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_topology_bad_min_key,
 102                                        "btree node with incorrect min_key at btree %s level %u:\n"
 103                                        "  prev %s\n"
 104                                        "  cur %s",
 105                                        bch2_btree_id_str(b->c.btree_id), b->c.level,
 106                                        buf1.buf, buf2.buf) && 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, FSCK_CAN_FIX|FSCK_CAN_IGNORE|FSCK_NO_RATELIMIT,
 126                           btree_node_topology_bad_max_key,
 127                           "btree node with incorrect max_key at btree %s level %u:\n"
 128                           "  %s\n"
 129                           "  expected %s",
 130                           bch2_btree_id_str(b->c.btree_id), b->c.level,
 131                           buf1.buf, buf2.buf) &&
 132                     should_restart_for_topology_repair(c)) {
 133                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 134                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 135                         goto err;
 136                 } else {
 137                         set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
 138                 }
 139         }
 140
 141         bch2_bkey_buf_copy(prev, c, cur.k);
 142 err:
 143 fsck_err:
 144         printbuf_exit(&buf2);
 145         printbuf_exit(&buf1);
 146         return ret;
 147 }
 148
 149 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 150 {
 151         switch (b->key.k.type) {
 152         case KEY_TYPE_btree_ptr: {
 153                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 154
 155                 dst->k.p                = src->k.p;
 156                 dst->v.mem_ptr          = 0;
 157                 dst->v.seq              = b->data->keys.seq;
 158                 dst->v.sectors_written  = 0;
 159                 dst->v.flags            = 0;
 160                 dst->v.min_key          = b->data->min_key;
 161                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 162                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 163                 break;
 164         }
 165         case KEY_TYPE_btree_ptr_v2:
 166                 bkey_copy(&dst->k_i, &b->key);
 167                 break;
 168         default:
 169                 BUG();
 170         }
 171 }
 172
 173 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
 174                                              enum btree_id btree, unsigned level,
 175                                              struct bkey_s_c old, struct bkey_i *new)
 176 {
 177         struct bch_fs *c = trans->c;
 178         struct btree *b;
 179         struct bkey_buf tmp;
 180         int ret;
 181
 182         bch2_bkey_buf_init(&tmp);
 183         bch2_bkey_buf_reassemble(&tmp, c, old);
 184
 185         b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
 186         if (!IS_ERR_OR_NULL(b)) {
 187                 mutex_lock(&c->btree_cache.lock);
 188
 189                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 190
 191                 bkey_copy(&b->key, new);
 192                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 193                 BUG_ON(ret);
 194
 195                 mutex_unlock(&c->btree_cache.lock);
 196                 six_unlock_read(&b->c.lock);
 197         }
 198
 199         bch2_bkey_buf_exit(&tmp, c);
 200 }
 201
 202 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 203 {
 204         struct bkey_i_btree_ptr_v2 *new;
 205         int ret;
 206
 207         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 208         if (!new)
 209                 return -BCH_ERR_ENOMEM_gc_repair_key;
 210
 211         btree_ptr_to_v2(b, new);
 212         b->data->min_key        = new_min;
 213         new->v.min_key          = new_min;
 214         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 215
 216         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 217         if (ret) {
 218                 kfree(new);
 219                 return ret;
 220         }
 221
 222         bch2_btree_node_drop_keys_outside_node(b);
 223         bkey_copy(&b->key, &new->k_i);
 224         return 0;
 225 }
 226
 227 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 228 {
 229         struct bkey_i_btree_ptr_v2 *new;
 230         int ret;
 231
 232         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 233         if (ret)
 234                 return ret;
 235
 236         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 237         if (!new)
 238                 return -BCH_ERR_ENOMEM_gc_repair_key;
 239
 240         btree_ptr_to_v2(b, new);
 241         b->data->max_key        = new_max;
 242         new->k.p                = new_max;
 243         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 244
 245         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 246         if (ret) {
 247                 kfree(new);
 248                 return ret;
 249         }
 250
 251         bch2_btree_node_drop_keys_outside_node(b);
 252
 253         mutex_lock(&c->btree_cache.lock);
 254         bch2_btree_node_hash_remove(&c->btree_cache, b);
 255
 256         bkey_copy(&b->key, &new->k_i);
 257         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 258         BUG_ON(ret);
 259         mutex_unlock(&c->btree_cache.lock);
 260         return 0;
 261 }
 262
 263 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 264                                         struct btree *prev, struct btree *cur)
 265 {
 266         struct bpos expected_start = !prev
 267                 ? b->data->min_key
 268                 : bpos_successor(prev->key.k.p);
 269         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 270         int ret = 0;
 271
 272         if (!prev) {
 273                 prt_printf(&buf1, "start of node: ");
 274                 bch2_bpos_to_text(&buf1, b->data->min_key);
 275         } else {
 276                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
 277         }
 278
 279         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
 280
 281         if (prev &&
 282             bpos_gt(expected_start, cur->data->min_key) &&
 283             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 284                 /* cur overwrites prev: */
 285
 286                 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
 287                                                 cur->data->min_key), c,
 288                                 btree_node_topology_overwritten_by_next_node,
 289                                 "btree node overwritten by next node at btree %s level %u:\n"
 290                                 "  node %s\n"
 291                                 "  next %s",
 292                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 293                                 buf1.buf, buf2.buf)) {
 294                         ret = DROP_PREV_NODE;
 295                         goto out;
 296                 }
 297
 298                 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
 299                                                  bpos_predecessor(cur->data->min_key)), c,
 300                                 btree_node_topology_bad_max_key,
 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_topology_overwritten_by_prev_node,
 314                                 "btree node overwritten by prev node at btree %s level %u:\n"
 315                                 "  prev %s\n"
 316                                 "  node %s",
 317                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 318                                 buf1.buf, buf2.buf)) {
 319                         ret = DROP_THIS_NODE;
 320                         goto out;
 321                 }
 322
 323                 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
 324                                 btree_node_topology_bad_min_key,
 325                                 "btree node with incorrect min_key at btree %s level %u:\n"
 326                                 "  prev %s\n"
 327                                 "  node %s",
 328                                 bch2_btree_id_str(b->c.btree_id), b->c.level,
 329                                 buf1.buf, buf2.buf))
 330                         ret = set_node_min(c, cur, expected_start);
 331         }
 332 out:
 333 fsck_err:
 334         printbuf_exit(&buf2);
 335         printbuf_exit(&buf1);
 336         return ret;
 337 }
 338
 339 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 340                                  struct btree *child)
 341 {
 342         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 343         int ret = 0;
 344
 345         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
 346         bch2_bpos_to_text(&buf2, b->key.k.p);
 347
 348         if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
 349                                 btree_node_topology_bad_max_key,
 350                         "btree node with incorrect max_key at btree %s level %u:\n"
 351                         "  %s\n"
 352                         "  expected %s",
 353                         bch2_btree_id_str(b->c.btree_id), b->c.level,
 354                         buf1.buf, buf2.buf)) {
 355                 ret = set_node_max(c, child, b->key.k.p);
 356                 if (ret)
 357                         goto err;
 358         }
 359 err:
 360 fsck_err:
 361         printbuf_exit(&buf2);
 362         printbuf_exit(&buf1);
 363         return ret;
 364 }
 365
 366 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
 367 {
 368         struct bch_fs *c = trans->c;
 369         struct btree_and_journal_iter iter;
 370         struct bkey_s_c k;
 371         struct bkey_buf prev_k, cur_k;
 372         struct btree *prev = NULL, *cur = NULL;
 373         bool have_child, dropped_children = false;
 374         struct printbuf buf = PRINTBUF;
 375         int ret = 0;
 376
 377         if (!b->c.level)
 378                 return 0;
 379 again:
