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_locking.h"
  13 #include "btree_update_interior.h"
  14 #include "btree_io.h"
  15 #include "btree_gc.h"
  16 #include "buckets.h"
  17 #include "clock.h"
  18 #include "debug.h"
  19 #include "ec.h"
  20 #include "error.h"
  21 #include "extents.h"
  22 #include "journal.h"
  23 #include "keylist.h"
  24 #include "move.h"
  25 #include "recovery.h"
  26 #include "reflink.h"
  27 #include "replicas.h"
  28 #include "super-io.h"
  29
  30 #include <linux/slab.h>
  31 #include <linux/bitops.h>
  32 #include <linux/freezer.h>
  33 #include <linux/kthread.h>
  34 #include <linux/preempt.h>
  35 #include <linux/rcupdate.h>
  36 #include <linux/sched/task.h>
  37 #include <trace/events/bcachefs.h>
  38
  39 #define DROP_THIS_NODE          10
  40 #define DROP_PREV_NODE          11
  41
  42 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  43 {
  44         preempt_disable();
  45         write_seqcount_begin(&c->gc_pos_lock);
  46         c->gc_pos = new_pos;
  47         write_seqcount_end(&c->gc_pos_lock);
  48         preempt_enable();
  49 }
  50
  51 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  52 {
  53         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  54         __gc_pos_set(c, new_pos);
  55 }
  56
  57 /*
  58  * Missing: if an interior btree node is empty, we need to do something -
  59  * perhaps just kill it
  60  */
  61 static int bch2_gc_check_topology(struct bch_fs *c,
  62                                   struct btree *b,
  63                                   struct bkey_buf *prev,
  64                                   struct bkey_buf cur,
  65                                   bool is_last)
  66 {
  67         struct bpos node_start  = b->data->min_key;
  68         struct bpos node_end    = b->data->max_key;
  69         struct bpos expected_start = bkey_deleted(&prev->k->k)
  70                 ? node_start
  71                 : bpos_successor(prev->k->k.p);
  72         char buf1[200], buf2[200];
  73         int ret = 0;
  74
  75         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  76                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  77
  78                 if (bkey_deleted(&prev->k->k)) {
  79                         struct printbuf out = PBUF(buf1);
  80                         pr_buf(&out, "start of node: ");
  81                         bch2_bpos_to_text(&out, node_start);
  82                 } else {
  83                         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
  84                 }
  85
  86                 if (bpos_cmp(expected_start, bp->v.min_key)) {
  87                         bch2_topology_error(c);
  88
  89                         if (__fsck_err(c,
  90                                   FSCK_CAN_FIX|
  91                                   FSCK_CAN_IGNORE|
  92                                   FSCK_NO_RATELIMIT,
  93                                   "btree node with incorrect min_key at btree %s level %u:\n"
  94                                   "  prev %s\n"
  95                                   "  cur %s",
  96                                   bch2_btree_ids[b->c.btree_id], b->c.level,
  97                                   buf1,
  98                                   (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2)) &&
  99                             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 100                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 101                                 return FSCK_ERR_START_TOPOLOGY_REPAIR;
 102                         } else {
 103                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 104                         }
 105                 }
 106         }
 107
 108         if (is_last && bpos_cmp(cur.k->k.p, node_end)) {
 109                 bch2_topology_error(c);
 110
 111                 if (__fsck_err(c,
 112                           FSCK_CAN_FIX|
 113                           FSCK_CAN_IGNORE|
 114                           FSCK_NO_RATELIMIT,
 115                           "btree node with incorrect max_key at btree %s level %u:\n"
 116                           "  %s\n"
 117                           "  expected %s",
 118                           bch2_btree_ids[b->c.btree_id], b->c.level,
 119                           (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
 120                           (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2)) &&
 121                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 122                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 123                         return FSCK_ERR_START_TOPOLOGY_REPAIR;
 124                 } else {
 125                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 126                 }
 127         }
 128
 129         bch2_bkey_buf_copy(prev, c, cur.k);
 130 fsck_err:
 131         return ret;
 132 }
 133
 134 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 135 {
 136         switch (b->key.k.type) {
 137         case KEY_TYPE_btree_ptr: {
 138                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 139
 140                 dst->k.p                = src->k.p;
 141                 dst->v.mem_ptr          = 0;
 142                 dst->v.seq              = b->data->keys.seq;
 143                 dst->v.sectors_written  = 0;
 144                 dst->v.flags            = 0;
 145                 dst->v.min_key          = b->data->min_key;
 146                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 147                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 148                 break;
 149         }
 150         case KEY_TYPE_btree_ptr_v2:
 151                 bkey_copy(&dst->k_i, &b->key);
 152                 break;
 153         default:
 154                 BUG();
 155         }
 156 }
 157
 158 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 159 {
 160         struct bkey_i_btree_ptr_v2 *new;
 161         int ret;
 162
 163         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 164         if (!new)
 165                 return -ENOMEM;
 166
 167         btree_ptr_to_v2(b, new);
 168         b->data->min_key        = new_min;
 169         new->v.min_key          = new_min;
 170         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 171
 172         ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 173         if (ret) {
 174                 kfree(new);
 175                 return ret;
 176         }
 177
 178         bch2_btree_node_drop_keys_outside_node(b);
 179
 180         return 0;
 181 }
 182
 183 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 184 {
 185         struct bkey_i_btree_ptr_v2 *new;
 186         int ret;
 187
 188         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 189         if (ret)
 190                 return 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->max_key        = new_max;
 198         new->k.p                = new_max;
 199         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 200
 201         ret = bch2_journal_key_insert(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         mutex_lock(&c->btree_cache.lock);
 210         bch2_btree_node_hash_remove(&c->btree_cache, b);
 211
 212         bkey_copy(&b->key, &new->k_i);
 213         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 214         BUG_ON(ret);
 215         mutex_unlock(&c->btree_cache.lock);
 216         return 0;
 217 }
 218
 219 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 220                                         struct btree *prev, struct btree *cur)
 221 {
 222         struct bpos expected_start = !prev
 223                 ? b->data->min_key
 224                 : bpos_successor(prev->key.k.p);
 225         char buf1[200], buf2[200];
 226         int ret = 0;
 227
 228         if (!prev) {
 229                 struct printbuf out = PBUF(buf1);
 230                 pr_buf(&out, "start of node: ");
 231                 bch2_bpos_to_text(&out, b->data->min_key);
 232         } else {
 233                 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
 234         }
 235
 236         bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key));
 237
 238         if (prev &&
 239             bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 240             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 241                 /* cur overwrites prev: */
 242
 243                 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
 244                                                  cur->data->min_key) >= 0, c,
 245                                 "btree node overwritten by next node at btree %s level %u:\n"
