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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  * Copyright (C) 2014 Datera Inc.
   5  */
   6
   7 #include "bcachefs.h"
   8 #include "alloc_background.h"
   9 #include "alloc_foreground.h"
  10 #include "bkey_methods.h"
  11 #include "bkey_buf.h"
  12 #include "btree_locking.h"
  13 #include "btree_update_interior.h"
  14 #include "btree_io.h"
  15 #include "btree_gc.h"
  16 #include "buckets.h"
  17 #include "clock.h"
  18 #include "debug.h"
  19 #include "ec.h"
  20 #include "error.h"
  21 #include "extents.h"
  22 #include "journal.h"
  23 #include "keylist.h"
  24 #include "move.h"
  25 #include "recovery.h"
  26 #include "replicas.h"
  27 #include "super-io.h"
  28
  29 #include <linux/slab.h>
  30 #include <linux/bitops.h>
  31 #include <linux/freezer.h>
  32 #include <linux/kthread.h>
  33 #include <linux/preempt.h>
  34 #include <linux/rcupdate.h>
  35 #include <linux/sched/task.h>
  36 #include <trace/events/bcachefs.h>
  37
  38 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  39 {
  40         preempt_disable();
  41         write_seqcount_begin(&c->gc_pos_lock);
  42         c->gc_pos = new_pos;
  43         write_seqcount_end(&c->gc_pos_lock);
  44         preempt_enable();
  45 }
  46
  47 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  48 {
  49         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  50         __gc_pos_set(c, new_pos);
  51 }
  52
  53 /*
  54  * Missing: if an interior btree node is empty, we need to do something -
  55  * perhaps just kill it
  56  */
  57 static int bch2_gc_check_topology(struct bch_fs *c,
  58                                   struct btree *b,
  59                                   struct bkey_buf *prev,
  60                                   struct bkey_buf cur,
  61                                   bool is_last)
  62 {
  63         struct bpos node_start  = b->data->min_key;
  64         struct bpos node_end    = b->data->max_key;
  65         struct bpos expected_start = bkey_deleted(&prev->k->k)
  66                 ? node_start
  67                 : bpos_successor(prev->k->k.p);
  68         char buf1[200], buf2[200];
  69         bool update_min = false;
  70         bool update_max = false;
  71         int ret = 0;
  72
  73         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  74                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  75
  76                 if (bkey_deleted(&prev->k->k)) {
  77                         struct printbuf out = PBUF(buf1);
  78                         pr_buf(&out, "start of node: ");
  79                         bch2_bpos_to_text(&out, node_start);
  80                 } else {
  81                         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
  82                 }
  83
  84                 if (fsck_err_on(bpos_cmp(expected_start, bp->v.min_key), c,
  85                                 "btree node with incorrect min_key at btree %s level %u:\n"
  86                                 "  prev %s\n"
  87                                 "  cur %s",
  88                                 bch2_btree_ids[b->c.btree_id], b->c.level,
  89                                 buf1,
  90                                 (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2)))
  91                         update_min = true;
  92         }
  93
  94         if (fsck_err_on(is_last &&
  95                         bpos_cmp(cur.k->k.p, node_end), c,
  96                         "btree node with incorrect max_key at btree %s level %u:\n"
  97                         "  %s\n"
  98                         "  expected %s",
  99                         bch2_btree_ids[b->c.btree_id], b->c.level,
 100                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
 101                         (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2)))
 102                 update_max = true;
 103
 104         bch2_bkey_buf_copy(prev, c, cur.k);
 105
 106         if (update_min || update_max) {
 107                 struct bkey_i *new;
 108                 struct bkey_i_btree_ptr_v2 *bp = NULL;
 109                 struct btree *n;
 110
 111                 if (update_max) {
 112                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 113                                                       b->c.level, cur.k->k.p);
 114                         if (ret)
 115                                 return ret;
 116                 }
 117
 118                 new = kmalloc(bkey_bytes(&cur.k->k), GFP_KERNEL);
 119                 if (!new) {
 120                         bch_err(c, "%s: error allocating new key", __func__);
 121                         return -ENOMEM;
 122                 }
 123
 124                 bkey_copy(new, cur.k);
 125
 126                 if (new->k.type == KEY_TYPE_btree_ptr_v2)
 127                         bp = bkey_i_to_btree_ptr_v2(new);
 128
 129                 if (update_min)
 130                         bp->v.min_key = expected_start;
 131                 if (update_max)
 132                         new->k.p = node_end;
 133                 if (bp)
 134                         SET_BTREE_PTR_RANGE_UPDATED(&bp->v, true);
 135
 136                 ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level, new);
 137                 if (ret) {
 138                         kfree(new);
 139                         return ret;
 140                 }
 141
 142                 n = bch2_btree_node_get_noiter(c, cur.k, b->c.btree_id,
 143                                                b->c.level - 1, true);
 144                 if (n) {
 145                         mutex_lock(&c->btree_cache.lock);
 146                         bch2_btree_node_hash_remove(&c->btree_cache, n);
 147
 148                         bkey_copy(&n->key, new);
 149                         if (update_min)
 150                                 n->data->min_key = expected_start;
 151                         if (update_max)
 152                                 n->data->max_key = node_end;
 153
 154                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, n);
 155                         BUG_ON(ret);
 156                         mutex_unlock(&c->btree_cache.lock);
 157                         six_unlock_read(&n->c.lock);
 158                 }
 159         }
 160 fsck_err:
 161         return ret;
 162 }
 163
 164 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 165                                unsigned level, bool is_root,
 166                                struct bkey_s_c *k)
 167 {
 168         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 169         const union bch_extent_entry *entry;
 170         struct extent_ptr_decoded p = { 0 };
 171         bool do_update = false;
 172         int ret = 0;
 173
 174         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 175                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 176                 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
 177                 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