 380         prev = NULL;
 381         have_child = dropped_children = false;
 382         bch2_bkey_buf_init(&prev_k);
 383         bch2_bkey_buf_init(&cur_k);
 384         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 385
 386         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 387                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 388                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 389
 390                 bch2_btree_and_journal_iter_advance(&iter);
 391                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 392
 393                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 394                                         b->c.btree_id, b->c.level - 1,
 395                                         false);
 396                 ret = PTR_ERR_OR_ZERO(cur);
 397
 398                 printbuf_reset(&buf);
 399                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 400
 401                 if (mustfix_fsck_err_on(ret == -EIO, c,
 402                                 btree_node_unreadable,
 403                                 "Topology repair: unreadable btree node at btree %s level %u:\n"
 404                                 "  %s",
 405                                 bch2_btree_id_str(b->c.btree_id),
 406                                 b->c.level - 1,
 407                                 buf.buf)) {
 408                         bch2_btree_node_evict(trans, cur_k.k);
 409                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 410                                                       b->c.level, cur_k.k->k.p);
 411                         cur = NULL;
 412                         if (ret)
 413                                 break;
 414                         continue;
 415                 }
 416
 417                 if (ret) {
 418                         bch_err_msg(c, ret, "getting btree node");
 419                         break;
 420                 }
 421
 422                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 423
 424                 if (ret == DROP_THIS_NODE) {
 425                         six_unlock_read(&cur->c.lock);
 426                         bch2_btree_node_evict(trans, cur_k.k);
 427                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 428                                                       b->c.level, cur_k.k->k.p);
 429                         cur = NULL;
 430                         if (ret)
 431                                 break;
 432                         continue;
 433                 }
 434
 435                 if (prev)
 436                         six_unlock_read(&prev->c.lock);
 437                 prev = NULL;
 438
 439                 if (ret == DROP_PREV_NODE) {
 440                         bch2_btree_node_evict(trans, prev_k.k);
 441                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 442                                                       b->c.level, prev_k.k->k.p);
 443                         if (ret)
 444                                 break;
 445
 446                         bch2_btree_and_journal_iter_exit(&iter);
 447                         bch2_bkey_buf_exit(&prev_k, c);
 448                         bch2_bkey_buf_exit(&cur_k, c);
 449                         goto again;
 450                 } else if (ret)
 451                         break;
 452
 453                 prev = cur;
 454                 cur = NULL;
 455                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 456         }
 457
 458         if (!ret && !IS_ERR_OR_NULL(prev)) {
 459                 BUG_ON(cur);
 460                 ret = btree_repair_node_end(c, b, prev);
 461         }
 462
 463         if (!IS_ERR_OR_NULL(prev))
 464                 six_unlock_read(&prev->c.lock);
 465         prev = NULL;
 466         if (!IS_ERR_OR_NULL(cur))
 467                 six_unlock_read(&cur->c.lock);
 468         cur = NULL;
 469
 470         if (ret)
 471                 goto err;
 472
 473         bch2_btree_and_journal_iter_exit(&iter);
 474         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 475
 476         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 477                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 478                 bch2_btree_and_journal_iter_advance(&iter);
 479
 480                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 481                                         b->c.btree_id, b->c.level - 1,
 482                                         false);
 483                 ret = PTR_ERR_OR_ZERO(cur);
 484
 485                 if (ret) {
 486                         bch_err_msg(c, ret, "getting btree node");
 487                         goto err;
 488                 }
 489
 490                 ret = bch2_btree_repair_topology_recurse(trans, cur);
 491                 six_unlock_read(&cur->c.lock);
 492                 cur = NULL;
 493
 494                 if (ret == DROP_THIS_NODE) {
 495                         bch2_btree_node_evict(trans, cur_k.k);
 496                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 497                                                       b->c.level, cur_k.k->k.p);
 498                         dropped_children = true;
 499                 }
 500
 501                 if (ret)
 502                         goto err;
 503
 504                 have_child = true;
 505         }
 506
 507         printbuf_reset(&buf);
 508         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 509
 510         if (mustfix_fsck_err_on(!have_child, c,
 511                         btree_node_topology_interior_node_empty,
 512                         "empty interior btree node at btree %s level %u\n"
 513                         "  %s",
 514                         bch2_btree_id_str(b->c.btree_id),
 515                         b->c.level, buf.buf))
 516                 ret = DROP_THIS_NODE;
 517 err:
 518 fsck_err:
 519         if (!IS_ERR_OR_NULL(prev))
 520                 six_unlock_read(&prev->c.lock);
 521         if (!IS_ERR_OR_NULL(cur))
 522                 six_unlock_read(&cur->c.lock);
 523
 524         bch2_btree_and_journal_iter_exit(&iter);
 525         bch2_bkey_buf_exit(&prev_k, c);
 526         bch2_bkey_buf_exit(&cur_k, c);
 527
 528         if (!ret && dropped_children)
 529                 goto again;
 530
 531         printbuf_exit(&buf);
 532         return ret;
 533 }
 534
 535 int bch2_check_topology(struct bch_fs *c)
 536 {
 537         struct btree_trans *trans = bch2_trans_get(c);
 538         struct btree *b;
 539         unsigned i;
 540         int ret = 0;
 541
 542         for (i = 0; i < btree_id_nr_alive(c) && !ret; i++) {
 543                 struct btree_root *r = bch2_btree_id_root(c, i);
 544
 545                 if (!r->alive)
 546                         continue;
 547
 548                 b = r->b;
 549                 if (btree_node_fake(b))
 550                         continue;
 551
 552                 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 553                 ret = bch2_btree_repair_topology_recurse(trans, b);
 554                 six_unlock_read(&b->c.lock);
 555
 556                 if (ret == DROP_THIS_NODE) {
 557                         bch_err(c, "empty btree root - repair unimplemented");
 558                         ret = -BCH_ERR_fsck_repair_unimplemented;
 559                 }
 560         }
 561
 562         bch2_trans_put(trans);
 563
 564         return ret;
 565 }
 566
 567 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
 568                                unsigned level, bool is_root,
 569                                struct bkey_s_c *k)
 570 {
 571         struct bch_fs *c = trans->c;
 572         struct bkey_ptrs_c ptrs_c = bch2_bkey_ptrs_c(*k);
 573         const union bch_extent_entry *entry_c;
 574         struct extent_ptr_decoded p = { 0 };
 575         bool do_update = false;
 576         struct printbuf buf = PRINTBUF;
 577         int ret = 0;
 578
 579         /*
 580          * XXX
 581          * use check_bucket_ref here
 582          */
 583         bkey_for_each_ptr_decode(k->k, ptrs_c, p, entry_c) {
 584                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 585                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 586                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry_c->ptr);
 587
 588                 if (!g->gen_valid &&
 589                     (c->opts.reconstruct_alloc ||
 590                      fsck_err(c, ptr_to_missing_alloc_key,
 591                               "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 592                               "while marking %s",
 593                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 594                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 595                               p.ptr.gen,
 596                               (printbuf_reset(&buf),
 597                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 598                         if (!p.ptr.cached) {
 599                                 g->gen_valid            = true;
 600                                 g->gen                  = p.ptr.gen;
 601                         } else {
 602                                 do_update = true;
 603                         }
 604                 }
 605
 606                 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
 607                     (c->opts.reconstruct_alloc ||
 608                      fsck_err(c, ptr_gen_newer_than_bucket_gen,
 609                               "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 610                               "while marking %s",
 611                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 612                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 613                               p.ptr.gen, g->gen,
 614                               (printbuf_reset(&buf),
 615                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 616                         if (!p.ptr.cached) {
 617                                 g->gen_valid            = true;
