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_journal_iter.h"
  13 #include "btree_key_cache.h"
  14 #include "btree_locking.h"
  15 #include "btree_update_interior.h"
  16 #include "btree_io.h"
  17 #include "btree_gc.h"
  18 #include "buckets.h"
  19 #include "clock.h"
  20 #include "debug.h"
  21 #include "ec.h"
  22 #include "error.h"
  23 #include "extents.h"
  24 #include "journal.h"
  25 #include "keylist.h"
  26 #include "move.h"
  27 #include "recovery.h"
  28 #include "reflink.h"
  29 #include "replicas.h"
  30 #include "super-io.h"
  31 #include "trace.h"
  32
  33 #include <linux/slab.h>
  34 #include <linux/bitops.h>
  35 #include <linux/freezer.h>
  36 #include <linux/kthread.h>
  37 #include <linux/preempt.h>
  38 #include <linux/rcupdate.h>
  39 #include <linux/sched/task.h>
  40
  41 #define DROP_THIS_NODE          10
  42 #define DROP_PREV_NODE          11
  43
  44 static struct bkey_s unsafe_bkey_s_c_to_s(struct bkey_s_c k)
  45 {
  46         return (struct bkey_s) {{{
  47                 (struct bkey *) k.k,
  48                 (struct bch_val *) k.v
  49         }}};
  50 }
  51
  52 static bool should_restart_for_topology_repair(struct bch_fs *c)
  53 {
  54         return c->opts.fix_errors != FSCK_FIX_no &&
  55                 !(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
  56 }
  57
  58 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  59 {
  60         preempt_disable();
  61         write_seqcount_begin(&c->gc_pos_lock);
  62         c->gc_pos = new_pos;
  63         write_seqcount_end(&c->gc_pos_lock);
  64         preempt_enable();
  65 }
  66
  67 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  68 {
  69         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  70         __gc_pos_set(c, new_pos);
  71 }
  72
  73 /*
  74  * Missing: if an interior btree node is empty, we need to do something -
  75  * perhaps just kill it
  76  */
  77 static int bch2_gc_check_topology(struct bch_fs *c,
  78                                   struct btree *b,
  79                                   struct bkey_buf *prev,
  80                                   struct bkey_buf cur,
  81                                   bool is_last)
  82 {
  83         struct bpos node_start  = b->data->min_key;
  84         struct bpos node_end    = b->data->max_key;
  85         struct bpos expected_start = bkey_deleted(&prev->k->k)
  86                 ? node_start
  87                 : bpos_successor(prev->k->k.p);
  88         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  89         int ret = 0;
  90
  91         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  92                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  93
  94                 if (!bpos_eq(expected_start, bp->v.min_key)) {
  95                         bch2_topology_error(c);
  96
  97                         if (bkey_deleted(&prev->k->k)) {
  98                                 prt_printf(&buf1, "start of node: ");
  99                                 bch2_bpos_to_text(&buf1, node_start);
 100                         } else {
 101                                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
 102                         }
 103                         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
 104
 105                         if (__fsck_err(c,
 106                                        FSCK_CAN_FIX|
 107                                        FSCK_CAN_IGNORE|
 108                                        FSCK_NO_RATELIMIT,
 109                                        btree_node_topology_bad_min_key,
 110                                        "btree node with incorrect min_key at btree %s level %u:\n"
 111                                        "  prev %s\n"
 112                                        "  cur %s",
 113                                        bch2_btree_id_str(b->c.btree_id), b->c.level,
 114                                        buf1.buf, buf2.buf) && should_restart_for_topology_repair(c)) {
 115                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 116                                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 117                                 goto err;
 118                         } else {
 119                                 set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
 120                         }
 121                 }
 122         }
 123
 124         if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
 125                 bch2_topology_error(c);
 126
 127                 printbuf_reset(&buf1);
 128                 printbuf_reset(&buf2);
 129
 130                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
 131                 bch2_bpos_to_text(&buf2, node_end);
 132
 133                 if (__fsck_err(c, FSCK_CAN_FIX|FSCK_CAN_IGNORE|FSCK_NO_RATELIMIT,
 134                           btree_node_topology_bad_max_key,
 135                           "btree node with incorrect max_key at btree %s level %u:\n"
 136                           "  %s\n"
 137                           "  expected %s",
 138                           bch2_btree_id_str(b->c.btree_id), b->c.level,
 139                           buf1.buf, buf2.buf) &&
 140                     should_restart_for_topology_repair(c)) {
 141                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 142                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 143                         goto err;
 144                 } else {
 145                         set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
 146                 }
 147         }
 148
 149         bch2_bkey_buf_copy(prev, c, cur.k);
 150 err:
 151 fsck_err:
 152         printbuf_exit(&buf2);
 153         printbuf_exit(&buf1);
 154         return ret;
 155 }
 156
 157 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 158 {
 159         switch (b->key.k.type) {
 160         case KEY_TYPE_btree_ptr: {
 161                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 162
 163                 dst->k.p                = src->k.p;
 164                 dst->v.mem_ptr          = 0;
 165                 dst->v.seq              = b->data->keys.seq;
 166                 dst->v.sectors_written  = 0;
 167                 dst->v.flags            = 0;
 168                 dst->v.min_key          = b->data->min_key;
 169                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 170                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 171                 break;
 172         }
 173         case KEY_TYPE_btree_ptr_v2:
 174                 bkey_copy(&dst->k_i, &b->key);
 175                 break;
 176         default:
 177                 BUG();
 178         }
 179 }
 180
 181 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
 182                                              enum btree_id btree, unsigned level,
 183                                              struct bkey_s_c old, struct bkey_i *new)
 184 {
 185         struct bch_fs *c = trans->c;
 186         struct btree *b;
 187         struct bkey_buf tmp;
 188         int ret;
 189
 190         bch2_bkey_buf_init(&tmp);
 191         bch2_bkey_buf_reassemble(&tmp, c, old);
 192
 193         b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
 194         if (!IS_ERR_OR_NULL(b)) {
 195                 mutex_lock(&c->btree_cache.lock);
 196
 197                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 198
 199                 bkey_copy(&b->key, new);
 200                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 201                 BUG_ON(ret);
 202
 203                 mutex_unlock(&c->btree_cache.lock);
 204                 six_unlock_read(&b->c.lock);
 205         }
 206
 207         bch2_bkey_buf_exit(&tmp, c);
 208 }
 209
 210 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 211 {
 212         struct bkey_i_btree_ptr_v2 *new;
 213         int ret;
 214
 215         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 216         if (!new)
 217                 return -BCH_ERR_ENOMEM_gc_repair_key;
 218
 219         btree_ptr_to_v2(b, new);
 220         b->data->min_key        = new_min;
 221         new->v.min_key          = new_min;
 222         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 223
 224         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 225         if (ret) {
 226                 kfree(new);
 227                 return ret;
 228         }
 229
 230         bch2_btree_node_drop_keys_outside_node(b);
 231         bkey_copy(&b->key, &new->k_i);
 232         return 0;
 233 }
 234
 235 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 236 {
 237         struct bkey_i_btree_ptr_v2 *new;
 238         int ret;
 239
 240         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 241         if (ret)
 242                 return ret;
 243
 244         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 245         if (!new)
 246                 return -BCH_ERR_ENOMEM_gc_repair_key;
 247
 248         btree_ptr_to_v2(b, new);
 249         b->data->max_key        = new_max;
 250         new->k.p                = new_max;
 251         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 252
 253         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 254         if (ret) {
 255                 kfree(new);
 256                 return ret;
 257         }
 258
 259         bch2_btree_node_drop_keys_outside_node(b);
 260
 261         mutex_lock(&c->btree_cache.lock);
 262         bch2_btree_node_hash_remove(&c->btree_cache, b);
 263
 264         bkey_copy(&b->key, &new->k_i);
 265         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 266         BUG_ON(ret);
 267         mutex_unlock(&c->btree_cache.lock);
 268         return 0;
 269 }
 270
 271 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 272                                         struct btree *prev, struct btree *cur)
 273 {
 274         struct bpos expected_start = !prev
 275                 ? b->data->min_key
 276                 : bpos_successor(prev->key.k.p);
 277         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 278         int ret = 0;
 279
 280         if (!prev) {
 281                 prt_printf(&buf1, "start of node: ");
 282                 bch2_bpos_to_text(&buf1, b->data->min_key);
 283         } else {
 284                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
