git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/buckets.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for manipulating bucket marks for garbage collection.
   4  *
   5  * Copyright 2014 Datera, Inc.
   6  *
   7  * Bucket states:
   8  * - free bucket: mark == 0
   9  *   The bucket contains no data and will not be read
  10  *
  11  * - allocator bucket: owned_by_allocator == 1
  12  *   The bucket is on a free list, or it is an open bucket
  13  *
  14  * - cached bucket: owned_by_allocator == 0 &&
  15  *                  dirty_sectors == 0 &&
  16  *                  cached_sectors > 0
  17  *   The bucket contains data but may be safely discarded as there are
  18  *   enough replicas of the data on other cache devices, or it has been
  19  *   written back to the backing device
  20  *
  21  * - dirty bucket: owned_by_allocator == 0 &&
  22  *                 dirty_sectors > 0
  23  *   The bucket contains data that we must not discard (either only copy,
  24  *   or one of the 'main copies' for data requiring multiple replicas)
  25  *
  26  * - metadata bucket: owned_by_allocator == 0 && is_metadata == 1
  27  *   This is a btree node, journal or gen/prio bucket
  28  *
  29  * Lifecycle:
  30  *
  31  * bucket invalidated => bucket on freelist => open bucket =>
  32  *     [dirty bucket =>] cached bucket => bucket invalidated => ...
  33  *
  34  * Note that cache promotion can skip the dirty bucket step, as data
  35  * is copied from a deeper tier to a shallower tier, onto a cached
  36  * bucket.
  37  * Note also that a cached bucket can spontaneously become dirty --
  38  * see below.
  39  *
  40  * Only a traversal of the key space can determine whether a bucket is
  41  * truly dirty or cached.
  42  *
  43  * Transitions:
  44  *
  45  * - free => allocator: bucket was invalidated
  46  * - cached => allocator: bucket was invalidated
  47  *
  48  * - allocator => dirty: open bucket was filled up
  49  * - allocator => cached: open bucket was filled up
  50  * - allocator => metadata: metadata was allocated
  51  *
  52  * - dirty => cached: dirty sectors were copied to a deeper tier
  53  * - dirty => free: dirty sectors were overwritten or moved (copy gc)
  54  * - cached => free: cached sectors were overwritten
  55  *
  56  * - metadata => free: metadata was freed
  57  *
  58  * Oddities:
  59  * - cached => dirty: a device was removed so formerly replicated data
  60  *                    is no longer sufficiently replicated
  61  * - free => cached: cannot happen
  62  * - free => dirty: cannot happen
  63  * - free => metadata: cannot happen
  64  */
  65
  66 #include "bcachefs.h"
  67 #include "alloc_background.h"
  68 #include "bset.h"
  69 #include "btree_gc.h"
  70 #include "btree_update.h"
  71 #include "buckets.h"
  72 #include "ec.h"
  73 #include "error.h"
  74 #include "movinggc.h"
  75 #include "replicas.h"
  76
  77 #include <linux/preempt.h>
  78 #include <trace/events/bcachefs.h>
  79
  80 /*
  81  * Clear journal_seq_valid for buckets for which it's not needed, to prevent
  82  * wraparound:
  83  */
  84 void bch2_bucket_seq_cleanup(struct bch_fs *c)
  85 {
  86         u64 journal_seq = atomic64_read(&c->journal.seq);
  87         u16 last_seq_ondisk = c->journal.last_seq_ondisk;
  88         struct bch_dev *ca;
  89         struct bucket_array *buckets;
  90         struct bucket *g;
  91         struct bucket_mark m;
  92         unsigned i;
  93
  94         if (journal_seq - c->last_bucket_seq_cleanup <
  95             (1U << (BUCKET_JOURNAL_SEQ_BITS - 2)))
  96                 return;
  97
  98         c->last_bucket_seq_cleanup = journal_seq;
  99
 100         for_each_member_device(ca, c, i) {
 101                 down_read(&ca->bucket_lock);
 102                 buckets = bucket_array(ca);
 103
 104                 for_each_bucket(g, buckets) {
 105                         bucket_cmpxchg(g, m, ({
 106                                 if (!m.journal_seq_valid ||
 107                                     bucket_needs_journal_commit(m, last_seq_ondisk))
 108                                         break;
 109
 110                                 m.journal_seq_valid = 0;
 111                         }));
 112                 }
 113                 up_read(&ca->bucket_lock);
 114         }
 115 }
 116
 117 void bch2_fs_usage_initialize(struct bch_fs *c)
 118 {
 119         struct bch_fs_usage *usage;
 120         unsigned i;
 121
 122         percpu_down_write(&c->mark_lock);
 123         usage = c->usage_base;
 124
 125         bch2_fs_usage_acc_to_base(c, 0);
 126         bch2_fs_usage_acc_to_base(c, 1);
 127
 128         for (i = 0; i < BCH_REPLICAS_MAX; i++)
 129                 usage->reserved += usage->persistent_reserved[i];
 130
 131         for (i = 0; i < c->replicas.nr; i++) {
 132                 struct bch_replicas_entry *e =
 133                         cpu_replicas_entry(&c->replicas, i);
 134
 135                 switch (e->data_type) {
 136                 case BCH_DATA_BTREE:
 137                         usage->btree    += usage->replicas[i];
 138                         break;
 139                 case BCH_DATA_USER:
 140                         usage->data     += usage->replicas[i];
 141                         break;
 142                 case BCH_DATA_CACHED:
 143                         usage->cached   += usage->replicas[i];
 144                         break;
 145                 }
 146         }
 147
 148         percpu_up_write(&c->mark_lock);
 149 }
 150
 151 void bch2_fs_usage_scratch_put(struct bch_fs *c, struct bch_fs_usage *fs_usage)
 152 {
 153         if (fs_usage == c->usage_scratch)
 154                 mutex_unlock(&c->usage_scratch_lock);
 155         else
 156                 kfree(fs_usage);
 157 }
 158
 159 struct bch_fs_usage *bch2_fs_usage_scratch_get(struct bch_fs *c)
 160 {
 161         struct bch_fs_usage *ret;
 162         unsigned bytes = fs_usage_u64s(c) * sizeof(u64);
 163
 164         ret = kzalloc(bytes, GFP_NOWAIT|__GFP_NOWARN);
 165         if (ret)
 166                 return ret;
 167
 168         if (mutex_trylock(&c->usage_scratch_lock))
 169                 goto out_pool;
 170
 171         ret = kzalloc(bytes, GFP_NOFS);
 172         if (ret)
 173                 return ret;
 174
 175         mutex_lock(&c->usage_scratch_lock);
 176 out_pool:
 177         ret = c->usage_scratch;
 178         memset(ret, 0, bytes);
 179         return ret;
 180 }
 181
 182 struct bch_dev_usage bch2_dev_usage_read(struct bch_fs *c, struct bch_dev *ca)
 183 {
 184         struct bch_dev_usage ret;
 185
 186         memset(&ret, 0, sizeof(ret));
 187         acc_u64s_percpu((u64 *) &ret,
 188                         (u64 __percpu *) ca->usage[0],
 189                         sizeof(ret) / sizeof(u64));
 190
 191         return ret;
 192 }
 193
 194 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 195                                                 unsigned journal_seq,
 196                                                 bool gc)
 197 {
 198         return this_cpu_ptr(gc
 199                             ? c->usage_gc
 200                             : c->usage[journal_seq & 1]);
 201 }
 202
 203 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 204 {
 205         ssize_t offset = v - (u64 *) c->usage_base;
 206         unsigned seq;
 207         u64 ret;
 208
 209         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 210         percpu_rwsem_assert_held(&c->mark_lock);
 211
 212         do {
 213                 seq = read_seqcount_begin(&c->usage_lock);
 214                 ret = *v +
 215                         percpu_u64_get((u64 __percpu *) c->usage[0] + offset) +
 216                         percpu_u64_get((u64 __percpu *) c->usage[1] + offset);
 217         } while (read_seqcount_retry(&c->usage_lock, seq));
 218
 219         return ret;
 220 }
 221
 222 struct bch_fs_usage *bch2_fs_usage_read(struct bch_fs *c)
 223 {
 224         struct bch_fs_usage *ret;
 225         unsigned seq, v, u64s = fs_usage_u64s(c);
 226 retry:
 227         ret = kmalloc(u64s * sizeof(u64), GFP_NOFS);
 228         if (unlikely(!ret))
 229                 return NULL;
 230
 231         percpu_down_read(&c->mark_lock);
 232
 233         v = fs_usage_u64s(c);
 234         if (unlikely(u64s != v)) {
 235                 u64s = v;
 236                 percpu_up_read(&c->mark_lock);
 237                 kfree(ret);
 238                 goto retry;
 239         }
 240
 241         do {
 242                 seq = read_seqcount_begin(&c->usage_lock);
 243                 memcpy(ret, c->usage_base, u64s * sizeof(u64));
 244                 acc_u64s_percpu((u64 *) ret, (u64 __percpu *) c->usage[0], u64s);
 245                 acc_u64s_percpu((u64 *) ret, (u64 __percpu *) c->usage[1], u64s);
 246         } while (read_seqcount_retry(&c->usage_lock, seq));
 247
 248         return ret;
 249 }
 250
 251 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 252 {
 253         unsigned u64s = fs_usage_u64s(c);
 254
 255         BUG_ON(idx >= 2);
 256
 257         write_seqcount_begin(&c->usage_lock);
 258
 259         acc_u64s_percpu((u64 *) c->usage_base,
 260                         (u64 __percpu *) c->usage[idx], u64s);
 261         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 262
 263         write_seqcount_end(&c->usage_lock);
 264 }
 265
 266 void bch2_fs_usage_to_text(struct printbuf *out,
 267                            struct bch_fs *c,
 268                            struct bch_fs_usage *fs_usage)
 269 {
 270         unsigned i;
 271
 272         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 273
 274         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 275                fs_usage->hidden);
 276         pr_buf(out, "data:\t\t\t\t%llu\n",
 277                fs_usage->data);
 278         pr_buf(out, "cached:\t\t\t\t%llu\n",
 279                fs_usage->cached);
 280         pr_buf(out, "reserved:\t\t\t%llu\n",
 281                fs_usage->reserved);
 282         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 283                fs_usage->nr_inodes);
 284         pr_buf(out, "online reserved:\t\t%llu\n",
 285                fs_usage->online_reserved);
 286
 287         for (i = 0;
 288              i < ARRAY_SIZE(fs_usage->persistent_reserved);