 178
 179                 if (fsck_err_on(!g->gen_valid, c,
 180                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree",
 181                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 182                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 183                                 p.ptr.gen)) {
 184                         if (p.ptr.cached) {
 185                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 186                                 g2->gen_valid   = g->gen_valid          = true;
 187                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 188                         } else {
 189                                 do_update = true;
 190                         }
 191                 }
 192
 193                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 194                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u",
 195                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 196                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 197                                 p.ptr.gen, g->mark.gen)) {
 198                         if (p.ptr.cached) {
 199                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 200                                 g2->gen_valid   = g->gen_valid  = true;
 201                                 g2->_mark.data_type             = 0;
 202                                 g2->_mark.dirty_sectors         = 0;
 203                                 g2->_mark.cached_sectors        = 0;
 204                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 205                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 206                         } else {
 207                                 do_update = true;
 208                         }
 209                 }
 210
 211                 if (fsck_err_on(!p.ptr.cached &&
 212                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 213                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u",
 214                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 215                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 216                                 p.ptr.gen, g->mark.gen))
 217                         do_update = true;
 218
 219                 if (p.has_ec) {
 220                         struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
 221
 222                         if (fsck_err_on(!m || !m->alive, c,
 223                                         "pointer to nonexistent stripe %llu",
 224                                         (u64) p.ec.idx))
 225                                 do_update = true;
 226
 227                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 228                                         "pointer does not match stripe %llu",
 229                                         (u64) p.ec.idx))
 230                                 do_update = true;
 231                 }
 232         }
 233
 234         if (do_update) {
 235                 struct bkey_ptrs ptrs;
 236                 union bch_extent_entry *entry;
 237                 struct bch_extent_ptr *ptr;
 238                 struct bkey_i *new;
 239
 240                 if (is_root) {
 241                         bch_err(c, "cannot update btree roots yet");
 242                         return -EINVAL;
 243                 }
 244
 245                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 246                 if (!new) {
 247                         bch_err(c, "%s: error allocating new key", __func__);
 248                         return -ENOMEM;
 249                 }
 250
 251                 bkey_reassemble(new, *k);
 252
 253                 if (level) {
 254                         /*
 255                          * We don't want to drop btree node pointers - if the
 256                          * btree node isn't there anymore, the read path will
 257                          * sort it out:
 258                          */
 259                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 260                         bkey_for_each_ptr(ptrs, ptr) {
 261                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 262                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 263
 264                                 ptr->gen = g->mark.gen;
 265                         }
 266                 } else {
 267                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 268                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 269                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 270
 271                                 (ptr->cached &&
 272                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 273                                 (!ptr->cached &&
 274                                  gen_cmp(ptr->gen, g->mark.gen) < 0);
 275                         }));
 276 again:
 277                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 278                         bkey_extent_entry_for_each(ptrs, entry) {
 279                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 280                                         struct stripe *m = genradix_ptr(&c->stripes[true],
 281                                                                         entry->stripe_ptr.idx);
 282                                         union bch_extent_entry *next_ptr;
 283
 284                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 285                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 286                                                         goto found;
 287                                         next_ptr = NULL;
 288 found:
 289                                         if (!next_ptr) {
 290                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 291                                                 continue;
 292                                         }
 293
 294                                         if (!m || !m->alive ||
 295                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 296                                                                        &next_ptr->ptr,
 297                                                                        m->sectors)) {
 298                                                 bch2_bkey_extent_entry_drop(new, entry);
 299                                                 goto again;
 300                                         }
 301                                 }
 302                         }
 303                 }
 304
 305                 ret = bch2_journal_key_insert(c, btree_id, level, new);
 306                 if (ret)
 307                         kfree(new);
 308                 else
 309                         *k = bkey_i_to_s_c(new);
 310         }
 311 fsck_err:
 312         return ret;
 313 }
 314
 315 /* marking of btree keys/nodes: */
 316
 317 static int bch2_gc_mark_key(struct bch_fs *c, enum btree_id btree_id,
