git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_gc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  * Copyright (C) 2014 Datera Inc.
   5  */
   6
   7 #include "bcachefs.h"
   8 #include "alloc_background.h"
   9 #include "alloc_foreground.h"
  10 #include "bkey_methods.h"
  11 #include "bkey_buf.h"
  12 #include "btree_journal_iter.h"
  13 #include "btree_key_cache.h"
  14 #include "btree_locking.h"
  15 #include "btree_update_interior.h"
  16 #include "btree_io.h"
  17 #include "btree_gc.h"
  18 #include "buckets.h"
  19 #include "clock.h"
  20 #include "debug.h"
  21 #include "ec.h"
  22 #include "error.h"
  23 #include "extents.h"
  24 #include "journal.h"
  25 #include "keylist.h"
  26 #include "move.h"
  27 #include "recovery.h"
  28 #include "reflink.h"
  29 #include "replicas.h"
  30 #include "super-io.h"
  31 #include "trace.h"
  32
  33 #include <linux/slab.h>
  34 #include <linux/bitops.h>
  35 #include <linux/freezer.h>
  36 #include <linux/kthread.h>
  37 #include <linux/preempt.h>
  38 #include <linux/rcupdate.h>
  39 #include <linux/sched/task.h>
  40
  41 #define DROP_THIS_NODE          10
  42 #define DROP_PREV_NODE          11
  43
  44 static bool should_restart_for_topology_repair(struct bch_fs *c)
  45 {
  46         return c->opts.fix_errors != FSCK_FIX_no &&
  47                 !(c->recovery_passes_complete & BIT_ULL(BCH_RECOVERY_PASS_check_topology));
  48 }
  49
  50 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  51 {
  52         preempt_disable();
  53         write_seqcount_begin(&c->gc_pos_lock);
  54         c->gc_pos = new_pos;
  55         write_seqcount_end(&c->gc_pos_lock);
  56         preempt_enable();
  57 }
  58
  59 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  60 {
  61         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  62         __gc_pos_set(c, new_pos);
  63 }
  64
  65 /*
  66  * Missing: if an interior btree node is empty, we need to do something -
  67  * perhaps just kill it
  68  */
  69 static int bch2_gc_check_topology(struct bch_fs *c,
  70                                   struct btree *b,
  71                                   struct bkey_buf *prev,
  72                                   struct bkey_buf cur,
  73                                   bool is_last)
  74 {
  75         struct bpos node_start  = b->data->min_key;
  76         struct bpos node_end    = b->data->max_key;
  77         struct bpos expected_start = bkey_deleted(&prev->k->k)
  78                 ? node_start
  79                 : bpos_successor(prev->k->k.p);
  80         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  81         int ret = 0;
  82
  83         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  84                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  85
  86                 if (!bpos_eq(expected_start, bp->v.min_key)) {
  87                         bch2_topology_error(c);
  88
  89                         if (bkey_deleted(&prev->k->k)) {
  90                                 prt_printf(&buf1, "start of node: ");
  91                                 bch2_bpos_to_text(&buf1, node_start);
  92                         } else {
  93                                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
  94                         }
  95                         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
  96
  97                         if (__fsck_err(c,
  98                                   FSCK_CAN_FIX|
  99                                   FSCK_CAN_IGNORE|
 100                                   FSCK_NO_RATELIMIT,
 101                                   "btree node with incorrect min_key at btree %s level %u:\n"
 102                                   "  prev %s\n"
 103                                   "  cur %s",
 104                                   bch2_btree_ids[b->c.btree_id], b->c.level,
 105                                   buf1.buf, buf2.buf) &&
 106                             should_restart_for_topology_repair(c)) {
 107                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 108                                 ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 109                                 goto err;
 110                         } else {
 111                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 112                         }
 113                 }
 114         }
 115
 116         if (is_last && !bpos_eq(cur.k->k.p, node_end)) {
 117                 bch2_topology_error(c);
 118
 119                 printbuf_reset(&buf1);
 120                 printbuf_reset(&buf2);
 121
 122                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
 123                 bch2_bpos_to_text(&buf2, node_end);
 124
 125                 if (__fsck_err(c,
 126                           FSCK_CAN_FIX|
 127                           FSCK_CAN_IGNORE|
 128                           FSCK_NO_RATELIMIT,
 129                           "btree node with incorrect max_key at btree %s level %u:\n"
 130                           "  %s\n"
 131                           "  expected %s",
 132                           bch2_btree_ids[b->c.btree_id], b->c.level,
 133                           buf1.buf, buf2.buf) &&
 134                     should_restart_for_topology_repair(c)) {
 135                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 136                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 137                         goto err;
 138                 } else {
 139                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 140                 }
 141         }
 142
 143         bch2_bkey_buf_copy(prev, c, cur.k);
 144 err:
 145 fsck_err:
 146         printbuf_exit(&buf2);
 147         printbuf_exit(&buf1);
 148         return ret;
 149 }
 150
 151 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 152 {
 153         switch (b->key.k.type) {
 154         case KEY_TYPE_btree_ptr: {
 155                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 156
 157                 dst->k.p                = src->k.p;
 158                 dst->v.mem_ptr          = 0;
 159                 dst->v.seq              = b->data->keys.seq;
 160                 dst->v.sectors_written  = 0;
 161                 dst->v.flags            = 0;
 162                 dst->v.min_key          = b->data->min_key;
 163                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 164                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 165                 break;
 166         }
 167         case KEY_TYPE_btree_ptr_v2:
 168                 bkey_copy(&dst->k_i, &b->key);
 169                 break;
 170         default:
 171                 BUG();
 172         }
 173 }
 174
 175 static void bch2_btree_node_update_key_early(struct btree_trans *trans,
 176                                              enum btree_id btree, unsigned level,
 177                                              struct bkey_s_c old, struct bkey_i *new)
 178 {
 179         struct bch_fs *c = trans->c;
 180         struct btree *b;
 181         struct bkey_buf tmp;
 182         int ret;
 183
 184         bch2_bkey_buf_init(&tmp);
 185         bch2_bkey_buf_reassemble(&tmp, c, old);
 186
 187         b = bch2_btree_node_get_noiter(trans, tmp.k, btree, level, true);
 188         if (!IS_ERR_OR_NULL(b)) {
 189                 mutex_lock(&c->btree_cache.lock);
 190
 191                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 192
 193                 bkey_copy(&b->key, new);
 194                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 195                 BUG_ON(ret);
 196
 197                 mutex_unlock(&c->btree_cache.lock);
 198                 six_unlock_read(&b->c.lock);
 199         }
 200
 201         bch2_bkey_buf_exit(&tmp, c);
 202 }
 203
 204 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 205 {
 206         struct bkey_i_btree_ptr_v2 *new;
 207         int ret;
 208
 209         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 210         if (!new)
 211                 return -BCH_ERR_ENOMEM_gc_repair_key;
 212
 213         btree_ptr_to_v2(b, new);
 214         b->data->min_key        = new_min;
 215         new->v.min_key          = new_min;
 216         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 217
 218         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 219         if (ret) {
 220                 kfree(new);
 221                 return ret;
 222         }
 223
 224         bch2_btree_node_drop_keys_outside_node(b);
 225         bkey_copy(&b->key, &new->k_i);
 226         return 0;
 227 }
 228
 229 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 230 {
 231         struct bkey_i_btree_ptr_v2 *new;
 232         int ret;
 233
 234         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 235         if (ret)
 236                 return ret;
 237
 238         new = kmalloc_array(BKEY_BTREE_PTR_U64s_MAX, sizeof(u64), GFP_KERNEL);
 239         if (!new)
 240                 return -BCH_ERR_ENOMEM_gc_repair_key;
 241
 242         btree_ptr_to_v2(b, new);
 243         b->data->max_key        = new_max;
 244         new->k.p                = new_max;
 245         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 246
 247         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 248         if (ret) {
 249                 kfree(new);
 250                 return ret;
 251         }
 252
 253         bch2_btree_node_drop_keys_outside_node(b);
 254
 255         mutex_lock(&c->btree_cache.lock);
 256         bch2_btree_node_hash_remove(&c->btree_cache, b);
 257
 258         bkey_copy(&b->key, &new->k_i);
 259         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 260         BUG_ON(ret);
 261         mutex_unlock(&c->btree_cache.lock);
 262         return 0;
 263 }
 264
 265 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 266                                         struct btree *prev, struct btree *cur)
 267 {
 268         struct bpos expected_start = !prev
 269                 ? b->data->min_key
 270                 : bpos_successor(prev->key.k.p);
 271         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 272         int ret = 0;
 273
 274         if (!prev) {
 275                 prt_printf(&buf1, "start of node: ");
 276                 bch2_bpos_to_text(&buf1, b->data->min_key);
 277         } else {