 289              i++) {
 290                 pr_buf(out, "%u replicas:\n", i + 1);
 291                 pr_buf(out, "\treserved:\t\t%llu\n",
 292                        fs_usage->persistent_reserved[i]);
 293         }
 294
 295         for (i = 0; i < c->replicas.nr; i++) {
 296                 struct bch_replicas_entry *e =
 297                         cpu_replicas_entry(&c->replicas, i);
 298
 299                 pr_buf(out, "\t");
 300                 bch2_replicas_entry_to_text(out, e);
 301                 pr_buf(out, ":\t%llu\n", fs_usage->replicas[i]);
 302         }
 303 }
 304
 305 #define RESERVE_FACTOR  6
 306
 307 static u64 reserve_factor(u64 r)
 308 {
 309         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 310 }
 311
 312 static u64 avail_factor(u64 r)
 313 {
 314         return (r << RESERVE_FACTOR) / ((1 << RESERVE_FACTOR) + 1);
 315 }
 316
 317 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage *fs_usage)
 318 {
 319         return min(fs_usage->hidden +
 320                    fs_usage->btree +
 321                    fs_usage->data +
 322                    reserve_factor(fs_usage->reserved +
 323                                   fs_usage->online_reserved),
 324                    c->capacity);
 325 }
 326
 327 static struct bch_fs_usage_short
 328 __bch2_fs_usage_read_short(struct bch_fs *c)
 329 {
 330         struct bch_fs_usage_short ret;
 331         u64 data, reserved;
 332
 333         ret.capacity = c->capacity -
 334                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 335
 336         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 337                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 338         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 339                 bch2_fs_usage_read_one(c, &c->usage_base->online_reserved);
 340
 341         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 342         ret.free        = ret.capacity - ret.used;
 343
 344         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 345
 346         return ret;
 347 }
 348
 349 struct bch_fs_usage_short
 350 bch2_fs_usage_read_short(struct bch_fs *c)
 351 {
 352         struct bch_fs_usage_short ret;
 353
 354         percpu_down_read(&c->mark_lock);
 355         ret = __bch2_fs_usage_read_short(c);
 356         percpu_up_read(&c->mark_lock);
 357
 358         return ret;
 359 }
 360
 361 static inline int is_unavailable_bucket(struct bucket_mark m)
 362 {
 363         return !is_available_bucket(m);
 364 }
 365
 366 static inline int is_fragmented_bucket(struct bucket_mark m,
 367                                        struct bch_dev *ca)
 368 {
 369         if (!m.owned_by_allocator &&
 370             m.data_type == BCH_DATA_USER &&
 371             bucket_sectors_used(m))
 372                 return max_t(int, 0, (int) ca->mi.bucket_size -
 373                              bucket_sectors_used(m));
 374         return 0;
 375 }
 376
 377 static inline enum bch_data_type bucket_type(struct bucket_mark m)
 378 {
 379         return m.cached_sectors && !m.dirty_sectors
 380                 ? BCH_DATA_CACHED
 381                 : m.data_type;
 382 }
 383
 384 static bool bucket_became_unavailable(struct bucket_mark old,
 385                                       struct bucket_mark new)
 386 {
 387         return is_available_bucket(old) &&
 388                !is_available_bucket(new);
 389 }
 390
 391 int bch2_fs_usage_apply(struct bch_fs *c,
 392                         struct bch_fs_usage *fs_usage,
 393                         struct disk_reservation *disk_res,
 394                         unsigned journal_seq)
 395 {
 396         s64 added = fs_usage->data + fs_usage->reserved;
 397         s64 should_not_have_added;
 398         int ret = 0;
 399
 400         percpu_rwsem_assert_held(&c->mark_lock);
 401
 402         /*
 403          * Not allowed to reduce sectors_available except by getting a
 404          * reservation:
 405          */
 406         should_not_have_added = added - (s64) (disk_res ? disk_res->sectors : 0);
 407         if (WARN_ONCE(should_not_have_added > 0,
 408                       "disk usage increased by %lli without a reservation",
 409                       should_not_have_added)) {
 410                 atomic64_sub(should_not_have_added, &c->sectors_available);
 411                 added -= should_not_have_added;
 412                 ret = -1;
 413         }
 414
 415         if (added > 0) {
 416                 disk_res->sectors               -= added;
 417                 fs_usage->online_reserved       -= added;
 418         }
 419
 420         preempt_disable();
 421         acc_u64s((u64 *) fs_usage_ptr(c, journal_seq, false),
 422                  (u64 *) fs_usage, fs_usage_u64s(c));
 423         preempt_enable();
 424
 425         return ret;
 426 }
 427
 428 static inline void account_bucket(struct bch_fs_usage *fs_usage,
 429                                   struct bch_dev_usage *dev_usage,
 430                                   enum bch_data_type type,
 431                                   int nr, s64 size)
 432 {
 433         if (type == BCH_DATA_SB || type == BCH_DATA_JOURNAL)
 434                 fs_usage->hidden        += size;
 435
 436         dev_usage->buckets[type]        += nr;
 437 }
 438
 439 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 440                                   struct bch_fs_usage *fs_usage,
 441                                   struct bucket_mark old, struct bucket_mark new,
 442                                   bool gc)
 443 {
 444         struct bch_dev_usage *dev_usage;
 445
 446         percpu_rwsem_assert_held(&c->mark_lock);
 447
 448         preempt_disable();
 449         dev_usage = this_cpu_ptr(ca->usage[gc]);
 450
 451         if (bucket_type(old))
 452                 account_bucket(fs_usage, dev_usage, bucket_type(old),
 453                                -1, -ca->mi.bucket_size);
 454
 455         if (bucket_type(new))
 456                 account_bucket(fs_usage, dev_usage, bucket_type(new),
 457                                1, ca->mi.bucket_size);
 458
 459         dev_usage->buckets_alloc +=
 460                 (int) new.owned_by_allocator - (int) old.owned_by_allocator;
 461         dev_usage->buckets_ec +=
 462                 (int) new.stripe - (int) old.stripe;
 463         dev_usage->buckets_unavailable +=
 464                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 465
 466         dev_usage->sectors[old.data_type] -= old.dirty_sectors;
 467         dev_usage->sectors[new.data_type] += new.dirty_sectors;
 468         dev_usage->sectors[BCH_DATA_CACHED] +=
 469                 (int) new.cached_sectors - (int) old.cached_sectors;
 470         dev_usage->sectors_fragmented +=
 471                 is_fragmented_bucket(new, ca) - is_fragmented_bucket(old, ca);
 472         preempt_enable();
 473
 474         if (!is_available_bucket(old) && is_available_bucket(new))
 475                 bch2_wake_allocator(ca);
 476 }
 477
 478 void bch2_dev_usage_from_buckets(struct bch_fs *c)
 479 {
 480         struct bch_dev *ca;
 481         struct bucket_mark old = { .v.counter = 0 };
 482         struct bucket_array *buckets;
 483         struct bucket *g;
 484         unsigned i;
 485         int cpu;
 486
 487         c->usage_base->hidden = 0;
 488
 489         for_each_member_device(ca, c, i) {
 490                 for_each_possible_cpu(cpu)
 491                         memset(per_cpu_ptr(ca->usage[0], cpu), 0,
 492                                sizeof(*ca->usage[0]));
 493
 494                 buckets = bucket_array(ca);
 495
 496                 for_each_bucket(g, buckets)
 497                         bch2_dev_usage_update(c, ca, c->usage_base,
 498                                               old, g->mark, false);
 499         }
 500 }
 501
 502 static inline int update_replicas(struct bch_fs *c,
 503                                   struct bch_fs_usage *fs_usage,
 504                                   struct bch_replicas_entry *r,
 505                                   s64 sectors)
 506 {
 507         int idx = bch2_replicas_entry_idx(c, r);
 508
 509         if (idx < 0)
 510                 return -1;
 511
 512         if (!fs_usage)
 513                 return 0;
 514
 515         switch (r->data_type) {
 516         case BCH_DATA_BTREE:
 517                 fs_usage->btree         += sectors;
 518                 break;
 519         case BCH_DATA_USER:
 520                 fs_usage->data          += sectors;
 521                 break;
 522         case BCH_DATA_CACHED:
 523                 fs_usage->cached        += sectors;
 524                 break;
 525         }
 526         fs_usage->replicas[idx]         += sectors;
 527         return 0;
 528 }
 529
 530 static inline void update_cached_sectors(struct bch_fs *c,
 531                                          struct bch_fs_usage *fs_usage,
 532                                          unsigned dev, s64 sectors)
 533 {
 534         struct bch_replicas_padded r;
 535
 536         bch2_replicas_entry_cached(&r.e, dev);
 537
 538         update_replicas(c, fs_usage, &r.e, sectors);
 539 }
 540
 541 static struct replicas_delta_list *
 542 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 543 {
 544         struct replicas_delta_list *d = trans->fs_usage_deltas;
 545         unsigned new_size = d ? (d->size + more) * 2 : 128;
 546
 547         if (!d || d->used + more > d->size) {
 548                 d = krealloc(d, sizeof(*d) + new_size, GFP_NOIO|__GFP_ZERO);
 549                 BUG_ON(!d);
 550
 551                 d->size = new_size;
 552                 trans->fs_usage_deltas = d;
 553         }
 554         return d;
 555 }
 556
 557 static inline void update_replicas_list(struct btree_trans *trans,
 558                                         struct bch_replicas_entry *r,
 559                                         s64 sectors)
 560 {
 561         struct replicas_delta_list *d;
 562         struct replicas_delta *n;
 563         unsigned b;
 564
 565         if (!sectors)
 566                 return;
 567
 568         b = replicas_entry_bytes(r) + 8;
 569         d = replicas_deltas_realloc(trans, b);
 570
 571         n = (void *) d->d + d->used;
 572         n->delta = sectors;