 246                                 "  node %s\n"
 247                                 "  next %s",
 248                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 249                                 buf1, buf2))
 250                         return DROP_PREV_NODE;
 251
 252                 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
 253                                                  bpos_predecessor(cur->data->min_key)), c,
 254                                 "btree node with incorrect max_key at btree %s level %u:\n"
 255                                 "  node %s\n"
 256                                 "  next %s",
 257                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 258                                 buf1, buf2))
 259                         ret = set_node_max(c, prev,
 260                                            bpos_predecessor(cur->data->min_key));
 261         } else {
 262                 /* prev overwrites cur: */
 263
 264                 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
 265                                                  cur->data->max_key) >= 0, c,
 266                                 "btree node overwritten by prev node at btree %s level %u:\n"
 267                                 "  prev %s\n"
 268                                 "  node %s",
 269                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 270                                 buf1, buf2))
 271                         return DROP_THIS_NODE;
 272
 273                 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 274                                 "btree node with incorrect min_key at btree %s level %u:\n"
 275                                 "  prev %s\n"
 276                                 "  node %s",
 277                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 278                                 buf1, buf2))
 279                     ret = set_node_min(c, cur, expected_start);
 280         }
 281 fsck_err:
 282         return ret;
 283 }
 284
 285 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 286                                  struct btree *child)
 287 {
 288         char buf1[200], buf2[200];
 289         int ret = 0;
 290
 291         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 292                         "btree node with incorrect max_key at btree %s level %u:\n"
 293                         "  %s\n"
 294                         "  expected %s",
 295                         bch2_btree_ids[b->c.btree_id], b->c.level,
 296                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
 297                         (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
 298                 ret = set_node_max(c, child, b->key.k.p);
 299                 if (ret)
 300                         return ret;
 301         }
 302 fsck_err:
 303         return ret;
 304 }
 305
 306 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 307 {
 308         struct btree_and_journal_iter iter;
 309         struct bkey_s_c k;
 310         struct bkey_buf prev_k, cur_k;
 311         struct btree *prev = NULL, *cur = NULL;
 312         bool have_child, dropped_children = false;
 313         char buf[200];
 314         int ret = 0;
 315
 316         if (!b->c.level)
 317                 return 0;
 318 again:
 319         prev = NULL;
 320         have_child = dropped_children = false;
 321         bch2_bkey_buf_init(&prev_k);
 322         bch2_bkey_buf_init(&cur_k);
 323         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 324
 325         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 326                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 327                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 328
 329                 bch2_btree_and_journal_iter_advance(&iter);
 330                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 331
 332                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 333                                         b->c.btree_id, b->c.level - 1,
 334                                         false);
 335                 ret = PTR_ERR_OR_ZERO(cur);
 336
 337                 if (mustfix_fsck_err_on(ret == -EIO, c,
 338                                 "Unreadable btree node at btree %s level %u:\n"
 339                                 "  %s",
 340                                 bch2_btree_ids[b->c.btree_id],
 341                                 b->c.level - 1,
 342                                 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
 343                         bch2_btree_node_evict(c, cur_k.k);
 344                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 345                                                       b->c.level, cur_k.k->k.p);
 346                         if (ret)
 347                                 break;
 348                         continue;
 349                 }
 350
 351                 if (ret) {
 352                         bch_err(c, "%s: error %i getting btree node",
 353                                 __func__, ret);
 354                         break;
 355                 }
 356
 357                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 358
 359                 if (ret == DROP_THIS_NODE) {
 360                         six_unlock_read(&cur->c.lock);
 361                         bch2_btree_node_evict(c, cur_k.k);
 362                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 363                                                       b->c.level, cur_k.k->k.p);
 364                         if (ret)
 365                                 break;
 366                         continue;
 367                 }
 368
 369                 if (prev)
 370                         six_unlock_read(&prev->c.lock);
 371                 prev = NULL;
 372
 373                 if (ret == DROP_PREV_NODE) {
 374                         bch2_btree_node_evict(c, prev_k.k);
 375                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 376                                                       b->c.level, prev_k.k->k.p);
 377                         if (ret)
 378                                 break;
 379
 380                         bch2_btree_and_journal_iter_exit(&iter);
 381                         bch2_bkey_buf_exit(&prev_k, c);
 382                         bch2_bkey_buf_exit(&cur_k, c);
 383                         goto again;
 384                 } else if (ret)
 385                         break;
 386
 387                 prev = cur;
 388                 cur = NULL;
 389                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 390         }
 391
 392         if (!ret && !IS_ERR_OR_NULL(prev)) {
 393                 BUG_ON(cur);
 394                 ret = btree_repair_node_end(c, b, prev);
 395         }
 396
 397         if (!IS_ERR_OR_NULL(prev))
 398                 six_unlock_read(&prev->c.lock);
 399         prev = NULL;
 400         if (!IS_ERR_OR_NULL(cur))
 401                 six_unlock_read(&cur->c.lock);
 402         cur = NULL;
 403
 404         if (ret)
 405                 goto err;
 406
 407         bch2_btree_and_journal_iter_exit(&iter);
 408         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 409
 410         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 411                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 412                 bch2_btree_and_journal_iter_advance(&iter);
 413
 414                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 415                                         b->c.btree_id, b->c.level - 1,
 416                                         false);
 417                 ret = PTR_ERR_OR_ZERO(cur);
 418
 419                 if (ret) {
 420                         bch_err(c, "%s: error %i getting btree node",
 421                                 __func__, ret);
 422                         goto err;
 423                 }
 424
 425                 ret = bch2_btree_repair_topology_recurse(c, cur);
 426                 six_unlock_read(&cur->c.lock);
 427                 cur = NULL;
 428
 429                 if (ret == DROP_THIS_NODE) {
 430                         bch2_btree_node_evict(c, cur_k.k);
 431                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 432                                                       b->c.level, cur_k.k->k.p);
 433                         dropped_children = true;
 434                 }
 435
 436                 if (ret)
 437                         goto err;
 438
 439                 have_child = true;
 440         }
 441
 442         if (mustfix_fsck_err_on(!have_child, c,
 443                         "empty interior btree node at btree %s level %u\n"
