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 int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 160 {
 161         struct bkey_i_btree_ptr_v2 *new;
 162         int ret;
 163
 164         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 165         if (!new)
 166                 return -ENOMEM;
 167
 168         btree_ptr_to_v2(b, new);
 169         b->data->min_key        = new_min;
 170         new->v.min_key          = new_min;
 171         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 172
 173         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 174         if (ret) {
 175                 kfree(new);
 176                 return ret;
 177         }
 178
 179         bch2_btree_node_drop_keys_outside_node(b);
 180
 181         return 0;
 182 }
 183
 184 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 185 {
 186         struct bkey_i_btree_ptr_v2 *new;
 187         int ret;
 188
 189         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 190         if (ret)
 191                 return ret;
 192
 193         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 194         if (!new)
 195                 return -ENOMEM;
 196
 197         btree_ptr_to_v2(b, new);
 198         b->data->max_key        = new_max;
 199         new->k.p                = new_max;
 200         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 201
 202         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 203         if (ret) {
 204                 kfree(new);
 205                 return ret;
 206         }
 207
 208         bch2_btree_node_drop_keys_outside_node(b);
 209
 210         mutex_lock(&c->btree_cache.lock);
 211         bch2_btree_node_hash_remove(&c->btree_cache, b);
 212
 213         bkey_copy(&b->key, &new->k_i);
 214         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 215         BUG_ON(ret);
 216         mutex_unlock(&c->btree_cache.lock);
 217         return 0;
 218 }
 219
 220 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 221                                         struct btree *prev, struct btree *cur)
 222 {
 223         struct bpos expected_start = !prev
 224                 ? b->data->min_key
 225                 : bpos_successor(prev->key.k.p);
 226         char buf1[200], buf2[200];
 227         int ret = 0;
 228
 229         if (!prev) {
 230                 struct printbuf out = PBUF(buf1);
 231                 pr_buf(&out, "start of node: ");
 232                 bch2_bpos_to_text(&out, b->data->min_key);
 233         } else {
 234                 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
 235         }
 236
 237         bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key));
 238
 239         if (prev &&
 240             bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 241             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 242                 /* cur overwrites prev: */
 243
 244                 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
 245                                                  cur->data->min_key) >= 0, c,
 246                                 "btree node overwritten by next node at btree %s level %u:\n"
 247                                 "  node %s\n"
 248                                 "  next %s",
 249                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 250                                 buf1, buf2))
 251                         return DROP_PREV_NODE;
 252
 253                 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
 254                                                  bpos_predecessor(cur->data->min_key)), c,
 255                                 "btree node with incorrect max_key at btree %s level %u:\n"
 256                                 "  node %s\n"
 257                                 "  next %s",
 258                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 259                                 buf1, buf2))
 260                         ret = set_node_max(c, prev,
 261                                            bpos_predecessor(cur->data->min_key));
 262         } else {
 263                 /* prev overwrites cur: */
 264
 265                 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
 266                                                  cur->data->max_key) >= 0, c,
 267                                 "btree node overwritten by prev node at btree %s level %u:\n"
 268                                 "  prev %s\n"
 269                                 "  node %s",
 270                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 271                                 buf1, buf2))
 272                         return DROP_THIS_NODE;
 273
 274                 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 275                                 "btree node with incorrect min_key at btree %s level %u:\n"
 276                                 "  prev %s\n"
 277                                 "  node %s",
 278                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 279                                 buf1, buf2))
 280                     ret = set_node_min(c, cur, expected_start);
 281         }
 282 fsck_err:
 283         return ret;
 284 }
 285
 286 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 287                                  struct btree *child)
 288 {
 289         char buf1[200], buf2[200];
 290         int ret = 0;
 291
 292         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 293                         "btree node with incorrect max_key at btree %s level %u:\n"
 294                         "  %s\n"
 295                         "  expected %s",
 296                         bch2_btree_ids[b->c.btree_id], b->c.level,
 297                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
 298                         (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
 299                 ret = set_node_max(c, child, b->key.k.p);
 300                 if (ret)
 301                         return ret;
 302         }
 303 fsck_err:
 304         return ret;
 305 }
 306
 307 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 308 {
 309         struct btree_and_journal_iter iter;
 310         struct bkey_s_c k;
 311         struct bkey_buf prev_k, cur_k;
 312         struct btree *prev = NULL, *cur = NULL;
 313         bool have_child, dropped_children = false;
 314         char buf[200];
 315         int ret = 0;
 316
 317         if (!b->c.level)
 318                 return 0;
 319 again:
 320         prev = NULL;
 321         have_child = dropped_children = false;
 322         bch2_bkey_buf_init(&prev_k);
 323         bch2_bkey_buf_init(&cur_k);
 324         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 325
 326         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 327                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 328                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 329
 330                 bch2_btree_and_journal_iter_advance(&iter);
 331                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 332
 333                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 334                                         b->c.btree_id, b->c.level - 1,
 335                                         false);
 336                 ret = PTR_ERR_OR_ZERO(cur);
 337
 338                 if (mustfix_fsck_err_on(ret == -EIO, c,
 339                                 "Unreadable btree node at btree %s level %u:\n"
 340                                 "  %s",
 341                                 bch2_btree_ids[b->c.btree_id],
 342                                 b->c.level - 1,
 343                                 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
 344                         bch2_btree_node_evict(c, cur_k.k);
 345                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 346                                                       b->c.level, cur_k.k->k.p);
 347                         if (ret)
 348                                 break;
 349                         continue;
 350                 }
 351
 352                 if (ret) {
 353                         bch_err(c, "%s: error %i getting btree node",
 354                                 __func__, ret);
 355                         break;
 356                 }
 357
 358                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 359
 360                 if (ret == DROP_THIS_NODE) {
 361                         six_unlock_read(&cur->c.lock);
 362                         bch2_btree_node_evict(c, cur_k.k);
 363                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 364                                                       b->c.level, cur_k.k->k.p);
 365                         if (ret)
 366                                 break;
 367                         continue;
 368                 }
 369
 370                 if (prev)
 371                         six_unlock_read(&prev->c.lock);
 372                 prev = NULL;
 373
 374                 if (ret == DROP_PREV_NODE) {
 375                         bch2_btree_node_evict(c, prev_k.k);
