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_key_cache.h"
  13 #include "btree_locking.h"
  14 #include "btree_update_interior.h"
  15 #include "btree_io.h"
  16 #include "btree_gc.h"
  17 #include "buckets.h"
  18 #include "clock.h"
  19 #include "debug.h"
  20 #include "ec.h"
  21 #include "error.h"
  22 #include "extents.h"
  23 #include "journal.h"
  24 #include "keylist.h"
  25 #include "move.h"
  26 #include "recovery.h"
  27 #include "reflink.h"
  28 #include "replicas.h"
  29 #include "super-io.h"
  30 #include "trace.h"
  31
  32 #include <linux/slab.h>
  33 #include <linux/bitops.h>
  34 #include <linux/freezer.h>
  35 #include <linux/kthread.h>
  36 #include <linux/preempt.h>
  37 #include <linux/rcupdate.h>
  38 #include <linux/sched/task.h>
  39
  40 #define DROP_THIS_NODE          10
  41 #define DROP_PREV_NODE          11
  42
  43 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  44 {
  45         preempt_disable();
  46         write_seqcount_begin(&c->gc_pos_lock);
  47         c->gc_pos = new_pos;
  48         write_seqcount_end(&c->gc_pos_lock);
  49         preempt_enable();
  50 }
  51
  52 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  53 {
  54         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  55         __gc_pos_set(c, new_pos);
  56 }
  57
  58 /*
  59  * Missing: if an interior btree node is empty, we need to do something -
  60  * perhaps just kill it
  61  */
  62 static int bch2_gc_check_topology(struct bch_fs *c,
  63                                   struct btree *b,
  64                                   struct bkey_buf *prev,
  65                                   struct bkey_buf cur,
  66                                   bool is_last)
  67 {
  68         struct bpos node_start  = b->data->min_key;
  69         struct bpos node_end    = b->data->max_key;
  70         struct bpos expected_start = bkey_deleted(&prev->k->k)
  71                 ? node_start
  72                 : bpos_successor(prev->k->k.p);
  73         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  74         int ret = 0;
  75
  76         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  77                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  78
  79                 if (!bpos_eq(expected_start, bp->v.min_key)) {
  80                         bch2_topology_error(c);
  81
  82                         if (bkey_deleted(&prev->k->k)) {
  83                                 prt_printf(&buf1, "start of node: ");
  84                                 bch2_bpos_to_text(&buf1, node_start);
  85                         } else {
  86                                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
  87                         }
  88                         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
  89
  90                         if (__fsck_err(c,
  91                                   FSCK_CAN_FIX|
  92                                   FSCK_CAN_IGNORE|
  93                                   FSCK_NO_RATELIMIT,
  94                                   "btree node with incorrect min_key at btree %s level %u:\n"
  95                                   "  prev %s\n"
  96                                   "  cur %s",
  97                                   bch2_btree_ids[b->c.btree_id], b->c.level,
  98                                   buf1.buf, buf2.buf) &&
  99                             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 100                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 101                                 ret = -BCH_ERR_need_topology_repair;
 102                                 goto err;
 103                         } else {
 104                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 105                         }
 106                 }
 107         }
 108
 109         if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
 110                 bch2_topology_error(c);
 111
 112                 printbuf_reset(&buf1);
 113                 printbuf_reset(&buf2);
 114
 115                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
 116                 bch2_bpos_to_text(&buf2, node_end);
 117
 118                 if (__fsck_err(c,
 119                           FSCK_CAN_FIX|
 120                           FSCK_CAN_IGNORE|
 121                           FSCK_NO_RATELIMIT,
 122                           "btree node with incorrect max_key at btree %s level %u:\n"
 123                           "  %s\n"
 124                           "  expected %s",
 125                           bch2_btree_ids[b->c.btree_id], b->c.level,
 126                           buf1.buf, buf2.buf) &&
 127                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 128                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 129                         ret = -BCH_ERR_need_topology_repair;
 130                         goto err;
 131                 } else {
 132                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 133                 }
 134         }
 135
 136         bch2_bkey_buf_copy(prev, c, cur.k);
 137 err:
 138 fsck_err:
 139         printbuf_exit(&buf2);
 140         printbuf_exit(&buf1);
 141         return ret;
 142 }
 143
 144 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 145 {
 146         switch (b->key.k.type) {
 147         case KEY_TYPE_btree_ptr: {
 148                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 149
 150                 dst->k.p                = src->k.p;
 151                 dst->v.mem_ptr          = 0;
 152                 dst->v.seq              = b->data->keys.seq;
 153                 dst->v.sectors_written  = 0;
 154                 dst->v.flags            = 0;
 155                 dst->v.min_key          = b->data->min_key;
 156                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 157                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 158                 break;
 159         }
 160         case KEY_TYPE_btree_ptr_v2:
 161                 bkey_copy(&dst->k_i, &b->key);
 162                 break;
 163         default:
 164                 BUG();
 165         }
 166 }
 167
 168 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
 169                                              enum btree_id btree, unsigned level,
 170                                              struct bkey_s_c old, struct bkey_i *new)
 171 {
 172         struct bch_fs *c = trans->c;
 173         struct btree *b;
 174         struct bkey_buf tmp;
 175         int ret;
 176
 177         bch2_bkey_buf_init(&tmp);
 178         bch2_bkey_buf_reassemble(&tmp, c, old);
 179
 180         b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
 181         if (!IS_ERR_OR_NULL(b)) {
 182                 mutex_lock(&c->btree_cache.lock);
 183
 184                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 185
 186                 bkey_copy(&b->key, new);
 187                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 188                 BUG_ON(ret);
 189
 190                 mutex_unlock(&c->btree_cache.lock);
 191                 six_unlock_read(&b->c.lock);
 192         }
 193
 194         bch2_bkey_buf_exit(&tmp, c);
 195 }
 196
 197 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 198 {
 199         struct bkey_i_btree_ptr_v2 *new;
 200         int ret;
 201
 202         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 203         if (!new)
 204                 return -BCH_ERR_ENOMEM_gc_repair_key;
 205
 206         btree_ptr_to_v2(b, new);
 207         b->data->min_key        = new_min;
 208         new->v.min_key          = new_min;
 209         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 210
 211         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 212         if (ret) {
 213                 kfree(new);
 214                 return ret;
 215         }
 216
 217         bch2_btree_node_drop_keys_outside_node(b);
 218         bkey_copy(&b->key, &new->k_i);
 219         return 0;
 220 }
 221
 222 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 223 {
 224         struct bkey_i_btree_ptr_v2 *new;
 225         int ret;
 226
 227         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 228         if (ret)
 229                 return ret;
 230
 231         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 232         if (!new)
 233                 return -BCH_ERR_ENOMEM_gc_repair_key;
 234
 235         btree_ptr_to_v2(b, new);
 236         b->data->max_key        = new_max;
 237         new->k.p                = new_max;
 238         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 239
 240         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 241         if (ret) {
 242                 kfree(new);
 243                 return ret;
 244         }
 245
 246         bch2_btree_node_drop_keys_outside_node(b);
 247
 248         mutex_lock(&c->btree_cache.lock);
 249         bch2_btree_node_hash_remove(&c->btree_cache, b);
 250
 251         bkey_copy(&b->key, &new->k_i);
 252         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 253         BUG_ON(ret);
 254         mutex_unlock(&c->btree_cache.lock);
 255         return 0;
 256 }
 257
 258 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 259                                         struct btree *prev, struct btree *cur)
 260 {
 261         struct bpos expected_start = !prev
 262                 ? b->data->min_key
 263                 : bpos_successor(prev->key.k.p);
 264         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 265         int ret = 0;
 266
 267         if (!prev) {
 268                 prt_printf(&buf1, "start of node: ");
 269                 bch2_bpos_to_text(&buf1, b->data->min_key);
 270         } else {
 271                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
 272         }
 273
 274         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
 275
 276         if (prev &&
 277             bpos_gt(expected_start, cur->data->min_key) &&
 278             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 279                 /* cur overwrites prev: */
 280
 281                 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
 282                                                 cur->data->min_key), c,
 283                                 "btree node overwritten by next node at btree %s level %u:\n"
 284                                 "  node %s\n"
 285                                 "  next %s",