 444                         "  %s",
 445                         bch2_btree_ids[b->c.btree_id],
 446                         b->c.level,
 447                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
 448                 ret = DROP_THIS_NODE;
 449 err:
 450 fsck_err:
 451         if (!IS_ERR_OR_NULL(prev))
 452                 six_unlock_read(&prev->c.lock);
 453         if (!IS_ERR_OR_NULL(cur))
 454                 six_unlock_read(&cur->c.lock);
 455
 456         bch2_btree_and_journal_iter_exit(&iter);
 457         bch2_bkey_buf_exit(&prev_k, c);
 458         bch2_bkey_buf_exit(&cur_k, c);
 459
 460         if (!ret && dropped_children)
 461                 goto again;
 462
 463         return ret;
 464 }
 465
 466 static int bch2_repair_topology(struct bch_fs *c)
 467 {
 468         struct btree *b;
 469         unsigned i;
 470         int ret = 0;
 471
 472         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 473                 b = c->btree_roots[i].b;
 474                 if (btree_node_fake(b))
 475                         continue;
 476
 477                 six_lock_read(&b->c.lock, NULL, NULL);
 478                 ret = bch2_btree_repair_topology_recurse(c, b);
 479                 six_unlock_read(&b->c.lock);
 480
 481                 if (ret == DROP_THIS_NODE) {
 482                         bch_err(c, "empty btree root - repair unimplemented");
 483                         ret = FSCK_ERR_EXIT;
 484                 }
 485         }
 486
 487         return ret;
 488 }
 489
 490 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 491                                unsigned level, bool is_root,
 492                                struct bkey_s_c *k)
 493 {
 494         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 495         const union bch_extent_entry *entry;
 496         struct extent_ptr_decoded p = { 0 };
 497         bool do_update = false;
 498         char buf[200];
 499         int ret = 0;
 500
 501         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 502                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 503                 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
 504                 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
 505                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 506
 507                 if (fsck_err_on(!g->gen_valid, c,
 508                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 509                                 "while marking %s",
 510                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 511                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 512                                 p.ptr.gen,
 513                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 514                         if (!p.ptr.cached) {
 515                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 516                                 g2->gen_valid   = g->gen_valid          = true;
 517                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 518                         } else {
 519                                 do_update = true;
 520                         }
 521                 }
 522
 523                 if (fsck_err_on(data_type == BCH_DATA_btree &&
 524                                 g->mark.gen != p.ptr.gen, c,
 525                                 "bucket %u:%zu data type %s has metadata but wrong gen: %u != %u\n"
 526                                 "while marking %s",
 527                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 528                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 529                                 p.ptr.gen, g->mark.gen,
 530                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 531                         g2->_mark.data_type     = g->_mark.data_type    = data_type;
 532                         g2->gen_valid           = g->gen_valid          = true;
 533                         set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 534                 }
 535
 536                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 537                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 538                                 "while marking %s",
 539                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 540                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 541                                 p.ptr.gen, g->mark.gen,
 542                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 543                         if (!p.ptr.cached) {
 544                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 545                                 g2->gen_valid   = g->gen_valid  = true;
 546                                 g2->_mark.data_type             = 0;
 547                                 g2->_mark.dirty_sectors         = 0;
 548                                 g2->_mark.cached_sectors        = 0;
 549                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 550                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 551                         } else {
 552                                 do_update = true;
 553                         }
 554                 }
 555
 556                 if (fsck_err_on(!p.ptr.cached &&
 557                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 558                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 559                                 "while marking %s",
 560                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 561                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 562                                 p.ptr.gen, g->mark.gen,
 563                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 564                         do_update = true;
 565
 566                 if (p.ptr.gen != g->mark.gen)
 567                         continue;
 568
 569                 if (fsck_err_on(g->mark.data_type &&
 570                                 g->mark.data_type != data_type, c,
 571                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 572                                 "while marking %s",
 573                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 574                                 bch2_data_types[g->mark.data_type],
 575                                 bch2_data_types[data_type],
 576                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 577                         if (data_type == BCH_DATA_btree) {
 578                                 g2->_mark.data_type     = g->_mark.data_type    = data_type;
 579                                 g2->gen_valid           = g->gen_valid          = true;
 580                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 581                         } else {
 582                                 do_update = true;
 583                         }
 584                 }
 585
 586                 if (p.has_ec) {
 587                         struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
 588
 589                         if (fsck_err_on(!m || !m->alive, c,
 590                                         "pointer to nonexistent stripe %llu\n"
 591                                         "while marking %s",
 592                                         (u64) p.ec.idx,
 593                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 594                                 do_update = true;
 595
 596                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 597                                         "pointer does not match stripe %llu\n"
 598                                         "while marking %s",
 599                                         (u64) p.ec.idx,
 600                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 601                                 do_update = true;
 602                 }
 603         }
 604
 605         if (do_update) {
 606                 struct bkey_ptrs ptrs;
 607                 union bch_extent_entry *entry;
 608                 struct bch_extent_ptr *ptr;
 609                 struct bkey_i *new;
 610
 611                 if (is_root) {
 612                         bch_err(c, "cannot update btree roots yet");
 613                         return -EINVAL;
 614                 }
 615
 616                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 617                 if (!new) {
 618                         bch_err(c, "%s: error allocating new key", __func__);
 619                         return -ENOMEM;
 620                 }
 621
 622                 bkey_reassemble(new, *k);
 623
 624                 if (level) {