 285         }
 286
 287         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
 288
 289         if (prev &&
 290             bpos_gt(expected_start, cur->data->min_key) &&
 291             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 292                 /* cur overwrites prev: */
 293
 294                 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
 295                                                 cur->data->min_key), c,
 296                                 btree_node_topology_overwritten_by_next_node,
 297                                 "btree node overwritten by next node at btree %s level %u:\n"
 298                                 "  node %s\n"
 299                                 "  next %s",
 300                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 301                                 buf1.buf, buf2.buf)) {
 302                         ret = DROP_PREV_NODE;
 303                         goto out;
 304                 }
 305
 306                 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
 307                                                  bpos_predecessor(cur->data->min_key)), c,
 308                                 btree_node_topology_bad_max_key,
 309                                 "btree node with incorrect max_key at btree %s level %u:\n"
 310                                 "  node %s\n"
 311                                 "  next %s",
 312                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 313                                 buf1.buf, buf2.buf))
 314                         ret = set_node_max(c, prev,
 315                                            bpos_predecessor(cur->data->min_key));
 316         } else {
 317                 /* prev overwrites cur: */
 318
 319                 if (mustfix_fsck_err_on(bpos_ge(expected_start,
 320                                                 cur->data->max_key), c,
 321                                 btree_node_topology_overwritten_by_prev_node,
 322                                 "btree node overwritten by prev node at btree %s level %u:\n"
 323                                 "  prev %s\n"
 324                                 "  node %s",
 325                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 326                                 buf1.buf, buf2.buf)) {
 327                         ret = DROP_THIS_NODE;
 328                         goto out;
 329                 }
 330
 331                 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
 332                                 btree_node_topology_bad_min_key,
 333                                 "btree node with incorrect min_key at btree %s level %u:\n"
 334                                 "  prev %s\n"
 335                                 "  node %s",
 336                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 337                                 buf1.buf, buf2.buf))
 338                         ret = set_node_min(c, cur, expected_start);
 339         }
 340 out:
 341 fsck_err:
 342         printbuf_exit(&buf2);
 343         printbuf_exit(&buf1);
 344         return ret;
 345 }
 346
 347 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 348                                  struct btree *child)
 349 {
 350         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 351         int ret = 0;
 352
 353         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
 354         bch2_bpos_to_text(&buf2, b->key.k.p);
 355
 356         if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
 357                                 btree_node_topology_bad_max_key,
 358                         "btree node with incorrect max_key at btree %s level %u:\n"
 359                         "  %s\n"
 360                         "  expected %s",
 361                         bch2_btree_id_str(b->c.btree_id), b->c.level,
 362                         buf1.buf, buf2.buf)) {
 363                 ret = set_node_max(c, child, b->key.k.p);
 364                 if (ret)
 365                         goto err;
 366         }
 367 err:
 368 fsck_err:
 369         printbuf_exit(&buf2);
 370         printbuf_exit(&buf1);
 371         return ret;
 372 }
 373
 374 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
 375 {
 376         struct bch_fs *c = trans->c;
 377         struct btree_and_journal_iter iter;
 378         struct bkey_s_c k;
 379         struct bkey_buf prev_k, cur_k;
 380         struct btree *prev = NULL, *cur = NULL;
 381         bool have_child, dropped_children = false;
 382         struct printbuf buf = PRINTBUF;
 383         int ret = 0;
 384
 385         if (!b->c.level)
 386                 return 0;
 387 again:
 388         prev = NULL;
 389         have_child = dropped_children = false;
 390         bch2_bkey_buf_init(&prev_k);
 391         bch2_bkey_buf_init(&cur_k);
 392         bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
 393         iter.prefetch = true;
 394
 395         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 396                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 397                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 398
 399                 bch2_btree_and_journal_iter_advance(&iter);
 400                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 401
 402                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 403                                         b->c.btree_id, b->c.level - 1,
 404                                         false);
 405                 ret = PTR_ERR_OR_ZERO(cur);
 406
 407                 printbuf_reset(&buf);
 408                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 409
 410                 if (mustfix_fsck_err_on(bch2_err_matches(ret, EIO), c,
 411                                 btree_node_unreadable,
 412                                 "Topology repair: unreadable btree node at btree %s level %u:\n"
 413                                 "  %s",
 414                                 bch2_btree_id_str(b->c.btree_id),
 415                                 b->c.level - 1,
 416                                 buf.buf)) {
 417                         bch2_btree_node_evict(trans, cur_k.k);
 418                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 419                                                       b->c.level, cur_k.k->k.p);
 420                         cur = NULL;
 421                         if (ret)
 422                                 break;
 423                         continue;
 424                 }
 425
 426                 bch_err_msg(c, ret, "getting btree node");
 427                 if (ret)
 428                         break;
 429
 430                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 431
 432                 if (ret == DROP_THIS_NODE) {
 433                         six_unlock_read(&cur->c.lock);
 434                         bch2_btree_node_evict(trans, cur_k.k);
 435                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 436                                                       b->c.level, cur_k.k->k.p);
 437                         cur = NULL;
 438                         if (ret)
 439                                 break;
 440                         continue;
 441                 }
 442
 443                 if (prev)
 444                         six_unlock_read(&prev->c.lock);
 445                 prev = NULL;
 446
 447                 if (ret == DROP_PREV_NODE) {
 448                         bch2_btree_node_evict(trans, prev_k.k);
 449                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 450                                                       b->c.level, prev_k.k->k.p);
 451                         if (ret)
 452                                 break;
 453
 454                         bch2_btree_and_journal_iter_exit(&iter);
 455                         bch2_bkey_buf_exit(&prev_k, c);
 456                         bch2_bkey_buf_exit(&cur_k, c);
 457                         goto again;
 458                 } else if (ret)
 459                         break;
 460
 461                 prev = cur;
 462                 cur = NULL;
 463                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 464         }
 465
 466         if (!ret && !IS_ERR_OR_NULL(prev)) {
 467                 BUG_ON(cur);
 468                 ret = btree_repair_node_end(c, b, prev);
 469         }
 470
 471         if (!IS_ERR_OR_NULL(prev))
 472                 six_unlock_read(&prev->c.lock);
 473         prev = NULL;
 474         if (!IS_ERR_OR_NULL(cur))
 475                 six_unlock_read(&cur->c.lock);
 476         cur = NULL;
 477
 478         if (ret)
 479                 goto err;
 480
 481         bch2_btree_and_journal_iter_exit(&iter);
 482         bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
 483         iter.prefetch = true;
 484
 485         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 486                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 487                 bch2_btree_and_journal_iter_advance(&iter);
 488
 489                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 490                                         b->c.btree_id, b->c.level - 1,
 491                                         false);
 492                 ret = PTR_ERR_OR_ZERO(cur);
 493
 494                 bch_err_msg(c, ret, "getting btree node");
 495                 if (ret)
 496                         goto err;
 497
 498                 ret = bch2_btree_repair_topology_recurse(trans, cur);
 499                 six_unlock_read(&cur->c.lock);
 500                 cur = NULL;
 501
 502                 if (ret == DROP_THIS_NODE) {
 503                         bch2_btree_node_evict(trans, cur_k.k);
 504                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 505                                                       b->c.level, cur_k.k->k.p);
 506                         dropped_children = true;
 507                 }
 508
 509                 if (ret)
 510                         goto err;
 511
 512                 have_child = true;
 513         }
 514
 515         printbuf_reset(&buf);
 516         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 517
 518         if (mustfix_fsck_err_on(!have_child, c,
 519                         btree_node_topology_interior_node_empty,
 520                         "empty interior btree node at btree %s level %u\n"