 286                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 287                                 buf1.buf, buf2.buf)) {
 288                         ret = DROP_PREV_NODE;
 289                         goto out;
 290                 }
 291
 292                 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
 293                                                  bpos_predecessor(cur->data->min_key)), c,
 294                                 "btree node with incorrect max_key at btree %s level %u:\n"
 295                                 "  node %s\n"
 296                                 "  next %s",
 297                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 298                                 buf1.buf, buf2.buf))
 299                         ret = set_node_max(c, prev,
 300                                            bpos_predecessor(cur->data->min_key));
 301         } else {
 302                 /* prev overwrites cur: */
 303
 304                 if (mustfix_fsck_err_on(bpos_ge(expected_start,
 305                                                 cur->data->max_key), c,
 306                                 "btree node overwritten by prev node at btree %s level %u:\n"
 307                                 "  prev %s\n"
 308                                 "  node %s",
 309                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 310                                 buf1.buf, buf2.buf)) {
 311                         ret = DROP_THIS_NODE;
 312                         goto out;
 313                 }
 314
 315                 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
 316                                 "btree node with incorrect min_key at btree %s level %u:\n"
 317                                 "  prev %s\n"
 318                                 "  node %s",
 319                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 320                                 buf1.buf, buf2.buf))
 321                         ret = set_node_min(c, cur, expected_start);
 322         }
 323 out:
 324 fsck_err:
 325         printbuf_exit(&buf2);
 326         printbuf_exit(&buf1);
 327         return ret;
 328 }
 329
 330 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 331                                  struct btree *child)
 332 {
 333         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 334         int ret = 0;
 335
 336         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
 337         bch2_bpos_to_text(&buf2, b->key.k.p);
 338
 339         if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
 340                         "btree node with incorrect max_key at btree %s level %u:\n"
 341                         "  %s\n"
 342                         "  expected %s",
 343                         bch2_btree_ids[b->c.btree_id], b->c.level,
 344                         buf1.buf, buf2.buf)) {
 345                 ret = set_node_max(c, child, b->key.k.p);
 346                 if (ret)
 347                         goto err;
 348         }
 349 err:
 350 fsck_err:
 351         printbuf_exit(&buf2);
 352         printbuf_exit(&buf1);
 353         return ret;
 354 }
 355
 356 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
 357 {
 358         struct bch_fs *c = trans->c;
 359         struct btree_and_journal_iter iter;
 360         struct bkey_s_c k;
 361         struct bkey_buf prev_k, cur_k;
 362         struct btree *prev = NULL, *cur = NULL;
 363         bool have_child, dropped_children = false;
 364         struct printbuf buf = PRINTBUF;
 365         int ret = 0;
 366
 367         if (!b->c.level)
 368                 return 0;
 369 again:
 370         prev = NULL;
 371         have_child = dropped_children = false;
 372         bch2_bkey_buf_init(&prev_k);
 373         bch2_bkey_buf_init(&cur_k);
 374         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 375
 376         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 377                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 378                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 379
 380                 bch2_btree_and_journal_iter_advance(&iter);
 381                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 382
 383                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 384                                         b->c.btree_id, b->c.level - 1,
 385                                         false);
 386                 ret = PTR_ERR_OR_ZERO(cur);
 387
 388                 printbuf_reset(&buf);
 389                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 390
 391                 if (mustfix_fsck_err_on(ret == -EIO, c,
 392                                 "Topology repair: unreadable btree node at btree %s level %u:\n"
 393                                 "  %s",
 394                                 bch2_btree_ids[b->c.btree_id],
 395                                 b->c.level - 1,
 396                                 buf.buf)) {
 397                         bch2_btree_node_evict(trans, cur_k.k);
 398                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 399                                                       b->c.level, cur_k.k->k.p);
 400                         cur = NULL;
 401                         if (ret)
 402                                 break;
 403                         continue;
 404                 }
 405
 406                 if (ret) {
 407                         bch_err_msg(c, ret, "getting btree node");
 408                         break;
 409                 }
 410
 411                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 412
 413                 if (ret == DROP_THIS_NODE) {
 414                         six_unlock_read(&cur->c.lock);
 415                         bch2_btree_node_evict(trans, cur_k.k);
 416                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 417                                                       b->c.level, cur_k.k->k.p);
 418                         cur = NULL;
 419                         if (ret)
 420                                 break;
 421                         continue;
 422                 }
 423
 424                 if (prev)
 425                         six_unlock_read(&prev->c.lock);
 426                 prev = NULL;
 427
 428                 if (ret == DROP_PREV_NODE) {
 429                         bch2_btree_node_evict(trans, prev_k.k);
 430                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 431                                                       b->c.level, prev_k.k->k.p);
 432                         if (ret)
 433                                 break;
 434
 435                         bch2_btree_and_journal_iter_exit(&iter);
 436                         bch2_bkey_buf_exit(&prev_k, c);
 437                         bch2_bkey_buf_exit(&cur_k, c);
 438                         goto again;
 439                 } else if (ret)
 440                         break;
 441
 442                 prev = cur;
 443                 cur = NULL;
 444                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 445         }
 446
 447         if (!ret && !IS_ERR_OR_NULL(prev)) {
 448                 BUG_ON(cur);
 449                 ret = btree_repair_node_end(c, b, prev);
 450         }
 451
 452         if (!IS_ERR_OR_NULL(prev))
 453                 six_unlock_read(&prev->c.lock);
 454         prev = NULL;
 455         if (!IS_ERR_OR_NULL(cur))
 456                 six_unlock_read(&cur->c.lock);
 457         cur = NULL;
 458
 459         if (ret)
 460                 goto err;
 461
 462         bch2_btree_and_journal_iter_exit(&iter);
 463         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 464
 465         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 466                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 467                 bch2_btree_and_journal_iter_advance(&iter);
 468
 469                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 470                                         b->c.btree_id, b->c.level - 1,
 471                                         false);
 472                 ret = PTR_ERR_OR_ZERO(cur);
 473
 474                 if (ret) {
 475                         bch_err_msg(c, ret, "getting btree node");
 476                         goto err;
 477                 }
 478
 479                 ret = bch2_btree_repair_topology_recurse(trans, cur);
 480                 six_unlock_read(&cur->c.lock);
 481                 cur = NULL;
 482
 483                 if (ret == DROP_THIS_NODE) {
 484                         bch2_btree_node_evict(trans, cur_k.k);
 485                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 486                                                       b->c.level, cur_k.k->k.p);
 487                         dropped_children = true;
 488                 }
 489
 490                 if (ret)
 491                         goto err;
 492
 493                 have_child = true;
 494         }
 495
 496         printbuf_reset(&buf);
 497         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 498
 499         if (mustfix_fsck_err_on(!have_child, c,
 500                         "empty interior btree node at btree %s level %u\n"
 501                         "  %s",
 502                         bch2_btree_ids[b->c.btree_id],
 503                         b->c.level, buf.buf))
 504                 ret = DROP_THIS_NODE;
 505 err:
 506 fsck_err:
 507         if (!IS_ERR_OR_NULL(prev))
 508                 six_unlock_read(&prev->c.lock);
 509         if (!IS_ERR_OR_NULL(cur))
 510                 six_unlock_read(&cur->c.lock);
 511
 512         bch2_btree_and_journal_iter_exit(&iter);
 513         bch2_bkey_buf_exit(&prev_k, c);
 514         bch2_bkey_buf_exit(&cur_k, c);
 515
 516         if (!ret && dropped_children)
 517                 goto again;
 518
 519         printbuf_exit(&buf);
 520         return ret;
 521 }
 522
 523 static int bch2_repair_topology(struct bch_fs *c)
 524 {
 525         struct btree_trans trans;
 526         struct btree *b;
 527         unsigned i;
 528         int ret = 0;
 529
 530         bch2_trans_init(&trans, c, 0, 0);
 531
 532         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 533                 b = c->btree_roots[i].b;
 534                 if (btree_node_fake(b))
 535                         continue;
 536
 537                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 538                 ret = bch2_btree_repair_topology_recurse(&trans, b);
 539                 six_unlock_read(&b->c.lock);