 278                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
 279         }
 280
 281         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
 282
 283         if (prev &&
 284             bpos_gt(expected_start, cur->data->min_key) &&
 285             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 286                 /* cur overwrites prev: */
 287
 288                 if (mustfix_fsck_err_on(bpos_ge(prev->data->min_key,
 289                                                 cur->data->min_key), c,
 290                                 "btree node overwritten by next node at btree %s level %u:\n"
 291                                 "  node %s\n"
 292                                 "  next %s",
 293                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 294                                 buf1.buf, buf2.buf)) {
 295                         ret = DROP_PREV_NODE;
 296                         goto out;
 297                 }
 298
 299                 if (mustfix_fsck_err_on(!bpos_eq(prev->key.k.p,
 300                                                  bpos_predecessor(cur->data->min_key)), c,
 301                                 "btree node with incorrect max_key at btree %s level %u:\n"
 302                                 "  node %s\n"
 303                                 "  next %s",
 304                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 305                                 buf1.buf, buf2.buf))
 306                         ret = set_node_max(c, prev,
 307                                            bpos_predecessor(cur->data->min_key));
 308         } else {
 309                 /* prev overwrites cur: */
 310
 311                 if (mustfix_fsck_err_on(bpos_ge(expected_start,
 312                                                 cur->data->max_key), c,
 313                                 "btree node overwritten by prev node at btree %s level %u:\n"
 314                                 "  prev %s\n"
 315                                 "  node %s",
 316                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 317                                 buf1.buf, buf2.buf)) {
 318                         ret = DROP_THIS_NODE;
 319                         goto out;
 320                 }
 321
 322                 if (mustfix_fsck_err_on(!bpos_eq(expected_start, cur->data->min_key), c,
 323                                 "btree node with incorrect min_key at btree %s level %u:\n"
 324                                 "  prev %s\n"
 325                                 "  node %s",
 326                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 327                                 buf1.buf, buf2.buf))
 328                         ret = set_node_min(c, cur, expected_start);
 329         }
 330 out:
 331 fsck_err:
 332         printbuf_exit(&buf2);
 333         printbuf_exit(&buf1);
 334         return ret;
 335 }
 336
 337 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 338                                  struct btree *child)
 339 {
 340         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 341         int ret = 0;
 342
 343         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
 344         bch2_bpos_to_text(&buf2, b->key.k.p);
 345
 346         if (mustfix_fsck_err_on(!bpos_eq(child->key.k.p, b->key.k.p), c,
 347                         "btree node with incorrect max_key at btree %s level %u:\n"
 348                         "  %s\n"
 349                         "  expected %s",
 350                         bch2_btree_ids[b->c.btree_id], b->c.level,
 351                         buf1.buf, buf2.buf)) {
 352                 ret = set_node_max(c, child, b->key.k.p);
 353                 if (ret)
 354                         goto err;
 355         }
 356 err:
 357 fsck_err:
 358         printbuf_exit(&buf2);
 359         printbuf_exit(&buf1);
 360         return ret;
 361 }
 362
 363 static int bch2_btree_repair_topology_recurse(struct btree_trans *trans, struct btree *b)
 364 {
 365         struct bch_fs *c = trans->c;
 366         struct btree_and_journal_iter iter;
 367         struct bkey_s_c k;
 368         struct bkey_buf prev_k, cur_k;
 369         struct btree *prev = NULL, *cur = NULL;
 370         bool have_child, dropped_children = false;
 371         struct printbuf buf = PRINTBUF;
 372         int ret = 0;
 373
 374         if (!b->c.level)
 375                 return 0;
 376 again:
 377         prev = NULL;
 378         have_child = dropped_children = false;
 379         bch2_bkey_buf_init(&prev_k);
 380         bch2_bkey_buf_init(&cur_k);
 381         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 382
 383         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 384                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 385                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 386
 387                 bch2_btree_and_journal_iter_advance(&iter);
 388                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 389
 390                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 391                                         b->c.btree_id, b->c.level - 1,
 392                                         false);
 393                 ret = PTR_ERR_OR_ZERO(cur);
 394
 395                 printbuf_reset(&buf);
 396                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 397
 398                 if (mustfix_fsck_err_on(ret == -EIO, c,
 399                                 "Topology repair: unreadable btree node at btree %s level %u:\n"
 400                                 "  %s",
 401                                 bch2_btree_ids[b->c.btree_id],
 402                                 b->c.level - 1,
 403                                 buf.buf)) {
 404                         bch2_btree_node_evict(trans, cur_k.k);
 405                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 406                                                       b->c.level, cur_k.k->k.p);
 407                         cur = NULL;
 408                         if (ret)
 409                                 break;
 410                         continue;
 411                 }
 412
 413                 if (ret) {
 414                         bch_err_msg(c, ret, "getting btree node");
 415                         break;
 416                 }
 417
 418                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 419
 420                 if (ret == DROP_THIS_NODE) {
 421                         six_unlock_read(&cur->c.lock);
 422                         bch2_btree_node_evict(trans, cur_k.k);
 423                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 424                                                       b->c.level, cur_k.k->k.p);
 425                         cur = NULL;
 426                         if (ret)
 427                                 break;
 428                         continue;
 429                 }
 430
 431                 if (prev)
 432                         six_unlock_read(&prev->c.lock);
 433                 prev = NULL;
 434
 435                 if (ret == DROP_PREV_NODE) {
 436                         bch2_btree_node_evict(trans, prev_k.k);
 437                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 438                                                       b->c.level, prev_k.k->k.p);
 439                         if (ret)
 440                                 break;
 441
 442                         bch2_btree_and_journal_iter_exit(&iter);
 443                         bch2_bkey_buf_exit(&prev_k, c);
 444                         bch2_bkey_buf_exit(&cur_k, c);
 445                         goto again;
 446                 } else if (ret)
 447                         break;
 448
 449                 prev = cur;
 450                 cur = NULL;
 451                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 452         }
 453
 454         if (!ret && !IS_ERR_OR_NULL(prev)) {
 455                 BUG_ON(cur);
 456                 ret = btree_repair_node_end(c, b, prev);
 457         }
 458
 459         if (!IS_ERR_OR_NULL(prev))
 460                 six_unlock_read(&prev->c.lock);
 461         prev = NULL;
 462         if (!IS_ERR_OR_NULL(cur))
 463                 six_unlock_read(&cur->c.lock);
 464         cur = NULL;
 465
 466         if (ret)
 467                 goto err;
 468
 469         bch2_btree_and_journal_iter_exit(&iter);
 470         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 471
 472         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 473                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 474                 bch2_btree_and_journal_iter_advance(&iter);
 475
 476                 cur = bch2_btree_node_get_noiter(trans, cur_k.k,
 477                                         b->c.btree_id, b->c.level - 1,
 478                                         false);
 479                 ret = PTR_ERR_OR_ZERO(cur);
 480
 481                 if (ret) {
 482                         bch_err_msg(c, ret, "getting btree node");
 483                         goto err;
 484                 }
 485
 486                 ret = bch2_btree_repair_topology_recurse(trans, cur);
 487                 six_unlock_read(&cur->c.lock);
 488                 cur = NULL;
 489
 490                 if (ret == DROP_THIS_NODE) {
 491                         bch2_btree_node_evict(trans, cur_k.k);
 492                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 493                                                       b->c.level, cur_k.k->k.p);
 494                         dropped_children = true;
 495                 }
 496
 497                 if (ret)
 498                         goto err;
 499
 500                 have_child = true;
 501         }
 502
 503         printbuf_reset(&buf);
 504         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 505
 506         if (mustfix_fsck_err_on(!have_child, c,
 507                         "empty interior btree node at btree %s level %u\n"
 508                         "  %s",
 509                         bch2_btree_ids[b->c.btree_id],
 510                         b->c.level, buf.buf))
 511                 ret = DROP_THIS_NODE;
 512 err:
 513 fsck_err:
 514         if (!IS_ERR_OR_NULL(prev))
 515                 six_unlock_read(&prev->c.lock);
 516         if (!IS_ERR_OR_NULL(cur))