 618                                 g->gen                  = p.ptr.gen;
 619                                 g->data_type            = 0;
 620                                 g->dirty_sectors        = 0;
 621                                 g->cached_sectors       = 0;
 622                                 set_bit(BCH_FS_need_another_gc, &c->flags);
 623                         } else {
 624                                 do_update = true;
 625                         }
 626                 }
 627
 628                 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
 629                     (c->opts.reconstruct_alloc ||
 630                      fsck_err(c, ptr_gen_newer_than_bucket_gen,
 631                               "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 632                               "while marking %s",
 633                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 634                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 635                               p.ptr.gen,
 636                               (printbuf_reset(&buf),
 637                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 638                         do_update = true;
 639
 640                 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
 641                     (c->opts.reconstruct_alloc ||
 642                      fsck_err(c, stale_dirty_ptr,
 643                               "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 644                               "while marking %s",
 645                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 646                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 647                               p.ptr.gen, g->gen,
 648                               (printbuf_reset(&buf),
 649                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 650                         do_update = true;
 651
 652                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 653                         continue;
 654
 655                 if (fsck_err_on(bucket_data_type(g->data_type) &&
 656                                 bucket_data_type(g->data_type) != data_type, c,
 657                                 ptr_bucket_data_type_mismatch,
 658                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 659                                 "while marking %s",
 660                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 661                                 bch2_data_types[g->data_type],
 662                                 bch2_data_types[data_type],
 663                                 (printbuf_reset(&buf),
 664                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 665                         if (data_type == BCH_DATA_btree) {
 666                                 g->data_type    = data_type;
 667                                 set_bit(BCH_FS_need_another_gc, &c->flags);
 668                         } else {
 669                                 do_update = true;
 670                         }
 671                 }
 672
 673                 if (p.has_ec) {
 674                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 675
 676                         if (fsck_err_on(!m || !m->alive, c,
 677                                         ptr_to_missing_stripe,
 678                                         "pointer to nonexistent stripe %llu\n"
 679                                         "while marking %s",
 680                                         (u64) p.ec.idx,
 681                                         (printbuf_reset(&buf),
 682                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 683                                 do_update = true;
 684
 685                         if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
 686                                         ptr_to_incorrect_stripe,
 687                                         "pointer does not match stripe %llu\n"
 688                                         "while marking %s",
 689                                         (u64) p.ec.idx,
 690                                         (printbuf_reset(&buf),
 691                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 692                                 do_update = true;
 693                 }
 694         }
 695
 696         if (do_update) {
 697                 struct bkey_ptrs ptrs;
 698                 union bch_extent_entry *entry;
 699                 struct bch_extent_ptr *ptr;
 700                 struct bkey_i *new;
 701
 702                 if (is_root) {
 703                         bch_err(c, "cannot update btree roots yet");
 704                         ret = -EINVAL;
 705                         goto err;
 706                 }
 707
 708                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 709                 if (!new) {
 710                         bch_err_msg(c, ret, "allocating new key");
 711                         ret = -BCH_ERR_ENOMEM_gc_repair_key;
 712                         goto err;
 713                 }
 714
 715                 bkey_reassemble(new, *k);
 716
 717                 if (level) {
 718                         /*
 719                          * We don't want to drop btree node pointers - if the
 720                          * btree node isn't there anymore, the read path will
 721                          * sort it out:
 722                          */
 723                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 724                         bkey_for_each_ptr(ptrs, ptr) {
 725                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 726                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 727
 728                                 ptr->gen = g->gen;
 729                         }
 730                 } else {
 731                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 732                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 733                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 734                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 735
 736                                 (ptr->cached &&
 737                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 738                                 (!ptr->cached &&
 739                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 740                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 741                                 (g->data_type &&
 742                                  g->data_type != data_type);
 743                         }));
 744 again:
 745                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 746                         bkey_extent_entry_for_each(ptrs, entry) {
 747                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 748                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 749                                                                         entry->stripe_ptr.idx);
 750                                         union bch_extent_entry *next_ptr;
 751
 752                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 753                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 754                                                         goto found;
 755                                         next_ptr = NULL;
 756 found:
 757                                         if (!next_ptr) {
 758                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 759                                                 continue;
 760                                         }
 761
 762                                         if (!m || !m->alive ||
 763                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 764                                                                        &next_ptr->ptr,
 765                                                                        m->sectors)) {
 766                                                 bch2_bkey_extent_entry_drop(new, entry);
 767                                                 goto again;
 768                                         }
 769                                 }
 770                         }
 771                 }
 772
 773                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 774                 if (ret) {
 775                         kfree(new);
 776                         goto err;
 777                 }
 778
 779                 if (level)
 780                         bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
 781
 782                 if (0) {
 783                         printbuf_reset(&buf);
 784                         bch2_bkey_val_to_text(&buf, c, *k);
 785                         bch_info(c, "updated %s", buf.buf);
 786
 787                         printbuf_reset(&buf);
 788                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 789                         bch_info(c, "new key %s", buf.buf);
 790                 }
 791
 792                 *k = bkey_i_to_s_c(new);
 793         }
 794 err:
 795 fsck_err:
 796         printbuf_exit(&buf);
 797         return ret;
 798 }
 799
 800 /* marking of btree keys/nodes: */
 801
 802 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 803                             unsigned level, bool is_root,
 804                             struct bkey_s_c *k,
 805                             bool initial)
 806 {
 807         struct bch_fs *c = trans->c;
 808         struct bkey deleted = KEY(0, 0, 0);
 809         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 810         unsigned flags =
 811                 BTREE_TRIGGER_GC|
 812                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 813         int ret = 0;
 814
 815         deleted.p = k->k->p;
 816
 817         if (initial) {
 818                 BUG_ON(bch2_journal_seq_verify &&
 819                        k->k->version.lo > atomic64_read(&c->journal.seq));
 820
 821                 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
 822                 if (ret)
