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         char buf1[200], buf2[200];
  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 (bkey_deleted(&prev->k->k)) {
  80                         struct printbuf out = PBUF(buf1);
  81                         pr_buf(&out, "start of node: ");
  82                         bch2_bpos_to_text(&out, node_start);
  83                 } else {
  84                         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
  85                 }
  86
  87                 if (bpos_cmp(expected_start, bp->v.min_key)) {
  88                         bch2_topology_error(c);
  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,
  99                                   (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2)) &&
 100                             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 101                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 102                                 return FSCK_ERR_START_TOPOLOGY_REPAIR;
 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                 if (__fsck_err(c,
 113                           FSCK_CAN_FIX|
 114                           FSCK_CAN_IGNORE|
 115                           FSCK_NO_RATELIMIT,
 116                           "btree node with incorrect max_key at btree %s level %u:\n"
 117                           "  %s\n"
 118                           "  expected %s",
 119                           bch2_btree_ids[b->c.btree_id], b->c.level,
 120                           (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
 121                           (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2)) &&
 122                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 123                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 124                         return FSCK_ERR_START_TOPOLOGY_REPAIR;
 125                 } else {
 126                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 127                 }
 128         }
 129
 130         bch2_bkey_buf_copy(prev, c, cur.k);
 131 fsck_err:
 132         return ret;
 133 }
 134
 135 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 136 {
 137         switch (b->key.k.type) {
 138         case KEY_TYPE_btree_ptr: {
 139                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 140
 141                 dst->k.p                = src->k.p;
 142                 dst->v.mem_ptr          = 0;
 143                 dst->v.seq              = b->data->keys.seq;
 144                 dst->v.sectors_written  = 0;
 145                 dst->v.flags            = 0;
 146                 dst->v.min_key          = b->data->min_key;
 147                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 148                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 149                 break;
 150         }
 151         case KEY_TYPE_btree_ptr_v2:
 152                 bkey_copy(&dst->k_i, &b->key);
 153                 break;
 154         default:
 155                 BUG();
 156         }
 157 }
 158
 159 static void bch2_btree_node_update_key_early(struct bch_fs *c,
 160                                              enum btree_id btree, unsigned level,
 161                                              struct bkey_s_c old, struct bkey_i *new)
 162 {
 163         struct btree *b;
 164         struct bkey_buf tmp;
 165         int ret;
 166
 167         bch2_bkey_buf_init(&tmp);
 168         bch2_bkey_buf_reassemble(&tmp, c, old);
 169
 170         b = bch2_btree_node_get_noiter(c, tmp.k, btree, level, true);
 171         if (!IS_ERR_OR_NULL(b)) {
 172                 mutex_lock(&c->btree_cache.lock);
 173
 174                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 175
 176                 bkey_copy(&b->key, new);
 177                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 178                 BUG_ON(ret);
 179
 180                 mutex_unlock(&c->btree_cache.lock);
 181                 six_unlock_read(&b->c.lock);
 182         }
 183
 184         bch2_bkey_buf_exit(&tmp, c);
 185 }
 186
 187 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 188 {
 189         struct bkey_i_btree_ptr_v2 *new;
 190         int ret;
 191
 192         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 193         if (!new)
 194                 return -ENOMEM;
 195
 196         btree_ptr_to_v2(b, new);
 197         b->data->min_key        = new_min;
 198         new->v.min_key          = new_min;
 199         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 200
 201         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 202         if (ret) {
 203                 kfree(new);
 204                 return ret;
 205         }
 206
 207         bch2_btree_node_drop_keys_outside_node(b);
 208
 209         return 0;
 210 }
 211
 212 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 213 {
 214         struct bkey_i_btree_ptr_v2 *new;
 215         int ret;
 216
 217         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 218         if (ret)
 219                 return ret;
 220
 221         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 222         if (!new)
 223                 return -ENOMEM;
 224
 225         btree_ptr_to_v2(b, new);
 226         b->data->max_key        = new_max;
 227         new->k.p                = new_max;
 228         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 229
 230         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 231         if (ret) {
 232                 kfree(new);
 233                 return ret;
 234         }
 235
 236         bch2_btree_node_drop_keys_outside_node(b);
 237
 238         mutex_lock(&c->btree_cache.lock);
 239         bch2_btree_node_hash_remove(&c->btree_cache, b);
 240
 241         bkey_copy(&b->key, &new->k_i);
 242         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 243         BUG_ON(ret);
 244         mutex_unlock(&c->btree_cache.lock);
 245         return 0;
 246 }
 247
 248 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 249                                         struct btree *prev, struct btree *cur)
 250 {
 251         struct bpos expected_start = !prev
 252                 ? b->data->min_key
 253                 : bpos_successor(prev->key.k.p);
 254         char buf1[200], buf2[200];
 255         int ret = 0;
 256
 257         if (!prev) {
 258                 struct printbuf out = PBUF(buf1);
 259                 pr_buf(&out, "start of node: ");
 260                 bch2_bpos_to_text(&out, b->data->min_key);
 261         } else {
 262                 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
 263         }
 264
 265         bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key));
 266
 267         if (prev &&
 268             bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 269             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 270                 /* cur overwrites prev: */
 271
 272                 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
 273                                                  cur->data->min_key) >= 0, c,
 274                                 "btree node overwritten by next node at btree %s level %u:\n"
 275                                 "  node %s\n"
 276                                 "  next %s",
 277                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 278                                 buf1, buf2))
 279                         return DROP_PREV_NODE;
 280
 281                 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
 282                                                  bpos_predecessor(cur->data->min_key)), c,
 283                                 "btree node with incorrect max_key at btree %s level %u:\n"
 284                                 "  node %s\n"
 285                                 "  next %s",
 286                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 287                                 buf1, buf2))
 288                         ret = set_node_max(c, prev,
 289                                            bpos_predecessor(cur->data->min_key));
 290         } else {
 291                 /* prev overwrites cur: */
 292
 293                 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
 294                                                  cur->data->max_key) >= 0, c,
 295                                 "btree node overwritten by prev node at btree %s level %u:\n"
 296                                 "  prev %s\n"
 297                                 "  node %s",
 298                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 299                                 buf1, buf2))
 300                         return DROP_THIS_NODE;
 301
 302                 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 303                                 "btree node with incorrect min_key at btree %s level %u:\n"
 304                                 "  prev %s\n"
 305                                 "  node %s",
 306                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 307                                 buf1, buf2))
 308                     ret = set_node_min(c, cur, expected_start);
 309         }
 310 fsck_err:
 311         return ret;
 312 }
 313
 314 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 315                                  struct btree *child)
 316 {
 317         char buf1[200], buf2[200];
 318         int ret = 0;
 319
 320         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 321                         "btree node with incorrect max_key at btree %s level %u:\n"
 322                         "  %s\n"
 323                         "  expected %s",
 324                         bch2_btree_ids[b->c.btree_id], b->c.level,
 325                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
