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