 823                         goto err;
 824
 825                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 826                                 bkey_version_in_future,
 827                                 "key version number higher than recorded: %llu > %llu",
 828                                 k->k->version.lo,
 829                                 atomic64_read(&c->key_version)))
 830                         atomic64_set(&c->key_version, k->k->version.lo);
 831         }
 832
 833         ret = commit_do(trans, NULL, NULL, 0,
 834                         bch2_mark_key(trans, btree_id, level, old, *k, flags));
 835 fsck_err:
 836 err:
 837         if (ret)
 838                 bch_err_fn(c, ret);
 839         return ret;
 840 }
 841
 842 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 843 {
 844         struct bch_fs *c = trans->c;
 845         struct btree_node_iter iter;
 846         struct bkey unpacked;
 847         struct bkey_s_c k;
 848         struct bkey_buf prev, cur;
 849         int ret = 0;
 850
 851         if (!btree_node_type_needs_gc(btree_node_type(b)))
 852                 return 0;
 853
 854         bch2_btree_node_iter_init_from_start(&iter, b);
 855         bch2_bkey_buf_init(&prev);
 856         bch2_bkey_buf_init(&cur);
 857         bkey_init(&prev.k->k);
 858
 859         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 860                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 861                                        &k, initial);
 862                 if (ret)
 863                         break;
 864
 865                 bch2_btree_node_iter_advance(&iter, b);
 866
 867                 if (b->c.level) {
 868                         bch2_bkey_buf_reassemble(&cur, c, k);
 869
 870                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 871                                         bch2_btree_node_iter_end(&iter));
 872                         if (ret)
 873                                 break;
 874                 }
 875         }
 876
 877         bch2_bkey_buf_exit(&cur, c);
 878         bch2_bkey_buf_exit(&prev, c);
 879         return ret;
 880 }
 881
 882 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 883                          bool initial, bool metadata_only)
 884 {
 885         struct bch_fs *c = trans->c;
 886         struct btree_iter iter;
 887         struct btree *b;
 888         unsigned depth = metadata_only ? 1 : 0;
 889         int ret = 0;
 890
 891         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 892
 893         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 894                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 895                 bch2_verify_btree_nr_keys(b);
 896
 897                 gc_pos_set(c, gc_pos_btree_node(b));
 898
 899                 ret = btree_gc_mark_node(trans, b, initial);
 900                 if (ret)
 901                         break;
 902         }
 903         bch2_trans_iter_exit(trans, &iter);
 904
 905         if (ret)
 906                 return ret;
 907
 908         mutex_lock(&c->btree_root_lock);
 909         b = bch2_btree_id_root(c, btree_id)->b;
 910         if (!btree_node_fake(b)) {
 911                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 912
 913                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
 914                                        true, &k, initial);
 915         }
 916         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 917         mutex_unlock(&c->btree_root_lock);
 918
 919         return ret;
 920 }
 921
 922 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 923                                       unsigned target_depth)
 924 {
 925         struct bch_fs *c = trans->c;
 926         struct btree_and_journal_iter iter;
 927         struct bkey_s_c k;
 928         struct bkey_buf cur, prev;
 929         struct printbuf buf = PRINTBUF;
 930         int ret = 0;
 931
 932         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 933         bch2_bkey_buf_init(&prev);
 934         bch2_bkey_buf_init(&cur);
 935         bkey_init(&prev.k->k);
 936
 937         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 938                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 939                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 940
 941                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 942                                        false, &k, true);
 943                 if (ret)
 944                         goto fsck_err;
 945
 946                 if (b->c.level) {
 947                         bch2_bkey_buf_reassemble(&cur, c, k);
 948                         k = bkey_i_to_s_c(cur.k);
 949
 950                         bch2_btree_and_journal_iter_advance(&iter);
 951
 952                         ret = bch2_gc_check_topology(c, b,
 953                                         &prev, cur,
 954                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 955                         if (ret)
 956                                 goto fsck_err;
 957                 } else {
 958                         bch2_btree_and_journal_iter_advance(&iter);
 959                 }
 960         }
 961
 962         if (b->c.level > target_depth) {
 963                 bch2_btree_and_journal_iter_exit(&iter);
 964                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 965
 966                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 967                         struct btree *child;
 968
 969                         bch2_bkey_buf_reassemble(&cur, c, k);
 970                         bch2_btree_and_journal_iter_advance(&iter);
 971
 972                         child = bch2_btree_node_get_noiter(trans, cur.k,
 973                                                 b->c.btree_id, b->c.level - 1,
 974                                                 false);
 975                         ret = PTR_ERR_OR_ZERO(child);
 976
 977                         if (ret == -EIO) {
 978                                 bch2_topology_error(c);
 979
 980                                 if (__fsck_err(c,
 981                                           FSCK_CAN_FIX|
 982                                           FSCK_CAN_IGNORE|
 983                                           FSCK_NO_RATELIMIT,
 984                                           btree_node_read_error,
 985                                           "Unreadable btree node at btree %s level %u:\n"
 986                                           "  %s",
 987                                           bch2_btree_id_str(b->c.btree_id),
 988                                           b->c.level - 1,
 989                                           (printbuf_reset(&buf),
 990                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 991                                     should_restart_for_topology_repair(c)) {
 992                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 993                                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 994                                         goto fsck_err;
 995                                 } else {
 996                                         /* Continue marking when opted to not
 997                                          * fix the error: */
 998                                         ret = 0;
 999                                         set_bit(BCH_FS_initial_gc_unfixed, &c->flags);
1000                                         continue;
1001                                 }
1002                         } else if (ret) {
1003                                 bch_err_msg(c, ret, "getting btree node");
1004                                 break;
1005                         }
1006
1007                         ret = bch2_gc_btree_init_recurse(trans, child,
1008                                                          target_depth);
1009                         six_unlock_read(&child->c.lock);
1010
1011                         if (ret)
1012                                 break;
1013                 }
1014         }
1015 fsck_err:
1016         bch2_bkey_buf_exit(&cur, c);
1017         bch2_bkey_buf_exit(&prev, c);
1018         bch2_btree_and_journal_iter_exit(&iter);
1019         printbuf_exit(&buf);
1020         return ret;
1021 }
1022
1023 static int bch2_gc_btree_init(struct btree_trans *trans,
1024                               enum btree_id btree_id,
1025                               bool metadata_only)
1026 {
1027         struct bch_fs *c = trans->c;
1028         struct btree *b;
1029         unsigned target_depth = metadata_only ? 1 : 0;
1030         struct printbuf buf = PRINTBUF;
1031         int ret = 0;
1032
1033         b = bch2_btree_id_root(c, btree_id)->b;
1034
1035         if (btree_node_fake(b))
1036                 return 0;
1037
1038         six_lock_read(&b->c.lock, NULL, NULL);
1039         printbuf_reset(&buf);
1040         bch2_bpos_to_text(&buf, b->data->min_key);
1041         if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1042                                 btree_root_bad_min_key,
1043                         "btree root with incorrect min_key: %s", buf.buf)) {
1044                 bch_err(c, "repair unimplemented");
1045                 ret = -BCH_ERR_fsck_repair_unimplemented;
1046                 goto fsck_err;
1047         }
1048
1049         printbuf_reset(&buf);
1050         bch2_bpos_to_text(&buf, b->data->max_key);
1051         if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1052                                 btree_root_bad_max_key,
1053                         "btree root with incorrect max_key: %s", buf.buf)) {
1054                 bch_err(c, "repair unimplemented");
1055                 ret = -BCH_ERR_fsck_repair_unimplemented;
1056                 goto fsck_err;
1057         }
1058
1059         if (b->c.level >= target_depth)
1060                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1061
1062         if (!ret) {
1063                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1064