 326                         (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
 327                 ret = set_node_max(c, child, b->key.k.p);
 328                 if (ret)
 329                         return ret;
 330         }
 331 fsck_err:
 332         return ret;
 333 }
 334
 335 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 336 {
 337         struct btree_and_journal_iter iter;
 338         struct bkey_s_c k;
 339         struct bkey_buf prev_k, cur_k;
 340         struct btree *prev = NULL, *cur = NULL;
 341         bool have_child, dropped_children = false;
 342         char buf[200];
 343         int ret = 0;
 344
 345         if (!b->c.level)
 346                 return 0;
 347 again:
 348         prev = NULL;
 349         have_child = dropped_children = false;
 350         bch2_bkey_buf_init(&prev_k);
 351         bch2_bkey_buf_init(&cur_k);
 352         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 353
 354         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 355                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 356                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 357
 358                 bch2_btree_and_journal_iter_advance(&iter);
 359                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 360
 361                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 362                                         b->c.btree_id, b->c.level - 1,
 363                                         false);
 364                 ret = PTR_ERR_OR_ZERO(cur);
 365
 366                 if (mustfix_fsck_err_on(ret == -EIO, c,
 367                                 "Unreadable btree node at btree %s level %u:\n"
 368                                 "  %s",
 369                                 bch2_btree_ids[b->c.btree_id],
 370                                 b->c.level - 1,
 371                                 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
 372                         bch2_btree_node_evict(c, cur_k.k);
 373                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 374                                                       b->c.level, cur_k.k->k.p);
 375                         if (ret)
 376                                 break;
 377                         continue;
 378                 }
 379
 380                 if (ret) {
 381                         bch_err(c, "%s: error %i getting btree node",
 382                                 __func__, ret);
 383                         break;
 384                 }
 385
 386                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 387
 388                 if (ret == DROP_THIS_NODE) {
 389                         six_unlock_read(&cur->c.lock);
 390                         bch2_btree_node_evict(c, cur_k.k);
 391                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 392                                                       b->c.level, cur_k.k->k.p);
 393                         if (ret)
 394                                 break;
 395                         continue;
 396                 }
 397
 398                 if (prev)
 399                         six_unlock_read(&prev->c.lock);
 400                 prev = NULL;
 401
 402                 if (ret == DROP_PREV_NODE) {
 403                         bch2_btree_node_evict(c, prev_k.k);
 404                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 405                                                       b->c.level, prev_k.k->k.p);
 406                         if (ret)
 407                                 break;
 408
 409                         bch2_btree_and_journal_iter_exit(&iter);
 410                         bch2_bkey_buf_exit(&prev_k, c);
 411                         bch2_bkey_buf_exit(&cur_k, c);
 412                         goto again;
 413                 } else if (ret)
 414                         break;
 415
 416                 prev = cur;
 417                 cur = NULL;
 418                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 419         }
 420
 421         if (!ret && !IS_ERR_OR_NULL(prev)) {
 422                 BUG_ON(cur);
 423                 ret = btree_repair_node_end(c, b, prev);
 424         }
 425
 426         if (!IS_ERR_OR_NULL(prev))
 427                 six_unlock_read(&prev->c.lock);
 428         prev = NULL;
 429         if (!IS_ERR_OR_NULL(cur))
 430                 six_unlock_read(&cur->c.lock);
 431         cur = NULL;
 432
 433         if (ret)
 434                 goto err;
 435
 436         bch2_btree_and_journal_iter_exit(&iter);
 437         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 438
 439         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 440                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 441                 bch2_btree_and_journal_iter_advance(&iter);
 442
 443                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 444                                         b->c.btree_id, b->c.level - 1,
 445                                         false);
 446                 ret = PTR_ERR_OR_ZERO(cur);
 447
 448                 if (ret) {
 449                         bch_err(c, "%s: error %i getting btree node",
 450                                 __func__, ret);
 451                         goto err;
 452                 }
 453
 454                 ret = bch2_btree_repair_topology_recurse(c, cur);
 455                 six_unlock_read(&cur->c.lock);
 456                 cur = NULL;
 457
 458                 if (ret == DROP_THIS_NODE) {
 459                         bch2_btree_node_evict(c, cur_k.k);
 460                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 461                                                       b->c.level, cur_k.k->k.p);
 462                         dropped_children = true;
 463                 }
 464
 465                 if (ret)
 466                         goto err;
 467
 468                 have_child = true;
 469         }
 470
 471         if (mustfix_fsck_err_on(!have_child, c,
 472                         "empty interior btree node at btree %s level %u\n"
 473                         "  %s",
 474                         bch2_btree_ids[b->c.btree_id],
 475                         b->c.level,
 476                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
 477                 ret = DROP_THIS_NODE;
 478 err:
 479 fsck_err:
 480         if (!IS_ERR_OR_NULL(prev))
 481                 six_unlock_read(&prev->c.lock);
 482         if (!IS_ERR_OR_NULL(cur))
 483                 six_unlock_read(&cur->c.lock);
 484
 485         bch2_btree_and_journal_iter_exit(&iter);
 486         bch2_bkey_buf_exit(&prev_k, c);
 487         bch2_bkey_buf_exit(&cur_k, c);
 488
 489         if (!ret && dropped_children)
 490                 goto again;
 491
 492         return ret;
 493 }
 494
 495 static int bch2_repair_topology(struct bch_fs *c)
 496 {
 497         struct btree *b;
 498         unsigned i;
 499         int ret = 0;
 500
 501         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 502                 b = c->btree_roots[i].b;
 503                 if (btree_node_fake(b))
 504                         continue;
 505
 506                 six_lock_read(&b->c.lock, NULL, NULL);
 507                 ret = bch2_btree_repair_topology_recurse(c, b);
 508                 six_unlock_read(&b->c.lock);
 509
 510                 if (ret == DROP_THIS_NODE) {
 511                         bch_err(c, "empty btree root - repair unimplemented");
 512                         ret = FSCK_ERR_EXIT;
 513                 }
 514         }
 515
 516         return ret;
 517 }
 518
 519 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 520                                unsigned level, bool is_root,
 521                                struct bkey_s_c *k)
 522 {
 523         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 524         const union bch_extent_entry *entry;
 525         struct extent_ptr_decoded p = { 0 };
 526         bool do_update = false;
 527         char buf[200];
 528         int ret = 0;
 529
 530         /*
 531          * XXX
 532          * use check_bucket_ref here
 533          */
 534         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 535                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 536                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 537                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 538
 539                 if (fsck_err_on(!g->gen_valid, c,
 540                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 541                                 "while marking %s",
 542                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 543                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 544                                 p.ptr.gen,
 545                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 546                         if (!p.ptr.cached) {
 547                                 g->_mark.gen            = p.ptr.gen;
 548                                 g->gen_valid            = true;
 549                         } else {
 550                                 do_update = true;
 551                         }
 552                 }
 553
 554                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 555                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 556                                 "while marking %s",
 557                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 558                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 559                                 p.ptr.gen, g->mark.gen,
 560                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 561                         if (!p.ptr.cached) {