 521                         "  %s",
 522                         bch2_btree_id_str(b->c.btree_id),
 523                         b->c.level, buf.buf))
 524                 ret = DROP_THIS_NODE;
 525 err:
 526 fsck_err:
 527         if (!IS_ERR_OR_NULL(prev))
 528                 six_unlock_read(&prev->c.lock);
 529         if (!IS_ERR_OR_NULL(cur))
 530                 six_unlock_read(&cur->c.lock);
 531
 532         bch2_btree_and_journal_iter_exit(&iter);
 533         bch2_bkey_buf_exit(&prev_k, c);
 534         bch2_bkey_buf_exit(&cur_k, c);
 535
 536         if (!ret && dropped_children)
 537                 goto again;
 538
 539         printbuf_exit(&buf);
 540         return ret;
 541 }
 542
 543 int bch2_check_topology(struct bch_fs *c)
 544 {
 545         struct btree_trans *trans = bch2_trans_get(c);
 546         struct btree *b;
 547         unsigned i;
 548         int ret = 0;
 549
 550         for (i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
 551                 struct btree_root *r = bch2_btree_id_root(c, i);
 552
 553                 if (!r->alive)
 554                         continue;
 555
 556                 b = r->b;
 557                 if (btree_node_fake(b))
 558                         continue;
 559
 560                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 561                 ret = bch2_btree_repair_topology_recurse(trans, b);
 562                 six_unlock_read(&b->c.lock);
 563
 564                 if (ret == DROP_THIS_NODE) {
 565                         bch_err(c, "empty btree root - repair unimplemented");
 566                         ret = -BCH_ERR_fsck_repair_unimplemented;
 567                 }
 568         }
 569
 570         bch2_trans_put(trans);
 571
 572         return ret;
 573 }
 574
 575 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
 576                                unsigned level, bool is_root,
 577                                struct bkey_s_c *k)
 578 {
 579         struct bch_fs *c = trans->c;
 580         struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
 581         const union bch_extent_entry *entry_c;
 582         struct extent_ptr_decoded p = { 0 };
 583         bool do_update = false;
 584         struct printbuf buf = PRINTBUF;
 585         int ret = 0;
 586
 587         /*
 588          * XXX
 589          * use check_bucket_ref here
 590          */
 591         bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
 592                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 593                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 594                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry_c->ptr);
 595
 596                 if (!g->gen_valid &&
 597                     (c->opts.reconstruct_alloc ||
 598                      fsck_err(c, ptr_to_missing_alloc_key,
 599                               "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 600                               "while marking %s",
 601                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 602                               bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
 603                               p.ptr.gen,
 604                               (printbuf_reset(&buf),
 605                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 606                         if (!p.ptr.cached) {
 607                                 g->gen_valid            = true;
 608                                 g->gen                  = p.ptr.gen;
 609                         } else {
 610                                 do_update = true;
 611                         }
 612                 }
 613
 614                 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
 615                     (c->opts.reconstruct_alloc ||
 616                      fsck_err(c, ptr_gen_newer_than_bucket_gen,
 617                               "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 618                               "while marking %s",
 619                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 620                               bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
 621                               p.ptr.gen, g->gen,
 622                               (printbuf_reset(&buf),
 623                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 624                         if (!p.ptr.cached) {
 625                                 g->gen_valid            = true;
 626                                 g->gen                  = p.ptr.gen;
 627                                 g->data_type            = 0;
 628                                 g->dirty_sectors        = 0;
 629                                 g->cached_sectors       = 0;
 630                                 set_bit(BCH_FS_need_another_gc, &c->flags);
 631                         } else {
 632                                 do_update = true;
 633                         }
 634                 }
 635
 636                 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
 637                     (c->opts.reconstruct_alloc ||
 638                      fsck_err(c, ptr_gen_newer_than_bucket_gen,
 639                               "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 640                               "while marking %s",
 641                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 642                               bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
 643                               p.ptr.gen,
 644                               (printbuf_reset(&buf),
 645                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 646                         do_update = true;
 647
 648                 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
 649                     (c->opts.reconstruct_alloc ||
 650                      fsck_err(c, stale_dirty_ptr,
 651                               "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 652                               "while marking %s",
 653                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 654                               bch2_data_type_str(ptr_data_type(k->k, &p.ptr)),
 655                               p.ptr.gen, g->gen,
 656                               (printbuf_reset(&buf),
 657                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 658                         do_update = true;
 659
 660                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 661                         continue;
 662
 663                 if (fsck_err_on(bucket_data_type(g->data_type) &&
 664                                 bucket_data_type(g->data_type) != data_type, c,
 665                                 ptr_bucket_data_type_mismatch,
 666                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 667                                 "while marking %s",
 668                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 669                                 bch2_data_type_str(g->data_type),
 670                                 bch2_data_type_str(data_type),
 671                                 (printbuf_reset(&buf),
 672                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 673                         if (data_type == BCH_DATA_btree) {
 674                                 g->data_type    = data_type;
 675                                 set_bit(BCH_FS_need_another_gc, &c->flags);
 676                         } else {
 677                                 do_update = true;
 678                         }
 679                 }
 680
 681                 if (p.has_ec) {
 682                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 683
 684                         if (fsck_err_on(!m || !m->alive, c,
 685                                         ptr_to_missing_stripe,
 686                                         "pointer to nonexistent stripe %llu\n"
 687                                         "while marking %s",
 688                                         (u64) p.ec.idx,
 689                                         (printbuf_reset(&buf),
 690                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 691                                 do_update = true;
 692
 693                         if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
 694                                         ptr_to_incorrect_stripe,
 695                                         "pointer does not match stripe %llu\n"
 696                                         "while marking %s",
 697                                         (u64) p.ec.idx,
 698                                         (printbuf_reset(&buf),
 699                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 700                                 do_update = true;
 701                 }
 702         }
 703
 704         if (do_update) {
 705                 struct bkey_ptrs ptrs;
 706                 union bch_extent_entry *entry;
 707                 struct bch_extent_ptr *ptr;
 708                 struct bkey_i *new;
 709
 710                 if (is_root) {
 711                         bch_err(c, "cannot update btree roots yet");
 712                         ret = -EINVAL;
 713                         goto err;
 714                 }
 715
 716                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 717                 if (!new) {
 718                         ret = -BCH_ERR_ENOMEM_gc_repair_key;
 719                         bch_err_msg(c, ret, "allocating new key");
 720                         goto err;
 721                 }
 722
 723                 bkey_reassemble(new, *k);
 724
 725                 if (level) {
 726                         /*
 727                          * We don't want to drop btree node pointers - if the
 728                          * btree node isn't there anymore, the read path will
 729                          * sort it out:
 730                          */
 731                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 732                         bkey_for_each_ptr(ptrs, ptr) {
 733                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 734                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 735
 736                                 ptr->gen = g->gen;
 737                         }
 738                 } else {