 625                         /*
 626                          * We don't want to drop btree node pointers - if the
 627                          * btree node isn't there anymore, the read path will
 628                          * sort it out:
 629                          */
 630                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 631                         bkey_for_each_ptr(ptrs, ptr) {
 632                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 633                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 634
 635                                 ptr->gen = g->mark.gen;
 636                         }
 637                 } else {
 638                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 639                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 640                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 641                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 642
 643                                 (ptr->cached &&
 644                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 645                                 (!ptr->cached &&
 646                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 647                                 (g->mark.data_type &&
 648                                  g->mark.data_type != data_type);
 649                         }));
 650 again:
 651                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 652                         bkey_extent_entry_for_each(ptrs, entry) {
 653                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 654                                         struct stripe *m = genradix_ptr(&c->stripes[true],
 655                                                                         entry->stripe_ptr.idx);
 656                                         union bch_extent_entry *next_ptr;
 657
 658                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 659                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 660                                                         goto found;
 661                                         next_ptr = NULL;
 662 found:
 663                                         if (!next_ptr) {
 664                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 665                                                 continue;
 666                                         }
 667
 668                                         if (!m || !m->alive ||
 669                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 670                                                                        &next_ptr->ptr,
 671                                                                        m->sectors)) {
 672                                                 bch2_bkey_extent_entry_drop(new, entry);
 673                                                 goto again;
 674                                         }
 675                                 }
 676                         }
 677                 }
 678
 679                 ret = bch2_journal_key_insert(c, btree_id, level, new);
 680                 if (ret)
 681                         kfree(new);
 682                 else
 683                         *k = bkey_i_to_s_c(new);
 684         }
 685 fsck_err:
 686         return ret;
 687 }
 688
 689 /* marking of btree keys/nodes: */
 690
 691 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 692                             unsigned level, bool is_root,
 693                             struct bkey_s_c *k,
 694                             u8 *max_stale, bool initial)
 695 {
 696         struct bch_fs *c = trans->c;
 697         struct bkey_ptrs_c ptrs;
 698         const struct bch_extent_ptr *ptr;
 699         unsigned flags =
 700                 BTREE_TRIGGER_GC|
 701                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 702         char buf[200];
 703         int ret = 0;
 704
 705         if (initial) {
 706                 BUG_ON(bch2_journal_seq_verify &&
 707                        k->k->version.lo > journal_cur_seq(&c->journal));
 708
 709                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 710                 if (ret)
 711                         goto err;
 712
 713                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 714                                 "key version number higher than recorded: %llu > %llu",
 715                                 k->k->version.lo,
 716                                 atomic64_read(&c->key_version)))
 717                         atomic64_set(&c->key_version, k->k->version.lo);
 718
 719                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 720                     fsck_err_on(!bch2_bkey_replicas_marked(c, *k), c,
 721                                 "superblock not marked as containing replicas\n"
 722                                 "  while marking %s",
 723                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 724                         ret = bch2_mark_bkey_replicas(c, *k);
 725                         if (ret) {
 726                                 bch_err(c, "error marking bkey replicas: %i", ret);
 727                                 goto err;
 728                         }
 729                 }
 730         }
 731
 732         ptrs = bch2_bkey_ptrs_c(*k);
 733         bkey_for_each_ptr(ptrs, ptr) {
 734                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 735                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 736
 737                 if (gen_after(g->oldest_gen, ptr->gen))
 738                         g->oldest_gen = ptr->gen;
 739
 740                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 741         }
 742
 743         ret = bch2_mark_key(trans, *k, flags);
 744 fsck_err:
 745 err:
 746         if (ret)
 747                 bch_err(c, "%s: ret %i", __func__, ret);
 748         return ret;
 749 }
 750
 751 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, u8 *max_stale,
 752                               bool initial)
 753 {
 754         struct bch_fs *c = trans->c;
 755         struct btree_node_iter iter;
 756         struct bkey unpacked;
 757         struct bkey_s_c k;
 758         struct bkey_buf prev, cur;
 759         int ret = 0;
 760
 761         *max_stale = 0;
 762
 763         if (!btree_node_type_needs_gc(btree_node_type(b)))
 764                 return 0;
 765
 766         bch2_btree_node_iter_init_from_start(&iter, b);
 767         bch2_bkey_buf_init(&prev);
 768         bch2_bkey_buf_init(&cur);
 769         bkey_init(&prev.k->k);
 770
 771         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 772                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 773                                        &k, max_stale, initial);
 774                 if (ret)
 775                         break;
 776
 777                 bch2_btree_node_iter_advance(&iter, b);
 778
 779                 if (b->c.level) {
 780                         bch2_bkey_buf_reassemble(&cur, c, k);
 781
 782                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 783                                         bch2_btree_node_iter_end(&iter));
 784                         if (ret)
 785                                 break;
 786                 }
 787         }
 788
 789         bch2_bkey_buf_exit(&cur, c);
 790         bch2_bkey_buf_exit(&prev, c);
 791         return ret;
 792 }
 793
 794 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 795                          bool initial, bool metadata_only)
 796 {
 797         struct bch_fs *c = trans->c;
 798         struct btree_iter iter;
 799         struct btree *b;
 800         unsigned depth = metadata_only                  ? 1
 801                 : bch2_expensive_debug_checks           ? 0
 802                 : !btree_node_type_needs_gc(btree_id)   ? 1
 803                 : 0;
 804         u8 max_stale = 0;
 805         int ret = 0;
 806
 807         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 808
 809         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 810                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 811                 bch2_verify_btree_nr_keys(b);
 812
 813                 gc_pos_set(c, gc_pos_btree_node(b));
 814
 815                 ret = btree_gc_mark_node(trans, b, &max_stale, initial);
 816                 if (ret)
 817                         break;
 818
 819                 if (!initial) {
 820                         if (max_stale > 64)
 821                                 bch2_btree_node_rewrite(trans, &iter, b,
 822                                                 BTREE_INSERT_NOWAIT|