 562                                 g->_mark.gen            = p.ptr.gen;
 563                                 g->gen_valid            = true;
 564                                 g->_mark.data_type      = 0;
 565                                 g->_mark.dirty_sectors  = 0;
 566                                 g->_mark.cached_sectors = 0;
 567                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 568                         } else {
 569                                 do_update = true;
 570                         }
 571                 }
 572
 573                 if (fsck_err_on(gen_cmp(g->mark.gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
 574                                 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 575                                 "while marking %s",
 576                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->mark.gen,
 577                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 578                                 p.ptr.gen,
 579                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 580                         do_update = true;
 581
 582                 if (fsck_err_on(!p.ptr.cached &&
 583                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 584                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 585                                 "while marking %s",
 586                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 587                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 588                                 p.ptr.gen, g->mark.gen,
 589                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 590                         do_update = true;
 591
 592                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->mark.gen)
 593                         continue;
 594
 595                 if (fsck_err_on(g->mark.data_type &&
 596                                 g->mark.data_type != data_type, c,
 597                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 598                                 "while marking %s",
 599                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 600                                 bch2_data_types[g->mark.data_type],
 601                                 bch2_data_types[data_type],
 602                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 603                         if (data_type == BCH_DATA_btree) {
 604                                 g->_mark.data_type      = data_type;
 605                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 606                         } else {
 607                                 do_update = true;
 608                         }
 609                 }
 610
 611                 if (p.has_ec) {
 612                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 613
 614                         if (fsck_err_on(!m || !m->alive, c,
 615                                         "pointer to nonexistent stripe %llu\n"
 616                                         "while marking %s",
 617                                         (u64) p.ec.idx,
 618                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 619                                 do_update = true;
 620
 621                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 622                                         "pointer does not match stripe %llu\n"
 623                                         "while marking %s",
 624                                         (u64) p.ec.idx,
 625                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 626                                 do_update = true;
 627                 }
 628         }
 629
 630         if (do_update) {
 631                 struct bkey_ptrs ptrs;
 632                 union bch_extent_entry *entry;
 633                 struct bch_extent_ptr *ptr;
 634                 struct bkey_i *new;
 635
 636                 if (is_root) {
 637                         bch_err(c, "cannot update btree roots yet");
 638                         return -EINVAL;
 639                 }
 640
 641                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 642                 if (!new) {
 643                         bch_err(c, "%s: error allocating new key", __func__);
 644                         return -ENOMEM;
 645                 }
 646
 647                 bkey_reassemble(new, *k);
 648
 649                 if (level) {
 650                         /*
 651                          * We don't want to drop btree node pointers - if the
 652                          * btree node isn't there anymore, the read path will
 653                          * sort it out:
 654                          */
 655                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 656                         bkey_for_each_ptr(ptrs, ptr) {
 657                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 658                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 659
 660                                 ptr->gen = g->mark.gen;
 661                         }
 662                 } else {
 663                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 664                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 665                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 666                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 667
 668                                 (ptr->cached &&
 669                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 670                                 (!ptr->cached &&
 671                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 672                                 gen_cmp(g->mark.gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 673                                 (g->mark.data_type &&
 674                                  g->mark.data_type != data_type);
 675                         }));
 676 again:
 677                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 678                         bkey_extent_entry_for_each(ptrs, entry) {
 679                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 680                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 681                                                                         entry->stripe_ptr.idx);
 682                                         union bch_extent_entry *next_ptr;
 683
 684                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 685                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 686                                                         goto found;
 687                                         next_ptr = NULL;
 688 found:
 689                                         if (!next_ptr) {
 690                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 691                                                 continue;
 692                                         }
 693
 694                                         if (!m || !m->alive ||
 695                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 696                                                                        &next_ptr->ptr,
 697                                                                        m->sectors)) {
 698                                                 bch2_bkey_extent_entry_drop(new, entry);
 699                                                 goto again;
 700                                         }
 701                                 }
 702                         }
 703                 }
 704
 705                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 706                 if (ret) {
 707                         kfree(new);
 708                         return ret;
 709                 }
 710
 711                 if (level)
 712                         bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
 713
 714                 bch2_bkey_val_to_text(&PBUF(buf), c, *k);
 715                 bch_info(c, "updated %s", buf);
 716                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(new));
 717                 bch_info(c, "new key %s", buf);
 718                 *k = bkey_i_to_s_c(new);
 719         }
 720 fsck_err:
 721         return ret;
 722 }
 723
 724 /* marking of btree keys/nodes: */
 725
 726 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 727                             unsigned level, bool is_root,
 728                             struct bkey_s_c *k,
 729                             bool initial)
 730 {
 731         struct bch_fs *c = trans->c;
 732         struct bkey deleted = KEY(0, 0, 0);
 733         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 734         unsigned flags =
 735                 BTREE_TRIGGER_GC|
 736                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 737         int ret = 0;
 738
 739         deleted.p = k->k->p;
 740
 741         if (initial) {
 742                 BUG_ON(bch2_journal_seq_verify &&
 743                        k->k->version.lo > journal_cur_seq(&c->journal));
 744
 745                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 746                 if (ret)
 747                         goto err;
 748
 749                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 750                                 "key version number higher than recorded: %llu > %llu",
 751                                 k->k->version.lo,
 752                                 atomic64_read(&c->key_version)))
 753                         atomic64_set(&c->key_version, k->k->version.lo);
 754         }
 755
 756         ret = bch2_mark_key(trans, old, *k, flags);
 757 fsck_err:
 758 err:
 759         if (ret)
 760                 bch_err(c, "%s: ret %i", __func__, ret);
 761         return ret;
 762 }