 739                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 740                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 741                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 742                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 743
 744                                 (ptr->cached &&
 745                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 746                                 (!ptr->cached &&
 747                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 748                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 749                                 (g->data_type &&
 750                                  g->data_type != data_type);
 751                         }));
 752 again:
 753                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 754                         bkey_extent_entry_for_each(ptrs, entry) {
 755                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 756                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 757                                                                         entry->stripe_ptr.idx);
 758                                         union bch_extent_entry *next_ptr;
 759
 760                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 761                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 762                                                         goto found;
 763                                         next_ptr = NULL;
 764 found:
 765                                         if (!next_ptr) {
 766                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 767                                                 continue;
 768                                         }
 769
 770                                         if (!m || !m->alive ||
 771                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 772                                                                        &next_ptr->ptr,
 773                                                                        m->sectors)) {
 774                                                 bch2_bkey_extent_entry_drop(new, entry);
 775                                                 goto again;
 776                                         }
 777                                 }
 778                         }
 779                 }
 780
 781                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 782                 if (ret) {
 783                         kfree(new);
 784                         goto err;
 785                 }
 786
 787                 if (level)
 788                         bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
 789
 790                 if (0) {
 791                         printbuf_reset(&buf);
 792                         bch2_bkey_val_to_text(&buf, c, *k);
 793                         bch_info(c, "updated %s", buf.buf);
 794
 795                         printbuf_reset(&buf);
 796                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 797                         bch_info(c, "new key %s", buf.buf);
 798                 }
 799
 800                 *k = bkey_i_to_s_c(new);
 801         }
 802 err:
 803 fsck_err:
 804         printbuf_exit(&buf);
 805         return ret;
 806 }
 807
 808 /* marking of btree keys/nodes: */
 809
 810 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 811                             unsigned level, bool is_root,
 812                             struct bkey_s_c *k,
 813                             bool initial)
 814 {
 815         struct bch_fs *c = trans->c;
 816         struct bkey deleted = KEY(0, 0, 0);
 817         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 818         int ret = 0;
 819
 820         deleted.p = k->k->p;
 821
 822         if (initial) {
 823                 BUG_ON(bch2_journal_seq_verify &&
 824                        k->k->version.lo > atomic64_read(&c->journal.seq));
 825
 826                 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
 827                 if (ret)
 828                         goto err;
 829
 830                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 831                                 bkey_version_in_future,
 832                                 "key version number higher than recorded: %llu > %llu",
 833                                 k->k->version.lo,
 834                                 atomic64_read(&c->key_version)))
 835                         atomic64_set(&c->key_version, k->k->version.lo);
 836         }
 837
 838         ret = commit_do(trans, NULL, NULL, 0,
 839                         bch2_key_trigger(trans, btree_id, level, old, unsafe_bkey_s_c_to_s(*k), BTREE_TRIGGER_GC));
 840 fsck_err:
 841 err:
 842         bch_err_fn(c, ret);
 843         return ret;
 844 }
 845
 846 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 847 {
 848         struct bch_fs *c = trans->c;
 849         struct btree_node_iter iter;
 850         struct bkey unpacked;
 851         struct bkey_s_c k;
 852         struct bkey_buf prev, cur;
 853         int ret = 0;
 854
 855         if (!btree_node_type_needs_gc(btree_node_type(b)))
 856                 return 0;
 857
 858         bch2_btree_node_iter_init_from_start(&iter, b);
 859         bch2_bkey_buf_init(&prev);
 860         bch2_bkey_buf_init(&cur);
 861         bkey_init(&prev.k->k);
 862
 863         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 864                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 865                                        &k, initial);
 866                 if (ret)
 867                         break;
 868
 869                 bch2_btree_node_iter_advance(&iter, b);
 870
 871                 if (b->c.level) {
 872                         bch2_bkey_buf_reassemble(&cur, c, k);
 873
 874                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 875                                         bch2_btree_node_iter_end(&iter));
 876                         if (ret)
 877                                 break;
 878                 }
 879         }
 880
 881         bch2_bkey_buf_exit(&cur, c);
 882         bch2_bkey_buf_exit(&prev, c);
 883         return ret;
 884 }
 885
 886 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 887                          bool initial, bool metadata_only)
 888 {
 889         struct bch_fs *c = trans->c;
 890         struct btree_iter iter;
 891         struct btree *b;
 892         unsigned depth = metadata_only ? 1 : 0;
 893         int ret = 0;
 894
 895         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 896
 897         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 898                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 899                 bch2_verify_btree_nr_keys(b);
 900
 901                 gc_pos_set(c, gc_pos_btree_node(b));
 902
 903                 ret = btree_gc_mark_node(trans, b, initial);
 904                 if (ret)
 905                         break;
 906         }
 907         bch2_trans_iter_exit(trans, &iter);
 908
 909         if (ret)
 910                 return ret;
 911
 912         mutex_lock(&c->btree_root_lock);
 913         b = bch2_btree_id_root(c, btree_id)->b;
 914         if (!btree_node_fake(b)) {
 915                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 916
 917                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
 918                                        true, &k, initial);
 919         }
 920         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 921         mutex_unlock(&c->btree_root_lock);
 922
 923         return ret;
 924 }
 925
 926 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 927                                       unsigned target_depth)
 928 {
 929         struct bch_fs *c = trans->c;
 930         struct btree_and_journal_iter iter;
 931         struct bkey_s_c k;
 932         struct bkey_buf cur, prev;
 933         struct printbuf buf = PRINTBUF;
 934         int ret = 0;
 935
 936         bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
 937         bch2_bkey_buf_init(&prev);
 938         bch2_bkey_buf_init(&cur);
 939         bkey_init(&prev.k->k);
 940
 941         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 942                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 943                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 944
 945                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 946                                        false, &k, true);
 947                 if (ret)
 948                         goto fsck_err;
 949
 950                 if (b->c.level) {
 951                         bch2_bkey_buf_reassemble(&cur, c, k);
 952                         k = bkey_i_to_s_c(cur.k);
 953
 954                         bch2_btree_and_journal_iter_advance(&iter);
 955
 956                         ret = bch2_gc_check_topology(c, b,
 957                                         &prev, cur,
 958                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 959                         if (ret)
 960                                 goto fsck_err;
 961                 } else {
 962                         bch2_btree_and_journal_iter_advance(&iter);
 963                 }
 964         }
 965
 966         if (b->c.level > target_depth) {
 967                 bch2_btree_and_journal_iter_exit(&iter);
 968                 bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
 969                 iter.prefetch = true;
 970
 971                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 972                         struct btree *child;
 973
 974                         bch2_bkey_buf_reassemble(&cur, c, k);
 975                         bch2_btree_and_journal_iter_advance(&iter);
 976
 977                         child = bch2_btree_node_get_noiter(trans, cur.k,
 978                                                 b->c.btree_id, b->c.level - 1,
 979                                                 false);
 980                         ret = PTR_ERR_OR_ZERO(child);
 981
 982                         if (bch2_err_matches(ret, EIO)) {
 983                                 bch2_topology_error(c);
 984
 985                                 if (__fsck_err(c,
 986                                           FSCK_CAN_FIX|
 987                                           FSCK_CAN_IGNORE|
 988                                           FSCK_NO_RATELIMIT,