 376                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 377                                                       b->c.level, prev_k.k->k.p);
 378                         if (ret)
 379                                 break;
 380
 381                         bch2_btree_and_journal_iter_exit(&iter);
 382                         bch2_bkey_buf_exit(&prev_k, c);
 383                         bch2_bkey_buf_exit(&cur_k, c);
 384                         goto again;
 385                 } else if (ret)
 386                         break;
 387
 388                 prev = cur;
 389                 cur = NULL;
 390                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 391         }
 392
 393         if (!ret && !IS_ERR_OR_NULL(prev)) {
 394                 BUG_ON(cur);
 395                 ret = btree_repair_node_end(c, b, prev);
 396         }
 397
 398         if (!IS_ERR_OR_NULL(prev))
 399                 six_unlock_read(&prev->c.lock);
 400         prev = NULL;
 401         if (!IS_ERR_OR_NULL(cur))
 402                 six_unlock_read(&cur->c.lock);
 403         cur = NULL;
 404
 405         if (ret)
 406                 goto err;
 407
 408         bch2_btree_and_journal_iter_exit(&iter);
 409         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 410
 411         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 412                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 413                 bch2_btree_and_journal_iter_advance(&iter);
 414
 415                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 416                                         b->c.btree_id, b->c.level - 1,
 417                                         false);
 418                 ret = PTR_ERR_OR_ZERO(cur);
 419
 420                 if (ret) {
 421                         bch_err(c, "%s: error %i getting btree node",
 422                                 __func__, ret);
 423                         goto err;
 424                 }
 425
 426                 ret = bch2_btree_repair_topology_recurse(c, cur);
 427                 six_unlock_read(&cur->c.lock);
 428                 cur = NULL;
 429
 430                 if (ret == DROP_THIS_NODE) {
 431                         bch2_btree_node_evict(c, cur_k.k);
 432                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 433                                                       b->c.level, cur_k.k->k.p);
 434                         dropped_children = true;
 435                 }
 436
 437                 if (ret)
 438                         goto err;
 439
 440                 have_child = true;
 441         }
 442
 443         if (mustfix_fsck_err_on(!have_child, c,
 444                         "empty interior btree node at btree %s level %u\n"
 445                         "  %s",
 446                         bch2_btree_ids[b->c.btree_id],
 447                         b->c.level,
 448                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
 449                 ret = DROP_THIS_NODE;
 450 err:
 451 fsck_err:
 452         if (!IS_ERR_OR_NULL(prev))
 453                 six_unlock_read(&prev->c.lock);
 454         if (!IS_ERR_OR_NULL(cur))
 455                 six_unlock_read(&cur->c.lock);
 456
 457         bch2_btree_and_journal_iter_exit(&iter);
 458         bch2_bkey_buf_exit(&prev_k, c);
 459         bch2_bkey_buf_exit(&cur_k, c);
 460
 461         if (!ret && dropped_children)
 462                 goto again;
 463
 464         return ret;
 465 }
 466
 467 static int bch2_repair_topology(struct bch_fs *c)
 468 {
 469         struct btree *b;
 470         unsigned i;
 471         int ret = 0;
 472
 473         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 474                 b = c->btree_roots[i].b;
 475                 if (btree_node_fake(b))
 476                         continue;
 477
 478                 six_lock_read(&b->c.lock, NULL, NULL);
 479                 ret = bch2_btree_repair_topology_recurse(c, b);
 480                 six_unlock_read(&b->c.lock);
 481
 482                 if (ret == DROP_THIS_NODE) {
 483                         bch_err(c, "empty btree root - repair unimplemented");
 484                         ret = FSCK_ERR_EXIT;
 485                 }
 486         }
 487
 488         return ret;
 489 }
 490
 491 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 492                                unsigned level, bool is_root,
 493                                struct bkey_s_c *k)
 494 {
 495         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 496         const union bch_extent_entry *entry;
 497         struct extent_ptr_decoded p = { 0 };
 498         bool do_update = false;
 499         char buf[200];
 500         int ret = 0;
 501
 502         /*
 503          * XXX
 504          * use check_bucket_ref here
 505          */
 506         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 507                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 508                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 509                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 510
 511                 if (fsck_err_on(!g->gen_valid, c,
 512                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 513                                 "while marking %s",
 514                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 515                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 516                                 p.ptr.gen,
 517                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 518                         if (!p.ptr.cached) {
 519                                 g->_mark.gen            = p.ptr.gen;
 520                                 g->gen_valid            = true;
 521                         } else {
 522                                 do_update = true;
 523                         }
 524                 }
 525
 526                 if (fsck_err_on(data_type == BCH_DATA_btree &&
 527                                 g->mark.gen != p.ptr.gen, c,
 528                                 "bucket %u:%zu data type %s has metadata but wrong gen: %u != %u\n"
 529                                 "while marking %s",
 530                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 531                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 532                                 p.ptr.gen, g->mark.gen,
 533                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 534                         g->_mark.data_type      = data_type;
 535                         g->gen_valid            = true;
 536                 }
 537
 538                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 539                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 540                                 "while marking %s",
 541                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 542                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 543                                 p.ptr.gen, g->mark.gen,
 544                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 545                         if (!p.ptr.cached) {
 546                                 g->_mark.gen            = p.ptr.gen;
 547                                 g->gen_valid            = true;
 548                                 g->_mark.data_type      = 0;
 549                                 g->_mark.dirty_sectors  = 0;
 550                                 g->_mark.cached_sectors = 0;
 551                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 552                         } else {
 553                                 do_update = true;
 554                         }
 555                 }
 556
 557                 if (fsck_err_on(gen_cmp(g->mark.gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
 558                                 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 559                                 "while marking %s",
 560                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->mark.gen,
 561                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 562                                 p.ptr.gen,
 563                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 564                         do_update = true;
 565
 566                 if (fsck_err_on(!p.ptr.cached &&
 567                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 568                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 569                                 "while marking %s",
 570                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 571                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 572                                 p.ptr.gen, g->mark.gen,
 573                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 574                         do_update = true;
 575
 576                 if (p.ptr.gen != g->mark.gen)
 577                         continue;
 578
 579                 if (fsck_err_on(g->mark.data_type &&
 580                                 g->mark.data_type != data_type, c,
 581                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 582                                 "while marking %s",
