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 bool should_restart_for_topology_repair(struct bch_fs *c)
  45 {
  46         return c->opts.fix_errors != FSCK_FIX_no &&
  47                 !(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
  48 }
  49
  50 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  51 {
  52         preempt_disable();
  53         write_seqcount_begin(&c->gc_pos_lock);
  54         c->gc_pos = new_pos;
  55         write_seqcount_end(&c->gc_pos_lock);
  56         preempt_enable();
  57 }
  58
  59 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  60 {
  61         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  62         __gc_pos_set(c, new_pos);
  63 }
  64
  65 /*
  66  * Missing: if an interior btree node is empty, we need to do something -
  67  * perhaps just kill it
  68  */
  69 static int bch2_gc_check_topology(struct bch_fs *c,
  70                                   struct btree *b,
  71                                   struct bkey_buf *prev,
  72                                   struct bkey_buf cur,
  73                                   bool is_last)
  74 {
  75         struct bpos node_start  = b->data->min_key;
  76         struct bpos node_end    = b->data->max_key;
  77         struct bpos expected_start = bkey_deleted(&prev->k->k)
  78                 ? node_start
  79                 : bpos_successor(prev->k->k.p);
  80         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  81         int ret = 0;
  82
  83         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  84                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  85
  86                 if (!bpos_eq(expected_start, bp->v.min_key)) {
  87                         bch2_topology_error(c);
  88
  89                         if (bkey_deleted(&prev->k->k)) {
  90                                 prt_printf(&buf1, "start of node: ");
  91                                 bch2_bpos_to_text(&buf1, node_start);
  92                         } else {
  93                                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
  94                         }
  95                         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
  96
  97                         if (__fsck_err(c,
  98                                   FSCK_CAN_FIX|
  99                                   FSCK_CAN_IGNORE|
 100                                   FSCK_NO_RATELIMIT,
 101                                   "btree node with incorrect min_key at btree %s level %u:\n"
 102                                   "  prev %s\n"
 103                                   "  cur %s",
 104                                   bch2_btree_id_str(b->c.btree_id), b->c.level,
 105                                   buf1.buf, buf2.buf) &&
 106                             should_restart_for_topology_repair(c)) {
 107                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 108                                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 109                                 goto err;
 110                         } else {
 111                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 112                         }
 113                 }
 114         }
 115
 116         if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
 117                 bch2_topology_error(c);
 118
 119                 printbuf_reset(&buf1);
 120                 printbuf_reset(&buf2);
 121
 122                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
 123                 bch2_bpos_to_text(&buf2, node_end);
 124
 125                 if (__fsck_err(c,
 126                           FSCK_CAN_FIX|
 127                           FSCK_CAN_IGNORE|
 128                           FSCK_NO_RATELIMIT,
 129                           "btree node with incorrect max_key at btree %s level %u:\n"
 130                           "  %s\n"
 131                           "  expected %s",
 132                           bch2_btree_id_str(b->c.btree_id), b->c.level,
 133                           buf1.buf, buf2.buf) &&
 134                     should_restart_for_topology_repair(c)) {
 135                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 136                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 137                         goto err;
 138                 } else {
 139                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 140                 }
 141         }
 142
 143         bch2_bkey_buf_copy(prev, c, cur.k);
 144 err:
 145 fsck_err:
 146         printbuf_exit(&buf2);
 147         printbuf_exit(&buf1);
 148         return ret;
 149 }
 150
 151 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 152 {
 153         switch (b->key.k.type) {
 154         case KEY_TYPE_btree_ptr: {
 155                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 156
 157                 dst->k.p                = src->k.p;
 158                 dst->v.mem_ptr          = 0;
 159                 dst->v.seq              = b->data->keys.seq;
 160                 dst->v.sectors_written  = 0;
 161                 dst->v.flags            = 0;
 162                 dst->v.min_key          = b->data->min_key;
 163                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 164                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 165                 break;
 166         }
 167         case KEY_TYPE_btree_ptr_v2:
 168                 bkey_copy(&dst->k_i, &b->key);
 169                 break;
 170         default:
 171                 BUG();
 172         }
 173 }
 174
 175 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
 176                                              enum btree_id btree, unsigned level,
 177                                              struct bkey_s_c old, struct bkey_i *new)
 178 {
 179         struct bch_fs *c = trans->c;
 180         struct btree *b;
 181         struct bkey_buf tmp;
 182         int ret;
 183
 184         bch2_bkey_buf_init(&tmp);
 185         bch2_bkey_buf_reassemble(&tmp, c, old);
 186
 187         b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
 188         if (!IS_ERR_OR_NULL(b)) {
 189                 mutex_lock(&c->btree_cache.lock);
 190
 191                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 192
 193                 bkey_copy(&b->key, new);
 194                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 195                 BUG_ON(ret);
 196
 197                 mutex_unlock(&c->btree_cache.lock);
 198                 six_unlock_read(&b->c.lock);
 199         }
 200
 201         bch2_bkey_buf_exit(&tmp, c);
 202 }
 203
 204 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 205 {
 206         struct bkey_i_btree_ptr_v2 *new;
 207         int ret;
 208
 209         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 210         if (!new)
 211                 return -BCH_ERR_ENOMEM_gc_repair_key;
 212
 213         btree_ptr_to_v2(b, new);
 214         b->data->min_key        = new_min;
 215         new->v.min_key          = new_min;
 216         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 217
 218         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 219         if (ret) {
 220                 kfree(new);
 221                 return ret;
 222         }
 223
 224         bch2_btree_node_drop_keys_outside_node(b);
 225         bkey_copy(&b->key, &new->k_i);
 226         return 0;
 227 }
 228
 229 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 230 {
 231         struct bkey_i_btree_ptr_v2 *new;
 232         int ret;
 233
 234         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 235         if (ret)
 236                 return ret;
 237
 238         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 239         if (!new)
 240                 return -BCH_ERR_ENOMEM_gc_repair_key;
 241
 242         btree_ptr_to_v2(b, new);
 243         b->data->max_key        = new_max;
 244         new->k.p                = new_max;
 245         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 246
 247         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 248         if (ret) {
 249                 kfree(new);
 250                 return ret;
 251         }
 252
 253         bch2_btree_node_drop_keys_outside_node(b);
 254
 255         mutex_lock(&c->btree_cache.lock);
 256         bch2_btree_node_hash_remove(&c->btree_cache, b);
 257
 258         bkey_copy(&b->key, &new->k_i);
 259         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 260         BUG_ON(ret);
 261         mutex_unlock(&c->btree_cache.lock);
 262         return 0;
 263 }
 264
 265 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 266                                         struct btree *prev, struct btree *cur)
 267 {
 268         struct bpos expected_start = !prev
 269                 ? b->data->min_key
 270                 : bpos_successor(prev->key.k.p);
 271         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 272         int ret = 0;
 273
 274         if (!prev) {
 275                 prt_printf(&buf1, "start of node: ");
 276                 bch2_bpos_to_text(&buf1, b->data->min_key);
 277         } else {
 278                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
 279         }
 280
 281         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
 282
 283         if (prev &&
 284             bpos_gt(expected_start, cur->data->min_key) &&
 285             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 286                 /* cur overwrites prev: */
 287
 288                 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
 289                                                 cur->data->min_key), c,
 290                                 "btree node overwritten by next node at btree %s level %u:\n"
 291                                 "  node %s\n"
 292                                 "  next %s",
 293                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 294                                 buf1.buf, buf2.buf)) {
 295                         ret = DROP_PREV_NODE;
 296                         goto out;