 989                                           btree_node_read_error,
 990                                           "Unreadable btree node at btree %s level %u:\n"
 991                                           "  %s",
 992                                           bch2_btree_id_str(b->c.btree_id),
 993                                           b->c.level - 1,
 994                                           (printbuf_reset(&buf),
 995                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 996                                     should_restart_for_topology_repair(c)) {
 997                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 998                                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 999                                         goto fsck_err;
1000                                 } else {
1001                                         /* Continue marking when opted to not
1002                                          * fix the error: */
1003                                         ret = 0;
1004                                         set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
1005                                         continue;
1006                                 }
1007                         } else if (ret) {
1008                                 bch_err_msg(c, ret, "getting btree node");
1009                                 break;
1010                         }
1011
1012                         ret = bch2_gc_btree_init_recurse(trans, child,
1013                                                          target_depth);
1014                         six_unlock_read(&child->c.lock);
1015
1016                         if (ret)
1017                                 break;
1018                 }
1019         }
1020 fsck_err:
1021         bch2_bkey_buf_exit(&cur, c);
1022         bch2_bkey_buf_exit(&prev, c);
1023         bch2_btree_and_journal_iter_exit(&iter);
1024         printbuf_exit(&buf);
1025         return ret;
1026 }
1027
1028 static int bch2_gc_btree_init(struct btree_trans *trans,
1029                               enum btree_id btree_id,
1030                               bool metadata_only)
1031 {
1032         struct bch_fs *c = trans->c;
1033         struct btree *b;
1034         unsigned target_depth = metadata_only ? 1 : 0;
1035         struct printbuf buf = PRINTBUF;
1036         int ret = 0;
1037
1038         b = bch2_btree_id_root(c, btree_id)->b;
1039
1040         if (btree_node_fake(b))
1041                 return 0;
1042
1043         six_lock_read(&b->c.lock, NULL, NULL);
1044         printbuf_reset(&buf);
1045         bch2_bpos_to_text(&buf, b->data->min_key);
1046         if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1047                                 btree_root_bad_min_key,
1048                         "btree root with incorrect min_key: %s", buf.buf)) {
1049                 bch_err(c, "repair unimplemented");
1050                 ret = -BCH_ERR_fsck_repair_unimplemented;
1051                 goto fsck_err;
1052         }
1053
1054         printbuf_reset(&buf);
1055         bch2_bpos_to_text(&buf, b->data->max_key);
1056         if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1057                                 btree_root_bad_max_key,
1058                         "btree root with incorrect max_key: %s", buf.buf)) {
1059                 bch_err(c, "repair unimplemented");
1060                 ret = -BCH_ERR_fsck_repair_unimplemented;
1061                 goto fsck_err;
1062         }
1063
1064         if (b->c.level >= target_depth)
1065                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1066
1067         if (!ret) {
1068                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1069
1070                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1071                                        &k, true);
1072         }
1073 fsck_err:
1074         six_unlock_read(&b->c.lock);
1075
1076         bch_err_fn(c, ret);
1077         printbuf_exit(&buf);
1078         return ret;
1079 }
1080
1081 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1082 {
1083         return  (int) btree_id_to_gc_phase(l) -
1084                 (int) btree_id_to_gc_phase(r);
1085 }
1086
1087 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1088 {
1089         struct btree_trans *trans = bch2_trans_get(c);
1090         enum btree_id ids[BTREE_ID_NR];
1091         unsigned i;
1092         int ret = 0;
1093
1094         for (i = 0; i < BTREE_ID_NR; i++)
1095                 ids[i] = i;
1096         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1097
1098         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1099                 ret = initial
1100                         ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1101                         : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1102
1103         for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1104                 if (!bch2_btree_id_root(c, i)->alive)
1105                         continue;
1106
1107                 ret = initial
1108                         ? bch2_gc_btree_init(trans, i, metadata_only)
1109                         : bch2_gc_btree(trans, i, initial, metadata_only);
1110         }
1111
1112         bch2_trans_put(trans);
1113         bch_err_fn(c, ret);
1114         return ret;
1115 }
1116
1117 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1118                                   u64 start, u64 end,
1119                                   enum bch_data_type type,
1120                                   unsigned flags)
1121 {
1122         u64 b = sector_to_bucket(ca, start);
1123
1124         do {
1125                 unsigned sectors =
1126                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1127
1128                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1129                                           gc_phase(GC_PHASE_SB), flags);
1130                 b++;
1131                 start += sectors;
1132         } while (start < end);
1133 }
1134
1135 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1136                                      unsigned flags)
1137 {
1138         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1139         unsigned i;
1140         u64 b;
1141
1142         for (i = 0; i < layout->nr_superblocks; i++) {
1143                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1144
1145                 if (offset == BCH_SB_SECTOR)
1146                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1147                                               BCH_DATA_sb, flags);
1148
1149                 mark_metadata_sectors(c, ca, offset,
1150                                       offset + (1 << layout->sb_max_size_bits),
1151                                       BCH_DATA_sb, flags);
1152         }
1153
1154         for (i = 0; i < ca->journal.nr; i++) {
1155                 b = ca->journal.buckets[i];
1156                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1157                                           ca->mi.bucket_size,
1158                                           gc_phase(GC_PHASE_SB), flags);
1159         }
1160 }
1161
1162 static void bch2_mark_superblocks(struct bch_fs *c)
1163 {
1164         mutex_lock(&c->sb_lock);
1165         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1166
1167         for_each_online_member(c, ca)
1168                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1169         mutex_unlock(&c->sb_lock);
1170 }
1171
1172 #if 0
1173 /* Also see bch2_pending_btree_node_free_insert_done() */
1174 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1175 {
1176         struct btree_update *as;
1177         struct pending_btree_node_free *d;
1178
1179         mutex_lock(&c->btree_interior_update_lock);
1180         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1181
1182         for_each_pending_btree_node_free(c, as, d)
1183                 if (d->index_update_done)
1184                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1185
1186         mutex_unlock(&c->btree_interior_update_lock);
1187 }
1188 #endif
1189
1190 static void bch2_gc_free(struct bch_fs *c)
1191 {
1192         genradix_free(&c->reflink_gc_table);
1193         genradix_free(&c->gc_stripes);
1194
1195         for_each_member_device(c, ca) {
1196                 kvfree(rcu_dereference_protected(ca->buckets_gc, 1));
1197                 ca->buckets_gc = NULL;
1198
1199                 free_percpu(ca->usage_gc);
1200                 ca->usage_gc = NULL;
1201         }
1202
1203         free_percpu(c->usage_gc);
1204         c->usage_gc = NULL;
1205 }
1206
1207 static int bch2_gc_done(struct bch_fs *c,
1208                         bool initial, bool metadata_only)
1209 {
1210         struct bch_dev *ca = NULL;
1211         struct printbuf buf = PRINTBUF;
1212         bool verify = !metadata_only &&
1213                 !c->opts.reconstruct_alloc &&
1214                 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1215         unsigned i;
1216         int ret = 0;
1217
1218         percpu_down_write(&c->mark_lock);
1219
1220 #define copy_field(_err, _f, _msg, ...)                                 \
1221         if (dst->_f != src->_f &&                                       \
1222             (!verify ||                                                 \
1223              fsck_err(c, _err, _msg ": got %llu, should be %llu"        \
1224                       , ##__VA_ARGS__, dst->_f, src->_f)))              \
1225                 dst->_f = src->_f
1226 #define copy_dev_field(_err, _f, _msg, ...)                             \
1227         copy_field(_err, _f, "dev %u has wrong " _msg, ca->dev_idx, ##__VA_ARGS__)
1228 #define copy_fs_field(_err, _f, _msg, ...)                              \
1229         copy_field(_err, _f, "fs has wrong " _msg, ##__VA_ARGS__)
1230
1231         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1232                 bch2_fs_usage_acc_to_base(c, i);
1233
1234         __for_each_member_device(c, ca) {
1235                 struct bch_dev_usage *dst = ca->usage_base;
1236                 struct bch_dev_usage *src = (void *)
1237                         bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1238                                              dev_usage_u64s());
1239
1240                 for (i = 0; i < BCH_DATA_NR; i++) {
1241                         copy_dev_field(dev_usage_buckets_wrong,