 583                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 584                                 bch2_data_types[g->mark.data_type],
 585                                 bch2_data_types[data_type],
 586                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 587                         if (data_type == BCH_DATA_btree) {
 588                                 g->_mark.data_type      = data_type;
 589                                 g->gen_valid            = true;
 590                         } else {
 591                                 do_update = true;
 592                         }
 593                 }
 594
 595                 if (p.has_ec) {
 596                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 597
 598                         if (fsck_err_on(!m || !m->alive, c,
 599                                         "pointer to nonexistent stripe %llu\n"
 600                                         "while marking %s",
 601                                         (u64) p.ec.idx,
 602                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 603                                 do_update = true;
 604
 605                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 606                                         "pointer does not match stripe %llu\n"
 607                                         "while marking %s",
 608                                         (u64) p.ec.idx,
 609                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 610                                 do_update = true;
 611                 }
 612         }
 613
 614         if (do_update) {
 615                 struct bkey_ptrs ptrs;
 616                 union bch_extent_entry *entry;
 617                 struct bch_extent_ptr *ptr;
 618                 struct bkey_i *new;
 619
 620                 if (is_root) {
 621                         bch_err(c, "cannot update btree roots yet");
 622                         return -EINVAL;
 623                 }
 624
 625                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 626                 if (!new) {
 627                         bch_err(c, "%s: error allocating new key", __func__);
 628                         return -ENOMEM;
 629                 }
 630
 631                 bkey_reassemble(new, *k);
 632
 633                 if (level) {
 634                         /*
 635                          * We don't want to drop btree node pointers - if the
 636                          * btree node isn't there anymore, the read path will
 637                          * sort it out:
 638                          */
 639                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 640                         bkey_for_each_ptr(ptrs, ptr) {
 641                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 642                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 643
 644                                 ptr->gen = g->mark.gen;
 645                         }
 646                 } else {
 647                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 648                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 649                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 650                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 651
 652                                 (ptr->cached &&
 653                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 654                                 (!ptr->cached &&
 655                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 656                                 gen_cmp(g->mark.gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 657                                 (g->mark.data_type &&
 658                                  g->mark.data_type != data_type);
 659                         }));
 660 again:
 661                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 662                         bkey_extent_entry_for_each(ptrs, entry) {
 663                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 664                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 665                                                                         entry->stripe_ptr.idx);
 666                                         union bch_extent_entry *next_ptr;
 667
 668                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 669                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 670                                                         goto found;
 671                                         next_ptr = NULL;
 672 found:
 673                                         if (!next_ptr) {
 674                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 675                                                 continue;
 676                                         }
 677
 678                                         if (!m || !m->alive ||
 679                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 680                                                                        &next_ptr->ptr,
 681                                                                        m->sectors)) {
 682                                                 bch2_bkey_extent_entry_drop(new, entry);
 683                                                 goto again;
 684                                         }
 685                                 }
 686                         }
 687                 }
 688
 689                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 690
 691                 if (ret)
 692                         kfree(new);
 693                 else {
 694                         bch2_bkey_val_to_text(&PBUF(buf), c, *k);
 695                         bch_info(c, "updated %s", buf);
 696                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(new));
 697                         bch_info(c, "new key %s", buf);
 698                         *k = bkey_i_to_s_c(new);
 699                 }
 700         }
 701 fsck_err:
 702         return ret;
 703 }
 704
 705 /* marking of btree keys/nodes: */
 706
 707 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 708                             unsigned level, bool is_root,
 709                             struct bkey_s_c *k,
 710                             u8 *max_stale, bool initial)
 711 {
 712         struct bch_fs *c = trans->c;
 713         struct bkey_ptrs_c ptrs;
 714         const struct bch_extent_ptr *ptr;
 715         struct bkey deleted = KEY(0, 0, 0);
 716         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 717         unsigned flags =
 718                 BTREE_TRIGGER_GC|
 719                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 720         int ret = 0;
 721
 722         deleted.p = k->k->p;
 723
 724         if (initial) {
 725                 BUG_ON(bch2_journal_seq_verify &&
 726                        k->k->version.lo > journal_cur_seq(&c->journal));
 727
 728                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 729                 if (ret)
 730                         goto err;
 731
 732                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 733                                 "key version number higher than recorded: %llu > %llu",
 734                                 k->k->version.lo,
 735                                 atomic64_read(&c->key_version)))
 736                         atomic64_set(&c->key_version, k->k->version.lo);
 737         }
 738
 739         ptrs = bch2_bkey_ptrs_c(*k);
 740         bkey_for_each_ptr(ptrs, ptr) {
 741                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 742                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 743
 744                 if (gen_after(g->oldest_gen, ptr->gen))
 745                         g->oldest_gen = ptr->gen;
 746
 747                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 748         }
 749
 750         ret = bch2_mark_key(trans, old, *k, flags);
 751 fsck_err:
 752 err:
 753         if (ret)
 754                 bch_err(c, "%s: ret %i", __func__, ret);
 755         return ret;
 756 }
 757
 758 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, u8 *max_stale,
 759                               bool initial)
 760 {
 761         struct bch_fs *c = trans->c;
 762         struct btree_node_iter iter;
 763         struct bkey unpacked;
 764         struct bkey_s_c k;
 765         struct bkey_buf prev, cur;
 766         int ret = 0;
 767
 768         *max_stale = 0;
 769
 770         if (!btree_node_type_needs_gc(btree_node_type(b)))
 771                 return 0;
 772
 773         bch2_btree_node_iter_init_from_start(&iter, b);
 774         bch2_bkey_buf_init(&prev);
 775         bch2_bkey_buf_init(&cur);
 776         bkey_init(&prev.k->k);
 777
 778         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 779                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 780                                        &k, max_stale, initial);
 781                 if (ret)
 782                         break;
 783
 784                 bch2_btree_node_iter_advance(&iter, b);
 785
 786                 if (b->c.level) {
 787                         bch2_bkey_buf_reassemble(&cur, c, k);