 297                 }
 298
 299                 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
 300                                                  bpos_predecessor(cur->data->min_key)), c,
 301                                 "btree node with incorrect max_key at btree %s level %u:\n"
 302                                 "  node %s\n"
 303                                 "  next %s",
 304                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 305                                 buf1.buf, buf2.buf))
 306                         ret = set_node_max(c, prev,
 307                                            bpos_predecessor(cur->data->min_key));
 308         } else {
 309                 /* prev overwrites cur: */
 310
 311                 if (mustfix_fsck_err_on(bpos_ge(expected_start,
 312                                                 cur->data->max_key), c,
 313                                 "btree node overwritten by prev node at btree %s level %u:\n"
 314                                 "  prev %s\n"
 315                                 "  node %s",
 316                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 317                                 buf1.buf, buf2.buf)) {
 318                         ret = DROP_THIS_NODE;
 319                         goto out;
 320                 }
 321
 322                 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
 323                                 "btree node with incorrect min_key at btree %s level %u:\n"
 324                                 "  prev %s\n"
 325                                 "  node %s",
 326                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 327                                 buf1.buf, buf2.buf))
 328                         ret = set_node_min(c, cur, expected_start);
 329         }
 330 out:
 331 fsck_err:
 332         printbuf_exit(&buf2);
 333         printbuf_exit(&buf1);
 334         return ret;
 335 }
 336
 337 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 338                                  struct btree *child)
 339 {
 340         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 341         int ret = 0;
 342
 343         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
 344         bch2_bpos_to_text(&buf2, b->key.k.p);
 345
 346         if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
 347                         "btree node with incorrect max_key at btree %s level %u:\n"
 348                         "  %s\n"
 349                         "  expected %s",
 350                         bch2_btree_id_str(b->c.btree_id), b->c.level,
 351                         buf1.buf, buf2.buf)) {
 352                 ret = set_node_max(c, child, b->key.k.p);
 353                 if (ret)
 354                         goto err;
 355         }
 356 err:
 357 fsck_err:
 358         printbuf_exit(&buf2);
 359         printbuf_exit(&buf1);
 360         return ret;
 361 }
 362
 363 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
 364 {
 365         struct bch_fs *c = trans->c;
 366         struct btree_and_journal_iter iter;
 367         struct bkey_s_c k;
 368         struct bkey_buf prev_k, cur_k;
 369         struct btree *prev = NULL, *cur = NULL;
 370         bool have_child, dropped_children = false;
 371         struct printbuf buf = PRINTBUF;
 372         int ret = 0;
 373
 374         if (!b->c.level)
 375                 return 0;
 376 again:
 377         prev = NULL;
 378         have_child = dropped_children = false;
 379         bch2_bkey_buf_init(&prev_k);
 380         bch2_bkey_buf_init(&cur_k);
 381         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 382
 383         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 384                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 385                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 386
 387                 bch2_btree_and_journal_iter_advance(&iter);
 388                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 389
 390                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 391                                         b->c.btree_id, b->c.level - 1,
 392                                         false);
 393                 ret = PTR_ERR_OR_ZERO(cur);
 394
 395                 printbuf_reset(&buf);
 396                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 397
 398                 if (mustfix_fsck_err_on(ret == -EIO, c,
 399                                 "Topology repair: unreadable btree node at btree %s level %u:\n"
 400                                 "  %s",
 401                                 bch2_btree_id_str(b->c.btree_id),
 402                                 b->c.level - 1,
 403                                 buf.buf)) {
 404                         bch2_btree_node_evict(trans, cur_k.k);
 405                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 406                                                       b->c.level, cur_k.k->k.p);
 407                         cur = NULL;
 408                         if (ret)
 409                                 break;
 410                         continue;
 411                 }
 412
 413                 if (ret) {
 414                         bch_err_msg(c, ret, "getting btree node");
 415                         break;
 416                 }
 417
 418                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 419
 420                 if (ret == DROP_THIS_NODE) {
 421                         six_unlock_read(&cur->c.lock);
 422                         bch2_btree_node_evict(trans, cur_k.k);
 423                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 424                                                       b->c.level, cur_k.k->k.p);
 425                         cur = NULL;
 426                         if (ret)
 427                                 break;
 428                         continue;
 429                 }
 430
 431                 if (prev)
 432                         six_unlock_read(&prev->c.lock);
 433                 prev = NULL;
 434
 435                 if (ret == DROP_PREV_NODE) {
 436                         bch2_btree_node_evict(trans, prev_k.k);
 437                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 438                                                       b->c.level, prev_k.k->k.p);
 439                         if (ret)
 440                                 break;
 441
 442                         bch2_btree_and_journal_iter_exit(&iter);
 443                         bch2_bkey_buf_exit(&prev_k, c);
 444                         bch2_bkey_buf_exit(&cur_k, c);
 445                         goto again;
 446                 } else if (ret)
 447                         break;
 448
 449                 prev = cur;
 450                 cur = NULL;
 451                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 452         }
 453
 454         if (!ret && !IS_ERR_OR_NULL(prev)) {
 455                 BUG_ON(cur);
 456                 ret = btree_repair_node_end(c, b, prev);
 457         }
 458
 459         if (!IS_ERR_OR_NULL(prev))
 460                 six_unlock_read(&prev->c.lock);
 461         prev = NULL;
 462         if (!IS_ERR_OR_NULL(cur))
 463                 six_unlock_read(&cur->c.lock);
 464         cur = NULL;
 465
 466         if (ret)
 467                 goto err;
 468
 469         bch2_btree_and_journal_iter_exit(&iter);
 470         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 471
 472         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 473                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 474                 bch2_btree_and_journal_iter_advance(&iter);
 475
 476                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 477                                         b->c.btree_id, b->c.level - 1,
 478                                         false);
 479                 ret = PTR_ERR_OR_ZERO(cur);
 480
 481                 if (ret) {
 482                         bch_err_msg(c, ret, "getting btree node");
 483                         goto err;
 484                 }
 485
 486                 ret = bch2_btree_repair_topology_recurse(trans, cur);
 487                 six_unlock_read(&cur->c.lock);
 488                 cur = NULL;
 489
 490                 if (ret == DROP_THIS_NODE) {
 491                         bch2_btree_node_evict(trans, cur_k.k);
 492                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 493                                                       b->c.level, cur_k.k->k.p);
 494                         dropped_children = true;
 495                 }
 496
 497                 if (ret)
 498                         goto err;
 499
 500                 have_child = true;
 501         }
 502
 503         printbuf_reset(&buf);
 504         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 505
 506         if (mustfix_fsck_err_on(!have_child, c,
 507                         "empty interior btree node at btree %s level %u\n"
 508                         "  %s",
 509                         bch2_btree_id_str(b->c.btree_id),
 510                         b->c.level, buf.buf))
 511                 ret = DROP_THIS_NODE;
 512 err:
 513 fsck_err:
 514         if (!IS_ERR_OR_NULL(prev))
 515                 six_unlock_read(&prev->c.lock);
 516         if (!IS_ERR_OR_NULL(cur))
 517                 six_unlock_read(&cur->c.lock);
 518
 519         bch2_btree_and_journal_iter_exit(&iter);
 520         bch2_bkey_buf_exit(&prev_k, c);
 521         bch2_bkey_buf_exit(&cur_k, c);
 522
 523         if (!ret && dropped_children)
 524                 goto again;
 525
 526         printbuf_exit(&buf);
 527         return ret;
 528 }
 529
 530 int bch2_check_topology(struct bch_fs *c)
 531 {
 532         struct btree_trans *trans = bch2_trans_get(c);
 533         struct btree *b;
 534         unsigned i;
 535         int ret = 0;
 536
 537         for (i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
 538                 struct btree_root *r = bch2_btree_id_root(c, i);
 539
 540                 if (!r->alive)