 540
 541                 if (ret == DROP_THIS_NODE) {
 542                         bch_err(c, "empty btree root - repair unimplemented");
 543                         ret = -BCH_ERR_fsck_repair_unimplemented;
 544                 }
 545         }
 546
 547         bch2_trans_exit(&trans);
 548
 549         return ret;
 550 }
 551
 552 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
 553                                unsigned level, bool is_root,
 554                                struct bkey_s_c *k)
 555 {
 556         struct bch_fs *c = trans->c;
 557         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 558         const union bch_extent_entry *entry;
 559         struct extent_ptr_decoded p = { 0 };
 560         bool do_update = false;
 561         struct printbuf buf = PRINTBUF;
 562         int ret = 0;
 563
 564         /*
 565          * XXX
 566          * use check_bucket_ref here
 567          */
 568         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 569                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 570                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 571                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 572
 573                 if (!g->gen_valid &&
 574                     (c->opts.reconstruct_alloc ||
 575                      fsck_err(c, "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 576                               "while marking %s",
 577                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 578                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 579                               p.ptr.gen,
 580                               (printbuf_reset(&buf),
 581                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 582                         if (!p.ptr.cached) {
 583                                 g->gen_valid            = true;
 584                                 g->gen                  = p.ptr.gen;
 585                         } else {
 586                                 do_update = true;
 587                         }
 588                 }
 589
 590                 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
 591                     (c->opts.reconstruct_alloc ||
 592                      fsck_err(c, "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 593                               "while marking %s",
 594                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 595                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 596                               p.ptr.gen, g->gen,
 597                               (printbuf_reset(&buf),
 598                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 599                         if (!p.ptr.cached) {
 600                                 g->gen_valid            = true;
 601                                 g->gen                  = p.ptr.gen;
 602                                 g->data_type            = 0;
 603                                 g->dirty_sectors        = 0;
 604                                 g->cached_sectors       = 0;
 605                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 606                         } else {
 607                                 do_update = true;
 608                         }
 609                 }
 610
 611                 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
 612                     (c->opts.reconstruct_alloc ||
 613                      fsck_err(c, "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 614                               "while marking %s",
 615                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 616                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 617                               p.ptr.gen,
 618                               (printbuf_reset(&buf),
 619                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 620                         do_update = true;
 621
 622                 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
 623                     (c->opts.reconstruct_alloc ||
 624                      fsck_err(c, "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 625                               "while marking %s",
 626                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 627                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 628                               p.ptr.gen, g->gen,
 629                               (printbuf_reset(&buf),
 630                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 631                         do_update = true;
 632
 633                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 634                         continue;
 635
 636                 if (fsck_err_on(bucket_data_type(g->data_type) &&
 637                                 bucket_data_type(g->data_type) != data_type, c,
 638                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 639                                 "while marking %s",
 640                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 641                                 bch2_data_types[g->data_type],
 642                                 bch2_data_types[data_type],
 643                                 (printbuf_reset(&buf),
 644                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 645                         if (data_type == BCH_DATA_btree) {
 646                                 g->data_type    = data_type;
 647                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 648                         } else {
 649                                 do_update = true;
 650                         }
 651                 }
 652
 653                 if (p.has_ec) {
 654                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 655
 656                         if (fsck_err_on(!m || !m->alive, c,
 657                                         "pointer to nonexistent stripe %llu\n"
 658                                         "while marking %s",
 659                                         (u64) p.ec.idx,
 660                                         (printbuf_reset(&buf),
 661                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 662                                 do_update = true;
 663
 664                         if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
 665                                         "pointer does not match stripe %llu\n"
 666                                         "while marking %s",
 667                                         (u64) p.ec.idx,
 668                                         (printbuf_reset(&buf),
 669                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 670                                 do_update = true;
 671                 }
 672         }
 673
 674         if (do_update) {
 675                 struct bkey_ptrs ptrs;
 676                 union bch_extent_entry *entry;
 677                 struct bch_extent_ptr *ptr;
 678                 struct bkey_i *new;
 679
 680                 if (is_root) {
 681                         bch_err(c, "cannot update btree roots yet");
 682                         ret = -EINVAL;
 683                         goto err;
 684                 }
 685
 686                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 687                 if (!new) {
 688                         bch_err_msg(c, ret, "allocating new key");
 689                         ret = -BCH_ERR_ENOMEM_gc_repair_key;
 690                         goto err;
 691                 }
 692
 693                 bkey_reassemble(new, *k);
 694
 695                 if (level) {
 696                         /*
 697                          * We don't want to drop btree node pointers - if the
 698                          * btree node isn't there anymore, the read path will
 699                          * sort it out:
 700                          */
 701                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 702                         bkey_for_each_ptr(ptrs, ptr) {
 703                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 704                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 705
 706                                 ptr->gen = g->gen;
 707                         }
 708                 } else {
 709                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 710                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 711                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 712                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 713
 714                                 (ptr->cached &&
 715                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 716                                 (!ptr->cached &&
 717                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 718                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 719                                 (g->data_type &&
 720                                  g->data_type != data_type);
 721                         }));
 722 again:
 723                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 724                         bkey_extent_entry_for_each(ptrs, entry) {
 725                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 726                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 727                                                                         entry->stripe_ptr.idx);
 728                                         union bch_extent_entry *next_ptr;
 729
 730                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 731                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 732                                                         goto found;
 733                                         next_ptr = NULL;
 734 found:
 735                                         if (!next_ptr) {
 736                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 737                                                 continue;
 738                                         }
 739
 740                                         if (!m || !m->alive ||
 741                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 742                                                                        &next_ptr->ptr,
 743                                                                        m->sectors)) {
 744                                                 bch2_bkey_extent_entry_drop(new, entry);