 517                 six_unlock_read(&cur->c.lock);
 518
 519         bch2_btree_and_journal_iter_exit(&iter);
 520         bch2_bkey_buf_exit(&prev_k, c);
 521         bch2_bkey_buf_exit(&cur_k, c);
 522
 523         if (!ret && dropped_children)
 524                 goto again;
 525
 526         printbuf_exit(&buf);
 527         return ret;
 528 }
 529
 530 int bch2_check_topology(struct bch_fs *c)
 531 {
 532         struct btree_trans trans;
 533         struct btree *b;
 534         unsigned i;
 535         int ret = 0;
 536
 537         bch2_trans_init(&trans, c, 0, 0);
 538
 539         for (i = 0; i < btree_id_nr_alive(c)&& !ret; i++) {
 540                 struct btree_root *r = bch2_btree_id_root(c, i);
 541
 542                 if (!r->alive)
 543                         continue;
 544
 545                 b = r->b;
 546                 if (btree_node_fake(b))
 547                         continue;
 548
 549                 btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 550                 ret = bch2_btree_repair_topology_recurse(&trans, b);
 551                 six_unlock_read(&b->c.lock);
 552
 553                 if (ret == DROP_THIS_NODE) {
 554                         bch_err(c, "empty btree root - repair unimplemented");
 555                         ret = -BCH_ERR_fsck_repair_unimplemented;
 556                 }
 557         }
 558
 559         bch2_trans_exit(&trans);
 560
 561         return ret;
 562 }
 563
 564 static int bch2_check_fix_ptrs(struct btree_trans *trans, enum btree_id btree_id,
 565                                unsigned level, bool is_root,
 566                                struct bkey_s_c *k)
 567 {
 568         struct bch_fs *c = trans->c;
 569         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 570         const union bch_extent_entry *entry;
 571         struct extent_ptr_decoded p = { 0 };
 572         bool do_update = false;
 573         struct printbuf buf = PRINTBUF;
 574         int ret = 0;
 575
 576         /*
 577          * XXX
 578          * use check_bucket_ref here
 579          */
 580         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 581                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 582                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 583                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 584
 585                 if (!g->gen_valid &&
 586                     (c->opts.reconstruct_alloc ||
 587                      fsck_err(c, "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 588                               "while marking %s",
 589                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 590                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 591                               p.ptr.gen,
 592                               (printbuf_reset(&buf),
 593                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 594                         if (!p.ptr.cached) {
 595                                 g->gen_valid            = true;
 596                                 g->gen                  = p.ptr.gen;
 597                         } else {
 598                                 do_update = true;
 599                         }
 600                 }
 601
 602                 if (gen_cmp(p.ptr.gen, g->gen) > 0 &&
 603                     (c->opts.reconstruct_alloc ||
 604                      fsck_err(c, "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 605                               "while marking %s",
 606                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 607                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 608                               p.ptr.gen, g->gen,
 609                               (printbuf_reset(&buf),
 610                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))) {
 611                         if (!p.ptr.cached) {
 612                                 g->gen_valid            = true;
 613                                 g->gen                  = p.ptr.gen;
 614                                 g->data_type            = 0;
 615                                 g->dirty_sectors        = 0;
 616                                 g->cached_sectors       = 0;
 617                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 618                         } else {
 619                                 do_update = true;
 620                         }
 621                 }
 622
 623                 if (gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX &&
 624                     (c->opts.reconstruct_alloc ||
 625                      fsck_err(c, "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 626                               "while marking %s",
 627                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 628                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 629                               p.ptr.gen,
 630                               (printbuf_reset(&buf),
 631                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 632                         do_update = true;
 633
 634                 if (!p.ptr.cached && gen_cmp(p.ptr.gen, g->gen) < 0 &&
 635                     (c->opts.reconstruct_alloc ||
 636                      fsck_err(c, "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 637                               "while marking %s",
 638                               p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 639                               bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 640                               p.ptr.gen, g->gen,
 641                               (printbuf_reset(&buf),
 642                                bch2_bkey_val_to_text(&buf, c, *k), buf.buf))))
 643                         do_update = true;
 644
 645                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 646                         continue;
 647
 648                 if (fsck_err_on(bucket_data_type(g->data_type) &&
 649                                 bucket_data_type(g->data_type) != data_type, c,
 650                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 651                                 "while marking %s",
 652                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 653                                 bch2_data_types[g->data_type],
 654                                 bch2_data_types[data_type],
 655                                 (printbuf_reset(&buf),
 656                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 657                         if (data_type == BCH_DATA_btree) {
 658                                 g->data_type    = data_type;
 659                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 660                         } else {
 661                                 do_update = true;
 662                         }
 663                 }
 664
 665                 if (p.has_ec) {
 666                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 667
 668                         if (fsck_err_on(!m || !m->alive, c,
 669                                         "pointer to nonexistent stripe %llu\n"
 670                                         "while marking %s",
 671                                         (u64) p.ec.idx,
 672                                         (printbuf_reset(&buf),
 673                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 674                                 do_update = true;
 675
 676                         if (fsck_err_on(m && m->alive && !bch2_ptr_matches_stripe_m(m, p), c,
 677                                         "pointer does not match stripe %llu\n"
 678                                         "while marking %s",
 679                                         (u64) p.ec.idx,
 680                                         (printbuf_reset(&buf),
 681                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 682                                 do_update = true;
 683                 }
 684         }
 685
 686         if (do_update) {
 687                 struct bkey_ptrs ptrs;
 688                 union bch_extent_entry *entry;
 689                 struct bch_extent_ptr *ptr;
 690                 struct bkey_i *new;
 691
 692                 if (is_root) {
 693                         bch_err(c, "cannot update btree roots yet");
 694                         ret = -EINVAL;
 695                         goto err;
 696                 }
 697
 698                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 699                 if (!new) {
 700                         bch_err_msg(c, ret, "allocating new key");
 701                         ret = -BCH_ERR_ENOMEM_gc_repair_key;
 702                         goto err;
 703                 }
 704
 705                 bkey_reassemble(new, *k);
 706
 707                 if (level) {
 708                         /*
 709                          * We don't want to drop btree node pointers - if the
 710                          * btree node isn't there anymore, the read path will
 711                          * sort it out:
 712                          */
 713                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 714                         bkey_for_each_ptr(ptrs, ptr) {
 715                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 716                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 717
 718                                 ptr->gen = g->gen;
 719                         }
 720                 } else {
 721                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 722                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 723                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 724                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 725
 726                                 (ptr->cached &&
 727                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 728                                 (!ptr->cached &&
 729                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 730                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 731                                 (g->data_type &&
 732                                  g->data_type != data_type);
 733                         }));
 734 again:
 735                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 736                         bkey_extent_entry_for_each(ptrs, entry) {
 737                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 738                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 739                                                                         entry->stripe_ptr.idx);
 740                                         union bch_extent_entry *next_ptr;
 741
 742                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 743                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 744                                                         goto found;
 745                                         next_ptr = NULL;
 746 found:
 747                                         if (!next_ptr) {
 748                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 749                                                 continue;
 750                                         }
 751
 752                                         if (!m || !m->alive ||
 753                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 754                                                                        &next_ptr->ptr,
 755                                                                        m->sectors)) {
 756                                                 bch2_bkey_extent_entry_drop(new, entry);
 757                                                 goto again;
 758                                         }
 759                                 }
 760                         }
 761                 }
 762
 763                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 764                 if (ret) {
 765                         kfree(new);
 766                         goto err;
 767                 }
 768
 769                 if (level)
 770                         bch2_btree_node_update_key_early(trans, btree_id, level - 1, *k, new);
 771
 772                 if (0) {
 773                         printbuf_reset(&buf);
 774                         bch2_bkey_val_to_text(&buf, c, *k);
 775                         bch_info(c, "updated %s", buf.buf);
 776
 777                         printbuf_reset(&buf);
 778                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 779                         bch_info(c, "new key %s", buf.buf);
 780                 }
 781
 782                 *k = bkey_i_to_s_c(new);
 783         }
 784 err:
 785 fsck_err:
 786         printbuf_exit(&buf);
 787         return ret;
 788 }
 789
 790 /* marking of btree keys/nodes: */
 791
 792 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 793                             unsigned level, bool is_root,
 794                             struct bkey_s_c *k,
 795                             bool initial)
 796 {
 797         struct bch_fs *c = trans->c;
 798         struct bkey deleted = KEY(0, 0, 0);
 799         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 800         unsigned flags =
 801                 BTREE_TRIGGER_GC|
 802                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 803         int ret = 0;
 804
 805         deleted.p = k->k->p;
 806
 807         if (initial) {
 808                 BUG_ON(bch2_journal_seq_verify &&
 809                        k->k->version.lo > atomic64_read(&c->journal.seq));
 810
 811                 ret = bch2_check_fix_ptrs(trans, btree_id, level, is_root, k);
 812                 if (ret)
 813                         goto err;
 814
 815                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 816                                 "key version number higher than recorded: %llu > %llu",
 817                                 k->k->version.lo,
 818                                 atomic64_read(&c->key_version)))
 819                         atomic64_set(&c->key_version, k->k->version.lo);
 820         }
 821
 822         ret = commit_do(trans, NULL, NULL, 0,
 823                         bch2_mark_key(trans, btree_id, level, old, *k, flags));
 824 fsck_err:
 825 err:
 826         if (ret)
 827                 bch_err_fn(c, ret);
 828         return ret;
 829 }
 830
 831 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 832 {
 833         struct bch_fs *c = trans->c;
 834         struct btree_node_iter iter;
 835         struct bkey unpacked;
 836         struct bkey_s_c k;
 837         struct bkey_buf prev, cur;
 838         int ret = 0;
 839
 840         if (!btree_node_type_needs_gc(btree_node_type(b)))
 841                 return 0;
 842
 843         bch2_btree_node_iter_init_from_start(&iter, b);
 844         bch2_bkey_buf_init(&prev);
 845         bch2_bkey_buf_init(&cur);
 846         bkey_init(&prev.k->k);
 847
 848         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 849                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 850                                        &k, initial);
 851                 if (ret)
 852                         break;
 853
 854                 bch2_btree_node_iter_advance(&iter, b);
 855
 856                 if (b->c.level) {
 857                         bch2_bkey_buf_reassemble(&cur, c, k);
 858
 859                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 860                                         bch2_btree_node_iter_end(&iter));
 861                         if (ret)
 862                                 break;
 863                 }
 864         }
 865
 866         bch2_bkey_buf_exit(&cur, c);
 867         bch2_bkey_buf_exit(&prev, c);
 868         return ret;
 869 }
 870
 871 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 872                          bool initial, bool metadata_only)
 873 {
 874         struct bch_fs *c = trans->c;
 875         struct btree_iter iter;
 876         struct btree *b;
 877         unsigned depth = metadata_only ? 1 : 0;
 878         int ret = 0;
 879
 880         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 881
 882         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 883                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 884                 bch2_verify_btree_nr_keys(b);
 885
 886                 gc_pos_set(c, gc_pos_btree_node(b));
 887
 888                 ret = btree_gc_mark_node(trans, b, initial);
 889                 if (ret)
 890                         break;
 891         }
 892         bch2_trans_iter_exit(trans, &iter);
 893
 894         if (ret)
 895                 return ret;
 896
 897         mutex_lock(&c->btree_root_lock);
 898         b = bch2_btree_id_root(c, btree_id)->b;
 899         if (!btree_node_fake(b)) {
 900                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 901
 902                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1,
 903                                        true, &k, initial);
 904         }
 905         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 906         mutex_unlock(&c->btree_root_lock);
 907
 908         return ret;
 909 }
 910
 911 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 912                                       unsigned target_depth)
 913 {
 914         struct bch_fs *c = trans->c;
 915         struct btree_and_journal_iter iter;
 916         struct bkey_s_c k;
 917         struct bkey_buf cur, prev;
 918         struct printbuf buf = PRINTBUF;
 919         int ret = 0;
 920
 921         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 922         bch2_bkey_buf_init(&prev);
 923         bch2_bkey_buf_init(&cur);
 924         bkey_init(&prev.k->k);
 925
 926         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 927                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 928                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 929
 930                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 931                                        false, &k, true);
 932                 if (ret)
 933                         goto fsck_err;
 934
 935                 if (b->c.level) {
 936                         bch2_bkey_buf_reassemble(&cur, c, k);
 937                         k = bkey_i_to_s_c(cur.k);
 938
 939                         bch2_btree_and_journal_iter_advance(&iter);
 940
 941                         ret = bch2_gc_check_topology(c, b,
 942                                         &prev, cur,
 943                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 944                         if (ret)
 945                                 goto fsck_err;
 946                 } else {
 947                         bch2_btree_and_journal_iter_advance(&iter);
 948                 }
 949         }
 950
 951         if (b->c.level > target_depth) {
 952                 bch2_btree_and_journal_iter_exit(&iter);
 953                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 954
 955                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 956                         struct btree *child;
 957
 958                         bch2_bkey_buf_reassemble(&cur, c, k);
 959                         bch2_btree_and_journal_iter_advance(&iter);
 960
 961                         child = bch2_btree_node_get_noiter(trans, cur.k,
 962                                                 b->c.btree_id, b->c.level - 1,
 963                                                 false);
 964                         ret = PTR_ERR_OR_ZERO(child);
 965
 966                         if (ret == -EIO) {
 967                                 bch2_topology_error(c);
 968
 969                                 if (__fsck_err(c,
 970                                           FSCK_CAN_FIX|
 971                                           FSCK_CAN_IGNORE|
 972                                           FSCK_NO_RATELIMIT,
 973                                           "Unreadable btree node at btree %s level %u:\n"
 974                                           "  %s",
 975                                           bch2_btree_ids[b->c.btree_id],
 976                                           b->c.level - 1,
 977                                           (printbuf_reset(&buf),
 978                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 979                                     should_restart_for_topology_repair(c)) {
 980                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 981                                         ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 982                                         goto fsck_err;
