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 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  81                                               enum bch_data_type data_type,
  82                                               s64 sectors)
  83 {
  84         switch (data_type) {
  85         case BCH_DATA_btree:
  86                 fs_usage->btree         += sectors;
  87                 break;
  88         case BCH_DATA_user:
  89         case BCH_DATA_parity:
  90                 fs_usage->data          += sectors;
  91                 break;
  92         case BCH_DATA_cached:
  93                 fs_usage->cached        += sectors;
  94                 break;
  95         default:
  96                 break;
  97         }
  98 }
  99
 100 /*
 101  * Clear journal_seq_valid for buckets for which it's not needed, to prevent
 102  * wraparound:
 103  */
 104 void bch2_bucket_seq_cleanup(struct bch_fs *c)
 105 {
 106         u64 journal_seq = atomic64_read(&c->journal.seq);
 107         u16 last_seq_ondisk = c->journal.last_seq_ondisk;
 108         struct bch_dev *ca;
 109         struct bucket_array *buckets;
 110         struct bucket *g;
 111         struct bucket_mark m;
 112         unsigned i;
 113
 114         if (journal_seq - c->last_bucket_seq_cleanup <
 115             (1U << (BUCKET_JOURNAL_SEQ_BITS - 2)))
 116                 return;
 117
 118         c->last_bucket_seq_cleanup = journal_seq;
 119
 120         for_each_member_device(ca, c, i) {
 121                 down_read(&ca->bucket_lock);
 122                 buckets = bucket_array(ca);
 123
 124                 for_each_bucket(g, buckets) {
 125                         bucket_cmpxchg(g, m, ({
 126                                 if (!m.journal_seq_valid ||
 127                                     bucket_needs_journal_commit(m, last_seq_ondisk))
 128                                         break;
 129
 130                                 m.journal_seq_valid = 0;
 131                         }));
 132                 }
 133                 up_read(&ca->bucket_lock);
 134         }
 135 }
 136
 137 void bch2_fs_usage_initialize(struct bch_fs *c)
 138 {
 139         struct bch_fs_usage *usage;
 140         unsigned i;
 141
 142         percpu_down_write(&c->mark_lock);
 143         usage = c->usage_base;
 144
 145         bch2_fs_usage_acc_to_base(c, 0);
 146         bch2_fs_usage_acc_to_base(c, 1);
 147
 148         for (i = 0; i < BCH_REPLICAS_MAX; i++)
 149                 usage->reserved += usage->persistent_reserved[i];
 150
 151         for (i = 0; i < c->replicas.nr; i++) {
 152                 struct bch_replicas_entry *e =
 153                         cpu_replicas_entry(&c->replicas, i);
 154
 155                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
 156         }
 157
 158         percpu_up_write(&c->mark_lock);
 159 }
 160
 161 void bch2_fs_usage_scratch_put(struct bch_fs *c, struct bch_fs_usage *fs_usage)
 162 {
 163         if (fs_usage == c->usage_scratch)
 164                 mutex_unlock(&c->usage_scratch_lock);
 165         else
 166                 kfree(fs_usage);
 167 }
 168
 169 struct bch_fs_usage *bch2_fs_usage_scratch_get(struct bch_fs *c)
 170 {
 171         struct bch_fs_usage *ret;
 172         unsigned bytes = fs_usage_u64s(c) * sizeof(u64);
 173
 174         ret = kzalloc(bytes, GFP_NOWAIT|__GFP_NOWARN);
 175         if (ret)
 176                 return ret;
 177
 178         if (mutex_trylock(&c->usage_scratch_lock))
 179                 goto out_pool;
 180
 181         ret = kzalloc(bytes, GFP_NOFS);
 182         if (ret)
 183                 return ret;
 184
 185         mutex_lock(&c->usage_scratch_lock);
 186 out_pool:
 187         ret = c->usage_scratch;
 188         memset(ret, 0, bytes);
 189         return ret;
 190 }
 191
 192 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
 193 {
 194         struct bch_dev_usage ret;
 195
 196         memset(&ret, 0, sizeof(ret));
 197         acc_u64s_percpu((u64 *) &ret,
 198                         (u64 __percpu *) ca->usage[0],
 199                         sizeof(ret) / sizeof(u64));
 200
 201         return ret;
 202 }
 203
 204 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 205                                                 unsigned journal_seq,
 206                                                 bool gc)
 207 {
 208         return this_cpu_ptr(gc
 209                             ? c->usage_gc
 210                             : c->usage[journal_seq & 1]);
 211 }
 212
 213 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 214 {
 215         ssize_t offset = v - (u64 *) c->usage_base;
 216         unsigned seq;
 217         u64 ret;
 218
 219         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 220         percpu_rwsem_assert_held(&c->mark_lock);
 221
 222         do {
 223                 seq = read_seqcount_begin(&c->usage_lock);
 224                 ret = *v +
 225                         percpu_u64_get((u64 __percpu *) c->usage[0] + offset) +
 226                         percpu_u64_get((u64 __percpu *) c->usage[1] + offset);
 227         } while (read_seqcount_retry(&c->usage_lock, seq));
 228
 229         return ret;
 230 }
 231
 232 struct bch_fs_usage *bch2_fs_usage_read(struct bch_fs *c)
 233 {
 234         struct bch_fs_usage *ret;
 235         unsigned seq, v, u64s = fs_usage_u64s(c);
 236 retry:
 237         ret = kmalloc(u64s * sizeof(u64), GFP_NOFS);
 238         if (unlikely(!ret))
 239                 return NULL;
 240
 241         percpu_down_read(&c->mark_lock);
 242
 243         v = fs_usage_u64s(c);
 244         if (unlikely(u64s != v)) {
 245                 u64s = v;
 246                 percpu_up_read(&c->mark_lock);
 247                 kfree(ret);
 248                 goto retry;
 249         }
 250
 251         do {
 252                 seq = read_seqcount_begin(&c->usage_lock);
 253                 memcpy(ret, c->usage_base, u64s * sizeof(u64));
 254                 acc_u64s_percpu((u64 *) ret, (u64 __percpu *) c->usage[0], u64s);
 255                 acc_u64s_percpu((u64 *) ret, (u64 __percpu *) c->usage[1], u64s);
 256         } while (read_seqcount_retry(&c->usage_lock, seq));
 257
 258         return ret;
 259 }
 260
 261 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 262 {
 263         unsigned u64s = fs_usage_u64s(c);
 264
 265         BUG_ON(idx >= 2);
 266
 267         preempt_disable();
 268         write_seqcount_begin(&c->usage_lock);
 269
 270         acc_u64s_percpu((u64 *) c->usage_base,
 271                         (u64 __percpu *) c->usage[idx], u64s);
 272         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 273
 274         write_seqcount_end(&c->usage_lock);
 275         preempt_enable();
 276 }
 277
 278 void bch2_fs_usage_to_text(struct printbuf *out,
 279                            struct bch_fs *c,
 280                            struct bch_fs_usage *fs_usage)
 281 {
 282         unsigned i;
 283
 284         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 285
 286         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 287                fs_usage->hidden);
 288         pr_buf(out, "data:\t\t\t\t%llu\n",
 289                fs_usage->data);
 290         pr_buf(out, "cached:\t\t\t\t%llu\n",
 291                fs_usage->cached);
 292         pr_buf(out, "reserved:\t\t\t%llu\n",
 293                fs_usage->reserved);
 294         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 295                fs_usage->nr_inodes);
 296         pr_buf(out, "online reserved:\t\t%llu\n",
 297                fs_usage->online_reserved);
 298
 299         for (i = 0;
 300              i < ARRAY_SIZE(fs_usage->persistent_reserved);
 301              i++) {
 302                 pr_buf(out, "%u replicas:\n", i + 1);
 303                 pr_buf(out, "\treserved:\t\t%llu\n",
 304                        fs_usage->persistent_reserved[i]);
 305         }
 306
 307         for (i = 0; i < c->replicas.nr; i++) {
 308                 struct bch_replicas_entry *e =
 309                         cpu_replicas_entry(&c->replicas, i);
 310
 311                 pr_buf(out, "\t");
 312                 bch2_replicas_entry_to_text(out, e);
 313                 pr_buf(out, ":\t%llu\n", fs_usage->replicas[i]);
 314         }
 315 }
 316
 317 #define RESERVE_FACTOR  6
 318
 319 static u64 reserve_factor(u64 r)
 320 {
 321         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 322 }
 323
 324 static u64 avail_factor(u64 r)
 325 {
 326         return div_u64(r << RESERVE_FACTOR, (1 << RESERVE_FACTOR) + 1);
 327 }
 328
 329 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage *fs_usage)
 330 {
 331         return min(fs_usage->hidden +
 332                    fs_usage->btree +
 333                    fs_usage->data +
 334                    reserve_factor(fs_usage->reserved +
 335                                   fs_usage->online_reserved),
 336                    c->capacity);
 337 }
 338
 339 static struct bch_fs_usage_short
 340 __bch2_fs_usage_read_short(struct bch_fs *c)
 341 {
 342         struct bch_fs_usage_short ret;
 343         u64 data, reserved;
 344
 345         ret.capacity = c->capacity -
 346                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 347
 348         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 349                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 350         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 351                 bch2_fs_usage_read_one(c, &c->usage_base->online_reserved);
 352
 353         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 354         ret.free        = ret.capacity - ret.used;
 355
 356         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 357
 358         return ret;
 359 }
 360
 361 struct bch_fs_usage_short
 362 bch2_fs_usage_read_short(struct bch_fs *c)
 363 {
 364         struct bch_fs_usage_short ret;
 365
 366         percpu_down_read(&c->mark_lock);
 367         ret = __bch2_fs_usage_read_short(c);
 368         percpu_up_read(&c->mark_lock);
 369
 370         return ret;
 371 }
 372
 373 static inline int is_unavailable_bucket(struct bucket_mark m)
 374 {
 375         return !is_available_bucket(m);
 376 }
 377
 378 static inline int is_fragmented_bucket(struct bucket_mark m,
 379                                        struct bch_dev *ca)
 380 {
 381         if (!m.owned_by_allocator &&
 382             m.data_type == BCH_DATA_user &&
 383             bucket_sectors_used(m))
 384                 return max_t(int, 0, (int) ca->mi.bucket_size -
 385                              bucket_sectors_used(m));
 386         return 0;
 387 }
 388
 389 static inline int is_stripe_data_bucket(struct bucket_mark m)
 390 {
 391         return m.stripe && m.data_type != BCH_DATA_parity;