 318                             unsigned level, bool is_root,
 319                             struct bkey_s_c *k,
 320                             u8 *max_stale, bool initial)
 321 {
 322         struct bkey_ptrs_c ptrs;
 323         const struct bch_extent_ptr *ptr;
 324         unsigned flags =
 325                 BTREE_TRIGGER_GC|
 326                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 327         int ret = 0;
 328
 329         if (initial) {
 330                 BUG_ON(bch2_journal_seq_verify &&
 331                        k->k->version.lo > journal_cur_seq(&c->journal));
 332
 333                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 334                                 "key version number higher than recorded: %llu > %llu",
 335                                 k->k->version.lo,
 336                                 atomic64_read(&c->key_version)))
 337                         atomic64_set(&c->key_version, k->k->version.lo);
 338
 339                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 340                     fsck_err_on(!bch2_bkey_replicas_marked(c, *k), c,
 341                                 "superblock not marked as containing replicas (type %u)",
 342                                 k->k->type)) {
 343                         ret = bch2_mark_bkey_replicas(c, *k);
 344                         if (ret) {
 345                                 bch_err(c, "error marking bkey replicas: %i", ret);
 346                                 goto err;
 347                         }
 348                 }
 349
 350                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 351         }
 352
 353         ptrs = bch2_bkey_ptrs_c(*k);
 354         bkey_for_each_ptr(ptrs, ptr) {
 355                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 356                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 357
 358                 if (gen_after(g->oldest_gen, ptr->gen))
 359                         g->oldest_gen = ptr->gen;
 360
 361                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 362         }
 363
 364         bch2_mark_key(c, *k, 0, k->k->size, NULL, 0, flags);
 365 fsck_err:
 366 err:
 367         if (ret)
 368                 bch_err(c, "%s: ret %i", __func__, ret);
 369         return ret;
 370 }
 371
 372 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
 373                               bool initial)
 374 {
 375         struct btree_node_iter iter;
 376         struct bkey unpacked;
 377         struct bkey_s_c k;
 378         struct bkey_buf prev, cur;
 379         int ret = 0;
 380
 381         *max_stale = 0;
 382
 383         if (!btree_node_type_needs_gc(btree_node_type(b)))
 384                 return 0;
 385
 386         bch2_btree_node_iter_init_from_start(&iter, b);
 387         bch2_bkey_buf_init(&prev);
 388         bch2_bkey_buf_init(&cur);
 389         bkey_init(&prev.k->k);
 390
 391         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 392                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 393                                        &k, max_stale, initial);
 394                 if (ret)
 395                         break;
 396
 397                 bch2_btree_node_iter_advance(&iter, b);
 398
 399                 if (b->c.level) {
 400                         bch2_bkey_buf_reassemble(&cur, c, k);
 401
 402                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 403                                         bch2_btree_node_iter_end(&iter));
 404                         if (ret)
 405                                 break;
 406                 }
 407         }
 408
 409         bch2_bkey_buf_exit(&cur, c);
 410         bch2_bkey_buf_exit(&prev, c);
 411         return ret;
 412 }
 413
 414 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 415                          bool initial, bool metadata_only)
 416 {
 417         struct btree_trans trans;
 418         struct btree_iter *iter;
 419         struct btree *b;
 420         unsigned depth = metadata_only                  ? 1
 421                 : bch2_expensive_debug_checks           ? 0
 422                 : !btree_node_type_needs_gc(btree_id)   ? 1
 423                 : 0;
 424         u8 max_stale = 0;
 425         int ret = 0;
 426
 427         bch2_trans_init(&trans, c, 0, 0);
 428
 429         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 430
 431         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 432                               0, depth, BTREE_ITER_PREFETCH, b) {
 433                 bch2_verify_btree_nr_keys(b);
 434
 435                 gc_pos_set(c, gc_pos_btree_node(b));
 436
 437                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 438                 if (ret)
 439                         break;
 440
 441                 if (!initial) {
 442                         if (max_stale > 64)
 443                                 bch2_btree_node_rewrite(c, iter,
 444                                                 b->data->keys.seq,
 445                                                 BTREE_INSERT_NOWAIT|
 446                                                 BTREE_INSERT_GC_LOCK_HELD);
 447                         else if (!bch2_btree_gc_rewrite_disabled &&
 448                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 449                                 bch2_btree_node_rewrite(c, iter,
 450                                                 b->data->keys.seq,
 451                                                 BTREE_INSERT_NOWAIT|
 452                                                 BTREE_INSERT_GC_LOCK_HELD);
 453                 }
 454
 455                 bch2_trans_cond_resched(&trans);
 456         }
 457         bch2_trans_iter_put(&trans, iter);
 458
 459         ret = bch2_trans_exit(&trans) ?: ret;
 460         if (ret)
 461                 return ret;
 462
 463         mutex_lock(&c->btree_root_lock);
 464         b = c->btree_roots[btree_id].b;
 465         if (!btree_node_fake(b)) {
 466                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 467
 468                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 469                                        &k, &max_stale, initial);
 470         }
 471         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 472         mutex_unlock(&c->btree_root_lock);
 473
 474         return ret;
 475 }
 476
 477 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
 478                                       unsigned target_depth)
 479 {
 480         struct btree_and_journal_iter iter;
 481         struct bkey_s_c k;
 482         struct bkey_buf cur, prev;
 483         u8 max_stale = 0;
 484         int ret = 0;
 485
 486         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 487         bch2_bkey_buf_init(&prev);
 488         bch2_bkey_buf_init(&cur);
 489         bkey_init(&prev.k->k);
 490
 491         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 492                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 493                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 494