 788
 789                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 790                                         bch2_btree_node_iter_end(&iter));
 791                         if (ret)
 792                                 break;
 793                 }
 794         }
 795
 796         bch2_bkey_buf_exit(&cur, c);
 797         bch2_bkey_buf_exit(&prev, c);
 798         return ret;
 799 }
 800
 801 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 802                          bool initial, bool metadata_only)
 803 {
 804         struct bch_fs *c = trans->c;
 805         struct btree_iter iter;
 806         struct btree *b;
 807         unsigned depth = metadata_only                  ? 1
 808                 : bch2_expensive_debug_checks           ? 0
 809                 : !btree_node_type_needs_gc(btree_id)   ? 1
 810                 : 0;
 811         u8 max_stale = 0;
 812         int ret = 0;
 813
 814         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 815
 816         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 817                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 818                 bch2_verify_btree_nr_keys(b);
 819
 820                 gc_pos_set(c, gc_pos_btree_node(b));
 821
 822                 ret = btree_gc_mark_node(trans, b, &max_stale, initial);
 823                 if (ret)
 824                         break;
 825
 826                 if (!initial) {
 827                         if (max_stale > 64)
 828                                 bch2_btree_node_rewrite(trans, &iter, b,
 829                                                 BTREE_INSERT_NOWAIT|
 830                                                 BTREE_INSERT_GC_LOCK_HELD);
 831                         else if (!bch2_btree_gc_rewrite_disabled &&
 832                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 833                                 bch2_btree_node_rewrite(trans, &iter,
 834                                                 b, BTREE_INSERT_NOWAIT|
 835                                                 BTREE_INSERT_GC_LOCK_HELD);
 836                 }
 837         }
 838         bch2_trans_iter_exit(trans, &iter);
 839
 840         if (ret)
 841                 return ret;
 842
 843         mutex_lock(&c->btree_root_lock);
 844         b = c->btree_roots[btree_id].b;
 845         if (!btree_node_fake(b)) {
 846                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 847
 848                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
 849                                        &k, &max_stale, initial);
 850         }
 851         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 852         mutex_unlock(&c->btree_root_lock);
 853
 854         return ret;
 855 }
 856
 857 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 858                                       unsigned target_depth)
 859 {
 860         struct bch_fs *c = trans->c;
 861         struct btree_and_journal_iter iter;
 862         struct bkey_s_c k;
 863         struct bkey_buf cur, prev;
 864         u8 max_stale = 0;
 865         char buf[200];
 866         int ret = 0;
 867
 868         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 869         bch2_bkey_buf_init(&prev);
 870         bch2_bkey_buf_init(&cur);
 871         bkey_init(&prev.k->k);
 872
 873         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 874                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 875                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 876
 877                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 878                                        &k, &max_stale, true);
 879                 if (ret) {
 880                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 881                         goto fsck_err;
 882                 }
 883
 884                 if (b->c.level) {
 885                         bch2_bkey_buf_reassemble(&cur, c, k);
 886                         k = bkey_i_to_s_c(cur.k);
 887
 888                         bch2_btree_and_journal_iter_advance(&iter);
 889
 890                         ret = bch2_gc_check_topology(c, b,
 891                                         &prev, cur,
 892                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 893                         if (ret)
 894                                 goto fsck_err;
 895                 } else {
 896                         bch2_btree_and_journal_iter_advance(&iter);
 897                 }
 898         }
 899
 900         if (b->c.level > target_depth) {
 901                 bch2_btree_and_journal_iter_exit(&iter);
 902                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 903
 904                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 905                         struct btree *child;
 906
 907                         bch2_bkey_buf_reassemble(&cur, c, k);
 908                         bch2_btree_and_journal_iter_advance(&iter);
 909
 910                         child = bch2_btree_node_get_noiter(c, cur.k,
 911                                                 b->c.btree_id, b->c.level - 1,
 912                                                 false);
 913                         ret = PTR_ERR_OR_ZERO(child);
 914
 915                         if (ret == -EIO) {
 916                                 bch2_topology_error(c);
 917
 918                                 if (__fsck_err(c,
 919                                           FSCK_CAN_FIX|
 920                                           FSCK_CAN_IGNORE|
 921                                           FSCK_NO_RATELIMIT,
 922                                           "Unreadable btree node at btree %s level %u:\n"
 923                                           "  %s",
 924                                           bch2_btree_ids[b->c.btree_id],
 925                                           b->c.level - 1,
 926                                           (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
 927                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 928                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 929                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 930                                         goto fsck_err;
 931                                 } else {
 932                                         /* Continue marking when opted to not
 933                                          * fix the error: */
 934                                         ret = 0;
 935                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 936                                         continue;
 937                                 }
 938                         } else if (ret) {
 939                                 bch_err(c, "%s: error %i getting btree node",
 940                                         __func__, ret);
 941                                 break;
 942                         }
 943
 944                         ret = bch2_gc_btree_init_recurse(trans, child,
 945                                                          target_depth);
 946                         six_unlock_read(&child->c.lock);
 947
 948                         if (ret)
 949                                 break;
 950                 }
 951         }
 952 fsck_err:
 953         bch2_bkey_buf_exit(&cur, c);
 954         bch2_bkey_buf_exit(&prev, c);
 955         bch2_btree_and_journal_iter_exit(&iter);
 956         return ret;
 957 }
 958
 959 static int bch2_gc_btree_init(struct btree_trans *trans,
 960                               enum btree_id btree_id,
 961                               bool metadata_only)
 962 {
 963         struct bch_fs *c = trans->c;
 964         struct btree *b;
 965         unsigned target_depth = metadata_only           ? 1
 966                 : bch2_expensive_debug_checks           ? 0
 967                 : !btree_node_type_needs_gc(btree_id)   ? 1
 968                 : 0;
 969         u8 max_stale = 0;
 970         char buf[100];
 971         int ret = 0;
 972
 973         b = c->btree_roots[btree_id].b;
 974
 975         if (btree_node_fake(b))
 976                 return 0;
 977
 978         six_lock_read(&b->c.lock, NULL, NULL);
 979         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 980                         "btree root with incorrect min_key: %s",
 981                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 982                 bch_err(c, "repair unimplemented");
 983                 ret = FSCK_ERR_EXIT;
 984                 goto fsck_err;
 985         }
 986
 987         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
 988                         "btree root with incorrect max_key: %s",
 989                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 990                 bch_err(c, "repair unimplemented");
 991                 ret = FSCK_ERR_EXIT;
 992                 goto fsck_err;
 993         }
 994
 995         if (b->c.level >= target_depth)
 996                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
 997
 998         if (!ret) {
 999                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1000
