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
  31 #include <linux/slab.h>
  32 #include <linux/bitops.h>
  33 #include <linux/freezer.h>
  34 #include <linux/kthread.h>
  35 #include <linux/preempt.h>
  36 #include <linux/rcupdate.h>
  37 #include <linux/sched/task.h>
  38 #include <trace/events/bcachefs.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(c, "%s: error getting btree node: %s",
 408                                 __func__, bch2_err_str(ret));
 409                         break;
 410                 }
 411
 412                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 413
 414                 if (ret == DROP_THIS_NODE) {
 415                         six_unlock_read(&cur->c.lock);
 416                         bch2_btree_node_evict(trans, cur_k.k);
 417                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 418                                                       b->c.level, cur_k.k->k.p);
 419                         cur = NULL;
 420                         if (ret)
 421                                 break;
 422                         continue;
 423                 }
 424
 425                 if (prev)
 426                         six_unlock_read(&prev->c.lock);
 427                 prev = NULL;
 428
 429                 if (ret == DROP_PREV_NODE) {
 430                         bch2_btree_node_evict(trans, prev_k.k);
 431                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 432                                                       b->c.level, prev_k.k->k.p);
 433                         if (ret)
 434                                 break;
 435
 436                         bch2_btree_and_journal_iter_exit(&iter);
 437                         bch2_bkey_buf_exit(&prev_k, c);
 438                         bch2_bkey_buf_exit(&cur_k, c);
 439                         goto again;
 440                 } else if (ret)
 441                         break;
 442
 443                 prev = cur;
 444                 cur = NULL;
 445                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 446         }
 447
 448         if (!ret && !IS_ERR_OR_NULL(prev)) {
 449                 BUG_ON(cur);
 450                 ret = btree_repair_node_end(c, b, prev);
 451         }
 452
 453         if (!IS_ERR_OR_NULL(prev))
 454                 six_unlock_read(&prev->c.lock);
 455         prev = NULL;
 456         if (!IS_ERR_OR_NULL(cur))
 457                 six_unlock_read(&cur->c.lock);
 458         cur = NULL;
 459
 460         if (ret)
 461                 goto err;
 462
 463         bch2_btree_and_journal_iter_exit(&iter);
 464         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 465
 466         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 467                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 468                 bch2_btree_and_journal_iter_advance(&iter);
 469
 470                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 471                                         b->c.btree_id, b->c.level - 1,
 472                                         false);
 473                 ret = PTR_ERR_OR_ZERO(cur);
 474
 475                 if (ret) {
 476                         bch_err(c, "%s: error getting btree node: %s",
 477                                 __func__, bch2_err_str(ret));
 478                         goto err;
 479                 }
 480
 481                 ret = bch2_btree_repair_topology_recurse(trans, cur);
 482                 six_unlock_read(&cur->c.lock);
 483                 cur = NULL;
 484
 485                 if (ret == DROP_THIS_NODE) {
 486                         bch2_btree_node_evict(trans, cur_k.k);
 487                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 488                                                       b->c.level, cur_k.k->k.p);
 489                         dropped_children = true;
 490                 }
 491
 492                 if (ret)
 493                         goto err;
 494
 495                 have_child = true;
 496         }
 497
 498         printbuf_reset(&buf);
 499         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 500
 501         if (mustfix_fsck_err_on(!have_child, c,
 502                         "empty interior btree node at btree %s level %u\n"
 503                         "  %s",
 504                         bch2_btree_ids[b->c.btree_id],
 505                         b->c.level, buf.buf))
 506                 ret = DROP_THIS_NODE;
 507 err:
 508 fsck_err:
 509         if (!IS_ERR_OR_NULL(prev))
 510                 six_unlock_read(&prev->c.lock);
 511         if (!IS_ERR_OR_NULL(cur))
 512                 six_unlock_read(&cur->c.lock);
 513
 514         bch2_btree_and_journal_iter_exit(&iter);
 515         bch2_bkey_buf_exit(&prev_k, c);
 516         bch2_bkey_buf_exit(&cur_k, c);
 517
 518         if (!ret && dropped_children)
 519                 goto again;
 520
 521         printbuf_exit(&buf);
 522         return ret;
 523 }
 524
 525 static int bch2_repair_topology(struct bch_fs *c)
 526 {
 527         struct btree_trans trans;
 528         struct btree *b;
 529         unsigned i;
 530         int ret = 0;
 531
 532         bch2_trans_init(&trans, c, 0, 0);
 533
 534         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 535                 b = c->btree_roots[i].b;
 536                 if (btree_node_fake(b))
 537                         continue;
 538
 539                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 540                 ret = bch2_btree_repair_topology_recurse(&trans, b);
 541                 six_unlock_read(&b->c.lock);
 542
 543                 if (ret == DROP_THIS_NODE) {
 544                         bch_err(c, "empty btree root - repair unimplemented");
 545                         ret = -BCH_ERR_fsck_repair_unimplemented;
 546                 }
 547         }
 548
 549         bch2_trans_exit(&trans);
 550
 551         return ret;
 552 }
 553
 554 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
 555                                unsigned level, bool is_root,
 556                                struct bkey_s_c *k)
 557 {
 558         struct bch_fs *c = trans->c;
 559         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 560         const union bch_extent_entry *entry;
 561         struct extent_ptr_decoded p = { 0 };
 562         bool do_update = false;
 563         struct printbuf buf = PRINTBUF;
 564         int ret = 0;
 565
 566         /*
 567          * XXX
 568          * use check_bucket_ref here
 569          */
 570         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 571                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 572                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 573                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 574
 575                 if (!g->gen_valid &&
 576                     (c->opts.reconstruct_alloc ||
 577                      fsck_err(c, "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 578                               "while marking %s",
 579                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 580                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 581                               p.ptr.gen,
 582                               (printbuf_reset(&buf),
 583                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 584                         if (!p.ptr.cached) {
 585                                 g->gen_valid            = true;
 586                                 g->gen                  = p.ptr.gen;
 587                         } else {
 588                                 do_update = true;
 589                         }
 590                 }
 591
 592                 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
 593                     (c->opts.reconstruct_alloc ||
 594                      fsck_err(c, "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 595                               "while marking %s",
 596                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 597                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 598                               p.ptr.gen, g->gen,
 599                               (printbuf_reset(&buf),
 600                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 601                         if (!p.ptr.cached) {
 602                                 g->gen_valid            = true;
 603                                 g->gen                  = p.ptr.gen;
 604                                 g->data_type            = 0;
 605                                 g->dirty_sectors        = 0;