 763
 764 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 765 {
 766         struct bch_fs *c = trans->c;
 767         struct btree_node_iter iter;
 768         struct bkey unpacked;
 769         struct bkey_s_c k;
 770         struct bkey_buf prev, cur;
 771         int ret = 0;
 772
 773         if (!btree_node_type_needs_gc(btree_node_type(b)))
 774                 return 0;
 775
 776         bch2_btree_node_iter_init_from_start(&iter, b);
 777         bch2_bkey_buf_init(&prev);
 778         bch2_bkey_buf_init(&cur);
 779         bkey_init(&prev.k->k);
 780
 781         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 782                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 783                                        &k, initial);
 784                 if (ret)
 785                         break;
 786
 787                 bch2_btree_node_iter_advance(&iter, b);
 788
 789                 if (b->c.level) {
 790                         bch2_bkey_buf_reassemble(&cur, c, k);
 791
 792                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 793                                         bch2_btree_node_iter_end(&iter));
 794                         if (ret)
 795                                 break;
 796                 }
 797         }
 798
 799         bch2_bkey_buf_exit(&cur, c);
 800         bch2_bkey_buf_exit(&prev, c);
 801         return ret;
 802 }
 803
 804 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 805                          bool initial, bool metadata_only)
 806 {
 807         struct bch_fs *c = trans->c;
 808         struct btree_iter iter;
 809         struct btree *b;
 810         unsigned depth = metadata_only                  ? 1
 811                 : bch2_expensive_debug_checks           ? 0
 812                 : !btree_node_type_needs_gc(btree_id)   ? 1
 813                 : 0;
 814         int ret = 0;
 815
 816         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 817
 818         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 819                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 820                 bch2_verify_btree_nr_keys(b);
 821
 822                 gc_pos_set(c, gc_pos_btree_node(b));
 823
 824                 ret = btree_gc_mark_node(trans, b, initial);
 825                 if (ret)
 826                         break;
 827         }
 828         bch2_trans_iter_exit(trans, &iter);
 829
 830         if (ret)
 831                 return ret;
 832
 833         mutex_lock(&c->btree_root_lock);
 834         b = c->btree_roots[btree_id].b;
 835         if (!btree_node_fake(b)) {
 836                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 837
 838                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 839                                        true, &k, initial);
 840         }
 841         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 842         mutex_unlock(&c->btree_root_lock);
 843
 844         return ret;
 845 }
 846
 847 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 848                                       unsigned target_depth)
 849 {
 850         struct bch_fs *c = trans->c;
 851         struct btree_and_journal_iter iter;
 852         struct bkey_s_c k;
 853         struct bkey_buf cur, prev;
 854         char buf[200];
 855         int ret = 0;
 856
 857         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 858         bch2_bkey_buf_init(&prev);
 859         bch2_bkey_buf_init(&cur);
 860         bkey_init(&prev.k->k);
 861
 862         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 863                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 864                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 865
 866                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 867                                        false, &k, true);
 868                 if (ret) {
 869                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 870                         goto fsck_err;
 871                 }
 872
 873                 if (b->c.level) {
 874                         bch2_bkey_buf_reassemble(&cur, c, k);
 875                         k = bkey_i_to_s_c(cur.k);
 876
 877                         bch2_btree_and_journal_iter_advance(&iter);
 878
 879                         ret = bch2_gc_check_topology(c, b,
 880                                         &prev, cur,
 881                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 882                         if (ret)
 883                                 goto fsck_err;
 884                 } else {
 885                         bch2_btree_and_journal_iter_advance(&iter);
 886                 }
 887         }
 888
 889         if (b->c.level > target_depth) {
 890                 bch2_btree_and_journal_iter_exit(&iter);
 891                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 892
 893                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 894                         struct btree *child;
 895
 896                         bch2_bkey_buf_reassemble(&cur, c, k);
 897                         bch2_btree_and_journal_iter_advance(&iter);
 898
 899                         child = bch2_btree_node_get_noiter(c, cur.k,
 900                                                 b->c.btree_id, b->c.level - 1,
 901                                                 false);
 902                         ret = PTR_ERR_OR_ZERO(child);
 903
 904                         if (ret == -EIO) {
 905                                 bch2_topology_error(c);
 906
 907                                 if (__fsck_err(c,
 908                                           FSCK_CAN_FIX|
 909                                           FSCK_CAN_IGNORE|
 910                                           FSCK_NO_RATELIMIT,
 911                                           "Unreadable btree node at btree %s level %u:\n"
 912                                           "  %s",
 913                                           bch2_btree_ids[b->c.btree_id],
 914                                           b->c.level - 1,
 915                                           (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
 916                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 917                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 918                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 919                                         goto fsck_err;
 920                                 } else {
 921                                         /* Continue marking when opted to not
 922                                          * fix the error: */
 923                                         ret = 0;
 924                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 925                                         continue;
 926                                 }
 927                         } else if (ret) {
 928                                 bch_err(c, "%s: error %i getting btree node",
 929                                         __func__, ret);
 930                                 break;
 931                         }
 932
 933                         ret = bch2_gc_btree_init_recurse(trans, child,
 934                                                          target_depth);
 935                         six_unlock_read(&child->c.lock);
 936
 937                         if (ret)
 938                                 break;
 939                 }
 940         }
 941 fsck_err:
 942         bch2_bkey_buf_exit(&cur, c);
 943         bch2_bkey_buf_exit(&prev, c);
 944         bch2_btree_and_journal_iter_exit(&iter);
 945         return ret;
 946 }
 947
 948 static int bch2_gc_btree_init(struct btree_trans *trans,
 949                               enum btree_id btree_id,
 950                               bool metadata_only)
 951 {
 952         struct bch_fs *c = trans->c;
 953         struct btree *b;
 954         unsigned target_depth = metadata_only           ? 1
 955                 : bch2_expensive_debug_checks           ? 0
 956                 : !btree_node_type_needs_gc(btree_id)   ? 1
 957                 : 0;
 958         char buf[100];
 959         int ret = 0;
 960
 961         b = c->btree_roots[btree_id].b;
 962
 963         if (btree_node_fake(b))
 964                 return 0;
 965
 966         six_lock_read(&b->c.lock, NULL, NULL);
 967         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 968                         "btree root with incorrect min_key: %s",
 969                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 970                 bch_err(c, "repair unimplemented");
 971                 ret = FSCK_ERR_EXIT;
 972                 goto fsck_err;
 973         }
 974
 975         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
 976                         "btree root with incorrect max_key: %s",
 977                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 978                 bch_err(c, "repair unimplemented");
 979                 ret = FSCK_ERR_EXIT;
 980                 goto fsck_err;
 981         }
 982
 983         if (b->c.level >= target_depth)
 984                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
 985
 986         if (!ret) {
 987                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 988
 989                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
 990                                        &k, true);
 991         }
 992 fsck_err:
 993         six_unlock_read(&b->c.lock);
 994
 995         if (ret < 0)
 996                 bch_err(c, "%s: ret %i", __func__, ret);
 997         return ret;
 998 }
 999
1000 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1001 {
1002         return  (int) btree_id_to_gc_phase(l) -