 541                         continue;
 542
 543                 b = r->b;
 544                 if (btree_node_fake(b))
 545                         continue;
 546
 547                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 548                 ret = bch2_btree_repair_topology_recurse(trans, b);
 549                 six_unlock_read(&b->c.lock);
 550
 551                 if (ret == DROP_THIS_NODE) {
 552                         bch_err(c, "empty btree root - repair unimplemented");
 553                         ret = -BCH_ERR_fsck_repair_unimplemented;
 554                 }
 555         }
 556
 557         bch2_trans_put(trans);
 558
 559         return ret;
 560 }
 561
 562 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
 563                                unsigned level, bool is_root,
 564                                struct bkey_s_c *k)
 565 {
 566         struct bch_fs *c = trans->c;
 567         struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
 568         const union bch_extent_entry *entry_c;
 569         struct extent_ptr_decoded p = { 0 };
 570         bool do_update = false;
 571         struct printbuf buf = PRINTBUF;
 572         int ret = 0;
 573
 574         /*
 575          * XXX
 576          * use check_bucket_ref here
 577          */
 578         bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
 579                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 580                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 581                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry_c->ptr);
 582
 583                 if (!g->gen_valid &&
 584                     (c->opts.reconstruct_alloc ||
 585                      fsck_err(c, "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 586                               "while marking %s",
 587                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 588                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 589                               p.ptr.gen,
 590                               (printbuf_reset(&buf),
 591                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 592                         if (!p.ptr.cached) {
 593                                 g->gen_valid            = true;
 594                                 g->gen                  = p.ptr.gen;
 595                         } else {
 596                                 do_update = true;
 597                         }
 598                 }
 599
 600                 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
 601                     (c->opts.reconstruct_alloc ||
 602                      fsck_err(c, "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 603                               "while marking %s",
 604                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 605                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 606                               p.ptr.gen, g->gen,
 607                               (printbuf_reset(&buf),
 608                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 609                         if (!p.ptr.cached) {
 610                                 g->gen_valid            = true;
 611                                 g->gen                  = p.ptr.gen;
 612                                 g->data_type            = 0;
 613                                 g->dirty_sectors        = 0;
 614                                 g->cached_sectors       = 0;
 615                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 616                         } else {
 617                                 do_update = true;
 618                         }
 619                 }
 620
 621                 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
 622                     (c->opts.reconstruct_alloc ||
 623                      fsck_err(c, "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 624                               "while marking %s",
 625                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 626                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 627                               p.ptr.gen,
 628                               (printbuf_reset(&buf),
 629                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 630                         do_update = true;
 631
 632                 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
 633                     (c->opts.reconstruct_alloc ||
 634                      fsck_err(c, "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 635                               "while marking %s",
 636                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 637                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 638                               p.ptr.gen, g->gen,
 639                               (printbuf_reset(&buf),
 640                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 641                         do_update = true;
 642
 643                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 644                         continue;
 645
 646                 if (fsck_err_on(bucket_data_type(g->data_type) &&
 647                                 bucket_data_type(g->data_type) != data_type, c,
 648                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 649                                 "while marking %s",
 650                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 651                                 bch2_data_types[g->data_type],
 652                                 bch2_data_types[data_type],
 653                                 (printbuf_reset(&buf),
 654                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 655                         if (data_type == BCH_DATA_btree) {
 656                                 g->data_type    = data_type;
 657                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 658                         } else {
 659                                 do_update = true;
 660                         }
 661                 }
 662
 663                 if (p.has_ec) {
 664                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 665
 666                         if (fsck_err_on(!m || !m->alive, c,
 667                                         "pointer to nonexistent 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                         if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
 675                                         "pointer does not match stripe %llu\n"
 676                                         "while marking %s",
 677                                         (u64) p.ec.idx,
 678                                         (printbuf_reset(&buf),
 679                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 680                                 do_update = true;
 681                 }
 682         }
 683
 684         if (do_update) {
 685                 struct bkey_ptrs ptrs;
 686                 union bch_extent_entry *entry;
 687                 struct bch_extent_ptr *ptr;
 688                 struct bkey_i *new;
 689
 690                 if (is_root) {
 691                         bch_err(c, "cannot update btree roots yet");
 692                         ret = -EINVAL;
 693                         goto err;
 694                 }
 695
 696                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 697                 if (!new) {
 698                         bch_err_msg(c, ret, "allocating new key");
 699                         ret = -BCH_ERR_ENOMEM_gc_repair_key;
 700                         goto err;
 701                 }
 702
 703                 bkey_reassemble(new, *k);
 704
 705                 if (level) {
 706                         /*
 707                          * We don't want to drop btree node pointers - if the
 708                          * btree node isn't there anymore, the read path will
 709                          * sort it out:
 710                          */
 711                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 712                         bkey_for_each_ptr(ptrs, ptr) {
 713                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 714                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 715
 716                                 ptr->gen = g->gen;
 717                         }
 718                 } else {
 719                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 720                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 721                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 722                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 723
 724                                 (ptr->cached &&
 725                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 726                                 (!ptr->cached &&
 727                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 728                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 729                                 (g->data_type &&
 730                                  g->data_type != data_type);
 731                         }));
 732 again:
 733                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 734                         bkey_extent_entry_for_each(ptrs, entry) {
 735                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 736                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 737                                                                         entry->stripe_ptr.idx);
 738                                         union bch_extent_entry *next_ptr;
 739
 740                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 741                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 742                                                         goto found;
 743                                         next_ptr = NULL;
 744 found:
 745                                         if (!next_ptr) {
 746                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 747                                                 continue;