1001                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1002                                        &k, &max_stale, true);
1003         }
1004 fsck_err:
1005         six_unlock_read(&b->c.lock);
1006
1007         if (ret < 0)
1008                 bch_err(c, "%s: ret %i", __func__, ret);
1009         return ret;
1010 }
1011
1012 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1013 {
1014         return  (int) btree_id_to_gc_phase(l) -
1015                 (int) btree_id_to_gc_phase(r);
1016 }
1017
1018 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1019 {
1020         struct btree_trans trans;
1021         enum btree_id ids[BTREE_ID_NR];
1022         unsigned i;
1023         int ret = 0;
1024
1025         bch2_trans_init(&trans, c, 0, 0);
1026
1027         for (i = 0; i < BTREE_ID_NR; i++)
1028                 ids[i] = i;
1029         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1030
1031         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1032                 ret = initial
1033                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1034                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1035
1036         if (ret < 0)
1037                 bch_err(c, "%s: ret %i", __func__, ret);
1038
1039         bch2_trans_exit(&trans);
1040         return ret;
1041 }
1042
1043 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1044                                   u64 start, u64 end,
1045                                   enum bch_data_type type,
1046                                   unsigned flags)
1047 {
1048         u64 b = sector_to_bucket(ca, start);
1049
1050         do {
1051                 unsigned sectors =
1052                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1053
1054                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1055                                           gc_phase(GC_PHASE_SB), flags);
1056                 b++;
1057                 start += sectors;
1058         } while (start < end);
1059 }
1060
1061 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1062                                      unsigned flags)
1063 {
1064         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1065         unsigned i;
1066         u64 b;
1067
1068         for (i = 0; i < layout->nr_superblocks; i++) {
1069                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1070
1071                 if (offset == BCH_SB_SECTOR)
1072                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1073                                               BCH_DATA_sb, flags);
1074
1075                 mark_metadata_sectors(c, ca, offset,
1076                                       offset + (1 << layout->sb_max_size_bits),
1077                                       BCH_DATA_sb, flags);
1078         }
1079
1080         for (i = 0; i < ca->journal.nr; i++) {
1081                 b = ca->journal.buckets[i];
1082                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1083                                           ca->mi.bucket_size,
1084                                           gc_phase(GC_PHASE_SB), flags);
1085         }
1086 }
1087
1088 static void bch2_mark_superblocks(struct bch_fs *c)
1089 {
1090         struct bch_dev *ca;
1091         unsigned i;
1092
1093         mutex_lock(&c->sb_lock);
1094         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1095
1096         for_each_online_member(ca, c, i)
1097                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1098         mutex_unlock(&c->sb_lock);
1099 }
1100
1101 #if 0
1102 /* Also see bch2_pending_btree_node_free_insert_done() */
1103 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1104 {
1105         struct btree_update *as;
1106         struct pending_btree_node_free *d;
1107
1108         mutex_lock(&c->btree_interior_update_lock);
1109         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1110
1111         for_each_pending_btree_node_free(c, as, d)
1112                 if (d->index_update_done)
1113                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1114
1115         mutex_unlock(&c->btree_interior_update_lock);
1116 }
1117 #endif
1118
1119 static void bch2_gc_free(struct bch_fs *c)
1120 {
1121         struct bch_dev *ca;
1122         unsigned i;
1123
1124         genradix_free(&c->reflink_gc_table);
1125         genradix_free(&c->gc_stripes);
1126
1127         for_each_member_device(ca, c, i) {
1128                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1129                         sizeof(struct bucket_array) +
1130                         ca->mi.nbuckets * sizeof(struct bucket));
1131                 ca->buckets[1] = NULL;
1132
1133                 free_percpu(ca->usage_gc);
1134                 ca->usage_gc = NULL;
1135         }
1136
1137         free_percpu(c->usage_gc);
1138         c->usage_gc = NULL;
1139 }
1140
1141 static int bch2_gc_done(struct bch_fs *c,
1142                         bool initial, bool metadata_only)
1143 {
1144         struct bch_dev *ca = NULL;
1145         bool verify = !metadata_only && (!initial ||
1146                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1147         unsigned i, dev;
1148         int ret = 0;
1149
1150         percpu_down_write(&c->mark_lock);
1151
1152 #define copy_field(_f, _msg, ...)                                       \
1153         if (dst->_f != src->_f) {                                       \
1154                 if (verify)                                             \
1155                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1156                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1157                 dst->_f = src->_f;                                      \
1158         }
1159 #define copy_stripe_field(_f, _msg, ...)                                \
1160         if (dst->_f != src->_f) {                                       \
1161                 if (verify)                                             \
1162                         fsck_err(c, "stripe %zu has wrong "_msg         \
1163                                 ": got %u, should be %u",               \
1164                                 iter.pos, ##__VA_ARGS__,                \
1165                                 dst->_f, src->_f);                      \
1166                 dst->_f = src->_f;                                      \
1167         }
1168 #define copy_dev_field(_f, _msg, ...)                                   \
1169         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1170 #define copy_fs_field(_f, _msg, ...)                                    \
1171         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1172
1173         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1174                 bch2_fs_usage_acc_to_base(c, i);
1175
1176         for_each_member_device(ca, c, dev) {
1177                 struct bch_dev_usage *dst = ca->usage_base;
1178                 struct bch_dev_usage *src = (void *)
1179                         bch2_acc_percpu_u64s((void *) ca->usage_gc,
1180                                              dev_usage_u64s());
1181
1182                 copy_dev_field(buckets_ec,              "buckets_ec");
1183                 copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1184
1185                 for (i = 0; i < BCH_DATA_NR; i++) {
1186                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1187                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1188                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1189                 }
1190         };
1191
1192         {
1193                 unsigned nr = fs_usage_u64s(c);
1194                 struct bch_fs_usage *dst = c->usage_base;
1195                 struct bch_fs_usage *src = (void *)
1196                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1197
1198                 copy_fs_field(hidden,           "hidden");
1199                 copy_fs_field(btree,            "btree");
1200
1201                 if (!metadata_only) {
1202                         copy_fs_field(data,     "data");
1203                         copy_fs_field(cached,   "cached");
1204                         copy_fs_field(reserved, "reserved");
1205                         copy_fs_field(nr_inodes,"nr_inodes");
1206
1207                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1208                                 copy_fs_field(persistent_reserved[i],
1209                                               "persistent_reserved[%i]", i);
1210                 }
1211
1212                 for (i = 0; i < c->replicas.nr; i++) {
1213                         struct bch_replicas_entry *e =
1214                                 cpu_replicas_entry(&c->replicas, i);
1215                         char buf[80];
1216
1217                         if (metadata_only &&
1218                             (e->data_type == BCH_DATA_user ||
1219                              e->data_type == BCH_DATA_cached))
1220                                 continue;
1221
1222                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1223
1224                         copy_fs_field(replicas[i], "%s", buf);
1225                 }
1226         }
1227