 983                                 } else {
 984                                         /* Continue marking when opted to not
 985                                          * fix the error: */
 986                                         ret = 0;
 987                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 988                                         continue;
 989                                 }
 990                         } else if (ret) {
 991                                 bch_err_msg(c, ret, "getting btree node");
 992                                 break;
 993                         }
 994
 995                         ret = bch2_gc_btree_init_recurse(trans, child,
 996                                                          target_depth);
 997                         six_unlock_read(&child->c.lock);
 998
 999                         if (ret)
1000                                 break;
1001                 }
1002         }
1003 fsck_err:
1004         bch2_bkey_buf_exit(&cur, c);
1005         bch2_bkey_buf_exit(&prev, c);
1006         bch2_btree_and_journal_iter_exit(&iter);
1007         printbuf_exit(&buf);
1008         return ret;
1009 }
1010
1011 static int bch2_gc_btree_init(struct btree_trans *trans,
1012                               enum btree_id btree_id,
1013                               bool metadata_only)
1014 {
1015         struct bch_fs *c = trans->c;
1016         struct btree *b;
1017         unsigned target_depth = metadata_only ? 1 : 0;
1018         struct printbuf buf = PRINTBUF;
1019         int ret = 0;
1020
1021         b = bch2_btree_id_root(c, btree_id)->b;
1022
1023         if (btree_node_fake(b))
1024                 return 0;
1025
1026         six_lock_read(&b->c.lock, NULL, NULL);
1027         printbuf_reset(&buf);
1028         bch2_bpos_to_text(&buf, b->data->min_key);
1029         if (mustfix_fsck_err_on(!bpos_eq(b->data->min_key, POS_MIN), c,
1030                         "btree root with incorrect min_key: %s", buf.buf)) {
1031                 bch_err(c, "repair unimplemented");
1032                 ret = -BCH_ERR_fsck_repair_unimplemented;
1033                 goto fsck_err;
1034         }
1035
1036         printbuf_reset(&buf);
1037         bch2_bpos_to_text(&buf, b->data->max_key);
1038         if (mustfix_fsck_err_on(!bpos_eq(b->data->max_key, SPOS_MAX), c,
1039                         "btree root with incorrect max_key: %s", buf.buf)) {
1040                 bch_err(c, "repair unimplemented");
1041                 ret = -BCH_ERR_fsck_repair_unimplemented;
1042                 goto fsck_err;
1043         }
1044
1045         if (b->c.level >= target_depth)
1046                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1047
1048         if (!ret) {
1049                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1050
1051                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level + 1, true,
1052                                        &k, true);
1053         }
1054 fsck_err:
1055         six_unlock_read(&b->c.lock);
1056
1057         if (ret < 0)
1058                 bch_err_fn(c, ret);
1059         printbuf_exit(&buf);
1060         return ret;
1061 }
1062
1063 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1064 {
1065         return  (int) btree_id_to_gc_phase(l) -
1066                 (int) btree_id_to_gc_phase(r);
1067 }
1068
1069 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1070 {
1071         struct btree_trans trans;
1072         enum btree_id ids[BTREE_ID_NR];
1073         unsigned i;
1074         int ret = 0;
1075
1076         bch2_trans_init(&trans, c, 0, 0);
1077
1078         if (initial)
1079                 trans.is_initial_gc = true;
1080
1081         for (i = 0; i < BTREE_ID_NR; i++)
1082                 ids[i] = i;
1083         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1084
1085         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1086                 ret = initial
1087                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1088                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1089
1090         for (i = BTREE_ID_NR; i < btree_id_nr_alive(c) && !ret; i++) {
1091                 if (!bch2_btree_id_root(c, i)->alive)
1092                         continue;
1093
1094                 ret = initial
1095                         ? bch2_gc_btree_init(&trans, i, metadata_only)
1096                         : bch2_gc_btree(&trans, i, initial, metadata_only);
1097         }
1098
1099         if (ret < 0)
1100                 bch_err_fn(c, ret);
1101
1102         bch2_trans_exit(&trans);
1103         return ret;
1104 }
1105
1106 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1107                                   u64 start, u64 end,
1108                                   enum bch_data_type type,
1109                                   unsigned flags)
1110 {
1111         u64 b = sector_to_bucket(ca, start);
1112
1113         do {
1114                 unsigned sectors =
1115                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1116
1117                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1118                                           gc_phase(GC_PHASE_SB), flags);
1119                 b++;
1120                 start += sectors;
1121         } while (start < end);
1122 }
1123
1124 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1125                                      unsigned flags)
1126 {
1127         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1128         unsigned i;
1129         u64 b;
1130
1131         for (i = 0; i < layout->nr_superblocks; i++) {
1132                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1133
1134                 if (offset == BCH_SB_SECTOR)
1135                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1136                                               BCH_DATA_sb, flags);
1137
1138                 mark_metadata_sectors(c, ca, offset,
1139                                       offset + (1 << layout->sb_max_size_bits),
1140                                       BCH_DATA_sb, flags);
1141         }
1142
1143         for (i = 0; i < ca->journal.nr; i++) {
1144                 b = ca->journal.buckets[i];
1145                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1146                                           ca->mi.bucket_size,
1147                                           gc_phase(GC_PHASE_SB), flags);
1148         }
1149 }
1150
1151 static void bch2_mark_superblocks(struct bch_fs *c)
1152 {
1153         struct bch_dev *ca;
1154         unsigned i;
1155
1156         mutex_lock(&c->sb_lock);
1157         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1158
1159         for_each_online_member(ca, c, i)
1160                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1161         mutex_unlock(&c->sb_lock);
1162 }
1163
1164 #if 0
1165 /* Also see bch2_pending_btree_node_free_insert_done() */
1166 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1167 {
1168         struct btree_update *as;
1169         struct pending_btree_node_free *d;
1170
1171         mutex_lock(&c->btree_interior_update_lock);
1172         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1173
1174         for_each_pending_btree_node_free(c, as, d)
1175                 if (d->index_update_done)
1176                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1177
1178         mutex_unlock(&c->btree_interior_update_lock);
1179 }
1180 #endif
1181
1182 static void bch2_gc_free(struct bch_fs *c)
1183 {
1184         struct bch_dev *ca;
1185         unsigned i;
1186
1187         genradix_free(&c->reflink_gc_table);
1188         genradix_free(&c->gc_stripes);
1189
1190         for_each_member_device(ca, c, i) {
1191                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1192                         sizeof(struct bucket_array) +
1193                         ca->mi.nbuckets * sizeof(struct bucket));
1194                 ca->buckets_gc = NULL;
1195
1196                 free_percpu(ca->usage_gc);
1197                 ca->usage_gc = NULL;
1198         }
1199
1200         free_percpu(c->usage_gc);
1201         c->usage_gc = NULL;
1202 }
1203
1204 static int bch2_gc_done(struct bch_fs *c,
1205                         bool initial, bool metadata_only)
1206 {
1207         struct bch_dev *ca = NULL;
1208         struct printbuf buf = PRINTBUF;
1209         bool verify = !metadata_only &&
1210                 !c->opts.reconstruct_alloc &&
1211                 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1212         unsigned i, dev;
1213         int ret = 0;
1214
1215         percpu_down_write(&c->mark_lock);
1216
1217 #define copy_field(_f, _msg, ...)                                       \
1218         if (dst->_f != src->_f &&                                       \
1219             (!verify ||                                                 \
1220              fsck_err(c, _msg ": got %llu, should be %llu"              \
1221                       , ##__VA_ARGS__, dst->_f, src->_f)))              \
1222                 dst->_f = src->_f
1223 #define copy_stripe_field(_f, _msg, ...)                                \
1224         if (dst->_f != src->_f &&                                       \
1225             (!verify ||                                                 \
1226              fsck_err(c, "stripe %zu has wrong "_msg                    \
1227                       ": got %u, should be %u",                         \
1228                       iter.pos, ##__VA_ARGS__,                          \
1229                       dst->_f, src->_f)))                               \
1230                 dst->_f = src->_f
1231 #define copy_dev_field(_f, _msg, ...)                                   \
1232         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1233 #define copy_fs_field(_f, _msg, ...)                                    \
1234         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1235
1236         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1237                 bch2_fs_usage_acc_to_base(c, i);
1238
1239         for_each_member_device(ca, c, dev) {
1240                 struct bch_dev_usage *dst = ca->usage_base;
1241                 struct bch_dev_usage *src = (void *)