 748                                         }
 749
 750                                         if (!m || !m->alive ||
 751                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 752                                                                        &next_ptr->ptr,
 753                                                                        m->sectors)) {
 754                                                 bch2_bkey_extent_entry_drop(new, entry);
 755                                                 goto again;
 756                                         }
 757                                 }
 758                         }
 759                 }
 760
 761                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 762                 if (ret) {
 763                         kfree(new);
 764                         goto err;
 765                 }
 766
 767                 if (level)
 768                         bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
 769
 770                 if (0) {
 771                         printbuf_reset(&buf);
 772                         bch2_bkey_val_to_text(&buf, c, *k);
 773                         bch_info(c, "updated %s", buf.buf);
 774
 775                         printbuf_reset(&buf);
 776                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 777                         bch_info(c, "new key %s", buf.buf);
 778                 }
 779
 780                 *k = bkey_i_to_s_c(new);
 781         }
 782 err:
 783 fsck_err:
 784         printbuf_exit(&buf);
 785         return ret;
 786 }
 787
 788 /* marking of btree keys/nodes: */
 789
 790 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 791                             unsigned level, bool is_root,
 792                             struct bkey_s_c *k,
 793                             bool initial)
 794 {
 795         struct bch_fs *c = trans->c;
 796         struct bkey deleted = KEY(0, 0, 0);
 797         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 798         unsigned flags =
 799                 BTREE_TRIGGER_GC|
 800                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 801         int ret = 0;
 802
 803         deleted.p = k->k->p;
 804
 805         if (initial) {
 806                 BUG_ON(bch2_journal_seq_verify &&
 807                        k->k->version.lo > atomic64_read(&c->journal.seq));
 808
 809                 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
 810                 if (ret)
 811                         goto err;
 812
 813                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 814                                 "key version number higher than recorded: %llu > %llu",
 815                                 k->k->version.lo,
 816                                 atomic64_read(&c->key_version)))
 817                         atomic64_set(&c->key_version, k->k->version.lo);
 818         }
 819
 820         ret = commit_do(trans, NULL, NULL, 0,
 821                         bch2_mark_key(trans, btree_id, level, old, *k, flags));
 822 fsck_err:
 823 err:
 824         if (ret)
 825                 bch_err_fn(c, ret);
 826         return ret;
 827 }
 828
 829 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 830 {
 831         struct bch_fs *c = trans->c;
 832         struct btree_node_iter iter;
 833         struct bkey unpacked;
 834         struct bkey_s_c k;
 835         struct bkey_buf prev, cur;
 836         int ret = 0;
 837
 838         if (!btree_node_type_needs_gc(btree_node_type(b)))
 839                 return 0;
 840
 841         bch2_btree_node_iter_init_from_start(&iter, b);
 842         bch2_bkey_buf_init(&prev);
 843         bch2_bkey_buf_init(&cur);
 844         bkey_init(&prev.k->k);
 845
 846         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 847                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 848                                        &k, initial);
 849                 if (ret)
 850                         break;
 851
 852                 bch2_btree_node_iter_advance(&iter, b);
 853
 854                 if (b->c.level) {
 855                         bch2_bkey_buf_reassemble(&cur, c, k);
 856
 857                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 858                                         bch2_btree_node_iter_end(&iter));
 859                         if (ret)
 860                                 break;
 861                 }
 862         }
 863
 864         bch2_bkey_buf_exit(&cur, c);
 865         bch2_bkey_buf_exit(&prev, c);
 866         return ret;
 867 }
 868
 869 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 870                          bool initial, bool metadata_only)
 871 {
 872         struct bch_fs *c = trans->c;
 873         struct btree_iter iter;
 874         struct btree *b;
 875         unsigned depth = metadata_only ? 1 : 0;
 876         int ret = 0;
 877
 878         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 879
 880         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 881                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 882                 bch2_verify_btree_nr_keys(b);
 883
 884                 gc_pos_set(c, gc_pos_btree_node(b));
 885
 886                 ret = btree_gc_mark_node(trans, b, initial);
 887                 if (ret)
 888                         break;
 889         }
 890         bch2_trans_iter_exit(trans, &iter);
 891
 892         if (ret)
 893                 return ret;
 894
 895         mutex_lock(&c->btree_root_lock);
 896         b = bch2_btree_id_root(c, btree_id)->b;
 897         if (!btree_node_fake(b)) {
 898                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 899
 900                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
 901                                        true, &k, initial);
 902         }
 903         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 904         mutex_unlock(&c->btree_root_lock);
 905
 906         return ret;
 907 }
 908
 909 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 910                                       unsigned target_depth)
 911 {
 912         struct bch_fs *c = trans->c;
 913         struct btree_and_journal_iter iter;
 914         struct bkey_s_c k;
 915         struct bkey_buf cur, prev;
 916         struct printbuf buf = PRINTBUF;
 917         int ret = 0;
 918
 919         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 920         bch2_bkey_buf_init(&prev);
 921         bch2_bkey_buf_init(&cur);
 922         bkey_init(&prev.k->k);
 923
 924         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 925                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 926                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 927
 928                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 929                                        false, &k, true);
 930                 if (ret)
 931                         goto fsck_err;
 932
 933                 if (b->c.level) {
 934                         bch2_bkey_buf_reassemble(&cur, c, k);
 935                         k = bkey_i_to_s_c(cur.k);
 936
 937                         bch2_btree_and_journal_iter_advance(&iter);
 938
 939                         ret = bch2_gc_check_topology(c, b,
 940                                         &prev, cur,
 941                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 942                         if (ret)
 943                                 goto fsck_err;
 944                 } else {
 945                         bch2_btree_and_journal_iter_advance(&iter);
 946                 }
 947         }
 948
 949         if (b->c.level > target_depth) {
 950                 bch2_btree_and_journal_iter_exit(&iter);
 951                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 952
 953                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 954                         struct btree *child;
 955
 956                         bch2_bkey_buf_reassemble(&cur, c, k);
 957                         bch2_btree_and_journal_iter_advance(&iter);
 958
 959                         child = bch2_btree_node_get_noiter(trans, cur.k,
 960                                                 b->c.btree_id, b->c.level - 1,
 961                                                 false);
 962                         ret = PTR_ERR_OR_ZERO(child);
 963
 964                         if (ret == -EIO) {
 965                                 bch2_topology_error(c);
 966
 967                                 if (__fsck_err(c,
 968                                           FSCK_CAN_FIX|
 969                                           FSCK_CAN_IGNORE|
 970                                           FSCK_NO_RATELIMIT,
 971                                           "Unreadable btree node at btree %s level %u:\n"
 972                                           "  %s",
 973                                           bch2_btree_id_str(b->c.btree_id),
 974                                           b->c.level - 1,
 975                                           (printbuf_reset(&buf),
 976                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 977                                     should_restart_for_topology_repair(c)) {
 978                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 979                                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 980                                         goto fsck_err;
 981                                 } else {
 982                                         /* Continue marking when opted to not
 983                                          * fix the error: */
 984                                         ret = 0;
 985                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 986                                         continue;
 987                                 }
 988                         } else if (ret) {
 989                                 bch_err_msg(c, ret, "getting btree node");
 990                                 break;