 392 }
 393
 394 static inline int bucket_stripe_sectors(struct bucket_mark m)
 395 {
 396         return is_stripe_data_bucket(m) ? m.dirty_sectors : 0;
 397 }
 398
 399 static inline enum bch_data_type bucket_type(struct bucket_mark m)
 400 {
 401         return m.cached_sectors && !m.dirty_sectors
 402                 ? BCH_DATA_cached
 403                 : m.data_type;
 404 }
 405
 406 static bool bucket_became_unavailable(struct bucket_mark old,
 407                                       struct bucket_mark new)
 408 {
 409         return is_available_bucket(old) &&
 410                !is_available_bucket(new);
 411 }
 412
 413 int bch2_fs_usage_apply(struct bch_fs *c,
 414                         struct bch_fs_usage *fs_usage,
 415                         struct disk_reservation *disk_res,
 416                         unsigned journal_seq)
 417 {
 418         s64 added = fs_usage->data + fs_usage->reserved;
 419         s64 should_not_have_added;
 420         int ret = 0;
 421
 422         percpu_rwsem_assert_held(&c->mark_lock);
 423
 424         /*
 425          * Not allowed to reduce sectors_available except by getting a
 426          * reservation:
 427          */
 428         should_not_have_added = added - (s64) (disk_res ? disk_res->sectors : 0);
 429         if (WARN_ONCE(should_not_have_added > 0,
 430                       "disk usage increased by %lli more than reservation of %llu",
 431                       added, disk_res ? disk_res->sectors : 0)) {
 432                 atomic64_sub(should_not_have_added, &c->sectors_available);
 433                 added -= should_not_have_added;
 434                 ret = -1;
 435         }
 436
 437         if (added > 0) {
 438                 disk_res->sectors               -= added;
 439                 fs_usage->online_reserved       -= added;
 440         }
 441
 442         preempt_disable();
 443         acc_u64s((u64 *) fs_usage_ptr(c, journal_seq, false),
 444                  (u64 *) fs_usage, fs_usage_u64s(c));
 445         preempt_enable();
 446
 447         return ret;
 448 }
 449
 450 static inline void account_bucket(struct bch_fs_usage *fs_usage,
 451                                   struct bch_dev_usage *dev_usage,
 452                                   enum bch_data_type type,
 453                                   int nr, s64 size)
 454 {
 455         if (type == BCH_DATA_sb || type == BCH_DATA_journal)
 456                 fs_usage->hidden        += size;
 457
 458         dev_usage->buckets[type]        += nr;
 459 }
 460
 461 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 462                                   struct bch_fs_usage *fs_usage,
 463                                   struct bucket_mark old, struct bucket_mark new,
 464                                   bool gc)
 465 {
 466         struct bch_dev_usage *u;
 467
 468         percpu_rwsem_assert_held(&c->mark_lock);
 469
 470         preempt_disable();
 471         u = this_cpu_ptr(ca->usage[gc]);
 472
 473         if (bucket_type(old))
 474                 account_bucket(fs_usage, u, bucket_type(old),
 475                                -1, -ca->mi.bucket_size);
 476
 477         if (bucket_type(new))
 478                 account_bucket(fs_usage, u, bucket_type(new),
 479                                1, ca->mi.bucket_size);
 480
 481         u->buckets_alloc +=
 482                 (int) new.owned_by_allocator - (int) old.owned_by_allocator;
 483         u->buckets_unavailable +=
 484                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 485
 486         u->buckets_ec += (int) new.stripe - (int) old.stripe;
 487         u->sectors_ec += bucket_stripe_sectors(new) -
 488                          bucket_stripe_sectors(old);
 489
 490         u->sectors[old.data_type] -= old.dirty_sectors;
 491         u->sectors[new.data_type] += new.dirty_sectors;
 492         u->sectors[BCH_DATA_cached] +=
 493                 (int) new.cached_sectors - (int) old.cached_sectors;
 494         u->sectors_fragmented +=
 495                 is_fragmented_bucket(new, ca) - is_fragmented_bucket(old, ca);
 496         preempt_enable();
 497
 498         if (!is_available_bucket(old) && is_available_bucket(new))
 499                 bch2_wake_allocator(ca);
 500 }
 501
 502 __flatten
 503 void bch2_dev_usage_from_buckets(struct bch_fs *c)
 504 {
 505         struct bch_dev *ca;
 506         struct bucket_mark old = { .v.counter = 0 };
 507         struct bucket_array *buckets;
 508         struct bucket *g;
 509         unsigned i;
 510         int cpu;
 511
 512         c->usage_base->hidden = 0;
 513
 514         for_each_member_device(ca, c, i) {
 515                 for_each_possible_cpu(cpu)
 516                         memset(per_cpu_ptr(ca->usage[0], cpu), 0,
 517                                sizeof(*ca->usage[0]));
 518
 519                 buckets = bucket_array(ca);
 520
 521                 for_each_bucket(g, buckets)
 522                         bch2_dev_usage_update(c, ca, c->usage_base,
 523                                               old, g->mark, false);
 524         }
 525 }
 526
 527 static inline int update_replicas(struct bch_fs *c,
 528                                   struct bch_fs_usage *fs_usage,
 529                                   struct bch_replicas_entry *r,
 530                                   s64 sectors)
 531 {
 532         int idx = bch2_replicas_entry_idx(c, r);
 533
 534         if (idx < 0)
 535                 return -1;
 536
 537         if (!fs_usage)
 538                 return 0;
 539
 540         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 541         fs_usage->replicas[idx]         += sectors;
 542         return 0;
 543 }
 544
 545 static inline void update_cached_sectors(struct bch_fs *c,
 546                                          struct bch_fs_usage *fs_usage,
 547                                          unsigned dev, s64 sectors)
 548 {
 549         struct bch_replicas_padded r;
 550
 551         bch2_replicas_entry_cached(&r.e, dev);
 552
 553         update_replicas(c, fs_usage, &r.e, sectors);
 554 }
 555
 556 static struct replicas_delta_list *
 557 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 558 {
 559         struct replicas_delta_list *d = trans->fs_usage_deltas;
 560         unsigned new_size = d ? (d->size + more) * 2 : 128;
 561
 562         if (!d || d->used + more > d->size) {
 563                 d = krealloc(d, sizeof(*d) + new_size, GFP_NOIO|__GFP_ZERO);
 564                 BUG_ON(!d);
 565
 566                 d->size = new_size;
 567                 trans->fs_usage_deltas = d;
 568         }
 569         return d;
 570 }
 571
 572 static inline void update_replicas_list(struct btree_trans *trans,
 573                                         struct bch_replicas_entry *r,
 574                                         s64 sectors)
 575 {
 576         struct replicas_delta_list *d;
 577         struct replicas_delta *n;
 578         unsigned b;
 579
 580         if (!sectors)
 581                 return;
 582
 583         b = replicas_entry_bytes(r) + 8;
 584         d = replicas_deltas_realloc(trans, b);
 585
 586         n = (void *) d->d + d->used;
 587         n->delta = sectors;
 588         memcpy(&n->r, r, replicas_entry_bytes(r));
 589         d->used += b;
 590 }
 591
 592 static inline void update_cached_sectors_list(struct btree_trans *trans,
 593                                               unsigned dev, s64 sectors)
 594 {
 595         struct bch_replicas_padded r;
 596
 597         bch2_replicas_entry_cached(&r.e, dev);
 598
 599         update_replicas_list(trans, &r.e, sectors);
 600 }
 601
 602 static inline struct replicas_delta *
 603 replicas_delta_next(struct replicas_delta *d)
 604 {
 605         return (void *) d + replicas_entry_bytes(&d->r) + 8;
 606 }
 607
 608 int bch2_replicas_delta_list_apply(struct bch_fs *c,
 609                                    struct bch_fs_usage *fs_usage,
 610                                    struct replicas_delta_list *r)
 611 {
 612         struct replicas_delta *d = r->d;
 613         struct replicas_delta *top = (void *) r->d + r->used;
 614         unsigned i;
 615
 616         for (d = r->d; d != top; d = replicas_delta_next(d))
 617                 if (update_replicas(c, fs_usage, &d->r, d->delta)) {
 618                         top = d;
 619                         goto unwind;
 620                 }
 621
 622         if (!fs_usage)
 623                 return 0;
 624
 625         fs_usage->nr_inodes += r->nr_inodes;
 626
 627         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
 628                 fs_usage->reserved += r->persistent_reserved[i];
 629                 fs_usage->persistent_reserved[i] += r->persistent_reserved[i];
 630         }
 631
 632         return 0;
 633 unwind:
 634         for (d = r->d; d != top; d = replicas_delta_next(d))
 635                 update_replicas(c, fs_usage, &d->r, -d->delta);
 636         return -1;
 637 }
 638
 639 #define do_mark_fn(fn, c, pos, flags, ...)                              \
 640 ({                                                                      \
 641         int gc, ret = 0;                                                \
 642                                                                         \
 643         percpu_rwsem_assert_held(&c->mark_lock);                        \
 644                                                                         \
 645         for (gc = 0; gc < 2 && !ret; gc++)                              \
 646                 if (!gc == !(flags & BTREE_TRIGGER_GC) ||               \
 647                     (gc && gc_visited(c, pos)))                         \
 648                         ret = fn(c, __VA_ARGS__, gc);                   \
 649         ret;                                                            \
 650 })
 651
 652 static int __bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 653                                     size_t b, struct bucket_mark *ret,
 654                                     bool gc)
 655 {
 656         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 657         struct bucket *g = __bucket(ca, b, gc);
 658         struct bucket_mark old, new;
 659
 660         old = bucket_cmpxchg(g, new, ({
 661                 BUG_ON(!is_available_bucket(new));
 662
 663                 new.owned_by_allocator  = true;
 664                 new.data_type           = 0;
 665                 new.cached_sectors      = 0;
 666                 new.dirty_sectors       = 0;
 667                 new.gen++;
 668         }));
 669
 670         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 671
 672         if (old.cached_sectors)