 573         memcpy(&n->r, r, replicas_entry_bytes(r));
 574         d->used += b;
 575 }
 576
 577 static inline void update_cached_sectors_list(struct btree_trans *trans,
 578                                               unsigned dev, s64 sectors)
 579 {
 580         struct bch_replicas_padded r;
 581
 582         bch2_replicas_entry_cached(&r.e, dev);
 583
 584         update_replicas_list(trans, &r.e, sectors);
 585 }
 586
 587 static inline struct replicas_delta *
 588 replicas_delta_next(struct replicas_delta *d)
 589 {
 590         return (void *) d + replicas_entry_bytes(&d->r) + 8;
 591 }
 592
 593 int bch2_replicas_delta_list_apply(struct bch_fs *c,
 594                                    struct bch_fs_usage *fs_usage,
 595                                    struct replicas_delta_list *r)
 596 {
 597         struct replicas_delta *d = r->d;
 598         struct replicas_delta *top = (void *) r->d + r->used;
 599         unsigned i;
 600
 601         for (d = r->d; d != top; d = replicas_delta_next(d))
 602                 if (update_replicas(c, fs_usage, &d->r, d->delta)) {
 603                         top = d;
 604                         goto unwind;
 605                 }
 606
 607         if (!fs_usage)
 608                 return 0;
 609
 610         fs_usage->nr_inodes += r->nr_inodes;
 611
 612         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
 613                 fs_usage->reserved += r->persistent_reserved[i];
 614                 fs_usage->persistent_reserved[i] += r->persistent_reserved[i];
 615         }
 616
 617         return 0;
 618 unwind:
 619         for (d = r->d; d != top; d = replicas_delta_next(d))
 620                 update_replicas(c, fs_usage, &d->r, -d->delta);
 621         return -1;
 622 }
 623
 624 #define do_mark_fn(fn, c, pos, flags, ...)                              \
 625 ({                                                                      \
 626         int gc, ret = 0;                                                \
 627                                                                         \
 628         percpu_rwsem_assert_held(&c->mark_lock);                        \
 629                                                                         \
 630         for (gc = 0; gc < 2 && !ret; gc++)                              \
 631                 if (!gc == !(flags & BTREE_TRIGGER_GC) ||               \
 632                     (gc && gc_visited(c, pos)))                         \
 633                         ret = fn(c, __VA_ARGS__, gc);                   \
 634         ret;                                                            \
 635 })
 636
 637 static int __bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 638                                     size_t b, struct bucket_mark *ret,
 639                                     bool gc)
 640 {
 641         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 642         struct bucket *g = __bucket(ca, b, gc);
 643         struct bucket_mark old, new;
 644
 645         old = bucket_cmpxchg(g, new, ({
 646                 BUG_ON(!is_available_bucket(new));
 647
 648                 new.owned_by_allocator  = true;
 649                 new.data_type           = 0;
 650                 new.cached_sectors      = 0;
 651                 new.dirty_sectors       = 0;
 652                 new.gen++;
 653         }));
 654
 655         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 656
 657         if (old.cached_sectors)
 658                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 659                                       -((s64) old.cached_sectors));
 660
 661         if (!gc)
 662                 *ret = old;
 663         return 0;
 664 }
 665
 666 void bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 667                             size_t b, struct bucket_mark *old)
 668 {
 669         do_mark_fn(__bch2_invalidate_bucket, c, gc_phase(GC_PHASE_START), 0,
 670                    ca, b, old);
 671
 672         if (!old->owned_by_allocator && old->cached_sectors)
 673                 trace_invalidate(ca, bucket_to_sector(ca, b),
 674                                  old->cached_sectors);
 675 }
 676
 677 static int __bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 678                                     size_t b, bool owned_by_allocator,
 679                                     bool gc)
 680 {
 681         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 682         struct bucket *g = __bucket(ca, b, gc);
 683         struct bucket_mark old, new;
 684
 685         old = bucket_cmpxchg(g, new, ({
 686                 new.owned_by_allocator  = owned_by_allocator;
 687         }));
 688
 689         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 690
 691         BUG_ON(!gc &&
 692                !owned_by_allocator && !old.owned_by_allocator);
 693
 694         return 0;
 695 }
 696
 697 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 698                             size_t b, bool owned_by_allocator,
 699                             struct gc_pos pos, unsigned flags)
 700 {
 701         preempt_disable();
 702
 703         do_mark_fn(__bch2_mark_alloc_bucket, c, pos, flags,
 704                    ca, b, owned_by_allocator);
 705
 706         preempt_enable();
 707 }
 708
 709 static int bch2_mark_alloc(struct bch_fs *c, struct bkey_s_c k,
 710                            struct bch_fs_usage *fs_usage,
 711                            u64 journal_seq, unsigned flags)
 712 {
 713         bool gc = flags & BTREE_TRIGGER_GC;
 714         struct bkey_alloc_unpacked u;
 715         struct bch_dev *ca;
 716         struct bucket *g;
 717         struct bucket_mark old, m;
 718
 719         /*
 720          * alloc btree is read in by bch2_alloc_read, not gc:
 721          */
 722         if ((flags & BTREE_TRIGGER_GC) &&
 723             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 724                 return 0;
 725
 726         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 727
 728         if (k.k->p.offset >= ca->mi.nbuckets)
 729                 return 0;
 730
 731         g = __bucket(ca, k.k->p.offset, gc);
 732         u = bch2_alloc_unpack(k);
 733
 734         old = bucket_cmpxchg(g, m, ({
 735                 m.gen                   = u.gen;
 736                 m.data_type             = u.data_type;
 737                 m.dirty_sectors         = u.dirty_sectors;
 738                 m.cached_sectors        = u.cached_sectors;
 739
 740                 if (journal_seq) {
 741                         m.journal_seq_valid     = 1;
 742                         m.journal_seq           = journal_seq;
 743                 }
 744         }));
 745
 746         if (!(flags & BTREE_TRIGGER_ALLOC_READ))
 747                 bch2_dev_usage_update(c, ca, fs_usage, old, m, gc);
 748
 749         g->io_time[READ]        = u.read_time;
 750         g->io_time[WRITE]       = u.write_time;
 751         g->oldest_gen           = u.oldest_gen;
 752         g->gen_valid            = 1;
 753
 754         /*
 755          * need to know if we're getting called from the invalidate path or
 756          * not:
 757          */
 758
 759         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 760             old.cached_sectors) {
 761                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 762                                       -old.cached_sectors);
 763                 trace_invalidate(ca, bucket_to_sector(ca, k.k->p.offset),
 764                                  old.cached_sectors);
 765         }
 766
 767         return 0;
 768 }
 769
 770 #define checked_add(a, b)                                       \
 771 ({                                                              \
 772         unsigned _res = (unsigned) (a) + (b);                   \
 773         bool overflow = _res > U16_MAX;                         \
 774         if (overflow)                                           \
 775                 _res = U16_MAX;                                 \
 776         (a) = _res;                                             \
 777         overflow;                                               \
 778 })
 779
 780 static int __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 781                                        size_t b, enum bch_data_type data_type,
 782                                        unsigned sectors, bool gc)
 783 {
 784         struct bucket *g = __bucket(ca, b, gc);
 785         struct bucket_mark old, new;
 786         bool overflow;
 787
 788         BUG_ON(data_type != BCH_DATA_SB &&
 789                data_type != BCH_DATA_JOURNAL);
 790
 791         old = bucket_cmpxchg(g, new, ({
 792                 new.data_type   = data_type;
 793                 overflow = checked_add(new.dirty_sectors, sectors);
 794         }));
 795
 796         bch2_fs_inconsistent_on(old.data_type &&
 797                                 old.data_type != data_type, c,
 798                 "different types of data in same bucket: %s, %s",
 799                 bch2_data_types[old.data_type],
 800                 bch2_data_types[data_type]);
 801
 802         bch2_fs_inconsistent_on(overflow, c,
 803                 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
 804                 ca->dev_idx, b, new.gen,
 805                 bch2_data_types[old.data_type ?: data_type],
 806                 old.dirty_sectors, sectors);
 807
 808         if (c)
 809                 bch2_dev_usage_update(c, ca, fs_usage_ptr(c, 0, gc),
 810                                       old, new, gc);
 811
 812         return 0;
 813 }
 814
 815 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 816                                size_t b, enum bch_data_type type,
 817                                unsigned sectors, struct gc_pos pos,
 818                                unsigned flags)
 819 {
 820         BUG_ON(type != BCH_DATA_SB &&
 821                type != BCH_DATA_JOURNAL);
 822
 823         preempt_disable();
 824
 825         if (likely(c)) {
 826                 do_mark_fn(__bch2_mark_metadata_bucket, c, pos, flags,
 827                            ca, b, type, sectors);
 828         } else {
 829                 __bch2_mark_metadata_bucket(c, ca, b, type, sectors, 0);
 830         }
 831