 991                         }
 992
 993                         ret = bch2_gc_btree_init_recurse(trans, child,
 994                                                          target_depth);
 995                         six_unlock_read(&child->c.lock);
 996
 997                         if (ret)
 998                                 break;
 999                 }
1000         }
1001 fsck_err:
1002         bch2_bkey_buf_exit(&cur, c);
1003         bch2_bkey_buf_exit(&prev, c);
1004         bch2_btree_and_journal_iter_exit(&iter);
1005         printbuf_exit(&buf);
1006         return ret;
1007 }
1008
1009 static int bch2_gc_btree_init(struct btree_trans *trans,
1010                               enum btree_id btree_id,
1011                               bool metadata_only)
1012 {
1013         struct bch_fs *c = trans->c;
1014         struct btree *b;
1015         unsigned target_depth = metadata_only ? 1 : 0;
1016         struct printbuf buf = PRINTBUF;
1017         int ret = 0;
1018
1019         b = bch2_btree_id_root(c, btree_id)->b;
1020
1021         if (btree_node_fake(b))
1022                 return 0;
1023
1024         six_lock_read(&b->c.lock, NULL, NULL);
1025         printbuf_reset(&buf);
1026         bch2_bpos_to_text(&buf, b->data->min_key);
1027         if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1028                         "btree root with incorrect min_key: %s", buf.buf)) {
1029                 bch_err(c, "repair unimplemented");
1030                 ret = -BCH_ERR_fsck_repair_unimplemented;
1031                 goto fsck_err;
1032         }
1033
1034         printbuf_reset(&buf);
1035         bch2_bpos_to_text(&buf, b->data->max_key);
1036         if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1037                         "btree root with incorrect max_key: %s", buf.buf)) {
1038                 bch_err(c, "repair unimplemented");
1039                 ret = -BCH_ERR_fsck_repair_unimplemented;
1040                 goto fsck_err;
1041         }
1042
1043         if (b->c.level >= target_depth)
1044                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1045
1046         if (!ret) {
1047                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1048
1049                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1050                                        &k, true);
1051         }
1052 fsck_err:
1053         six_unlock_read(&b->c.lock);
1054
1055         if (ret < 0)
1056                 bch_err_fn(c, ret);
1057         printbuf_exit(&buf);
1058         return ret;
1059 }
1060
1061 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1062 {
1063         return  (int) btree_id_to_gc_phase(l) -
1064                 (int) btree_id_to_gc_phase(r);
1065 }
1066
1067 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1068 {
1069         struct btree_trans *trans = bch2_trans_get(c);
1070         enum btree_id ids[BTREE_ID_NR];
1071         unsigned i;
1072         int ret = 0;
1073
1074         if (initial)
1075                 trans->is_initial_gc = true;
1076
1077         for (i = 0; i < BTREE_ID_NR; i++)
1078                 ids[i] = i;
1079         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1080
1081         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1082                 ret = initial
1083                         ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1084                         : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1085
1086         for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1087                 if (!bch2_btree_id_root(c, i)->alive)
1088                         continue;
1089
1090                 ret = initial
1091                         ? bch2_gc_btree_init(trans, i, metadata_only)
1092                         : bch2_gc_btree(trans, i, initial, metadata_only);
1093         }
1094
1095         if (ret < 0)
1096                 bch_err_fn(c, ret);
1097
1098         bch2_trans_put(trans);
1099         return ret;
1100 }
1101
1102 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1103                                   u64 start, u64 end,
1104                                   enum bch_data_type type,
1105                                   unsigned flags)
1106 {
1107         u64 b = sector_to_bucket(ca, start);
1108
1109         do {
1110                 unsigned sectors =
1111                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1112
1113                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1114                                           gc_phase(GC_PHASE_SB), flags);
1115                 b++;
1116                 start += sectors;
1117         } while (start < end);
1118 }
1119
1120 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1121                                      unsigned flags)
1122 {
1123         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1124         unsigned i;
1125         u64 b;
1126
1127         for (i = 0; i < layout->nr_superblocks; i++) {
1128                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1129
1130                 if (offset == BCH_SB_SECTOR)
1131                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1132                                               BCH_DATA_sb, flags);
1133
1134                 mark_metadata_sectors(c, ca, offset,
1135                                       offset + (1 << layout->sb_max_size_bits),
1136                                       BCH_DATA_sb, flags);
1137         }
1138
1139         for (i = 0; i < ca->journal.nr; i++) {
1140                 b = ca->journal.buckets[i];
1141                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1142                                           ca->mi.bucket_size,
1143                                           gc_phase(GC_PHASE_SB), flags);
1144         }
1145 }
1146
1147 static void bch2_mark_superblocks(struct bch_fs *c)
1148 {
1149         struct bch_dev *ca;
1150         unsigned i;
1151
1152         mutex_lock(&c->sb_lock);
1153         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1154
1155         for_each_online_member(ca, c, i)
1156                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1157         mutex_unlock(&c->sb_lock);
1158 }
1159
1160 #if 0
1161 /* Also see bch2_pending_btree_node_free_insert_done() */
1162 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1163 {
1164         struct btree_update *as;
1165         struct pending_btree_node_free *d;
1166
1167         mutex_lock(&c->btree_interior_update_lock);
1168         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1169
1170         for_each_pending_btree_node_free(c, as, d)
1171                 if (d->index_update_done)
1172                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1173
1174         mutex_unlock(&c->btree_interior_update_lock);
1175 }
1176 #endif
1177
1178 static void bch2_gc_free(struct bch_fs *c)
1179 {
1180         struct bch_dev *ca;
1181         unsigned i;
1182
1183         genradix_free(&c->reflink_gc_table);
1184         genradix_free(&c->gc_stripes);
1185
1186         for_each_member_device(ca, c, i) {
1187                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1188                         sizeof(struct bucket_array) +
1189                         ca->mi.nbuckets * sizeof(struct bucket));
1190                 ca->buckets_gc = NULL;
1191
1192                 free_percpu(ca->usage_gc);
1193                 ca->usage_gc = NULL;
1194         }
1195
1196         free_percpu(c->usage_gc);
1197         c->usage_gc = NULL;
1198 }
1199
1200 static int bch2_gc_done(struct bch_fs *c,
1201                         bool initial, bool metadata_only)
1202 {
1203         struct bch_dev *ca = NULL;
1204         struct printbuf buf = PRINTBUF;
1205         bool verify = !metadata_only &&
1206                 !c->opts.reconstruct_alloc &&
1207                 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1208         unsigned i, dev;
1209         int ret = 0;
1210
1211         percpu_down_write(&c->mark_lock);
1212
1213 #define copy_field(_f, _msg, ...)                                       \
1214         if (dst->_f != src->_f &&                                       \
1215             (!verify ||                                                 \
1216              fsck_err(c, _msg ": got %llu, should be %llu"              \
1217                       , ##__VA_ARGS__, dst->_f, src->_f)))              \
1218                 dst->_f = src->_f
1219 #define copy_dev_field(_f, _msg, ...)                                   \
1220         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1221 #define copy_fs_field(_f, _msg, ...)                                    \
1222         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1223
1224         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1225                 bch2_fs_usage_acc_to_base(c, i);
1226
1227         for_each_member_device(ca, c, dev) {
1228                 struct bch_dev_usage *dst = ca->usage_base;
1229                 struct bch_dev_usage *src = (void *)
1230                         bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1231                                              dev_usage_u64s());
1232
1233                 copy_dev_field(buckets_ec,              "buckets_ec");
1234
1235                 for (i = 0; i < BCH_DATA_NR; i++) {
1236                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1237                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1238                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1239                 }
1240         }
1241
1242         {
1243                 unsigned nr = fs_usage_u64s(c);
1244                 struct bch_fs_usage *dst = c->usage_base;
1245                 struct bch_fs_usage *src = (void *)
1246                         bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1247
1248                 copy_fs_field(hidden,           "hidden");
1249                 copy_fs_field(btree,            "btree");
1250
1251                 if (!metadata_only) {
1252                         copy_fs_field(data,     "data");