 823                                                 BTREE_INSERT_GC_LOCK_HELD);
 824                         else if (!bch2_btree_gc_rewrite_disabled &&
 825                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 826                                 bch2_btree_node_rewrite(trans, &iter,
 827                                                 b, BTREE_INSERT_NOWAIT|
 828                                                 BTREE_INSERT_GC_LOCK_HELD);
 829                 }
 830         }
 831         bch2_trans_iter_exit(trans, &iter);
 832
 833         if (ret)
 834                 return ret;
 835
 836         mutex_lock(&c->btree_root_lock);
 837         b = c->btree_roots[btree_id].b;
 838         if (!btree_node_fake(b)) {
 839                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 840
 841                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
 842                                        &k, &max_stale, initial);
 843         }
 844         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 845         mutex_unlock(&c->btree_root_lock);
 846
 847         return ret;
 848 }
 849
 850 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 851                                       unsigned target_depth)
 852 {
 853         struct bch_fs *c = trans->c;
 854         struct btree_and_journal_iter iter;
 855         struct bkey_s_c k;
 856         struct bkey_buf cur, prev;
 857         u8 max_stale = 0;
 858         char buf[200];
 859         int ret = 0;
 860
 861         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 862         bch2_bkey_buf_init(&prev);
 863         bch2_bkey_buf_init(&cur);
 864         bkey_init(&prev.k->k);
 865
 866         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 867                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 868                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 869
 870                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 871                                        &k, &max_stale, true);
 872                 if (ret) {
 873                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 874                         goto fsck_err;
 875                 }
 876
 877                 if (b->c.level) {
 878                         bch2_bkey_buf_reassemble(&cur, c, k);
 879                         k = bkey_i_to_s_c(cur.k);
 880
 881                         bch2_btree_and_journal_iter_advance(&iter);
 882
 883                         ret = bch2_gc_check_topology(c, b,
 884                                         &prev, cur,
 885                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 886                         if (ret)
 887                                 goto fsck_err;
 888                 } else {
 889                         bch2_btree_and_journal_iter_advance(&iter);
 890                 }
 891         }
 892
 893         if (b->c.level > target_depth) {
 894                 bch2_btree_and_journal_iter_exit(&iter);
 895                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 896
 897                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 898                         struct btree *child;
 899
 900                         bch2_bkey_buf_reassemble(&cur, c, k);
 901                         bch2_btree_and_journal_iter_advance(&iter);
 902
 903                         child = bch2_btree_node_get_noiter(c, cur.k,
 904                                                 b->c.btree_id, b->c.level - 1,
 905                                                 false);
 906                         ret = PTR_ERR_OR_ZERO(child);
 907
 908                         if (ret == -EIO) {
 909                                 bch2_topology_error(c);
 910
 911                                 if (__fsck_err(c,
 912                                           FSCK_CAN_FIX|
 913                                           FSCK_CAN_IGNORE|
 914                                           FSCK_NO_RATELIMIT,
 915                                           "Unreadable btree node at btree %s level %u:\n"
 916                                           "  %s",
 917                                           bch2_btree_ids[b->c.btree_id],
 918                                           b->c.level - 1,
 919                                           (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
 920                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 921                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 922                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 923                                         goto fsck_err;
 924                                 } else {
 925                                         /* Continue marking when opted to not
 926                                          * fix the error: */
 927                                         ret = 0;
 928                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 929                                         continue;
 930                                 }
 931                         } else if (ret) {
 932                                 bch_err(c, "%s: error %i getting btree node",
 933                                         __func__, ret);
 934                                 break;
 935                         }
 936
 937                         ret = bch2_gc_btree_init_recurse(trans, child,
 938                                                          target_depth);
 939                         six_unlock_read(&child->c.lock);
 940
 941                         if (ret)
 942                                 break;
 943                 }
 944         }
 945 fsck_err:
 946         bch2_bkey_buf_exit(&cur, c);
 947         bch2_bkey_buf_exit(&prev, c);
 948         bch2_btree_and_journal_iter_exit(&iter);
 949         return ret;
 950 }
 951
 952 static int bch2_gc_btree_init(struct btree_trans *trans,
 953                               enum btree_id btree_id,
 954                               bool metadata_only)
 955 {
 956         struct bch_fs *c = trans->c;
 957         struct btree *b;
 958         unsigned target_depth = metadata_only           ? 1
 959                 : bch2_expensive_debug_checks           ? 0
 960                 : !btree_node_type_needs_gc(btree_id)   ? 1
 961                 : 0;
 962         u8 max_stale = 0;
 963         char buf[100];
 964         int ret = 0;
 965
 966         b = c->btree_roots[btree_id].b;
 967
 968         if (btree_node_fake(b))
 969                 return 0;
 970
 971         six_lock_read(&b->c.lock, NULL, NULL);
 972         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 973                         "btree root with incorrect min_key: %s",
 974                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 975                 bch_err(c, "repair unimplemented");
 976                 ret = FSCK_ERR_EXIT;
 977                 goto fsck_err;
 978         }
 979
 980         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
 981                         "btree root with incorrect max_key: %s",
 982                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 983                 bch_err(c, "repair unimplemented");
 984                 ret = FSCK_ERR_EXIT;
 985                 goto fsck_err;
 986         }
 987
 988         if (b->c.level >= target_depth)
 989                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
 990
 991         if (!ret) {
 992                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 993
 994                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
 995                                        &k, &max_stale, true);
 996         }
 997 fsck_err:
 998         six_unlock_read(&b->c.lock);
 999
1000         if (ret < 0)
1001                 bch_err(c, "%s: ret %i", __func__, ret);
1002         return ret;
1003 }
1004
1005 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1006 {
1007         return  (int) btree_id_to_gc_phase(l) -
1008                 (int) btree_id_to_gc_phase(r);
1009 }
1010
1011 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1012 {
1013         struct btree_trans trans;
1014         enum btree_id ids[BTREE_ID_NR];
1015         unsigned i;
1016         int ret = 0;
1017
1018         bch2_trans_init(&trans, c, 0, 0);
1019
1020         for (i = 0; i < BTREE_ID_NR; i++)
1021                 ids[i] = i;
1022         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1023
1024         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1025                 ret = initial
1026                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1027                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1028