 495                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 496                                        &k, &max_stale, true);
 497                 if (ret) {
 498                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 499                         break;
 500                 }
 501
 502                 if (b->c.level) {
 503                         bch2_bkey_buf_reassemble(&cur, c, k);
 504                         k = bkey_i_to_s_c(cur.k);
 505
 506                         bch2_btree_and_journal_iter_advance(&iter);
 507
 508                         ret = bch2_gc_check_topology(c, b,
 509                                         &prev, cur,
 510                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 511                         if (ret)
 512                                 break;
 513                 } else {
 514                         bch2_btree_and_journal_iter_advance(&iter);
 515                 }
 516         }
 517
 518         if (b->c.level > target_depth) {
 519                 bch2_btree_and_journal_iter_exit(&iter);
 520                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 521
 522                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 523                         struct btree *child;
 524
 525                         bch2_bkey_buf_reassemble(&cur, c, k);
 526                         bch2_btree_and_journal_iter_advance(&iter);
 527
 528                         child = bch2_btree_node_get_noiter(c, cur.k,
 529                                                 b->c.btree_id, b->c.level - 1,
 530                                                 false);
 531                         ret = PTR_ERR_OR_ZERO(child);
 532
 533                         if (fsck_err_on(ret == -EIO, c,
 534                                         "unreadable btree node")) {
 535                                 ret = bch2_journal_key_delete(c, b->c.btree_id,
 536                                                               b->c.level, cur.k->k.p);
 537                                 if (ret)
 538                                         return ret;
 539
 540                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 541                                 continue;
 542                         }
 543
 544                         if (ret) {
 545                                 bch_err(c, "%s: error %i getting btree node",
 546                                         __func__, ret);
 547                                 break;
 548                         }
 549
 550                         ret = bch2_gc_btree_init_recurse(c, child,
 551                                                          target_depth);
 552                         six_unlock_read(&child->c.lock);
 553
 554                         if (ret)
 555                                 break;
 556                 }
 557         }
 558 fsck_err:
 559         bch2_bkey_buf_exit(&cur, c);
 560         bch2_bkey_buf_exit(&prev, c);
 561         bch2_btree_and_journal_iter_exit(&iter);
 562         return ret;
 563 }
 564
 565 static int bch2_gc_btree_init(struct bch_fs *c,
 566                               enum btree_id btree_id,
 567                               bool metadata_only)
 568 {
 569         struct btree *b;
 570         unsigned target_depth = metadata_only           ? 1
 571                 : bch2_expensive_debug_checks           ? 0
 572                 : !btree_node_type_needs_gc(btree_id)   ? 1
 573                 : 0;
 574         u8 max_stale = 0;
 575         char buf[100];
 576         int ret = 0;
 577
 578         b = c->btree_roots[btree_id].b;
 579
 580         if (btree_node_fake(b))
 581                 return 0;
 582
 583         six_lock_read(&b->c.lock, NULL, NULL);
 584         if (fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 585                         "btree root with incorrect min_key: %s",
 586                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 587                 BUG();
 588         }
 589
 590         if (fsck_err_on(bpos_cmp(b->data->max_key, POS_MAX), c,
 591                         "btree root with incorrect max_key: %s",
 592                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 593                 BUG();
 594         }
 595
 596         if (b->c.level >= target_depth)
 597                 ret = bch2_gc_btree_init_recurse(c, b, target_depth);
 598
 599         if (!ret) {
 600                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 601
 602                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 603                                        &k, &max_stale, true);
 604         }
 605 fsck_err:
 606         six_unlock_read(&b->c.lock);
 607
 608         if (ret)
 609                 bch_err(c, "%s: ret %i", __func__, ret);
 610         return ret;
 611 }
 612
 613 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 614 {
 615         return  (int) btree_id_to_gc_phase(l) -
 616                 (int) btree_id_to_gc_phase(r);
 617 }
 618
 619 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
 620 {
 621         enum btree_id ids[BTREE_ID_NR];
 622         unsigned i;
 623
 624         for (i = 0; i < BTREE_ID_NR; i++)
 625                 ids[i] = i;
 626         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
 627
 628         for (i = 0; i < BTREE_ID_NR; i++) {
 629                 enum btree_id id = ids[i];
 630                 int ret = initial
 631                         ? bch2_gc_btree_init(c, id, metadata_only)
 632                         : bch2_gc_btree(c, id, initial, metadata_only);
 633                 if (ret) {
 634                         bch_err(c, "%s: ret %i", __func__, ret);
 635                         return ret;
 636                 }
 637         }
 638
 639         return 0;
 640 }
 641
 642 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 643                                   u64 start, u64 end,
 644                                   enum bch_data_type type,
 645                                   unsigned flags)
 646 {
 647         u64 b = sector_to_bucket(ca, start);
 648
 649         do {
 650                 unsigned sectors =
 651                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
 652
 653                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
 654                                           gc_phase(GC_PHASE_SB), flags);
 655                 b++;
 656                 start += sectors;
 657         } while (start < end);
 658 }
 659
 660 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 661                               unsigned flags)
 662 {
 663         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 664         unsigned i;
 665         u64 b;
 666
 667         /*
 668          * This conditional is kind of gross, but we may be called from the
 669          * device add path, before the new device has actually been added to the
 670          * running filesystem:
 671          */
 672         if (c) {
 673                 lockdep_assert_held(&c->sb_lock);
