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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  * Copyright (C) 2014 Datera Inc.
   5  */
   6
   7 #include "bcachefs.h"
   8 #include "alloc_background.h"
   9 #include "alloc_foreground.h"
  10 #include "bkey_methods.h"
  11 #include "bkey_buf.h"
  12 #include "btree_key_cache.h"
  13 #include "btree_locking.h"
  14 #include "btree_update_interior.h"
  15 #include "btree_io.h"
  16 #include "btree_gc.h"
  17 #include "buckets.h"
  18 #include "clock.h"
  19 #include "debug.h"
  20 #include "ec.h"
  21 #include "error.h"
  22 #include "extents.h"
  23 #include "journal.h"
  24 #include "keylist.h"
  25 #include "move.h"
  26 #include "recovery.h"
  27 #include "reflink.h"
  28 #include "replicas.h"
  29 #include "super-io.h"
  30
  31 #include <linux/slab.h>
  32 #include <linux/bitops.h>
  33 #include <linux/freezer.h>
  34 #include <linux/kthread.h>
  35 #include <linux/preempt.h>
  36 #include <linux/rcupdate.h>
  37 #include <linux/sched/task.h>
  38 #include <trace/events/bcachefs.h>
  39
  40 #define DROP_THIS_NODE          10
  41 #define DROP_PREV_NODE          11
  42
  43 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  44 {
  45         preempt_disable();
  46         write_seqcount_begin(&c->gc_pos_lock);
  47         c->gc_pos = new_pos;
  48         write_seqcount_end(&c->gc_pos_lock);
  49         preempt_enable();
  50 }
  51
  52 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  53 {
  54         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  55         __gc_pos_set(c, new_pos);
  56 }
  57
  58 /*
  59  * Missing: if an interior btree node is empty, we need to do something -
  60  * perhaps just kill it
  61  */
  62 static int bch2_gc_check_topology(struct bch_fs *c,
  63                                   struct btree *b,
  64                                   struct bkey_buf *prev,
  65                                   struct bkey_buf cur,
  66                                   bool is_last)
  67 {
  68         struct bpos node_start  = b->data->min_key;
  69         struct bpos node_end    = b->data->max_key;
  70         struct bpos expected_start = bkey_deleted(&prev->k->k)
  71                 ? node_start
  72                 : bpos_successor(prev->k->k.p);
  73         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  74         int ret = 0;
  75
  76         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  77                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  78
  79                 if (bpos_cmp(expected_start, bp->v.min_key)) {
  80                         bch2_topology_error(c);
  81
  82                         if (bkey_deleted(&prev->k->k)) {
  83                                 prt_printf(&buf1, "start of node: ");
  84                                 bch2_bpos_to_text(&buf1, node_start);
  85                         } else {
  86                                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
  87                         }
  88                         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
  89
  90                         if (__fsck_err(c,
  91                                   FSCK_CAN_FIX|
  92                                   FSCK_CAN_IGNORE|
  93                                   FSCK_NO_RATELIMIT,
  94                                   "btree node with incorrect min_key at btree %s level %u:\n"
  95                                   "  prev %s\n"
  96                                   "  cur %s",
  97                                   bch2_btree_ids[b->c.btree_id], b->c.level,
  98                                   buf1.buf, buf2.buf) &&
  99                             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 100                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 101                                 ret = -BCH_ERR_need_topology_repair;
 102                                 goto err;
 103                         } else {
 104                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 105                         }
 106                 }
 107         }
 108
 109         if (is_last && bpos_cmp(cur.k->k.p, node_end)) {
 110                 bch2_topology_error(c);
 111
 112                 printbuf_reset(&buf1);
 113                 printbuf_reset(&buf2);
 114
 115                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
 116                 bch2_bpos_to_text(&buf2, node_end);
 117
 118                 if (__fsck_err(c,
 119                           FSCK_CAN_FIX|
 120                           FSCK_CAN_IGNORE|
 121                           FSCK_NO_RATELIMIT,
 122                           "btree node with incorrect max_key at btree %s level %u:\n"
 123                           "  %s\n"
 124                           "  expected %s",
 125                           bch2_btree_ids[b->c.btree_id], b->c.level,
 126                           buf1.buf, buf2.buf) &&
 127                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 128                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 129                         ret = -BCH_ERR_need_topology_repair;
 130                         goto err;
 131                 } else {
 132                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 133                 }
 134         }
 135
 136         bch2_bkey_buf_copy(prev, c, cur.k);
 137 err:
 138 fsck_err:
 139         printbuf_exit(&buf2);
 140         printbuf_exit(&buf1);
 141         return ret;
 142 }
 143
 144 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 145 {
 146         switch (b->key.k.type) {
 147         case KEY_TYPE_btree_ptr: {
 148                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 149
 150                 dst->k.p                = src->k.p;
 151                 dst->v.mem_ptr          = 0;
 152                 dst->v.seq              = b->data->keys.seq;
 153                 dst->v.sectors_written  = 0;
 154                 dst->v.flags            = 0;
 155                 dst->v.min_key          = b->data->min_key;
 156                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 157                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 158                 break;
 159         }
 160         case KEY_TYPE_btree_ptr_v2:
 161                 bkey_copy(&dst->k_i, &b->key);
 162                 break;
 163         default:
 164                 BUG();
 165         }
 166 }
 167
 168 static void bch2_btree_node_update_key_early(struct bch_fs *c,
 169                                              enum btree_id btree, unsigned level,
 170                                              struct bkey_s_c old, struct bkey_i *new)
 171 {
 172         struct btree *b;
 173         struct bkey_buf tmp;
 174         int ret;
 175
 176         bch2_bkey_buf_init(&tmp);
 177         bch2_bkey_buf_reassemble(&tmp, c, old);
 178
 179         b = bch2_btree_node_get_noiter(c, tmp.k, btree, level, true);
 180         if (!IS_ERR_OR_NULL(b)) {
 181                 mutex_lock(&c->btree_cache.lock);
 182
 183                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 184
 185                 bkey_copy(&b->key, new);
 186                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 187                 BUG_ON(ret);
 188
 189                 mutex_unlock(&c->btree_cache.lock);
 190                 six_unlock_read(&b->c.lock);
 191         }
 192
 193         bch2_bkey_buf_exit(&tmp, c);
 194 }
 195
 196 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 197 {
 198         struct bkey_i_btree_ptr_v2 *new;
 199         int ret;
 200
 201         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 202         if (!new)
 203                 return -ENOMEM;
 204
 205         btree_ptr_to_v2(b, new);
 206         b->data->min_key        = new_min;
 207         new->v.min_key          = new_min;
 208         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 209
 210         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 211         if (ret) {
 212                 kfree(new);
 213                 return ret;
 214         }
 215
 216         bch2_btree_node_drop_keys_outside_node(b);
 217         bkey_copy(&b->key, &new->k_i);
 218         return 0;
 219 }
 220
 221 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 222 {
 223         struct bkey_i_btree_ptr_v2 *new;
 224         int ret;
 225
 226         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 227         if (ret)
 228                 return ret;
 229
 230         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 231         if (!new)
 232                 return -ENOMEM;
 233
 234         btree_ptr_to_v2(b, new);
 235         b->data->max_key        = new_max;
 236         new->k.p                = new_max;
 237         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 238
 239         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 240         if (ret) {
 241                 kfree(new);
 242                 return ret;
 243         }
 244
 245         bch2_btree_node_drop_keys_outside_node(b);
 246
 247         mutex_lock(&c->btree_cache.lock);
 248         bch2_btree_node_hash_remove(&c->btree_cache, b);
 249
 250         bkey_copy(&b->key, &new->k_i);
 251         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 252         BUG_ON(ret);
 253         mutex_unlock(&c->btree_cache.lock);
 254         return 0;
 255 }
 256
 257 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 258                                         struct btree *prev, struct btree *cur)
 259 {
 260         struct bpos expected_start = !prev
 261                 ? b->data->min_key
 262                 : bpos_successor(prev->key.k.p);
 263         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 264         int ret = 0;
 265
 266         if (!prev) {
 267                 prt_printf(&buf1, "start of node: ");
 268                 bch2_bpos_to_text(&buf1, b->data->min_key);
 269         } else {
 270                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
 271         }
 272
 273         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
 274
 275         if (prev &&
 276             bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 277             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 278                 /* cur overwrites prev: */
 279
 280                 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