 673                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 674                                       -((s64) old.cached_sectors));
 675
 676         if (!gc)
 677                 *ret = old;
 678         return 0;
 679 }
 680
 681 void bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 682                             size_t b, struct bucket_mark *old)
 683 {
 684         do_mark_fn(__bch2_invalidate_bucket, c, gc_phase(GC_PHASE_START), 0,
 685                    ca, b, old);
 686
 687         if (!old->owned_by_allocator && old->cached_sectors)
 688                 trace_invalidate(ca, bucket_to_sector(ca, b),
 689                                  old->cached_sectors);
 690 }
 691
 692 static int __bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 693                                     size_t b, bool owned_by_allocator,
 694                                     bool gc)
 695 {
 696         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 697         struct bucket *g = __bucket(ca, b, gc);
 698         struct bucket_mark old, new;
 699
 700         old = bucket_cmpxchg(g, new, ({
 701                 new.owned_by_allocator  = owned_by_allocator;
 702         }));
 703
 704         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 705
 706         BUG_ON(!gc &&
 707                !owned_by_allocator && !old.owned_by_allocator);
 708
 709         return 0;
 710 }
 711
 712 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 713                             size_t b, bool owned_by_allocator,
 714                             struct gc_pos pos, unsigned flags)
 715 {
 716         preempt_disable();
 717
 718         do_mark_fn(__bch2_mark_alloc_bucket, c, pos, flags,
 719                    ca, b, owned_by_allocator);
 720
 721         preempt_enable();
 722 }
 723
 724 static int bch2_mark_alloc(struct bch_fs *c,
 725                            struct bkey_s_c old, struct bkey_s_c new,
 726                            struct bch_fs_usage *fs_usage,
 727                            u64 journal_seq, unsigned flags)
 728 {
 729         bool gc = flags & BTREE_TRIGGER_GC;
 730         struct bkey_alloc_unpacked u;
 731         struct bch_dev *ca;
 732         struct bucket *g;
 733         struct bucket_mark old_m, m;
 734
 735         /* We don't do anything for deletions - do we?: */
 736         if (new.k->type != KEY_TYPE_alloc)
 737                 return 0;
 738
 739         /*
 740          * alloc btree is read in by bch2_alloc_read, not gc:
 741          */
 742         if ((flags & BTREE_TRIGGER_GC) &&
 743             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 744                 return 0;
 745
 746         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 747
 748         if (new.k->p.offset >= ca->mi.nbuckets)
 749                 return 0;
 750
 751         g = __bucket(ca, new.k->p.offset, gc);
 752         u = bch2_alloc_unpack(new);
 753
 754         old_m = bucket_cmpxchg(g, m, ({
 755                 m.gen                   = u.gen;
 756                 m.data_type             = u.data_type;
 757                 m.dirty_sectors         = u.dirty_sectors;
 758                 m.cached_sectors        = u.cached_sectors;
 759
 760                 if (journal_seq) {
 761                         m.journal_seq_valid     = 1;
 762                         m.journal_seq           = journal_seq;
 763                 }
 764         }));
 765
 766         bch2_dev_usage_update(c, ca, fs_usage, old_m, m, gc);
 767
 768         g->io_time[READ]        = u.read_time;
 769         g->io_time[WRITE]       = u.write_time;
 770         g->oldest_gen           = u.oldest_gen;
 771         g->gen_valid            = 1;
 772
 773         /*
 774          * need to know if we're getting called from the invalidate path or
 775          * not:
 776          */
 777
 778         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 779             old_m.cached_sectors) {
 780                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 781                                       -old_m.cached_sectors);
 782                 trace_invalidate(ca, bucket_to_sector(ca, new.k->p.offset),
 783                                  old_m.cached_sectors);
 784         }
 785
 786         return 0;
 787 }
 788
 789 #define checked_add(a, b)                                       \
 790 ({                                                              \
 791         unsigned _res = (unsigned) (a) + (b);                   \
 792         bool overflow = _res > U16_MAX;                         \
 793         if (overflow)                                           \
 794                 _res = U16_MAX;                                 \
 795         (a) = _res;                                             \
 796         overflow;                                               \
 797 })
 798
 799 static int __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 800                                        size_t b, enum bch_data_type data_type,
 801                                        unsigned sectors, bool gc)
 802 {
 803         struct bucket *g = __bucket(ca, b, gc);
 804         struct bucket_mark old, new;
 805         bool overflow;
 806
 807         BUG_ON(data_type != BCH_DATA_sb &&
 808                data_type != BCH_DATA_journal);
 809
 810         old = bucket_cmpxchg(g, new, ({
 811                 new.data_type   = data_type;
 812                 overflow = checked_add(new.dirty_sectors, sectors);
 813         }));
 814
 815         bch2_fs_inconsistent_on(old.data_type &&
 816                                 old.data_type != data_type, c,
 817                 "different types of data in same bucket: %s, %s",
 818                 bch2_data_types[old.data_type],
 819                 bch2_data_types[data_type]);
 820
 821         bch2_fs_inconsistent_on(overflow, c,
 822                 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
 823                 ca->dev_idx, b, new.gen,
 824                 bch2_data_types[old.data_type ?: data_type],
 825                 old.dirty_sectors, sectors);
 826
 827         if (c)
 828                 bch2_dev_usage_update(c, ca, fs_usage_ptr(c, 0, gc),
 829                                       old, new, gc);
 830
 831         return 0;
 832 }
 833
 834 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 835                                size_t b, enum bch_data_type type,
 836                                unsigned sectors, struct gc_pos pos,
 837                                unsigned flags)
 838 {
 839         BUG_ON(type != BCH_DATA_sb &&
 840                type != BCH_DATA_journal);
 841
 842         preempt_disable();
 843
 844         if (likely(c)) {
 845                 do_mark_fn(__bch2_mark_metadata_bucket, c, pos, flags,
 846                            ca, b, type, sectors);
 847         } else {
 848                 __bch2_mark_metadata_bucket(c, ca, b, type, sectors, 0);
 849         }
 850
 851         preempt_enable();
 852 }
 853
 854 static s64 disk_sectors_scaled(unsigned n, unsigned d, unsigned sectors)
 855 {
 856         return DIV_ROUND_UP(sectors * n, d);
 857 }
 858
 859 static s64 __ptr_disk_sectors_delta(unsigned old_size,
 860                                     unsigned offset, s64 delta,
 861                                     unsigned flags,
 862                                     unsigned n, unsigned d)
 863 {
 864         BUG_ON(!n || !d);
 865
 866         if (flags & BTREE_TRIGGER_OVERWRITE_SPLIT) {
 867                 BUG_ON(offset + -delta > old_size);
 868
 869                 return -disk_sectors_scaled(n, d, old_size) +
 870                         disk_sectors_scaled(n, d, offset) +
 871                         disk_sectors_scaled(n, d, old_size - offset + delta);
 872         } else if (flags & BTREE_TRIGGER_OVERWRITE) {
 873                 BUG_ON(offset + -delta > old_size);
 874
 875                 return -disk_sectors_scaled(n, d, old_size) +
 876                         disk_sectors_scaled(n, d, old_size + delta);
 877         } else {
 878                 return  disk_sectors_scaled(n, d, delta);
 879         }
 880 }
 881
 882 static s64 ptr_disk_sectors_delta(struct extent_ptr_decoded p,
 883                                   unsigned offset, s64 delta,
 884                                   unsigned flags)
 885 {
 886         return __ptr_disk_sectors_delta(p.crc.live_size,
 887                                         offset, delta, flags,
 888                                         p.crc.compressed_size,
 889                                         p.crc.uncompressed_size);
 890 }
 891
 892 static int check_bucket_ref(struct bch_fs *c, struct bkey_s_c k,
 893                             const struct bch_extent_ptr *ptr,
 894                             s64 sectors, enum bch_data_type ptr_data_type,
 895                             u8 bucket_gen, u8 bucket_data_type,
 896                             u16 dirty_sectors, u16 cached_sectors)
 897 {
 898         size_t bucket_nr = PTR_BUCKET_NR(bch_dev_bkey_exists(c, ptr->dev), ptr);
 899         u16 bucket_sectors = !ptr->cached
 900                 ? dirty_sectors
 901                 : cached_sectors;
 902         char buf[200];
 903
 904         if (gen_after(ptr->gen, bucket_gen)) {
 905                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 906                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 907                         "while marking %s",
 908                         ptr->dev, bucket_nr, bucket_gen,
 909                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 910                         ptr->gen,
 911                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 912                 return -EIO;
 913         }
 914
 915         if (gen_cmp(bucket_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 916                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 917                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 918                         "while marking %s",
 919                         ptr->dev, bucket_nr, bucket_gen,
 920                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 921                         ptr->gen,
 922                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 923                 return -EIO;
 924         }
 925
 926         if (bucket_gen != ptr->gen && !ptr->cached) {
 927                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 928                         "bucket %u:%zu gen %u data type %s: stale dirty ptr (gen %u)\n"
 929                         "while marking %s",
 930                         ptr->dev, bucket_nr, bucket_gen,
 931                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 932                         ptr->gen,