 674                 percpu_down_read(&c->mark_lock);
 675         }
 676
 677         for (i = 0; i < layout->nr_superblocks; i++) {
 678                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 679
 680                 if (offset == BCH_SB_SECTOR)
 681                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 682                                               BCH_DATA_sb, flags);
 683
 684                 mark_metadata_sectors(c, ca, offset,
 685                                       offset + (1 << layout->sb_max_size_bits),
 686                                       BCH_DATA_sb, flags);
 687         }
 688
 689         for (i = 0; i < ca->journal.nr; i++) {
 690                 b = ca->journal.buckets[i];
 691                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
 692                                           ca->mi.bucket_size,
 693                                           gc_phase(GC_PHASE_SB), flags);
 694         }
 695
 696         if (c)
 697                 percpu_up_read(&c->mark_lock);
 698 }
 699
 700 static void bch2_mark_superblocks(struct bch_fs *c)
 701 {
 702         struct bch_dev *ca;
 703         unsigned i;
 704
 705         mutex_lock(&c->sb_lock);
 706         gc_pos_set(c, gc_phase(GC_PHASE_SB));
 707
 708         for_each_online_member(ca, c, i)
 709                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
 710         mutex_unlock(&c->sb_lock);
 711 }
 712
 713 #if 0
 714 /* Also see bch2_pending_btree_node_free_insert_done() */
 715 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 716 {
 717         struct btree_update *as;
 718         struct pending_btree_node_free *d;
 719
 720         mutex_lock(&c->btree_interior_update_lock);
 721         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 722
 723         for_each_pending_btree_node_free(c, as, d)
 724                 if (d->index_update_done)
 725                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
 726                                       0, 0, NULL, 0,
 727                                       BTREE_TRIGGER_GC);
 728
 729         mutex_unlock(&c->btree_interior_update_lock);
 730 }
 731 #endif
 732
 733 static void bch2_gc_free(struct bch_fs *c)
 734 {
 735         struct bch_dev *ca;
 736         unsigned i;
 737
 738         genradix_free(&c->stripes[1]);
 739
 740         for_each_member_device(ca, c, i) {
 741                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
 742                         sizeof(struct bucket_array) +
 743                         ca->mi.nbuckets * sizeof(struct bucket));
 744                 ca->buckets[1] = NULL;
 745
 746                 free_percpu(ca->usage_gc);
 747                 ca->usage_gc = NULL;
 748         }
 749
 750         free_percpu(c->usage_gc);
 751         c->usage_gc = NULL;
 752 }
 753
 754 static int bch2_gc_done(struct bch_fs *c,
 755                         bool initial, bool metadata_only)
 756 {
 757         struct bch_dev *ca;
 758         bool verify = !metadata_only && (!initial ||
 759                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
 760         unsigned i, dev;
 761         int ret = 0;
 762
 763 #define copy_field(_f, _msg, ...)                                       \
 764         if (dst->_f != src->_f) {                                       \
 765                 if (verify)                                             \
 766                         fsck_err(c, _msg ": got %llu, should be %llu"   \
 767                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
 768                 dst->_f = src->_f;                                      \
 769                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 770         }
 771 #define copy_stripe_field(_f, _msg, ...)                                \
 772         if (dst->_f != src->_f) {                                       \
 773                 if (verify)                                             \
 774                         fsck_err(c, "stripe %zu has wrong "_msg         \
 775                                 ": got %u, should be %u",               \
 776                                 iter.pos, ##__VA_ARGS__,                \
 777                                 dst->_f, src->_f);                      \
 778                 dst->_f = src->_f;                                      \
 779                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 780         }
 781 #define copy_bucket_field(_f)                                           \
 782         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
 783                 if (verify)                                             \
 784                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
 785                                 ": got %u, should be %u", dev, b,       \
 786                                 dst->b[b].mark.gen,                     \
 787                                 bch2_data_types[dst->b[b].mark.data_type],\
 788                                 dst->b[b].mark._f, src->b[b].mark._f);  \
 789                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
 790                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 791         }
 792 #define copy_dev_field(_f, _msg, ...)                                   \
 793         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
 794 #define copy_fs_field(_f, _msg, ...)                                    \
 795         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
 796
 797         if (!metadata_only) {
 798                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
 799                 struct stripe *dst, *src;
 800
 801                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
 802                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
 803
 804                         if (dst->alive          != src->alive ||
 805                             dst->sectors        != src->sectors ||
 806                             dst->algorithm      != src->algorithm ||
 807                             dst->nr_blocks      != src->nr_blocks ||
 808                             dst->nr_redundant   != src->nr_redundant) {
 809                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
 810                                 ret = -EINVAL;
 811                                 goto fsck_err;
 812                         }
 813
 814                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
 815                                 copy_stripe_field(block_sectors[i],
 816                                                   "block_sectors[%u]", i);
 817
 818                         dst->blocks_nonempty = 0;
 819                         for (i = 0; i < dst->nr_blocks; i++)
 820                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
 821
 822                         genradix_iter_advance(&iter, &c->stripes[1]);
 823                 }
 824         }
 825
 826         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 827                 bch2_fs_usage_acc_to_base(c, i);