 832         preempt_enable();
 833 }
 834
 835 static s64 disk_sectors_scaled(unsigned n, unsigned d, unsigned sectors)
 836 {
 837         return DIV_ROUND_UP(sectors * n, d);
 838 }
 839
 840 static s64 __ptr_disk_sectors_delta(unsigned old_size,
 841                                     unsigned offset, s64 delta,
 842                                     unsigned flags,
 843                                     unsigned n, unsigned d)
 844 {
 845         BUG_ON(!n || !d);
 846
 847         if (flags & BTREE_TRIGGER_OVERWRITE_SPLIT) {
 848                 BUG_ON(offset + -delta > old_size);
 849
 850                 return -disk_sectors_scaled(n, d, old_size) +
 851                         disk_sectors_scaled(n, d, offset) +
 852                         disk_sectors_scaled(n, d, old_size - offset + delta);
 853         } else if (flags & BTREE_TRIGGER_OVERWRITE) {
 854                 BUG_ON(offset + -delta > old_size);
 855
 856                 return -disk_sectors_scaled(n, d, old_size) +
 857                         disk_sectors_scaled(n, d, old_size + delta);
 858         } else {
 859                 return  disk_sectors_scaled(n, d, delta);
 860         }
 861 }
 862
 863 static s64 ptr_disk_sectors_delta(struct extent_ptr_decoded p,
 864                                   unsigned offset, s64 delta,
 865                                   unsigned flags)
 866 {
 867         return __ptr_disk_sectors_delta(p.crc.live_size,
 868                                         offset, delta, flags,
 869                                         p.crc.compressed_size,
 870                                         p.crc.uncompressed_size);
 871 }
 872
 873 static void bucket_set_stripe(struct bch_fs *c,
 874                               const struct bch_stripe *v,
 875                               struct bch_fs_usage *fs_usage,
 876                               u64 journal_seq,
 877                               unsigned flags)
 878 {
 879         bool enabled = !(flags & BTREE_TRIGGER_OVERWRITE);
 880         bool gc = flags & BTREE_TRIGGER_GC;
 881         unsigned i;
 882
 883         for (i = 0; i < v->nr_blocks; i++) {
 884                 const struct bch_extent_ptr *ptr = v->ptrs + i;
 885                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 886                 struct bucket *g = PTR_BUCKET(ca, ptr, gc);
 887                 struct bucket_mark new, old;
 888
 889                 old = bucket_cmpxchg(g, new, ({
 890                         new.stripe                      = enabled;
 891                         if (journal_seq) {
 892                                 new.journal_seq_valid   = 1;
 893                                 new.journal_seq         = journal_seq;
 894                         }
 895                 }));
 896
 897                 bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 898
 899                 /*
 900                  * XXX write repair code for these, flag stripe as possibly bad
 901                  */
 902                 if (old.gen != ptr->gen)
 903                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 904                                       "stripe with stale pointer");
 905 #if 0
 906                 /*
 907                  * We'd like to check for these, but these checks don't work
 908                  * yet:
 909                  */
 910                 if (old.stripe && enabled)
 911                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 912                                       "multiple stripes using same bucket");
 913
 914                 if (!old.stripe && !enabled)
 915                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 916                                       "deleting stripe but bucket not marked as stripe bucket");
 917 #endif
 918         }
 919 }
 920
 921 static int __mark_pointer(struct bch_fs *c, struct bkey_s_c k,
 922                           struct extent_ptr_decoded p,
 923                           s64 sectors, enum bch_data_type ptr_data_type,
 924                           u8 bucket_gen, u8 *bucket_data_type,
 925                           u16 *dirty_sectors, u16 *cached_sectors)
 926 {
 927         u16 *dst_sectors = !p.ptr.cached
 928                 ? dirty_sectors
 929                 : cached_sectors;
 930         u16 orig_sectors = *dst_sectors;
 931         char buf[200];
 932
 933         if (gen_after(p.ptr.gen, bucket_gen)) {
 934                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 935                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 936                         "while marking %s",
 937                         p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
 938                         bucket_gen,
 939                         bch2_data_types[*bucket_data_type ?: ptr_data_type],
 940                         p.ptr.gen,
 941                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 942                 return -EIO;
 943         }
 944
 945         if (gen_cmp(bucket_gen, p.ptr.gen) >= 96U) {
 946                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 947                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 948                         "while marking %s",
 949                         p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
 950                         bucket_gen,
 951                         bch2_data_types[*bucket_data_type ?: ptr_data_type],
 952                         p.ptr.gen,
 953                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 954                 return -EIO;
 955         }
 956
 957         if (bucket_gen != p.ptr.gen && !p.ptr.cached) {
 958                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 959                         "bucket %u:%zu gen %u data type %s: stale dirty ptr (gen %u)\n"
 960                         "while marking %s",
 961                         p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
 962                         bucket_gen,
 963                         bch2_data_types[*bucket_data_type ?: ptr_data_type],
 964                         p.ptr.gen,
 965                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 966                 return -EIO;
 967         }
 968
 969         if (bucket_gen != p.ptr.gen)
 970                 return 1;
 971
 972         if (*bucket_data_type && *bucket_data_type != ptr_data_type) {
 973                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 974                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 975                         "while marking %s",
 976                         p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
 977                         bucket_gen,
 978                         bch2_data_types[*bucket_data_type],
 979                         bch2_data_types[ptr_data_type],
 980                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 981                 return -EIO;
 982         }
 983
 984         if (checked_add(*dst_sectors, sectors)) {
 985                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 986                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 987                         "while marking %s",
 988                         p.ptr.dev, PTR_BUCKET_NR(bch_dev_bkey_exists(c, p.ptr.dev), &p.ptr),
 989                         bucket_gen,
 990                         bch2_data_types[*bucket_data_type ?: ptr_data_type],
 991                         orig_sectors, sectors,
 992                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 993                 return -EIO;
 994         }
 995
 996         *bucket_data_type = *dirty_sectors || *cached_sectors
 997                 ? ptr_data_type : 0;
 998         return 0;
 999 }
1000
1001 static int bch2_mark_pointer(struct bch_fs *c, struct bkey_s_c k,
1002                              struct extent_ptr_decoded p,
1003                              s64 sectors, enum bch_data_type data_type,
1004                              struct bch_fs_usage *fs_usage,
1005                              u64 journal_seq, unsigned flags)
1006 {
1007         bool gc = flags & BTREE_TRIGGER_GC;
1008         struct bucket_mark old, new;
1009         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
1010         struct bucket *g = PTR_BUCKET(ca, &p.ptr, gc);
1011         u8 bucket_data_type;
1012         u64 v;
1013         int ret;
1014
1015         v = atomic64_read(&g->_mark.v);
1016         do {
1017                 new.v.counter = old.v.counter = v;
1018                 bucket_data_type = new.data_type;
1019
1020                 ret = __mark_pointer(c, k, p, sectors, data_type, new.gen,
1021                                      &bucket_data_type,
1022                                      &new.dirty_sectors,
1023                                      &new.cached_sectors);
1024                 if (ret)
1025                         return ret;
1026
1027                 new.data_type = bucket_data_type;
1028
1029                 if (journal_seq) {
1030                         new.journal_seq_valid = 1;
1031                         new.journal_seq = journal_seq;
1032                 }
1033
1034                 if (flags & BTREE_TRIGGER_NOATOMIC) {
1035                         g->_mark = new;
1036                         break;
1037                 }
1038         } while ((v = atomic64_cmpxchg(&g->_mark.v,
1039                               old.v.counter,
1040                               new.v.counter)) != old.v.counter);
1041
1042         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
1043
1044         BUG_ON(!gc && bucket_became_unavailable(old, new));
1045
1046         return 0;
1047 }
1048
1049 static int bch2_mark_stripe_ptr(struct bch_fs *c,
1050                                 struct bch_extent_stripe_ptr p,
1051                                 enum bch_data_type data_type,
1052                                 struct bch_fs_usage *fs_usage,
1053                                 s64 sectors, unsigned flags,
1054                                 struct bch_replicas_padded *r,
1055                                 unsigned *nr_data,
1056                                 unsigned *nr_parity)
1057 {
1058         bool gc = flags & BTREE_TRIGGER_GC;
1059         struct stripe *m;