1065                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1066                                        &k, true);
1067         }
1068 fsck_err:
1069         six_unlock_read(&b->c.lock);
1070
1071         if (ret < 0)
1072                 bch_err_fn(c, ret);
1073         printbuf_exit(&buf);
1074         return ret;
1075 }
1076
1077 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1078 {
1079         return  (int) btree_id_to_gc_phase(l) -
1080                 (int) btree_id_to_gc_phase(r);
1081 }
1082
1083 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1084 {
1085         struct btree_trans *trans = bch2_trans_get(c);
1086         enum btree_id ids[BTREE_ID_NR];
1087         unsigned i;
1088         int ret = 0;
1089
1090         for (i = 0; i < BTREE_ID_NR; i++)
1091                 ids[i] = i;
1092         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1093
1094         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1095                 ret = initial
1096                         ? bch2_gc_btree_init(trans, ids[i], metadata_only)
1097                         : bch2_gc_btree(trans, ids[i], initial, metadata_only);
1098
1099         for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1100                 if (!bch2_btree_id_root(c, i)->alive)
1101                         continue;
1102
1103                 ret = initial
1104                         ? bch2_gc_btree_init(trans, i, metadata_only)
1105                         : bch2_gc_btree(trans, i, initial, metadata_only);
1106         }
1107
1108         if (ret < 0)
1109                 bch_err_fn(c, ret);
1110
1111         bch2_trans_put(trans);
1112         return ret;
1113 }
1114
1115 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1116                                   u64 start, u64 end,
1117                                   enum bch_data_type type,
1118                                   unsigned flags)
1119 {
1120         u64 b = sector_to_bucket(ca, start);
1121
1122         do {
1123                 unsigned sectors =
1124                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1125
1126                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1127                                           gc_phase(GC_PHASE_SB), flags);
1128                 b++;
1129                 start += sectors;
1130         } while (start < end);
1131 }
1132
1133 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1134                                      unsigned flags)
1135 {
1136         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1137         unsigned i;
1138         u64 b;
1139
1140         for (i = 0; i < layout->nr_superblocks; i++) {
1141                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1142
1143                 if (offset == BCH_SB_SECTOR)
1144                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1145                                               BCH_DATA_sb, flags);
1146
1147                 mark_metadata_sectors(c, ca, offset,
1148                                       offset + (1 << layout->sb_max_size_bits),
1149                                       BCH_DATA_sb, flags);
1150         }
1151
1152         for (i = 0; i < ca->journal.nr; i++) {
1153                 b = ca->journal.buckets[i];
1154                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1155                                           ca->mi.bucket_size,
1156                                           gc_phase(GC_PHASE_SB), flags);
1157         }
1158 }
1159
1160 static void bch2_mark_superblocks(struct bch_fs *c)
1161 {
1162         struct bch_dev *ca;
1163         unsigned i;
1164
1165         mutex_lock(&c->sb_lock);
1166         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1167
1168         for_each_online_member(ca, c, i)
1169                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1170         mutex_unlock(&c->sb_lock);
1171 }
1172
1173 #if 0
1174 /* Also see bch2_pending_btree_node_free_insert_done() */
1175 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1176 {
1177         struct btree_update *as;
1178         struct pending_btree_node_free *d;
1179
1180         mutex_lock(&c->btree_interior_update_lock);
1181         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1182
1183         for_each_pending_btree_node_free(c, as, d)
1184                 if (d->index_update_done)
1185                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1186
1187         mutex_unlock(&c->btree_interior_update_lock);
1188 }
1189 #endif
1190
1191 static void bch2_gc_free(struct bch_fs *c)
1192 {
1193         struct bch_dev *ca;
1194         unsigned i;
1195
1196         genradix_free(&c->reflink_gc_table);
1197         genradix_free(&c->gc_stripes);
1198
1199         for_each_member_device(ca, c, i) {
1200                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1201                         sizeof(struct bucket_array) +
1202                         ca->mi.nbuckets * sizeof(struct bucket));
1203                 ca->buckets_gc = NULL;
1204
1205                 free_percpu(ca->usage_gc);
1206                 ca->usage_gc = NULL;
1207         }
1208
1209         free_percpu(c->usage_gc);
1210         c->usage_gc = NULL;
1211 }
1212
1213 static int bch2_gc_done(struct bch_fs *c,
1214                         bool initial, bool metadata_only)
1215 {
1216         struct bch_dev *ca = NULL;
1217         struct printbuf buf = PRINTBUF;
1218         bool verify = !metadata_only &&
1219                 !c->opts.reconstruct_alloc &&
1220                 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1221         unsigned i, dev;
1222         int ret = 0;
1223
1224         percpu_down_write(&c->mark_lock);
1225
1226 #define copy_field(_err, _f, _msg, ...)                                 \
1227         if (dst->_f != src->_f &&                                       \
1228             (!verify ||                                                 \
1229              fsck_err(c, _err, _msg ": got %llu, should be %llu"        \
1230                       , ##__VA_ARGS__, dst->_f, src->_f)))              \
1231                 dst->_f = src->_f
1232 #define copy_dev_field(_err, _f, _msg, ...)                             \
1233         copy_field(_err, _f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1234 #define copy_fs_field(_err, _f, _msg, ...)                              \
1235         copy_field(_err, _f, "fs has wrong " _msg, ##__VA_ARGS__)
1236
1237         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1238                 bch2_fs_usage_acc_to_base(c, i);
1239
1240         for_each_member_device(ca, c, dev) {
1241                 struct bch_dev_usage *dst = ca->usage_base;
1242                 struct bch_dev_usage *src = (void *)
1243                         bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1244                                              dev_usage_u64s());
1245
1246                 for (i = 0; i < BCH_DATA_NR; i++) {
1247                         copy_dev_field(dev_usage_buckets_wrong,
1248                                        d[i].buckets,    "%s buckets", bch2_data_types[i]);
1249                         copy_dev_field(dev_usage_sectors_wrong,
1250                                        d[i].sectors,    "%s sectors", bch2_data_types[i]);
1251                         copy_dev_field(dev_usage_fragmented_wrong,
1252                                        d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1253                 }
1254         }
1255
1256         {
1257                 unsigned nr = fs_usage_u64s(c);
1258                 struct bch_fs_usage *dst = c->usage_base;
1259                 struct bch_fs_usage *src = (void *)
1260                         bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1261
1262                 copy_fs_field(fs_usage_hidden_wrong,
1263                               hidden,           "hidden");
1264                 copy_fs_field(fs_usage_btree_wrong,
1265                               btree,            "btree");
1266
1267                 if (!metadata_only) {
1268                         copy_fs_field(fs_usage_data_wrong,
1269                                       data,     "data");
1270                         copy_fs_field(fs_usage_cached_wrong,
1271                                       cached,   "cached");
1272                         copy_fs_field(fs_usage_reserved_wrong,
1273                                       reserved, "reserved");
1274                         copy_fs_field(fs_usage_nr_inodes_wrong,
1275                                       nr_inodes,"nr_inodes");
1276
1277                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1278                                 copy_fs_field(fs_usage_persistent_reserved_wrong,
1279                                               persistent_reserved[i],
1280                                               "persistent_reserved[%i]", i);
1281                 }
1282
1283                 for (i = 0; i < c->replicas.nr; i++) {
1284                         struct bch_replicas_entry_v1 *e =
1285                                 cpu_replicas_entry(&c->replicas, i);
1286
1287                         if (metadata_only &&
1288                             (e->data_type == BCH_DATA_user ||
1289                              e->data_type == BCH_DATA_cached))
1290                                 continue;
1291
1292                         printbuf_reset(&buf);
1293                         bch2_replicas_entry_to_text(&buf, e);
1294
1295                         copy_fs_field(fs_usage_replicas_wrong,
1296                                       replicas[i], "%s", buf.buf);
1297                 }
1298         }
1299
1300 #undef copy_fs_field
1301 #undef copy_dev_field
1302 #undef copy_stripe_field
1303 #undef copy_field
1304 fsck_err:
1305         if (ca)
1306                 percpu_ref_put(&ca->ref);
1307         if (ret)
1308                 bch_err_fn(c, ret);
1309
1310         percpu_up_write(&c->mark_lock);
1311         printbuf_exit(&buf);
1312         return ret;
1313 }
1314
1315 static int bch2_gc_start(struct bch_fs *c)
1316 {
1317         struct bch_dev *ca = NULL;