 745                                                 goto again;
 746                                         }
 747                                 }
 748                         }
 749                 }
 750
 751                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 752                 if (ret) {
 753                         kfree(new);
 754                         goto err;
 755                 }
 756
 757                 if (level)
 758                         bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
 759
 760                 if (0) {
 761                         printbuf_reset(&buf);
 762                         bch2_bkey_val_to_text(&buf, c, *k);
 763                         bch_info(c, "updated %s", buf.buf);
 764
 765                         printbuf_reset(&buf);
 766                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 767                         bch_info(c, "new key %s", buf.buf);
 768                 }
 769
 770                 *k = bkey_i_to_s_c(new);
 771         }
 772 err:
 773 fsck_err:
 774         printbuf_exit(&buf);
 775         return ret;
 776 }
 777
 778 /* marking of btree keys/nodes: */
 779
 780 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 781                             unsigned level, bool is_root,
 782                             struct bkey_s_c *k,
 783                             bool initial)
 784 {
 785         struct bch_fs *c = trans->c;
 786         struct bkey deleted = KEY(0, 0, 0);
 787         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 788         unsigned flags =
 789                 BTREE_TRIGGER_GC|
 790                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 791         int ret = 0;
 792
 793         deleted.p = k->k->p;
 794
 795         if (initial) {
 796                 BUG_ON(bch2_journal_seq_verify &&
 797                        k->k->version.lo > atomic64_read(&c->journal.seq));
 798
 799                 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
 800                 if (ret)
 801                         goto err;
 802
 803                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 804                                 "key version number higher than recorded: %llu > %llu",
 805                                 k->k->version.lo,
 806                                 atomic64_read(&c->key_version)))
 807                         atomic64_set(&c->key_version, k->k->version.lo);
 808         }
 809
 810         ret = commit_do(trans, NULL, NULL, 0,
 811                         bch2_mark_key(trans, btree_id, level, old, *k, flags));
 812 fsck_err:
 813 err:
 814         if (ret)
 815                 bch_err_fn(c, ret);
 816         return ret;
 817 }
 818
 819 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 820 {
 821         struct bch_fs *c = trans->c;
 822         struct btree_node_iter iter;
 823         struct bkey unpacked;
 824         struct bkey_s_c k;
 825         struct bkey_buf prev, cur;
 826         int ret = 0;
 827
 828         if (!btree_node_type_needs_gc(btree_node_type(b)))
 829                 return 0;
 830
 831         bch2_btree_node_iter_init_from_start(&iter, b);
 832         bch2_bkey_buf_init(&prev);
 833         bch2_bkey_buf_init(&cur);
 834         bkey_init(&prev.k->k);
 835
 836         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 837                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 838                                        &k, initial);
 839                 if (ret)
 840                         break;
 841
 842                 bch2_btree_node_iter_advance(&iter, b);
 843
 844                 if (b->c.level) {
 845                         bch2_bkey_buf_reassemble(&cur, c, k);
 846
 847                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 848                                         bch2_btree_node_iter_end(&iter));
 849                         if (ret)
 850                                 break;
 851                 }
 852         }
 853
 854         bch2_bkey_buf_exit(&cur, c);
 855         bch2_bkey_buf_exit(&prev, c);
 856         return ret;
 857 }
 858
 859 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 860                          bool initial, bool metadata_only)
 861 {
 862         struct bch_fs *c = trans->c;
 863         struct btree_iter iter;
 864         struct btree *b;
 865         unsigned depth = metadata_only ? 1 : 0;
 866         int ret = 0;
 867
 868         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 869
 870         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 871                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 872                 bch2_verify_btree_nr_keys(b);
 873
 874                 gc_pos_set(c, gc_pos_btree_node(b));
 875
 876                 ret = btree_gc_mark_node(trans, b, initial);
 877                 if (ret)
 878                         break;
 879         }
 880         bch2_trans_iter_exit(trans, &iter);
 881
 882         if (ret)
 883                 return ret;
 884
 885         mutex_lock(&c->btree_root_lock);
 886         b = c->btree_roots[btree_id].b;
 887         if (!btree_node_fake(b)) {
 888                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 889
 890                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
 891                                        true, &k, initial);
 892         }
 893         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 894         mutex_unlock(&c->btree_root_lock);
 895
 896         return ret;
 897 }
 898
 899 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 900                                       unsigned target_depth)
 901 {
 902         struct bch_fs *c = trans->c;
 903         struct btree_and_journal_iter iter;
 904         struct bkey_s_c k;
 905         struct bkey_buf cur, prev;
 906         struct printbuf buf = PRINTBUF;
 907         int ret = 0;
 908
 909         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 910         bch2_bkey_buf_init(&prev);
 911         bch2_bkey_buf_init(&cur);
 912         bkey_init(&prev.k->k);
 913
 914         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 915                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 916                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 917
 918                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 919                                        false, &k, true);
 920                 if (ret)
 921                         goto fsck_err;
 922
 923                 if (b->c.level) {
 924                         bch2_bkey_buf_reassemble(&cur, c, k);
 925                         k = bkey_i_to_s_c(cur.k);
 926
 927                         bch2_btree_and_journal_iter_advance(&iter);
 928
 929                         ret = bch2_gc_check_topology(c, b,
 930                                         &prev, cur,
 931                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 932                         if (ret)
 933                                 goto fsck_err;
 934                 } else {
 935                         bch2_btree_and_journal_iter_advance(&iter);
 936                 }
 937         }
 938
 939         if (b->c.level > target_depth) {
 940                 bch2_btree_and_journal_iter_exit(&iter);
 941                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 942
 943                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 944                         struct btree *child;
 945
 946                         bch2_bkey_buf_reassemble(&cur, c, k);
 947                         bch2_btree_and_journal_iter_advance(&iter);
 948
 949                         child = bch2_btree_node_get_noiter(trans, cur.k,
 950                                                 b->c.btree_id, b->c.level - 1,
 951                                                 false);
 952                         ret = PTR_ERR_OR_ZERO(child);
 953
 954                         if (ret == -EIO) {
 955                                 bch2_topology_error(c);
 956
 957                                 if (__fsck_err(c,
 958                                           FSCK_CAN_FIX|
 959                                           FSCK_CAN_IGNORE|
 960                                           FSCK_NO_RATELIMIT,
 961                                           "Unreadable btree node at btree %s level %u:\n"
 962                                           "  %s",
 963                                           bch2_btree_ids[b->c.btree_id],
 964                                           b->c.level - 1,
 965                                           (printbuf_reset(&buf),
 966                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 967                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 968                                         ret = -BCH_ERR_need_topology_repair;
 969                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 970                                         goto fsck_err;
 971                                 } else {
 972                                         /* Continue marking when opted to not
 973                                          * fix the error: */
 974                                         ret = 0;
 975                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 976                                         continue;
 977                                 }
 978                         } else if (ret) {
 979                                 bch_err_msg(c, ret, "getting btree node");
 980                                 break;
 981                         }
 982
 983                         ret = bch2_gc_btree_init_recurse(trans, child,
 984                                                          target_depth);
 985                         six_unlock_read(&child->c.lock);
 986
 987                         if (ret)
 988                                 break;
 989                 }
 990         }
 991 fsck_err:
 992         bch2_bkey_buf_exit(&cur, c);
 993         bch2_bkey_buf_exit(&prev, c);
 994         bch2_btree_and_journal_iter_exit(&iter);
 995         printbuf_exit(&buf);
 996         return ret;
 997 }
 998
 999 static int bch2_gc_btree_init(struct btree_trans *trans,
1000                               enum btree_id btree_id,
1001                               bool metadata_only)
1002 {
1003         struct bch_fs *c = trans->c;
1004         struct btree *b;
1005         unsigned target_depth = metadata_only ? 1 : 0;