 281                                                  cur->data->min_key) >= 0, c,
 282                                 "btree node overwritten by next node at btree %s level %u:\n"
 283                                 "  node %s\n"
 284                                 "  next %s",
 285                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 286                                 buf1.buf, buf2.buf)) {
 287                         ret = DROP_PREV_NODE;
 288                         goto out;
 289                 }
 290
 291                 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
 292                                                  bpos_predecessor(cur->data->min_key)), c,
 293                                 "btree node with incorrect max_key at btree %s level %u:\n"
 294                                 "  node %s\n"
 295                                 "  next %s",
 296                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 297                                 buf1.buf, buf2.buf))
 298                         ret = set_node_max(c, prev,
 299                                            bpos_predecessor(cur->data->min_key));
 300         } else {
 301                 /* prev overwrites cur: */
 302
 303                 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
 304                                                  cur->data->max_key) >= 0, c,
 305                                 "btree node overwritten by prev node at btree %s level %u:\n"
 306                                 "  prev %s\n"
 307                                 "  node %s",
 308                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 309                                 buf1.buf, buf2.buf)) {
 310                         ret = DROP_THIS_NODE;
 311                         goto out;
 312                 }
 313
 314                 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 315                                 "btree node with incorrect min_key at btree %s level %u:\n"
 316                                 "  prev %s\n"
 317                                 "  node %s",
 318                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 319                                 buf1.buf, buf2.buf))
 320                     ret = set_node_min(c, cur, expected_start);
 321         }
 322 out:
 323 fsck_err:
 324         printbuf_exit(&buf2);
 325         printbuf_exit(&buf1);
 326         return ret;
 327 }
 328
 329 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 330                                  struct btree *child)
 331 {
 332         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 333         int ret = 0;
 334
 335         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
 336         bch2_bpos_to_text(&buf2, b->key.k.p);
 337
 338         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 339                         "btree node with incorrect max_key at btree %s level %u:\n"
 340                         "  %s\n"
 341                         "  expected %s",
 342                         bch2_btree_ids[b->c.btree_id], b->c.level,
 343                         buf1.buf, buf2.buf)) {
 344                 ret = set_node_max(c, child, b->key.k.p);
 345                 if (ret)
 346                         goto err;
 347         }
 348 err:
 349 fsck_err:
 350         printbuf_exit(&buf2);
 351         printbuf_exit(&buf1);
 352         return ret;
 353 }
 354
 355 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 356 {
 357         struct btree_and_journal_iter iter;
 358         struct bkey_s_c k;
 359         struct bkey_buf prev_k, cur_k;
 360         struct btree *prev = NULL, *cur = NULL;
 361         bool have_child, dropped_children = false;
 362         struct printbuf buf = PRINTBUF;
 363         int ret = 0;
 364
 365         if (!b->c.level)
 366                 return 0;
 367 again:
 368         prev = NULL;
 369         have_child = dropped_children = false;
 370         bch2_bkey_buf_init(&prev_k);
 371         bch2_bkey_buf_init(&cur_k);
 372         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 373
 374         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 375                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 376                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 377
 378                 bch2_btree_and_journal_iter_advance(&iter);
 379                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 380
 381                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 382                                         b->c.btree_id, b->c.level - 1,
 383                                         false);
 384                 ret = PTR_ERR_OR_ZERO(cur);
 385
 386                 printbuf_reset(&buf);
 387                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 388
 389                 if (mustfix_fsck_err_on(ret == -EIO, c,
 390                                 "Topology repair: unreadable btree node at btree %s level %u:\n"
 391                                 "  %s",
 392                                 bch2_btree_ids[b->c.btree_id],
 393                                 b->c.level - 1,
 394                                 buf.buf)) {
 395                         bch2_btree_node_evict(c, cur_k.k);
 396                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 397                                                       b->c.level, cur_k.k->k.p);
 398                         cur = NULL;
 399                         if (ret)
 400                                 break;
 401                         continue;
 402                 }
 403
 404                 if (ret) {
 405                         bch_err(c, "%s: error getting btree node: %s",
 406                                 __func__, bch2_err_str(ret));
 407                         break;
 408                 }
 409
 410                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 411
 412                 if (ret == DROP_THIS_NODE) {
 413                         six_unlock_read(&cur->c.lock);
 414                         bch2_btree_node_evict(c, cur_k.k);
 415                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 416                                                       b->c.level, cur_k.k->k.p);
 417                         cur = NULL;
 418                         if (ret)
 419                                 break;
 420                         continue;
 421                 }
 422
 423                 if (prev)
 424                         six_unlock_read(&prev->c.lock);
 425                 prev = NULL;
 426
 427                 if (ret == DROP_PREV_NODE) {
 428                         bch2_btree_node_evict(c, prev_k.k);
 429                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 430                                                       b->c.level, prev_k.k->k.p);
 431                         if (ret)
 432                                 break;
 433
 434                         bch2_btree_and_journal_iter_exit(&iter);
 435                         bch2_bkey_buf_exit(&prev_k, c);
 436                         bch2_bkey_buf_exit(&cur_k, c);
 437                         goto again;
 438                 } else if (ret)
 439                         break;
 440
 441                 prev = cur;
 442                 cur = NULL;
 443                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 444         }
 445
 446         if (!ret && !IS_ERR_OR_NULL(prev)) {
 447                 BUG_ON(cur);
 448                 ret = btree_repair_node_end(c, b, prev);
 449         }
 450
 451         if (!IS_ERR_OR_NULL(prev))
 452                 six_unlock_read(&prev->c.lock);
 453         prev = NULL;
 454         if (!IS_ERR_OR_NULL(cur))
 455                 six_unlock_read(&cur->c.lock);
 456         cur = NULL;
 457
 458         if (ret)
 459                 goto err;
 460
 461         bch2_btree_and_journal_iter_exit(&iter);
 462         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 463
 464         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 465                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 466                 bch2_btree_and_journal_iter_advance(&iter);
 467
 468                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 469                                         b->c.btree_id, b->c.level - 1,
 470                                         false);
 471                 ret = PTR_ERR_OR_ZERO(cur);
 472
 473                 if (ret) {
 474                         bch_err(c, "%s: error getting btree node: %s",
 475                                 __func__, bch2_err_str(ret));
 476                         goto err;
 477                 }
 478
 479                 ret = bch2_btree_repair_topology_recurse(c, cur);
 480                 six_unlock_read(&cur->c.lock);
 481                 cur = NULL;
 482
 483                 if (ret == DROP_THIS_NODE) {
 484                         bch2_btree_node_evict(c, cur_k.k);
 485                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 486                                                       b->c.level, cur_k.k->k.p);
 487                         dropped_children = true;
 488                 }
 489
 490                 if (ret)
 491                         goto err;
 492
 493                 have_child = true;
 494         }
 495
 496         printbuf_reset(&buf);
 497         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 498
 499         if (mustfix_fsck_err_on(!have_child, c,
 500                         "empty interior btree node at btree %s level %u\n"
 501                         "  %s",
 502                         bch2_btree_ids[b->c.btree_id],
 503                         b->c.level, buf.buf))
 504                 ret = DROP_THIS_NODE;
 505 err:
 506 fsck_err:
 507         if (!IS_ERR_OR_NULL(prev))
 508                 six_unlock_read(&prev->c.lock);
 509         if (!IS_ERR_OR_NULL(cur))
 510                 six_unlock_read(&cur->c.lock);
 511
 512         bch2_btree_and_journal_iter_exit(&iter);
 513         bch2_bkey_buf_exit(&prev_k, c);
 514         bch2_bkey_buf_exit(&cur_k, c);
 515
 516         if (!ret && dropped_children)
 517                 goto again;
 518
 519         printbuf_exit(&buf);
 520         return ret;
 521 }
 522
 523 static int bch2_repair_topology(struct bch_fs *c)
 524 {
 525         struct btree *b;
 526         unsigned i;
