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

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