 933                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 934                 return -EIO;
 935         }
 936
 937         if (bucket_gen != ptr->gen)
 938                 return 1;
 939
 940         if (bucket_data_type && ptr_data_type &&
 941             bucket_data_type != ptr_data_type) {
 942                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 943                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 944                         "while marking %s",
 945                         ptr->dev, bucket_nr, bucket_gen,
 946                         bch2_data_types[bucket_data_type],
 947                         bch2_data_types[ptr_data_type],
 948                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 949                 return -EIO;
 950         }
 951
 952         if ((unsigned) (bucket_sectors + sectors) > U16_MAX) {
 953                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 954                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 955                         "while marking %s",
 956                         ptr->dev, bucket_nr, bucket_gen,
 957                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 958                         bucket_sectors, sectors,
 959                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 960                 return -EIO;
 961         }
 962
 963         return 0;
 964 }
 965
 966 static int bucket_set_stripe(struct bch_fs *c, struct bkey_s_c k,
 967                              unsigned ptr_idx,
 968                              struct bch_fs_usage *fs_usage,
 969                              u64 journal_seq, unsigned flags,
 970                              bool enabled)
 971 {
 972         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 973         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 974         bool parity = ptr_idx >= nr_data;
 975         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 976         bool gc = flags & BTREE_TRIGGER_GC;
 977         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 978         struct bucket *g = PTR_BUCKET(ca, ptr, gc);
 979         struct bucket_mark new, old;
 980         char buf[200];
 981         int ret;
 982
 983         if (enabled)
 984                 g->ec_redundancy = s->nr_redundant;
 985
 986         old = bucket_cmpxchg(g, new, ({
 987                 ret = check_bucket_ref(c, k, ptr, 0, 0, new.gen, new.data_type,
 988                                        new.dirty_sectors, new.cached_sectors);
 989                 if (ret)
 990                         return ret;
 991
 992                 if (new.stripe && enabled)
 993                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 994                                       "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 995                                       ptr->dev, PTR_BUCKET_NR(ca, ptr), new.gen,
 996                                       (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 997
 998                 if (!new.stripe && !enabled)
 999                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1000                                       "bucket %u:%zu gen %u: deleting stripe but not marked\n%s",
1001                                       ptr->dev, PTR_BUCKET_NR(ca, ptr), new.gen,
1002                                       (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
1003
1004                 new.stripe                      = enabled;
1005
1006                 if ((flags & BTREE_TRIGGER_GC) && parity) {
1007                         new.data_type = enabled ? BCH_DATA_parity : 0;
1008                         new.dirty_sectors = enabled ? le16_to_cpu(s->sectors): 0;
1009                 }
1010
1011                 if (journal_seq) {
1012                         new.journal_seq_valid   = 1;
1013                         new.journal_seq         = journal_seq;
1014                 }
1015         }));
1016
1017         if (!enabled)
1018                 g->ec_redundancy = 0;
1019
1020         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
1021         return 0;
1022 }
1023
1024 static int __mark_pointer(struct bch_fs *c, struct bkey_s_c k,
1025                           const struct bch_extent_ptr *ptr,
1026                           s64 sectors, enum bch_data_type ptr_data_type,
1027                           u8 bucket_gen, u8 *bucket_data_type,
1028                           u16 *dirty_sectors, u16 *cached_sectors)
1029 {
1030         u16 *dst_sectors = !ptr->cached
1031                 ? dirty_sectors
1032                 : cached_sectors;
1033         int ret = check_bucket_ref(c, k, ptr, sectors, ptr_data_type,
1034                                    bucket_gen, *bucket_data_type,
1035                                    *dirty_sectors, *cached_sectors);
1036
1037         if (ret)
1038                 return ret;
1039
1040         *dst_sectors += sectors;
1041         *bucket_data_type = *dirty_sectors || *cached_sectors
1042                 ? ptr_data_type : 0;
1043         return 0;
1044 }
1045
1046 static int bch2_mark_pointer(struct bch_fs *c, struct bkey_s_c k,
1047                              struct extent_ptr_decoded p,
1048                              s64 sectors, enum bch_data_type data_type,
1049                              struct bch_fs_usage *fs_usage,
1050                              u64 journal_seq, unsigned flags)
1051 {
1052         bool gc = flags & BTREE_TRIGGER_GC;
1053         struct bucket_mark old, new;
1054         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
1055         struct bucket *g = PTR_BUCKET(ca, &p.ptr, gc);
1056         u8 bucket_data_type;
1057         u64 v;
1058         int ret;
1059
1060         v = atomic64_read(&g->_mark.v);
1061         do {
1062                 new.v.counter = old.v.counter = v;
1063                 bucket_data_type = new.data_type;
1064
1065                 ret = __mark_pointer(c, k, &p.ptr, sectors, data_type, new.gen,
1066                                      &bucket_data_type,
1067                                      &new.dirty_sectors,
1068                                      &new.cached_sectors);
1069                 if (ret)
1070                         return ret;
1071
1072                 new.data_type = bucket_data_type;
1073
1074                 if (journal_seq) {
1075                         new.journal_seq_valid = 1;
1076                         new.journal_seq = journal_seq;
1077                 }
1078
1079                 if (flags & BTREE_TRIGGER_NOATOMIC) {
1080                         g->_mark = new;
1081                         break;
1082                 }
1083         } while ((v = atomic64_cmpxchg(&g->_mark.v,
1084                               old.v.counter,
1085                               new.v.counter)) != old.v.counter);
1086
1087         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
1088
1089         BUG_ON(!gc && bucket_became_unavailable(old, new));
1090
1091         return 0;
1092 }
1093
1094 static int bch2_mark_stripe_ptr(struct bch_fs *c,
1095                                 struct bch_extent_stripe_ptr p,
1096                                 enum bch_data_type data_type,
1097                                 struct bch_fs_usage *fs_usage,
1098                                 s64 sectors, unsigned flags)
1099 {
1100         bool gc = flags & BTREE_TRIGGER_GC;
1101         struct bch_replicas_padded r;
1102         struct stripe *m;
1103         unsigned i, blocks_nonempty = 0;
1104
1105         m = genradix_ptr(&c->stripes[gc], p.idx);
1106
1107         spin_lock(&c->ec_stripes_heap_lock);
1108
1109         if (!m || !m->alive) {
1110                 spin_unlock(&c->ec_stripes_heap_lock);
1111                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
1112                                     (u64) p.idx);
1113                 return -EIO;
1114         }
1115
1116         m->block_sectors[p.block] += sectors;
1117
1118         r = m->r;
1119
1120         for (i = 0; i < m->nr_blocks; i++)
1121                 blocks_nonempty += m->block_sectors[i] != 0;
1122
1123         if (m->blocks_nonempty != blocks_nonempty) {
1124                 m->blocks_nonempty = blocks_nonempty;
1125                 if (!gc)
1126                         bch2_stripes_heap_update(c, m, p.idx);
1127         }
1128
1129         spin_unlock(&c->ec_stripes_heap_lock);
1130
1131         r.e.data_type = data_type;
1132         update_replicas(c, fs_usage, &r.e, sectors);
1133
1134         return 0;
1135 }
1136
1137 static int bch2_mark_extent(struct bch_fs *c,
1138                             struct bkey_s_c old, struct bkey_s_c new,
1139                             unsigned offset, s64 sectors,
1140                             enum bch_data_type data_type,
1141                             struct bch_fs_usage *fs_usage,
1142                             unsigned journal_seq, unsigned flags)
1143 {
1144         struct bkey_s_c k = flags & BTREE_TRIGGER_INSERT ? new : old;
1145         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1146         const union bch_extent_entry *entry;
1147         struct extent_ptr_decoded p;
1148         struct bch_replicas_padded r;
1149         s64 dirty_sectors = 0;
1150         bool stale;
1151         int ret;
1152
1153         r.e.data_type   = data_type;
1154         r.e.nr_devs     = 0;
1155         r.e.nr_required = 1;
1156
1157         BUG_ON(!sectors);
1158
1159         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1160                 s64 disk_sectors = data_type == BCH_DATA_btree
1161                         ? sectors
1162                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1163
1164                 ret = bch2_mark_pointer(c, k, p, disk_sectors, data_type,
1165                                         fs_usage, journal_seq, flags);
1166                 if (ret < 0)
1167                         return ret;
1168
1169                 stale = ret > 0;
1170
1171                 if (p.ptr.cached) {
1172                         if (!stale)
1173                                 update_cached_sectors(c, fs_usage, p.ptr.dev,
1174                                                       disk_sectors);
1175                 } else if (!p.has_ec) {
1176                         dirty_sectors          += disk_sectors;
1177                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1178                 } else {
1179                         ret = bch2_mark_stripe_ptr(c, p.ec, data_type,
1180                                         fs_usage, disk_sectors, flags);
1181                         if (ret)
1182                                 return ret;
1183
1184                         /*
1185                          * There may be other dirty pointers in this extent, but