 527         int ret = 0;
 528
 529         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 530                 b = c->btree_roots[i].b;
 531                 if (btree_node_fake(b))
 532                         continue;
 533
 534                 six_lock_read(&b->c.lock, NULL, NULL);
 535                 ret = bch2_btree_repair_topology_recurse(c, b);
 536                 six_unlock_read(&b->c.lock);
 537
 538                 if (ret == DROP_THIS_NODE) {
 539                         bch_err(c, "empty btree root - repair unimplemented");
 540                         ret = -BCH_ERR_fsck_repair_unimplemented;
 541                 }
 542         }
 543
 544         return ret;
 545 }
 546
 547 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 548                                unsigned level, bool is_root,
 549                                struct bkey_s_c *k)
 550 {
 551         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 552         const union bch_extent_entry *entry;
 553         struct extent_ptr_decoded p = { 0 };
 554         bool do_update = false;
 555         struct printbuf buf = PRINTBUF;
 556         int ret = 0;
 557
 558         /*
 559          * XXX
 560          * use check_bucket_ref here
 561          */
 562         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 563                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 564                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 565                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 566
 567                 if (c->opts.reconstruct_alloc ||
 568                     fsck_err_on(!g->gen_valid, c,
 569                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 570                                 "while marking %s",
 571                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 572                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 573                                 p.ptr.gen,
 574                                 (printbuf_reset(&buf),
 575                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 576                         if (!p.ptr.cached) {
 577                                 g->gen_valid            = true;
 578                                 g->gen                  = p.ptr.gen;
 579                         } else {
 580                                 do_update = true;
 581                         }
 582                 }
 583
 584                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->gen) > 0, c,
 585                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 586                                 "while marking %s",
 587                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 588                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 589                                 p.ptr.gen, g->gen,
 590                                 (printbuf_reset(&buf),
 591                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 592                         if (!p.ptr.cached) {
 593                                 g->gen_valid            = true;
 594                                 g->gen                  = p.ptr.gen;
 595                                 g->data_type            = 0;
 596                                 g->dirty_sectors        = 0;
 597                                 g->cached_sectors       = 0;
 598                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 599                         } else {
 600                                 do_update = true;
 601                         }
 602                 }
 603
 604                 if (fsck_err_on(gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
 605                                 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 606                                 "while marking %s",
 607                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 608                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 609                                 p.ptr.gen,
 610                                 (printbuf_reset(&buf),
 611                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 612                         do_update = true;
 613
 614                 if (fsck_err_on(!p.ptr.cached &&
 615                                 gen_cmp(p.ptr.gen, g->gen) < 0, c,
 616                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 617                                 "while marking %s",
 618                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 619                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 620                                 p.ptr.gen, g->gen,
 621                                 (printbuf_reset(&buf),
 622                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 623                         do_update = true;
 624
 625                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 626                         continue;
 627
 628                 if (fsck_err_on(g->data_type &&
 629                                 g->data_type != data_type, c,
 630                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 631                                 "while marking %s",
 632                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 633                                 bch2_data_types[g->data_type],
 634                                 bch2_data_types[data_type],
 635                                 (printbuf_reset(&buf),
 636                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 637                         if (data_type == BCH_DATA_btree) {
 638                                 g->data_type    = data_type;
 639                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 640                         } else {
 641                                 do_update = true;
 642                         }
 643                 }
 644
 645                 if (p.has_ec) {
 646                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 647
 648                         if (fsck_err_on(!m || !m->alive, c,
 649                                         "pointer to nonexistent stripe %llu\n"
 650                                         "while marking %s",
 651                                         (u64) p.ec.idx,
 652                                         (printbuf_reset(&buf),
 653                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 654                                 do_update = true;
 655
 656                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 657                                         "pointer does not match stripe %llu\n"
 658                                         "while marking %s",
 659                                         (u64) p.ec.idx,
 660                                         (printbuf_reset(&buf),
 661                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 662                                 do_update = true;
 663                 }
 664         }
 665
 666         if (do_update) {
 667                 struct bkey_ptrs ptrs;
 668                 union bch_extent_entry *entry;
 669                 struct bch_extent_ptr *ptr;
 670                 struct bkey_i *new;
 671
 672                 if (is_root) {
 673                         bch_err(c, "cannot update btree roots yet");
 674                         ret = -EINVAL;
 675                         goto err;
 676                 }
 677
 678                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 679                 if (!new) {
 680                         bch_err(c, "%s: error allocating new key", __func__);
 681                         ret = -ENOMEM;
 682                         goto err;
 683                 }
 684
 685                 bkey_reassemble(new, *k);
 686
 687                 if (level) {
 688                         /*
 689                          * We don't want to drop btree node pointers - if the
 690                          * btree node isn't there anymore, the read path will
 691                          * sort it out:
 692                          */
 693                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 694                         bkey_for_each_ptr(ptrs, ptr) {
 695                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 696                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 697
 698                                 ptr->gen = g->gen;
 699                         }
 700                 } else {
 701                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 702                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 703                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 704                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 705
 706                                 (ptr->cached &&
 707                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 708                                 (!ptr->cached &&
 709                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 710                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 711                                 (g->data_type &&
 712                                  g->data_type != data_type);
 713                         }));
 714 again:
 715                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 716                         bkey_extent_entry_for_each(ptrs, entry) {
 717                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 718                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 719                                                                         entry->stripe_ptr.idx);
 720                                         union bch_extent_entry *next_ptr;
 721
 722                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 723                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 724                                                         goto found;
 725                                         next_ptr = NULL;
 726 found:
 727                                         if (!next_ptr) {
 728                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 729                                                 continue;
 730                                         }
 731
 732                                         if (!m || !m->alive ||
