lock(&dev->fs_info->reada_lock);
1083         clear_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state);
1084         spin_unlock(&dev->fs_info->reada_lock);
1085 }