1060         unsigned old, new;
1061         int blocks_nonempty_delta;
1062
1063         m = genradix_ptr(&c->stripes[gc], p.idx);
1064
1065         spin_lock(&c->ec_stripes_heap_lock);
1066
1067         if (!m || !m->alive) {
1068                 spin_unlock(&c->ec_stripes_heap_lock);
1069                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
1070                                     (u64) p.idx);
1071                 return -EIO;
1072         }
1073
1074         BUG_ON(m->r.e.data_type != data_type);
1075
1076         *nr_data        = m->nr_blocks - m->nr_redundant;
1077         *nr_parity      = m->nr_redundant;
1078         *r = m->r;
1079
1080         old = m->block_sectors[p.block];
1081         m->block_sectors[p.block] += sectors;
1082         new = m->block_sectors[p.block];
1083
1084         blocks_nonempty_delta = (int) !!new - (int) !!old;
1085         if (blocks_nonempty_delta) {
1086                 m->blocks_nonempty += blocks_nonempty_delta;
1087
1088                 if (!gc)
1089                         bch2_stripes_heap_update(c, m, p.idx);
1090         }
1091
1092         m->dirty = true;
1093
1094         spin_unlock(&c->ec_stripes_heap_lock);
1095
1096         return 0;
1097 }
1098
1099 static int bch2_mark_extent(struct bch_fs *c, struct bkey_s_c k,
1100                             unsigned offset, s64 sectors,
1101                             enum bch_data_type data_type,
1102                             struct bch_fs_usage *fs_usage,
1103                             unsigned journal_seq, unsigned flags)
1104 {
1105         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1106         const union bch_extent_entry *entry;
1107         struct extent_ptr_decoded p;
1108         struct bch_replicas_padded r;
1109         s64 dirty_sectors = 0;
1110         bool stale;
1111         int ret;
1112
1113         r.e.data_type   = data_type;
1114         r.e.nr_devs     = 0;
1115         r.e.nr_required = 1;
1116
1117         BUG_ON(!sectors);
1118
1119         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1120                 s64 disk_sectors = data_type == BCH_DATA_BTREE
1121                         ? sectors
1122                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1123
1124                 ret = bch2_mark_pointer(c, k, p, disk_sectors, data_type,
1125                                         fs_usage, journal_seq, flags);
1126                 if (ret < 0)
1127                         return ret;
1128
1129                 stale = ret > 0;
1130
1131                 if (p.ptr.cached) {
1132                         if (!stale)
1133                                 update_cached_sectors(c, fs_usage, p.ptr.dev,
1134                                                       disk_sectors);
1135                 } else if (!p.has_ec) {
1136                         dirty_sectors          += disk_sectors;
1137                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1138                 } else {
1139                         struct bch_replicas_padded ec_r;
1140                         unsigned nr_data, nr_parity;
1141                         s64 parity_sectors;
1142
1143                         ret = bch2_mark_stripe_ptr(c, p.ec, data_type,
1144                                         fs_usage, disk_sectors, flags,
1145                                         &ec_r, &nr_data, &nr_parity);
1146                         if (ret)
1147                                 return ret;
1148
1149                         parity_sectors =
1150                                 __ptr_disk_sectors_delta(p.crc.live_size,
1151                                         offset, sectors, flags,
1152                                         p.crc.compressed_size * nr_parity,
1153                                         p.crc.uncompressed_size * nr_data);
1154
1155                         update_replicas(c, fs_usage, &ec_r.e,
1156                                         disk_sectors + parity_sectors);
1157
1158                         /*
1159                          * There may be other dirty pointers in this extent, but
1160                          * if so they're not required for mounting if we have an
1161                          * erasure coded pointer in this extent:
1162                          */
1163                         r.e.nr_required = 0;
1164                 }
1165         }
1166
1167         if (r.e.nr_devs)
1168                 update_replicas(c, fs_usage, &r.e, dirty_sectors);
1169
1170         return 0;
1171 }
1172
1173 static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
1174                             struct bch_fs_usage *fs_usage,
1175                             u64 journal_seq, unsigned flags)
1176 {
1177         bool gc = flags & BTREE_TRIGGER_GC;
1178         struct bkey_s_c_stripe s = bkey_s_c_to_stripe(k);
1179         size_t idx = s.k->p.offset;
1180         struct stripe *m = genradix_ptr(&c->stripes[gc], idx);
1181         unsigned i;
1182
1183         spin_lock(&c->ec_stripes_heap_lock);
1184
1185         if (!m || ((flags & BTREE_TRIGGER_OVERWRITE) && !m->alive)) {
1186                 spin_unlock(&c->ec_stripes_heap_lock);
1187                 bch_err_ratelimited(c, "error marking nonexistent stripe %zu",
1188                                     idx);
1189                 return -1;
1190         }
1191
1192         if (!(flags & BTREE_TRIGGER_OVERWRITE)) {
1193                 m->sectors      = le16_to_cpu(s.v->sectors);
1194                 m->algorithm    = s.v->algorithm;
1195                 m->nr_blocks    = s.v->nr_blocks;
1196                 m->nr_redundant = s.v->nr_redundant;
1197
1198                 bch2_bkey_to_replicas(&m->r.e, k);
1199
1200                 /*
1201                  * XXX: account for stripes somehow here
1202                  */
1203 #if 0
1204                 update_replicas(c, fs_usage, &m->r.e, stripe_sectors);
1205 #endif
1206
1207                 /* gc recalculates these fields: */
1208                 if (!(flags & BTREE_TRIGGER_GC)) {
1209                         for (i = 0; i < s.v->nr_blocks; i++) {
1210                                 m->block_sectors[i] =
1211                                         stripe_blockcount_get(s.v, i);
1212                                 m->blocks_nonempty += !!m->block_sectors[i];
1213                         }
1214                 }
1215
1216                 if (!gc)
1217                         bch2_stripes_heap_update(c, m, idx);
1218                 m->alive        = true;
1219         } else {
1220                 if (!gc)
1221                         bch2_stripes_heap_del(c, m, idx);
1222                 memset(m, 0, sizeof(*m));
1223         }
1224
1225         spin_unlock(&c->ec_stripes_heap_lock);
1226
1227         bucket_set_stripe(c, s.v, fs_usage, 0, flags);
1228         return 0;
1229 }
1230
1231 static int bch2_mark_key_locked(struct bch_fs *c,
1232                    struct bkey_s_c k,
1233                    unsigned offset, s64 sectors,
1234                    struct bch_fs_usage *fs_usage,
1235                    u64 journal_seq, unsigned flags)
1236 {
1237         int ret = 0;
1238
1239         preempt_disable();
1240
1241         if (!fs_usage || (flags & BTREE_TRIGGER_GC))
1242                 fs_usage = fs_usage_ptr(c, journal_seq,
1243                                         flags & BTREE_TRIGGER_GC);
1244
1245         switch (k.k->type) {
1246         case KEY_TYPE_alloc:
1247                 ret = bch2_mark_alloc(c, k, fs_usage, journal_seq, flags);
1248                 break;
1249         case KEY_TYPE_btree_ptr:
1250         case KEY_TYPE_btree_ptr_v2:
1251                 sectors = !(flags & BTREE_TRIGGER_OVERWRITE)
1252                         ?  c->opts.btree_node_size
1253                         : -c->opts.btree_node_size;
1254
1255                 ret = bch2_mark_extent(c, k, offset, sectors, BCH_DATA_BTREE,
1256                                 fs_usage, journal_seq, flags);
1257                 break;
1258         case KEY_TYPE_extent:
1259         case KEY_TYPE_reflink_v:
1260                 ret = bch2_mark_extent(c, k, offset, sectors, BCH_DATA_USER,
1261                                 fs_usage, journal_seq, flags);
1262                 break;
1263         case KEY_TYPE_stripe:
1264                 ret = bch2_mark_stripe(c, k, fs_usage, journal_seq, flags);
1265                 break;
1266         case KEY_TYPE_inode:
1267                 if (!(flags & BTREE_TRIGGER_OVERWRITE))
1268                         fs_usage->nr_inodes++;
1269                 else
1270                         fs_usage->nr_inodes--;
1271                 break;
1272         case KEY_TYPE_reservation: {
1273                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1274
1275                 sectors *= replicas;
1276                 replicas = clamp_t(unsigned, replicas, 1,
1277                                    ARRAY_SIZE(fs_usage->persistent_reserved));
1278
1279                 fs_usage->reserved                              += sectors;
1280                 fs_usage->persistent_reserved[replicas - 1]     += sectors;
1281                 break;
1282         }
1283         }
1284
1285         preempt_enable();
1286
1287         return ret;
1288 }
1289
1290 int bch2_mark_key(struct bch_fs *c, struct bkey_s_c k,
1291                   unsigned offset, s64 sectors,
1292                   struct bch_fs_usage *fs_usage,
1293                   u64 journal_seq, unsigned flags)
1294 {
1295         int ret;
1296
1297         percpu_down_read(&c->mark_lock);
1298         ret = bch2_mark_key_locked(c, k, offset, sectors,
1299                                    fs_usage, journal_seq, flags);
1300         percpu_up_read(&c->mark_lock);
1301
1302         return ret;
1303 }
1304
1305 inline int bch2_mark_overwrite(struct btree_trans *trans,
1306                                struct btree_iter *iter,
1307                                struct bkey_s_c old,
1308                                struct bkey_i *new,
1309                                struct bch_fs_usage *fs_usage,
1310                                unsigned flags,
1311                                bool is_extents)
1312 {
1313         struct bch_fs           *c = trans->c;
1314         unsigned                offset = 0;
1315         s64                     sectors = -((s64) old.k->size);
