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