1318         unsigned i;
1319
1320         BUG_ON(c->usage_gc);
1321
1322         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1323                                          sizeof(u64), GFP_KERNEL);
1324         if (!c->usage_gc) {
1325                 bch_err(c, "error allocating c->usage_gc");
1326                 return -BCH_ERR_ENOMEM_gc_start;
1327         }
1328
1329         for_each_member_device(ca, c, i) {
1330                 BUG_ON(ca->usage_gc);
1331
1332                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1333                 if (!ca->usage_gc) {
1334                         bch_err(c, "error allocating ca->usage_gc");
1335                         percpu_ref_put(&ca->ref);
1336                         return -BCH_ERR_ENOMEM_gc_start;
1337                 }
1338
1339                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1340                                ca->mi.nbuckets - ca->mi.first_bucket);
1341         }
1342
1343         return 0;
1344 }
1345
1346 static int bch2_gc_reset(struct bch_fs *c)
1347 {
1348         struct bch_dev *ca;
1349         unsigned i;
1350
1351         for_each_member_device(ca, c, i) {
1352                 free_percpu(ca->usage_gc);
1353                 ca->usage_gc = NULL;
1354         }
1355
1356         free_percpu(c->usage_gc);
1357         c->usage_gc = NULL;
1358
1359         return bch2_gc_start(c);
1360 }
1361
1362 /* returns true if not equal */
1363 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1364                                      struct bch_alloc_v4 r)
1365 {
1366         return  l.gen != r.gen                          ||
1367                 l.oldest_gen != r.oldest_gen            ||
1368                 l.data_type != r.data_type              ||
1369                 l.dirty_sectors != r.dirty_sectors      ||
1370                 l.cached_sectors != r.cached_sectors     ||
1371                 l.stripe_redundancy != r.stripe_redundancy ||
1372                 l.stripe != r.stripe;
1373 }
1374
1375 static int bch2_alloc_write_key(struct btree_trans *trans,
1376                                 struct btree_iter *iter,
1377                                 struct bkey_s_c k,
1378                                 bool metadata_only)
1379 {
1380         struct bch_fs *c = trans->c;
1381         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1382         struct bucket gc, *b;
1383         struct bkey_i_alloc_v4 *a;
1384         struct bch_alloc_v4 old_convert, new;
1385         const struct bch_alloc_v4 *old;
1386         enum bch_data_type type;
1387         int ret;
1388
1389         if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
1390                 return 1;
1391
1392         old = bch2_alloc_to_v4(k, &old_convert);
1393         new = *old;
1394
1395         percpu_down_read(&c->mark_lock);
1396         b = gc_bucket(ca, iter->pos.offset);
1397
1398         /*
1399          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1400          * fix that here:
1401          */
1402         type = __alloc_data_type(b->dirty_sectors,
1403                                  b->cached_sectors,
1404                                  b->stripe,
1405                                  *old,
1406                                  b->data_type);
1407         if (b->data_type != type) {
1408                 struct bch_dev_usage *u;
1409
1410                 preempt_disable();
1411                 u = this_cpu_ptr(ca->usage_gc);
1412                 u->d[b->data_type].buckets--;
1413                 b->data_type = type;
1414                 u->d[b->data_type].buckets++;
1415                 preempt_enable();
1416         }
1417
1418         gc = *b;
1419         percpu_up_read(&c->mark_lock);
1420
1421         if (metadata_only &&
1422             gc.data_type != BCH_DATA_sb &&
1423             gc.data_type != BCH_DATA_journal &&
1424             gc.data_type != BCH_DATA_btree)
1425                 return 0;
1426
1427         if (gen_after(old->gen, gc.gen))
1428                 return 0;
1429
1430         if (c->opts.reconstruct_alloc ||
1431             fsck_err_on(new.data_type != gc.data_type, c,
1432                         alloc_key_data_type_wrong,
1433                         "bucket %llu:%llu gen %u has wrong data_type"
1434                         ": got %s, should be %s",
1435                         iter->pos.inode, iter->pos.offset,
1436                         gc.gen,
1437                         bch2_data_types[new.data_type],
1438                         bch2_data_types[gc.data_type]))
1439                 new.data_type = gc.data_type;
1440
1441 #define copy_bucket_field(_errtype, _f)                                 \
1442         if (c->opts.reconstruct_alloc ||                                \
1443             fsck_err_on(new._f != gc._f, c, _errtype,                   \
1444                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1445                         ": got %u, should be %u",                       \
1446                         iter->pos.inode, iter->pos.offset,              \
1447                         gc.gen,                                         \
1448                         bch2_data_types[gc.data_type],                  \
1449                         new._f, gc._f))                                 \
1450                 new._f = gc._f;                                         \
1451
1452         copy_bucket_field(alloc_key_gen_wrong,
1453                           gen);
1454         copy_bucket_field(alloc_key_dirty_sectors_wrong,
1455                           dirty_sectors);
1456         copy_bucket_field(alloc_key_cached_sectors_wrong,
1457                           cached_sectors);
1458         copy_bucket_field(alloc_key_stripe_wrong,
1459                           stripe);
1460         copy_bucket_field(alloc_key_stripe_redundancy_wrong,
1461                           stripe_redundancy);
1462 #undef copy_bucket_field
1463
1464         if (!bch2_alloc_v4_cmp(*old, new))
1465                 return 0;
1466
1467         a = bch2_alloc_to_v4_mut(trans, k);
1468         ret = PTR_ERR_OR_ZERO(a);
1469         if (ret)
1470                 return ret;
1471
1472         a->v = new;
1473
1474         /*
1475          * The trigger normally makes sure this is set, but we're not running
1476          * triggers:
1477          */
1478         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1479                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1480
1481         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1482 fsck_err:
1483         return ret;
1484 }
1485
1486 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1487 {
1488         struct btree_trans *trans = bch2_trans_get(c);
1489         struct btree_iter iter;
1490         struct bkey_s_c k;
1491         struct bch_dev *ca;
1492         unsigned i;
1493         int ret = 0;
1494
1495         for_each_member_device(ca, c, i) {
1496                 ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1497                                 POS(ca->dev_idx, ca->mi.first_bucket),
1498                                 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1499                                 NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
1500                         bch2_alloc_write_key(trans, &iter, k, metadata_only));
1501
1502                 if (ret < 0) {
1503                         bch_err_fn(c, ret);
1504                         percpu_ref_put(&ca->ref);
1505                         break;
1506                 }
1507         }
1508
1509         bch2_trans_put(trans);
1510         return ret < 0 ? ret : 0;
1511 }
1512
1513 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1514 {
1515         struct bch_dev *ca;
1516         struct btree_trans *trans = bch2_trans_get(c);
1517         struct btree_iter iter;
1518         struct bkey_s_c k;
1519         struct bucket *g;
1520         struct bch_alloc_v4 a_convert;
1521         const struct bch_alloc_v4 *a;
1522         unsigned i;
1523         int ret;
1524
1525         for_each_member_device(ca, c, i) {
1526                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1527                                 ca->mi.nbuckets * sizeof(struct bucket),
1528                                 GFP_KERNEL|__GFP_ZERO);
1529                 if (!buckets) {
1530                         percpu_ref_put(&ca->ref);
1531                         bch_err(c, "error allocating ca->buckets[gc]");
1532                         ret = -BCH_ERR_ENOMEM_gc_alloc_start;
1533                         goto err;
1534                 }
1535
1536                 buckets->first_bucket   = ca->mi.first_bucket;
1537                 buckets->nbuckets       = ca->mi.nbuckets;
1538                 rcu_assign_pointer(ca->buckets_gc, buckets);
1539         }
1540
1541         ret = for_each_btree_key2(trans, iter, BTREE_ID_alloc, POS_MIN,
1542                                   BTREE_ITER_PREFETCH, k, ({
1543                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1544                 g = gc_bucket(ca, k.k->p.offset);
1545
1546                 a = bch2_alloc_to_v4(k, &a_convert);
1547
1548                 g->gen_valid    = 1;
1549                 g->gen          = a->gen;
1550
1551                 if (metadata_only &&
1552                     (a->data_type == BCH_DATA_user ||
1553                      a->data_type == BCH_DATA_cached ||
1554                      a->data_type == BCH_DATA_parity)) {
1555                         g->data_type            = a->data_type;
1556                         g->dirty_sectors        = a->dirty_sectors;
1557                         g->cached_sectors       = a->cached_sectors;
1558                         g->stripe               = a->stripe;
1559                         g->stripe_redundancy    = a->stripe_redundancy;
1560                 }
1561
1562                 0;
1563         }));
1564 err:
1565         bch2_trans_put(trans);
1566         if (ret)
1567                 bch_err_fn(c, ret);
1568         return ret;
1569 }
1570
1571 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1572 {