 606                                 g->cached_sectors       = 0;
 607                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 608                         } else {
 609                                 do_update = true;
 610                         }
 611                 }
 612
 613                 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
 614                     (c->opts.reconstruct_alloc ||
 615                      fsck_err(c, "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 616                               "while marking %s",
 617                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 618                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 619                               p.ptr.gen,
 620                               (printbuf_reset(&buf),
 621                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 622                         do_update = true;
 623
 624                 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
 625                     (c->opts.reconstruct_alloc ||
 626                      fsck_err(c, "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 627                               "while marking %s",
 628                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 629                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 630                               p.ptr.gen, g->gen,
 631                               (printbuf_reset(&buf),
 632                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 633                         do_update = true;
 634
 635                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 636                         continue;
 637
 638                 if (fsck_err_on(bucket_data_type(g->data_type) &&
 639                                 bucket_data_type(g->data_type) != data_type, c,
 640                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 641                                 "while marking %s",
 642                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 643                                 bch2_data_types[g->data_type],
 644                                 bch2_data_types[data_type],
 645                                 (printbuf_reset(&buf),
 646                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 647                         if (data_type == BCH_DATA_btree) {
 648                                 g->data_type    = data_type;
 649                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 650                         } else {
 651                                 do_update = true;
 652                         }
 653                 }
 654
 655                 if (p.has_ec) {
 656                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 657
 658                         if (fsck_err_on(!m || !m->alive, c,
 659                                         "pointer to nonexistent stripe %llu\n"
 660                                         "while marking %s",
 661                                         (u64) p.ec.idx,
 662                                         (printbuf_reset(&buf),
 663                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 664                                 do_update = true;
 665
 666                         if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
 667                                         "pointer does not match stripe %llu\n"
 668                                         "while marking %s",
 669                                         (u64) p.ec.idx,
 670                                         (printbuf_reset(&buf),
 671                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 672                                 do_update = true;
 673                 }
 674         }
 675
 676         if (do_update) {
 677                 struct bkey_ptrs ptrs;
 678                 union bch_extent_entry *entry;
 679                 struct bch_extent_ptr *ptr;
 680                 struct bkey_i *new;
 681
 682                 if (is_root) {
 683                         bch_err(c, "cannot update btree roots yet");
 684                         ret = -EINVAL;
 685                         goto err;
 686                 }
 687
 688                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 689                 if (!new) {
 690                         bch_err(c, "%s: error allocating new key", __func__);
 691                         ret = -BCH_ERR_ENOMEM_gc_repair_key;
 692                         goto err;
 693                 }
 694
 695                 bkey_reassemble(new, *k);
 696
 697                 if (level) {
 698                         /*
 699                          * We don't want to drop btree node pointers - if the
 700                          * btree node isn't there anymore, the read path will
 701                          * sort it out:
 702                          */
 703                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 704                         bkey_for_each_ptr(ptrs, ptr) {
 705                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 706                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 707
 708                                 ptr->gen = g->gen;
 709                         }
 710                 } else {
 711                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 712                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 713                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 714                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 715
 716                                 (ptr->cached &&
 717                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 718                                 (!ptr->cached &&
 719                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 720                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 721                                 (g->data_type &&
 722                                  g->data_type != data_type);
 723                         }));
 724 again:
 725                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 726                         bkey_extent_entry_for_each(ptrs, entry) {
 727                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 728                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 729                                                                         entry->stripe_ptr.idx);
 730                                         union bch_extent_entry *next_ptr;
 731
 732                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 733                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 734                                                         goto found;
 735                                         next_ptr = NULL;
 736 found:
 737                                         if (!next_ptr) {
 738                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 739                                                 continue;
 740                                         }
 741
 742                                         if (!m || !m->alive ||
 743                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 744                                                                        &next_ptr->ptr,
 745                                                                        m->sectors)) {
 746                                                 bch2_bkey_extent_entry_drop(new, entry);
 747                                                 goto again;
 748                                         }
 749                                 }
 750                         }
 751                 }
 752
 753                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 754                 if (ret) {
 755                         kfree(new);
 756                         goto err;
 757                 }
 758
 759                 if (level)
 760                         bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
 761
 762                 if (0) {
 763                         printbuf_reset(&buf);
 764                         bch2_bkey_val_to_text(&buf, c, *k);
 765                         bch_info(c, "updated %s", buf.buf);
 766
 767                         printbuf_reset(&buf);
 768                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 769                         bch_info(c, "new key %s", buf.buf);
 770                 }
 771
 772                 *k = bkey_i_to_s_c(new);
 773         }
 774 err:
 775 fsck_err:
 776         printbuf_exit(&buf);