1186                          * if so they're not required for mounting if we have an
1187                          * erasure coded pointer in this extent:
1188                          */
1189                         r.e.nr_required = 0;
1190                 }
1191         }
1192
1193         if (r.e.nr_devs)
1194                 update_replicas(c, fs_usage, &r.e, dirty_sectors);
1195
1196         return 0;
1197 }
1198
1199 static int bch2_mark_stripe(struct bch_fs *c,
1200                             struct bkey_s_c old, struct bkey_s_c new,
1201                             struct bch_fs_usage *fs_usage,
1202                             u64 journal_seq, unsigned flags)
1203 {
1204         bool gc = flags & BTREE_TRIGGER_GC;
1205         size_t idx = new.k->p.offset;
1206         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1207                 ? bkey_s_c_to_stripe(old).v : NULL;
1208         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1209                 ? bkey_s_c_to_stripe(new).v : NULL;
1210         struct stripe *m = genradix_ptr(&c->stripes[gc], idx);
1211         unsigned i;
1212         int ret;
1213
1214         if (!m || (old_s && !m->alive)) {
1215                 bch_err_ratelimited(c, "error marking nonexistent stripe %zu",
1216                                     idx);
1217                 return -1;
1218         }
1219
1220         if (!new_s) {
1221                 /* Deleting: */
1222                 for (i = 0; i < old_s->nr_blocks; i++) {
1223                         ret = bucket_set_stripe(c, old, i, fs_usage,
1224                                                 journal_seq, flags, false);
1225                         if (ret)
1226                                 return ret;
1227                 }
1228
1229                 if (!gc && m->on_heap) {
1230                         spin_lock(&c->ec_stripes_heap_lock);
1231                         bch2_stripes_heap_del(c, m, idx);
1232                         spin_unlock(&c->ec_stripes_heap_lock);
1233                 }
1234
1235                 if (gc)
1236                         update_replicas(c, fs_usage, &m->r.e,
1237                                         -((s64) m->sectors * m->nr_redundant));
1238
1239                 memset(m, 0, sizeof(*m));
1240         } else {
1241                 BUG_ON(old_s && new_s->nr_blocks != old_s->nr_blocks);
1242                 BUG_ON(old_s && new_s->nr_redundant != old_s->nr_redundant);
1243
1244                 for (i = 0; i < new_s->nr_blocks; i++) {
1245                         if (!old_s ||
1246                             memcmp(new_s->ptrs + i,
1247                                    old_s->ptrs + i,
1248                                    sizeof(struct bch_extent_ptr))) {
1249
1250                                 if (old_s) {
1251                                         bucket_set_stripe(c, old, i, fs_usage,
1252                                                           journal_seq, flags, false);
1253                                         if (ret)
1254                                                 return ret;
1255                                 }
1256                                 ret = bucket_set_stripe(c, new, i, fs_usage,
1257                                                         journal_seq, flags, true);
1258                                 if (ret)
1259                                         return ret;
1260                         }
1261                 }
1262
1263                 m->alive        = true;
1264                 m->sectors      = le16_to_cpu(new_s->sectors);
1265                 m->algorithm    = new_s->algorithm;
1266                 m->nr_blocks    = new_s->nr_blocks;
1267                 m->nr_redundant = new_s->nr_redundant;
1268                 m->blocks_nonempty = 0;
1269
1270                 for (i = 0; i < new_s->nr_blocks; i++) {
1271                         m->block_sectors[i] =
1272                                 stripe_blockcount_get(new_s, i);
1273                         m->blocks_nonempty += !!m->block_sectors[i];
1274                 }
1275
1276                 if (gc && old_s)
1277                         update_replicas(c, fs_usage, &m->r.e,
1278                                         -((s64) m->sectors * m->nr_redundant));
1279
1280                 bch2_bkey_to_replicas(&m->r.e, new);
1281
1282                 if (gc)
1283                         update_replicas(c, fs_usage, &m->r.e,
1284                                         ((s64) m->sectors * m->nr_redundant));
1285
1286                 if (!gc) {
1287                         spin_lock(&c->ec_stripes_heap_lock);
1288                         bch2_stripes_heap_update(c, m, idx);
1289                         spin_unlock(&c->ec_stripes_heap_lock);
1290                 }
1291         }
1292
1293         return 0;
1294 }
1295
1296 static int bch2_mark_key_locked(struct bch_fs *c,
1297                    struct bkey_s_c old,
1298                    struct bkey_s_c new,
1299                    unsigned offset, s64 sectors,
1300                    struct bch_fs_usage *fs_usage,
1301                    u64 journal_seq, unsigned flags)
1302 {
1303         struct bkey_s_c k = flags & BTREE_TRIGGER_INSERT ? new : old;
1304         int ret = 0;
1305
1306         BUG_ON(!(flags & (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE)));
1307
1308         preempt_disable();
1309
1310         if (!fs_usage || (flags & BTREE_TRIGGER_GC))
1311                 fs_usage = fs_usage_ptr(c, journal_seq,
1312                                         flags & BTREE_TRIGGER_GC);
1313
1314         switch (k.k->type) {
1315         case KEY_TYPE_alloc:
1316                 ret = bch2_mark_alloc(c, old, new, fs_usage, journal_seq, flags);
1317                 break;
1318         case KEY_TYPE_btree_ptr:
1319         case KEY_TYPE_btree_ptr_v2:
1320                 sectors = !(flags & BTREE_TRIGGER_OVERWRITE)
1321                         ?  c->opts.btree_node_size
1322                         : -c->opts.btree_node_size;
1323
1324                 ret = bch2_mark_extent(c, old, new, offset, sectors,
1325                                 BCH_DATA_btree, fs_usage, journal_seq, flags);
1326                 break;
1327         case KEY_TYPE_extent:
1328         case KEY_TYPE_reflink_v:
1329                 ret = bch2_mark_extent(c, old, new, offset, sectors,
1330                                 BCH_DATA_user, fs_usage, journal_seq, flags);
1331                 break;
1332         case KEY_TYPE_stripe:
1333                 ret = bch2_mark_stripe(c, old, new, fs_usage, journal_seq, flags);
1334                 break;
1335         case KEY_TYPE_inode:
1336                 if (!(flags & BTREE_TRIGGER_OVERWRITE))
1337                         fs_usage->nr_inodes++;
1338                 else
1339                         fs_usage->nr_inodes--;
1340                 break;
1341         case KEY_TYPE_reservation: {
1342                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1343
1344                 sectors *= replicas;
1345                 replicas = clamp_t(unsigned, replicas, 1,
1346                                    ARRAY_SIZE(fs_usage->persistent_reserved));
1347
1348                 fs_usage->reserved                              += sectors;
1349                 fs_usage->persistent_reserved[replicas - 1]     += sectors;
1350                 break;
1351         }
1352         }
1353
1354         preempt_enable();
1355
1356         return ret;
1357 }
1358
1359 int bch2_mark_key(struct bch_fs *c, struct bkey_s_c new,
1360                   unsigned offset, s64 sectors,
1361                   struct bch_fs_usage *fs_usage,
1362                   u64 journal_seq, unsigned flags)
1363 {
1364         struct bkey deleted;
1365         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
1366         int ret;
1367
1368         bkey_init(&deleted);
1369
1370         percpu_down_read(&c->mark_lock);
1371         ret = bch2_mark_key_locked(c, old, new, offset, sectors,
1372                                    fs_usage, journal_seq,
1373                                    BTREE_TRIGGER_INSERT|flags);
1374         percpu_up_read(&c->mark_lock);
1375
1376         return ret;
1377 }
1378
1379 int bch2_mark_update(struct btree_trans *trans,
1380                      struct btree_iter *iter,
1381                      struct bkey_i *new,
1382                      struct bch_fs_usage *fs_usage,
1383                      unsigned flags)
1384 {
1385         struct bch_fs           *c = trans->c;
1386         struct btree            *b = iter_l(iter)->b;
1387         struct btree_node_iter  node_iter = iter_l(iter)->iter;
1388         struct bkey_packed      *_old;
1389         struct bkey_s_c         old;
1390         struct bkey             unpacked;
1391         int ret = 0;
1392
1393         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1394                 return 0;
1395
1396         if (!btree_node_type_needs_gc(iter->btree_id))
1397                 return 0;
1398
1399         bkey_init(&unpacked);
1400         old = (struct bkey_s_c) { &unpacked, NULL };
1401
1402         if (!btree_node_type_is_extents(iter->btree_id)) {
1403                 if (btree_iter_type(iter) != BTREE_ITER_CACHED) {
1404                         _old = bch2_btree_node_iter_peek(&node_iter, b);
1405                         if (_old)
1406                                 old = bkey_disassemble(b, _old, &unpacked);
1407                 } else {
1408                         struct bkey_cached *ck = (void *) iter->l[0].b;
1409
1410                         if (ck->valid)
1411                                 old = bkey_i_to_s_c(ck->k);
1412                 }
1413
1414                 if (old.k->type == new->k.type) {
1415                         bch2_mark_key_locked(c, old, bkey_i_to_s_c(new), 0, 0,
1416                                 fs_usage, trans->journal_res.seq,
1417                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
1418
1419                 } else {
1420                         bch2_mark_key_locked(c, old, bkey_i_to_s_c(new), 0, 0,
1421                                 fs_usage, trans->journal_res.seq,
1422                                 BTREE_TRIGGER_INSERT|flags);
1423                         bch2_mark_key_locked(c, old, bkey_i_to_s_c(new), 0, 0,
1424                                 fs_usage, trans->journal_res.seq,
1425                                 BTREE_TRIGGER_OVERWRITE|flags);
1426                 }
1427         } else {
1428                 BUG_ON(btree_iter_type(iter) == BTREE_ITER_CACHED);
1429                 bch2_mark_key_locked(c, old, bkey_i_to_s_c(new),
1430                         0, new->k.size,
1431                         fs_usage, trans->journal_res.seq,