1006         struct printbuf buf = PRINTBUF;
1007         int ret = 0;
1008
1009         b = c->btree_roots[btree_id].b;
1010
1011         if (btree_node_fake(b))
1012                 return 0;
1013
1014         six_lock_read(&b->c.lock, NULL, NULL);
1015         printbuf_reset(&buf);
1016         bch2_bpos_to_text(&buf, b->data->min_key);
1017         if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1018                         "btree root with incorrect min_key: %s", buf.buf)) {
1019                 bch_err(c, "repair unimplemented");
1020                 ret = -BCH_ERR_fsck_repair_unimplemented;
1021                 goto fsck_err;
1022         }
1023
1024         printbuf_reset(&buf);
1025         bch2_bpos_to_text(&buf, b->data->max_key);
1026         if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1027                         "btree root with incorrect max_key: %s", buf.buf)) {
1028                 bch_err(c, "repair unimplemented");
1029                 ret = -BCH_ERR_fsck_repair_unimplemented;
1030                 goto fsck_err;
1031         }
1032
1033         if (b->c.level >= target_depth)
1034                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1035
1036         if (!ret) {
1037                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1038
1039                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1040                                        &k, true);
1041         }
1042 fsck_err:
1043         six_unlock_read(&b->c.lock);
1044
1045         if (ret < 0)
1046                 bch_err_fn(c, ret);
1047         printbuf_exit(&buf);
1048         return ret;
1049 }
1050
1051 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1052 {
1053         return  (int) btree_id_to_gc_phase(l) -
1054                 (int) btree_id_to_gc_phase(r);
1055 }
1056
1057 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1058 {
1059         struct btree_trans trans;
1060         enum btree_id ids[BTREE_ID_NR];
1061         unsigned i;
1062         int ret = 0;
1063
1064         bch2_trans_init(&trans, c, 0, 0);
1065
1066         if (initial)
1067                 trans.is_initial_gc = true;
1068
1069         for (i = 0; i < BTREE_ID_NR; i++)
1070                 ids[i] = i;
1071         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1072
1073         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1074                 ret = initial
1075                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1076                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1077
1078         if (ret < 0)
1079                 bch_err_fn(c, ret);
1080
1081         bch2_trans_exit(&trans);
1082         return ret;
1083 }
1084
1085 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1086                                   u64 start, u64 end,
1087                                   enum bch_data_type type,
1088                                   unsigned flags)
1089 {
1090         u64 b = sector_to_bucket(ca, start);
1091
1092         do {
1093                 unsigned sectors =
1094                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1095
1096                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1097                                           gc_phase(GC_PHASE_SB), flags);
1098                 b++;
1099                 start += sectors;
1100         } while (start < end);
1101 }
1102
1103 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1104                                      unsigned flags)
1105 {
1106         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1107         unsigned i;
1108         u64 b;
1109
1110         for (i = 0; i < layout->nr_superblocks; i++) {
1111                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1112
1113                 if (offset == BCH_SB_SECTOR)
1114                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1115                                               BCH_DATA_sb, flags);
1116
1117                 mark_metadata_sectors(c, ca, offset,
1118                                       offset + (1 << layout->sb_max_size_bits),
1119                                       BCH_DATA_sb, flags);
1120         }
1121
1122         for (i = 0; i < ca->journal.nr; i++) {
1123                 b = ca->journal.buckets[i];
1124                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1125                                           ca->mi.bucket_size,
1126                                           gc_phase(GC_PHASE_SB), flags);
1127         }
1128 }
1129
1130 static void bch2_mark_superblocks(struct bch_fs *c)
1131 {
1132         struct bch_dev *ca;
1133         unsigned i;
1134
1135         mutex_lock(&c->sb_lock);
1136         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1137
1138         for_each_online_member(ca, c, i)
1139                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1140         mutex_unlock(&c->sb_lock);
1141 }
1142
1143 #if 0
1144 /* Also see bch2_pending_btree_node_free_insert_done() */
1145 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1146 {
1147         struct btree_update *as;
1148         struct pending_btree_node_free *d;
1149
1150         mutex_lock(&c->btree_interior_update_lock);
1151         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1152
1153         for_each_pending_btree_node_free(c, as, d)
1154                 if (d->index_update_done)
1155                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1156
1157         mutex_unlock(&c->btree_interior_update_lock);
1158 }
1159 #endif
1160
1161 static void bch2_gc_free(struct bch_fs *c)
1162 {
1163         struct bch_dev *ca;
1164         unsigned i;
1165
1166         genradix_free(&c->reflink_gc_table);
1167         genradix_free(&c->gc_stripes);
1168
1169         for_each_member_device(ca, c, i) {
1170                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1171                         sizeof(struct bucket_array) +
1172                         ca->mi.nbuckets * sizeof(struct bucket));
1173                 ca->buckets_gc = NULL;
1174
1175                 free_percpu(ca->usage_gc);
1176                 ca->usage_gc = NULL;
1177         }
1178
1179         free_percpu(c->usage_gc);
1180         c->usage_gc = NULL;
1181 }
1182
1183 static int bch2_gc_done(struct bch_fs *c,
1184                         bool initial, bool metadata_only)
1185 {
1186         struct bch_dev *ca = NULL;
1187         struct printbuf buf = PRINTBUF;
1188         bool verify = !metadata_only &&
1189                 !c->opts.reconstruct_alloc &&
1190                 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1191         unsigned i, dev;
1192         int ret = 0;
1193
1194         percpu_down_write(&c->mark_lock);
1195
1196 #define copy_field(_f, _msg, ...)                                       \
1197         if (dst->_f != src->_f &&                                       \
1198             (!verify ||                                                 \
1199              fsck_err(c, _msg ": got %llu, should be %llu"              \
1200                       , ##__VA_ARGS__, dst->_f, src->_f)))              \
1201                 dst->_f = src->_f
1202 #define copy_stripe_field(_f, _msg, ...)                                \
1203         if (dst->_f != src->_f &&                                       \
1204             (!verify ||                                                 \
1205              fsck_err(c, "stripe %zu has wrong "_msg                    \
1206                       ": got %u, should be %u",                         \
1207                       iter.pos, ##__VA_ARGS__,                          \
1208                       dst->_f, src->_f)))                               \
1209                 dst->_f = src->_f
1210 #define copy_dev_field(_f, _msg, ...)                                   \
1211         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1212 #define copy_fs_field(_f, _msg, ...)                                    \
1213         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1214
1215         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1216                 bch2_fs_usage_acc_to_base(c, i);
1217
1218         for_each_member_device(ca, c, dev) {
1219                 struct bch_dev_usage *dst = ca->usage_base;
1220                 struct bch_dev_usage *src = (void *)
1221                         bch2_acc_percpu_u64s((void *) ca->usage_gc,
1222                                              dev_usage_u64s());
1223
1224                 copy_dev_field(buckets_ec,              "buckets_ec");
1225
1226                 for (i = 0; i < BCH_DATA_NR; i++) {
1227                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1228                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1229                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1230                 }
1231         };
1232
1233         {
1234                 unsigned nr = fs_usage_u64s(c);
1235                 struct bch_fs_usage *dst = c->usage_base;
1236                 struct bch_fs_usage *src = (void *)
1237                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1238
1239                 copy_fs_field(hidden,           "hidden");
1240                 copy_fs_field(btree,            "btree");
1241
1242                 if (!metadata_only) {
1243                         copy_fs_field(data,     "data");
1244                         copy_fs_field(cached,   "cached");
1245                         copy_fs_field(reserved, "reserved");
1246                         copy_fs_field(nr_inodes,"nr_inodes");
1247
1248                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1249                                 copy_fs_field(persistent_reserved[i],
1250                                               "persistent_reserved[%i]", i);
1251                 }
1252
1253                 for (i = 0; i < c->replicas.nr; i++) {
1254                         struct bch_replicas_entry *e =
1255                                 cpu_replicas_entry(&c->replicas, i);
1256
1257                         if (metadata_only &&
1258                             (e->data_type == BCH_DATA_user ||
1259                              e->data_type == BCH_DATA_cached))
1260                                 continue;
1261
1262                         printbuf_reset(&buf);
1263                         bch2_replicas_entry_to_text(&buf, e);
1264
1265                         copy_fs_field(replicas[i], "%s", buf.buf);
1266                 }
1267         }