1228 #undef copy_fs_field
1229 #undef copy_dev_field
1230 #undef copy_stripe_field
1231 #undef copy_field
1232 fsck_err:
1233         if (ca)
1234                 percpu_ref_put(&ca->ref);
1235         if (ret)
1236                 bch_err(c, "%s: ret %i", __func__, ret);
1237
1238         percpu_up_write(&c->mark_lock);
1239         return ret;
1240 }
1241
1242 static int bch2_gc_start(struct bch_fs *c,
1243                          bool metadata_only)
1244 {
1245         struct bch_dev *ca = NULL;
1246         unsigned i;
1247
1248         BUG_ON(c->usage_gc);
1249
1250         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1251                                          sizeof(u64), GFP_KERNEL);
1252         if (!c->usage_gc) {
1253                 bch_err(c, "error allocating c->usage_gc");
1254                 return -ENOMEM;
1255         }
1256
1257         for_each_member_device(ca, c, i) {
1258                 BUG_ON(ca->buckets[1]);
1259                 BUG_ON(ca->usage_gc);
1260
1261                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1262                 if (!ca->usage_gc) {
1263                         bch_err(c, "error allocating ca->usage_gc");
1264                         percpu_ref_put(&ca->ref);
1265                         return -ENOMEM;
1266                 }
1267         }
1268
1269         return 0;
1270 }
1271
1272 static int bch2_alloc_write_key(struct btree_trans *trans,
1273                                 struct btree_iter *iter,
1274                                 bool initial, bool metadata_only)
1275 {
1276         struct bch_fs *c = trans->c;
1277         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1278         struct bucket *g;
1279         struct bkey_s_c k;
1280         struct bkey_alloc_unpacked old_u, new_u, gc_u;
1281         struct bkey_alloc_buf *a;
1282         int ret;
1283
1284         /*
1285          * For this to be correct at runtime, we'll need to figure out a way for
1286          * it to actually lock the key in the btree key cache:
1287          */
1288
1289         if (!initial) {
1290                 ret = bch2_btree_key_cache_flush(trans,
1291                                 BTREE_ID_alloc, iter->pos);
1292                 if (ret)
1293                         return ret;
1294         }
1295
1296         k = bch2_btree_iter_peek_slot(iter);
1297         ret = bkey_err(k);
1298         if (ret)
1299                 return ret;
1300
1301         old_u = new_u   = bch2_alloc_unpack(k);
1302
1303         percpu_down_read(&c->mark_lock);
1304         g       = gc_bucket(ca, iter->pos.offset);
1305         gc_u = (struct bkey_alloc_unpacked) {
1306                 .dev            = iter->pos.inode,
1307                 .bucket         = iter->pos.offset,
1308                 .gen            = g->mark.gen,
1309                 .oldest_gen     = g->oldest_gen,
1310                 .data_type      = g->mark.data_type,
1311                 .dirty_sectors  = g->mark.dirty_sectors,
1312                 .cached_sectors = g->mark.cached_sectors,
1313                 .read_time      = g->io_time[READ],
1314                 .write_time     = g->io_time[WRITE],
1315                 .stripe         = g->stripe,
1316                 .stripe_redundancy = g->stripe_redundancy,
1317         };
1318         percpu_up_read(&c->mark_lock);
1319
1320         if (metadata_only &&
1321             gc_u.data_type != BCH_DATA_sb &&
1322             gc_u.data_type != BCH_DATA_journal &&
1323             gc_u.data_type != BCH_DATA_btree)
1324                 return 0;
1325
1326         if (!bkey_alloc_unpacked_cmp(old_u, gc_u) ||
1327             gen_after(old_u.gen, gc_u.gen))
1328                 return 0;
1329
1330 #define copy_bucket_field(_f)                                           \
1331         if (fsck_err_on(new_u._f != gc_u._f, c,                         \
1332                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1333                         ": got %u, should be %u",                       \
1334                         iter->pos.inode, iter->pos.offset,              \
1335                         new_u.gen,                                      \
1336                         bch2_data_types[new_u.data_type],               \
1337                         new_u._f, gc_u._f))                             \
1338                 new_u._f = gc_u._f;                                     \
1339
1340         copy_bucket_field(gen);
1341         copy_bucket_field(data_type);
1342         copy_bucket_field(stripe);
1343         copy_bucket_field(dirty_sectors);
1344         copy_bucket_field(cached_sectors);
1345         copy_bucket_field(stripe_redundancy);
1346         copy_bucket_field(stripe);
1347 #undef copy_bucket_field
1348
1349         new_u.oldest_gen = gc_u.oldest_gen;
1350
1351         if (!bkey_alloc_unpacked_cmp(old_u, new_u))
1352                 return 0;
1353
1354         a = bch2_alloc_pack(trans, new_u);
1355         if (IS_ERR(a))
1356                 return PTR_ERR(a);
1357
1358         ret = initial
1359                 ? bch2_journal_key_insert(c, BTREE_ID_alloc, 0, &a->k)
1360                 : bch2_trans_update(trans, iter, &a->k, BTREE_TRIGGER_NORUN);
1361 fsck_err:
1362         return ret;
1363 }
1364
1365 static int bch2_gc_alloc_done(struct bch_fs *c, bool initial, bool metadata_only)
1366 {
1367         struct btree_trans trans;
1368         struct btree_iter iter;
1369         struct bkey_s_c k;
1370         struct bch_dev *ca;
1371         unsigned i;
1372         int ret = 0;
1373
1374         bch2_trans_init(&trans, c, 0, 0);
1375
1376         for_each_member_device(ca, c, i) {
1377                 for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1378                                    POS(ca->dev_idx, ca->mi.first_bucket),
1379                                    BTREE_ITER_SLOTS|
1380                                    BTREE_ITER_PREFETCH, k, ret) {
1381                         if (bkey_cmp(iter.pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1382                                 break;
1383
1384                         ret = __bch2_trans_do(&trans, NULL, NULL,
1385                                               BTREE_INSERT_LAZY_RW,
1386                                         bch2_alloc_write_key(&trans, &iter,
1387                                                              initial, metadata_only));
1388                         if (ret)
1389                                 break;
1390                 }
1391                 bch2_trans_iter_exit(&trans, &iter);
1392
1393                 if (ret) {
1394                         bch_err(c, "error writing alloc info: %i", ret);
1395                         percpu_ref_put(&ca->ref);
1396                         break;
1397                 }
1398         }
1399
1400         bch2_trans_exit(&trans);
1401         return ret;
1402 }
1403
1404 static int bch2_gc_alloc_start(struct bch_fs *c, bool initial, bool metadata_only)
1405 {
1406         struct bch_dev *ca;
1407         unsigned i;
1408
1409         for_each_member_device(ca, c, i) {
1410                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1411                                 ca->mi.nbuckets * sizeof(struct bucket),
1412                                 GFP_KERNEL|__GFP_ZERO);
1413                 if (!buckets) {
1414                         percpu_ref_put(&ca->ref);
1415                         percpu_up_write(&c->mark_lock);
1416                         bch_err(c, "error allocating ca->buckets[gc]");
1417                         return -ENOMEM;
1418                 }
1419
1420                 buckets->first_bucket   = ca->mi.first_bucket;
1421                 buckets->nbuckets       = ca->mi.nbuckets;
1422                 rcu_assign_pointer(ca->buckets[1], buckets);
1423         };
1424
1425         return bch2_alloc_read(c, true, metadata_only);
1426 }
1427
1428 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1429                                 bool metadata_only)
1430 {
1431         struct btree_trans trans;
1432         struct btree_iter iter;
1433         struct bkey_s_c k;
1434         struct reflink_gc *r;
1435         size_t idx = 0;
1436         char buf[200];
1437         int ret = 0;
1438
1439         if (metadata_only)
1440                 return 0;
1441
1442         bch2_trans_init(&trans, c, 0, 0);
1443
1444         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1445                            BTREE_ITER_PREFETCH, k, ret) {
1446                 const __le64 *refcount = bkey_refcount_c(k);
1447
1448                 if (!refcount)
1449                         continue;
1450
1451                 r = genradix_ptr(&c->reflink_gc_table, idx++);
1452                 if (!r ||
1453                     r->offset != k.k->p.offset ||
1454                     r->size != k.k->size) {
1455                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1456                         ret = -EINVAL;
1457                         break;
1458                 }
1459
1460                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1461                                 "reflink key has wrong refcount:\n"
1462                                 "  %s\n"
1463                                 "  should be %u",