1242                                        d[i].buckets,    "%s buckets", bch2_data_type_str(i));
1243                         copy_dev_field(dev_usage_sectors_wrong,
1244                                        d[i].sectors,    "%s sectors", bch2_data_type_str(i));
1245                         copy_dev_field(dev_usage_fragmented_wrong,
1246                                        d[i].fragmented, "%s fragmented", bch2_data_type_str(i));
1247                 }
1248         }
1249
1250         {
1251                 unsigned nr = fs_usage_u64s(c);
1252                 struct bch_fs_usage *dst = c->usage_base;
1253                 struct bch_fs_usage *src = (void *)
1254                         bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1255
1256                 copy_fs_field(fs_usage_hidden_wrong,
1257                               b.hidden,         "hidden");
1258                 copy_fs_field(fs_usage_btree_wrong,
1259                               b.btree,          "btree");
1260
1261                 if (!metadata_only) {
1262                         copy_fs_field(fs_usage_data_wrong,
1263                                       b.data,   "data");
1264                         copy_fs_field(fs_usage_cached_wrong,
1265                                       b.cached, "cached");
1266                         copy_fs_field(fs_usage_reserved_wrong,
1267                                       b.reserved,       "reserved");
1268                         copy_fs_field(fs_usage_nr_inodes_wrong,
1269                                       b.nr_inodes,"nr_inodes");
1270
1271                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1272                                 copy_fs_field(fs_usage_persistent_reserved_wrong,
1273                                               persistent_reserved[i],
1274                                               "persistent_reserved[%i]", i);
1275                 }
1276
1277                 for (i = 0; i < c->replicas.nr; i++) {
1278                         struct bch_replicas_entry_v1 *e =
1279                                 cpu_replicas_entry(&c->replicas, i);
1280
1281                         if (metadata_only &&
1282                             (e->data_type == BCH_DATA_user ||
1283                              e->data_type == BCH_DATA_cached))
1284                                 continue;
1285
1286                         printbuf_reset(&buf);
1287                         bch2_replicas_entry_to_text(&buf, e);
1288
1289                         copy_fs_field(fs_usage_replicas_wrong,
1290                                       replicas[i], "%s", buf.buf);
1291                 }
1292         }
1293
1294 #undef copy_fs_field
1295 #undef copy_dev_field
1296 #undef copy_stripe_field
1297 #undef copy_field
1298 fsck_err:
1299         if (ca)
1300                 percpu_ref_put(&ca->ref);
1301         bch_err_fn(c, ret);
1302
1303         percpu_up_write(&c->mark_lock);
1304         printbuf_exit(&buf);
1305         return ret;
1306 }
1307
1308 static int bch2_gc_start(struct bch_fs *c)
1309 {
1310         BUG_ON(c->usage_gc);
1311
1312         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1313                                          sizeof(u64), GFP_KERNEL);
1314         if (!c->usage_gc) {
1315                 bch_err(c, "error allocating c->usage_gc");
1316                 return -BCH_ERR_ENOMEM_gc_start;
1317         }
1318
1319         for_each_member_device(c, ca) {
1320                 BUG_ON(ca->usage_gc);
1321
1322                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1323                 if (!ca->usage_gc) {
1324                         bch_err(c, "error allocating ca->usage_gc");
1325                         percpu_ref_put(&ca->ref);
1326                         return -BCH_ERR_ENOMEM_gc_start;
1327                 }
1328
1329                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1330                                ca->mi.nbuckets - ca->mi.first_bucket);
1331         }
1332
1333         return 0;
1334 }
1335
1336 static int bch2_gc_reset(struct bch_fs *c)
1337 {
1338         for_each_member_device(c, ca) {
1339                 free_percpu(ca->usage_gc);
1340                 ca->usage_gc = NULL;
1341         }
1342
1343         free_percpu(c->usage_gc);
1344         c->usage_gc = NULL;
1345
1346         return bch2_gc_start(c);
1347 }
1348
1349 /* returns true if not equal */
1350 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1351                                      struct bch_alloc_v4 r)
1352 {
1353         return  l.gen != r.gen                          ||
1354                 l.oldest_gen != r.oldest_gen            ||
1355                 l.data_type != r.data_type              ||
1356                 l.dirty_sectors != r.dirty_sectors      ||
1357                 l.cached_sectors != r.cached_sectors     ||
1358                 l.stripe_redundancy != r.stripe_redundancy ||
1359                 l.stripe != r.stripe;
1360 }
1361
1362 static int bch2_alloc_write_key(struct btree_trans *trans,
1363                                 struct btree_iter *iter,
1364                                 struct bkey_s_c k,
1365                                 bool metadata_only)
1366 {
1367         struct bch_fs *c = trans->c;
1368         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1369         struct bucket gc, *b;
1370         struct bkey_i_alloc_v4 *a;
1371         struct bch_alloc_v4 old_convert, new;
1372         const struct bch_alloc_v4 *old;
1373         enum bch_data_type type;
1374         int ret;
1375
1376         old = bch2_alloc_to_v4(k, &old_convert);
1377         new = *old;
1378
1379         percpu_down_read(&c->mark_lock);
1380         b = gc_bucket(ca, iter->pos.offset);
1381
1382         /*
1383          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1384          * fix that here:
1385          */
1386         type = __alloc_data_type(b->dirty_sectors,
1387                                  b->cached_sectors,
1388                                  b->stripe,
1389                                  *old,
1390                                  b->data_type);
1391         if (b->data_type != type) {
1392                 struct bch_dev_usage *u;
1393
1394                 preempt_disable();
1395                 u = this_cpu_ptr(ca->usage_gc);
1396                 u->d[b->data_type].buckets--;
1397                 b->data_type = type;
1398                 u->d[b->data_type].buckets++;
1399                 preempt_enable();
1400         }
1401
1402         gc = *b;
1403         percpu_up_read(&c->mark_lock);
1404
1405         if (metadata_only &&
1406             gc.data_type != BCH_DATA_sb &&
1407             gc.data_type != BCH_DATA_journal &&
1408             gc.data_type != BCH_DATA_btree)
1409                 return 0;
1410
1411         if (gen_after(old->gen, gc.gen))
1412                 return 0;
1413
1414         if (c->opts.reconstruct_alloc ||
1415             fsck_err_on(new.data_type != gc.data_type, c,
1416                         alloc_key_data_type_wrong,
1417                         "bucket %llu:%llu gen %u has wrong data_type"
1418                         ": got %s, should be %s",
1419                         iter->pos.inode, iter->pos.offset,
1420                         gc.gen,
1421                         bch2_data_type_str(new.data_type),
1422                         bch2_data_type_str(gc.data_type)))
1423                 new.data_type = gc.data_type;
1424
1425 #define copy_bucket_field(_errtype, _f)                                 \
1426         if (c->opts.reconstruct_alloc ||                                \
1427             fsck_err_on(new._f != gc._f, c, _errtype,                   \
1428                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1429                         ": got %u, should be %u",                       \
1430                         iter->pos.inode, iter->pos.offset,              \
1431                         gc.gen,                                         \
1432                         bch2_data_type_str(gc.data_type),               \
1433                         new._f, gc._f))                                 \
1434                 new._f = gc._f;                                         \
1435
1436         copy_bucket_field(alloc_key_gen_wrong,
1437                           gen);
1438         copy_bucket_field(alloc_key_dirty_sectors_wrong,
1439                           dirty_sectors);
1440         copy_bucket_field(alloc_key_cached_sectors_wrong,
1441                           cached_sectors);
1442         copy_bucket_field(alloc_key_stripe_wrong,
1443                           stripe);
1444         copy_bucket_field(alloc_key_stripe_redundancy_wrong,
1445                           stripe_redundancy);
1446 #undef copy_bucket_field
1447
1448         if (!bch2_alloc_v4_cmp(*old, new))
1449                 return 0;
1450
1451         a = bch2_alloc_to_v4_mut(trans, k);
1452         ret = PTR_ERR_OR_ZERO(a);
1453         if (ret)
1454                 return ret;
1455
1456         a->v = new;
1457
1458         /*
1459          * The trigger normally makes sure this is set, but we're not running
1460          * triggers:
1461          */
1462         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1463                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1464
1465         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1466 fsck_err:
1467         return ret;
1468 }
1469
1470 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1471 {
1472         int ret = 0;
1473
1474         for_each_member_device(c, ca) {
1475                 ret = bch2_trans_run(c,
1476                         for_each_btree_key_upto_commit(trans, iter, BTREE_ID_alloc,
1477                                         POS(ca->dev_idx, ca->mi.first_bucket),
1478                                         POS(ca->dev_idx, ca->mi.nbuckets - 1),
1479                                         BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1480                                         NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
1481                                 bch2_alloc_write_key(trans, &iter, k, metadata_only)));
1482                 if (ret) {
1483                         percpu_ref_put(&ca->ref);
1484                         break;
1485                 }
1486         }
1487
1488         bch_err_fn(c, ret);
1489         return ret;