1268
1269 #undef copy_fs_field
1270 #undef copy_dev_field
1271 #undef copy_stripe_field
1272 #undef copy_field
1273 fsck_err:
1274         if (ca)
1275                 percpu_ref_put(&ca->ref);
1276         if (ret)
1277                 bch_err_fn(c, ret);
1278
1279         percpu_up_write(&c->mark_lock);
1280         printbuf_exit(&buf);
1281         return ret;
1282 }
1283
1284 static int bch2_gc_start(struct bch_fs *c)
1285 {
1286         struct bch_dev *ca = NULL;
1287         unsigned i;
1288
1289         BUG_ON(c->usage_gc);
1290
1291         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1292                                          sizeof(u64), GFP_KERNEL);
1293         if (!c->usage_gc) {
1294                 bch_err(c, "error allocating c->usage_gc");
1295                 return -BCH_ERR_ENOMEM_gc_start;
1296         }
1297
1298         for_each_member_device(ca, c, i) {
1299                 BUG_ON(ca->usage_gc);
1300
1301                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1302                 if (!ca->usage_gc) {
1303                         bch_err(c, "error allocating ca->usage_gc");
1304                         percpu_ref_put(&ca->ref);
1305                         return -BCH_ERR_ENOMEM_gc_start;
1306                 }
1307
1308                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1309                                ca->mi.nbuckets - ca->mi.first_bucket);
1310         }
1311
1312         return 0;
1313 }
1314
1315 static int bch2_gc_reset(struct bch_fs *c)
1316 {
1317         struct bch_dev *ca;
1318         unsigned i;
1319
1320         for_each_member_device(ca, c, i) {
1321                 free_percpu(ca->usage_gc);
1322                 ca->usage_gc = NULL;
1323         }
1324
1325         free_percpu(c->usage_gc);
1326         c->usage_gc = NULL;
1327
1328         return bch2_gc_start(c);
1329 }
1330
1331 /* returns true if not equal */
1332 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1333                                      struct bch_alloc_v4 r)
1334 {
1335         return  l.gen != r.gen                          ||
1336                 l.oldest_gen != r.oldest_gen            ||
1337                 l.data_type != r.data_type              ||
1338                 l.dirty_sectors != r.dirty_sectors      ||
1339                 l.cached_sectors != r.cached_sectors     ||
1340                 l.stripe_redundancy != r.stripe_redundancy ||
1341                 l.stripe != r.stripe;
1342 }
1343
1344 static int bch2_alloc_write_key(struct btree_trans *trans,
1345                                 struct btree_iter *iter,
1346                                 struct bkey_s_c k,
1347                                 bool metadata_only)
1348 {
1349         struct bch_fs *c = trans->c;
1350         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1351         struct bucket gc, *b;
1352         struct bkey_i_alloc_v4 *a;
1353         struct bch_alloc_v4 old_convert, new;
1354         const struct bch_alloc_v4 *old;
1355         enum bch_data_type type;
1356         int ret;
1357
1358         if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
1359                 return 1;
1360
1361         old = bch2_alloc_to_v4(k, &old_convert);
1362         new = *old;
1363
1364         percpu_down_read(&c->mark_lock);
1365         b = gc_bucket(ca, iter->pos.offset);
1366
1367         /*
1368          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1369          * fix that here:
1370          */
1371         type = __alloc_data_type(b->dirty_sectors,
1372                                  b->cached_sectors,
1373                                  b->stripe,
1374                                  *old,
1375                                  b->data_type);
1376         if (b->data_type != type) {
1377                 struct bch_dev_usage *u;
1378
1379                 preempt_disable();
1380                 u = this_cpu_ptr(ca->usage_gc);
1381                 u->d[b->data_type].buckets--;
1382                 b->data_type = type;
1383                 u->d[b->data_type].buckets++;
1384                 preempt_enable();
1385         }
1386
1387         gc = *b;
1388         percpu_up_read(&c->mark_lock);
1389
1390         if (metadata_only &&
1391             gc.data_type != BCH_DATA_sb &&
1392             gc.data_type != BCH_DATA_journal &&
1393             gc.data_type != BCH_DATA_btree)
1394                 return 0;
1395
1396         if (gen_after(old->gen, gc.gen))
1397                 return 0;
1398
1399         if (c->opts.reconstruct_alloc ||
1400             fsck_err_on(new.data_type != gc.data_type, c,
1401                         "bucket %llu:%llu gen %u has wrong data_type"
1402                         ": got %s, should be %s",
1403                         iter->pos.inode, iter->pos.offset,
1404                         gc.gen,
1405                         bch2_data_types[new.data_type],
1406                         bch2_data_types[gc.data_type]))
1407                 new.data_type = gc.data_type;
1408
1409 #define copy_bucket_field(_f)                                           \
1410         if (c->opts.reconstruct_alloc ||                                \
1411             fsck_err_on(new._f != gc._f, c,                             \
1412                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1413                         ": got %u, should be %u",                       \
1414                         iter->pos.inode, iter->pos.offset,              \
1415                         gc.gen,                                         \
1416                         bch2_data_types[gc.data_type],                  \
1417                         new._f, gc._f))                                 \
1418                 new._f = gc._f;                                         \
1419
1420         copy_bucket_field(gen);
1421         copy_bucket_field(dirty_sectors);
1422         copy_bucket_field(cached_sectors);
1423         copy_bucket_field(stripe_redundancy);
1424         copy_bucket_field(stripe);
1425 #undef copy_bucket_field
1426
1427         if (!bch2_alloc_v4_cmp(*old, new))
1428                 return 0;
1429
1430         a = bch2_alloc_to_v4_mut(trans, k);
1431         ret = PTR_ERR_OR_ZERO(a);
1432         if (ret)
1433                 return ret;
1434
1435         a->v = new;
1436
1437         /*
1438          * The trigger normally makes sure this is set, but we're not running
1439          * triggers:
1440          */
1441         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1442                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1443
1444         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1445 fsck_err:
1446         return ret;
1447 }
1448
1449 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1450 {
1451         struct btree_trans trans;
1452         struct btree_iter iter;
1453         struct bkey_s_c k;
1454         struct bch_dev *ca;
1455         unsigned i;
1456         int ret = 0;
1457
1458         bch2_trans_init(&trans, c, 0, 0);
1459
1460         for_each_member_device(ca, c, i) {
1461                 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1462                                 POS(ca->dev_idx, ca->mi.first_bucket),
1463                                 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1464                                 NULL, NULL, BTREE_INSERT_LAZY_RW,
1465                         bch2_alloc_write_key(&trans, &iter, k, metadata_only));
1466
1467                 if (ret < 0) {
1468                         bch_err(c, "error writing alloc info: %s", bch2_err_str(ret));
1469                         percpu_ref_put(&ca->ref);
1470                         break;
1471                 }
1472         }
1473
1474         bch2_trans_exit(&trans);
1475         return ret < 0 ? ret : 0;
1476 }
1477
1478 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1479 {
1480         struct bch_dev *ca;
1481         struct btree_trans trans;
1482         struct btree_iter iter;
1483         struct bkey_s_c k;
1484         struct bucket *g;
1485         struct bch_alloc_v4 a_convert;
1486         const struct bch_alloc_v4 *a;
1487         unsigned i;
1488         int ret;
1489
1490         for_each_member_device(ca, c, i) {
1491                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1492                                 ca->mi.nbuckets * sizeof(struct bucket),
1493                                 GFP_KERNEL|__GFP_ZERO);
1494                 if (!buckets) {
1495                         percpu_ref_put(&ca->ref);
1496                         bch_err(c, "error allocating ca->buckets[gc]");
1497                         return -BCH_ERR_ENOMEM_gc_alloc_start;
1498                 }
1499
1500                 buckets->first_bucket   = ca->mi.first_bucket;
1501                 buckets->nbuckets       = ca->mi.nbuckets;
1502                 rcu_assign_pointer(ca->buckets_gc, buckets);
1503         };
1504
1505         bch2_trans_init(&trans, c, 0, 0);
1506
1507         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1508                            BTREE_ITER_PREFETCH, k, ret) {
1509                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1510                 g = gc_bucket(ca, k.k->p.offset);
1511
1512                 a = bch2_alloc_to_v4(k, &a_convert);
1513
1514                 g->gen_valid    = 1;
1515                 g->gen          = a->gen;
1516
1517                 if (metadata_only &&
1518                     (a->data_type == BCH_DATA_user ||
1519                      a->data_type == BCH_DATA_cached ||
1520                      a->data_type == BCH_DATA_parity)) {
1521                         g->data_type            = a->data_type;
1522                         g->dirty_sectors        = a->dirty_sectors;
1523                         g->cached_sectors       = a->cached_sectors;
1524                         g->stripe               = a->stripe;
1525                         g->stripe_redundancy    = a->stripe_redundancy;
1526                 }
1527         }
1528         bch2_trans_iter_exit(&trans, &iter);
1529
1530         bch2_trans_exit(&trans);
1531
1532         if (ret)
1533                 bch_err(c, "error reading alloc info at gc start: %s", bch2_err_str(ret));
1534
1535         return ret;
1536 }
1537
1538 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1539 {
1540         struct bch_dev *ca;
1541         unsigned i;
1542
1543         for_each_member_device(ca, c, i) {
1544                 struct bucket_array *buckets = gc_bucket_array(ca);