1242                         bch2_acc_percpu_u64s((u64 __percpu *) ca->usage_gc,
1243                                              dev_usage_u64s());
1244
1245                 copy_dev_field(buckets_ec,              "buckets_ec");
1246
1247                 for (i = 0; i < BCH_DATA_NR; i++) {
1248                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1249                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1250                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1251                 }
1252         };
1253
1254         {
1255                 unsigned nr = fs_usage_u64s(c);
1256                 struct bch_fs_usage *dst = c->usage_base;
1257                 struct bch_fs_usage *src = (void *)
1258                         bch2_acc_percpu_u64s((u64 __percpu *) c->usage_gc, nr);
1259
1260                 copy_fs_field(hidden,           "hidden");
1261                 copy_fs_field(btree,            "btree");
1262
1263                 if (!metadata_only) {
1264                         copy_fs_field(data,     "data");
1265                         copy_fs_field(cached,   "cached");
1266                         copy_fs_field(reserved, "reserved");
1267                         copy_fs_field(nr_inodes,"nr_inodes");
1268
1269                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1270                                 copy_fs_field(persistent_reserved[i],
1271                                               "persistent_reserved[%i]", i);
1272                 }
1273
1274                 for (i = 0; i < c->replicas.nr; i++) {
1275                         struct bch_replicas_entry *e =
1276                                 cpu_replicas_entry(&c->replicas, i);
1277
1278                         if (metadata_only &&
1279                             (e->data_type == BCH_DATA_user ||
1280                              e->data_type == BCH_DATA_cached))
1281                                 continue;
1282
1283                         printbuf_reset(&buf);
1284                         bch2_replicas_entry_to_text(&buf, e);
1285
1286                         copy_fs_field(replicas[i], "%s", buf.buf);
1287                 }
1288         }
1289
1290 #undef copy_fs_field
1291 #undef copy_dev_field
1292 #undef copy_stripe_field
1293 #undef copy_field
1294 fsck_err:
1295         if (ca)
1296                 percpu_ref_put(&ca->ref);
1297         if (ret)
1298                 bch_err_fn(c, ret);
1299
1300         percpu_up_write(&c->mark_lock);
1301         printbuf_exit(&buf);
1302         return ret;
1303 }
1304
1305 static int bch2_gc_start(struct bch_fs *c)
1306 {
1307         struct bch_dev *ca = NULL;
1308         unsigned i;
1309
1310         BUG_ON(c->usage_gc);
1311
1312         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1313                                          sizeof(u64), GFP_KERNEL);
1314         if (!c->usage_gc) {
1315                 bch_err(c, "error allocating c->usage_gc");
1316                 return -BCH_ERR_ENOMEM_gc_start;
1317         }
1318
1319         for_each_member_device(ca, c, i) {
1320                 BUG_ON(ca->usage_gc);
1321
1322                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1323                 if (!ca->usage_gc) {
1324                         bch_err(c, "error allocating ca->usage_gc");
1325                         percpu_ref_put(&ca->ref);
1326                         return -BCH_ERR_ENOMEM_gc_start;
1327                 }
1328
1329                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1330                                ca->mi.nbuckets - ca->mi.first_bucket);
1331         }
1332
1333         return 0;
1334 }
1335
1336 static int bch2_gc_reset(struct bch_fs *c)
1337 {
1338         struct bch_dev *ca;
1339         unsigned i;
1340
1341         for_each_member_device(ca, c, i) {
1342                 free_percpu(ca->usage_gc);
1343                 ca->usage_gc = NULL;
1344         }
1345
1346         free_percpu(c->usage_gc);
1347         c->usage_gc = NULL;
1348
1349         return bch2_gc_start(c);
1350 }
1351
1352 /* returns true if not equal */
1353 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1354                                      struct bch_alloc_v4 r)
1355 {
1356         return  l.gen != r.gen                          ||
1357                 l.oldest_gen != r.oldest_gen            ||
1358                 l.data_type != r.data_type              ||
1359                 l.dirty_sectors != r.dirty_sectors      ||
1360                 l.cached_sectors != r.cached_sectors     ||
1361                 l.stripe_redundancy != r.stripe_redundancy ||
1362                 l.stripe != r.stripe;
1363 }
1364
1365 static int bch2_alloc_write_key(struct btree_trans *trans,
1366                                 struct btree_iter *iter,
1367                                 struct bkey_s_c k,
1368                                 bool metadata_only)
1369 {
1370         struct bch_fs *c = trans->c;
1371         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1372         struct bucket gc, *b;
1373         struct bkey_i_alloc_v4 *a;
1374         struct bch_alloc_v4 old_convert, new;
1375         const struct bch_alloc_v4 *old;
1376         enum bch_data_type type;
1377         int ret;
1378
1379         if (bkey_ge(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)))
1380                 return 1;
1381
1382         old = bch2_alloc_to_v4(k, &old_convert);
1383         new = *old;
1384
1385         percpu_down_read(&c->mark_lock);
1386         b = gc_bucket(ca, iter->pos.offset);
1387
1388         /*
1389          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1390          * fix that here:
1391          */
1392         type = __alloc_data_type(b->dirty_sectors,
1393                                  b->cached_sectors,
1394                                  b->stripe,
1395                                  *old,
1396                                  b->data_type);
1397         if (b->data_type != type) {
1398                 struct bch_dev_usage *u;
1399
1400                 preempt_disable();
1401                 u = this_cpu_ptr(ca->usage_gc);
1402                 u->d[b->data_type].buckets--;
1403                 b->data_type = type;
1404                 u->d[b->data_type].buckets++;
1405                 preempt_enable();
1406         }
1407
1408         gc = *b;
1409         percpu_up_read(&c->mark_lock);
1410
1411         if (metadata_only &&
1412             gc.data_type != BCH_DATA_sb &&
1413             gc.data_type != BCH_DATA_journal &&
1414             gc.data_type != BCH_DATA_btree)
1415                 return 0;
1416
1417         if (gen_after(old->gen, gc.gen))
1418                 return 0;
1419
1420         if (c->opts.reconstruct_alloc ||
1421             fsck_err_on(new.data_type != gc.data_type, c,
1422                         "bucket %llu:%llu gen %u has wrong data_type"
1423                         ": got %s, should be %s",
1424                         iter->pos.inode, iter->pos.offset,
1425                         gc.gen,
1426                         bch2_data_types[new.data_type],
1427                         bch2_data_types[gc.data_type]))
1428                 new.data_type = gc.data_type;
1429
1430 #define copy_bucket_field(_f)                                           \
1431         if (c->opts.reconstruct_alloc ||                                \
1432             fsck_err_on(new._f != gc._f, c,                             \
1433                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1434                         ": got %u, should be %u",                       \
1435                         iter->pos.inode, iter->pos.offset,              \
1436                         gc.gen,                                         \
1437                         bch2_data_types[gc.data_type],                  \
1438                         new._f, gc._f))                                 \
1439                 new._f = gc._f;                                         \
1440
1441         copy_bucket_field(gen);
1442         copy_bucket_field(dirty_sectors);
1443         copy_bucket_field(cached_sectors);
1444         copy_bucket_field(stripe_redundancy);
1445         copy_bucket_field(stripe);
1446 #undef copy_bucket_field
1447
1448         if (!bch2_alloc_v4_cmp(*old, new))
1449                 return 0;
1450
1451         a = bch2_alloc_to_v4_mut(trans, k);
1452         ret = PTR_ERR_OR_ZERO(a);
1453         if (ret)
1454                 return ret;
1455
1456         a->v = new;
1457
1458         /*
1459          * The trigger normally makes sure this is set, but we're not running
1460          * triggers:
1461          */
1462         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1463                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1464
1465         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1466 fsck_err:
1467         return ret;
1468 }
1469
1470 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1471 {
1472         struct btree_trans trans;
1473         struct btree_iter iter;
1474         struct bkey_s_c k;
1475         struct bch_dev *ca;
1476         unsigned i;
1477         int ret = 0;
1478
1479         bch2_trans_init(&trans, c, 0, 0);
1480
1481         for_each_member_device(ca, c, i) {
1482                 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1483                                 POS(ca->dev_idx, ca->mi.first_bucket),
1484                                 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1485                                 NULL, NULL, BTREE_INSERT_LAZY_RW,
1486                         bch2_alloc_write_key(&trans, &iter, k, metadata_only));
1487
1488                 if (ret < 0) {
1489                         bch_err(c, "error writing alloc info: %s", bch2_err_str(ret));
1490                         percpu_ref_put(&ca->ref);
1491                         break;
1492                 }
1493         }
1494
1495         bch2_trans_exit(&trans);
1496         return ret < 0 ? ret : 0;
1497 }
1498
1499 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1500 {
1501         struct bch_dev *ca;
1502         struct btree_trans trans;
1503         struct btree_iter iter;
1504         struct bkey_s_c k;
1505         struct bucket *g;
1506         struct bch_alloc_v4 a_convert;
1507         const struct bch_alloc_v4 *a;