1573         struct bch_dev *ca;
1574         unsigned i;
1575
1576         for_each_member_device(ca, c, i) {
1577                 struct bucket_array *buckets = gc_bucket_array(ca);
1578                 struct bucket *g;
1579
1580                 for_each_bucket(g, buckets) {
1581                         if (metadata_only &&
1582                             (g->data_type == BCH_DATA_user ||
1583                              g->data_type == BCH_DATA_cached ||
1584                              g->data_type == BCH_DATA_parity))
1585                                 continue;
1586                         g->data_type = 0;
1587                         g->dirty_sectors = 0;
1588                         g->cached_sectors = 0;
1589                 }
1590         }
1591 }
1592
1593 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1594                                      struct btree_iter *iter,
1595                                      struct bkey_s_c k,
1596                                      size_t *idx)
1597 {
1598         struct bch_fs *c = trans->c;
1599         const __le64 *refcount = bkey_refcount_c(k);
1600         struct printbuf buf = PRINTBUF;
1601         struct reflink_gc *r;
1602         int ret = 0;
1603
1604         if (!refcount)
1605                 return 0;
1606
1607         while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1608                r->offset < k.k->p.offset)
1609                 ++*idx;
1610
1611         if (!r ||
1612             r->offset != k.k->p.offset ||
1613             r->size != k.k->size) {
1614                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1615                 return -EINVAL;
1616         }
1617
1618         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1619                         reflink_v_refcount_wrong,
1620                         "reflink key has wrong refcount:\n"
1621                         "  %s\n"
1622                         "  should be %u",
1623                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1624                         r->refcount)) {
1625                 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1626
1627                 ret = PTR_ERR_OR_ZERO(new);
1628                 if (ret)
1629                         return ret;
1630
1631                 if (!r->refcount)
1632                         new->k.type = KEY_TYPE_deleted;
1633                 else
1634                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1635         }
1636 fsck_err:
1637         printbuf_exit(&buf);
1638         return ret;
1639 }
1640
1641 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1642 {
1643         struct btree_trans *trans;
1644         struct btree_iter iter;
1645         struct bkey_s_c k;
1646         size_t idx = 0;
1647         int ret = 0;
1648
1649         if (metadata_only)
1650                 return 0;
1651
1652         trans = bch2_trans_get(c);
1653
1654         ret = for_each_btree_key_commit(trans, iter,
1655                         BTREE_ID_reflink, POS_MIN,
1656                         BTREE_ITER_PREFETCH, k,
1657                         NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1658                 bch2_gc_write_reflink_key(trans, &iter, k, &idx));
1659
1660         c->reflink_gc_nr = 0;
1661         bch2_trans_put(trans);
1662         return ret;
1663 }
1664
1665 static int bch2_gc_reflink_start(struct bch_fs *c,
1666                                  bool metadata_only)
1667 {
1668         struct btree_trans *trans;
1669         struct btree_iter iter;
1670         struct bkey_s_c k;
1671         struct reflink_gc *r;
1672         int ret = 0;
1673
1674         if (metadata_only)
1675                 return 0;
1676
1677         trans = bch2_trans_get(c);
1678         c->reflink_gc_nr = 0;
1679
1680         for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
1681                            BTREE_ITER_PREFETCH, k, ret) {
1682                 const __le64 *refcount = bkey_refcount_c(k);
1683
1684                 if (!refcount)
1685                         continue;
1686
1687                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1688                                        GFP_KERNEL);
1689                 if (!r) {
1690                         ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1691                         break;
1692                 }
1693
1694                 r->offset       = k.k->p.offset;
1695                 r->size         = k.k->size;
1696                 r->refcount     = 0;
1697         }
1698         bch2_trans_iter_exit(trans, &iter);
1699
1700         bch2_trans_put(trans);
1701         return ret;
1702 }
1703
1704 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1705 {
1706         struct genradix_iter iter;
1707         struct reflink_gc *r;
1708
1709         genradix_for_each(&c->reflink_gc_table, iter, r)
1710                 r->refcount = 0;
1711 }
1712
1713 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1714                                      struct btree_iter *iter,
1715                                      struct bkey_s_c k)
1716 {
1717         struct bch_fs *c = trans->c;
1718         struct printbuf buf = PRINTBUF;
1719         const struct bch_stripe *s;
1720         struct gc_stripe *m;
1721         bool bad = false;
1722         unsigned i;
1723         int ret = 0;
1724
1725         if (k.k->type != KEY_TYPE_stripe)
1726                 return 0;
1727
1728         s = bkey_s_c_to_stripe(k).v;
1729         m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1730
1731         for (i = 0; i < s->nr_blocks; i++) {
1732                 u32 old = stripe_blockcount_get(s, i);
1733                 u32 new = (m ? m->block_sectors[i] : 0);
1734
1735                 if (old != new) {
1736                         prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1737                                    i, old, new);
1738                         bad = true;
1739                 }
1740         }
1741
1742         if (bad)
1743                 bch2_bkey_val_to_text(&buf, c, k);
1744
1745         if (fsck_err_on(bad, c, stripe_sector_count_wrong,
1746                         "%s", buf.buf)) {
1747                 struct bkey_i_stripe *new;
1748
1749                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1750                 ret = PTR_ERR_OR_ZERO(new);
1751                 if (ret)
1752                         return ret;
1753
1754                 bkey_reassemble(&new->k_i, k);
1755
1756                 for (i = 0; i < new->v.nr_blocks; i++)
1757                         stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1758
1759                 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1760         }
1761 fsck_err:
1762         printbuf_exit(&buf);
1763         return ret;
1764 }
1765
1766 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1767 {
1768         struct btree_trans *trans;
1769         struct btree_iter iter;
1770         struct bkey_s_c k;
1771         int ret = 0;
1772
1773         if (metadata_only)
1774                 return 0;
1775
1776         trans = bch2_trans_get(c);
1777
1778         ret = for_each_btree_key_commit(trans, iter,
1779                         BTREE_ID_stripes, POS_MIN,
1780                         BTREE_ITER_PREFETCH, k,
1781                         NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1782                 bch2_gc_write_stripes_key(trans, &iter, k));
1783
1784         bch2_trans_put(trans);
1785         return ret;
1786 }
1787
1788 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1789 {
1790         genradix_free(&c->gc_stripes);
1791 }
1792
1793 /**
1794  * bch2_gc - walk _all_ references to buckets, and recompute them:
1795  *
1796  * @c:                  filesystem object
1797  * @initial:            are we in recovery?
1798  * @metadata_only:      are we just checking metadata references, or everything?
1799  *
1800  * Returns: 0 on success, or standard errcode on failure
1801  *
1802  * Order matters here:
1803  *  - Concurrent GC relies on the fact that we have a total ordering for
1804  *    everything that GC walks - see  gc_will_visit_node(),
1805  *    gc_will_visit_root()
1806  *
1807  *  - also, references move around in the course of index updates and
1808  *    various other crap: everything needs to agree on the ordering
1809  *    references are allowed to move around in - e.g., we're allowed to
1810  *    start with a reference owned by an open_bucket (the allocator) and
1811  *    move it to the btree, but not the reverse.
1812  *
1813  *    This is necessary to ensure that gc doesn't miss references that
1814  *    move around - if references move backwards in the ordering GC
1815  *    uses, GC could skip past them
1816  */
1817 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1818 {
1819         unsigned iter = 0;
1820         int ret;
1821
1822         lockdep_assert_held(&c->state_lock);
1823
1824         down_write(&c->gc_lock);
1825
1826         bch2_btree_interior_updates_flush(c);
1827
1828         ret   = bch2_gc_start(c) ?:
1829                 bch2_gc_alloc_start(c, metadata_only) ?:
1830                 bch2_gc_reflink_start(c, metadata_only);
1831         if (ret)
1832                 goto out;
1833 again:
1834         gc_pos_set(c, gc_phase(GC_PHASE_START));
1835
1836         bch2_mark_superblocks(c);
1837
1838         ret = bch2_gc_btrees(c, initial, metadata_only);
1839
1840         if (ret)
1841                 goto out;
1842
1843 #if 0
1844         bch2_mark_pending_btree_node_frees(c);
1845 #endif
1846         c->gc_count++;
1847
1848         if (test_bit(BCH_FS_need_another_gc, &c->flags) ||
1849             (!iter && bch2_test_restart_gc)) {
1850                 if (iter++ > 2) {
1851                         bch_info(c, "Unable to fix bucket gens, looping");
1852                         ret = -EINVAL;
1853                         goto out;
1854                 }
1855
1856                 /*
1857                  * XXX: make sure gens we fixed got saved
1858                  */
1859                 bch_info(c, "Second GC pass needed, restarting:");