1316
1317         flags |= BTREE_TRIGGER_OVERWRITE;
1318
1319         if (is_extents
1320             ? bkey_cmp(new->k.p, bkey_start_pos(old.k)) <= 0
1321             : bkey_cmp(new->k.p, old.k->p))
1322                 return 0;
1323
1324         if (is_extents) {
1325                 switch (bch2_extent_overlap(&new->k, old.k)) {
1326                 case BCH_EXTENT_OVERLAP_ALL:
1327                         offset = 0;
1328                         sectors = -((s64) old.k->size);
1329                         break;
1330                 case BCH_EXTENT_OVERLAP_BACK:
1331                         offset = bkey_start_offset(&new->k) -
1332                                 bkey_start_offset(old.k);
1333                         sectors = bkey_start_offset(&new->k) -
1334                                 old.k->p.offset;
1335                         break;
1336                 case BCH_EXTENT_OVERLAP_FRONT:
1337                         offset = 0;
1338                         sectors = bkey_start_offset(old.k) -
1339                                 new->k.p.offset;
1340                         break;
1341                 case BCH_EXTENT_OVERLAP_MIDDLE:
1342                         offset = bkey_start_offset(&new->k) -
1343                                 bkey_start_offset(old.k);
1344                         sectors = -((s64) new->k.size);
1345                         flags |= BTREE_TRIGGER_OVERWRITE_SPLIT;
1346                         break;
1347                 }
1348
1349                 BUG_ON(sectors >= 0);
1350         }
1351
1352         return bch2_mark_key_locked(c, old, offset, sectors, fs_usage,
1353                                     trans->journal_res.seq, flags) ?: 1;
1354 }
1355
1356 int bch2_mark_update(struct btree_trans *trans,
1357                      struct btree_iter *iter,
1358                      struct bkey_i *insert,
1359                      struct bch_fs_usage *fs_usage,
1360                      unsigned flags)
1361 {
1362         struct bch_fs           *c = trans->c;
1363         struct btree            *b = iter->l[0].b;
1364         struct btree_node_iter  node_iter = iter->l[0].iter;
1365         struct bkey_packed      *_k;
1366         int ret = 0;
1367
1368         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1369                 return 0;
1370
1371         if (!btree_node_type_needs_gc(iter->btree_id))
1372                 return 0;
1373
1374         bch2_mark_key_locked(c, bkey_i_to_s_c(insert),
1375                 0, insert->k.size,
1376                 fs_usage, trans->journal_res.seq,
1377                 BTREE_TRIGGER_INSERT|flags);
1378
1379         if (unlikely(flags & BTREE_TRIGGER_NOOVERWRITES))
1380                 return 0;
1381
1382         /*
1383          * For non extents, we only mark the new key, not the key being
1384          * overwritten - unless we're actually deleting:
1385          */
1386         if ((iter->btree_id == BTREE_ID_ALLOC ||
1387              iter->btree_id == BTREE_ID_EC) &&
1388             !bkey_deleted(&insert->k))
1389                 return 0;
1390
1391         while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
1392                 struct bkey             unpacked;
1393                 struct bkey_s_c         k = bkey_disassemble(b, _k, &unpacked);
1394
1395                 ret = bch2_mark_overwrite(trans, iter, k, insert,
1396                                           fs_usage, flags,
1397                                           btree_node_type_is_extents(iter->btree_id));
1398                 if (ret <= 0)
1399                         break;
1400
1401                 bch2_btree_node_iter_advance(&node_iter, b);
1402         }
1403
1404         return ret;
1405 }
1406
1407 void bch2_trans_fs_usage_apply(struct btree_trans *trans,
1408                                struct bch_fs_usage *fs_usage)
1409 {
1410         struct bch_fs *c = trans->c;
1411         struct btree_insert_entry *i;
1412         static int warned_disk_usage = 0;
1413         u64 disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1414         char buf[200];
1415
1416         if (!bch2_fs_usage_apply(c, fs_usage, trans->disk_res,
1417                                  trans->journal_res.seq) ||
1418             warned_disk_usage ||
1419             xchg(&warned_disk_usage, 1))
1420                 return;
1421
1422         bch_err(c, "disk usage increased more than %llu sectors reserved",
1423                 disk_res_sectors);
1424
1425         trans_for_each_update(trans, i) {
1426                 struct btree_iter       *iter = i->iter;
1427                 struct btree            *b = iter->l[0].b;
1428                 struct btree_node_iter  node_iter = iter->l[0].iter;
1429                 struct bkey_packed      *_k;
1430
1431                 pr_err("while inserting");
1432                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1433                 pr_err("%s", buf);
1434                 pr_err("overlapping with");
1435
1436                 node_iter = iter->l[0].iter;
1437                 while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
1438                         struct bkey             unpacked;
1439                         struct bkey_s_c         k;
1440
1441                         k = bkey_disassemble(b, _k, &unpacked);
1442
1443                         if (btree_node_is_extents(b)
1444                             ? bkey_cmp(i->k->k.p, bkey_start_pos(k.k)) <= 0
1445                             : bkey_cmp(i->k->k.p, k.k->p))
1446                                 break;
1447
1448                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
1449                         pr_err("%s", buf);
1450
1451                         bch2_btree_node_iter_advance(&node_iter, b);
1452                 }
1453         }
1454 }
1455
1456 /* trans_mark: */
1457
1458 static int trans_get_key(struct btree_trans *trans,
1459                          enum btree_id btree_id, struct bpos pos,
1460                          struct btree_iter **iter,
1461                          struct bkey_s_c *k)
1462 {
1463         struct btree_insert_entry *i;
1464         int ret;
1465
1466         trans_for_each_update(trans, i)
1467                 if (i->iter->btree_id == btree_id &&
1468                     (btree_node_type_is_extents(btree_id)
1469                      ? bkey_cmp(pos, bkey_start_pos(&i->k->k)) >= 0 &&
1470                        bkey_cmp(pos, i->k->k.p) < 0
1471                      : !bkey_cmp(pos, i->iter->pos))) {
1472                         *iter   = i->iter;
1473                         *k      = bkey_i_to_s_c(i->k);
1474                         return 1;
1475                 }
1476
1477         *iter = bch2_trans_get_iter(trans, btree_id, pos,
1478                                     BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1479         if (IS_ERR(*iter))
1480                 return PTR_ERR(*iter);
1481
1482         *k = bch2_btree_iter_peek_slot(*iter);
1483         ret = bkey_err(*k);
1484         if (ret)
1485                 bch2_trans_iter_put(trans, *iter);
1486         return ret;
1487 }
1488
1489 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1490                         struct bkey_s_c k, struct extent_ptr_decoded p,
1491                         s64 sectors, enum bch_data_type data_type)
1492 {
1493         struct bch_fs *c = trans->c;
1494         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
1495         struct btree_iter *iter;
1496         struct bkey_s_c k_a;
1497         struct bkey_alloc_unpacked u;
1498         struct bkey_i_alloc *a;
1499         int ret;
1500
1501         ret = trans_get_key(trans, BTREE_ID_ALLOC,
1502                             POS(p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr)),
1503                             &iter, &k_a);
1504         if (ret < 0)
1505                 return ret;
1506
1507         if (k_a.k->type != KEY_TYPE_alloc ||
1508             (!ret && unlikely(!test_bit(BCH_FS_ALLOC_WRITTEN, &c->flags)))) {
1509                 /*
1510                  * During journal replay, and if gc repairs alloc info at
1511                  * runtime, the alloc info in the btree might not be up to date
1512                  * yet - so, trust the in memory mark - unless we're already
1513                  * updating that key:
1514                  */
1515                 struct bucket *g;
1516                 struct bucket_mark m;
1517
1518                 percpu_down_read(&c->mark_lock);
1519                 g       = bucket(ca, iter->pos.offset);
1520                 m       = READ_ONCE(g->mark);
1521                 u       = alloc_mem_to_key(g, m);
1522                 percpu_up_read(&c->mark_lock);
1523         } else {
1524                 u = bch2_alloc_unpack(k_a);
1525         }
1526
1527         ret = __mark_pointer(c, k, p, sectors, data_type, u.gen, &u.data_type,
1528                              &u.dirty_sectors, &u.cached_sectors);
1529         if (ret)
1530                 goto out;
1531
1532         a = bch2_trans_kmalloc(trans, BKEY_ALLOC_U64s_MAX * 8);
1533         ret = PTR_ERR_OR_ZERO(a);
1534         if (ret)
1535                 goto out;
1536
1537         bkey_alloc_init(&a->k_i);
1538         a->k.p = iter->pos;
1539         bch2_alloc_pack(a, u);
1540         bch2_trans_update(trans, iter, &a->k_i, 0);
1541 out:
1542         bch2_trans_iter_put(trans, iter);
1543         return ret;
1544 }
1545
1546 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1547                         struct bch_extent_stripe_ptr p,
1548                         s64 sectors, enum bch_data_type data_type,
1549                         struct bch_replicas_padded *r,
1550                         unsigned *nr_data,
1551                         unsigned *nr_parity)
1552 {
1553         struct bch_fs *c = trans->c;
1554         struct btree_iter *iter;
1555         struct bkey_s_c k;
1556         struct bkey_i_stripe *s;
1557         int ret = 0;
1558
1559         ret = trans_get_key(trans, BTREE_ID_EC, POS(0, p.idx), &iter, &k);
1560         if (ret < 0)
1561                 return ret;
1562
1563         if (k.k->type != KEY_TYPE_stripe) {
1564                 bch2_fs_inconsistent(c,
1565                         "pointer to nonexistent stripe %llu",