1545                 struct bucket *g;
1546
1547                 for_each_bucket(g, buckets) {
1548                         if (metadata_only &&
1549                             (g->data_type == BCH_DATA_user ||
1550                              g->data_type == BCH_DATA_cached ||
1551                              g->data_type == BCH_DATA_parity))
1552                                 continue;
1553                         g->data_type = 0;
1554                         g->dirty_sectors = 0;
1555                         g->cached_sectors = 0;
1556                 }
1557         };
1558 }
1559
1560 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1561                                      struct btree_iter *iter,
1562                                      struct bkey_s_c k,
1563                                      size_t *idx)
1564 {
1565         struct bch_fs *c = trans->c;
1566         const __le64 *refcount = bkey_refcount_c(k);
1567         struct printbuf buf = PRINTBUF;
1568         struct reflink_gc *r;
1569         int ret = 0;
1570
1571         if (!refcount)
1572                 return 0;
1573
1574         while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1575                r->offset < k.k->p.offset)
1576                 ++*idx;
1577
1578         if (!r ||
1579             r->offset != k.k->p.offset ||
1580             r->size != k.k->size) {
1581                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1582                 return -EINVAL;
1583         }
1584
1585         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1586                         "reflink key has wrong refcount:\n"
1587                         "  %s\n"
1588                         "  should be %u",
1589                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1590                         r->refcount)) {
1591                 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1592
1593                 ret = PTR_ERR_OR_ZERO(new);
1594                 if (ret)
1595                         return ret;
1596
1597                 if (!r->refcount)
1598                         new->k.type = KEY_TYPE_deleted;
1599                 else
1600                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1601         }
1602 fsck_err:
1603         printbuf_exit(&buf);
1604         return ret;
1605 }
1606
1607 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1608 {
1609         struct btree_trans trans;
1610         struct btree_iter iter;
1611         struct bkey_s_c k;
1612         size_t idx = 0;
1613         int ret = 0;
1614
1615         if (metadata_only)
1616                 return 0;
1617
1618         bch2_trans_init(&trans, c, 0, 0);
1619
1620         ret = for_each_btree_key_commit(&trans, iter,
1621                         BTREE_ID_reflink, POS_MIN,
1622                         BTREE_ITER_PREFETCH, k,
1623                         NULL, NULL, BTREE_INSERT_NOFAIL,
1624                 bch2_gc_write_reflink_key(&trans, &iter, k, &idx));
1625
1626         c->reflink_gc_nr = 0;
1627         bch2_trans_exit(&trans);
1628         return ret;
1629 }
1630
1631 static int bch2_gc_reflink_start(struct bch_fs *c,
1632                                  bool metadata_only)
1633 {
1634         struct btree_trans trans;
1635         struct btree_iter iter;
1636         struct bkey_s_c k;
1637         struct reflink_gc *r;
1638         int ret = 0;
1639
1640         if (metadata_only)
1641                 return 0;
1642
1643         bch2_trans_init(&trans, c, 0, 0);
1644         c->reflink_gc_nr = 0;
1645
1646         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1647                            BTREE_ITER_PREFETCH, k, ret) {
1648                 const __le64 *refcount = bkey_refcount_c(k);
1649
1650                 if (!refcount)
1651                         continue;
1652
1653                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1654                                        GFP_KERNEL);
1655                 if (!r) {
1656                         ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1657                         break;
1658                 }
1659
1660                 r->offset       = k.k->p.offset;
1661                 r->size         = k.k->size;
1662                 r->refcount     = 0;
1663         }
1664         bch2_trans_iter_exit(&trans, &iter);
1665
1666         bch2_trans_exit(&trans);
1667         return ret;
1668 }
1669
1670 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1671 {
1672         struct genradix_iter iter;
1673         struct reflink_gc *r;
1674
1675         genradix_for_each(&c->reflink_gc_table, iter, r)
1676                 r->refcount = 0;
1677 }
1678
1679 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1680                                      struct btree_iter *iter,
1681                                      struct bkey_s_c k)
1682 {
1683         struct bch_fs *c = trans->c;
1684         struct printbuf buf = PRINTBUF;
1685         const struct bch_stripe *s;
1686         struct gc_stripe *m;
1687         bool bad = false;
1688         unsigned i;
1689         int ret = 0;
1690
1691         if (k.k->type != KEY_TYPE_stripe)
1692                 return 0;
1693
1694         s = bkey_s_c_to_stripe(k).v;
1695         m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1696
1697         for (i = 0; i < s->nr_blocks; i++) {
1698                 u32 old = stripe_blockcount_get(s, i);
1699                 u32 new = (m ? m->block_sectors[i] : 0);
1700
1701                 if (old != new) {
1702                         prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1703                                    i, old, new);
1704                         bad = true;
1705                 }
1706         }
1707
1708         if (bad)
1709                 bch2_bkey_val_to_text(&buf, c, k);
1710
1711         if (fsck_err_on(bad, c, "%s", buf.buf)) {
1712                 struct bkey_i_stripe *new;
1713
1714                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1715                 ret = PTR_ERR_OR_ZERO(new);
1716                 if (ret)
1717                         return ret;
1718
1719                 bkey_reassemble(&new->k_i, k);
1720
1721                 for (i = 0; i < new->v.nr_blocks; i++)
1722                         stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1723
1724                 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1725         }
1726 fsck_err:
1727         printbuf_exit(&buf);
1728         return ret;
1729 }
1730
1731 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1732 {
1733         struct btree_trans trans;
1734         struct btree_iter iter;
1735         struct bkey_s_c k;
1736         int ret = 0;
1737
1738         if (metadata_only)
1739                 return 0;
1740
1741         bch2_trans_init(&trans, c, 0, 0);
1742
1743         ret = for_each_btree_key_commit(&trans, iter,
1744                         BTREE_ID_stripes, POS_MIN,
1745                         BTREE_ITER_PREFETCH, k,
1746                         NULL, NULL, BTREE_INSERT_NOFAIL,
1747                 bch2_gc_write_stripes_key(&trans, &iter, k));
1748
1749         bch2_trans_exit(&trans);
1750         return ret;
1751 }
1752
1753 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1754 {
1755         genradix_free(&c->gc_stripes);
1756 }
1757
1758 /**
1759  * bch2_gc - walk _all_ references to buckets, and recompute them:
1760  *
1761  * Order matters here:
1762  *  - Concurrent GC relies on the fact that we have a total ordering for
1763  *    everything that GC walks - see  gc_will_visit_node(),
1764  *    gc_will_visit_root()
1765  *
1766  *  - also, references move around in the course of index updates and
1767  *    various other crap: everything needs to agree on the ordering
1768  *    references are allowed to move around in - e.g., we're allowed to
1769  *    start with a reference owned by an open_bucket (the allocator) and
1770  *    move it to the btree, but not the reverse.
1771  *
1772  *    This is necessary to ensure that gc doesn't miss references that
1773  *    move around - if references move backwards in the ordering GC
1774  *    uses, GC could skip past them
1775  */
1776 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1777 {
1778         unsigned iter = 0;
1779         int ret;
1780
1781         lockdep_assert_held(&c->state_lock);
1782
1783         down_write(&c->gc_lock);
1784
1785         bch2_btree_interior_updates_flush(c);
1786
1787         ret   = bch2_gc_start(c) ?:
1788                 bch2_gc_alloc_start(c, metadata_only) ?:
1789                 bch2_gc_reflink_start(c, metadata_only);
1790         if (ret)
1791                 goto out;
1792 again:
1793         gc_pos_set(c, gc_phase(GC_PHASE_START));
1794
1795         bch2_mark_superblocks(c);
1796
1797         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) ||
1798             (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1799              !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1800              c->opts.fix_errors != FSCK_OPT_NO)) {
1801                 bch_info(c, "Starting topology repair pass");
1802                 ret = bch2_repair_topology(c);
1803                 if (ret)
1804                         goto out;
1805                 bch_info(c, "Topology repair pass done");
1806
1807                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1808         }
1809
1810         ret = bch2_gc_btrees(c, initial, metadata_only);
1811
1812         if (ret == -BCH_ERR_need_topology_repair &&
1813             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1814             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1815                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1816                 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1817                 ret = 0;
1818         }
1819
1820         if (ret == -BCH_ERR_need_topology_repair)
1821                 ret = -BCH_ERR_fsck_errors_not_fixed;
1822
1823         if (ret)
1824                 goto out;
1825
1826 #if 0
1827         bch2_mark_pending_btree_node_frees(c);
1828 #endif
1829         c->gc_count++;
1830
1831         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1832             (!iter && bch2_test_restart_gc)) {
1833                 if (iter++ > 2) {
1834                         bch_info(c, "Unable to fix bucket gens, looping");
1835                         ret = -EINVAL;
1836                         goto out;
1837                 }
1838
1839                 /*