1490 }
1491
1492 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1493 {
1494         for_each_member_device(c, ca) {
1495                 struct bucket_array *buckets = kvmalloc(sizeof(struct bucket_array) +
1496                                 ca->mi.nbuckets * sizeof(struct bucket),
1497                                 GFP_KERNEL|__GFP_ZERO);
1498                 if (!buckets) {
1499                         percpu_ref_put(&ca->ref);
1500                         bch_err(c, "error allocating ca->buckets[gc]");
1501                         return -BCH_ERR_ENOMEM_gc_alloc_start;
1502                 }
1503
1504                 buckets->first_bucket   = ca->mi.first_bucket;
1505                 buckets->nbuckets       = ca->mi.nbuckets;
1506                 rcu_assign_pointer(ca->buckets_gc, buckets);
1507         }
1508
1509         int ret = bch2_trans_run(c,
1510                 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
1511                                          BTREE_ITER_PREFETCH, k, ({
1512                         struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
1513                         struct bucket *g = gc_bucket(ca, k.k->p.offset);
1514
1515                         struct bch_alloc_v4 a_convert;
1516                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1517
1518                         g->gen_valid    = 1;
1519                         g->gen          = a->gen;
1520
1521                         if (metadata_only &&
1522                             (a->data_type == BCH_DATA_user ||
1523                              a->data_type == BCH_DATA_cached ||
1524                              a->data_type == BCH_DATA_parity)) {
1525                                 g->data_type            = a->data_type;
1526                                 g->dirty_sectors        = a->dirty_sectors;
1527                                 g->cached_sectors       = a->cached_sectors;
1528                                 g->stripe               = a->stripe;
1529                                 g->stripe_redundancy    = a->stripe_redundancy;
1530                         }
1531
1532                         0;
1533                 })));
1534         bch_err_fn(c, ret);
1535         return ret;
1536 }
1537
1538 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1539 {
1540         for_each_member_device(c, ca) {
1541                 struct bucket_array *buckets = gc_bucket_array(ca);
1542                 struct bucket *g;
1543
1544                 for_each_bucket(g, buckets) {
1545                         if (metadata_only &&
1546                             (g->data_type == BCH_DATA_user ||
1547                              g->data_type == BCH_DATA_cached ||
1548                              g->data_type == BCH_DATA_parity))
1549                                 continue;
1550                         g->data_type = 0;
1551                         g->dirty_sectors = 0;
1552                         g->cached_sectors = 0;
1553                 }
1554         }
1555 }
1556
1557 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1558                                      struct btree_iter *iter,
1559                                      struct bkey_s_c k,
1560                                      size_t *idx)
1561 {
1562         struct bch_fs *c = trans->c;
1563         const __le64 *refcount = bkey_refcount_c(k);
1564         struct printbuf buf = PRINTBUF;
1565         struct reflink_gc *r;
1566         int ret = 0;
1567
1568         if (!refcount)
1569                 return 0;
1570
1571         while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1572                r->offset < k.k->p.offset)
1573                 ++*idx;
1574
1575         if (!r ||
1576             r->offset != k.k->p.offset ||
1577             r->size != k.k->size) {
1578                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1579                 return -EINVAL;
1580         }
1581
1582         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1583                         reflink_v_refcount_wrong,
1584                         "reflink key has wrong refcount:\n"
1585                         "  %s\n"
1586                         "  should be %u",
1587                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1588                         r->refcount)) {
1589                 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1590
1591                 ret = PTR_ERR_OR_ZERO(new);
1592                 if (ret)
1593                         return ret;
1594
1595                 if (!r->refcount)
1596                         new->k.type = KEY_TYPE_deleted;
1597                 else
1598                         *bkey_refcount(bkey_i_to_s(new)) = cpu_to_le64(r->refcount);
1599         }
1600 fsck_err:
1601         printbuf_exit(&buf);
1602         return ret;
1603 }
1604
1605 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1606 {
1607         size_t idx = 0;
1608
1609         if (metadata_only)
1610                 return 0;
1611
1612         int ret = bch2_trans_run(c,
1613                 for_each_btree_key_commit(trans, iter,
1614                                 BTREE_ID_reflink, POS_MIN,
1615                                 BTREE_ITER_PREFETCH, k,
1616                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1617                         bch2_gc_write_reflink_key(trans, &iter, k, &idx)));
1618         c->reflink_gc_nr = 0;
1619         return ret;
1620 }
1621
1622 static int bch2_gc_reflink_start(struct bch_fs *c,
1623                                  bool metadata_only)
1624 {
1625
1626         if (metadata_only)
1627                 return 0;
1628
1629         c->reflink_gc_nr = 0;
1630
1631         int ret = bch2_trans_run(c,
1632                 for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1633                                    BTREE_ITER_PREFETCH, k, ({
1634                         const __le64 *refcount = bkey_refcount_c(k);
1635
1636                         if (!refcount)
1637                                 continue;
1638
1639                         struct reflink_gc *r = genradix_ptr_alloc(&c->reflink_gc_table,
1640                                                         c->reflink_gc_nr++, GFP_KERNEL);
1641                         if (!r) {
1642                                 ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1643                                 break;
1644                         }
1645
1646                         r->offset       = k.k->p.offset;
1647                         r->size         = k.k->size;
1648                         r->refcount     = 0;
1649                         0;
1650                 })));
1651
1652         bch_err_fn(c, ret);
1653         return ret;
1654 }
1655
1656 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1657 {
1658         struct genradix_iter iter;
1659         struct reflink_gc *r;
1660
1661         genradix_for_each(&c->reflink_gc_table, iter, r)
1662                 r->refcount = 0;
1663 }
1664
1665 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1666                                      struct btree_iter *iter,
1667                                      struct bkey_s_c k)
1668 {
1669         struct bch_fs *c = trans->c;
1670         struct printbuf buf = PRINTBUF;
1671         const struct bch_stripe *s;
1672         struct gc_stripe *m;
1673         bool bad = false;
1674         unsigned i;
1675         int ret = 0;
1676
1677         if (k.k->type != KEY_TYPE_stripe)
1678                 return 0;
1679
1680         s = bkey_s_c_to_stripe(k).v;
1681         m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1682
1683         for (i = 0; i < s->nr_blocks; i++) {
1684                 u32 old = stripe_blockcount_get(s, i);
1685                 u32 new = (m ? m->block_sectors[i] : 0);
1686
1687                 if (old != new) {
1688                         prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1689                                    i, old, new);
1690                         bad = true;
1691                 }
1692         }
1693
1694         if (bad)
1695                 bch2_bkey_val_to_text(&buf, c, k);
1696
1697         if (fsck_err_on(bad, c, stripe_sector_count_wrong,
1698                         "%s", buf.buf)) {
1699                 struct bkey_i_stripe *new;
1700
1701                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1702                 ret = PTR_ERR_OR_ZERO(new);
1703                 if (ret)
1704                         return ret;
1705
1706                 bkey_reassemble(&new->k_i, k);
1707
1708                 for (i = 0; i < new->v.nr_blocks; i++)
1709                         stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1710
1711                 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1712         }
1713 fsck_err:
1714         printbuf_exit(&buf);
1715         return ret;
1716 }
1717
1718 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1719 {
1720         if (metadata_only)
1721                 return 0;
1722
1723         return bch2_trans_run(c,
1724                 for_each_btree_key_commit(trans, iter,
1725                                 BTREE_ID_stripes, POS_MIN,
1726                                 BTREE_ITER_PREFETCH, k,
1727                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1728                         bch2_gc_write_stripes_key(trans, &iter, k)));
1729 }
1730
1731 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1732 {
1733         genradix_free(&c->gc_stripes);
1734 }
1735
1736 /**
1737  * bch2_gc - walk _all_ references to buckets, and recompute them:
1738  *
1739  * @c:                  filesystem object
1740  * @initial:            are we in recovery?
1741  * @metadata_only:      are we just checking metadata references, or everything?
1742  *
1743  * Returns: 0 on success, or standard errcode on failure
1744  *
1745  * Order matters here:
1746  *  - Concurrent GC relies on the fact that we have a total ordering for
1747  *    everything that GC walks - see  gc_will_visit_node(),
1748  *    gc_will_visit_root()
1749  *
1750  *  - also, references move around in the course of index updates and
1751  *    various other crap: everything needs to agree on the ordering
1752  *    references are allowed to move around in - e.g., we're allowed to
1753  *    start with a reference owned by an open_bucket (the allocator) and
1754  *    move it to the btree, but not the reverse.