1566                         (u64) p.idx);
1567                 ret = -EIO;
1568                 goto out;
1569         }
1570
1571         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1572         ret = PTR_ERR_OR_ZERO(s);
1573         if (ret)
1574                 goto out;
1575
1576         bkey_reassemble(&s->k_i, k);
1577
1578         stripe_blockcount_set(&s->v, p.block,
1579                 stripe_blockcount_get(&s->v, p.block) +
1580                 sectors);
1581
1582         *nr_data        = s->v.nr_blocks - s->v.nr_redundant;
1583         *nr_parity      = s->v.nr_redundant;
1584         bch2_bkey_to_replicas(&r->e, bkey_i_to_s_c(&s->k_i));
1585         bch2_trans_update(trans, iter, &s->k_i, 0);
1586 out:
1587         bch2_trans_iter_put(trans, iter);
1588         return ret;
1589 }
1590
1591 static int bch2_trans_mark_extent(struct btree_trans *trans,
1592                         struct bkey_s_c k, unsigned offset,
1593                         s64 sectors, unsigned flags,
1594                         enum bch_data_type data_type)
1595 {
1596         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1597         const union bch_extent_entry *entry;
1598         struct extent_ptr_decoded p;
1599         struct bch_replicas_padded r;
1600         s64 dirty_sectors = 0;
1601         bool stale;
1602         int ret;
1603
1604         r.e.data_type   = data_type;
1605         r.e.nr_devs     = 0;
1606         r.e.nr_required = 1;
1607
1608         BUG_ON(!sectors);
1609
1610         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1611                 s64 disk_sectors = data_type == BCH_DATA_BTREE
1612                         ? sectors
1613                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1614
1615                 ret = bch2_trans_mark_pointer(trans, k, p, disk_sectors,
1616                                               data_type);
1617                 if (ret < 0)
1618                         return ret;
1619
1620                 stale = ret > 0;
1621
1622                 if (p.ptr.cached) {
1623                         if (!stale)
1624                                 update_cached_sectors_list(trans, p.ptr.dev,
1625                                                            disk_sectors);
1626                 } else if (!p.has_ec) {
1627                         dirty_sectors          += disk_sectors;
1628                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1629                 } else {
1630                         struct bch_replicas_padded ec_r;
1631                         unsigned nr_data, nr_parity;
1632                         s64 parity_sectors;
1633
1634                         ret = bch2_trans_mark_stripe_ptr(trans, p.ec,
1635                                         disk_sectors, data_type,
1636                                         &ec_r, &nr_data, &nr_parity);
1637                         if (ret)
1638                                 return ret;
1639
1640                         parity_sectors =
1641                                 __ptr_disk_sectors_delta(p.crc.live_size,
1642                                         offset, sectors, flags,
1643                                         p.crc.compressed_size * nr_parity,
1644                                         p.crc.uncompressed_size * nr_data);
1645
1646                         update_replicas_list(trans, &ec_r.e,
1647                                              disk_sectors + parity_sectors);
1648
1649                         r.e.nr_required = 0;
1650                 }
1651         }
1652
1653         if (r.e.nr_devs)
1654                 update_replicas_list(trans, &r.e, dirty_sectors);
1655
1656         return 0;
1657 }
1658
1659 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1660                         struct bkey_s_c_reflink_p p,
1661                         u64 idx, unsigned sectors,
1662                         unsigned flags)
1663 {
1664         struct bch_fs *c = trans->c;
1665         struct btree_iter *iter;
1666         struct bkey_s_c k;
1667         struct bkey_i_reflink_v *r_v;
1668         s64 ret;
1669
1670         ret = trans_get_key(trans, BTREE_ID_REFLINK,
1671                             POS(0, idx), &iter, &k);
1672         if (ret < 0)
1673                 return ret;
1674
1675         if (k.k->type != KEY_TYPE_reflink_v) {
1676                 bch2_fs_inconsistent(c,
1677                         "%llu:%llu len %u points to nonexistent indirect extent %llu",
1678                         p.k->p.inode, p.k->p.offset, p.k->size, idx);
1679                 ret = -EIO;
1680                 goto err;
1681         }
1682
1683         if ((flags & BTREE_TRIGGER_OVERWRITE) &&
1684             (bkey_start_offset(k.k) < idx ||
1685              k.k->p.offset > idx + sectors))
1686                 goto out;
1687
1688         sectors = k.k->p.offset - idx;
1689
1690         r_v = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1691         ret = PTR_ERR_OR_ZERO(r_v);
1692         if (ret)
1693                 goto err;
1694
1695         bkey_reassemble(&r_v->k_i, k);
1696
1697         le64_add_cpu(&r_v->v.refcount,
1698                      !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1);
1699
1700         if (!r_v->v.refcount) {
1701                 r_v->k.type = KEY_TYPE_deleted;
1702                 set_bkey_val_u64s(&r_v->k, 0);
1703         }
1704
1705         bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
1706         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
1707
1708         bch2_trans_update(trans, iter, &r_v->k_i, 0);
1709 out:
1710         ret = sectors;
1711 err:
1712         bch2_trans_iter_put(trans, iter);
1713         return ret;
1714 }
1715
1716 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1717                         struct bkey_s_c_reflink_p p, unsigned offset,
1718                         s64 sectors, unsigned flags)
1719 {
1720         u64 idx = le64_to_cpu(p.v->idx) + offset;
1721         s64 ret = 0;
1722
1723         sectors = abs(sectors);
1724         BUG_ON(offset + sectors > p.k->size);
1725
1726         while (sectors) {
1727                 ret = __bch2_trans_mark_reflink_p(trans, p, idx, sectors, flags);
1728                 if (ret < 0)
1729                         break;
1730
1731                 idx += ret;
1732                 sectors = max_t(s64, 0LL, sectors - ret);
1733                 ret = 0;
1734         }
1735
1736         return ret;
1737 }
1738
1739 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c k,
1740                         unsigned offset, s64 sectors, unsigned flags)
1741 {
1742         struct replicas_delta_list *d;
1743         struct bch_fs *c = trans->c;
1744
1745         switch (k.k->type) {
1746         case KEY_TYPE_btree_ptr:
1747         case KEY_TYPE_btree_ptr_v2:
1748                 sectors = !(flags & BTREE_TRIGGER_OVERWRITE)
1749                         ?  c->opts.btree_node_size
1750                         : -c->opts.btree_node_size;
1751
1752                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1753                                               flags, BCH_DATA_BTREE);
1754         case KEY_TYPE_extent:
1755         case KEY_TYPE_reflink_v:
1756                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1757                                               flags, BCH_DATA_USER);
1758         case KEY_TYPE_inode:
1759                 d = replicas_deltas_realloc(trans, 0);
1760
1761                 if (!(flags & BTREE_TRIGGER_OVERWRITE))
1762                         d->nr_inodes++;
1763                 else
1764                         d->nr_inodes--;
1765                 return 0;
1766         case KEY_TYPE_reservation: {
1767                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1768
1769                 d = replicas_deltas_realloc(trans, 0);
1770
1771                 sectors *= replicas;
1772                 replicas = clamp_t(unsigned, replicas, 1,
1773                                    ARRAY_SIZE(d->persistent_reserved));
1774
1775                 d->persistent_reserved[replicas - 1] += sectors;
1776                 return 0;
1777         }
1778         case KEY_TYPE_reflink_p:
1779                 return bch2_trans_mark_reflink_p(trans,
1780                                         bkey_s_c_to_reflink_p(k),
1781                                         offset, sectors, flags);
1782         default:
1783                 return 0;
1784         }
1785 }
1786
1787 int bch2_trans_mark_update(struct btree_trans *trans,
1788                            struct btree_iter *iter,
1789                            struct bkey_i *insert,
1790                            unsigned flags)
1791 {
1792         struct btree            *b = iter->l[0].b;
1793         struct btree_node_iter  node_iter = iter->l[0].iter;
1794         struct bkey_packed      *_k;
1795         int ret;
1796
1797         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1798                 return 0;
1799
1800         if (!btree_node_type_needs_gc(iter->btree_id))
1801                 return 0;
1802
1803         ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(insert),
1804                         0, insert->k.size, BTREE_TRIGGER_INSERT);
1805         if (ret)
1806                 return ret;
1807
1808         if (unlikely(flags & BTREE_TRIGGER_NOOVERWRITES))
1809                 return 0;
1810
1811         while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
1812                 struct bkey             unpacked;
1813                 struct bkey_s_c         k;
1814                 unsigned                offset = 0;
1815                 s64                     sectors = 0;
1816                 unsigned                flags = BTREE_TRIGGER_OVERWRITE;
1817
1818                 k = bkey_disassemble(b, _k, &unpacked);
1819
1820                 if (btree_node_is_extents(b)
1821                     ? bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0
1822                     : bkey_cmp(insert->k.p, k.k->p))
1823                         break;
1824
1825                 if (btree_node_is_extents(b)) {
1826                         switch (bch2_extent_overlap(&insert->k, k.k)) {
1827                         case BCH_EXTENT_OVERLAP_ALL:
1828                                 offset = 0;