1029         if (ret < 0)
1030                 bch_err(c, "%s: ret %i", __func__, ret);
1031
1032         bch2_trans_exit(&trans);
1033         return ret;
1034 }
1035
1036 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1037                                   u64 start, u64 end,
1038                                   enum bch_data_type type,
1039                                   unsigned flags)
1040 {
1041         u64 b = sector_to_bucket(ca, start);
1042
1043         do {
1044                 unsigned sectors =
1045                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1046
1047                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1048                                           gc_phase(GC_PHASE_SB), flags);
1049                 b++;
1050                 start += sectors;
1051         } while (start < end);
1052 }
1053
1054 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1055                               unsigned flags)
1056 {
1057         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1058         unsigned i;
1059         u64 b;
1060
1061         /*
1062          * This conditional is kind of gross, but we may be called from the
1063          * device add path, before the new device has actually been added to the
1064          * running filesystem:
1065          */
1066         if (c) {
1067                 lockdep_assert_held(&c->sb_lock);
1068                 percpu_down_read(&c->mark_lock);
1069         }
1070
1071         for (i = 0; i < layout->nr_superblocks; i++) {
1072                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1073
1074                 if (offset == BCH_SB_SECTOR)
1075                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1076                                               BCH_DATA_sb, flags);
1077
1078                 mark_metadata_sectors(c, ca, offset,
1079                                       offset + (1 << layout->sb_max_size_bits),
1080                                       BCH_DATA_sb, flags);
1081         }
1082
1083         for (i = 0; i < ca->journal.nr; i++) {
1084                 b = ca->journal.buckets[i];
1085                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1086                                           ca->mi.bucket_size,
1087                                           gc_phase(GC_PHASE_SB), flags);
1088         }
1089
1090         if (c)
1091                 percpu_up_read(&c->mark_lock);
1092 }
1093
1094 static void bch2_mark_superblocks(struct bch_fs *c)
1095 {
1096         struct bch_dev *ca;
1097         unsigned i;
1098
1099         mutex_lock(&c->sb_lock);
1100         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1101
1102         for_each_online_member(ca, c, i)
1103                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1104         mutex_unlock(&c->sb_lock);
1105 }
1106
1107 #if 0
1108 /* Also see bch2_pending_btree_node_free_insert_done() */
1109 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1110 {
1111         struct btree_update *as;
1112         struct pending_btree_node_free *d;
1113
1114         mutex_lock(&c->btree_interior_update_lock);
1115         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1116
1117         for_each_pending_btree_node_free(c, as, d)
1118                 if (d->index_update_done)
1119                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1120
1121         mutex_unlock(&c->btree_interior_update_lock);
1122 }
1123 #endif
1124
1125 static void bch2_gc_free(struct bch_fs *c)
1126 {
1127         struct bch_dev *ca;
1128         unsigned i;
1129
1130         genradix_free(&c->stripes[1]);
1131
1132         for_each_member_device(ca, c, i) {
1133                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1134                         sizeof(struct bucket_array) +
1135                         ca->mi.nbuckets * sizeof(struct bucket));
1136                 ca->buckets[1] = NULL;
1137
1138                 free_percpu(ca->usage_gc);
1139                 ca->usage_gc = NULL;
1140         }
1141
1142         free_percpu(c->usage_gc);
1143         c->usage_gc = NULL;
1144 }
1145
1146 static int bch2_gc_done(struct bch_fs *c,
1147                         bool initial, bool metadata_only)
1148 {
1149         struct bch_dev *ca = NULL;
1150         bool verify = !metadata_only && (!initial ||
1151                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1152         unsigned i, dev;
1153         int ret = 0;
1154
1155 #define copy_field(_f, _msg, ...)                                       \
1156         if (dst->_f != src->_f) {                                       \
1157                 if (verify)                                             \
1158                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1159                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1160                 dst->_f = src->_f;                                      \
1161                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1162         }
1163 #define copy_stripe_field(_f, _msg, ...)                                \
1164         if (dst->_f != src->_f) {                                       \
1165                 if (verify)                                             \
1166                         fsck_err(c, "stripe %zu has wrong "_msg         \
1167                                 ": got %u, should be %u",               \
1168                                 iter.pos, ##__VA_ARGS__,                \
1169                                 dst->_f, src->_f);                      \
1170                 dst->_f = src->_f;                                      \
1171                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1172         }
1173 #define copy_bucket_field(_f)                                           \
1174         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
1175                 if (verify)                                             \
1176                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
1177                                 ": got %u, should be %u", dev, b,       \
1178                                 dst->b[b].mark.gen,                     \
1179                                 bch2_data_types[dst->b[b].mark.data_type],\
1180                                 dst->b[b].mark._f, src->b[b].mark._f);  \
1181                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
1182                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1183         }
1184 #define copy_dev_field(_f, _msg, ...)                                   \
1185         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1186 #define copy_fs_field(_f, _msg, ...)                                    \
1187         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1188
1189         if (!metadata_only) {
1190                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
1191                 struct stripe *dst, *src;
1192
1193                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
1194                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
1195
1196                         if (dst->alive          != src->alive ||
1197                             dst->sectors        != src->sectors ||
1198                             dst->algorithm      != src->algorithm ||
1199                             dst->nr_blocks      != src->nr_blocks ||
1200                             dst->nr_redundant   != src->nr_redundant) {
1201                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
1202                                 ret = -EINVAL;
1203                                 goto fsck_err;
1204                         }
1205
1206                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
1207                                 copy_stripe_field(block_sectors[i],
1208                                                   "block_sectors[%u]", i);
1209
1210                         dst->blocks_nonempty = 0;
1211                         for (i = 0; i < dst->nr_blocks; i++)
1212                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
1213
1214                         genradix_iter_advance(&iter, &c->stripes[1]);
1215                 }
1216         }
1217
1218         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1219                 bch2_fs_usage_acc_to_base(c, i);
1220
1221         for_each_member_device(ca, c, dev) {
1222                 struct bucket_array *dst = __bucket_array(ca, 0);
1223                 struct bucket_array *src = __bucket_array(ca, 1);
1224                 size_t b;
1225
1226                 for (b = 0; b < src->nbuckets; b++) {
1227                         copy_bucket_field(gen);