 828
 829         for_each_member_device(ca, c, dev) {
 830                 struct bucket_array *dst = __bucket_array(ca, 0);
 831                 struct bucket_array *src = __bucket_array(ca, 1);
 832                 size_t b;
 833
 834                 for (b = 0; b < src->nbuckets; b++) {
 835                         copy_bucket_field(gen);
 836                         copy_bucket_field(data_type);
 837                         copy_bucket_field(stripe);
 838                         copy_bucket_field(dirty_sectors);
 839                         copy_bucket_field(cached_sectors);
 840
 841                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
 842                 }
 843
 844                 {
 845                         struct bch_dev_usage *dst = ca->usage_base;
 846                         struct bch_dev_usage *src = (void *)
 847                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
 848                                                      dev_usage_u64s());
 849
 850                         copy_dev_field(buckets_ec,              "buckets_ec");
 851                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
 852
 853                         for (i = 0; i < BCH_DATA_NR; i++) {
 854                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
 855                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
 856                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
 857                         }
 858                 }
 859         };
 860
 861         {
 862                 unsigned nr = fs_usage_u64s(c);
 863                 struct bch_fs_usage *dst = c->usage_base;
 864                 struct bch_fs_usage *src = (void *)
 865                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
 866
 867                 copy_fs_field(hidden,           "hidden");
 868                 copy_fs_field(btree,            "btree");
 869
 870                 if (!metadata_only) {
 871                         copy_fs_field(data,     "data");
 872                         copy_fs_field(cached,   "cached");
 873                         copy_fs_field(reserved, "reserved");
 874                         copy_fs_field(nr_inodes,"nr_inodes");
 875
 876                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
 877                                 copy_fs_field(persistent_reserved[i],
 878                                               "persistent_reserved[%i]", i);
 879                 }
 880
 881                 for (i = 0; i < c->replicas.nr; i++) {
 882                         struct bch_replicas_entry *e =
 883                                 cpu_replicas_entry(&c->replicas, i);
 884                         char buf[80];
 885
 886                         if (metadata_only &&
 887                             (e->data_type == BCH_DATA_user ||
 888                              e->data_type == BCH_DATA_cached))
 889                                 continue;
 890
 891                         bch2_replicas_entry_to_text(&PBUF(buf), e);
 892
 893                         copy_fs_field(replicas[i], "%s", buf);
 894                 }
 895         }
 896
 897 #undef copy_fs_field
 898 #undef copy_dev_field
 899 #undef copy_bucket_field
 900 #undef copy_stripe_field
 901 #undef copy_field
 902 fsck_err:
 903         if (ret)
 904                 bch_err(c, "%s: ret %i", __func__, ret);
 905         return ret;
 906 }
 907
 908 static int bch2_gc_start(struct bch_fs *c,
 909                          bool metadata_only)
 910 {
 911         struct bch_dev *ca;
 912         unsigned i;
 913         int ret;
 914
 915         BUG_ON(c->usage_gc);
 916
 917         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
 918                                          sizeof(u64), GFP_KERNEL);
 919         if (!c->usage_gc) {
 920                 bch_err(c, "error allocating c->usage_gc");
 921                 return -ENOMEM;
 922         }
 923
 924         for_each_member_device(ca, c, i) {
 925                 BUG_ON(ca->buckets[1]);
 926                 BUG_ON(ca->usage_gc);
 927
 928                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
 929                                 ca->mi.nbuckets * sizeof(struct bucket),
 930                                 GFP_KERNEL|__GFP_ZERO);
 931                 if (!ca->buckets[1]) {
 932                         percpu_ref_put(&ca->ref);
 933                         bch_err(c, "error allocating ca->buckets[gc]");
 934                         return -ENOMEM;
 935                 }
 936
 937                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
 938                 if (!ca->usage_gc) {
 939                         bch_err(c, "error allocating ca->usage_gc");
 940                         percpu_ref_put(&ca->ref);
 941                         return -ENOMEM;
 942                 }
 943         }
 944
 945         ret = bch2_ec_mem_alloc(c, true);
 946         if (ret) {
 947                 bch_err(c, "error allocating ec gc mem");
 948                 return ret;
 949         }
 950
 951         percpu_down_write(&c->mark_lock);
 952
 953         /*
 954          * indicate to stripe code that we need to allocate for the gc stripes
 955          * radix tree, too
 956          */
 957         gc_pos_set(c, gc_phase(GC_PHASE_START));
 958
 959         for_each_member_device(ca, c, i) {
 960                 struct bucket_array *dst = __bucket_array(ca, 1);
 961                 struct bucket_array *src = __bucket_array(ca, 0);
 962                 size_t b;
 963
 964                 dst->first_bucket       = src->first_bucket;
 965                 dst->nbuckets           = src->nbuckets;
 966
 967                 for (b = 0; b < src->nbuckets; b++) {
 968                         struct bucket *d = &dst->b[b];
 969                         struct bucket *s = &src->b[b];
 970
 971                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
 972                         d->gen_valid = s->gen_valid;
 973
 974                         if (metadata_only &&
 975                             (s->mark.data_type == BCH_DATA_user ||
 976                              s->mark.data_type == BCH_DATA_cached))
 977                                 d->_mark = s->mark;
 978                 }
 979         };
 980
 981         percpu_up_write(&c->mark_lock);
 982
 983         return 0;
 984 }
 985
 986 /**
 987  * bch2_gc - walk _all_ references to buckets, and recompute them:
 988  *
 989  * Order matters here:
 990  *  - Concurrent GC relies on the fact that we have a total ordering for
 991  *    everything that GC walks - see  gc_will_visit_node(),
 992  *    gc_will_visit_root()
 993  *
 994  *  - also, references move around in the course of index updates and
 995  *    various other crap: everything needs to agree on the ordering
 996  *    references are allowed to move around in - e.g., we're allowed to
 997  *    start with a reference owned by an open_bucket (the allocator) and
 998  *    move it to the btree, but not the reverse.