1253                         copy_fs_field(cached,   "cached");
1254                         copy_fs_field(reserved, "reserved");
1255                         copy_fs_field(nr_inodes,"nr_inodes");
1256
1257                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1258                                 copy_fs_field(persistent_reserved[i],
1259                                               "persistent_reserved[%i]", i);
1260                 }
1261
1262                 for (i = 0; i < c->replicas.nr; i++) {
1263                         struct bch_replicas_entry *e =
1264                                 cpu_replicas_entry(&c->replicas, i);
1265
1266                         if (metadata_only &&
1267                             (e->data_type == BCH_DATA_user ||
1268                              e->data_type == BCH_DATA_cached))
1269                                 continue;
1270
1271                         printbuf_reset(&buf);
1272                         bch2_replicas_entry_to_text(&buf, e);
1273
1274                         copy_fs_field(replicas[i], "%s", buf.buf);
1275                 }
1276         }
1277
1278 #undef copy_fs_field
1279 #undef copy_dev_field
1280 #undef copy_stripe_field
1281 #undef copy_field
1282 fsck_err:
1283         if (ca)
1284                 percpu_ref_put(&ca->ref);
1285         if (ret)
1286                 bch_err_fn(c, ret);
1287
1288         percpu_up_write(&c->mark_lock);
1289         printbuf_exit(&buf);
1290         return ret;
1291 }
1292
1293 static int bch2_gc_start(struct bch_fs *c)
1294 {
1295         struct bch_dev *ca = NULL;
1296         unsigned i;
1297
1298         BUG_ON(c->usage_gc);
1299
1300         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1301                                          sizeof(u64), GFP_KERNEL);
1302         if (!c->usage_gc) {
1303                 bch_err(c, "error allocating c->usage_gc");
1304                 return -BCH_ERR_ENOMEM_gc_start;
1305         }
1306
1307         for_each_member_device(ca, c, i) {
1308                 BUG_ON(ca->usage_gc);
1309
1310                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1311                 if (!ca->usage_gc) {
1312                         bch_err(c, "error allocating ca->usage_gc");
1313                         percpu_ref_put(&ca->ref);
1314                         return -BCH_ERR_ENOMEM_gc_start;
1315                 }
1316
1317                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1318                                ca->mi.nbuckets - ca->mi.first_bucket);
1319         }
1320
1321         return 0;
1322 }
1323
1324 static int bch2_gc_reset(struct bch_fs *c)
1325 {
1326         struct bch_dev *ca;
1327         unsigned i;
1328
1329         for_each_member_device(ca, c, i) {
1330                 free_percpu(ca->usage_gc);
1331                 ca->usage_gc = NULL;
1332         }
1333
1334         free_percpu(c->usage_gc);
1335         c->usage_gc = NULL;
1336
1337         return bch2_gc_start(c);
1338 }
1339
1340 /* returns true if not equal */
1341 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1342                                      struct bch_alloc_v4 r)
1343 {
1344         return  l.gen != r.gen                          ||
1345                 l.oldest_gen != r.oldest_gen            ||
1346                 l.data_type != r.data_type              ||
1347                 l.dirty_sectors != r.dirty_sectors      ||
1348                 l.cached_sectors != r.cached_sectors     ||
1349                 l.stripe_redundancy != r.stripe_redundancy ||
1350                 l.stripe != r.stripe;
1351 }
1352
1353 static int bch2_alloc_write_key(struct btree_trans *trans,
1354                                 struct btree_iter *iter,
1355                                 struct bkey_s_c k,
1356                                 bool metadata_only)
1357 {
1358         struct bch_fs *c = trans->c;
1359         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1360         struct bucket gc, *b;
1361         struct bkey_i_alloc_v4 *a;
1362         struct bch_alloc_v4 old_convert, new;
1363         const struct bch_alloc_v4 *old;
1364         enum bch_data_type type;
1365         int ret;
1366
1367         if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
1368                 return 1;
1369
1370         old = bch2_alloc_to_v4(k, &old_convert);
1371         new = *old;
1372
1373         percpu_down_read(&c->mark_lock);
1374         b = gc_bucket(ca, iter->pos.offset);
1375
1376         /*
1377          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1378          * fix that here:
1379          */
1380         type = __alloc_data_type(b->dirty_sectors,
1381                                  b->cached_sectors,
1382                                  b->stripe,
1383                                  *old,
1384                                  b->data_type);
1385         if (b->data_type != type) {
1386                 struct bch_dev_usage *u;
1387
1388                 preempt_disable();
1389                 u = this_cpu_ptr(ca->usage_gc);
1390                 u->d[b->data_type].buckets--;
1391                 b->data_type = type;
1392                 u->d[b->data_type].buckets++;
1393                 preempt_enable();
1394         }
1395
1396         gc = *b;
1397         percpu_up_read(&c->mark_lock);
1398
1399         if (metadata_only &&
1400             gc.data_type != BCH_DATA_sb &&
1401             gc.data_type != BCH_DATA_journal &&
1402             gc.data_type != BCH_DATA_btree)
1403                 return 0;
1404
1405         if (gen_after(old->gen, gc.gen))
1406                 return 0;
1407
1408         if (c->opts.reconstruct_alloc ||
1409             fsck_err_on(new.data_type != gc.data_type, c,
1410                         "bucket %llu:%llu gen %u has wrong data_type"
1411                         ": got %s, should be %s",
1412                         iter->pos.inode, iter->pos.offset,
1413                         gc.gen,
1414                         bch2_data_types[new.data_type],
1415                         bch2_data_types[gc.data_type]))
1416                 new.data_type = gc.data_type;
1417
1418 #define copy_bucket_field(_f)                                           \
1419         if (c->opts.reconstruct_alloc ||                                \
1420             fsck_err_on(new._f != gc._f, c,                             \
1421                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1422                         ": got %u, should be %u",                       \
1423                         iter->pos.inode, iter->pos.offset,              \
1424                         gc.gen,                                         \
1425                         bch2_data_types[gc.data_type],                  \
1426                         new._f, gc._f))                                 \
1427                 new._f = gc._f;                                         \
1428
1429         copy_bucket_field(gen);
1430         copy_bucket_field(dirty_sectors);
1431         copy_bucket_field(cached_sectors);
1432         copy_bucket_field(stripe_redundancy);
1433         copy_bucket_field(stripe);
1434 #undef copy_bucket_field
1435
1436         if (!bch2_alloc_v4_cmp(*old, new))
1437                 return 0;
1438
1439         a = bch2_alloc_to_v4_mut(trans, k);
1440         ret = PTR_ERR_OR_ZERO(a);
1441         if (ret)
1442                 return ret;
1443
1444         a->v = new;
1445
1446         /*
1447          * The trigger normally makes sure this is set, but we're not running
1448          * triggers:
1449          */
1450         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1451                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1452
1453         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1454 fsck_err:
1455         return ret;
1456 }
1457
1458 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1459 {
1460         struct btree_trans *trans = bch2_trans_get(c);
1461         struct btree_iter iter;
1462         struct bkey_s_c k;
1463         struct bch_dev *ca;
1464         unsigned i;
1465         int ret = 0;
1466
1467         for_each_member_device(ca, c, i) {
1468                 ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1469                                 POS(ca->dev_idx, ca->mi.first_bucket),
1470                                 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1471                                 NULL, NULL, BTREE_INSERT_LAZY_RW,
1472                         bch2_alloc_write_key(trans, &iter, k, metadata_only));
1473
1474                 if (ret < 0) {
1475                         bch_err_fn(c, ret);
1476                         percpu_ref_put(&ca->ref);
1477                         break;
1478                 }
1479         }
1480
1481         bch2_trans_put(trans);
1482         return ret < 0 ? ret : 0;
1483 }
1484
1485 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1486 {
1487         struct bch_dev *ca;
1488         struct btree_trans *trans = bch2_trans_get(c);
1489         struct btree_iter iter;
1490         struct bkey_s_c k;
1491         struct bucket *g;
1492         struct bch_alloc_v4 a_convert;
1493         const struct bch_alloc_v4 *a;
1494         unsigned i;
1495         int ret;
1496
1497         for_each_member_device(ca, c, i) {
1498                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1499                                 ca->mi.nbuckets * sizeof(struct bucket),
1500                                 GFP_KERNEL|__GFP_ZERO);
1501                 if (!buckets) {
1502                         percpu_ref_put(&ca->ref);
1503                         bch_err(c, "error allocating ca->buckets[gc]");
1504                         ret = -BCH_ERR_ENOMEM_gc_alloc_start;
1505                         goto err;
1506                 }
1507
1508                 buckets->first_bucket   = ca->mi.first_bucket;
1509                 buckets->nbuckets       = ca->mi.nbuckets;
1510                 rcu_assign_pointer(ca->buckets_gc, buckets);
1511         }
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 err:
1536         bch2_trans_put(trans);