1003                 (int) btree_id_to_gc_phase(r);
1004 }
1005
1006 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1007 {
1008         struct btree_trans trans;
1009         enum btree_id ids[BTREE_ID_NR];
1010         unsigned i;
1011         int ret = 0;
1012
1013         bch2_trans_init(&trans, c, 0, 0);
1014
1015         for (i = 0; i < BTREE_ID_NR; i++)
1016                 ids[i] = i;
1017         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1018
1019         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1020                 ret = initial
1021                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1022                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1023
1024         if (ret < 0)
1025                 bch_err(c, "%s: ret %i", __func__, ret);
1026
1027         bch2_trans_exit(&trans);
1028         return ret;
1029 }
1030
1031 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1032                                   u64 start, u64 end,
1033                                   enum bch_data_type type,
1034                                   unsigned flags)
1035 {
1036         u64 b = sector_to_bucket(ca, start);
1037
1038         do {
1039                 unsigned sectors =
1040                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1041
1042                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1043                                           gc_phase(GC_PHASE_SB), flags);
1044                 b++;
1045                 start += sectors;
1046         } while (start < end);
1047 }
1048
1049 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1050                                      unsigned flags)
1051 {
1052         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1053         unsigned i;
1054         u64 b;
1055
1056         for (i = 0; i < layout->nr_superblocks; i++) {
1057                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1058
1059                 if (offset == BCH_SB_SECTOR)
1060                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1061                                               BCH_DATA_sb, flags);
1062
1063                 mark_metadata_sectors(c, ca, offset,
1064                                       offset + (1 << layout->sb_max_size_bits),
1065                                       BCH_DATA_sb, flags);
1066         }
1067
1068         for (i = 0; i < ca->journal.nr; i++) {
1069                 b = ca->journal.buckets[i];
1070                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1071                                           ca->mi.bucket_size,
1072                                           gc_phase(GC_PHASE_SB), flags);
1073         }
1074 }
1075
1076 static void bch2_mark_superblocks(struct bch_fs *c)
1077 {
1078         struct bch_dev *ca;
1079         unsigned i;
1080
1081         mutex_lock(&c->sb_lock);
1082         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1083
1084         for_each_online_member(ca, c, i)
1085                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1086         mutex_unlock(&c->sb_lock);
1087 }
1088
1089 #if 0
1090 /* Also see bch2_pending_btree_node_free_insert_done() */
1091 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1092 {
1093         struct btree_update *as;
1094         struct pending_btree_node_free *d;
1095
1096         mutex_lock(&c->btree_interior_update_lock);
1097         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1098
1099         for_each_pending_btree_node_free(c, as, d)
1100                 if (d->index_update_done)
1101                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1102
1103         mutex_unlock(&c->btree_interior_update_lock);
1104 }
1105 #endif
1106
1107 static void bch2_gc_free(struct bch_fs *c)
1108 {
1109         struct bch_dev *ca;
1110         unsigned i;
1111
1112         genradix_free(&c->reflink_gc_table);
1113         genradix_free(&c->gc_stripes);
1114
1115         for_each_member_device(ca, c, i) {
1116                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1117                         sizeof(struct bucket_array) +
1118                         ca->mi.nbuckets * sizeof(struct bucket));
1119                 ca->buckets[1] = NULL;
1120
1121                 free_percpu(ca->usage_gc);
1122                 ca->usage_gc = NULL;
1123         }
1124
1125         free_percpu(c->usage_gc);
1126         c->usage_gc = NULL;
1127 }
1128
1129 static int bch2_gc_done(struct bch_fs *c,
1130                         bool initial, bool metadata_only)
1131 {
1132         struct bch_dev *ca = NULL;
1133         bool verify = !metadata_only && (!initial ||
1134                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1135         unsigned i, dev;
1136         int ret = 0;
1137
1138         percpu_down_write(&c->mark_lock);
1139
1140 #define copy_field(_f, _msg, ...)                                       \
1141         if (dst->_f != src->_f) {                                       \
1142                 if (verify)                                             \
1143                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1144                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1145                 dst->_f = src->_f;                                      \
1146         }
1147 #define copy_stripe_field(_f, _msg, ...)                                \
1148         if (dst->_f != src->_f) {                                       \
1149                 if (verify)                                             \
1150                         fsck_err(c, "stripe %zu has wrong "_msg         \
1151                                 ": got %u, should be %u",               \
1152                                 iter.pos, ##__VA_ARGS__,                \
1153                                 dst->_f, src->_f);                      \
1154                 dst->_f = src->_f;                                      \
1155         }
1156 #define copy_dev_field(_f, _msg, ...)                                   \
1157         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1158 #define copy_fs_field(_f, _msg, ...)                                    \
1159         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1160
1161         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1162                 bch2_fs_usage_acc_to_base(c, i);
1163
1164         for_each_member_device(ca, c, dev) {
1165                 struct bch_dev_usage *dst = ca->usage_base;
1166                 struct bch_dev_usage *src = (void *)
1167                         bch2_acc_percpu_u64s((void *) ca->usage_gc,
1168                                              dev_usage_u64s());
1169
1170                 copy_dev_field(buckets_ec,              "buckets_ec");
1171                 copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1172
1173                 for (i = 0; i < BCH_DATA_NR; i++) {
1174                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1175                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1176                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1177                 }
1178         };
1179
1180         {
1181                 unsigned nr = fs_usage_u64s(c);
1182                 struct bch_fs_usage *dst = c->usage_base;
1183                 struct bch_fs_usage *src = (void *)
1184                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1185
1186                 copy_fs_field(hidden,           "hidden");
1187                 copy_fs_field(btree,            "btree");
1188
1189                 if (!metadata_only) {
1190                         copy_fs_field(data,     "data");
1191                         copy_fs_field(cached,   "cached");
1192                         copy_fs_field(reserved, "reserved");
1193                         copy_fs_field(nr_inodes,"nr_inodes");
1194
1195                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1196                                 copy_fs_field(persistent_reserved[i],
1197                                               "persistent_reserved[%i]", i);
1198                 }
1199
1200                 for (i = 0; i < c->replicas.nr; i++) {
1201                         struct bch_replicas_entry *e =
1202                                 cpu_replicas_entry(&c->replicas, i);
1203                         char buf[80];
1204
1205                         if (metadata_only &&
1206                             (e->data_type == BCH_DATA_user ||
1207                              e->data_type == BCH_DATA_cached))
1208                                 continue;
1209
1210                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1211
1212                         copy_fs_field(replicas[i], "%s", buf);
1213                 }
1214         }
1215
1216 #undef copy_fs_field
1217 #undef copy_dev_field
1218 #undef copy_stripe_field
1219 #undef copy_field
1220 fsck_err:
1221         if (ca)
1222                 percpu_ref_put(&ca->ref);
1223         if (ret)
1224                 bch_err(c, "%s: ret %i", __func__, ret);
1225
1226         percpu_up_write(&c->mark_lock);
1227         return ret;
1228 }
1229
1230 static int bch2_gc_start(struct bch_fs *c,
1231                          bool metadata_only)
1232 {
1233         struct bch_dev *ca = NULL;
1234         unsigned i;
1235
1236         BUG_ON(c->usage_gc);
1237
1238         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1239                                          sizeof(u64), GFP_KERNEL);
1240         if (!c->usage_gc) {
1241                 bch_err(c, "error allocating c->usage_gc");
1242                 return -ENOMEM;
1243         }
1244
1245         for_each_member_device(ca, c, i) {
1246                 BUG_ON(ca->buckets[1]);
1247                 BUG_ON(ca->usage_gc);
1248