1228                         copy_bucket_field(data_type);
1229                         copy_bucket_field(stripe);
1230                         copy_bucket_field(dirty_sectors);
1231                         copy_bucket_field(cached_sectors);
1232
1233                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
1234                 }
1235
1236                 {
1237                         struct bch_dev_usage *dst = ca->usage_base;
1238                         struct bch_dev_usage *src = (void *)
1239                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1240                                                      dev_usage_u64s());
1241
1242                         copy_dev_field(buckets_ec,              "buckets_ec");
1243                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1244
1245                         for (i = 0; i < BCH_DATA_NR; i++) {
1246                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1247                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1248                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1249                         }
1250                 }
1251         };
1252
1253         {
1254                 unsigned nr = fs_usage_u64s(c);
1255                 struct bch_fs_usage *dst = c->usage_base;
1256                 struct bch_fs_usage *src = (void *)
1257                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1258
1259                 copy_fs_field(hidden,           "hidden");
1260                 copy_fs_field(btree,            "btree");
1261
1262                 if (!metadata_only) {
1263                         copy_fs_field(data,     "data");
1264                         copy_fs_field(cached,   "cached");
1265                         copy_fs_field(reserved, "reserved");
1266                         copy_fs_field(nr_inodes,"nr_inodes");
1267
1268                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1269                                 copy_fs_field(persistent_reserved[i],
1270                                               "persistent_reserved[%i]", i);
1271                 }
1272
1273                 for (i = 0; i < c->replicas.nr; i++) {
1274                         struct bch_replicas_entry *e =
1275                                 cpu_replicas_entry(&c->replicas, i);
1276                         char buf[80];
1277
1278                         if (metadata_only &&
1279                             (e->data_type == BCH_DATA_user ||
1280                              e->data_type == BCH_DATA_cached))
1281                                 continue;
1282
1283                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1284
1285                         copy_fs_field(replicas[i], "%s", buf);
1286                 }
1287         }
1288
1289 #undef copy_fs_field
1290 #undef copy_dev_field
1291 #undef copy_bucket_field
1292 #undef copy_stripe_field
1293 #undef copy_field
1294 fsck_err:
1295         if (ca)
1296                 percpu_ref_put(&ca->ref);
1297         if (ret)
1298                 bch_err(c, "%s: ret %i", __func__, ret);
1299         return ret;
1300 }
1301
1302 static int bch2_gc_start(struct bch_fs *c,
1303                          bool metadata_only)
1304 {
1305         struct bch_dev *ca = NULL;
1306         unsigned i;
1307         int ret;
1308
1309         BUG_ON(c->usage_gc);
1310
1311         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1312                                          sizeof(u64), GFP_KERNEL);
1313         if (!c->usage_gc) {
1314                 bch_err(c, "error allocating c->usage_gc");
1315                 return -ENOMEM;
1316         }
1317
1318         for_each_member_device(ca, c, i) {
1319                 BUG_ON(ca->buckets[1]);
1320                 BUG_ON(ca->usage_gc);
1321
1322                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1323                                 ca->mi.nbuckets * sizeof(struct bucket),
1324                                 GFP_KERNEL|__GFP_ZERO);
1325                 if (!ca->buckets[1]) {
1326                         percpu_ref_put(&ca->ref);
1327                         bch_err(c, "error allocating ca->buckets[gc]");
1328                         return -ENOMEM;
1329                 }
1330
1331                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1332                 if (!ca->usage_gc) {
1333                         bch_err(c, "error allocating ca->usage_gc");
1334                         percpu_ref_put(&ca->ref);
1335                         return -ENOMEM;
1336                 }
1337         }
1338
1339         ret = bch2_ec_mem_alloc(c, true);
1340         if (ret) {
1341                 bch_err(c, "error allocating ec gc mem");
1342                 return ret;
1343         }
1344
1345         percpu_down_write(&c->mark_lock);
1346
1347         /*
1348          * indicate to stripe code that we need to allocate for the gc stripes
1349          * radix tree, too
1350          */
1351         gc_pos_set(c, gc_phase(GC_PHASE_START));
1352
1353         for_each_member_device(ca, c, i) {
1354                 struct bucket_array *dst = __bucket_array(ca, 1);
1355                 struct bucket_array *src = __bucket_array(ca, 0);
1356                 size_t b;
1357
1358                 dst->first_bucket       = src->first_bucket;
1359                 dst->nbuckets           = src->nbuckets;
1360
1361                 for (b = 0; b < src->nbuckets; b++) {
1362                         struct bucket *d = &dst->b[b];
1363                         struct bucket *s = &src->b[b];
1364
1365                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1366                         d->gen_valid = s->gen_valid;
1367
1368                         if (metadata_only &&
1369                             (s->mark.data_type == BCH_DATA_user ||
1370                              s->mark.data_type == BCH_DATA_cached))
1371                                 d->_mark = s->mark;
1372                 }
1373         };
1374
1375         percpu_up_write(&c->mark_lock);
1376
1377         return 0;
1378 }
1379
1380 static int bch2_gc_reflink_done_initial_fn(struct btree_trans *trans,
1381                                            struct bkey_s_c k)
1382 {
1383         struct bch_fs *c = trans->c;
1384         struct reflink_gc *r;
1385         const __le64 *refcount = bkey_refcount_c(k);
1386         char buf[200];
1387         int ret = 0;
1388
1389         if (!refcount)
1390                 return 0;
1391
1392         r = genradix_ptr(&c->reflink_gc_table, c->reflink_gc_idx++);
1393         if (!r)
1394                 return -ENOMEM;
1395
1396         if (!r ||
1397             r->offset != k.k->p.offset ||
1398             r->size != k.k->size) {
1399                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1400                 return -EINVAL;
1401         }
1402
1403         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1404                         "reflink key has wrong refcount:\n"
1405                         "  %s\n"
1406                         "  should be %u",
1407                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1408                         r->refcount)) {
1409                 struct bkey_i *new;
1410
1411                 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1412                 if (!new) {
1413                         ret = -ENOMEM;
1414                         goto fsck_err;
1415                 }
1416
1417                 bkey_reassemble(new, k);
1418
1419                 if (!r->refcount) {
1420                         new->k.type = KEY_TYPE_deleted;
1421                         new->k.size = 0;
1422                 } else {
1423                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1424                 }
1425
1426                 ret = bch2_journal_key_insert(c, BTREE_ID_reflink, 0, new);
1427                 if (ret)
1428                         kfree(new);
1429         }
1430 fsck_err:
1431         return ret;
1432 }
1433
1434 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1435                                 bool metadata_only)
1436 {
1437         struct btree_trans trans;
1438         struct btree_iter iter;
1439         struct bkey_s_c k;
1440         struct reflink_gc *r;
1441         size_t idx = 0;
1442         char buf[200];
1443         int ret = 0;
1444
1445         if (metadata_only)
1446                 return 0;
1447
1448         bch2_trans_init(&trans, c, 0, 0);
1449
1450         if (initial) {
1451                 c->reflink_gc_idx = 0;
1452
1453                 ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_reflink,
1454                                 bch2_gc_reflink_done_initial_fn);
1455                 goto out;