1508         unsigned i;
1509         int ret;
1510
1511         for_each_member_device(ca, c, i) {
1512                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1513                                 ca->mi.nbuckets * sizeof(struct bucket),
1514                                 GFP_KERNEL|__GFP_ZERO);
1515                 if (!buckets) {
1516                         percpu_ref_put(&ca->ref);
1517                         bch_err(c, "error allocating ca->buckets[gc]");
1518                         return -BCH_ERR_ENOMEM_gc_alloc_start;
1519                 }
1520
1521                 buckets->first_bucket   = ca->mi.first_bucket;
1522                 buckets->nbuckets       = ca->mi.nbuckets;
1523                 rcu_assign_pointer(ca->buckets_gc, buckets);
1524         };
1525
1526         bch2_trans_init(&trans, c, 0, 0);
1527
1528         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1529                            BTREE_ITER_PREFETCH, k, ret) {
1530                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1531                 g = gc_bucket(ca, k.k->p.offset);
1532
1533                 a = bch2_alloc_to_v4(k, &a_convert);
1534
1535                 g->gen_valid    = 1;
1536                 g->gen          = a->gen;
1537
1538                 if (metadata_only &&
1539                     (a->data_type == BCH_DATA_user ||
1540                      a->data_type == BCH_DATA_cached ||
1541                      a->data_type == BCH_DATA_parity)) {
1542                         g->data_type            = a->data_type;
1543                         g->dirty_sectors        = a->dirty_sectors;
1544                         g->cached_sectors       = a->cached_sectors;
1545                         g->stripe               = a->stripe;
1546                         g->stripe_redundancy    = a->stripe_redundancy;
1547                 }
1548         }
1549         bch2_trans_iter_exit(&trans, &iter);
1550
1551         bch2_trans_exit(&trans);
1552
1553         if (ret)
1554                 bch_err(c, "error reading alloc info at gc start: %s", bch2_err_str(ret));
1555
1556         return ret;
1557 }
1558
1559 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1560 {
1561         struct bch_dev *ca;
1562         unsigned i;
1563
1564         for_each_member_device(ca, c, i) {
1565                 struct bucket_array *buckets = gc_bucket_array(ca);
1566                 struct bucket *g;
1567
1568                 for_each_bucket(g, buckets) {
1569                         if (metadata_only &&
1570                             (g->data_type == BCH_DATA_user ||
1571                              g->data_type == BCH_DATA_cached ||
1572                              g->data_type == BCH_DATA_parity))
1573                                 continue;
1574                         g->data_type = 0;
1575                         g->dirty_sectors = 0;
1576                         g->cached_sectors = 0;
1577                 }
1578         };
1579 }
1580
1581 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1582                                      struct btree_iter *iter,
1583                                      struct bkey_s_c k,
1584                                      size_t *idx)
1585 {
1586         struct bch_fs *c = trans->c;
1587         const __le64 *refcount = bkey_refcount_c(k);
1588         struct printbuf buf = PRINTBUF;
1589         struct reflink_gc *r;
1590         int ret = 0;
1591
1592         if (!refcount)
1593                 return 0;
1594
1595         while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1596                r->offset < k.k->p.offset)
1597                 ++*idx;
1598
1599         if (!r ||
1600             r->offset != k.k->p.offset ||
1601             r->size != k.k->size) {
1602                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1603                 return -EINVAL;
1604         }
1605
1606         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1607                         "reflink key has wrong refcount:\n"
1608                         "  %s\n"
1609                         "  should be %u",
1610                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1611                         r->refcount)) {
1612                 struct bkey_i *new = bch2_bkey_make_mut(trans, iter, &k, 0);
1613
1614                 ret = PTR_ERR_OR_ZERO(new);
1615                 if (ret)
1616                         return ret;
1617
1618                 if (!r->refcount)
1619                         new->k.type = KEY_TYPE_deleted;
1620                 else
1621                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1622         }
1623 fsck_err:
1624         printbuf_exit(&buf);
1625         return ret;
1626 }
1627
1628 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1629 {
1630         struct btree_trans trans;
1631         struct btree_iter iter;
1632         struct bkey_s_c k;
1633         size_t idx = 0;
1634         int ret = 0;
1635
1636         if (metadata_only)
1637                 return 0;
1638
1639         bch2_trans_init(&trans, c, 0, 0);
1640
1641         ret = for_each_btree_key_commit(&trans, iter,
1642                         BTREE_ID_reflink, POS_MIN,
1643                         BTREE_ITER_PREFETCH, k,
1644                         NULL, NULL, BTREE_INSERT_NOFAIL,
1645                 bch2_gc_write_reflink_key(&trans, &iter, k, &idx));
1646
1647         c->reflink_gc_nr = 0;
1648         bch2_trans_exit(&trans);
1649         return ret;
1650 }
1651
1652 static int bch2_gc_reflink_start(struct bch_fs *c,
1653                                  bool metadata_only)
1654 {
1655         struct btree_trans trans;
1656         struct btree_iter iter;
1657         struct bkey_s_c k;
1658         struct reflink_gc *r;
1659         int ret = 0;
1660
1661         if (metadata_only)
1662                 return 0;
1663
1664         bch2_trans_init(&trans, c, 0, 0);
1665         c->reflink_gc_nr = 0;
1666
1667         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1668                            BTREE_ITER_PREFETCH, k, ret) {
1669                 const __le64 *refcount = bkey_refcount_c(k);
1670
1671                 if (!refcount)
1672                         continue;
1673
1674                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1675                                        GFP_KERNEL);
1676                 if (!r) {
1677                         ret = -BCH_ERR_ENOMEM_gc_reflink_start;
1678                         break;
1679                 }
1680
1681                 r->offset       = k.k->p.offset;
1682                 r->size         = k.k->size;
1683                 r->refcount     = 0;
1684         }
1685         bch2_trans_iter_exit(&trans, &iter);
1686
1687         bch2_trans_exit(&trans);
1688         return ret;
1689 }
1690
1691 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1692 {
1693         struct genradix_iter iter;
1694         struct reflink_gc *r;
1695
1696         genradix_for_each(&c->reflink_gc_table, iter, r)
1697                 r->refcount = 0;
1698 }
1699
1700 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1701                                      struct btree_iter *iter,
1702                                      struct bkey_s_c k)
1703 {
1704         struct bch_fs *c = trans->c;
1705         struct printbuf buf = PRINTBUF;
1706         const struct bch_stripe *s;
1707         struct gc_stripe *m;
1708         bool bad = false;
1709         unsigned i;
1710         int ret = 0;
1711
1712         if (k.k->type != KEY_TYPE_stripe)
1713                 return 0;
1714
1715         s = bkey_s_c_to_stripe(k).v;
1716         m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1717
1718         for (i = 0; i < s->nr_blocks; i++) {
1719                 u32 old = stripe_blockcount_get(s, i);
1720                 u32 new = (m ? m->block_sectors[i] : 0);
1721
1722                 if (old != new) {
1723                         prt_printf(&buf, "stripe block %u has wrong sector count: got %u, should be %u\n",
1724                                    i, old, new);
1725                         bad = true;
1726                 }
1727         }
1728
1729         if (bad)
1730                 bch2_bkey_val_to_text(&buf, c, k);
1731
1732         if (fsck_err_on(bad, c, "%s", buf.buf)) {
1733                 struct bkey_i_stripe *new;
1734
1735                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1736                 ret = PTR_ERR_OR_ZERO(new);
1737                 if (ret)
1738                         return ret;
1739
1740                 bkey_reassemble(&new->k_i, k);
1741
1742                 for (i = 0; i < new->v.nr_blocks; i++)
1743                         stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1744
1745                 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1746         }
1747 fsck_err:
1748         printbuf_exit(&buf);
1749         return ret;
1750 }
1751
1752 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1753 {
1754         struct btree_trans trans;
1755         struct btree_iter iter;
1756         struct bkey_s_c k;
1757         int ret = 0;
1758
1759         if (metadata_only)
1760                 return 0;
1761
1762         bch2_trans_init(&trans, c, 0, 0);
1763
1764         ret = for_each_btree_key_commit(&trans, iter,
1765                         BTREE_ID_stripes, POS_MIN,
1766                         BTREE_ITER_PREFETCH, k,
1767                         NULL, NULL, BTREE_INSERT_NOFAIL,
1768                 bch2_gc_write_stripes_key(&trans, &iter, k));
1769
1770         bch2_trans_exit(&trans);
1771         return ret;
1772 }
1773
1774 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1775 {
1776         genradix_free(&c->gc_stripes);
1777 }
1778
1779 /**
1780  * bch2_gc - walk _all_ references to buckets, and recompute them:
1781  *
1782  * Order matters here:
1783  *  - Concurrent GC relies on the fact that we have a total ordering for
1784  *    everything that GC walks - see  gc_will_visit_node(),
1785  *    gc_will_visit_root()
1786  *
1787  *  - also, references move around in the course of index updates and
1788  *    various other crap: everything needs to agree on the ordering
1789  *    references are allowed to move around in - e.g., we're allowed to
1790  *    start with a reference owned by an open_bucket (the allocator) and
1791  *    move it to the btree, but not the reverse.