1537         if (ret)
1538                 bch_err_fn(c, ret);
1539         return ret;
1540 }
1541
1542 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1543 {
1544         struct bch_dev *ca;
1545         unsigned i;
1546
1547         for_each_member_device(ca, c, i) {
1548                 struct bucket_array *buckets = gc_bucket_array(ca);
1549                 struct bucket *g;
1550
1551                 for_each_bucket(g, buckets) {
1552                         if (metadata_only &&
1553                             (g->data_type == BCH_DATA_user ||
1554                              g->data_type == BCH_DATA_cached ||
1555                              g->data_type == BCH_DATA_parity))
1556                                 continue;
1557                         g->data_type = 0;
1558                         g->dirty_sectors = 0;
1559                         g->cached_sectors = 0;
1560                 }
1561         }
1562 }
1563
1564 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1565                                      struct btree_iter *iter,
1566                                      struct bkey_s_c k,
1567                                      size_t *idx)
1568 {
1569         struct bch_fs *c = trans->c;
1570         const __le64 *refcount = bkey_refcount_c(k);
1571         struct printbuf buf = PRINTBUF;
1572         struct reflink_gc *r;
1573         int ret = 0;
1574
1575         if (!refcount)
1576                 return 0;
1577
1578         while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1579                r->offset < k.k->p.offset)
1580                 ++*idx;
1581
1582         if (!r ||
1583             r->offset != k.k->p.offset ||
1584             r->size != k.k->size) {
1585                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1586                 return -EINVAL;
1587         }
1588
1589         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1590                         "reflink key has wrong refcount:\n"
1591                         "  %s\n"
1592                         "  should be %u",
1593                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1594                         r->refcount)) {
1595                 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1596
1597                 ret = PTR_ERR_OR_ZERO(new);
1598                 if (ret)
1599                         return ret;
1600
1601                 if (!r->refcount)
1602                         new->k.type = KEY_TYPE_deleted;
1603                 else
1604                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1605         }
1606 fsck_err:
1607         printbuf_exit(&buf);
1608         return ret;
1609 }
1610
1611 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1612 {
1613         struct btree_trans *trans;
1614         struct btree_iter iter;
1615         struct bkey_s_c k;
1616         size_t idx = 0;
1617         int ret = 0;
1618
1619         if (metadata_only)
1620                 return 0;
1621
1622         trans = bch2_trans_get(c);
1623
1624         ret = for_each_btree_key_commit(trans, iter,
1625                         BTREE_ID_reflink, POS_MIN,
1626                         BTREE_ITER_PREFETCH, k,
1627                         NULL, NULL, BTREE_INSERT_NOFAIL,
1628                 bch2_gc_write_reflink_key(trans, &iter, k, &idx));
1629
1630         c->reflink_gc_nr = 0;
1631         bch2_trans_put(trans);
1632         return ret;
1633 }
1634
1635 static int bch2_gc_reflink_start(struct bch_fs *c,
1636                                  bool metadata_only)
1637 {
1638         struct btree_trans *trans;
1639         struct btree_iter iter;
1640         struct bkey_s_c k;
1641         struct reflink_gc *r;
1642         int ret = 0;
1643
1644         if (metadata_only)
1645                 return 0;
1646
1647         trans = bch2_trans_get(c);
1648         c->reflink_gc_nr = 0;
1649
1650         for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1651                            BTREE_ITER_PREFETCH, k, ret) {
1652                 const __le64 *refcount = bkey_refcount_c(k);
1653
1654                 if (!refcount)
1655                         continue;
1656
1657                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1658                                        GFP_KERNEL);
1659                 if (!r) {
1660                         ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1661                         break;
1662                 }
1663
1664                 r->offset       = k.k->p.offset;
1665                 r->size         = k.k->size;
1666                 r->refcount     = 0;
1667         }
1668         bch2_trans_iter_exit(trans, &iter);
1669
1670         bch2_trans_put(trans);
1671         return ret;
1672 }
1673
1674 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1675 {
1676         struct genradix_iter iter;
1677         struct reflink_gc *r;
1678
1679         genradix_for_each(&c->reflink_gc_table, iter, r)
1680                 r->refcount = 0;
1681 }
1682
1683 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1684                                      struct btree_iter *iter,
1685                                      struct bkey_s_c k)
1686 {
1687         struct bch_fs *c = trans->c;
1688         struct printbuf buf = PRINTBUF;
1689         const struct bch_stripe *s;
1690         struct gc_stripe *m;
1691         bool bad = false;
1692         unsigned i;
1693         int ret = 0;
1694
1695         if (k.k->type != KEY_TYPE_stripe)
1696                 return 0;
1697
1698         s = bkey_s_c_to_stripe(k).v;
1699         m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1700
1701         for (i = 0; i < s->nr_blocks; i++) {
1702                 u32 old = stripe_blockcount_get(s, i);
1703                 u32 new = (m ? m->block_sectors[i] : 0);
1704
1705                 if (old != new) {
1706                         prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1707                                    i, old, new);
1708                         bad = true;
1709                 }
1710         }
1711
1712         if (bad)
1713                 bch2_bkey_val_to_text(&buf, c, k);
1714
1715         if (fsck_err_on(bad, c, "%s", buf.buf)) {
1716                 struct bkey_i_stripe *new;
1717
1718                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1719                 ret = PTR_ERR_OR_ZERO(new);
1720                 if (ret)
1721                         return ret;
1722
1723                 bkey_reassemble(&new->k_i, k);
1724
1725                 for (i = 0; i < new->v.nr_blocks; i++)
1726                         stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1727
1728                 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1729         }
1730 fsck_err:
1731         printbuf_exit(&buf);
1732         return ret;
1733 }
1734
1735 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1736 {
1737         struct btree_trans *trans;
1738         struct btree_iter iter;
1739         struct bkey_s_c k;
1740         int ret = 0;
1741
1742         if (metadata_only)
1743                 return 0;
1744
1745         trans = bch2_trans_get(c);
1746
1747         ret = for_each_btree_key_commit(trans, iter,
1748                         BTREE_ID_stripes, POS_MIN,
1749                         BTREE_ITER_PREFETCH, k,
1750                         NULL, NULL, BTREE_INSERT_NOFAIL,
1751                 bch2_gc_write_stripes_key(trans, &iter, k));
1752
1753         bch2_trans_put(trans);
1754         return ret;
1755 }
1756
1757 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1758 {
1759         genradix_free(&c->gc_stripes);
1760 }
1761
1762 /**
1763  * bch2_gc - walk _all_ references to buckets, and recompute them:
1764  *
1765  * @c:                  filesystem object
1766  * @initial:            are we in recovery?
1767  * @metadata_only:      are we just checking metadata references, or everything?
1768  *
1769  * Returns: 0 on success, or standard errcode on failure
1770  *
1771  * Order matters here:
1772  *  - Concurrent GC relies on the fact that we have a total ordering for
1773  *    everything that GC walks - see  gc_will_visit_node(),
1774  *    gc_will_visit_root()
1775  *
1776  *  - also, references move around in the course of index updates and
1777  *    various other crap: everything needs to agree on the ordering
1778  *    references are allowed to move around in - e.g., we're allowed to
1779  *    start with a reference owned by an open_bucket (the allocator) and
1780  *    move it to the btree, but not the reverse.
1781  *
1782  *    This is necessary to ensure that gc doesn't miss references that
1783  *    move around - if references move backwards in the ordering GC
1784  *    uses, GC could skip past them
1785  */
1786 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1787 {
1788         unsigned iter = 0;
1789         int ret;
1790
1791         lockdep_assert_held(&c->state_lock);
1792
1793         down_write(&c->gc_lock);
1794
1795         bch2_btree_interior_updates_flush(c);
1796
1797         ret   = bch2_gc_start(c) ?:
1798                 bch2_gc_alloc_start(c, metadata_only) ?:
1799                 bch2_gc_reflink_start(c, metadata_only);
1800         if (ret)
1801                 goto out;
1802 again:
1803         gc_pos_set(c, gc_phase(GC_PHASE_START));
1804
1805         bch2_mark_superblocks(c);
1806
1807         ret = bch2_gc_btrees(c, initial, metadata_only);
1808
1809         if (ret)
1810                 goto out;
1811
1812 #if 0
1813         bch2_mark_pending_btree_node_frees(c);
1814 #endif
1815         c->gc_count++;
1816
1817         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1818             (!iter && bch2_test_restart_gc)) {
1819                 if (iter++ > 2) {
1820                         bch_info(c, "Unable to fix bucket gens, looping");
1821                         ret = -EINVAL;
1822                         goto out;
1823                 }
1824
1825                 /*
1826                  * XXX: make sure gens we fixed got saved
1827                  */
1828                 bch_info(c, "Second GC pass needed, restarting:");