1249                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1250                 if (!ca->usage_gc) {
1251                         bch_err(c, "error allocating ca->usage_gc");
1252                         percpu_ref_put(&ca->ref);
1253                         return -ENOMEM;
1254                 }
1255         }
1256
1257         return 0;
1258 }
1259
1260 static int bch2_alloc_write_key(struct btree_trans *trans,
1261                                 struct btree_iter *iter,
1262                                 bool initial, bool metadata_only)
1263 {
1264         struct bch_fs *c = trans->c;
1265         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1266         struct bucket *g;
1267         struct bkey_s_c k;
1268         struct bkey_alloc_unpacked old_u, new_u, gc_u;
1269         struct bkey_alloc_buf *a;
1270         int ret;
1271
1272         k = bch2_btree_iter_peek_slot(iter);
1273         ret = bkey_err(k);
1274         if (ret)
1275                 return ret;
1276
1277         old_u = new_u = bch2_alloc_unpack(k);
1278
1279         percpu_down_read(&c->mark_lock);
1280         g       = gc_bucket(ca, iter->pos.offset);
1281         gc_u = (struct bkey_alloc_unpacked) {
1282                 .dev            = iter->pos.inode,
1283                 .bucket         = iter->pos.offset,
1284                 .gen            = g->mark.gen,
1285                 .data_type      = g->mark.data_type,
1286                 .dirty_sectors  = g->mark.dirty_sectors,
1287                 .cached_sectors = g->mark.cached_sectors,
1288                 .read_time      = g->io_time[READ],
1289                 .write_time     = g->io_time[WRITE],
1290                 .stripe         = g->stripe,
1291                 .stripe_redundancy = g->stripe_redundancy,
1292         };
1293         percpu_up_read(&c->mark_lock);
1294
1295         if (metadata_only &&
1296             gc_u.data_type != BCH_DATA_sb &&
1297             gc_u.data_type != BCH_DATA_journal &&
1298             gc_u.data_type != BCH_DATA_btree)
1299                 return 0;
1300
1301         if (gen_after(old_u.gen, gc_u.gen))
1302                 return 0;
1303
1304 #define copy_bucket_field(_f)                                           \
1305         if (fsck_err_on(new_u._f != gc_u._f, c,                         \
1306                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1307                         ": got %u, should be %u",                       \
1308                         iter->pos.inode, iter->pos.offset,              \
1309                         new_u.gen,                                      \
1310                         bch2_data_types[new_u.data_type],               \
1311                         new_u._f, gc_u._f))                             \
1312                 new_u._f = gc_u._f;                                     \
1313
1314         copy_bucket_field(gen);
1315         copy_bucket_field(data_type);
1316         copy_bucket_field(stripe);
1317         copy_bucket_field(dirty_sectors);
1318         copy_bucket_field(cached_sectors);
1319         copy_bucket_field(stripe_redundancy);
1320         copy_bucket_field(stripe);
1321 #undef copy_bucket_field
1322
1323         if (!bkey_alloc_unpacked_cmp(old_u, new_u))
1324                 return 0;
1325
1326         a = bch2_alloc_pack(trans, new_u);
1327         if (IS_ERR(a))
1328                 return PTR_ERR(a);
1329
1330         ret = initial
1331                 ? bch2_journal_key_insert(c, BTREE_ID_alloc, 0, &a->k)
1332                 : bch2_trans_update(trans, iter, &a->k, BTREE_TRIGGER_NORUN);
1333 fsck_err:
1334         return ret;
1335 }
1336
1337 static int bch2_gc_alloc_done(struct bch_fs *c, bool initial, bool metadata_only)
1338 {
1339         struct btree_trans trans;
1340         struct btree_iter iter;
1341         struct bkey_s_c k;
1342         struct bch_dev *ca;
1343         unsigned i;
1344         int ret = 0;
1345
1346         bch2_trans_init(&trans, c, 0, 0);
1347
1348         for_each_member_device(ca, c, i) {
1349                 for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1350                                    POS(ca->dev_idx, ca->mi.first_bucket),
1351                                    BTREE_ITER_SLOTS|
1352                                    BTREE_ITER_PREFETCH, k, ret) {
1353                         if (bkey_cmp(iter.pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1354                                 break;
1355
1356                         ret = __bch2_trans_do(&trans, NULL, NULL,
1357                                               BTREE_INSERT_LAZY_RW,
1358                                         bch2_alloc_write_key(&trans, &iter,
1359                                                              initial, metadata_only));
1360                         if (ret)
1361                                 break;
1362                 }
1363                 bch2_trans_iter_exit(&trans, &iter);
1364
1365                 if (ret) {
1366                         bch_err(c, "error writing alloc info: %i", ret);
1367                         percpu_ref_put(&ca->ref);
1368                         break;
1369                 }
1370         }
1371
1372         bch2_trans_exit(&trans);
1373         return ret;
1374 }
1375
1376 static int bch2_gc_alloc_start(struct bch_fs *c, bool initial, bool metadata_only)
1377 {
1378         struct bch_dev *ca;
1379         unsigned i;
1380
1381         for_each_member_device(ca, c, i) {
1382                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1383                                 ca->mi.nbuckets * sizeof(struct bucket),
1384                                 GFP_KERNEL|__GFP_ZERO);
1385                 if (!buckets) {
1386                         percpu_ref_put(&ca->ref);
1387                         percpu_up_write(&c->mark_lock);
1388                         bch_err(c, "error allocating ca->buckets[gc]");
1389                         return -ENOMEM;
1390                 }
1391
1392                 buckets->first_bucket   = ca->mi.first_bucket;
1393                 buckets->nbuckets       = ca->mi.nbuckets;
1394                 rcu_assign_pointer(ca->buckets[1], buckets);
1395         };
1396
1397         return bch2_alloc_read(c, true, metadata_only);
1398 }
1399
1400 static void bch2_gc_alloc_reset(struct bch_fs *c, bool initial, bool metadata_only)
1401 {
1402         struct bch_dev *ca;
1403         unsigned i;
1404
1405         for_each_member_device(ca, c, i) {
1406                 struct bucket_array *buckets = __bucket_array(ca, true);
1407                 struct bucket *g;
1408
1409                 for_each_bucket(g, buckets) {
1410                         if (metadata_only &&
1411                             (g->mark.data_type == BCH_DATA_user ||
1412                              g->mark.data_type == BCH_DATA_cached ||
1413                              g->mark.data_type == BCH_DATA_parity))
1414                                 continue;
1415                         g->_mark.dirty_sectors = 0;
1416                         g->_mark.cached_sectors = 0;
1417                 }
1418         };
1419 }
1420
1421 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1422                                 bool metadata_only)
1423 {
1424         struct btree_trans trans;
1425         struct btree_iter iter;
1426         struct bkey_s_c k;
1427         struct reflink_gc *r;
1428         size_t idx = 0;
1429         char buf[200];
1430         int ret = 0;
1431
1432         if (metadata_only)
1433                 return 0;
1434
1435         bch2_trans_init(&trans, c, 0, 0);
1436
1437         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1438                            BTREE_ITER_PREFETCH, k, ret) {
1439                 const __le64 *refcount = bkey_refcount_c(k);
1440
1441                 if (!refcount)
1442                         continue;
1443
1444                 r = genradix_ptr(&c->reflink_gc_table, idx++);
1445                 if (!r ||
1446                     r->offset != k.k->p.offset ||
1447                     r->size != k.k->size) {
1448                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1449                         ret = -EINVAL;
1450                         break;
1451                 }
1452
1453                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1454                                 "reflink key has wrong refcount:\n"
1455                                 "  %s\n"
1456                                 "  should be %u",
1457                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1458                                 r->refcount)) {
1459                         struct bkey_i *new;
1460
1461                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1462                         if (!new) {
1463                                 ret = -ENOMEM;
1464                                 break;
1465                         }
1466
1467                         bkey_reassemble(new, k);
1468
1469                         if (!r->refcount) {
1470                                 new->k.type = KEY_TYPE_deleted;
1471                                 /*
1472                                  * XXX ugly: bch2_journal_key_insert() queues up
1473                                  * the key for the journal replay code, which
1474                                  * doesn't run the extent overwrite pass
1475                                  */
1476                                 if (initial)
1477                                         new->k.size = 0;
1478                         } else {
1479                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1480                         }
1481
1482                         ret = initial
1483                                ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, new)
1484                                : __bch2_trans_do(&trans, NULL, NULL, 0,