 777         return ret;
 778 }
 779
 780 /* marking of btree keys/nodes: */
 781
 782 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 783                             unsigned level, bool is_root,
 784                             struct bkey_s_c *k,
 785                             bool initial)
 786 {
 787         struct bch_fs *c = trans->c;
 788         struct bkey deleted = KEY(0, 0, 0);
 789         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 790         unsigned flags =
 791                 BTREE_TRIGGER_GC|
 792                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 793         int ret = 0;
 794
 795         deleted.p = k->k->p;
 796
 797         if (initial) {
 798                 BUG_ON(bch2_journal_seq_verify &&
 799                        k->k->version.lo > atomic64_read(&c->journal.seq));
 800
 801                 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
 802                 if (ret)
 803                         goto err;
 804
 805                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 806                                 "key version number higher than recorded: %llu > %llu",
 807                                 k->k->version.lo,
 808                                 atomic64_read(&c->key_version)))
 809                         atomic64_set(&c->key_version, k->k->version.lo);
 810         }
 811
 812         ret = commit_do(trans, NULL, NULL, 0,
 813                         bch2_mark_key(trans, btree_id, level, old, *k, flags));
 814 fsck_err:
 815 err:
 816         if (ret)
 817                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
 818         return ret;
 819 }
 820
 821 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 822 {
 823         struct bch_fs *c = trans->c;
 824         struct btree_node_iter iter;
 825         struct bkey unpacked;
 826         struct bkey_s_c k;
 827         struct bkey_buf prev, cur;
 828         int ret = 0;
 829
 830         if (!btree_node_type_needs_gc(btree_node_type(b)))
 831                 return 0;
 832
 833         bch2_btree_node_iter_init_from_start(&iter, b);
 834         bch2_bkey_buf_init(&prev);
 835         bch2_bkey_buf_init(&cur);
 836         bkey_init(&prev.k->k);
 837
 838         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 839                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 840                                        &k, initial);
 841                 if (ret)
 842                         break;
 843
 844                 bch2_btree_node_iter_advance(&iter, b);
 845
 846                 if (b->c.level) {
 847                         bch2_bkey_buf_reassemble(&cur, c, k);
 848
 849                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 850                                         bch2_btree_node_iter_end(&iter));
 851                         if (ret)
 852                                 break;
 853                 }
 854         }
 855
 856         bch2_bkey_buf_exit(&cur, c);
 857         bch2_bkey_buf_exit(&prev, c);
 858         return ret;
 859 }
 860
 861 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 862                          bool initial, bool metadata_only)
 863 {
 864         struct bch_fs *c = trans->c;
 865         struct btree_iter iter;
 866         struct btree *b;
 867         unsigned depth = metadata_only ? 1 : 0;
 868         int ret = 0;
 869
 870         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 871
 872         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 873                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 874                 bch2_verify_btree_nr_keys(b);
 875
 876                 gc_pos_set(c, gc_pos_btree_node(b));
 877
 878                 ret = btree_gc_mark_node(trans, b, initial);
 879                 if (ret)
 880                         break;
 881         }
 882         bch2_trans_iter_exit(trans, &iter);
 883
 884         if (ret)
 885                 return ret;
 886
 887         mutex_lock(&c->btree_root_lock);
 888         b = c->btree_roots[btree_id].b;
 889         if (!btree_node_fake(b)) {
 890                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 891
 892                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
 893                                        true, &k, initial);
 894         }
 895         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 896         mutex_unlock(&c->btree_root_lock);
 897
 898         return ret;
 899 }
 900
 901 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 902                                       unsigned target_depth)
 903 {
 904         struct bch_fs *c = trans->c;
 905         struct btree_and_journal_iter iter;
 906         struct bkey_s_c k;
 907         struct bkey_buf cur, prev;
 908         struct printbuf buf = PRINTBUF;
 909         int ret = 0;
 910
 911         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 912         bch2_bkey_buf_init(&prev);
 913         bch2_bkey_buf_init(&cur);
 914         bkey_init(&prev.k->k);
 915
 916         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 917                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 918                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 919
 920                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 921                                        false, &k, true);
 922                 if (ret) {
 923                         bch_err(c, "%s: error from bch2_gc_mark_key: %s",
 924                                 __func__, bch2_err_str(ret));
 925                         goto fsck_err;
 926                 }
 927
 928                 if (b->c.level) {
 929                         bch2_bkey_buf_reassemble(&cur, c, k);
 930                         k = bkey_i_to_s_c(cur.k);
 931
 932                         bch2_btree_and_journal_iter_advance(&iter);
 933
 934                         ret = bch2_gc_check_topology(c, b,
 935                                         &prev, cur,
 936                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 937                         if (ret)
 938                                 goto fsck_err;
 939                 } else {
 940                         bch2_btree_and_journal_iter_advance(&iter);
 941                 }
 942         }
 943
 944         if (b->c.level > target_depth) {
 945                 bch2_btree_and_journal_iter_exit(&iter);
 946                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 947
 948                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 949                         struct btree *child;
 950
 951                         bch2_bkey_buf_reassemble(&cur, c, k);
 952                         bch2_btree_and_journal_iter_advance(&iter);
 953
 954                         child = bch2_btree_node_get_noiter(trans, cur.k,
 955                                                 b->c.btree_id, b->c.level - 1,
 956                                                 false);
 957                         ret = PTR_ERR_OR_ZERO(child);
 958
 959                         if (ret == -EIO) {
 960                                 bch2_topology_error(c);
 961
 962                                 if (__fsck_err(c,
 963                                           FSCK_CAN_FIX|
 964                                           FSCK_CAN_IGNORE|
 965                                           FSCK_NO_RATELIMIT,
 966                                           "Unreadable btree node at btree %s level %u:\n"
 967                                           "  %s",
 968                                           bch2_btree_ids[b->c.btree_id],
 969                                           b->c.level - 1,
 970                                           (printbuf_reset(&buf),
 971                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 972                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 973                                         ret = -BCH_ERR_need_topology_repair;
 974                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 975                                         goto fsck_err;
 976                                 } else {
 977                                         /* Continue marking when opted to not
 978                                          * fix the error: */
 979                                         ret = 0;
 980                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 981                                         continue;
 982                                 }
 983                         } else if (ret) {
 984                                 bch_err(c, "%s: error getting btree node: %s",