 999  *
1000  *    This is necessary to ensure that gc doesn't miss references that
1001  *    move around - if references move backwards in the ordering GC
1002  *    uses, GC could skip past them
1003  */
1004 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1005 {
1006         struct bch_dev *ca;
1007         u64 start_time = local_clock();
1008         unsigned i, iter = 0;
1009         int ret;
1010
1011         lockdep_assert_held(&c->state_lock);
1012         trace_gc_start(c);
1013
1014         down_write(&c->gc_lock);
1015
1016         /* flush interior btree updates: */
1017         closure_wait_event(&c->btree_interior_update_wait,
1018                            !bch2_btree_interior_updates_nr_pending(c));
1019 again:
1020         ret = bch2_gc_start(c, metadata_only);
1021         if (ret)
1022                 goto out;
1023
1024         bch2_mark_superblocks(c);
1025
1026         ret = bch2_gc_btrees(c, initial, metadata_only);
1027         if (ret)
1028                 goto out;
1029
1030 #if 0
1031         bch2_mark_pending_btree_node_frees(c);
1032 #endif
1033         c->gc_count++;
1034
1035         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1036             (!iter && bch2_test_restart_gc)) {
1037                 /*
1038                  * XXX: make sure gens we fixed got saved
1039                  */
1040                 if (iter++ <= 2) {
1041                         bch_info(c, "Second GC pass needed, restarting:");
1042                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1043                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1044
1045                         percpu_down_write(&c->mark_lock);
1046                         bch2_gc_free(c);
1047                         percpu_up_write(&c->mark_lock);
1048                         /* flush fsck errors, reset counters */
1049                         bch2_flush_fsck_errs(c);
1050
1051                         goto again;
1052                 }
1053
1054                 bch_info(c, "Unable to fix bucket gens, looping");
1055                 ret = -EINVAL;
1056         }
1057 out:
1058         if (!ret) {
1059                 bch2_journal_block(&c->journal);
1060
1061                 percpu_down_write(&c->mark_lock);
1062                 ret = bch2_gc_done(c, initial, metadata_only);
1063
1064                 bch2_journal_unblock(&c->journal);
1065         } else {
1066                 percpu_down_write(&c->mark_lock);
1067         }
1068
1069         /* Indicates that gc is no longer in progress: */
1070         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1071
1072         bch2_gc_free(c);
1073         percpu_up_write(&c->mark_lock);
1074
1075         up_write(&c->gc_lock);
1076
1077         trace_gc_end(c);
1078         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1079
1080         /*
1081          * Wake up allocator in case it was waiting for buckets
1082          * because of not being able to inc gens
1083          */
1084         for_each_member_device(ca, c, i)
1085                 bch2_wake_allocator(ca);
1086
1087         /*
1088          * At startup, allocations can happen directly instead of via the
1089          * allocator thread - issue wakeup in case they blocked on gc_lock:
1090          */
1091         closure_wake_up(&c->freelist_wait);
1092         return ret;
1093 }
1094
1095 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1096 {
1097         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1098         const struct bch_extent_ptr *ptr;
1099
1100         percpu_down_read(&c->mark_lock);
1101         bkey_for_each_ptr(ptrs, ptr) {
1102                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1103                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1104
1105                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1106                         percpu_up_read(&c->mark_lock);
1107                         return true;
1108                 }
1109         }
1110
1111         bkey_for_each_ptr(ptrs, ptr) {
1112                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1113                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1114
1115                 if (gen_after(g->gc_gen, ptr->gen))
1116                         g->gc_gen = ptr->gen;
1117         }
1118         percpu_up_read(&c->mark_lock);
1119
1120         return false;
1121 }
1122
1123 /*
1124  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1125  * node pointers currently never have cached pointers that can become stale:
1126  */
1127 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1128 {
1129         struct btree_trans trans;
1130         struct btree_iter *iter;
1131         struct bkey_s_c k;
1132         struct bkey_buf sk;
1133         int ret = 0, commit_err = 0;
1134
1135         bch2_bkey_buf_init(&sk);
1136         bch2_trans_init(&trans, c, 0, 0);
1137
1138         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
1139                                    BTREE_ITER_PREFETCH|
1140                                    BTREE_ITER_NOT_EXTENTS|
1141                                    BTREE_ITER_ALL_SNAPSHOTS);
1142
1143         while ((k = bch2_btree_iter_peek(iter)).k &&
1144                !(ret = bkey_err(k))) {
1145                 c->gc_gens_pos = iter->pos;
1146
1147                 if (gc_btree_gens_key(c, k) && !commit_err) {
1148                         bch2_bkey_buf_reassemble(&sk, c, k);