1840                  * XXX: make sure gens we fixed got saved
1841                  */
1842                 bch_info(c, "Second GC pass needed, restarting:");
1843                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1844                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1845
1846                 bch2_gc_stripes_reset(c, metadata_only);
1847                 bch2_gc_alloc_reset(c, metadata_only);
1848                 bch2_gc_reflink_reset(c, metadata_only);
1849                 ret = bch2_gc_reset(c);
1850                 if (ret)
1851                         goto out;
1852
1853                 /* flush fsck errors, reset counters */
1854                 bch2_flush_fsck_errs(c);
1855                 goto again;
1856         }
1857 out:
1858         if (!ret) {
1859                 bch2_journal_block(&c->journal);
1860
1861                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1862                         bch2_gc_reflink_done(c, metadata_only) ?:
1863                         bch2_gc_alloc_done(c, metadata_only) ?:
1864                         bch2_gc_done(c, initial, metadata_only);
1865
1866                 bch2_journal_unblock(&c->journal);
1867         }
1868
1869         percpu_down_write(&c->mark_lock);
1870         /* Indicates that gc is no longer in progress: */
1871         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1872
1873         bch2_gc_free(c);
1874         percpu_up_write(&c->mark_lock);
1875
1876         up_write(&c->gc_lock);
1877
1878         /*
1879          * At startup, allocations can happen directly instead of via the
1880          * allocator thread - issue wakeup in case they blocked on gc_lock:
1881          */
1882         closure_wake_up(&c->freelist_wait);
1883
1884         if (ret)
1885                 bch_err_fn(c, ret);
1886         return ret;
1887 }
1888
1889 static int gc_btree_gens_key(struct btree_trans *trans,
1890                              struct btree_iter *iter,
1891                              struct bkey_s_c k)
1892 {
1893         struct bch_fs *c = trans->c;
1894         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1895         const struct bch_extent_ptr *ptr;
1896         struct bkey_i *u;
1897         int ret;
1898
1899         percpu_down_read(&c->mark_lock);
1900         bkey_for_each_ptr(ptrs, ptr) {
1901                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1902
1903                 if (ptr_stale(ca, ptr) > 16) {
1904                         percpu_up_read(&c->mark_lock);
1905                         goto update;
1906                 }
1907         }
1908
1909         bkey_for_each_ptr(ptrs, ptr) {
1910                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1911                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1912
1913                 if (gen_after(*gen, ptr->gen))
1914                         *gen = ptr->gen;
1915         }
1916         percpu_up_read(&c->mark_lock);
1917         return 0;
1918 update:
1919         u = bch2_bkey_make_mut(trans, iter, &k, 0);
1920         ret = PTR_ERR_OR_ZERO(u);
1921         if (ret)
1922                 return ret;
1923
1924         bch2_extent_normalize(c, bkey_i_to_s(u));
1925         return 0;
1926 }
1927
1928 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1929                                        struct bkey_s_c k)
1930 {
1931         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1932         struct bch_alloc_v4 a_convert;
1933         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1934         struct bkey_i_alloc_v4 *a_mut;
1935         int ret;
1936
1937         if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1938                 return 0;
1939
1940         a_mut = bch2_alloc_to_v4_mut(trans, k);
1941         ret = PTR_ERR_OR_ZERO(a_mut);
1942         if (ret)
1943                 return ret;
1944
1945         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1946         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1947
1948         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1949 }
1950
1951 int bch2_gc_gens(struct bch_fs *c)
1952 {
1953         struct btree_trans trans;
1954         struct btree_iter iter;
1955         struct bkey_s_c k;
1956         struct bch_dev *ca;
1957         u64 b, start_time = local_clock();
1958         unsigned i;
1959         int ret;
1960
1961         /*
1962          * Ideally we would be using state_lock and not gc_lock here, but that
1963          * introduces a deadlock in the RO path - we currently take the state
1964          * lock at the start of going RO, thus the gc thread may get stuck:
1965          */
1966         if (!mutex_trylock(&c->gc_gens_lock))
1967                 return 0;
1968
1969         trace_and_count(c, gc_gens_start, c);
1970         down_read(&c->gc_lock);
1971         bch2_trans_init(&trans, c, 0, 0);
1972
1973         for_each_member_device(ca, c, i) {
1974                 struct bucket_gens *gens;
1975
1976                 BUG_ON(ca->oldest_gen);
1977
1978                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1979                 if (!ca->oldest_gen) {
1980                         percpu_ref_put(&ca->ref);
1981                         ret = -BCH_ERR_ENOMEM_gc_gens;
1982                         goto err;
1983                 }
1984
1985                 gens = bucket_gens(ca);
1986
1987                 for (b = gens->first_bucket;
1988                      b < gens->nbuckets; b++)
1989                         ca->oldest_gen[b] = gens->b[b];
1990         }
1991
1992         for (i = 0; i < BTREE_ID_NR; i++)
1993                 if (btree_type_has_ptrs(i)) {
1994                         struct btree_iter iter;
1995                         struct bkey_s_c k;
1996
1997                         c->gc_gens_btree = i;
1998                         c->gc_gens_pos = POS_MIN;
1999                         ret = for_each_btree_key_commit(&trans, iter, i,
2000                                         POS_MIN,
2001                                         BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
2002                                         k,
2003                                         NULL, NULL,
2004                                         BTREE_INSERT_NOFAIL,
2005                                 gc_btree_gens_key(&trans, &iter, k));
2006                         if (ret && !bch2_err_matches(ret, EROFS))
2007                                 bch_err(c, "error recalculating oldest_gen: %s", bch2_err_str(ret));
2008                         if (ret)
2009                                 goto err;
2010                 }
2011
2012         ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
2013                         POS_MIN,
2014                         BTREE_ITER_PREFETCH,
2015                         k,
2016                         NULL, NULL,
2017                         BTREE_INSERT_NOFAIL,
2018                 bch2_alloc_write_oldest_gen(&trans, &iter, k));
2019         if (ret && !bch2_err_matches(ret, EROFS))
2020                 bch_err(c, "error writing oldest_gen: %s", bch2_err_str(ret));
2021         if (ret)
2022                 goto err;
2023
2024         c->gc_gens_btree        = 0;
2025         c->gc_gens_pos          = POS_MIN;
2026
2027         c->gc_count++;
2028
2029         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2030         trace_and_count(c, gc_gens_end, c);
2031 err:
2032         for_each_member_device(ca, c, i) {
2033                 kvfree(ca->oldest_gen);
2034                 ca->oldest_gen = NULL;
2035         }
2036
2037         bch2_trans_exit(&trans);
2038         up_read(&c->gc_lock);
2039         mutex_unlock(&c->gc_gens_lock);
2040         return ret;
2041 }
2042
2043 static int bch2_gc_thread(void *arg)
2044 {
2045         struct bch_fs *c = arg;
2046         struct io_clock *clock = &c->io_clock[WRITE];
2047         unsigned long last = atomic64_read(&clock->now);
2048         unsigned last_kick = atomic_read(&c->kick_gc);
2049         int ret;
2050
2051         set_freezable();
2052
2053         while (1) {
2054                 while (1) {
2055                         set_current_state(TASK_INTERRUPTIBLE);
2056
2057                         if (kthread_should_stop()) {
2058                                 __set_current_state(TASK_RUNNING);
2059                                 return 0;
2060                         }
2061
2062                         if (atomic_read(&c->kick_gc) != last_kick)
2063                                 break;
2064
2065                         if (c->btree_gc_periodic) {
2066                                 unsigned long next = last + c->capacity / 16;
2067
2068                                 if (atomic64_read(&clock->now) >= next)
2069                                         break;
2070
2071                                 bch2_io_clock_schedule_timeout(clock, next);
2072                         } else {
2073                                 schedule();
2074                         }
2075
2076                         try_to_freeze();
2077                 }
2078                 __set_current_state(TASK_RUNNING);
2079
2080                 last = atomic64_read(&clock->now);
2081                 last_kick = atomic_read(&c->kick_gc);
2082
2083                 /*
2084                  * Full gc is currently incompatible with btree key cache:
2085                  */
2086 #if 0
2087                 ret = bch2_gc(c, false, false);
2088 #else
2089                 ret = bch2_gc_gens(c);
2090 #endif
2091                 if (ret < 0)
2092                         bch_err(c, "btree gc failed: %s", bch2_err_str(ret));
2093
2094                 debug_check_no_locks_held();
2095         }
2096
2097         return 0;
2098 }
2099
2100 void bch2_gc_thread_stop(struct bch_fs *c)
2101 {
2102         struct task_struct *p;
2103
2104         p = c->gc_thread;
2105         c->gc_thread = NULL;
2106
2107         if (p) {
2108                 kthread_stop(p);
2109                 put_task_struct(p);
2110         }
2111 }
2112
2113 int bch2_gc_thread_start(struct bch_fs *c)
2114 {
2115         struct task_struct *p;
2116
2117         if (c->gc_thread)
2118                 return 0;
2119
2120         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2121         if (IS_ERR(p)) {
2122                 bch_err(c, "error creating gc thread: %s", bch2_err_str(PTR_ERR(p)));
2123                 return PTR_ERR(p);
2124         }
2125
2126         get_task_struct(p);
2127         c->gc_thread = p;
2128         wake_up_process(p);
2129         return 0;
2130 }