1860                 clear_bit(BCH_FS_need_another_gc, &c->flags);
1861                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1862
1863                 bch2_gc_stripes_reset(c, metadata_only);
1864                 bch2_gc_alloc_reset(c, metadata_only);
1865                 bch2_gc_reflink_reset(c, metadata_only);
1866                 ret = bch2_gc_reset(c);
1867                 if (ret)
1868                         goto out;
1869
1870                 /* flush fsck errors, reset counters */
1871                 bch2_flush_fsck_errs(c);
1872                 goto again;
1873         }
1874 out:
1875         if (!ret) {
1876                 bch2_journal_block(&c->journal);
1877
1878                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1879                         bch2_gc_reflink_done(c, metadata_only) ?:
1880                         bch2_gc_alloc_done(c, metadata_only) ?:
1881                         bch2_gc_done(c, initial, metadata_only);
1882
1883                 bch2_journal_unblock(&c->journal);
1884         }
1885
1886         percpu_down_write(&c->mark_lock);
1887         /* Indicates that gc is no longer in progress: */
1888         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1889
1890         bch2_gc_free(c);
1891         percpu_up_write(&c->mark_lock);
1892
1893         up_write(&c->gc_lock);
1894
1895         /*
1896          * At startup, allocations can happen directly instead of via the
1897          * allocator thread - issue wakeup in case they blocked on gc_lock:
1898          */
1899         closure_wake_up(&c->freelist_wait);
1900
1901         if (ret)
1902                 bch_err_fn(c, ret);
1903         return ret;
1904 }
1905
1906 static int gc_btree_gens_key(struct btree_trans *trans,
1907                              struct btree_iter *iter,
1908                              struct bkey_s_c k)
1909 {
1910         struct bch_fs *c = trans->c;
1911         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1912         const struct bch_extent_ptr *ptr;
1913         struct bkey_i *u;
1914         int ret;
1915
1916         percpu_down_read(&c->mark_lock);
1917         bkey_for_each_ptr(ptrs, ptr) {
1918                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1919
1920                 if (ptr_stale(ca, ptr) > 16) {
1921                         percpu_up_read(&c->mark_lock);
1922                         goto update;
1923                 }
1924         }
1925
1926         bkey_for_each_ptr(ptrs, ptr) {
1927                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1928                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1929
1930                 if (gen_after(*gen, ptr->gen))
1931                         *gen = ptr->gen;
1932         }
1933         percpu_up_read(&c->mark_lock);
1934         return 0;
1935 update:
1936         u = bch2_bkey_make_mut(trans, iter, &k, 0);
1937         ret = PTR_ERR_OR_ZERO(u);
1938         if (ret)
1939                 return ret;
1940
1941         bch2_extent_normalize(c, bkey_i_to_s(u));
1942         return 0;
1943 }
1944
1945 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1946                                        struct bkey_s_c k)
1947 {
1948         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1949         struct bch_alloc_v4 a_convert;
1950         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1951         struct bkey_i_alloc_v4 *a_mut;
1952         int ret;
1953
1954         if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1955                 return 0;
1956
1957         a_mut = bch2_alloc_to_v4_mut(trans, k);
1958         ret = PTR_ERR_OR_ZERO(a_mut);
1959         if (ret)
1960                 return ret;
1961
1962         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1963         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1964
1965         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1966 }
1967
1968 int bch2_gc_gens(struct bch_fs *c)
1969 {
1970         struct btree_trans *trans;
1971         struct btree_iter iter;
1972         struct bkey_s_c k;
1973         struct bch_dev *ca;
1974         u64 b, start_time = local_clock();
1975         unsigned i;
1976         int ret;
1977
1978         /*
1979          * Ideally we would be using state_lock and not gc_lock here, but that
1980          * introduces a deadlock in the RO path - we currently take the state
1981          * lock at the start of going RO, thus the gc thread may get stuck:
1982          */
1983         if (!mutex_trylock(&c->gc_gens_lock))
1984                 return 0;
1985
1986         trace_and_count(c, gc_gens_start, c);
1987         down_read(&c->gc_lock);
1988         trans = bch2_trans_get(c);
1989
1990         for_each_member_device(ca, c, i) {
1991                 struct bucket_gens *gens = bucket_gens(ca);
1992
1993                 BUG_ON(ca->oldest_gen);
1994
1995                 ca->oldest_gen = kvmalloc(gens->nbuckets, GFP_KERNEL);
1996                 if (!ca->oldest_gen) {
1997                         percpu_ref_put(&ca->ref);
1998                         ret = -BCH_ERR_ENOMEM_gc_gens;
1999                         goto err;
2000                 }
2001
2002                 for (b = gens->first_bucket;
2003                      b < gens->nbuckets; b++)
2004                         ca->oldest_gen[b] = gens->b[b];
2005         }
2006
2007         for (i = 0; i < BTREE_ID_NR; i++)
2008                 if (btree_type_has_ptrs(i)) {
2009                         c->gc_gens_btree = i;
2010                         c->gc_gens_pos = POS_MIN;
2011
2012                         ret = for_each_btree_key_commit(trans, iter, i,
2013                                         POS_MIN,
2014                                         BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
2015                                         k,
2016                                         NULL, NULL,
2017                                         BCH_TRANS_COMMIT_no_enospc,
2018                                 gc_btree_gens_key(trans, &iter, k));
2019                         if (ret && !bch2_err_matches(ret, EROFS))
2020                                 bch_err_fn(c, ret);
2021                         if (ret)
2022                                 goto err;
2023                 }
2024
2025         ret = for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
2026                         POS_MIN,
2027                         BTREE_ITER_PREFETCH,
2028                         k,
2029                         NULL, NULL,
2030                         BCH_TRANS_COMMIT_no_enospc,
2031                 bch2_alloc_write_oldest_gen(trans, &iter, k));
2032         if (ret && !bch2_err_matches(ret, EROFS))
2033                 bch_err_fn(c, ret);
2034         if (ret)
2035                 goto err;
2036
2037         c->gc_gens_btree        = 0;
2038         c->gc_gens_pos          = POS_MIN;
2039
2040         c->gc_count++;
2041
2042         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2043         trace_and_count(c, gc_gens_end, c);
2044 err:
2045         for_each_member_device(ca, c, i) {
2046                 kvfree(ca->oldest_gen);
2047                 ca->oldest_gen = NULL;
2048         }
2049
2050         bch2_trans_put(trans);
2051         up_read(&c->gc_lock);
2052         mutex_unlock(&c->gc_gens_lock);
2053         return ret;
2054 }
2055
2056 static int bch2_gc_thread(void *arg)
2057 {
2058         struct bch_fs *c = arg;
2059         struct io_clock *clock = &c->io_clock[WRITE];
2060         unsigned long last = atomic64_read(&clock->now);
2061         unsigned last_kick = atomic_read(&c->kick_gc);
2062         int ret;
2063
2064         set_freezable();
2065
2066         while (1) {
2067                 while (1) {
2068                         set_current_state(TASK_INTERRUPTIBLE);
2069
2070                         if (kthread_should_stop()) {
2071                                 __set_current_state(TASK_RUNNING);
2072                                 return 0;
2073                         }
2074
2075                         if (atomic_read(&c->kick_gc) != last_kick)
2076                                 break;
2077
2078                         if (c->btree_gc_periodic) {
2079                                 unsigned long next = last + c->capacity / 16;
2080
2081                                 if (atomic64_read(&clock->now) >= next)
2082                                         break;
2083
2084                                 bch2_io_clock_schedule_timeout(clock, next);
2085                         } else {
2086                                 schedule();
2087                         }
2088
2089                         try_to_freeze();
2090                 }
2091                 __set_current_state(TASK_RUNNING);
2092
2093                 last = atomic64_read(&clock->now);
2094                 last_kick = atomic_read(&c->kick_gc);
2095
2096                 /*
2097                  * Full gc is currently incompatible with btree key cache:
2098                  */
2099 #if 0
2100                 ret = bch2_gc(c, false, false);
2101 #else
2102                 ret = bch2_gc_gens(c);
2103 #endif
2104                 if (ret < 0)
2105                         bch_err_fn(c, ret);
2106
2107                 debug_check_no_locks_held();
2108         }
2109
2110         return 0;
2111 }
2112
2113 void bch2_gc_thread_stop(struct bch_fs *c)
2114 {
2115         struct task_struct *p;
2116
2117         p = c->gc_thread;
2118         c->gc_thread = NULL;
2119
2120         if (p) {
2121                 kthread_stop(p);
2122                 put_task_struct(p);
2123         }
2124 }
2125
2126 int bch2_gc_thread_start(struct bch_fs *c)
2127 {
2128         struct task_struct *p;
2129
2130         if (c->gc_thread)
2131                 return 0;
2132
2133         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2134         if (IS_ERR(p)) {
2135                 bch_err_fn(c, PTR_ERR(p));
2136                 return PTR_ERR(p);
2137         }
2138
2139         get_task_struct(p);
2140         c->gc_thread = p;
2141         wake_up_process(p);
2142         return 0;
2143 }