1432                         BTREE_TRIGGER_INSERT|flags);
1433
1434                 while ((_old = bch2_btree_node_iter_peek(&node_iter, b))) {
1435                         unsigned offset = 0;
1436                         s64 sectors;
1437
1438                         old = bkey_disassemble(b, _old, &unpacked);
1439                         sectors = -((s64) old.k->size);
1440
1441                         flags |= BTREE_TRIGGER_OVERWRITE;
1442
1443                         if (bkey_cmp(new->k.p, bkey_start_pos(old.k)) <= 0)
1444                                 return 0;
1445
1446                         switch (bch2_extent_overlap(&new->k, old.k)) {
1447                         case BCH_EXTENT_OVERLAP_ALL:
1448                                 offset = 0;
1449                                 sectors = -((s64) old.k->size);
1450                                 break;
1451                         case BCH_EXTENT_OVERLAP_BACK:
1452                                 offset = bkey_start_offset(&new->k) -
1453                                         bkey_start_offset(old.k);
1454                                 sectors = bkey_start_offset(&new->k) -
1455                                         old.k->p.offset;
1456                                 break;
1457                         case BCH_EXTENT_OVERLAP_FRONT:
1458                                 offset = 0;
1459                                 sectors = bkey_start_offset(old.k) -
1460                                         new->k.p.offset;
1461                                 break;
1462                         case BCH_EXTENT_OVERLAP_MIDDLE:
1463                                 offset = bkey_start_offset(&new->k) -
1464                                         bkey_start_offset(old.k);
1465                                 sectors = -((s64) new->k.size);
1466                                 flags |= BTREE_TRIGGER_OVERWRITE_SPLIT;
1467                                 break;
1468                         }
1469
1470                         BUG_ON(sectors >= 0);
1471
1472                         ret = bch2_mark_key_locked(c, old, bkey_i_to_s_c(new),
1473                                         offset, sectors, fs_usage,
1474                                         trans->journal_res.seq, flags) ?: 1;
1475                         if (ret <= 0)
1476                                 break;
1477
1478                         bch2_btree_node_iter_advance(&node_iter, b);
1479                 }
1480         }
1481
1482         return ret;
1483 }
1484
1485 void bch2_trans_fs_usage_apply(struct btree_trans *trans,
1486                                struct bch_fs_usage *fs_usage)
1487 {
1488         struct bch_fs *c = trans->c;
1489         struct btree_insert_entry *i;
1490         static int warned_disk_usage = 0;
1491         u64 disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1492         char buf[200];
1493
1494         if (!bch2_fs_usage_apply(c, fs_usage, trans->disk_res,
1495                                  trans->journal_res.seq) ||
1496             warned_disk_usage ||
1497             xchg(&warned_disk_usage, 1))
1498                 return;
1499
1500         bch_err(c, "disk usage increased more than %llu sectors reserved",
1501                 disk_res_sectors);
1502
1503         trans_for_each_update(trans, i) {
1504                 pr_err("while inserting");
1505                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1506                 pr_err("%s", buf);
1507                 pr_err("overlapping with");
1508
1509                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED) {
1510                         struct btree            *b = iter_l(i->iter)->b;
1511                         struct btree_node_iter  node_iter = iter_l(i->iter)->iter;
1512                         struct bkey_packed      *_k;
1513
1514                         while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
1515                                 struct bkey             unpacked;
1516                                 struct bkey_s_c         k;
1517
1518                                 pr_info("_k %px format %u", _k, _k->format);
1519                                 k = bkey_disassemble(b, _k, &unpacked);
1520
1521                                 if (btree_node_is_extents(b)
1522                                     ? bkey_cmp(i->k->k.p, bkey_start_pos(k.k)) <= 0
1523                                     : bkey_cmp(i->k->k.p, k.k->p))
1524                                         break;
1525
1526                                 bch2_bkey_val_to_text(&PBUF(buf), c, k);
1527                                 pr_err("%s", buf);
1528
1529                                 bch2_btree_node_iter_advance(&node_iter, b);
1530                         }
1531                 } else {
1532                         struct bkey_cached *ck = (void *) i->iter->l[0].b;
1533
1534                         if (ck->valid) {
1535                                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(ck->k));
1536                                 pr_err("%s", buf);
1537                         }
1538                 }
1539         }
1540 }
1541
1542 /* trans_mark: */
1543
1544 static struct btree_iter *trans_get_update(struct btree_trans *trans,
1545                             enum btree_id btree_id, struct bpos pos,
1546                             struct bkey_s_c *k)
1547 {
1548         struct btree_insert_entry *i;
1549
1550         trans_for_each_update(trans, i)
1551                 if (i->iter->btree_id == btree_id &&
1552                     (btree_node_type_is_extents(btree_id)
1553                      ? bkey_cmp(pos, bkey_start_pos(&i->k->k)) >= 0 &&
1554                        bkey_cmp(pos, i->k->k.p) < 0
1555                      : !bkey_cmp(pos, i->iter->pos))) {
1556                         *k = bkey_i_to_s_c(i->k);
1557                         return i->iter;
1558                 }
1559
1560         return NULL;
1561 }
1562
1563 static int trans_get_key(struct btree_trans *trans,
1564                          enum btree_id btree_id, struct bpos pos,
1565                          struct btree_iter **iter,
1566                          struct bkey_s_c *k)
1567 {
1568         unsigned flags = btree_id != BTREE_ID_ALLOC
1569                 ? BTREE_ITER_SLOTS
1570                 : BTREE_ITER_CACHED;
1571         int ret;
1572
1573         *iter = trans_get_update(trans, btree_id, pos, k);
1574         if (*iter)
1575                 return 1;
1576
1577         *iter = bch2_trans_get_iter(trans, btree_id, pos,
1578                                     flags|BTREE_ITER_INTENT);
1579         if (IS_ERR(*iter))
1580                 return PTR_ERR(*iter);
1581
1582         *k = __bch2_btree_iter_peek(*iter, flags);
1583         ret = bkey_err(*k);
1584         if (ret)
1585                 bch2_trans_iter_put(trans, *iter);
1586         return ret;
1587 }
1588
1589 static int bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter **_iter,
1590                                          const struct bch_extent_ptr *ptr,
1591                                          struct bkey_alloc_unpacked *u)
1592 {
1593         struct bch_fs *c = trans->c;
1594         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1595         struct bpos pos = POS(ptr->dev, PTR_BUCKET_NR(ca, ptr));
1596         struct bucket *g;
1597         struct btree_iter *iter;
1598         struct bkey_s_c k;
1599         int ret;
1600
1601         iter = trans_get_update(trans, BTREE_ID_ALLOC, pos, &k);
1602         if (iter) {
1603                 *u = bch2_alloc_unpack(k);
1604         } else {
1605                 iter = bch2_trans_get_iter(trans, BTREE_ID_ALLOC, pos,
1606                                            BTREE_ITER_CACHED|
1607                                            BTREE_ITER_CACHED_NOFILL|
1608                                            BTREE_ITER_INTENT);
1609                 if (IS_ERR(iter))
1610                         return PTR_ERR(iter);
1611
1612                 ret = bch2_btree_iter_traverse(iter);
1613                 if (ret) {
1614                         bch2_trans_iter_put(trans, iter);
1615                         return ret;
1616                 }
1617
1618                 percpu_down_read(&c->mark_lock);
1619                 g = bucket(ca, pos.offset);
1620                 *u = alloc_mem_to_key(g, READ_ONCE(g->mark));
1621                 percpu_up_read(&c->mark_lock);
1622         }
1623
1624         *_iter = iter;
1625         return 0;
1626 }
1627
1628 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1629                         struct bkey_s_c k, struct extent_ptr_decoded p,
1630                         s64 sectors, enum bch_data_type data_type)
1631 {
1632         struct bch_fs *c = trans->c;
1633         struct btree_iter *iter;
1634         struct bkey_alloc_unpacked u;
1635         struct bkey_i_alloc *a;
1636         int ret;
1637
1638         ret = bch2_trans_start_alloc_update(trans, &iter, &p.ptr, &u);
1639         if (ret)
1640                 return ret;
1641
1642         ret = __mark_pointer(c, k, &p.ptr, sectors, data_type, u.gen, &u.data_type,
1643                              &u.dirty_sectors, &u.cached_sectors);
1644         if (ret)
1645                 goto out;
1646
1647         a = bch2_trans_kmalloc(trans, BKEY_ALLOC_U64s_MAX * 8);
1648         ret = PTR_ERR_OR_ZERO(a);
1649         if (ret)
1650                 goto out;
1651
1652         bkey_alloc_init(&a->k_i);
1653         a->k.p = iter->pos;
1654         bch2_alloc_pack(a, u);
1655         bch2_trans_update(trans, iter, &a->k_i, 0);
1656 out:
1657         bch2_trans_iter_put(trans, iter);
1658         return ret;
1659 }
1660
1661 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1662                         struct bch_extent_stripe_ptr p,
1663                         s64 sectors, enum bch_data_type data_type)
1664 {
1665         struct bch_fs *c = trans->c;
1666         struct btree_iter *iter;
1667         struct bkey_s_c k;
1668         struct bkey_i_stripe *s;
1669         struct bch_replicas_padded r;
1670         int ret = 0;
1671
1672         ret = trans_get_key(trans, BTREE_ID_EC, POS(0, p.idx), &iter, &k);
1673         if (ret < 0)
1674                 return ret;
1675
1676         if (k.k->type != KEY_TYPE_stripe) {
1677                 bch2_fs_inconsistent(c,
1678                         "pointer to nonexistent stripe %llu",
1679                         (u64) p.idx);
1680                 ret = -EIO;
1681                 goto out;
1682         }
1683
1684         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1685         ret = PTR_ERR_OR_ZERO(s);
1686         if (ret)
1687                 goto out;
1688
1689         bkey_reassemble(&s->k_i, k);
1690         stripe_blockcount_set(&s->v, p.block,