1464                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1465                                 r->refcount)) {
1466                         struct bkey_i *new;
1467
1468                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1469                         if (!new) {
1470                                 ret = -ENOMEM;
1471                                 break;
1472                         }
1473
1474                         bkey_reassemble(new, k);
1475
1476                         if (!r->refcount)
1477                                 new->k.type = KEY_TYPE_deleted;
1478                         else
1479                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1480
1481                         ret = initial
1482                                ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, new)
1483                                : __bch2_trans_do(&trans, NULL, NULL, 0,
1484                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1485                         kfree(new);
1486
1487                         if (ret)
1488                                 break;
1489                 }
1490         }
1491 fsck_err:
1492         bch2_trans_iter_exit(&trans, &iter);
1493         c->reflink_gc_nr = 0;
1494         bch2_trans_exit(&trans);
1495         return ret;
1496 }
1497
1498 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1499                                  bool metadata_only)
1500 {
1501         struct btree_trans trans;
1502         struct btree_iter iter;
1503         struct bkey_s_c k;
1504         struct reflink_gc *r;
1505         int ret = 0;
1506
1507         if (metadata_only)
1508                 return 0;
1509
1510         bch2_trans_init(&trans, c, 0, 0);
1511         c->reflink_gc_nr = 0;
1512
1513         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1514                            BTREE_ITER_PREFETCH, k, ret) {
1515                 const __le64 *refcount = bkey_refcount_c(k);
1516
1517                 if (!refcount)
1518                         continue;
1519
1520                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1521                                        GFP_KERNEL);
1522                 if (!r) {
1523                         ret = -ENOMEM;
1524                         break;
1525                 }
1526
1527                 r->offset       = k.k->p.offset;
1528                 r->size         = k.k->size;
1529                 r->refcount     = 0;
1530         }
1531         bch2_trans_iter_exit(&trans, &iter);
1532
1533         bch2_trans_exit(&trans);
1534         return ret;
1535 }
1536
1537 static int bch2_gc_stripes_done(struct bch_fs *c, bool initial,
1538                                 bool metadata_only)
1539 {
1540         struct btree_trans trans;
1541         struct btree_iter iter;
1542         struct bkey_s_c k;
1543         struct gc_stripe *m;
1544         const struct bch_stripe *s;
1545         char buf[200];
1546         unsigned i;
1547         int ret = 0;
1548
1549         if (metadata_only)
1550                 return 0;
1551
1552         bch2_trans_init(&trans, c, 0, 0);
1553
1554         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1555                            BTREE_ITER_PREFETCH, k, ret) {
1556                 if (k.k->type != KEY_TYPE_stripe)
1557                         continue;
1558
1559                 s = bkey_s_c_to_stripe(k).v;
1560                 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1561
1562                 for (i = 0; i < s->nr_blocks; i++)
1563                         if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1564                                 goto inconsistent;
1565                 continue;
1566 inconsistent:
1567                 if (fsck_err_on(true, c,
1568                                 "stripe has wrong block sector count %u:\n"
1569                                 "  %s\n"
1570                                 "  should be %u", i,
1571                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1572                                 m ? m->block_sectors[i] : 0)) {
1573                         struct bkey_i_stripe *new;
1574
1575                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1576                         if (!new) {
1577                                 ret = -ENOMEM;
1578                                 break;
1579                         }
1580
1581                         bkey_reassemble(&new->k_i, k);
1582
1583                         for (i = 0; i < new->v.nr_blocks; i++)
1584                                 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1585
1586                         ret = initial
1587                                 ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, &new->k_i)
1588                                 : __bch2_trans_do(&trans, NULL, NULL, 0,
1589                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1590                         kfree(new);
1591                 }
1592         }
1593 fsck_err:
1594         bch2_trans_iter_exit(&trans, &iter);
1595
1596         bch2_trans_exit(&trans);
1597         return ret;
1598 }
1599
1600 /**
1601  * bch2_gc - walk _all_ references to buckets, and recompute them:
1602  *
1603  * Order matters here:
1604  *  - Concurrent GC relies on the fact that we have a total ordering for
1605  *    everything that GC walks - see  gc_will_visit_node(),
1606  *    gc_will_visit_root()
1607  *
1608  *  - also, references move around in the course of index updates and
1609  *    various other crap: everything needs to agree on the ordering
1610  *    references are allowed to move around in - e.g., we're allowed to
1611  *    start with a reference owned by an open_bucket (the allocator) and
1612  *    move it to the btree, but not the reverse.
1613  *
1614  *    This is necessary to ensure that gc doesn't miss references that
1615  *    move around - if references move backwards in the ordering GC
1616  *    uses, GC could skip past them
1617  */
1618 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1619 {
1620         struct bch_dev *ca;
1621         u64 start_time = local_clock();
1622         unsigned i, iter = 0;
1623         int ret;
1624
1625         lockdep_assert_held(&c->state_lock);
1626         trace_gc_start(c);
1627
1628         down_write(&c->gc_lock);
1629
1630         /* flush interior btree updates: */
1631         closure_wait_event(&c->btree_interior_update_wait,
1632                            !bch2_btree_interior_updates_nr_pending(c));
1633 again:
1634         ret   = bch2_gc_start(c, metadata_only) ?:
1635                 bch2_gc_alloc_start(c, initial, metadata_only) ?:
1636                 bch2_gc_reflink_start(c, initial, metadata_only);
1637         if (ret)
1638                 goto out;
1639
1640         gc_pos_set(c, gc_phase(GC_PHASE_START));
1641
1642         bch2_mark_superblocks(c);
1643
1644         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1645             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1646             c->opts.fix_errors != FSCK_OPT_NO) {
1647                 bch_info(c, "starting topology repair pass");
1648                 ret = bch2_repair_topology(c);
1649                 if (ret)
1650                         goto out;
1651                 bch_info(c, "topology repair pass done");
1652
1653                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1654         }
1655
1656         ret = bch2_gc_btrees(c, initial, metadata_only);
1657
1658         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1659             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1660             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1661                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1662                 ret = 0;
1663         }
1664
1665         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1666                 ret = FSCK_ERR_EXIT;
1667
1668         if (ret)
1669                 goto out;
1670
1671 #if 0
1672         bch2_mark_pending_btree_node_frees(c);
1673 #endif
1674         c->gc_count++;
1675
1676         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1677             (!iter && bch2_test_restart_gc)) {
1678                 /*
1679                  * XXX: make sure gens we fixed got saved
1680                  */
1681                 if (iter++ <= 2) {
1682                         bch_info(c, "Second GC pass needed, restarting:");
1683                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1684                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1685
1686                         percpu_down_write(&c->mark_lock);
1687                         bch2_gc_free(c);
1688                         percpu_up_write(&c->mark_lock);
1689                         /* flush fsck errors, reset counters */
1690                         bch2_flush_fsck_errs(c);
1691
1692                         goto again;
1693                 }
1694
1695                 bch_info(c, "Unable to fix bucket gens, looping");
1696                 ret = -EINVAL;
1697         }
1698 out:
1699         if (!ret) {
1700                 bch2_journal_block(&c->journal);
1701