 733                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 734                                                                        &next_ptr->ptr,
 735                                                                        m->sectors)) {
 736                                                 bch2_bkey_extent_entry_drop(new, entry);
 737                                                 goto again;
 738                                         }
 739                                 }
 740                         }
 741                 }
 742
 743                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 744                 if (ret) {
 745                         kfree(new);
 746                         goto err;
 747                 }
 748
 749                 if (level)
 750                         bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
 751
 752                 if (c->opts.verbose) {
 753                         printbuf_reset(&buf);
 754                         bch2_bkey_val_to_text(&buf, c, *k);
 755                         bch_info(c, "updated %s", buf.buf);
 756
 757                         printbuf_reset(&buf);
 758                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 759                         bch_info(c, "new key %s", buf.buf);
 760                 }
 761
 762                 *k = bkey_i_to_s_c(new);
 763         }
 764 err:
 765 fsck_err:
 766         printbuf_exit(&buf);
 767         return ret;
 768 }
 769
 770 /* marking of btree keys/nodes: */
 771
 772 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 773                             unsigned level, bool is_root,
 774                             struct bkey_s_c *k,
 775                             bool initial)
 776 {
 777         struct bch_fs *c = trans->c;
 778         struct bkey deleted = KEY(0, 0, 0);
 779         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 780         unsigned flags =
 781                 BTREE_TRIGGER_GC|
 782                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 783         int ret = 0;
 784
 785         deleted.p = k->k->p;
 786
 787         if (initial) {
 788                 BUG_ON(bch2_journal_seq_verify &&
 789                        k->k->version.lo > atomic64_read(&c->journal.seq));
 790
 791                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 792                 if (ret)
 793                         goto err;
 794
 795                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 796                                 "key version number higher than recorded: %llu > %llu",
 797                                 k->k->version.lo,
 798                                 atomic64_read(&c->key_version)))
 799                         atomic64_set(&c->key_version, k->k->version.lo);
 800         }
 801
 802         ret = commit_do(trans, NULL, NULL, 0,
 803                         bch2_mark_key(trans, old, *k, flags));
 804 fsck_err:
 805 err:
 806         if (ret)
 807                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
 808         return ret;
 809 }
 810
 811 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 812 {
 813         struct bch_fs *c = trans->c;
 814         struct btree_node_iter iter;
 815         struct bkey unpacked;
 816         struct bkey_s_c k;
 817         struct bkey_buf prev, cur;
 818         int ret = 0;
 819
 820         if (!btree_node_type_needs_gc(btree_node_type(b)))
 821                 return 0;
 822
 823         bch2_btree_node_iter_init_from_start(&iter, b);
 824         bch2_bkey_buf_init(&prev);
 825         bch2_bkey_buf_init(&cur);
 826         bkey_init(&prev.k->k);
 827
 828         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 829                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 830                                        &k, initial);
 831                 if (ret)
 832                         break;
 833
 834                 bch2_btree_node_iter_advance(&iter, b);
 835
 836                 if (b->c.level) {
 837                         bch2_bkey_buf_reassemble(&cur, c, k);
 838
 839                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 840                                         bch2_btree_node_iter_end(&iter));
 841                         if (ret)
 842                                 break;
 843                 }
 844         }
 845
 846         bch2_bkey_buf_exit(&cur, c);
 847         bch2_bkey_buf_exit(&prev, c);
 848         return ret;
 849 }
 850
 851 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 852                          bool initial, bool metadata_only)
 853 {
 854         struct bch_fs *c = trans->c;
 855         struct btree_iter iter;
 856         struct btree *b;
 857         unsigned depth = metadata_only ? 1 : 0;
 858         int ret = 0;
 859
 860         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 861
 862         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 863                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 864                 bch2_verify_btree_nr_keys(b);
 865
 866                 gc_pos_set(c, gc_pos_btree_node(b));
 867
 868                 ret = btree_gc_mark_node(trans, b, initial);
 869                 if (ret)
 870                         break;
 871         }
 872         bch2_trans_iter_exit(trans, &iter);
 873
 874         if (ret)
 875                 return ret;
 876
 877         mutex_lock(&c->btree_root_lock);
 878         b = c->btree_roots[btree_id].b;
 879         if (!btree_node_fake(b)) {
 880                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 881
 882                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 883                                        true, &k, initial);
 884         }
 885         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 886         mutex_unlock(&c->btree_root_lock);
 887
 888         return ret;
 889 }
 890
 891 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 892                                       unsigned target_depth)
 893 {
 894         struct bch_fs *c = trans->c;
 895         struct btree_and_journal_iter iter;
 896         struct bkey_s_c k;
 897         struct bkey_buf cur, prev;
 898         struct printbuf buf = PRINTBUF;
 899         int ret = 0;
 900
 901         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 902         bch2_bkey_buf_init(&prev);
 903         bch2_bkey_buf_init(&cur);
 904         bkey_init(&prev.k->k);
 905
 906         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 907                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 908                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 909
 910                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 911                                        false, &k, true);
 912                 if (ret) {
 913                         bch_err(c, "%s: error from bch2_gc_mark_key: %s",
 914                                 __func__, bch2_err_str(ret));
 915                         goto fsck_err;
 916                 }
 917
 918                 if (b->c.level) {
 919                         bch2_bkey_buf_reassemble(&cur, c, k);
 920                         k = bkey_i_to_s_c(cur.k);
 921
 922                         bch2_btree_and_journal_iter_advance(&iter);
 923
 924                         ret = bch2_gc_check_topology(c, b,
 925                                         &prev, cur,
 926                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 927                         if (ret)
 928                                 goto fsck_err;
 929                 } else {
 930                         bch2_btree_and_journal_iter_advance(&iter);
 931                 }
 932         }
 933
 934         if (b->c.level > target_depth) {
 935                 bch2_btree_and_journal_iter_exit(&iter);
 936                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 937
 938                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 939                         struct btree *child;
 940
 941                         bch2_bkey_buf_reassemble(&cur, c, k);
 942                         bch2_btree_and_journal_iter_advance(&iter);
 943
 944                         child = bch2_btree_node_get_noiter(c, cur.k,
 945                                                 b->c.btree_id, b->c.level - 1,
 946                                                 false);
 947                         ret = PTR_ERR_OR_ZERO(child);
 948
 949                         if (ret == -EIO) {
 950                                 bch2_topology_error(c);
 951
 952                                 if (__fsck_err(c,
 953                                           FSCK_CAN_FIX|
 954                                           FSCK_CAN_IGNORE|
 955                                           FSCK_NO_RATELIMIT,
 956                                           "Unreadable btree node at btree %s level %u:\n"
 957                                           "  %s",
 958                                           bch2_btree_ids[b->c.btree_id],
 959                                           b->c.level - 1,
 960                                           (printbuf_reset(&buf),
 961                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 962                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 963                                         ret = -BCH_ERR_need_topology_repair;
 964                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 965                                         goto fsck_err;
 966                                 } else {
 967                                         /* Continue marking when opted to not
 968                                          * fix the error: */
 969                                         ret = 0;
 970                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 971                                         continue;
 972                                 }
 973                         } else if (ret) {
 974                                 bch_err(c, "%s: error getting btree node: %s",
 975                                         __func__, bch2_err_str(ret));
 976                                 break;
 977                         }
 978
 979                         ret = bch2_gc_btree_init_recurse(trans, child,
 980                                                          target_depth);
 981                         six_unlock_read(&child->c.lock);
 982
 983                         if (ret)