1755  *
1756  *    This is necessary to ensure that gc doesn't miss references that
1757  *    move around - if references move backwards in the ordering GC
1758  *    uses, GC could skip past them
1759  */
1760 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1761 {
1762         unsigned iter = 0;
1763         int ret;
1764
1765         lockdep_assert_held(&c->state_lock);
1766
1767         down_write(&c->gc_lock);
1768
1769         bch2_btree_interior_updates_flush(c);
1770
1771         ret   = bch2_gc_start(c) ?:
1772                 bch2_gc_alloc_start(c, metadata_only) ?:
1773                 bch2_gc_reflink_start(c, metadata_only);
1774         if (ret)
1775                 goto out;
1776 again:
1777         gc_pos_set(c, gc_phase(GC_PHASE_START));
1778
1779         bch2_mark_superblocks(c);
1780
1781         ret = bch2_gc_btrees(c, initial, metadata_only);
1782
1783         if (ret)
1784                 goto out;
1785
1786 #if 0
1787         bch2_mark_pending_btree_node_frees(c);
1788 #endif
1789         c->gc_count++;
1790
1791         if (test_bit(BCH_FS_need_another_gc, &c->flags) ||
1792             (!iter && bch2_test_restart_gc)) {
1793                 if (iter++ > 2) {
1794                         bch_info(c, "Unable to fix bucket gens, looping");
1795                         ret = -EINVAL;
1796                         goto out;
1797                 }
1798
1799                 /*
1800                  * XXX: make sure gens we fixed got saved
1801                  */
1802                 bch_info(c, "Second GC pass needed, restarting:");
1803                 clear_bit(BCH_FS_need_another_gc, &c->flags);
1804                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1805
1806                 bch2_gc_stripes_reset(c, metadata_only);
1807                 bch2_gc_alloc_reset(c, metadata_only);
1808                 bch2_gc_reflink_reset(c, metadata_only);
1809                 ret = bch2_gc_reset(c);
1810                 if (ret)
1811                         goto out;
1812
1813                 /* flush fsck errors, reset counters */
1814                 bch2_flush_fsck_errs(c);
1815                 goto again;
1816         }
1817 out:
1818         if (!ret) {
1819                 bch2_journal_block(&c->journal);
1820
1821                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1822                         bch2_gc_reflink_done(c, metadata_only) ?:
1823                         bch2_gc_alloc_done(c, metadata_only) ?:
1824                         bch2_gc_done(c, initial, metadata_only);
1825
1826                 bch2_journal_unblock(&c->journal);
1827         }
1828
1829         percpu_down_write(&c->mark_lock);
1830         /* Indicates that gc is no longer in progress: */
1831         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1832
1833         bch2_gc_free(c);
1834         percpu_up_write(&c->mark_lock);
1835
1836         up_write(&c->gc_lock);
1837
1838         /*
1839          * At startup, allocations can happen directly instead of via the
1840          * allocator thread - issue wakeup in case they blocked on gc_lock:
1841          */
1842         closure_wake_up(&c->freelist_wait);
1843         bch_err_fn(c, ret);
1844         return ret;
1845 }
1846
1847 static int gc_btree_gens_key(struct btree_trans *trans,
1848                              struct btree_iter *iter,
1849                              struct bkey_s_c k)
1850 {
1851         struct bch_fs *c = trans->c;
1852         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1853         struct bkey_i *u;
1854         int ret;
1855
1856         percpu_down_read(&c->mark_lock);
1857         bkey_for_each_ptr(ptrs, ptr) {
1858                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1859
1860                 if (ptr_stale(ca, ptr) > 16) {
1861                         percpu_up_read(&c->mark_lock);
1862                         goto update;
1863                 }
1864         }
1865
1866         bkey_for_each_ptr(ptrs, ptr) {
1867                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1868                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1869
1870                 if (gen_after(*gen, ptr->gen))
1871                         *gen = ptr->gen;
1872         }
1873         percpu_up_read(&c->mark_lock);
1874         return 0;
1875 update:
1876         u = bch2_bkey_make_mut(trans, iter, &k, 0);
1877         ret = PTR_ERR_OR_ZERO(u);
1878         if (ret)
1879                 return ret;
1880
1881         bch2_extent_normalize(c, bkey_i_to_s(u));
1882         return 0;
1883 }
1884
1885 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1886                                        struct bkey_s_c k)
1887 {
1888         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1889         struct bch_alloc_v4 a_convert;
1890         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1891         struct bkey_i_alloc_v4 *a_mut;
1892         int ret;
1893
1894         if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1895                 return 0;
1896
1897         a_mut = bch2_alloc_to_v4_mut(trans, k);
1898         ret = PTR_ERR_OR_ZERO(a_mut);
1899         if (ret)
1900                 return ret;
1901
1902         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1903         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1904
1905         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1906 }
1907
1908 int bch2_gc_gens(struct bch_fs *c)
1909 {
1910         u64 b, start_time = local_clock();
1911         int ret;
1912
1913         /*
1914          * Ideally we would be using state_lock and not gc_lock here, but that
1915          * introduces a deadlock in the RO path - we currently take the state
1916          * lock at the start of going RO, thus the gc thread may get stuck:
1917          */
1918         if (!mutex_trylock(&c->gc_gens_lock))
1919                 return 0;
1920
1921         trace_and_count(c, gc_gens_start, c);
1922         down_read(&c->gc_lock);
1923
1924         for_each_member_device(c, ca) {
1925                 struct bucket_gens *gens = bucket_gens(ca);
1926
1927                 BUG_ON(ca->oldest_gen);
1928
1929                 ca->oldest_gen = kvmalloc(gens->nbuckets, GFP_KERNEL);
1930                 if (!ca->oldest_gen) {
1931                         percpu_ref_put(&ca->ref);
1932                         ret = -BCH_ERR_ENOMEM_gc_gens;
1933                         goto err;
1934                 }
1935
1936                 for (b = gens->first_bucket;
1937                      b < gens->nbuckets; b++)
1938                         ca->oldest_gen[b] = gens->b[b];
1939         }
1940
1941         for (unsigned i = 0; i < BTREE_ID_NR; i++)
1942                 if (btree_type_has_ptrs(i)) {
1943                         c->gc_gens_btree = i;
1944                         c->gc_gens_pos = POS_MIN;
1945
1946                         ret = bch2_trans_run(c,
1947                                 for_each_btree_key_commit(trans, iter, i,
1948                                                 POS_MIN,
1949                                                 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1950                                                 k,
1951                                                 NULL, NULL,
1952                                                 BCH_TRANS_COMMIT_no_enospc,
1953                                         gc_btree_gens_key(trans, &iter, k)));
1954                         if (ret)
1955                                 goto err;
1956                 }
1957
1958         ret = bch2_trans_run(c,
1959                 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1960                                 POS_MIN,
1961                                 BTREE_ITER_PREFETCH,
1962                                 k,
1963                                 NULL, NULL,
1964                                 BCH_TRANS_COMMIT_no_enospc,
1965                         bch2_alloc_write_oldest_gen(trans, &iter, k)));
1966         if (ret)
1967                 goto err;
1968
1969         c->gc_gens_btree        = 0;
1970         c->gc_gens_pos          = POS_MIN;
1971
1972         c->gc_count++;
1973
1974         time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1975         trace_and_count(c, gc_gens_end, c);
1976 err:
1977         for_each_member_device(c, ca) {
1978                 kvfree(ca->oldest_gen);
1979                 ca->oldest_gen = NULL;
1980         }
1981
1982         up_read(&c->gc_lock);
1983         mutex_unlock(&c->gc_gens_lock);
1984         if (!bch2_err_matches(ret, EROFS))
1985                 bch_err_fn(c, ret);
1986         return ret;
1987 }
1988
1989 static int bch2_gc_thread(void *arg)
1990 {
1991         struct bch_fs *c = arg;
1992         struct io_clock *clock = &c->io_clock[WRITE];
1993         unsigned long last = atomic64_read(&clock->now);
1994         unsigned last_kick = atomic_read(&c->kick_gc);
1995
1996         set_freezable();
1997
1998         while (1) {
1999                 while (1) {
2000                         set_current_state(TASK_INTERRUPTIBLE);
2001
2002                         if (kthread_should_stop()) {
2003                                 __set_current_state(TASK_RUNNING);
2004                                 return 0;
2005                         }
2006
2007                         if (atomic_read(&c->kick_gc) != last_kick)
2008                                 break;
2009
2010                         if (c->btree_gc_periodic) {
2011                                 unsigned long next = last + c->capacity / 16;
2012
2013                                 if (atomic64_read(&clock->now) >= next)
2014                                         break;
2015
2016                                 bch2_io_clock_schedule_timeout(clock, next);
2017                         } else {
2018                                 schedule();
2019                         }
2020
2021                         try_to_freeze();
2022                 }
2023                 __set_current_state(TASK_RUNNING);
2024
2025                 last = atomic64_read(&clock->now);
2026                 last_kick = atomic_read(&c->kick_gc);
2027
2028                 /*
2029                  * Full gc is currently incompatible with btree key cache:
2030                  */
2031 #if 0
2032                 ret = bch2_gc(c, false, false);
2033 #else
2034                 bch2_gc_gens(c);
2035 #endif
2036                 debug_check_no_locks_held();
2037         }
2038
2039         return 0;
2040 }
2041
2042 void bch2_gc_thread_stop(struct bch_fs *c)
2043 {
2044         struct task_struct *p;
2045
2046         p = c->gc_thread;
2047         c->gc_thread = NULL;
2048
2049         if (p) {
2050                 kthread_stop(p);
2051                 put_task_struct(p);
2052         }
2053 }
2054
2055 int bch2_gc_thread_start(struct bch_fs *c)
2056 {
2057         struct task_struct *p;
2058
2059         if (c->gc_thread)
2060                 return 0;
2061
2062         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2063         if (IS_ERR(p)) {
2064                 bch_err_fn(c, PTR_ERR(p));
2065                 return PTR_ERR(p);
2066         }
2067
2068         get_task_struct(p);
2069         c->gc_thread = p;
2070         wake_up_process(p);
2071         return 0;
2072 }