1485                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1486                         kfree(new);
1487
1488                         if (ret)
1489                                 break;
1490                 }
1491         }
1492 fsck_err:
1493         bch2_trans_iter_exit(&trans, &iter);
1494         c->reflink_gc_nr = 0;
1495         bch2_trans_exit(&trans);
1496         return ret;
1497 }
1498
1499 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1500                                  bool metadata_only)
1501 {
1502         struct btree_trans trans;
1503         struct btree_iter iter;
1504         struct bkey_s_c k;
1505         struct reflink_gc *r;
1506         int ret = 0;
1507
1508         if (metadata_only)
1509                 return 0;
1510
1511         bch2_trans_init(&trans, c, 0, 0);
1512         c->reflink_gc_nr = 0;
1513
1514         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1515                            BTREE_ITER_PREFETCH, k, ret) {
1516                 const __le64 *refcount = bkey_refcount_c(k);
1517
1518                 if (!refcount)
1519                         continue;
1520
1521                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1522                                        GFP_KERNEL);
1523                 if (!r) {
1524                         ret = -ENOMEM;
1525                         break;
1526                 }
1527
1528                 r->offset       = k.k->p.offset;
1529                 r->size         = k.k->size;
1530                 r->refcount     = 0;
1531         }
1532         bch2_trans_iter_exit(&trans, &iter);
1533
1534         bch2_trans_exit(&trans);
1535         return ret;
1536 }
1537
1538 static void bch2_gc_reflink_reset(struct bch_fs *c, bool initial,
1539                                   bool metadata_only)
1540 {
1541         struct genradix_iter iter;
1542         struct reflink_gc *r;
1543
1544         genradix_for_each(&c->reflink_gc_table, iter, r)
1545                 r->refcount = 0;
1546 }
1547
1548 static int bch2_gc_stripes_done(struct bch_fs *c, bool initial,
1549                                 bool metadata_only)
1550 {
1551         struct btree_trans trans;
1552         struct btree_iter iter;
1553         struct bkey_s_c k;
1554         struct gc_stripe *m;
1555         const struct bch_stripe *s;
1556         char buf[200];
1557         unsigned i;
1558         int ret = 0;
1559
1560         if (metadata_only)
1561                 return 0;
1562
1563         bch2_trans_init(&trans, c, 0, 0);
1564
1565         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1566                            BTREE_ITER_PREFETCH, k, ret) {
1567                 if (k.k->type != KEY_TYPE_stripe)
1568                         continue;
1569
1570                 s = bkey_s_c_to_stripe(k).v;
1571                 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1572
1573                 for (i = 0; i < s->nr_blocks; i++)
1574                         if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1575                                 goto inconsistent;
1576                 continue;
1577 inconsistent:
1578                 if (fsck_err_on(true, c,
1579                                 "stripe has wrong block sector count %u:\n"
1580                                 "  %s\n"
1581                                 "  should be %u", i,
1582                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1583                                 m ? m->block_sectors[i] : 0)) {
1584                         struct bkey_i_stripe *new;
1585
1586                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1587                         if (!new) {
1588                                 ret = -ENOMEM;
1589                                 break;
1590                         }
1591
1592                         bkey_reassemble(&new->k_i, k);
1593
1594                         for (i = 0; i < new->v.nr_blocks; i++)
1595                                 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1596
1597                         ret = initial
1598                                 ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, &new->k_i)
1599                                 : __bch2_trans_do(&trans, NULL, NULL, 0,
1600                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1601                         kfree(new);
1602                 }
1603         }
1604 fsck_err:
1605         bch2_trans_iter_exit(&trans, &iter);
1606
1607         bch2_trans_exit(&trans);
1608         return ret;
1609 }
1610
1611 static void bch2_gc_stripes_reset(struct bch_fs *c, bool initial,
1612                                 bool metadata_only)
1613 {
1614         genradix_free(&c->gc_stripes);
1615 }
1616
1617 /**
1618  * bch2_gc - walk _all_ references to buckets, and recompute them:
1619  *
1620  * Order matters here:
1621  *  - Concurrent GC relies on the fact that we have a total ordering for
1622  *    everything that GC walks - see  gc_will_visit_node(),
1623  *    gc_will_visit_root()
1624  *
1625  *  - also, references move around in the course of index updates and
1626  *    various other crap: everything needs to agree on the ordering
1627  *    references are allowed to move around in - e.g., we're allowed to
1628  *    start with a reference owned by an open_bucket (the allocator) and
1629  *    move it to the btree, but not the reverse.
1630  *
1631  *    This is necessary to ensure that gc doesn't miss references that
1632  *    move around - if references move backwards in the ordering GC
1633  *    uses, GC could skip past them
1634  */
1635 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1636 {
1637         struct bch_dev *ca;
1638         u64 start_time = local_clock();
1639         unsigned i, iter = 0;
1640         int ret;
1641
1642         lockdep_assert_held(&c->state_lock);
1643         trace_gc_start(c);
1644
1645         down_write(&c->gc_lock);
1646
1647         /* flush interior btree updates: */
1648         closure_wait_event(&c->btree_interior_update_wait,
1649                            !bch2_btree_interior_updates_nr_pending(c));
1650
1651         ret   = bch2_gc_start(c, metadata_only) ?:
1652                 bch2_gc_alloc_start(c, initial, metadata_only) ?:
1653                 bch2_gc_reflink_start(c, initial, metadata_only);
1654         if (ret)
1655                 goto out;
1656 again:
1657         gc_pos_set(c, gc_phase(GC_PHASE_START));
1658
1659         bch2_mark_superblocks(c);
1660
1661         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1662             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1663             c->opts.fix_errors != FSCK_OPT_NO) {
1664                 bch_info(c, "starting topology repair pass");
1665                 ret = bch2_repair_topology(c);
1666                 if (ret)
1667                         goto out;
1668                 bch_info(c, "topology repair pass done");
1669
1670                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1671         }
1672
1673         ret = bch2_gc_btrees(c, initial, metadata_only);
1674
1675         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1676             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1677             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1678                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1679                 ret = 0;
1680         }
1681
1682         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1683                 ret = FSCK_ERR_EXIT;
1684
1685         if (ret)
1686                 goto out;
1687
1688 #if 0
1689         bch2_mark_pending_btree_node_frees(c);
1690 #endif
1691         c->gc_count++;
1692
1693         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1694             (!iter && bch2_test_restart_gc)) {
1695                 if (iter++ > 2) {
1696                         bch_info(c, "Unable to fix bucket gens, looping");
1697                         ret = -EINVAL;
1698                         goto out;
1699                 }
1700
1701                 /*
1702                  * XXX: make sure gens we fixed got saved
1703                  */
1704                 bch_info(c, "Second GC pass needed, restarting:");
1705                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1706                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1707
1708                 bch2_gc_stripes_reset(c, initial, metadata_only);
1709                 bch2_gc_alloc_reset(c, initial, metadata_only);
1710                 bch2_gc_reflink_reset(c, initial, metadata_only);
1711
1712                 /* flush fsck errors, reset counters */
1713                 bch2_flush_fsck_errs(c);
1714                 goto again;
1715         }
1716 out:
1717         if (!ret) {
1718                 bch2_journal_block(&c->journal);
1719
1720                 ret   = bch2_gc_stripes_done(c, initial, metadata_only) ?:
1721                         bch2_gc_reflink_done(c, initial, metadata_only) ?:
1722                         bch2_gc_alloc_done(c, initial, metadata_only) ?:
1723                         bch2_gc_done(c, initial, metadata_only);
1724
1725                 bch2_journal_unblock(&c->journal);
1726         }
1727
1728         percpu_down_write(&c->mark_lock);
1729         /* Indicates that gc is no longer in progress: */
1730         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1731
1732         bch2_gc_free(c);
1733         percpu_up_write(&c->mark_lock);
1734
1735         up_write(&c->gc_lock);
1736
1737         trace_gc_end(c);
1738         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1739
1740         /*
1741          * Wake up allocator in case it was waiting for buckets
1742          * because of not being able to inc gens
1743          */