 984                                 break;
 985                 }
 986         }
 987 fsck_err:
 988         bch2_bkey_buf_exit(&cur, c);
 989         bch2_bkey_buf_exit(&prev, c);
 990         bch2_btree_and_journal_iter_exit(&iter);
 991         printbuf_exit(&buf);
 992         return ret;
 993 }
 994
 995 static int bch2_gc_btree_init(struct btree_trans *trans,
 996                               enum btree_id btree_id,
 997                               bool metadata_only)
 998 {
 999         struct bch_fs *c = trans->c;
1000         struct btree *b;
1001         unsigned target_depth = metadata_only ? 1 : 0;
1002         struct printbuf buf = PRINTBUF;
1003         int ret = 0;
1004
1005         b = c->btree_roots[btree_id].b;
1006
1007         if (btree_node_fake(b))
1008                 return 0;
1009
1010         six_lock_read(&b->c.lock, NULL, NULL);
1011         printbuf_reset(&buf);
1012         bch2_bpos_to_text(&buf, b->data->min_key);
1013         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1014                         "btree root with incorrect min_key: %s", buf.buf)) {
1015                 bch_err(c, "repair unimplemented");
1016                 ret = -BCH_ERR_fsck_repair_unimplemented;
1017                 goto fsck_err;
1018         }
1019
1020         printbuf_reset(&buf);
1021         bch2_bpos_to_text(&buf, b->data->max_key);
1022         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1023                         "btree root with incorrect max_key: %s", buf.buf)) {
1024                 bch_err(c, "repair unimplemented");
1025                 ret = -BCH_ERR_fsck_repair_unimplemented;
1026                 goto fsck_err;
1027         }
1028
1029         if (b->c.level >= target_depth)
1030                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1031
1032         if (!ret) {
1033                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1034
1035                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1036                                        &k, true);
1037         }
1038 fsck_err:
1039         six_unlock_read(&b->c.lock);
1040
1041         if (ret < 0)
1042                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1043         printbuf_exit(&buf);
1044         return ret;
1045 }
1046
1047 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1048 {
1049         return  (int) btree_id_to_gc_phase(l) -
1050                 (int) btree_id_to_gc_phase(r);
1051 }
1052
1053 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1054 {
1055         struct btree_trans trans;
1056         enum btree_id ids[BTREE_ID_NR];
1057         unsigned i;
1058         int ret = 0;
1059
1060         bch2_trans_init(&trans, c, 0, 0);
1061
1062         if (initial)
1063                 trans.is_initial_gc = true;
1064
1065         for (i = 0; i < BTREE_ID_NR; i++)
1066                 ids[i] = i;
1067         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1068
1069         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1070                 ret = initial
1071                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1072                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1073
1074         if (ret < 0)
1075                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1076
1077         bch2_trans_exit(&trans);
1078         return ret;
1079 }
1080
1081 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1082                                   u64 start, u64 end,
1083                                   enum bch_data_type type,
1084                                   unsigned flags)
1085 {
1086         u64 b = sector_to_bucket(ca, start);
1087
1088         do {
1089                 unsigned sectors =
1090                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1091
1092                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1093                                           gc_phase(GC_PHASE_SB), flags);
1094                 b++;
1095                 start += sectors;
1096         } while (start < end);
1097 }
1098
1099 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1100                                      unsigned flags)
1101 {
1102         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1103         unsigned i;
1104         u64 b;
1105
1106         for (i = 0; i < layout->nr_superblocks; i++) {
1107                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1108
1109                 if (offset == BCH_SB_SECTOR)
1110                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1111                                               BCH_DATA_sb, flags);
1112
1113                 mark_metadata_sectors(c, ca, offset,
1114                                       offset + (1 << layout->sb_max_size_bits),
1115                                       BCH_DATA_sb, flags);
1116         }
1117
1118         for (i = 0; i < ca->journal.nr; i++) {
1119                 b = ca->journal.buckets[i];
1120                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1121                                           ca->mi.bucket_size,
1122                                           gc_phase(GC_PHASE_SB), flags);
1123         }
1124 }
1125
1126 static void bch2_mark_superblocks(struct bch_fs *c)
1127 {
1128         struct bch_dev *ca;
1129         unsigned i;
1130
1131         mutex_lock(&c->sb_lock);
1132         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1133
1134         for_each_online_member(ca, c, i)
1135                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1136         mutex_unlock(&c->sb_lock);
1137 }
1138
1139 #if 0
1140 /* Also see bch2_pending_btree_node_free_insert_done() */
1141 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1142 {
1143         struct btree_update *as;
1144         struct pending_btree_node_free *d;
1145
1146         mutex_lock(&c->btree_interior_update_lock);
1147         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1148
1149         for_each_pending_btree_node_free(c, as, d)
1150                 if (d->index_update_done)
1151                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1152
1153         mutex_unlock(&c->btree_interior_update_lock);
1154 }
1155 #endif
1156
1157 static void bch2_gc_free(struct bch_fs *c)
1158 {
1159         struct bch_dev *ca;
1160         unsigned i;
1161
1162         genradix_free(&c->reflink_gc_table);
1163         genradix_free(&c->gc_stripes);
1164
1165         for_each_member_device(ca, c, i) {
1166                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1167                         sizeof(struct bucket_array) +
1168                         ca->mi.nbuckets * sizeof(struct bucket));
1169                 ca->buckets_gc = NULL;
1170
1171                 free_percpu(ca->usage_gc);
1172                 ca->usage_gc = NULL;
1173         }
1174
1175         free_percpu(c->usage_gc);
1176         c->usage_gc = NULL;
1177 }
1178
1179 static int bch2_gc_done(struct bch_fs *c,
1180                         bool initial, bool metadata_only)
1181 {
1182         struct bch_dev *ca = NULL;
1183         struct printbuf buf = PRINTBUF;
1184         bool verify = !metadata_only &&
1185                 !c->opts.reconstruct_alloc &&
1186                 (!initial || (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1187         unsigned i, dev;
1188         int ret = 0;
1189
1190         percpu_down_write(&c->mark_lock);
1191
1192 #define copy_field(_f, _msg, ...)                                       \
1193         if (dst->_f != src->_f &&                                       \
1194             (!verify ||                                                 \
1195              fsck_err(c, _msg ": got %llu, should be %llu"              \
1196                       , ##__VA_ARGS__, dst->_f, src->_f)))              \
1197                 dst->_f = src->_f
1198 #define copy_stripe_field(_f, _msg, ...)                                \
1199         if (dst->_f != src->_f &&                                       \
1200             (!verify ||                                                 \
1201              fsck_err(c, "stripe %zu has wrong "_msg                    \
1202                       ": got %u, should be %u",                         \
1203                       iter.pos, ##__VA_ARGS__,                          \
1204                       dst->_f, src->_f)))                               \
1205                 dst->_f = src->_f
1206 #define copy_dev_field(_f, _msg, ...)                                   \
1207         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1208 #define copy_fs_field(_f, _msg, ...)                                    \
1209         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1210
1211         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1212                 bch2_fs_usage_acc_to_base(c, i);
1213
1214         for_each_member_device(ca, c, dev) {
1215                 struct bch_dev_usage *dst = ca->usage_base;
1216                 struct bch_dev_usage *src = (void *)
1217                         bch2_acc_percpu_u64s((void *) ca->usage_gc,
1218                                              dev_usage_u64s());
1219
1220                 copy_dev_field(buckets_ec,              "buckets_ec");
1221
1222                 for (i = 0; i < BCH_DATA_NR; i++) {
1223                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1224                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1225                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1226                 }
1227         };
1228
1229         {
1230                 unsigned nr = fs_usage_u64s(c);
1231                 struct bch_fs_usage *dst = c->usage_base;
1232                 struct bch_fs_usage *src = (void *)
1233                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1234
1235                 copy_fs_field(hidden,           "hidden");
1236                 copy_fs_field(btree,            "btree");
1237
1238                 if (!metadata_only) {
1239                         copy_fs_field(data,     "data");
1240                         copy_fs_field(cached,   "cached");
1241                         copy_fs_field(reserved, "reserved");
1242                         copy_fs_field(nr_inodes,"nr_inodes");
1243
1244                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1245                                 copy_fs_field(persistent_reserved[i],