1829                                 sectors = -((s64) k.k->size);
1830                                 break;
1831                         case BCH_EXTENT_OVERLAP_BACK:
1832                                 offset = bkey_start_offset(&insert->k) -
1833                                         bkey_start_offset(k.k);
1834                                 sectors = bkey_start_offset(&insert->k) -
1835                                         k.k->p.offset;
1836                                 break;
1837                         case BCH_EXTENT_OVERLAP_FRONT:
1838                                 offset = 0;
1839                                 sectors = bkey_start_offset(k.k) -
1840                                         insert->k.p.offset;
1841                                 break;
1842                         case BCH_EXTENT_OVERLAP_MIDDLE:
1843                                 offset = bkey_start_offset(&insert->k) -
1844                                         bkey_start_offset(k.k);
1845                                 sectors = -((s64) insert->k.size);
1846                                 flags |= BTREE_TRIGGER_OVERWRITE_SPLIT;
1847                                 break;
1848                         }
1849
1850                         BUG_ON(sectors >= 0);
1851                 }
1852
1853                 ret = bch2_trans_mark_key(trans, k, offset, sectors, flags);
1854                 if (ret)
1855                         return ret;
1856
1857                 bch2_btree_node_iter_advance(&node_iter, b);
1858         }
1859
1860         return 0;
1861 }
1862
1863 /* Disk reservations: */
1864
1865 static u64 bch2_recalc_sectors_available(struct bch_fs *c)
1866 {
1867         percpu_u64_set(&c->pcpu->sectors_available, 0);
1868
1869         return avail_factor(__bch2_fs_usage_read_short(c).free);
1870 }
1871
1872 void __bch2_disk_reservation_put(struct bch_fs *c, struct disk_reservation *res)
1873 {
1874         percpu_down_read(&c->mark_lock);
1875         this_cpu_sub(c->usage[0]->online_reserved,
1876                      res->sectors);
1877         percpu_up_read(&c->mark_lock);
1878
1879         res->sectors = 0;
1880 }
1881
1882 #define SECTORS_CACHE   1024
1883
1884 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1885                               unsigned sectors, int flags)
1886 {
1887         struct bch_fs_pcpu *pcpu;
1888         u64 old, v, get;
1889         s64 sectors_available;
1890         int ret;
1891
1892         percpu_down_read(&c->mark_lock);
1893         preempt_disable();
1894         pcpu = this_cpu_ptr(c->pcpu);
1895
1896         if (sectors <= pcpu->sectors_available)
1897                 goto out;
1898
1899         v = atomic64_read(&c->sectors_available);
1900         do {
1901                 old = v;
1902                 get = min((u64) sectors + SECTORS_CACHE, old);
1903
1904                 if (get < sectors) {
1905                         preempt_enable();
1906                         percpu_up_read(&c->mark_lock);
1907                         goto recalculate;
1908                 }
1909         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1910                                        old, old - get)) != old);
1911
1912         pcpu->sectors_available         += get;
1913
1914 out:
1915         pcpu->sectors_available         -= sectors;
1916         this_cpu_add(c->usage[0]->online_reserved, sectors);
1917         res->sectors                    += sectors;
1918
1919         preempt_enable();
1920         percpu_up_read(&c->mark_lock);
1921         return 0;
1922
1923 recalculate:
1924         percpu_down_write(&c->mark_lock);
1925
1926         sectors_available = bch2_recalc_sectors_available(c);
1927
1928         if (sectors <= sectors_available ||
1929             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1930                 atomic64_set(&c->sectors_available,
1931                              max_t(s64, 0, sectors_available - sectors));
1932                 this_cpu_add(c->usage[0]->online_reserved, sectors);
1933                 res->sectors                    += sectors;
1934                 ret = 0;
1935         } else {
1936                 atomic64_set(&c->sectors_available, sectors_available);
1937                 ret = -ENOSPC;
1938         }
1939
1940         percpu_up_write(&c->mark_lock);
1941
1942         return ret;
1943 }
1944
1945 /* Startup/shutdown: */
1946
1947 static void buckets_free_rcu(struct rcu_head *rcu)
1948 {
1949         struct bucket_array *buckets =
1950                 container_of(rcu, struct bucket_array, rcu);
1951
1952         kvpfree(buckets,
1953                 sizeof(struct bucket_array) +
1954                 buckets->nbuckets * sizeof(struct bucket));
1955 }
1956
1957 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
1958 {
1959         struct bucket_array *buckets = NULL, *old_buckets = NULL;
1960         unsigned long *buckets_nouse = NULL;
1961         alloc_fifo      free[RESERVE_NR];
1962         alloc_fifo      free_inc;
1963         alloc_heap      alloc_heap;
1964         copygc_heap     copygc_heap;
1965
1966         size_t btree_reserve    = DIV_ROUND_UP(BTREE_NODE_RESERVE,
1967                              ca->mi.bucket_size / c->opts.btree_node_size);
1968         /* XXX: these should be tunable */
1969         size_t reserve_none     = max_t(size_t, 1, nbuckets >> 9);
1970         size_t copygc_reserve   = max_t(size_t, 2, nbuckets >> 7);
1971         size_t free_inc_nr      = max(max_t(size_t, 1, nbuckets >> 12),
1972                                       btree_reserve * 2);
1973         bool resize = ca->buckets[0] != NULL,
1974              start_copygc = ca->copygc_thread != NULL;
1975         int ret = -ENOMEM;
1976         unsigned i;
1977
1978         memset(&free,           0, sizeof(free));
1979         memset(&free_inc,       0, sizeof(free_inc));
1980         memset(&alloc_heap,     0, sizeof(alloc_heap));
1981         memset(&copygc_heap,    0, sizeof(copygc_heap));
1982
1983         if (!(buckets           = kvpmalloc(sizeof(struct bucket_array) +
1984                                             nbuckets * sizeof(struct bucket),
1985                                             GFP_KERNEL|__GFP_ZERO)) ||
1986             !(buckets_nouse     = kvpmalloc(BITS_TO_LONGS(nbuckets) *
1987                                             sizeof(unsigned long),
1988                                             GFP_KERNEL|__GFP_ZERO)) ||
1989             !init_fifo(&free[RESERVE_BTREE], btree_reserve, GFP_KERNEL) ||
1990             !init_fifo(&free[RESERVE_MOVINGGC],
1991                        copygc_reserve, GFP_KERNEL) ||
1992             !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
1993             !init_fifo(&free_inc,       free_inc_nr, GFP_KERNEL) ||
1994             !init_heap(&alloc_heap,     ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL) ||
1995             !init_heap(&copygc_heap,    copygc_reserve, GFP_KERNEL))
1996                 goto err;
1997
1998         buckets->first_bucket   = ca->mi.first_bucket;
1999         buckets->nbuckets       = nbuckets;
2000
2001         bch2_copygc_stop(ca);
2002
2003         if (resize) {
2004                 down_write(&c->gc_lock);
2005                 down_write(&ca->bucket_lock);
2006                 percpu_down_write(&c->mark_lock);
2007         }
2008
2009         old_buckets = bucket_array(ca);
2010
2011         if (resize) {
2012                 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
2013
2014                 memcpy(buckets->b,
2015                        old_buckets->b,
2016                        n * sizeof(struct bucket));
2017                 memcpy(buckets_nouse,
2018                        ca->buckets_nouse,
2019                        BITS_TO_LONGS(n) * sizeof(unsigned long));
2020         }
2021
2022         rcu_assign_pointer(ca->buckets[0], buckets);
2023         buckets = old_buckets;
2024
2025         swap(ca->buckets_nouse, buckets_nouse);
2026
2027         if (resize)
2028                 percpu_up_write(&c->mark_lock);
2029
2030         spin_lock(&c->freelist_lock);
2031         for (i = 0; i < RESERVE_NR; i++) {
2032                 fifo_move(&free[i], &ca->free[i]);
2033                 swap(ca->free[i], free[i]);
2034         }
2035         fifo_move(&free_inc, &ca->free_inc);
2036         swap(ca->free_inc, free_inc);
2037         spin_unlock(&c->freelist_lock);
2038
2039         /* with gc lock held, alloc_heap can't be in use: */
2040         swap(ca->alloc_heap, alloc_heap);
2041
2042         /* and we shut down copygc: */
2043         swap(ca->copygc_heap, copygc_heap);
2044
2045         nbuckets = ca->mi.nbuckets;
2046
2047         if (resize) {
2048                 up_write(&ca->bucket_lock);
2049                 up_write(&c->gc_lock);
2050         }
2051
2052         if (start_copygc &&
2053             bch2_copygc_start(c, ca))
2054                 bch_err(ca, "error restarting copygc thread");
2055
2056         ret = 0;
2057 err:
2058         free_heap(&copygc_heap);
2059         free_heap(&alloc_heap);
2060         free_fifo(&free_inc);
2061         for (i = 0; i < RESERVE_NR; i++)
2062                 free_fifo(&free[i]);
2063         kvpfree(buckets_nouse,
2064                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2065         if (buckets)
2066                 call_rcu(&old_buckets->rcu, buckets_free_rcu);
2067
2068         return ret;
2069 }
2070
2071 void bch2_dev_buckets_free(struct bch_dev *ca)
2072 {
2073         unsigned i;
2074
2075         free_heap(&ca->copygc_heap);
2076         free_heap(&ca->alloc_heap);
2077         free_fifo(&ca->free_inc);
2078         for (i = 0; i < RESERVE_NR; i++)
2079                 free_fifo(&ca->free[i]);
2080         kvpfree(ca->buckets_nouse,
2081                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2082         kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2083                 sizeof(struct bucket_array) +
2084                 ca->mi.nbuckets * sizeof(struct bucket));
2085
2086         free_percpu(ca->usage[0]);
2087 }
2088
2089 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2090 {
2091         if (!(ca->usage[0] = alloc_percpu(struct bch_dev_usage)))
2092                 return -ENOMEM;
2093
2094         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2095 }