1149                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1150
1151                         bch2_trans_update(&trans, iter, sk.k, 0);
1152
1153                         commit_err = bch2_trans_commit(&trans, NULL, NULL,
1154                                                        BTREE_INSERT_NOWAIT|
1155                                                        BTREE_INSERT_NOFAIL);
1156                         if (commit_err == -EINTR) {
1157                                 commit_err = 0;
1158                                 continue;
1159                         }
1160                 }
1161
1162                 bch2_btree_iter_advance(iter);
1163         }
1164         bch2_trans_iter_put(&trans, iter);
1165
1166         bch2_trans_exit(&trans);
1167         bch2_bkey_buf_exit(&sk, c);
1168
1169         return ret;
1170 }
1171
1172 int bch2_gc_gens(struct bch_fs *c)
1173 {
1174         struct bch_dev *ca;
1175         struct bucket_array *buckets;
1176         struct bucket *g;
1177         unsigned i;
1178         int ret;
1179
1180         /*
1181          * Ideally we would be using state_lock and not gc_lock here, but that
1182          * introduces a deadlock in the RO path - we currently take the state
1183          * lock at the start of going RO, thus the gc thread may get stuck:
1184          */
1185         down_read(&c->gc_lock);
1186
1187         for_each_member_device(ca, c, i) {
1188                 down_read(&ca->bucket_lock);
1189                 buckets = bucket_array(ca);
1190
1191                 for_each_bucket(g, buckets)
1192                         g->gc_gen = g->mark.gen;
1193                 up_read(&ca->bucket_lock);
1194         }
1195
1196         for (i = 0; i < BTREE_ID_NR; i++)
1197                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1198                         c->gc_gens_btree = i;
1199                         c->gc_gens_pos = POS_MIN;
1200                         ret = bch2_gc_btree_gens(c, i);
1201                         if (ret) {
1202                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1203                                 goto err;
1204                         }
1205                 }
1206
1207         for_each_member_device(ca, c, i) {
1208                 down_read(&ca->bucket_lock);
1209                 buckets = bucket_array(ca);
1210
1211                 for_each_bucket(g, buckets)
1212                         g->oldest_gen = g->gc_gen;
1213                 up_read(&ca->bucket_lock);
1214         }
1215
1216         c->gc_gens_btree        = 0;
1217         c->gc_gens_pos          = POS_MIN;
1218
1219         c->gc_count++;
1220 err:
1221         up_read(&c->gc_lock);
1222         return ret;
1223 }
1224
1225 static int bch2_gc_thread(void *arg)
1226 {
1227         struct bch_fs *c = arg;
1228         struct io_clock *clock = &c->io_clock[WRITE];
1229         unsigned long last = atomic64_read(&clock->now);
1230         unsigned last_kick = atomic_read(&c->kick_gc);
1231         int ret;
1232
1233         set_freezable();
1234
1235         while (1) {
1236                 while (1) {
1237                         set_current_state(TASK_INTERRUPTIBLE);
1238
1239                         if (kthread_should_stop()) {
1240                                 __set_current_state(TASK_RUNNING);
1241                                 return 0;
1242                         }
1243
1244                         if (atomic_read(&c->kick_gc) != last_kick)
1245                                 break;
1246
1247                         if (c->btree_gc_periodic) {
1248                                 unsigned long next = last + c->capacity / 16;
1249
1250                                 if (atomic64_read(&clock->now) >= next)
1251                                         break;
1252
1253                                 bch2_io_clock_schedule_timeout(clock, next);
1254                         } else {
1255                                 schedule();
1256                         }
1257
1258                         try_to_freeze();
1259                 }
1260                 __set_current_state(TASK_RUNNING);
1261
1262                 last = atomic64_read(&clock->now);
1263                 last_kick = atomic_read(&c->kick_gc);
1264
1265                 /*
1266                  * Full gc is currently incompatible with btree key cache:
1267                  */
1268 #if 0
1269                 ret = bch2_gc(c, false, false);
1270 #else
1271                 ret = bch2_gc_gens(c);
1272 #endif
1273                 if (ret < 0)
1274                         bch_err(c, "btree gc failed: %i", ret);
1275
1276                 debug_check_no_locks_held();
1277         }
1278
1279         return 0;
1280 }
1281
1282 void bch2_gc_thread_stop(struct bch_fs *c)
1283 {
1284         struct task_struct *p;
1285
1286         p = c->gc_thread;
1287         c->gc_thread = NULL;
1288
1289         if (p) {
1290                 kthread_stop(p);
1291                 put_task_struct(p);
1292         }
1293 }
1294
1295 int bch2_gc_thread_start(struct bch_fs *c)
1296 {
1297         struct task_struct *p;
1298
1299         if (c->gc_thread)
1300                 return 0;
1301
1302         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1303         if (IS_ERR(p)) {
1304                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1305                 return PTR_ERR(p);
1306         }
1307
1308         get_task_struct(p);
1309         c->gc_thread = p;
1310         wake_up_process(p);
1311         return 0;
1312 }