1246                                               "persistent_reserved[%i]", i);
1247                 }
1248
1249                 for (i = 0; i < c->replicas.nr; i++) {
1250                         struct bch_replicas_entry *e =
1251                                 cpu_replicas_entry(&c->replicas, i);
1252
1253                         if (metadata_only &&
1254                             (e->data_type == BCH_DATA_user ||
1255                              e->data_type == BCH_DATA_cached))
1256                                 continue;
1257
1258                         printbuf_reset(&buf);
1259                         bch2_replicas_entry_to_text(&buf, e);
1260
1261                         copy_fs_field(replicas[i], "%s", buf.buf);
1262                 }
1263         }
1264
1265 #undef copy_fs_field
1266 #undef copy_dev_field
1267 #undef copy_stripe_field
1268 #undef copy_field
1269 fsck_err:
1270         if (ca)
1271                 percpu_ref_put(&ca->ref);
1272         if (ret)
1273                 bch_err(c, "error from %s(): %s", __func__, bch2_err_str(ret));
1274
1275         percpu_up_write(&c->mark_lock);
1276         printbuf_exit(&buf);
1277         return ret;
1278 }
1279
1280 static int bch2_gc_start(struct bch_fs *c,
1281                          bool metadata_only)
1282 {
1283         struct bch_dev *ca = NULL;
1284         unsigned i;
1285
1286         BUG_ON(c->usage_gc);
1287
1288         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1289                                          sizeof(u64), GFP_KERNEL);
1290         if (!c->usage_gc) {
1291                 bch_err(c, "error allocating c->usage_gc");
1292                 return -ENOMEM;
1293         }
1294
1295         for_each_member_device(ca, c, i) {
1296                 BUG_ON(ca->buckets_gc);
1297                 BUG_ON(ca->usage_gc);
1298
1299                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1300                 if (!ca->usage_gc) {
1301                         bch_err(c, "error allocating ca->usage_gc");
1302                         percpu_ref_put(&ca->ref);
1303                         return -ENOMEM;
1304                 }
1305
1306                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1307                                ca->mi.nbuckets - ca->mi.first_bucket);
1308         }
1309
1310         return 0;
1311 }
1312
1313 /* returns true if not equal */
1314 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1315                                      struct bch_alloc_v4 r)
1316 {
1317         return  l.gen != r.gen                          ||
1318                 l.oldest_gen != r.oldest_gen            ||
1319                 l.data_type != r.data_type              ||
1320                 l.dirty_sectors != r.dirty_sectors      ||
1321                 l.cached_sectors != r.cached_sectors     ||
1322                 l.stripe_redundancy != r.stripe_redundancy ||
1323                 l.stripe != r.stripe;
1324 }
1325
1326 static int bch2_alloc_write_key(struct btree_trans *trans,
1327                                 struct btree_iter *iter,
1328                                 struct bkey_s_c k,
1329                                 bool metadata_only)
1330 {
1331         struct bch_fs *c = trans->c;
1332         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1333         struct bucket gc, *b;
1334         struct bkey_i_alloc_v4 *a;
1335         struct bch_alloc_v4 old, new;
1336         enum bch_data_type type;
1337         int ret;
1338
1339         if (bkey_cmp(iter->pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1340                 return 1;
1341
1342         bch2_alloc_to_v4(k, &old);
1343         new = old;
1344
1345         percpu_down_read(&c->mark_lock);
1346         b = gc_bucket(ca, iter->pos.offset);
1347
1348         /*
1349          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1350          * fix that here:
1351          */
1352         type = __alloc_data_type(b->dirty_sectors,
1353                                  b->cached_sectors,
1354                                  b->stripe,
1355                                  old,
1356                                  b->data_type);
1357         if (b->data_type != type) {
1358                 struct bch_dev_usage *u;
1359
1360                 preempt_disable();
1361                 u = this_cpu_ptr(ca->usage_gc);
1362                 u->d[b->data_type].buckets--;
1363                 b->data_type = type;
1364                 u->d[b->data_type].buckets++;
1365                 preempt_enable();
1366         }
1367
1368         gc = *b;
1369         percpu_up_read(&c->mark_lock);
1370
1371         if (metadata_only &&
1372             gc.data_type != BCH_DATA_sb &&
1373             gc.data_type != BCH_DATA_journal &&
1374             gc.data_type != BCH_DATA_btree)
1375                 return 0;
1376
1377         if (gen_after(old.gen, gc.gen))
1378                 return 0;
1379
1380 #define copy_bucket_field(_f)                                           \
1381         if (c->opts.reconstruct_alloc ||                                \
1382             fsck_err_on(new._f != gc._f, c,                             \
1383                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1384                         ": got %u, should be %u",                       \
1385                         iter->pos.inode, iter->pos.offset,              \
1386                         gc.gen,                                         \
1387                         bch2_data_types[gc.data_type],                  \
1388                         new._f, gc._f))                                 \
1389                 new._f = gc._f;                                         \
1390
1391         copy_bucket_field(gen);
1392         copy_bucket_field(data_type);
1393         copy_bucket_field(dirty_sectors);
1394         copy_bucket_field(cached_sectors);
1395         copy_bucket_field(stripe_redundancy);
1396         copy_bucket_field(stripe);
1397 #undef copy_bucket_field
1398
1399         if (!bch2_alloc_v4_cmp(old, new))
1400                 return 0;
1401
1402         a = bch2_alloc_to_v4_mut(trans, k);
1403         ret = PTR_ERR_OR_ZERO(a);
1404         if (ret)
1405                 return ret;
1406
1407         a->v = new;
1408
1409         /*
1410          * The trigger normally makes sure this is set, but we're not running
1411          * triggers:
1412          */
1413         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1414                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1415
1416         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1417 fsck_err:
1418         return ret;
1419 }
1420
1421 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1422 {
1423         struct btree_trans trans;
1424         struct btree_iter iter;
1425         struct bkey_s_c k;
1426         struct bch_dev *ca;
1427         unsigned i;
1428         int ret = 0;
1429
1430         bch2_trans_init(&trans, c, 0, 0);
1431
1432         for_each_member_device(ca, c, i) {
1433                 ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1434                                 POS(ca->dev_idx, ca->mi.first_bucket),
1435                                 BTREE_ITER_SLOTS|BTREE_ITER_PREFETCH, k,
1436                                 NULL, NULL, BTREE_INSERT_LAZY_RW,
1437                         bch2_alloc_write_key(&trans, &iter, k, metadata_only));
1438
1439                 if (ret < 0) {
1440                         bch_err(c, "error writing alloc info: %s", bch2_err_str(ret));
1441                         percpu_ref_put(&ca->ref);
1442                         break;
1443                 }
1444         }
1445
1446         bch2_trans_exit(&trans);
1447         return ret < 0 ? ret : 0;
1448 }
1449
1450 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1451 {
1452         struct bch_dev *ca;
1453         struct btree_trans trans;
1454         struct btree_iter iter;
1455         struct bkey_s_c k;
1456         struct bucket *g;
1457         struct bch_alloc_v4 a;
1458         unsigned i;
1459         int ret;
1460
1461         for_each_member_device(ca, c, i) {
1462                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1463                                 ca->mi.nbuckets * sizeof(struct bucket),
1464                                 GFP_KERNEL|__GFP_ZERO);
1465                 if (!buckets) {
1466                         percpu_ref_put(&ca->ref);
1467                         bch_err(c, "error allocating ca->buckets[gc]");
1468                         return -ENOMEM;
1469                 }
1470
1471                 buckets->first_bucket   = ca->mi.first_bucket;
1472                 buckets->nbuckets       = ca->mi.nbuckets;
1473                 rcu_assign_pointer(ca->buckets_gc, buckets);
1474         };
1475
1476         bch2_trans_init(&trans, c, 0, 0);
1477
1478         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1479                            BTREE_ITER_PREFETCH, k, ret) {
1480                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1481                 g = gc_bucket(ca, k.k->p.offset);
1482
1483                 bch2_alloc_to_v4(k, &a);
1484
1485                 g->gen_valid    = 1;
1486                 g->gen          = a.gen;
1487
1488                 if (metadata_only &&
1489                     (a.data_type == BCH_DATA_user ||
1490                      a.data_type == BCH_DATA_cached ||
1491                      a.data_type == BCH_DATA_parity)) {
1492                         g->data_type            = a.data_type;
1493                         g->dirty_sectors        = a.dirty_sectors;
1494                         g->cached_sectors       = a.cached_sectors;
1495                         g->stripe               = a.stripe;
1496                         g->stripe_redundancy    = a.stripe_redundancy;
1497                 }
1498         }
1499         bch2_trans_iter_exit(&trans, &iter);
1500
1501         bch2_trans_exit(&trans);
1502
1503         if (ret)
1504                 bch_err(c, "error reading alloc info at gc start: %s", bch2_err_str(ret));
1505
1506         return ret;
1507 }
1508
1509 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1510 {
1511         struct bch_dev *ca;
1512         unsigned i;
1513
1514         for_each_member_device(ca, c, i) {
1515                 struct bucket_array *buckets = gc_bucket_array(ca);
1516                 struct bucket *g;
1517