1792  *
1793  *    This is necessary to ensure that gc doesn't miss references that
1794  *    move around - if references move backwards in the ordering GC
1795  *    uses, GC could skip past them
1796  */
1797 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1798 {
1799         unsigned iter = 0;
1800         int ret;
1801
1802         lockdep_assert_held(&c->state_lock);
1803
1804         down_write(&c->gc_lock);
1805
1806         bch2_btree_interior_updates_flush(c);
1807
1808         ret   = bch2_gc_start(c) ?:
1809                 bch2_gc_alloc_start(c, metadata_only) ?:
1810                 bch2_gc_reflink_start(c, metadata_only);
1811         if (ret)
1812                 goto out;
1813 again:
1814         gc_pos_set(c, gc_phase(GC_PHASE_START));
1815
1816         bch2_mark_superblocks(c);
1817
1818         ret = bch2_gc_btrees(c, initial, metadata_only);
1819
1820         if (ret)
1821                 goto out;
1822
1823 #if 0
1824         bch2_mark_pending_btree_node_frees(c);
1825 #endif
1826         c->gc_count++;
1827
1828         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1829             (!iter && bch2_test_restart_gc)) {
1830                 if (iter++ > 2) {
1831                         bch_info(c, "Unable to fix bucket gens, looping");
1832                         ret = -EINVAL;
1833                         goto out;
1834                 }
1835
1836                 /*
1837                  * XXX: make sure gens we fixed got saved
1838                  */
1839                 bch_info(c, "Second GC pass needed, restarting:");
1840                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1841                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1842
1843                 bch2_gc_stripes_reset(c, metadata_only);
1844                 bch2_gc_alloc_reset(c, metadata_only);
1845                 bch2_gc_reflink_reset(c, metadata_only);
1846                 ret = bch2_gc_reset(c);
1847                 if (ret)
1848                         goto out;
1849
1850                 /* flush fsck errors, reset counters */
1851                 bch2_flush_fsck_errs(c);
1852                 goto again;
1853         }
1854 out:
1855         if (!ret) {
1856                 bch2_journal_block(&c->journal);
1857
1858                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1859                         bch2_gc_reflink_done(c, metadata_only) ?:
1860                         bch2_gc_alloc_done(c, metadata_only) ?:
1861                         bch2_gc_done(c, initial, metadata_only);
1862
1863                 bch2_journal_unblock(&c->journal);
1864         }
1865
1866         percpu_down_write(&c->mark_lock);
1867         /* Indicates that gc is no longer in progress: */
1868         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1869
1870         bch2_gc_free(c);
1871         percpu_up_write(&c->mark_lock);
1872
1873         up_write(&c->gc_lock);
1874
1875         /*
1876          * At startup, allocations can happen directly instead of via the
1877          * allocator thread - issue wakeup in case they blocked on gc_lock:
1878          */
1879         closure_wake_up(&c->freelist_wait);
1880
1881         if (ret)
1882                 bch_err_fn(c, ret);
1883         return ret;
1884 }
1885
1886 static int gc_btree_gens_key(struct btree_trans *trans,
1887                              struct btree_iter *iter,
1888                              struct bkey_s_c k)
1889 {
1890         struct bch_fs *c = trans->c;
1891         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1892         const struct bch_extent_ptr *ptr;
1893         struct bkey_i *u;
1894         int ret;
1895
1896         percpu_down_read(&c->mark_lock);
1897         bkey_for_each_ptr(ptrs, ptr) {
1898                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1899
1900                 if (ptr_stale(ca, ptr) > 16) {
1901                         percpu_up_read(&c->mark_lock);
1902                         goto update;
1903                 }
1904         }
1905
1906         bkey_for_each_ptr(ptrs, ptr) {
1907                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1908                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1909
1910                 if (gen_after(*gen, ptr->gen))
1911                         *gen = ptr->gen;
1912         }
1913         percpu_up_read(&c->mark_lock);
1914         return 0;
1915 update:
1916         u = bch2_bkey_make_mut(trans, iter, &k, 0);
1917         ret = PTR_ERR_OR_ZERO(u);
1918         if (ret)
1919                 return ret;
1920
1921         bch2_extent_normalize(c, bkey_i_to_s(u));
1922         return 0;
1923 }
1924
1925 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1926                                        struct bkey_s_c k)
1927 {
1928         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1929         struct bch_alloc_v4 a_convert;
1930         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1931         struct bkey_i_alloc_v4 *a_mut;
1932         int ret;
1933
1934         if (a->oldest_gen == ca->oldest_gen[iter->pos.offset])
1935                 return 0;
1936
1937         a_mut = bch2_alloc_to_v4_mut(trans, k);
1938         ret = PTR_ERR_OR_ZERO(a_mut);
1939         if (ret)
1940                 return ret;
1941
1942         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1943         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1944
1945         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1946 }
1947
1948 int bch2_gc_gens(struct bch_fs *c)
1949 {
1950         struct btree_trans trans;
1951         struct btree_iter iter;
1952         struct bkey_s_c k;
1953         struct bch_dev *ca;
1954         u64 b, start_time = local_clock();
1955         unsigned i;
1956         int ret;
1957
1958         /*
1959          * Ideally we would be using state_lock and not gc_lock here, but that
1960          * introduces a deadlock in the RO path - we currently take the state
1961          * lock at the start of going RO, thus the gc thread may get stuck:
1962          */
1963         if (!mutex_trylock(&c->gc_gens_lock))
1964                 return 0;
1965
1966         trace_and_count(c, gc_gens_start, c);
1967         down_read(&c->gc_lock);
1968         bch2_trans_init(&trans, c, 0, 0);
1969
1970         for_each_member_device(ca, c, i) {
1971                 struct bucket_gens *gens;
1972
1973                 BUG_ON(ca->oldest_gen);
1974
1975                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1976                 if (!ca->oldest_gen) {
1977                         percpu_ref_put(&ca->ref);
1978                         ret = -BCH_ERR_ENOMEM_gc_gens;
1979                         goto err;
1980                 }
1981
1982                 gens = bucket_gens(ca);
1983
1984                 for (b = gens->first_bucket;
1985                      b < gens->nbuckets; b++)
1986                         ca->oldest_gen[b] = gens->b[b];
1987         }
1988
1989         for (i = 0; i < BTREE_ID_NR; i++)
1990                 if (btree_type_has_ptrs(i)) {
1991                         struct btree_iter iter;
1992                         struct bkey_s_c k;
1993
1994                         c->gc_gens_btree = i;
1995                         c->gc_gens_pos = POS_MIN;
1996                         ret = for_each_btree_key_commit(&trans, iter, i,
1997                                         POS_MIN,
1998                                         BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1999                                         k,
2000                                         NULL, NULL,
2001                                         BTREE_INSERT_NOFAIL,
2002                                 gc_btree_gens_key(&trans, &iter, k));
2003                         if (ret && !bch2_err_matches(ret, EROFS))
2004                                 bch_err(c, "error recalculating oldest_gen: %s", bch2_err_str(ret));
2005                         if (ret)
2006                                 goto err;
2007                 }
2008
2009         ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
2010                         POS_MIN,
2011                         BTREE_ITER_PREFETCH,
2012                         k,
2013                         NULL, NULL,
2014                         BTREE_INSERT_NOFAIL,
2015                 bch2_alloc_write_oldest_gen(&trans, &iter, k));
2016         if (ret && !bch2_err_matches(ret, EROFS))
2017                 bch_err(c, "error writing oldest_gen: %s", bch2_err_str(ret));
2018         if (ret)
2019                 goto err;
2020
2021         c->gc_gens_btree        = 0;
2022         c->gc_gens_pos          = POS_MIN;
2023
2024         c->gc_count++;
2025
2026         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2027         trace_and_count(c, gc_gens_end, c);
2028 err:
2029         for_each_member_device(ca, c, i) {
2030                 kvfree(ca->oldest_gen);
2031                 ca->oldest_gen = NULL;
2032         }
2033
2034         bch2_trans_exit(&trans);
2035         up_read(&c->gc_lock);
2036         mutex_unlock(&c->gc_gens_lock);
2037         return ret;
2038 }
2039
2040 static int bch2_gc_thread(void *arg)
2041 {
2042         struct bch_fs *c = arg;
2043         struct io_clock *clock = &c->io_clock[WRITE];
2044         unsigned long last = atomic64_read(&clock->now);
2045         unsigned last_kick = atomic_read(&c->kick_gc);
2046         int ret;
2047
2048         set_freezable();
2049
2050         while (1) {
2051                 while (1) {
2052                         set_current_state(TASK_INTERRUPTIBLE);
2053
2054                         if (kthread_should_stop()) {
2055                                 __set_current_state(TASK_RUNNING);
2056                                 return 0;
2057                         }
2058
2059                         if (atomic_read(&c->kick_gc) != last_kick)
2060                                 break;
2061
2062                         if (c->btree_gc_periodic) {
2063                                 unsigned long next = last + c->capacity / 16;
2064
2065                                 if (atomic64_read(&clock->now) >= next)
2066                                         break;
2067
2068                                 bch2_io_clock_schedule_timeout(clock, next);
2069                         } else {
2070                                 schedule();
2071                         }
2072
2073                         try_to_freeze();
2074                 }
2075                 __set_current_state(TASK_RUNNING);
2076
2077                 last = atomic64_read(&clock->now);
2078                 last_kick = atomic_read(&c->kick_gc);
2079
2080                 /*
2081                  * Full gc is currently incompatible with btree key cache:
2082                  */
2083 #if 0
2084                 ret = bch2_gc(c, false, false);
2085 #else
2086                 ret = bch2_gc_gens(c);
2087 #endif
2088                 if (ret < 0)
2089                         bch_err(c, "btree gc failed: %s", bch2_err_str(ret));
2090
2091                 debug_check_no_locks_held();
2092         }
2093
2094         return 0;
2095 }
2096
2097 void bch2_gc_thread_stop(struct bch_fs *c)
2098 {
2099         struct task_struct *p;
2100
2101         p = c->gc_thread;
2102         c->gc_thread = NULL;
2103
2104         if (p) {
2105                 kthread_stop(p);
2106                 put_task_struct(p);
2107         }
2108 }
2109
2110 int bch2_gc_thread_start(struct bch_fs *c)
2111 {
2112         struct task_struct *p;
2113
2114         if (c->gc_thread)
2115                 return 0;
2116
2117         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2118         if (IS_ERR(p)) {
2119                 bch_err(c, "error creating gc thread: %s", bch2_err_str(PTR_ERR(p)));
2120                 return PTR_ERR(p);
2121         }
2122
2123         get_task_struct(p);
2124         c->gc_thread = p;
2125         wake_up_process(p);
2126         return 0;
2127 }