1702                 ret   = bch2_gc_stripes_done(c, initial, metadata_only) ?:
1703                         bch2_gc_reflink_done(c, initial, metadata_only) ?:
1704                         bch2_gc_alloc_done(c, initial, metadata_only) ?:
1705                         bch2_gc_done(c, initial, metadata_only);
1706
1707                 bch2_journal_unblock(&c->journal);
1708         }
1709
1710         percpu_down_write(&c->mark_lock);
1711         /* Indicates that gc is no longer in progress: */
1712         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1713
1714         bch2_gc_free(c);
1715         percpu_up_write(&c->mark_lock);
1716
1717         up_write(&c->gc_lock);
1718
1719         trace_gc_end(c);
1720         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1721
1722         /*
1723          * Wake up allocator in case it was waiting for buckets
1724          * because of not being able to inc gens
1725          */
1726         for_each_member_device(ca, c, i)
1727                 bch2_wake_allocator(ca);
1728
1729         /*
1730          * At startup, allocations can happen directly instead of via the
1731          * allocator thread - issue wakeup in case they blocked on gc_lock:
1732          */
1733         closure_wake_up(&c->freelist_wait);
1734         return ret;
1735 }
1736
1737 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1738 {
1739         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1740         const struct bch_extent_ptr *ptr;
1741
1742         percpu_down_read(&c->mark_lock);
1743         bkey_for_each_ptr(ptrs, ptr) {
1744                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1745                 struct bucket *g = PTR_BUCKET(ca, ptr);
1746
1747                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1748                         percpu_up_read(&c->mark_lock);
1749                         return true;
1750                 }
1751         }
1752
1753         bkey_for_each_ptr(ptrs, ptr) {
1754                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1755                 struct bucket *g = PTR_BUCKET(ca, ptr);
1756
1757                 if (gen_after(g->gc_gen, ptr->gen))
1758                         g->gc_gen = ptr->gen;
1759         }
1760         percpu_up_read(&c->mark_lock);
1761
1762         return false;
1763 }
1764
1765 /*
1766  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1767  * node pointers currently never have cached pointers that can become stale:
1768  */
1769 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1770 {
1771         struct btree_trans trans;
1772         struct btree_iter iter;
1773         struct bkey_s_c k;
1774         struct bkey_buf sk;
1775         int ret = 0, commit_err = 0;
1776
1777         bch2_bkey_buf_init(&sk);
1778         bch2_trans_init(&trans, c, 0, 0);
1779
1780         bch2_trans_iter_init(&trans, &iter, btree_id, POS_MIN,
1781                              BTREE_ITER_PREFETCH|
1782                              BTREE_ITER_NOT_EXTENTS|
1783                              BTREE_ITER_ALL_SNAPSHOTS);
1784
1785         while ((bch2_trans_begin(&trans),
1786                 k = bch2_btree_iter_peek(&iter)).k) {
1787                 ret = bkey_err(k);
1788
1789                 if (ret == -EINTR)
1790                         continue;
1791                 if (ret)
1792                         break;
1793
1794                 c->gc_gens_pos = iter.pos;
1795
1796                 if (gc_btree_gens_key(c, k) && !commit_err) {
1797                         bch2_bkey_buf_reassemble(&sk, c, k);
1798                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1799
1800                         commit_err =
1801                                 bch2_trans_update(&trans, &iter, sk.k, 0) ?:
1802                                 bch2_trans_commit(&trans, NULL, NULL,
1803                                                        BTREE_INSERT_NOWAIT|
1804                                                        BTREE_INSERT_NOFAIL);
1805                         if (commit_err == -EINTR) {
1806                                 commit_err = 0;
1807                                 continue;
1808                         }
1809                 }
1810
1811                 bch2_btree_iter_advance(&iter);
1812         }
1813         bch2_trans_iter_exit(&trans, &iter);
1814
1815         bch2_trans_exit(&trans);
1816         bch2_bkey_buf_exit(&sk, c);
1817
1818         return ret;
1819 }
1820
1821 int bch2_gc_gens(struct bch_fs *c)
1822 {
1823         struct bch_dev *ca;
1824         struct bucket_array *buckets;
1825         struct bucket *g;
1826         u64 start_time = local_clock();
1827         unsigned i;
1828         int ret;
1829
1830         /*
1831          * Ideally we would be using state_lock and not gc_lock here, but that
1832          * introduces a deadlock in the RO path - we currently take the state
1833          * lock at the start of going RO, thus the gc thread may get stuck:
1834          */
1835         down_read(&c->gc_lock);
1836
1837         for_each_member_device(ca, c, i) {
1838                 down_read(&ca->bucket_lock);
1839                 buckets = bucket_array(ca);
1840
1841                 for_each_bucket(g, buckets)
1842                         g->gc_gen = g->mark.gen;
1843                 up_read(&ca->bucket_lock);
1844         }
1845
1846         for (i = 0; i < BTREE_ID_NR; i++)
1847                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1848                         c->gc_gens_btree = i;
1849                         c->gc_gens_pos = POS_MIN;
1850                         ret = bch2_gc_btree_gens(c, i);
1851                         if (ret) {
1852                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1853                                 goto err;
1854                         }
1855                 }
1856
1857         for_each_member_device(ca, c, i) {
1858                 down_read(&ca->bucket_lock);
1859                 buckets = bucket_array(ca);
1860
1861                 for_each_bucket(g, buckets)
1862                         g->oldest_gen = g->gc_gen;
1863                 up_read(&ca->bucket_lock);
1864         }
1865
1866         c->gc_gens_btree        = 0;
1867         c->gc_gens_pos          = POS_MIN;
1868
1869         c->gc_count++;
1870
1871         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1872 err:
1873         up_read(&c->gc_lock);
1874         return ret;
1875 }
1876
1877 static int bch2_gc_thread(void *arg)
1878 {
1879         struct bch_fs *c = arg;
1880         struct io_clock *clock = &c->io_clock[WRITE];
1881         unsigned long last = atomic64_read(&clock->now);
1882         unsigned last_kick = atomic_read(&c->kick_gc);
1883         int ret;
1884
1885         set_freezable();
1886
1887         while (1) {
1888                 while (1) {
1889                         set_current_state(TASK_INTERRUPTIBLE);
1890
1891                         if (kthread_should_stop()) {
1892                                 __set_current_state(TASK_RUNNING);
1893                                 return 0;
1894                         }
1895
1896                         if (atomic_read(&c->kick_gc) != last_kick)
1897                                 break;
1898
1899                         if (c->btree_gc_periodic) {
1900                                 unsigned long next = last + c->capacity / 16;
1901
1902                                 if (atomic64_read(&clock->now) >= next)
1903                                         break;
1904
1905                                 bch2_io_clock_schedule_timeout(clock, next);
1906                         } else {
1907                                 schedule();
1908                         }
1909
1910                         try_to_freeze();
1911                 }
1912                 __set_current_state(TASK_RUNNING);
1913
1914                 last = atomic64_read(&clock->now);
1915                 last_kick = atomic_read(&c->kick_gc);
1916
1917                 /*
1918                  * Full gc is currently incompatible with btree key cache:
1919                  */
1920 #if 0
1921                 ret = bch2_gc(c, false, false);
1922 #else
1923                 ret = bch2_gc_gens(c);
1924 #endif
1925                 if (ret < 0)
1926                         bch_err(c, "btree gc failed: %i", ret);
1927
1928                 debug_check_no_locks_held();
1929         }
1930
1931         return 0;
1932 }
1933
1934 void bch2_gc_thread_stop(struct bch_fs *c)
1935 {
1936         struct task_struct *p;
1937
1938         p = c->gc_thread;
1939         c->gc_thread = NULL;
1940
1941         if (p) {
1942                 kthread_stop(p);
1943                 put_task_struct(p);
1944         }
1945 }
1946
1947 int bch2_gc_thread_start(struct bch_fs *c)
1948 {
1949         struct task_struct *p;
1950
1951         if (c->gc_thread)
1952                 return 0;
1953
1954         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1955         if (IS_ERR(p)) {
1956                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1957                 return PTR_ERR(p);
1958         }
1959
1960         get_task_struct(p);
1961         c->gc_thread = p;
1962         wake_up_process(p);
1963         return 0;
1964 }