1518                 for_each_bucket(g, buckets) {
1519                         if (metadata_only &&
1520                             (g->data_type == BCH_DATA_user ||
1521                              g->data_type == BCH_DATA_cached ||
1522                              g->data_type == BCH_DATA_parity))
1523                                 continue;
1524                         g->data_type = 0;
1525                         g->dirty_sectors = 0;
1526                         g->cached_sectors = 0;
1527                 }
1528         };
1529 }
1530
1531 static int bch2_gc_write_reflink_key(struct btree_trans *trans,
1532                                      struct btree_iter *iter,
1533                                      struct bkey_s_c k,
1534                                      size_t *idx)
1535 {
1536         struct bch_fs *c = trans->c;
1537         const __le64 *refcount = bkey_refcount_c(k);
1538         struct printbuf buf = PRINTBUF;
1539         struct reflink_gc *r;
1540         int ret = 0;
1541
1542         if (!refcount)
1543                 return 0;
1544
1545         while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
1546                r->offset < k.k->p.offset)
1547                 ++*idx;
1548
1549         if (!r ||
1550             r->offset != k.k->p.offset ||
1551             r->size != k.k->size) {
1552                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1553                 return -EINVAL;
1554         }
1555
1556         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1557                         "reflink key has wrong refcount:\n"
1558                         "  %s\n"
1559                         "  should be %u",
1560                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1561                         r->refcount)) {
1562                 struct bkey_i *new;
1563
1564                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1565                 ret = PTR_ERR_OR_ZERO(new);
1566                 if (ret)
1567                         return ret;
1568
1569                 bkey_reassemble(new, k);
1570
1571                 if (!r->refcount)
1572                         new->k.type = KEY_TYPE_deleted;
1573                 else
1574                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1575
1576                 ret = bch2_trans_update(trans, iter, new, 0);
1577         }
1578 fsck_err:
1579         printbuf_exit(&buf);
1580         return ret;
1581 }
1582
1583 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1584 {
1585         struct btree_trans trans;
1586         struct btree_iter iter;
1587         struct bkey_s_c k;
1588         size_t idx = 0;
1589         int ret = 0;
1590
1591         if (metadata_only)
1592                 return 0;
1593
1594         bch2_trans_init(&trans, c, 0, 0);
1595
1596         ret = for_each_btree_key_commit(&trans, iter,
1597                         BTREE_ID_reflink, POS_MIN,
1598                         BTREE_ITER_PREFETCH, k,
1599                         NULL, NULL, BTREE_INSERT_NOFAIL,
1600                 bch2_gc_write_reflink_key(&trans, &iter, k, &idx));
1601
1602         c->reflink_gc_nr = 0;
1603         bch2_trans_exit(&trans);
1604         return ret;
1605 }
1606
1607 static int bch2_gc_reflink_start(struct bch_fs *c,
1608                                  bool metadata_only)
1609 {
1610         struct btree_trans trans;
1611         struct btree_iter iter;
1612         struct bkey_s_c k;
1613         struct reflink_gc *r;
1614         int ret = 0;
1615
1616         if (metadata_only)
1617                 return 0;
1618
1619         bch2_trans_init(&trans, c, 0, 0);
1620         c->reflink_gc_nr = 0;
1621
1622         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1623                            BTREE_ITER_PREFETCH, k, ret) {
1624                 const __le64 *refcount = bkey_refcount_c(k);
1625
1626                 if (!refcount)
1627                         continue;
1628
1629                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1630                                        GFP_KERNEL);
1631                 if (!r) {
1632                         ret = -ENOMEM;
1633                         break;
1634                 }
1635
1636                 r->offset       = k.k->p.offset;
1637                 r->size         = k.k->size;
1638                 r->refcount     = 0;
1639         }
1640         bch2_trans_iter_exit(&trans, &iter);
1641
1642         bch2_trans_exit(&trans);
1643         return ret;
1644 }
1645
1646 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1647 {
1648         struct genradix_iter iter;
1649         struct reflink_gc *r;
1650
1651         genradix_for_each(&c->reflink_gc_table, iter, r)
1652                 r->refcount = 0;
1653 }
1654
1655 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
1656                                      struct btree_iter *iter,
1657                                      struct bkey_s_c k)
1658 {
1659         struct bch_fs *c = trans->c;
1660         struct printbuf buf = PRINTBUF;
1661         const struct bch_stripe *s;
1662         struct gc_stripe *m;
1663         unsigned i;
1664         int ret = 0;
1665
1666         if (k.k->type != KEY_TYPE_stripe)
1667                 return 0;
1668
1669         s = bkey_s_c_to_stripe(k).v;
1670         m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1671
1672         for (i = 0; i < s->nr_blocks; i++)
1673                 if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1674                         goto inconsistent;
1675         return 0;
1676 inconsistent:
1677         if (fsck_err_on(true, c,
1678                         "stripe has wrong block sector count %u:\n"
1679                         "  %s\n"
1680                         "  should be %u", i,
1681                         (printbuf_reset(&buf),
1682                          bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1683                         m ? m->block_sectors[i] : 0)) {
1684                 struct bkey_i_stripe *new;
1685
1686                 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1687                 ret = PTR_ERR_OR_ZERO(new);
1688                 if (ret)
1689                         return ret;
1690
1691                 bkey_reassemble(&new->k_i, k);
1692
1693                 for (i = 0; i < new->v.nr_blocks; i++)
1694                         stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1695
1696                 ret = bch2_trans_update(trans, iter, &new->k_i, 0);
1697         }
1698 fsck_err:
1699         printbuf_exit(&buf);
1700         return ret;
1701 }
1702
1703 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1704 {
1705         struct btree_trans trans;
1706         struct btree_iter iter;
1707         struct bkey_s_c k;
1708         int ret = 0;
1709
1710         if (metadata_only)
1711                 return 0;
1712
1713         bch2_trans_init(&trans, c, 0, 0);
1714
1715         ret = for_each_btree_key_commit(&trans, iter,
1716                         BTREE_ID_stripes, POS_MIN,
1717                         BTREE_ITER_PREFETCH, k,
1718                         NULL, NULL, BTREE_INSERT_NOFAIL,
1719                 bch2_gc_write_stripes_key(&trans, &iter, k));
1720
1721         bch2_trans_exit(&trans);
1722         return ret;
1723 }
1724
1725 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1726 {
1727         genradix_free(&c->gc_stripes);
1728 }
1729
1730 /**
1731  * bch2_gc - walk _all_ references to buckets, and recompute them:
1732  *
1733  * Order matters here:
1734  *  - Concurrent GC relies on the fact that we have a total ordering for
1735  *    everything that GC walks - see  gc_will_visit_node(),
1736  *    gc_will_visit_root()
1737  *
1738  *  - also, references move around in the course of index updates and
1739  *    various other crap: everything needs to agree on the ordering
1740  *    references are allowed to move around in - e.g., we're allowed to
1741  *    start with a reference owned by an open_bucket (the allocator) and
1742  *    move it to the btree, but not the reverse.
1743  *
1744  *    This is necessary to ensure that gc doesn't miss references that
1745  *    move around - if references move backwards in the ordering GC
1746  *    uses, GC could skip past them
1747  */
1748 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1749 {
1750         unsigned iter = 0;
1751         int ret;
1752
1753         lockdep_assert_held(&c->state_lock);
1754
1755         down_write(&c->gc_lock);
1756
1757         bch2_btree_interior_updates_flush(c);
1758
1759         ret   = bch2_gc_start(c, metadata_only) ?:
1760                 bch2_gc_alloc_start(c, metadata_only) ?:
1761                 bch2_gc_reflink_start(c, metadata_only);
1762         if (ret)
1763                 goto out;
1764 again:
1765         gc_pos_set(c, gc_phase(GC_PHASE_START));
1766
1767         bch2_mark_superblocks(c);
1768
1769         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1770             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1771             c->opts.fix_errors != FSCK_OPT_NO) {
1772                 bch_info(c, "Starting topology repair pass");
1773                 ret = bch2_repair_topology(c);
1774                 if (ret)
1775                         goto out;
1776                 bch_info(c, "Topology repair pass done");
1777
1778                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1779         }
1780
1781         ret = bch2_gc_btrees(c, initial, metadata_only);
1782
1783         if (ret == -BCH_ERR_need_topology_repair &&
1784             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1785             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1786                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1787                 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1788                 ret = 0;
1789         }
1790
1791         if (ret == -BCH_ERR_need_topology_repair)
1792                 ret = -BCH_ERR_fsck_errors_not_fixed;
1793
1794         if (ret)
1795                 goto out;
1796
1797 #if 0
1798         bch2_mark_pending_btree_node_frees(c);
1799 #endif
1800         c->gc_count++;
1801
1802         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1803             (!iter && bch2_test_restart_gc)) {
1804                 if (iter++ > 2) {
1805                         bch_info(c, "Unable to fix bucket gens, looping");
1806                         ret = -EINVAL;