1744         for_each_member_device(ca, c, i)
1745                 bch2_wake_allocator(ca);
1746
1747         /*
1748          * At startup, allocations can happen directly instead of via the
1749          * allocator thread - issue wakeup in case they blocked on gc_lock:
1750          */
1751         closure_wake_up(&c->freelist_wait);
1752         return ret;
1753 }
1754
1755 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1756 {
1757         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1758         const struct bch_extent_ptr *ptr;
1759
1760         percpu_down_read(&c->mark_lock);
1761         bkey_for_each_ptr(ptrs, ptr) {
1762                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1763
1764                 if (ptr_stale(ca, ptr) > 16) {
1765                         percpu_up_read(&c->mark_lock);
1766                         return true;
1767                 }
1768         }
1769
1770         bkey_for_each_ptr(ptrs, ptr) {
1771                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1772                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1773
1774                 if (gen_after(*gen, ptr->gen))
1775                         *gen = ptr->gen;
1776         }
1777         percpu_up_read(&c->mark_lock);
1778
1779         return false;
1780 }
1781
1782 /*
1783  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1784  * node pointers currently never have cached pointers that can become stale:
1785  */
1786 static int bch2_gc_btree_gens(struct btree_trans *trans, enum btree_id btree_id)
1787 {
1788         struct bch_fs *c = trans->c;
1789         struct btree_iter iter;
1790         struct bkey_s_c k;
1791         struct bkey_buf sk;
1792         int ret = 0, commit_err = 0;
1793
1794         bch2_bkey_buf_init(&sk);
1795
1796         bch2_trans_iter_init(trans, &iter, btree_id, POS_MIN,
1797                              BTREE_ITER_PREFETCH|
1798                              BTREE_ITER_NOT_EXTENTS|
1799                              BTREE_ITER_ALL_SNAPSHOTS);
1800
1801         while ((bch2_trans_begin(trans),
1802                 k = bch2_btree_iter_peek(&iter)).k) {
1803                 ret = bkey_err(k);
1804
1805                 if (ret == -EINTR)
1806                         continue;
1807                 if (ret)
1808                         break;
1809
1810                 c->gc_gens_pos = iter.pos;
1811
1812                 if (gc_btree_gens_key(c, k) && !commit_err) {
1813                         bch2_bkey_buf_reassemble(&sk, c, k);
1814                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1815
1816                         commit_err =
1817                                 bch2_trans_update(trans, &iter, sk.k, 0) ?:
1818                                 bch2_trans_commit(trans, NULL, NULL,
1819                                                   BTREE_INSERT_NOWAIT|
1820                                                   BTREE_INSERT_NOFAIL);
1821                         if (commit_err == -EINTR) {
1822                                 commit_err = 0;
1823                                 continue;
1824                         }
1825                 }
1826
1827                 bch2_btree_iter_advance(&iter);
1828         }
1829         bch2_trans_iter_exit(trans, &iter);
1830
1831         bch2_bkey_buf_exit(&sk, c);
1832
1833         return ret;
1834 }
1835
1836 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter)
1837 {
1838         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1839         struct bkey_s_c k;
1840         struct bkey_alloc_unpacked u;
1841         int ret;
1842
1843         k = bch2_btree_iter_peek_slot(iter);
1844         ret = bkey_err(k);
1845         if (ret)
1846                 return ret;
1847
1848         u = bch2_alloc_unpack(k);
1849
1850         if (u.oldest_gen == ca->oldest_gen[iter->pos.offset])
1851                 return 0;
1852
1853         u.oldest_gen = ca->oldest_gen[iter->pos.offset];
1854
1855         return bch2_alloc_write(trans, iter, &u, BTREE_TRIGGER_NORUN);
1856 }
1857
1858 int bch2_gc_gens(struct bch_fs *c)
1859 {
1860         struct btree_trans trans;
1861         struct btree_iter iter;
1862         struct bkey_s_c k;
1863         struct bch_dev *ca;
1864         u64 b, start_time = local_clock();
1865         unsigned i;
1866         int ret;
1867
1868         /*
1869          * Ideally we would be using state_lock and not gc_lock here, but that
1870          * introduces a deadlock in the RO path - we currently take the state
1871          * lock at the start of going RO, thus the gc thread may get stuck:
1872          */
1873         if (!mutex_trylock(&c->gc_gens_lock))
1874                 return 0;
1875
1876         down_read(&c->gc_lock);
1877         bch2_trans_init(&trans, c, 0, 0);
1878
1879         for_each_member_device(ca, c, i) {
1880                 struct bucket_gens *gens;
1881
1882                 BUG_ON(ca->oldest_gen);
1883
1884                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1885                 if (!ca->oldest_gen) {
1886                         percpu_ref_put(&ca->ref);
1887                         ret = -ENOMEM;
1888                         goto err;
1889                 }
1890
1891                 gens = bucket_gens(ca);
1892
1893                 for (b = gens->first_bucket;
1894                      b < gens->nbuckets; b++)
1895                         ca->oldest_gen[b] = gens->b[b];
1896         }
1897
1898         for (i = 0; i < BTREE_ID_NR; i++)
1899                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1900                         c->gc_gens_btree = i;
1901                         c->gc_gens_pos = POS_MIN;
1902                         ret = bch2_gc_btree_gens(&trans, i);
1903                         if (ret) {
1904                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1905                                 goto err;
1906                         }
1907                 }
1908
1909         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1910                            BTREE_ITER_PREFETCH, k, ret) {
1911                 ret = __bch2_trans_do(&trans, NULL, NULL,
1912                                       BTREE_INSERT_NOFAIL,
1913                                 bch2_alloc_write_oldest_gen(&trans, &iter));
1914                 if (ret) {
1915                         bch_err(c, "error writing oldest_gen: %i", ret);
1916                         break;
1917                 }
1918         }
1919         bch2_trans_iter_exit(&trans, &iter);
1920
1921         c->gc_gens_btree        = 0;
1922         c->gc_gens_pos          = POS_MIN;
1923
1924         c->gc_count++;
1925
1926         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1927 err:
1928         for_each_member_device(ca, c, i) {
1929                 kvfree(ca->oldest_gen);
1930                 ca->oldest_gen = NULL;
1931         }
1932
1933         bch2_trans_exit(&trans);
1934         up_read(&c->gc_lock);
1935         mutex_unlock(&c->gc_gens_lock);
1936         return ret;
1937 }
1938
1939 static int bch2_gc_thread(void *arg)
1940 {
1941         struct bch_fs *c = arg;
1942         struct io_clock *clock = &c->io_clock[WRITE];
1943         unsigned long last = atomic64_read(&clock->now);
1944         unsigned last_kick = atomic_read(&c->kick_gc);
1945         int ret;
1946
1947         set_freezable();
1948
1949         while (1) {
1950                 while (1) {
1951                         set_current_state(TASK_INTERRUPTIBLE);
1952
1953                         if (kthread_should_stop()) {
1954                                 __set_current_state(TASK_RUNNING);
1955                                 return 0;
1956                         }
1957
1958                         if (atomic_read(&c->kick_gc) != last_kick)
1959                                 break;
1960
1961                         if (c->btree_gc_periodic) {
1962                                 unsigned long next = last + c->capacity / 16;
1963
1964                                 if (atomic64_read(&clock->now) >= next)
1965                                         break;
1966
1967                                 bch2_io_clock_schedule_timeout(clock, next);
1968                         } else {
1969                                 schedule();
1970                         }
1971
1972                         try_to_freeze();
1973                 }
1974                 __set_current_state(TASK_RUNNING);
1975
1976                 last = atomic64_read(&clock->now);
1977                 last_kick = atomic_read(&c->kick_gc);
1978
1979                 /*
1980                  * Full gc is currently incompatible with btree key cache:
1981                  */
1982 #if 0
1983                 ret = bch2_gc(c, false, false);
1984 #else
1985                 ret = bch2_gc_gens(c);
1986 #endif
1987                 if (ret < 0)
1988                         bch_err(c, "btree gc failed: %i", ret);
1989
1990                 debug_check_no_locks_held();
1991         }
1992
1993         return 0;
1994 }
1995
1996 void bch2_gc_thread_stop(struct bch_fs *c)
1997 {
1998         struct task_struct *p;
1999
2000         p = c->gc_thread;
2001         c->gc_thread = NULL;
2002
2003         if (p) {
2004                 kthread_stop(p);
2005                 put_task_struct(p);
2006         }
2007 }
2008
2009 int bch2_gc_thread_start(struct bch_fs *c)
2010 {
2011         struct task_struct *p;
2012
2013         if (c->gc_thread)
2014                 return 0;
2015
2016         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2017         if (IS_ERR(p)) {
2018                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
2019                 return PTR_ERR(p);
2020         }
2021
2022         get_task_struct(p);
2023         c->gc_thread = p;
2024         wake_up_process(p);
2025         return 0;
2026 }