 985                                         __func__, bch2_err_str(ret));
 986                                 break;
 987                         }
 988
 989                         ret = bch2_gc_btree_init_recurse(trans, child,
 990                                                          target_depth);
 991                         six_unlock_read(&child->c.lock);
 992
 993                         if (ret)
 994                                 break;
 995                 }
 996         }
 997 fsck_err:
 998         bch2_bkey_buf_exit(&cur, c);
 999         bch2_bkey_buf_exit(&prev, c);
1000         bch2_btree_and_journal_iter_exit(&iter);
1001         printbuf_exit(&buf);
1002         return ret;
1003 }
1004
1005 static int bch2_gc_btree_init(struct btree_trans *trans,
1006                               enum btree_id btree_id,
1007                               bool metadata_only)
1008 {
1009         struct bch_fs *c = trans->c;
1010         struct btree *b;
1011         unsigned target_depth = metadata_only ? 1 : 0;
1012         struct printbuf buf = PRINTBUF;
1013         int ret = 0;
1014
1015         b = c->btree_roots[btree_id].b;
1016
1017         if (btree_node_fake(b))
1018                 return 0;
1019
1020         six_lock_read(&b->c.lock, NULL, NULL);
1021         printbuf_reset(&buf);
1022         bch2_bpos_to_text(&buf, b->data->min_key);
1023         if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1024                         "btree root with incorrect min_key: %s", buf.buf)) {
1025                 bch_err(c, "repair unimplemented");
1026                 ret = -BCH_ERR_fsck_repair_unimplemented;
1027                 goto fsck_err;
1028         }
1029
1030         printbuf_reset(&buf);
1031         bch2_bpos_to_text(&buf, b->data->max_key);
1032         if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1033                         "btree root with incorrect max_key: %s", buf.buf)) {
1034                 bch_err(c, "repair unimplemented");
1035                 ret = -BCH_ERR_fsck_repair_unimplemented;
1036                 goto fsck_err;
1037         }
1038
1039         if (b->c.level >= target_depth)
1040                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1041
1042         if (!ret) {
1043                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1044
1045                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1046                                        &k, true);
1047         }
1048 fsck_err:
1049         six_unlock_read(&b->c.lock);
1050
1051         if (ret < 0)
1052                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1053         printbuf_exit(&buf);
1054         return ret;
1055 }
1056
1057 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1058 {
1059         return  (int) btree_id_to_gc_phase(l) -
1060                 (int) btree_id_to_gc_phase(r);
1061 }
1062
1063 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1064 {
1065         struct btree_trans trans;
1066         enum btree_id ids[BTREE_ID_NR];
1067         unsigned i;
1068         int ret = 0;
1069
1070         bch2_trans_init(&trans, c, 0, 0);
1071
1072         if (initial)
1073                 trans.is_initial_gc = true;
1074
1075         for (i = 0; i < BTREE_ID_NR; i++)
1076                 ids[i] = i;
1077         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1078
1079         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1080                 ret = initial
1081                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1082                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1083
1084         if (ret < 0)
1085                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1086
1087         bch2_trans_exit(&trans);
1088         return ret;
1089 }
1090
1091 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1092                                   u64 start, u64 end,
1093                                   enum bch_data_type type,
1094                                   unsigned flags)
1095 {
1096         u64 b = sector_to_bucket(ca, start);
1097
1098         do {
1099                 unsigned sectors =
1100                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1101
1102                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1103                                           gc_phase(GC_PHASE_SB), flags);
1104                 b++;
1105                 start += sectors;
1106         } while (start < end);
1107 }
1108
1109 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1110                                      unsigned flags)
1111 {
1112         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1113         unsigned i;
1114         u64 b;
1115
1116         for (i = 0; i < layout->nr_superblocks; i++) {
1117                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1118
1119                 if (offset == BCH_SB_SECTOR)
1120                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1121                                               BCH_DATA_sb, flags);
1122
1123                 mark_metadata_sectors(c, ca, offset,
1124                                       offset + (1 << layout->sb_max_size_bits),
1125                                       BCH_DATA_sb, flags);
1126         }
1127
1128         for (i = 0; i < ca->journal.nr; i++) {
1129                 b = ca->journal.buckets[i];
1130                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1131                                           ca->mi.bucket_size,
1132                                           gc_phase(GC_PHASE_SB), flags);
1133         }
1134 }
1135
1136 static void bch2_mark_superblocks(struct bch_fs *c)
1137 {
1138         struct bch_dev *ca;
1139         unsigned i;
1140
1141         mutex_lock(&c->sb_lock);
1142         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1143
1144         for_each_online_member(ca, c, i)
1145                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1146         mutex_unlock(&c->sb_lock);
1147 }
1148
1149 #if 0
1150 /* Also see bch2_pending_btree_node_free_insert_done() */
1151 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1152 {
1153         struct btree_update *as;
1154         struct pending_btree_node_free *d;
1155
1156         mutex_lock(&c->btree_interior_update_lock);
1157         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1158
1159         for_each_pending_btree_node_free(c, as, d)
1160                 if (d->index_update_done)
1161                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1162
1163         mutex_unlock(&c->btree_interior_update_lock);
1164 }
1165 #endif
1166
1167 static void bch2_gc_free(struct bch_fs *c)
1168 {
1169         struct bch_dev *ca;
1170         unsigned i;
1171
1172         genradix_free(&c->reflink_gc_table);
1173         genradix_free(&c->gc_stripes);
1174
1175         for_each_member_device(ca, c, i) {
1176                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1177                         sizeof(struct bucket_array) +
1178                         ca->mi.nbuckets * sizeof(struct bucket));
1179                 ca->buckets_gc = NULL;
1180
1181                 free_percpu(ca->usage_gc);
1182                 ca->usage_gc = NULL;
1183         }
1184
1185         free_percpu(c->usage_gc);
1186         c->usage_gc = NULL;
1187 }
1188
1189 static int bch2_gc_done(struct bch_fs *c,
1190                         bool initial, bool metadata_only)
1191 {
1192         struct bch_dev *ca = NULL;
1193         struct printbuf buf = PRINTBUF;
1194         bool verify = !metadata_only &&
1195                 !c->opts.reconstruct_alloc &&
1196                 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1197         unsigned i, dev;
1198         int ret = 0;
1199
1200         percpu_down_write(&c->mark_lock);
1201
1202 #define copy_field(_f, _msg, ...)                                       \
1203         if (dst->_f != src->_f &&                                       \
1204             (!verify ||                                                 \
1205              fsck_err(c, _msg ": got %llu, should be %llu"              \
1206                       , ##__VA_ARGS__, dst->_f, src->_f)))              \
1207                 dst->_f = src->_f
1208 #define copy_stripe_field(_f, _msg, ...)                                \
1209         if (dst->_f != src->_f &&                                       \
1210             (!verify ||                                                 \
1211              fsck_err(c, "stripe %zu has wrong "_msg                    \