1691                 stripe_blockcount_get(&s->v, p.block) +
1692                 sectors);
1693         bch2_trans_update(trans, iter, &s->k_i, 0);
1694
1695         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1696         r.e.data_type = data_type;
1697         update_replicas_list(trans, &r.e, sectors);
1698 out:
1699         bch2_trans_iter_put(trans, iter);
1700         return ret;
1701 }
1702
1703 static int bch2_trans_mark_extent(struct btree_trans *trans,
1704                         struct bkey_s_c k, unsigned offset,
1705                         s64 sectors, unsigned flags,
1706                         enum bch_data_type data_type)
1707 {
1708         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1709         const union bch_extent_entry *entry;
1710         struct extent_ptr_decoded p;
1711         struct bch_replicas_padded r;
1712         s64 dirty_sectors = 0;
1713         bool stale;
1714         int ret;
1715
1716         r.e.data_type   = data_type;
1717         r.e.nr_devs     = 0;
1718         r.e.nr_required = 1;
1719
1720         BUG_ON(!sectors);
1721
1722         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1723                 s64 disk_sectors = data_type == BCH_DATA_btree
1724                         ? sectors
1725                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1726
1727                 ret = bch2_trans_mark_pointer(trans, k, p, disk_sectors,
1728                                               data_type);
1729                 if (ret < 0)
1730                         return ret;
1731
1732                 stale = ret > 0;
1733
1734                 if (p.ptr.cached) {
1735                         if (!stale)
1736                                 update_cached_sectors_list(trans, p.ptr.dev,
1737                                                            disk_sectors);
1738                 } else if (!p.has_ec) {
1739                         dirty_sectors          += disk_sectors;
1740                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1741                 } else {
1742                         ret = bch2_trans_mark_stripe_ptr(trans, p.ec,
1743                                         disk_sectors, data_type);
1744                         if (ret)
1745                                 return ret;
1746
1747                         r.e.nr_required = 0;
1748                 }
1749         }
1750
1751         if (r.e.nr_devs)
1752                 update_replicas_list(trans, &r.e, dirty_sectors);
1753
1754         return 0;
1755 }
1756
1757 static int bch2_trans_mark_stripe(struct btree_trans *trans,
1758                                   struct bkey_s_c k,
1759                                   unsigned flags)
1760 {
1761         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
1762         unsigned nr_data = s->nr_blocks - s->nr_redundant;
1763         struct bch_replicas_padded r;
1764         struct bkey_alloc_unpacked u;
1765         struct bkey_i_alloc *a;
1766         struct btree_iter *iter;
1767         bool deleting = flags & BTREE_TRIGGER_OVERWRITE;
1768         s64 sectors = le16_to_cpu(s->sectors);
1769         unsigned i;
1770         int ret = 0;
1771
1772         if (deleting)
1773                 sectors = -sectors;
1774
1775         bch2_bkey_to_replicas(&r.e, k);
1776         update_replicas_list(trans, &r.e, sectors * s->nr_redundant);
1777
1778         /*
1779          * The allocator code doesn't necessarily update bucket gens in the
1780          * btree when incrementing them, right before handing out new buckets -
1781          * we just need to persist those updates here along with the new stripe:
1782          */
1783
1784         for (i = 0; i < s->nr_blocks && !ret; i++) {
1785                 bool parity = i >= nr_data;
1786
1787                 ret = bch2_trans_start_alloc_update(trans, &iter,
1788                                                     &s->ptrs[i], &u);
1789                 if (ret)
1790                         break;
1791
1792                 if (parity) {
1793                         u.dirty_sectors += sectors;
1794                         u.data_type = u.dirty_sectors
1795                                 ? BCH_DATA_parity
1796                                 : 0;
1797                 }
1798
1799                 a = bch2_trans_kmalloc(trans, BKEY_ALLOC_U64s_MAX * 8);
1800                 ret = PTR_ERR_OR_ZERO(a);
1801                 if (ret)
1802                         goto put_iter;
1803
1804                 bkey_alloc_init(&a->k_i);
1805                 a->k.p = iter->pos;
1806                 bch2_alloc_pack(a, u);
1807                 bch2_trans_update(trans, iter, &a->k_i, 0);
1808 put_iter:
1809                 bch2_trans_iter_put(trans, iter);
1810         }
1811
1812         return ret;
1813 }
1814
1815 static __le64 *bkey_refcount(struct bkey_i *k)
1816 {
1817         switch (k->k.type) {
1818         case KEY_TYPE_reflink_v:
1819                 return &bkey_i_to_reflink_v(k)->v.refcount;
1820         case KEY_TYPE_indirect_inline_data:
1821                 return &bkey_i_to_indirect_inline_data(k)->v.refcount;
1822         default:
1823                 return NULL;
1824         }
1825 }
1826
1827 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1828                         struct bkey_s_c_reflink_p p,
1829                         u64 idx, unsigned sectors,
1830                         unsigned flags)
1831 {
1832         struct bch_fs *c = trans->c;
1833         struct btree_iter *iter;
1834         struct bkey_s_c k;
1835         struct bkey_i *n;
1836         __le64 *refcount;
1837         s64 ret;
1838
1839         ret = trans_get_key(trans, BTREE_ID_REFLINK,
1840                             POS(0, idx), &iter, &k);
1841         if (ret < 0)
1842                 return ret;
1843
1844         if ((flags & BTREE_TRIGGER_OVERWRITE) &&
1845             (bkey_start_offset(k.k) < idx ||
1846              k.k->p.offset > idx + sectors))
1847                 goto out;
1848
1849         sectors = k.k->p.offset - idx;
1850
1851         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1852         ret = PTR_ERR_OR_ZERO(n);
1853         if (ret)
1854                 goto err;
1855
1856         bkey_reassemble(n, k);
1857
1858         refcount = bkey_refcount(n);
1859         if (!refcount) {
1860                 bch2_fs_inconsistent(c,
1861                         "%llu:%llu len %u points to nonexistent indirect extent %llu",
1862                         p.k->p.inode, p.k->p.offset, p.k->size, idx);
1863                 ret = -EIO;
1864                 goto err;
1865         }
1866
1867         le64_add_cpu(refcount, !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1);
1868
1869         if (!*refcount) {
1870                 n->k.type = KEY_TYPE_deleted;
1871                 set_bkey_val_u64s(&n->k, 0);
1872         }
1873
1874         bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
1875         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
1876
1877         bch2_trans_update(trans, iter, n, 0);
1878 out:
1879         ret = sectors;
1880 err:
1881         bch2_trans_iter_put(trans, iter);
1882         return ret;
1883 }
1884
1885 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1886                         struct bkey_s_c_reflink_p p, unsigned offset,
1887                         s64 sectors, unsigned flags)
1888 {
1889         u64 idx = le64_to_cpu(p.v->idx) + offset;
1890         s64 ret = 0;
1891
1892         sectors = abs(sectors);
1893         BUG_ON(offset + sectors > p.k->size);
1894
1895         while (sectors) {
1896                 ret = __bch2_trans_mark_reflink_p(trans, p, idx, sectors, flags);
1897                 if (ret < 0)
1898                         break;
1899
1900                 idx += ret;
1901                 sectors = max_t(s64, 0LL, sectors - ret);
1902                 ret = 0;
1903         }
1904
1905         return ret;
1906 }
1907
1908 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c k,
1909                         unsigned offset, s64 sectors, unsigned flags)
1910 {
1911         struct replicas_delta_list *d;
1912         struct bch_fs *c = trans->c;
1913
1914         switch (k.k->type) {
1915         case KEY_TYPE_btree_ptr:
1916         case KEY_TYPE_btree_ptr_v2:
1917                 sectors = !(flags & BTREE_TRIGGER_OVERWRITE)
1918                         ?  c->opts.btree_node_size
1919                         : -c->opts.btree_node_size;
1920
1921                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1922                                               flags, BCH_DATA_btree);
1923         case KEY_TYPE_extent:
1924         case KEY_TYPE_reflink_v:
1925                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1926                                               flags, BCH_DATA_user);
1927         case KEY_TYPE_stripe:
1928                 return bch2_trans_mark_stripe(trans, k, flags);
1929         case KEY_TYPE_inode:
1930                 d = replicas_deltas_realloc(trans, 0);
1931
1932                 if (!(flags & BTREE_TRIGGER_OVERWRITE))
1933                         d->nr_inodes++;
1934                 else
1935                         d->nr_inodes--;
1936                 return 0;
1937         case KEY_TYPE_reservation: {
1938                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1939
1940                 d = replicas_deltas_realloc(trans, 0);
1941
1942                 sectors *= replicas;
1943                 replicas = clamp_t(unsigned, replicas, 1,
1944                                    ARRAY_SIZE(d->persistent_reserved));
1945
1946                 d->persistent_reserved[replicas - 1] += sectors;
1947                 return 0;
1948         }
1949         case KEY_TYPE_reflink_p:
1950                 return bch2_trans_mark_reflink_p(trans,
1951                                         bkey_s_c_to_reflink_p(k),
1952                                         offset, sectors, flags);
1953         default:
1954                 return 0;
1955         }
1956 }
1957
1958 int bch2_trans_mark_update(struct btree_trans *trans,
1959                            struct btree_iter *iter,
1960                            struct bkey_i *insert,
1961                            unsigned flags)
1962 {
1963         struct btree            *b = iter_l(iter)->b;
1964         struct btree_node_iter  node_iter = iter_l(iter)->iter;
1965         struct bkey_packed      *_k;
1966         int ret;
1967
1968         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1969                 return 0;
1970
1971         if (!btree_node_type_needs_gc(iter->btree_id))
1972                 return 0;
1973
1974         ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(insert),