1807                         goto out;
1808                 }
1809
1810                 /*
1811                  * XXX: make sure gens we fixed got saved
1812                  */
1813                 bch_info(c, "Second GC pass needed, restarting:");
1814                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1815                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1816
1817                 bch2_gc_stripes_reset(c, metadata_only);
1818                 bch2_gc_alloc_reset(c, metadata_only);
1819                 bch2_gc_reflink_reset(c, metadata_only);
1820
1821                 /* flush fsck errors, reset counters */
1822                 bch2_flush_fsck_errs(c);
1823                 goto again;
1824         }
1825 out:
1826         if (!ret) {
1827                 bch2_journal_block(&c->journal);
1828
1829                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1830                         bch2_gc_reflink_done(c, metadata_only) ?:
1831                         bch2_gc_alloc_done(c, metadata_only) ?:
1832                         bch2_gc_done(c, initial, metadata_only);
1833
1834                 bch2_journal_unblock(&c->journal);
1835         }
1836
1837         percpu_down_write(&c->mark_lock);
1838         /* Indicates that gc is no longer in progress: */
1839         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1840
1841         bch2_gc_free(c);
1842         percpu_up_write(&c->mark_lock);
1843
1844         up_write(&c->gc_lock);
1845
1846         /*
1847          * At startup, allocations can happen directly instead of via the
1848          * allocator thread - issue wakeup in case they blocked on gc_lock:
1849          */
1850         closure_wake_up(&c->freelist_wait);
1851         return ret;
1852 }
1853
1854 static int gc_btree_gens_key(struct btree_trans *trans,
1855                              struct btree_iter *iter,
1856                              struct bkey_s_c k)
1857 {
1858         struct bch_fs *c = trans->c;
1859         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1860         const struct bch_extent_ptr *ptr;
1861         struct bkey_i *u;
1862         int ret;
1863
1864         percpu_down_read(&c->mark_lock);
1865         bkey_for_each_ptr(ptrs, ptr) {
1866                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1867
1868                 if (ptr_stale(ca, ptr) > 16) {
1869                         percpu_up_read(&c->mark_lock);
1870                         goto update;
1871                 }
1872         }
1873
1874         bkey_for_each_ptr(ptrs, ptr) {
1875                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1876                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1877
1878                 if (gen_after(*gen, ptr->gen))
1879                         *gen = ptr->gen;
1880         }
1881         percpu_up_read(&c->mark_lock);
1882         return 0;
1883 update:
1884         u = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1885         ret = PTR_ERR_OR_ZERO(u);
1886         if (ret)
1887                 return ret;
1888
1889         bkey_reassemble(u, k);
1890
1891         bch2_extent_normalize(c, bkey_i_to_s(u));
1892         return bch2_trans_update(trans, iter, u, 0);
1893 }
1894
1895 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter,
1896                                        struct bkey_s_c k)
1897 {
1898         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1899         struct bch_alloc_v4 a;
1900         struct bkey_i_alloc_v4 *a_mut;
1901         int ret;
1902
1903         bch2_alloc_to_v4(k, &a);
1904
1905         if (a.oldest_gen == ca->oldest_gen[iter->pos.offset])
1906                 return 0;
1907
1908         a_mut = bch2_alloc_to_v4_mut(trans, k);
1909         ret = PTR_ERR_OR_ZERO(a_mut);
1910         if (ret)
1911                 return ret;
1912
1913         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1914         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1915
1916         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1917 }
1918
1919 int bch2_gc_gens(struct bch_fs *c)
1920 {
1921         struct btree_trans trans;
1922         struct btree_iter iter;
1923         struct bkey_s_c k;
1924         struct bch_dev *ca;
1925         u64 b, start_time = local_clock();
1926         unsigned i;
1927         int ret;
1928
1929         /*
1930          * Ideally we would be using state_lock and not gc_lock here, but that
1931          * introduces a deadlock in the RO path - we currently take the state
1932          * lock at the start of going RO, thus the gc thread may get stuck:
1933          */
1934         if (!mutex_trylock(&c->gc_gens_lock))
1935                 return 0;
1936
1937         trace_gc_gens_start(c);
1938         down_read(&c->gc_lock);
1939         bch2_trans_init(&trans, c, 0, 0);
1940
1941         for_each_member_device(ca, c, i) {
1942                 struct bucket_gens *gens;
1943
1944                 BUG_ON(ca->oldest_gen);
1945
1946                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1947                 if (!ca->oldest_gen) {
1948                         percpu_ref_put(&ca->ref);
1949                         ret = -ENOMEM;
1950                         goto err;
1951                 }
1952
1953                 gens = bucket_gens(ca);
1954
1955                 for (b = gens->first_bucket;
1956                      b < gens->nbuckets; b++)
1957                         ca->oldest_gen[b] = gens->b[b];
1958         }
1959
1960         for (i = 0; i < BTREE_ID_NR; i++)
1961                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1962                         struct btree_iter iter;
1963                         struct bkey_s_c k;
1964
1965                         c->gc_gens_btree = i;
1966                         c->gc_gens_pos = POS_MIN;
1967                         ret = for_each_btree_key_commit(&trans, iter, i,
1968                                         POS_MIN,
1969                                         BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS,
1970                                         k,
1971                                         NULL, NULL,
1972                                         BTREE_INSERT_NOFAIL,
1973                                 gc_btree_gens_key(&trans, &iter, k));
1974                         if (ret) {
1975                                 bch_err(c, "error recalculating oldest_gen: %s", bch2_err_str(ret));
1976                                 goto err;
1977                         }
1978                 }
1979
1980         ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1981                         POS_MIN,
1982                         BTREE_ITER_PREFETCH,
1983                         k,
1984                         NULL, NULL,
1985                         BTREE_INSERT_NOFAIL,
1986                 bch2_alloc_write_oldest_gen(&trans, &iter, k));
1987         if (ret) {
1988                 bch_err(c, "error writing oldest_gen: %s", bch2_err_str(ret));
1989                 goto err;
1990         }
1991
1992         c->gc_gens_btree        = 0;
1993         c->gc_gens_pos          = POS_MIN;
1994
1995         c->gc_count++;
1996
1997         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1998         trace_gc_gens_end(c);
1999 err:
2000         for_each_member_device(ca, c, i) {
2001                 kvfree(ca->oldest_gen);
2002                 ca->oldest_gen = NULL;
2003         }
2004
2005         bch2_trans_exit(&trans);
2006         up_read(&c->gc_lock);
2007         mutex_unlock(&c->gc_gens_lock);
2008         return ret;
2009 }
2010
2011 static int bch2_gc_thread(void *arg)
2012 {
2013         struct bch_fs *c = arg;
2014         struct io_clock *clock = &c->io_clock[WRITE];
2015         unsigned long last = atomic64_read(&clock->now);
2016         unsigned last_kick = atomic_read(&c->kick_gc);
2017         int ret;
2018
2019         set_freezable();
2020
2021         while (1) {
2022                 while (1) {
2023                         set_current_state(TASK_INTERRUPTIBLE);
2024
2025                         if (kthread_should_stop()) {
2026                                 __set_current_state(TASK_RUNNING);
2027                                 return 0;
2028                         }
2029
2030                         if (atomic_read(&c->kick_gc) != last_kick)
2031                                 break;
2032
2033                         if (c->btree_gc_periodic) {
2034                                 unsigned long next = last + c->capacity / 16;
2035
2036                                 if (atomic64_read(&clock->now) >= next)
2037                                         break;
2038
2039                                 bch2_io_clock_schedule_timeout(clock, next);
2040                         } else {
2041                                 schedule();
2042                         }
2043
2044                         try_to_freeze();
2045                 }
2046                 __set_current_state(TASK_RUNNING);
2047
2048                 last = atomic64_read(&clock->now);
2049                 last_kick = atomic_read(&c->kick_gc);
2050
2051                 /*
2052                  * Full gc is currently incompatible with btree key cache:
2053                  */
2054 #if 0
2055                 ret = bch2_gc(c, false, false);
2056 #else
2057                 ret = bch2_gc_gens(c);
2058 #endif
2059                 if (ret < 0)
2060                         bch_err(c, "btree gc failed: %s", bch2_err_str(ret));
2061
2062                 debug_check_no_locks_held();
2063         }
2064
2065         return 0;
2066 }
2067
2068 void bch2_gc_thread_stop(struct bch_fs *c)
2069 {
2070         struct task_struct *p;
2071
2072         p = c->gc_thread;
2073         c->gc_thread = NULL;
2074
2075         if (p) {
2076                 kthread_stop(p);
2077                 put_task_struct(p);
2078         }
2079 }
2080
2081 int bch2_gc_thread_start(struct bch_fs *c)
2082 {
2083         struct task_struct *p;
2084
2085         if (c->gc_thread)
2086                 return 0;
2087
2088         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2089         if (IS_ERR(p)) {
2090                 bch_err(c, "error creating gc thread: %s", bch2_err_str(PTR_ERR(p)));
2091                 return PTR_ERR(p);
2092         }
2093
2094         get_task_struct(p);
2095         c->gc_thread = p;
2096         wake_up_process(p);
2097         return 0;
2098 }