1212                       ": got %u, should be %u",                         \
1213                       iter.pos, ##__VA_ARGS__,                          \
1214                       dst->_f, src->_f)))                               \
1215                 dst->_f = src->_f
1216 #define copy_dev_field(_f, _msg, ...)                                   \
1217         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1218 #define copy_fs_field(_f, _msg, ...)                                    \
1219         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1220
1221         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1222                 bch2_fs_usage_acc_to_base(c, i);
1223
1224         for_each_member_device(ca, c, dev) {
1225                 struct bch_dev_usage *dst = ca->usage_base;
1226                 struct bch_dev_usage *src = (void *)
1227                         bch2_acc_percpu_u64s((void *) ca->usage_gc,
1228                                              dev_usage_u64s());
1229
1230                 copy_dev_field(buckets_ec,              "buckets_ec");
1231
1232                 for (i = 0; i < BCH_DATA_NR; i++) {
1233                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1234                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1235                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1236                 }
1237         };
1238
1239         {
1240                 unsigned nr = fs_usage_u64s(c);
1241                 struct bch_fs_usage *dst = c->usage_base;
1242                 struct bch_fs_usage *src = (void *)
1243                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1244
1245                 copy_fs_field(hidden,           "hidden");
1246                 copy_fs_field(btree,            "btree");
1247
1248                 if (!metadata_only) {
1249                         copy_fs_field(data,     "data");
1250                         copy_fs_field(cached,   "cached");
1251                         copy_fs_field(reserved, "reserved");
1252                         copy_fs_field(nr_inodes,"nr_inodes");
1253
1254                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1255                                 copy_fs_field(persistent_reserved[i],
1256                                               "persistent_reserved[%i]", i);
1257                 }
1258
1259                 for (i = 0; i < c->replicas.nr; i++) {
1260                         struct bch_replicas_entry *e =
1261                                 cpu_replicas_entry(&c->replicas, i);
1262
1263                         if (metadata_only &&
1264                             (e->data_type == BCH_DATA_user ||
1265                              e->data_type == BCH_DATA_cached))
1266                                 continue;
1267
1268                         printbuf_reset(&buf);
1269                         bch2_replicas_entry_to_text(&buf, e);
1270
1271                         copy_fs_field(replicas[i], "%s", buf.buf);
1272                 }
1273         }
1274
1275 #undef copy_fs_field
1276 #undef copy_dev_field
1277 #undef copy_stripe_field
1278 #undef copy_field
1279 fsck_err:
1280         if (ca)
1281                 percpu_ref_put(&ca->ref);
1282         if (ret)
1283                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1284
1285         percpu_up_write(&c->mark_lock);
1286         printbuf_exit(&buf);
1287         return ret;
1288 }
1289
1290 static int bch2_gc_start(struct bch_fs *c)
1291 {
1292         struct bch_dev *ca = NULL;
1293         unsigned i;
1294
1295         BUG_ON(c->usage_gc);
1296
1297         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1298                                          sizeof(u64), GFP_KERNEL);
1299         if (!c->usage_gc) {
1300                 bch_err(c, "error allocating c->usage_gc");
1301                 return -BCH_ERR_ENOMEM_gc_start;
1302         }
1303
1304         for_each_member_device(ca, c, i) {
1305                 BUG_ON(ca->usage_gc);
1306
1307                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1308                 if (!ca->usage_gc) {
1309                         bch_err(c, "error allocating ca->usage_gc");
1310                         percpu_ref_put(&ca->ref);
1311                         return -BCH_ERR_ENOMEM_gc_start;
1312                 }
1313
1314                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1315                                ca->mi.nbuckets - ca->mi.first_bucket);
1316         }
1317
1318         return 0;
1319 }
1320
1321 static int bch2_gc_reset(struct bch_fs *c)
1322 {
1323         struct bch_dev *ca;
1324         unsigned i;
1325
1326         for_each_member_device(ca, c, i) {
1327                 free_percpu(ca->usage_gc);
1328                 ca->usage_gc = NULL;
1329         }
1330
1331         free_percpu(c->usage_gc);
1332         c->usage_gc = NULL;
1333
1334         return bch2_gc_start(c);
1335 }
1336
1337 /* returns true if not equal */
1338 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1339                                      struct bch_alloc_v4 r)
1340 {
1341         return  l.gen != r.gen                          ||
1342                 l.oldest_gen != r.oldest_gen            ||
1343                 l.data_type != r.data_type              ||
1344                 l.dirty_sectors != r.dirty_sectors      ||
1345                 l.cached_sectors != r.cached_sectors     ||
1346                 l.stripe_redundancy != r.stripe_redundancy ||
1347                 l.stripe != r.stripe;
1348 }
1349
1350 static int bch2_alloc_write_key(struct btree_trans *trans,
1351                                 struct btree_iter *iter,
1352                                 struct bkey_s_c k,
1353                                 bool metadata_only)
1354 {
1355         struct bch_fs *c = trans->c;
1356         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1357         struct bucket gc, *b;
1358         struct bkey_i_alloc_v4 *a;
1359         struct bch_alloc_v4 old_convert, new;
1360         const struct bch_alloc_v4 *old;
1361         enum bch_data_type type;
1362         int ret;
1363
1364         if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
1365                 return 1;
1366
1367         old = bch2_alloc_to_v4(k, &old_convert);
1368         new = *old;
1369
1370         percpu_down_read(&c->mark_lock);
1371         b = gc_bucket(ca, iter->pos.offset);
1372
1373         /*
1374          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1375          * fix that here:
1376          */
1377         type = __alloc_data_type(b->dirty_sectors,
1378                                  b->cached_sectors,
1379                                  b->stripe,
1380                                  *old,
1381                                  b->data_type);
1382         if (b->data_type != type) {
1383                 struct bch_dev_usage *u;
1384
1385                 preempt_disable();
1386                 u = this_cpu_ptr(ca->usage_gc);
1387                 u->d[b->data_type].buckets--;
1388                 b->data_type = type;
1389                 u->d[b->data_type].buckets++;
1390                 preempt_enable();
1391         }
1392
1393         gc = *b;
1394         percpu_up_read(&c->mark_lock);
1395
1396         if (metadata_only &&
1397             gc.data_type != BCH_DATA_sb &&
1398             gc.data_type != BCH_DATA_journal &&
1399             gc.data_type != BCH_DATA_btree)
1400                 return 0;
1401
1402         if (gen_after(old->gen, gc.gen))
1403                 return 0;
1404
1405         if (c->opts.reconstruct_alloc ||
1406             fsck_err_on(new.data_type != gc.data_type, c,
1407                         "bucket %llu:%llu gen %u has wrong data_type"
1408                         ": got %s, should be %s",
1409                         iter->pos.inode, iter->pos.offset,
1410                         gc.gen,
1411                         bch2_data_types[new.data_type],
1412                         bch2_data_types[gc.data_type]))
1413                 new.data_type = gc.data_type;
1414
1415 #define copy_bucket_field(_f)                                           \
1416         if (c->opts.reconstruct_alloc ||                                \
1417             fsck_err_on(new._f != gc._f, c,                             \
1418                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1419                         ": got %u, should be %u",                       \
1420                         iter->pos.inode, iter->pos.offset,              \
1421                         gc.gen,                                         \
1422                         bch2_data_types[gc.data_type],                  \