1975                         0, insert->k.size, BTREE_TRIGGER_INSERT);
1976         if (ret)
1977                 return ret;
1978
1979         if (btree_iter_type(iter) == BTREE_ITER_CACHED) {
1980                 struct bkey_cached *ck = (void *) iter->l[0].b;
1981
1982                 return bch2_trans_mark_key(trans, bkey_i_to_s_c(ck->k),
1983                                            0, 0, BTREE_TRIGGER_OVERWRITE);
1984         }
1985
1986         while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
1987                 struct bkey             unpacked;
1988                 struct bkey_s_c         k;
1989                 unsigned                offset = 0;
1990                 s64                     sectors = 0;
1991                 unsigned                flags = BTREE_TRIGGER_OVERWRITE;
1992
1993                 k = bkey_disassemble(b, _k, &unpacked);
1994
1995                 if (btree_node_is_extents(b)
1996                     ? bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0
1997                     : bkey_cmp(insert->k.p, k.k->p))
1998                         break;
1999
2000                 if (btree_node_is_extents(b)) {
2001                         switch (bch2_extent_overlap(&insert->k, k.k)) {
2002                         case BCH_EXTENT_OVERLAP_ALL:
2003                                 offset = 0;
2004                                 sectors = -((s64) k.k->size);
2005                                 break;
2006                         case BCH_EXTENT_OVERLAP_BACK:
2007                                 offset = bkey_start_offset(&insert->k) -
2008                                         bkey_start_offset(k.k);
2009                                 sectors = bkey_start_offset(&insert->k) -
2010                                         k.k->p.offset;
2011                                 break;
2012                         case BCH_EXTENT_OVERLAP_FRONT:
2013                                 offset = 0;
2014                                 sectors = bkey_start_offset(k.k) -
2015                                         insert->k.p.offset;
2016                                 break;
2017                         case BCH_EXTENT_OVERLAP_MIDDLE:
2018                                 offset = bkey_start_offset(&insert->k) -
2019                                         bkey_start_offset(k.k);
2020                                 sectors = -((s64) insert->k.size);
2021                                 flags |= BTREE_TRIGGER_OVERWRITE_SPLIT;
2022                                 break;
2023                         }
2024
2025                         BUG_ON(sectors >= 0);
2026                 }
2027
2028                 ret = bch2_trans_mark_key(trans, k, offset, sectors, flags);
2029                 if (ret)
2030                         return ret;
2031
2032                 bch2_btree_node_iter_advance(&node_iter, b);
2033         }
2034
2035         return 0;
2036 }
2037
2038 /* Disk reservations: */
2039
2040 static u64 bch2_recalc_sectors_available(struct bch_fs *c)
2041 {
2042         percpu_u64_set(&c->pcpu->sectors_available, 0);
2043
2044         return avail_factor(__bch2_fs_usage_read_short(c).free);
2045 }
2046
2047 #define SECTORS_CACHE   1024
2048
2049 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
2050                               unsigned sectors, int flags)
2051 {
2052         struct bch_fs_pcpu *pcpu;
2053         u64 old, v, get;
2054         s64 sectors_available;
2055         int ret;
2056
2057         percpu_down_read(&c->mark_lock);
2058         preempt_disable();
2059         pcpu = this_cpu_ptr(c->pcpu);
2060
2061         if (sectors <= pcpu->sectors_available)
2062                 goto out;
2063
2064         v = atomic64_read(&c->sectors_available);
2065         do {
2066                 old = v;
2067                 get = min((u64) sectors + SECTORS_CACHE, old);
2068
2069                 if (get < sectors) {
2070                         preempt_enable();
2071                         percpu_up_read(&c->mark_lock);
2072                         goto recalculate;
2073                 }
2074         } while ((v = atomic64_cmpxchg(&c->sectors_available,
2075                                        old, old - get)) != old);
2076
2077         pcpu->sectors_available         += get;
2078
2079 out:
2080         pcpu->sectors_available         -= sectors;
2081         this_cpu_add(c->usage[0]->online_reserved, sectors);
2082         res->sectors                    += sectors;
2083
2084         preempt_enable();
2085         percpu_up_read(&c->mark_lock);
2086         return 0;
2087
2088 recalculate:
2089         percpu_down_write(&c->mark_lock);
2090
2091         sectors_available = bch2_recalc_sectors_available(c);
2092
2093         if (sectors <= sectors_available ||
2094             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2095                 atomic64_set(&c->sectors_available,
2096                              max_t(s64, 0, sectors_available - sectors));
2097                 this_cpu_add(c->usage[0]->online_reserved, sectors);
2098                 res->sectors                    += sectors;
2099                 ret = 0;
2100         } else {
2101                 atomic64_set(&c->sectors_available, sectors_available);
2102                 ret = -ENOSPC;
2103         }
2104
2105         percpu_up_write(&c->mark_lock);
2106
2107         return ret;
2108 }
2109
2110 /* Startup/shutdown: */
2111
2112 static void buckets_free_rcu(struct rcu_head *rcu)
2113 {
2114         struct bucket_array *buckets =
2115                 container_of(rcu, struct bucket_array, rcu);
2116
2117         kvpfree(buckets,
2118                 sizeof(struct bucket_array) +
2119                 buckets->nbuckets * sizeof(struct bucket));
2120 }
2121
2122 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2123 {
2124         struct bucket_array *buckets = NULL, *old_buckets = NULL;
2125         unsigned long *buckets_nouse = NULL;
2126         alloc_fifo      free[RESERVE_NR];
2127         alloc_fifo      free_inc;
2128         alloc_heap      alloc_heap;
2129
2130         size_t btree_reserve    = DIV_ROUND_UP(BTREE_NODE_RESERVE,
2131                              ca->mi.bucket_size / c->opts.btree_node_size);
2132         /* XXX: these should be tunable */
2133         size_t reserve_none     = max_t(size_t, 1, nbuckets >> 9);
2134         size_t copygc_reserve   = max_t(size_t, 2, nbuckets >> 7);
2135         size_t free_inc_nr      = max(max_t(size_t, 1, nbuckets >> 12),
2136                                       btree_reserve * 2);
2137         bool resize = ca->buckets[0] != NULL;
2138         int ret = -ENOMEM;
2139         unsigned i;
2140
2141         memset(&free,           0, sizeof(free));
2142         memset(&free_inc,       0, sizeof(free_inc));
2143         memset(&alloc_heap,     0, sizeof(alloc_heap));
2144
2145         if (!(buckets           = kvpmalloc(sizeof(struct bucket_array) +
2146                                             nbuckets * sizeof(struct bucket),
2147                                             GFP_KERNEL|__GFP_ZERO)) ||
2148             !(buckets_nouse     = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2149                                             sizeof(unsigned long),
2150                                             GFP_KERNEL|__GFP_ZERO)) ||
2151             !init_fifo(&free[RESERVE_BTREE], btree_reserve, GFP_KERNEL) ||
2152             !init_fifo(&free[RESERVE_MOVINGGC],
2153                        copygc_reserve, GFP_KERNEL) ||
2154             !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
2155             !init_fifo(&free_inc,       free_inc_nr, GFP_KERNEL) ||
2156             !init_heap(&alloc_heap,     ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL))
2157                 goto err;
2158
2159         buckets->first_bucket   = ca->mi.first_bucket;
2160         buckets->nbuckets       = nbuckets;
2161
2162         bch2_copygc_stop(c);
2163
2164         if (resize) {
2165                 down_write(&c->gc_lock);
2166                 down_write(&ca->bucket_lock);
2167                 percpu_down_write(&c->mark_lock);
2168         }
2169
2170         old_buckets = bucket_array(ca);
2171
2172         if (resize) {
2173                 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
2174
2175                 memcpy(buckets->b,
2176                        old_buckets->b,
2177                        n * sizeof(struct bucket));
2178                 memcpy(buckets_nouse,
2179                        ca->buckets_nouse,
2180                        BITS_TO_LONGS(n) * sizeof(unsigned long));
2181         }
2182
2183         rcu_assign_pointer(ca->buckets[0], buckets);
2184         buckets = old_buckets;
2185
2186         swap(ca->buckets_nouse, buckets_nouse);
2187
2188         if (resize) {
2189                 percpu_up_write(&c->mark_lock);
2190                 up_write(&c->gc_lock);
2191         }
2192
2193         spin_lock(&c->freelist_lock);
2194         for (i = 0; i < RESERVE_NR; i++) {
2195                 fifo_move(&free[i], &ca->free[i]);
2196                 swap(ca->free[i], free[i]);
2197         }
2198         fifo_move(&free_inc, &ca->free_inc);
2199         swap(ca->free_inc, free_inc);
2200         spin_unlock(&c->freelist_lock);
2201
2202         /* with gc lock held, alloc_heap can't be in use: */
2203         swap(ca->alloc_heap, alloc_heap);
2204
2205         nbuckets = ca->mi.nbuckets;
2206
2207         if (resize)
2208                 up_write(&ca->bucket_lock);
2209
2210         ret = 0;
2211 err:
2212         free_heap(&alloc_heap);
2213         free_fifo(&free_inc);
2214         for (i = 0; i < RESERVE_NR; i++)
2215                 free_fifo(&free[i]);
2216         kvpfree(buckets_nouse,
2217                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2218         if (buckets)
2219                 call_rcu(&old_buckets->rcu, buckets_free_rcu);
2220
2221         return ret;
2222 }
2223
2224 void bch2_dev_buckets_free(struct bch_dev *ca)
2225 {
2226         unsigned i;
2227
2228         free_heap(&ca->alloc_heap);
2229         free_fifo(&ca->free_inc);
2230         for (i = 0; i < RESERVE_NR; i++)
2231                 free_fifo(&ca->free[i]);
2232         kvpfree(ca->buckets_nouse,
2233                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2234         kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2235                 sizeof(struct bucket_array) +
2236                 ca->mi.nbuckets * sizeof(struct bucket));
2237
2238         free_percpu(ca->usage[0]);
2239 }
2240
2241 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2242 {
2243         if (!(ca->usage[0] = alloc_percpu(struct bch_dev_usage)))
2244                 return -ENOMEM;
2245
2246         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2247 }