1829                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1830                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1831
1832                 bch2_gc_stripes_reset(c, metadata_only);
1833                 bch2_gc_alloc_reset(c, metadata_only);
1834                 bch2_gc_reflink_reset(c, metadata_only);
1835                 ret = bch2_gc_reset(c);
1836                 if (ret)
1837                         goto out;
1838
1839                 /* flush fsck errors, reset counters */
1840                 bch2_flush_fsck_errs(c);
1841                 goto again;
1842         }
1843 out:
1844         if (!ret) {
1845                 bch2_journal_block(&c->journal);
1846
1847                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1848                         bch2_gc_reflink_done(c, metadata_only) ?:
1849                         bch2_gc_alloc_done(c, metadata_only) ?:
1850                         bch2_gc_done(c, initial, metadata_only);
1851
1852                 bch2_journal_unblock(&c->journal);
1853         }
1854
1855         percpu_down_write(&c->mark_lock);
1856         /* Indicates that gc is no longer in progress: */
1857         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1858
1859         bch2_gc_free(c);
1860         percpu_up_write(&c->mark_lock);
1861
1862         up_write(&c->gc_lock);
1863
1864         /*
1865          * At startup, allocations can happen directly instead of via the
1866          * allocator thread - issue wakeup in case they blocked on gc_lock:
1867          */
1868         closure_wake_up(&c->freelist_wait);
1869
1870         if (ret)
1871                 bch_err_fn(c, ret);
1872         return ret;
1873 }
1874
1875 static int gc_btree_gens_key(struct btree_trans *trans,
1876                              struct btree_iter *iter,
1877                              struct bkey_s_c k)
1878 {
1879         struct bch_fs *c = trans->c;
1880         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1881         const struct bch_extent_ptr *ptr;
1882         struct bkey_i *u;
1883         int ret;
1884
1885         percpu_down_read(&c->mark_lock);
1886         bkey_for_each_ptr(ptrs, ptr) {
1887                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1888
1889                 if (ptr_stale(ca, ptr) > 16) {
1890                         percpu_up_read(&c->mark_lock);
1891                         goto update;
1892                 }
1893         }
1894
1895         bkey_for_each_ptr(ptrs, ptr) {
1896                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1897                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1898
1899                 if (gen_after(*gen, ptr->gen))
1900                         *gen = ptr->gen;
1901         }
1902         percpu_up_read(&c->mark_lock);
1903         return 0;
1904 update:
1905         u = bch2_bkey_make_mut(trans, iter, &k, 0);
1906         ret = PTR_ERR_OR_ZERO(u);
1907         if (ret)
1908                 return ret;
1909
1910         bch2_extent_normalize(c, bkey_i_to_s(u));
1911         return 0;
1912 }
1913
1914 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1915                                        struct bkey_s_c k)
1916 {
1917         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1918         struct bch_alloc_v4 a_convert;
1919         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1920         struct bkey_i_alloc_v4 *a_mut;
1921         int ret;
1922
1923         if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1924                 return 0;
1925
1926         a_mut = bch2_alloc_to_v4_mut(trans, k);
1927         ret = PTR_ERR_OR_ZERO(a_mut);
1928         if (ret)
1929                 return ret;
1930
1931         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1932         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1933
1934         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1935 }
1936
1937 int bch2_gc_gens(struct bch_fs *c)
1938 {
1939         struct btree_trans *trans;
1940         struct btree_iter iter;
1941         struct bkey_s_c k;
1942         struct bch_dev *ca;
1943         u64 b, start_time = local_clock();
1944         unsigned i;
1945         int ret;
1946
1947         /*
1948          * Ideally we would be using state_lock and not gc_lock here, but that
1949          * introduces a deadlock in the RO path - we currently take the state
1950          * lock at the start of going RO, thus the gc thread may get stuck:
1951          */
1952         if (!mutex_trylock(&c->gc_gens_lock))
1953                 return 0;
1954
1955         trace_and_count(c, gc_gens_start, c);
1956         down_read(&c->gc_lock);
1957         trans = bch2_trans_get(c);
1958
1959         for_each_member_device(ca, c, i) {
1960                 struct bucket_gens *gens;
1961
1962                 BUG_ON(ca->oldest_gen);
1963
1964                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1965                 if (!ca->oldest_gen) {
1966                         percpu_ref_put(&ca->ref);
1967                         ret = -BCH_ERR_ENOMEM_gc_gens;
1968                         goto err;
1969                 }
1970
1971                 gens = bucket_gens(ca);
1972
1973                 for (b = gens->first_bucket;
1974                      b < gens->nbuckets; b++)
1975                         ca->oldest_gen[b] = gens->b[b];
1976         }
1977
1978         for (i = 0; i < BTREE_ID_NR; i++)
1979                 if (btree_type_has_ptrs(i)) {
1980                         c->gc_gens_btree = i;
1981                         c->gc_gens_pos = POS_MIN;
1982
1983                         ret = for_each_btree_key_commit(trans, iter, i,
1984                                         POS_MIN,
1985                                         BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1986                                         k,
1987                                         NULL, NULL,
1988                                         BTREE_INSERT_NOFAIL,
1989                                 gc_btree_gens_key(trans, &iter, k));
1990                         if (ret && !bch2_err_matches(ret, EROFS))
1991                                 bch_err_fn(c, ret);
1992                         if (ret)
1993                                 goto err;
1994                 }
1995
1996         ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1997                         POS_MIN,
1998                         BTREE_ITER_PREFETCH,
1999                         k,
2000                         NULL, NULL,
2001                         BTREE_INSERT_NOFAIL,
2002                 bch2_alloc_write_oldest_gen(trans, &iter, k));
2003         if (ret && !bch2_err_matches(ret, EROFS))
2004                 bch_err_fn(c, ret);
2005         if (ret)
2006                 goto err;
2007
2008         c->gc_gens_btree        = 0;
2009         c->gc_gens_pos          = POS_MIN;
2010
2011         c->gc_count++;
2012
2013         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2014         trace_and_count(c, gc_gens_end, c);
2015 err:
2016         for_each_member_device(ca, c, i) {
2017                 kvfree(ca->oldest_gen);
2018                 ca->oldest_gen = NULL;
2019         }
2020
2021         bch2_trans_put(trans);
2022         up_read(&c->gc_lock);
2023         mutex_unlock(&c->gc_gens_lock);
2024         return ret;
2025 }
2026
2027 static int bch2_gc_thread(void *arg)
2028 {
2029         struct bch_fs *c = arg;
2030         struct io_clock *clock = &c->io_clock[WRITE];
2031         unsigned long last = atomic64_read(&clock->now);
2032         unsigned last_kick = atomic_read(&c->kick_gc);
2033         int ret;
2034
2035         set_freezable();
2036
2037         while (1) {
2038                 while (1) {
2039                         set_current_state(TASK_INTERRUPTIBLE);
2040
2041                         if (kthread_should_stop()) {
2042                                 __set_current_state(TASK_RUNNING);
2043                                 return 0;
2044                         }
2045
2046                         if (atomic_read(&c->kick_gc) != last_kick)
2047                                 break;
2048
2049                         if (c->btree_gc_periodic) {
2050                                 unsigned long next = last + c->capacity / 16;
2051
2052                                 if (atomic64_read(&clock->now) >= next)
2053                                         break;
2054
2055                                 bch2_io_clock_schedule_timeout(clock, next);
2056                         } else {
2057                                 schedule();
2058                         }
2059
2060                         try_to_freeze();
2061                 }
2062                 __set_current_state(TASK_RUNNING);
2063
2064                 last = atomic64_read(&clock->now);
2065                 last_kick = atomic_read(&c->kick_gc);
2066
2067                 /*
2068                  * Full gc is currently incompatible with btree key cache:
2069                  */
2070 #if 0
2071                 ret = bch2_gc(c, false, false);
2072 #else
2073                 ret = bch2_gc_gens(c);
2074 #endif
2075                 if (ret < 0)
2076                         bch_err_fn(c, ret);
2077
2078                 debug_check_no_locks_held();
2079         }
2080
2081         return 0;
2082 }
2083
2084 void bch2_gc_thread_stop(struct bch_fs *c)
2085 {
2086         struct task_struct *p;
2087
2088         p = c->gc_thread;
2089         c->gc_thread = NULL;
2090
2091         if (p) {
2092                 kthread_stop(p);
2093                 put_task_struct(p);
2094         }
2095 }
2096
2097 int bch2_gc_thread_start(struct bch_fs *c)
2098 {
2099         struct task_struct *p;
2100
2101         if (c->gc_thread)
2102                 return 0;
2103
2104         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2105         if (IS_ERR(p)) {
2106                 bch_err_fn(c, PTR_ERR(p));
2107                 return PTR_ERR(p);
2108         }
2109
2110         get_task_struct(p);
2111         c->gc_thread = p;
2112         wake_up_process(p);
2113         return 0;
2114 }