1423                         new._f, gc._f))                                 \
1424                 new._f = gc._f;                                         \
1425
1426         copy_bucket_field(gen);
1427         copy_bucket_field(dirty_sectors);
1428         copy_bucket_field(cached_sectors);
1429         copy_bucket_field(stripe_redundancy);
1430         copy_bucket_field(stripe);
1431 #undef copy_bucket_field
1432
1433         if (!bch2_alloc_v4_cmp(*old, new))
1434                 return 0;
1435
1436         a = bch2_alloc_to_v4_mut(trans, k);
1437         ret = PTR_ERR_OR_ZERO(a);
1438         if (ret)
1439                 return ret;
1440
1441         a->v = new;
1442
1443         /*
1444          * The trigger normally makes sure this is set, but we're not running
1445          * triggers:
1446          */
1447         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1448                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1449
1450         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1451 fsck_err:
1452         return ret;
1453 }
1454
1455 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1456 {
1457         struct btree_trans trans;
1458         struct btree_iter iter;
1459         struct bkey_s_c k;
1460         struct bch_dev *ca;
1461         unsigned i;
1462         int ret = 0;
1463
1464         bch2_trans_init(&trans, c, 0, 0);
1465
1466         for_each_member_device(ca, c, i) {
1467                 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1468                                 POS(ca->dev_idx, ca->mi.first_bucket),
1469                                 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1470                                 NULL, NULL, BTREE_INSERT_LAZY_RW,
1471                         bch2_alloc_write_key(&trans, &iter, k, metadata_only));
1472
1473                 if (ret < 0) {
1474                         bch_err(c, "error writing alloc info: %s", bch2_err_str(ret));
1475                         percpu_ref_put(&ca->ref);
1476                         break;
1477                 }
1478         }
1479
1480         bch2_trans_exit(&trans);
1481         return ret < 0 ? ret : 0;
1482 }
1483
1484 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1485 {
1486         struct bch_dev *ca;
1487         struct btree_trans trans;
1488         struct btree_iter iter;
1489         struct bkey_s_c k;
1490         struct bucket *g;
1491         struct bch_alloc_v4 a_convert;
1492         const struct bch_alloc_v4 *a;
1493         unsigned i;
1494         int ret;
1495
1496         for_each_member_device(ca, c, i) {
1497                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1498                                 ca->mi.nbuckets * sizeof(struct bucket),
1499                                 GFP_KERNEL|__GFP_ZERO);
1500                 if (!buckets) {
1501                         percpu_ref_put(&ca->ref);
1502                         bch_err(c, "error allocating ca->buckets[gc]");
1503                         return -BCH_ERR_ENOMEM_gc_alloc_start;
1504                 }
1505
1506                 buckets->first_bucket   = ca->mi.first_bucket;
1507                 buckets->nbuckets       = ca->mi.nbuckets;
1508                 rcu_assign_pointer(ca->buckets_gc, buckets);
1509         };
1510
1511         bch2_trans_init(&trans, c, 0, 0);
1512
1513         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1514                            BTREE_ITER_PREFETCH, k, ret) {
1515                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1516                 g = gc_bucket(ca, k.k->p.offset);
1517
1518                 a = bch2_alloc_to_v4(k, &a_convert);
1519
1520                 g->gen_valid    = 1;
1521                 g->gen          = a->gen;
1522
1523                 if (metadata_only &&
1524                     (a->data_type == BCH_DATA_user ||
1525                      a->data_type == BCH_DATA_cached ||
1526                      a->data_type == BCH_DATA_parity)) {
1527                         g->data_type            = a->data_type;
1528                         g->dirty_sectors        = a->dirty_sectors;
1529                         g->cached_sectors       = a->cached_sectors;
1530                         g->stripe               = a->stripe;
1531                         g->stripe_redundancy    = a->stripe_redundancy;
1532                 }
1533         }
1534         bch2_trans_iter_exit(&trans, &iter);
1535
1536         bch2_trans_exit(&trans);
1537
1538         if (ret)
1539                 bch_err(c, "error reading alloc info at gc start: %s", bch2_err_str(ret));
1540
1541         return ret;
1542 }
1543
1544 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1545 {
1546         struct bch_dev *ca;
1547         unsigned i;
1548
1549         for_each_member_device(ca, c, i) {
1550                 struct bucket_array *buckets = gc_bucket_array(ca);
1551                 struct bucket *g;
1552
1553                 for_each_bucket(g, buckets) {
1554                         if (metadata_only &&
1555                             (g->data_type == BCH_DATA_user ||
1556                              g->data_type == BCH_DATA_cached ||
1557                              g->data_type == BCH_DATA_parity))
1558                                 continue;
1559                         g->data_type = 0;
1560                         g->dirty_sectors = 0;
1561                         g->cached_sectors = 0;
1562                 }
1563         };
1564 }
1565
1566 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1567                                      struct btree_iter *iter,
1568                                      struct bkey_s_c k,
1569                                      size_t *idx)
1570 {
1571         struct bch_fs *c = trans->c;
1572         const __le64 *refcount = bkey_refcount_c(k);
1573         struct printbuf buf = PRINTBUF;
1574         struct reflink_gc *r;
1575         int ret = 0;
1576
1577         if (!refcount)
1578                 return 0;
1579
1580         while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1581                r->offset < k.k->p.offset)
1582                 ++*idx;
1583
1584         if (!r ||
1585             r->offset != k.k->p.offset ||
1586             r->size != k.k->size) {
1587                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1588                 return -EINVAL;
1589         }
1590
1591         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1592                         "reflink key has wrong refcount:\n"
1593                         "  %s\n"
1594                         "  should be %u",
1595                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1596                         r->refcount)) {
1597                 struct bkey_i *new = bch2_bkey_make_mut(trans, k);
1598
1599                 ret = PTR_ERR_OR_ZERO(new);
1600                 if (ret)
1601                         return ret;
1602
1603                 if (!r->refcount)
1604                         new->k.type = KEY_TYPE_deleted;
1605                 else
1606                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1607
1608                 ret = bch2_trans_update(trans, iter, new, 0);
1609         }
1610 fsck_err:
1611         printbuf_exit(&buf);
1612         return ret;
1613 }
1614
1615 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1616 {
1617         struct btree_trans trans;
1618         struct btree_iter iter;
1619         struct bkey_s_c k;
1620         size_t idx = 0;
1621         int ret = 0;
1622
1623         if (metadata_only)
1624                 return 0;
1625
1626         bch2_trans_init(&trans, c, 0, 0);
1627
1628         ret = for_each_btree_key_commit(&trans, iter,
1629                         BTREE_ID_reflink, POS_MIN,
1630                         BTREE_ITER_PREFETCH, k,
1631                         NULL, NULL, BTREE_INSERT_NOFAIL,
1632                 bch2_gc_write_reflink_key(&trans, &iter, k, &idx));
1633
1634         c->reflink_gc_nr = 0;
1635         bch2_trans_exit(&trans);
1636         return ret;
1637 }
1638
1639 static int bch2_gc_reflink_start(struct bch_fs *c,
1640                                  bool metadata_only)
1641 {
1642         struct btree_trans trans;
1643         struct btree_iter iter;
1644         struct bkey_s_c k;
1645         struct reflink_gc *r;
1646         int ret = 0;
1647
1648         if (metadata_only)
1649                 return 0;
1650
1651         bch2_trans_init(&trans, c, 0, 0);
1652         c->reflink_gc_nr = 0;
1653
1654         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1655                            BTREE_ITER_PREFETCH, k, ret) {
1656                 const __le64 *refcount = bkey_refcount_c(k);
1657
1658                 if (!refcount)
1659                         continue;
1660
1661                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1662                                        GFP_KERNEL);
1663                 if (!r) {
1664                         ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1665                         break;
1666                 }
1667
1668                 r->offset       = k.k->p.offset;
1669                 r->size         = k.k->size;
1670                 r->refcount     = 0;
1671         }
1672         bch2_trans_iter_exit(&trans, &iter);
1673
1674         bch2_trans_exit(&trans);
1675         return ret;
1676 }
1677
1678 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1679 {
1680         struct genradix_iter iter;
1681         struct reflink_gc *r;
1682
1683         genradix_for_each(&c->reflink_gc_table, iter, r)
1684                 r->refcount = 0;
1685 }
1686
1687 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1688                                      struct btree_iter *iter,
1689                                      struct bkey_s_c k)
1690 {
1691         struct bch_fs *c = trans->c;
1692         struct printbuf buf = PRINTBUF;
1693         const struct bch_stripe *s;
1694         struct gc_stripe *m;
1695         bool bad = false;
1696         unsigned i;
1697         int ret = 0;
1698
1699         if (k.k->type != KEY_TYPE_stripe)
1700                 return 0;
1701
1702         s = bkey_s_c_to_stripe(k).v;
1703         m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1704
1705         for (i = 0; i < s->nr_blocks; i++) {
1706                 u32 old = stripe_blockcount_get(s, i);
1707                 u32 new = (m ? m->block_sectors[i] : 0);
1708
1709                 if (old != new) {
1710                         prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1711                                    i, old, new);
1712                         bad = true;
1713                 }
1714         }
1715
1716         if (bad)
1717                 bch2_bkey_val_to_text(&buf, c, k);
1718
1719         if (fsck_err_on(bad, c, "%s", buf.buf)) {
1720                 struct bkey_i_stripe *new;
1721
1722                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1723                 ret = PTR_ERR_OR_ZERO(new);
1724                 if (ret)
1725                         return ret;
1726
1727                 bkey_reassemble(&new->k_i, k);
1728
1729                 for (i = 0; i < new->v.nr_blocks; i++)
1730                         stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1731
1732                 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1733         }
1734 fsck_err:
1735         printbuf_exit(&buf);
1736         return ret;
1737 }
1738
1739 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1740 {
1741         struct btree_trans trans;
1742         struct btree_iter iter;
1743         struct bkey_s_c k;
1744         int ret = 0;
1745
1746         if (metadata_only)
1747                 return 0;
1748
1749         bch2_trans_init(&trans, c, 0, 0);
1750
1751         ret = for_each_btree_key_commit(&trans, iter,
1752                         BTREE_ID_stripes, POS_MIN,
1753                         BTREE_ITER_PREFETCH, k,
1754                         NULL, NULL, BTREE_INSERT_NOFAIL,
1755                 bch2_gc_write_stripes_key(&trans, &iter, k));
1756
1757         bch2_trans_exit(&trans);
1758         return ret;
1759 }
1760
1761 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1762 {
1763         genradix_free(&c->gc_stripes);
1764 }
1765
1766 /**
1767  * bch2_gc - walk _all_ references to buckets, and recompute them:
1768  *
1769  * Order matters here:
1770  *  - Concurrent GC relies on the fact that we have a total ordering for
1771  *    everything that GC walks - see  gc_will_visit_node(),
1772  *    gc_will_visit_root()
1773  *
1774  *  - also, references move around in the course of index updates and
1775  *    various other crap: everything needs to agree on the ordering
1776  *    references are allowed to move around in - e.g., we're allowed to
1777  *    start with a reference owned by an open_bucket (the allocator) and
1778  *    move it to the btree, but not the reverse.
1779  *
1780  *    This is necessary to ensure that gc doesn't miss references that
1781  *    move around - if references move backwards in the ordering GC
1782  *    uses, GC could skip past them
1783  */
1784 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1785 {
1786         unsigned iter = 0;
1787         int ret;
1788
1789         lockdep_assert_held(&c->state_lock);
1790
1791         down_write(&c->gc_lock);
1792
1793         bch2_btree_interior_updates_flush(c);
1794
1795         ret   = bch2_gc_start(c) ?:
1796                 bch2_gc_alloc_start(c, metadata_only) ?:
1797                 bch2_gc_reflink_start(c, metadata_only);
1798         if (ret)
1799                 goto out;
1800 again:
1801         gc_pos_set(c, gc_phase(GC_PHASE_START));
1802
1803         bch2_mark_superblocks(c);
1804
1805         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1806             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1807             c->opts.fix_errors != FSCK_OPT_NO) {
1808                 bch_info(c, "Starting topology repair pass");
1809                 ret = bch2_repair_topology(c);
1810                 if (ret)
1811                         goto out;
1812                 bch_info(c, "Topology repair pass done");
1813
1814                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1815         }
1816
1817         ret = bch2_gc_btrees(c, initial, metadata_only);
1818
1819         if (ret == -BCH_ERR_need_topology_repair &&
1820             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1821             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1822                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1823                 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1824                 ret = 0;
1825         }
1826
1827         if (ret == -BCH_ERR_need_topology_repair)
1828                 ret = -BCH_ERR_fsck_errors_not_fixed;
1829
1830         if (ret)
1831                 goto out;
1832
1833 #if 0
1834         bch2_mark_pending_btree_node_frees(c);
1835 #endif
1836         c->gc_count++;
1837
1838         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1839             (!iter && bch2_test_restart_gc)) {
1840                 if (iter++ > 2) {
1841                         bch_info(c, "Unable to fix bucket gens, looping");
1842                         ret = -EINVAL;
1843                         goto out;
1844                 }
1845
1846                 /*
1847                  * XXX: make sure gens we fixed got saved
1848                  */
1849                 bch_info(c, "Second GC pass needed, restarting:");
1850                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1851                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1852
1853                 bch2_gc_stripes_reset(c, metadata_only);
1854                 bch2_gc_alloc_reset(c, metadata_only);
1855                 bch2_gc_reflink_reset(c, metadata_only);
1856                 ret = bch2_gc_reset(c);
1857                 if (ret)
1858                         goto out;
1859
1860                 /* flush fsck errors, reset counters */
1861                 bch2_flush_fsck_errs(c);
1862                 goto again;
1863         }
1864 out:
1865         if (!ret) {
1866                 bch2_journal_block(&c->journal);
1867
1868                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1869                         bch2_gc_reflink_done(c, metadata_only) ?:
1870                         bch2_gc_alloc_done(c, metadata_only) ?:
1871                         bch2_gc_done(c, initial, metadata_only);
1872
1873                 bch2_journal_unblock(&c->journal);
1874         }
1875
1876         percpu_down_write(&c->mark_lock);
1877         /* Indicates that gc is no longer in progress: */
1878         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1879
1880         bch2_gc_free(c);
1881         percpu_up_write(&c->mark_lock);
1882
1883         up_write(&c->gc_lock);
1884
1885         /*
1886          * At startup, allocations can happen directly instead of via the
1887          * allocator thread - issue wakeup in case they blocked on gc_lock:
1888          */
1889         closure_wake_up(&c->freelist_wait);
1890         return ret;
1891 }
1892
1893 static int gc_btree_gens_key(struct btree_trans *trans,
1894                              struct btree_iter *iter,
1895                              struct bkey_s_c k)
1896 {
1897         struct bch_fs *c = trans->c;
1898         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1899         const struct bch_extent_ptr *ptr;
1900         struct bkey_i *u;
1901         int ret;
1902
1903         percpu_down_read(&c->mark_lock);
1904         bkey_for_each_ptr(ptrs, ptr) {
1905                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1906
1907                 if (ptr_stale(ca, ptr) > 16) {
1908                         percpu_up_read(&c->mark_lock);
1909                         goto update;
1910                 }
1911         }
1912
1913         bkey_for_each_ptr(ptrs, ptr) {
1914                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1915                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1916
1917                 if (gen_after(*gen, ptr->gen))
1918                         *gen = ptr->gen;
1919         }
1920         percpu_up_read(&c->mark_lock);
1921         return 0;
1922 update:
1923         u = bch2_bkey_make_mut(trans, k);
1924         ret = PTR_ERR_OR_ZERO(u);
1925         if (ret)
1926                 return ret;
1927
1928         bch2_extent_normalize(c, bkey_i_to_s(u));
1929         return bch2_trans_update(trans, iter, u, 0);
1930 }
1931
1932 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1933                                        struct bkey_s_c k)
1934 {
1935         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1936         struct bch_alloc_v4 a_convert;
1937         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1938         struct bkey_i_alloc_v4 *a_mut;
1939         int ret;
1940
1941         if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1942                 return 0;
1943
1944         a_mut = bch2_alloc_to_v4_mut(trans, k);
1945         ret = PTR_ERR_OR_ZERO(a_mut);
1946         if (ret)
1947                 return ret;
1948
1949         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1950         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1951
1952         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1953 }
1954
1955 int bch2_gc_gens(struct bch_fs *c)
1956 {
1957         struct btree_trans trans;
1958         struct btree_iter iter;
1959         struct bkey_s_c k;
1960         struct bch_dev *ca;
1961         u64 b, start_time = local_clock();
1962         unsigned i;
1963         int ret;
1964
1965         /*
1966          * Ideally we would be using state_lock and not gc_lock here, but that
1967          * introduces a deadlock in the RO path - we currently take the state
1968          * lock at the start of going RO, thus the gc thread may get stuck:
1969          */
1970         if (!mutex_trylock(&c->gc_gens_lock))
1971                 return 0;
1972
1973         trace_and_count(c, gc_gens_start, c);
1974         down_read(&c->gc_lock);
1975         bch2_trans_init(&trans, c, 0, 0);
1976
1977         for_each_member_device(ca, c, i) {
1978                 struct bucket_gens *gens;
1979
1980                 BUG_ON(ca->oldest_gen);
1981
1982                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1983                 if (!ca->oldest_gen) {
1984                         percpu_ref_put(&ca->ref);
1985                         ret = -BCH_ERR_ENOMEM_gc_gens;
1986                         goto err;
1987                 }
1988
1989                 gens = bucket_gens(ca);
1990
1991                 for (b = gens->first_bucket;
1992                      b < gens->nbuckets; b++)
1993                         ca->oldest_gen[b] = gens->b[b];
1994         }
1995
1996         for (i = 0; i < BTREE_ID_NR; i++)
1997                 if (btree_type_has_ptrs(i)) {
1998                         struct btree_iter iter;
1999                         struct bkey_s_c k;
2000
2001                         c->gc_gens_btree = i;
2002                         c->gc_gens_pos = POS_MIN;
2003                         ret = for_each_btree_key_commit(&trans, iter, i,
2004                                         POS_MIN,
2005                                         BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
2006                                         k,
2007                                         NULL, NULL,
2008                                         BTREE_INSERT_NOFAIL,
2009                                 gc_btree_gens_key(&trans, &iter, k));
2010                         if (ret && !bch2_err_matches(ret, EROFS))
2011                                 bch_err(c, "error recalculating oldest_gen: %s", bch2_err_str(ret));
2012                         if (ret)
2013                                 goto err;
2014                 }
2015
2016         ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
2017                         POS_MIN,
2018                         BTREE_ITER_PREFETCH,
2019                         k,
2020                         NULL, NULL,
2021                         BTREE_INSERT_NOFAIL,
2022                 bch2_alloc_write_oldest_gen(&trans, &iter, k));
2023         if (ret && !bch2_err_matches(ret, EROFS))
2024                 bch_err(c, "error writing oldest_gen: %s", bch2_err_str(ret));
2025         if (ret)
2026                 goto err;
2027
2028         c->gc_gens_btree        = 0;
2029         c->gc_gens_pos          = POS_MIN;
2030
2031         c->gc_count++;
2032
2033         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2034         trace_and_count(c, gc_gens_end, c);
2035 err:
2036         for_each_member_device(ca, c, i) {
2037                 kvfree(ca->oldest_gen);
2038                 ca->oldest_gen = NULL;
2039         }
2040
2041         bch2_trans_exit(&trans);
2042         up_read(&c->gc_lock);
2043         mutex_unlock(&c->gc_gens_lock);
2044         return ret;
2045 }
2046
2047 static int bch2_gc_thread(void *arg)
2048 {
2049         struct bch_fs *c = arg;
2050         struct io_clock *clock = &c->io_clock[WRITE];
2051         unsigned long last = atomic64_read(&clock->now);
2052         unsigned last_kick = atomic_read(&c->kick_gc);
2053         int ret;
2054
2055         set_freezable();
2056
2057         while (1) {
2058                 while (1) {
2059                         set_current_state(TASK_INTERRUPTIBLE);
2060
2061                         if (kthread_should_stop()) {
2062                                 __set_current_state(TASK_RUNNING);
2063                                 return 0;
2064                         }
2065
2066                         if (atomic_read(&c->kick_gc) != last_kick)
2067                                 break;
2068
2069                         if (c->btree_gc_periodic) {
2070                                 unsigned long next = last + c->capacity / 16;
2071
2072                                 if (atomic64_read(&clock->now) >= next)
2073                                         break;
2074
2075                                 bch2_io_clock_schedule_timeout(clock, next);
2076                         } else {
2077                                 schedule();
2078                         }
2079
2080                         try_to_freeze();
2081                 }
2082                 __set_current_state(TASK_RUNNING);
2083
2084                 last = atomic64_read(&clock->now);
2085                 last_kick = atomic_read(&c->kick_gc);
2086
2087                 /*
2088                  * Full gc is currently incompatible with btree key cache:
2089                  */
2090 #if 0
2091                 ret = bch2_gc(c, false, false);
2092 #else
2093                 ret = bch2_gc_gens(c);
2094 #endif
2095                 if (ret < 0)
2096                         bch_err(c, "btree gc failed: %s", bch2_err_str(ret));
2097
2098                 debug_check_no_locks_held();
2099         }
2100
2101         return 0;
2102 }
2103
2104 void bch2_gc_thread_stop(struct bch_fs *c)
2105 {
2106         struct task_struct *p;
2107
2108         p = c->gc_thread;
2109         c->gc_thread = NULL;
2110
2111         if (p) {
2112                 kthread_stop(p);
2113                 put_task_struct(p);
2114         }
2115 }
2116
2117 int bch2_gc_thread_start(struct bch_fs *c)
2118 {
2119         struct task_struct *p;
2120
2121         if (c->gc_thread)
2122                 return 0;
2123
2124         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2125         if (IS_ERR(p)) {
2126                 bch_err(c, "error creating gc thread: %s", bch2_err_str(PTR_ERR(p)));
2127                 return PTR_ERR(p);
2128         }
2129
2130         get_task_struct(p);
2131         c->gc_thread = p;
2132         wake_up_process(p);
2133         return 0;
2134 }