1456         }
1457
1458         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1459                            BTREE_ITER_PREFETCH, k, ret) {
1460                 const __le64 *refcount = bkey_refcount_c(k);
1461
1462                 if (!refcount)
1463                         continue;
1464
1465                 r = genradix_ptr(&c->reflink_gc_table, idx);
1466                 if (!r ||
1467                     r->offset != k.k->p.offset ||
1468                     r->size != k.k->size) {
1469                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1470                         ret = -EINVAL;
1471                         break;
1472                 }
1473
1474                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1475                                 "reflink key has wrong refcount:\n"
1476                                 "  %s\n"
1477                                 "  should be %u",
1478                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1479                                 r->refcount)) {
1480                         struct bkey_i *new;
1481
1482                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1483                         if (!new) {
1484                                 ret = -ENOMEM;
1485                                 break;
1486                         }
1487
1488                         bkey_reassemble(new, k);
1489
1490                         if (!r->refcount)
1491                                 new->k.type = KEY_TYPE_deleted;
1492                         else
1493                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1494
1495                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1496                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1497                         kfree(new);
1498
1499                         if (ret)
1500                                 break;
1501                 }
1502         }
1503 fsck_err:
1504         bch2_trans_iter_exit(&trans, &iter);
1505 out:
1506         genradix_free(&c->reflink_gc_table);
1507         c->reflink_gc_nr = 0;
1508         bch2_trans_exit(&trans);
1509         return ret;
1510 }
1511
1512 static int bch2_gc_reflink_start_initial_fn(struct btree_trans *trans,
1513                                             struct bkey_s_c k)
1514 {
1515
1516         struct bch_fs *c = trans->c;
1517         struct reflink_gc *r;
1518         const __le64 *refcount = bkey_refcount_c(k);
1519
1520         if (!refcount)
1521                 return 0;
1522
1523         r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1524                                GFP_KERNEL);
1525         if (!r)
1526                 return -ENOMEM;
1527
1528         r->offset       = k.k->p.offset;
1529         r->size         = k.k->size;
1530         r->refcount     = 0;
1531         return 0;
1532 }
1533
1534 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1535                                  bool metadata_only)
1536 {
1537         struct btree_trans trans;
1538         struct btree_iter iter;
1539         struct bkey_s_c k;
1540         struct reflink_gc *r;
1541         int ret = 0;
1542
1543         if (metadata_only)
1544                 return 0;
1545
1546         bch2_trans_init(&trans, c, 0, 0);
1547         genradix_free(&c->reflink_gc_table);
1548         c->reflink_gc_nr = 0;
1549
1550         if (initial) {
1551                 ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_reflink,
1552                                                 bch2_gc_reflink_start_initial_fn);
1553                 goto out;
1554         }
1555
1556         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1557                            BTREE_ITER_PREFETCH, k, ret) {
1558                 const __le64 *refcount = bkey_refcount_c(k);
1559
1560                 if (!refcount)
1561                         continue;
1562
1563                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1564                                        GFP_KERNEL);
1565                 if (!r) {
1566                         ret = -ENOMEM;
1567                         break;
1568                 }
1569
1570                 r->offset       = k.k->p.offset;
1571                 r->size         = k.k->size;
1572                 r->refcount     = 0;
1573         }
1574         bch2_trans_iter_exit(&trans, &iter);
1575 out:
1576         bch2_trans_exit(&trans);
1577         return ret;
1578 }
1579
1580 /**
1581  * bch2_gc - walk _all_ references to buckets, and recompute them:
1582  *
1583  * Order matters here:
1584  *  - Concurrent GC relies on the fact that we have a total ordering for
1585  *    everything that GC walks - see  gc_will_visit_node(),
1586  *    gc_will_visit_root()
1587  *
1588  *  - also, references move around in the course of index updates and
1589  *    various other crap: everything needs to agree on the ordering
1590  *    references are allowed to move around in - e.g., we're allowed to
1591  *    start with a reference owned by an open_bucket (the allocator) and
1592  *    move it to the btree, but not the reverse.
1593  *
1594  *    This is necessary to ensure that gc doesn't miss references that
1595  *    move around - if references move backwards in the ordering GC
1596  *    uses, GC could skip past them
1597  */
1598 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1599 {
1600         struct bch_dev *ca;
1601         u64 start_time = local_clock();
1602         unsigned i, iter = 0;
1603         int ret;
1604
1605         lockdep_assert_held(&c->state_lock);
1606         trace_gc_start(c);
1607
1608         down_write(&c->gc_lock);
1609
1610         /* flush interior btree updates: */
1611         closure_wait_event(&c->btree_interior_update_wait,
1612                            !bch2_btree_interior_updates_nr_pending(c));
1613 again:
1614         ret   = bch2_gc_start(c, metadata_only) ?:
1615                 bch2_gc_reflink_start(c, initial, metadata_only);
1616         if (ret)
1617                 goto out;
1618
1619         bch2_mark_superblocks(c);
1620
1621         if (test_bit(BCH_FS_TOPOLOGY_ERROR, &c->flags) &&
1622             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1623             c->opts.fix_errors != FSCK_OPT_NO) {
1624                 bch_info(c, "starting topology repair pass");
1625                 ret = bch2_repair_topology(c);
1626                 if (ret)
1627                         goto out;
1628                 bch_info(c, "topology repair pass done");
1629
1630                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1631         }
1632
1633         ret = bch2_gc_btrees(c, initial, metadata_only);
1634
1635         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1636             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1637             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1638                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1639                 ret = 0;
1640         }
1641
1642         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1643                 ret = FSCK_ERR_EXIT;
1644
1645         if (ret)
1646                 goto out;
1647
1648 #if 0
1649         bch2_mark_pending_btree_node_frees(c);
1650 #endif
1651         c->gc_count++;
1652
1653         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1654             (!iter && bch2_test_restart_gc)) {
1655                 /*
1656                  * XXX: make sure gens we fixed got saved
1657                  */
1658                 if (iter++ <= 2) {
1659                         bch_info(c, "Second GC pass needed, restarting:");
1660                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1661                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1662
1663                         percpu_down_write(&c->mark_lock);
1664                         bch2_gc_free(c);
1665                         percpu_up_write(&c->mark_lock);
1666                         /* flush fsck errors, reset counters */
1667                         bch2_flush_fsck_errs(c);
1668
1669                         goto again;
1670                 }
1671
1672                 bch_info(c, "Unable to fix bucket gens, looping");
1673                 ret = -EINVAL;
1674         }
1675 out:
1676         if (!ret) {
1677                 bch2_journal_block(&c->journal);
1678
1679                 percpu_down_write(&c->mark_lock);
1680                 ret   = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1681                         bch2_gc_done(c, initial, metadata_only);
1682
1683                 bch2_journal_unblock(&c->journal);
1684         } else {
1685                 percpu_down_write(&c->mark_lock);
1686         }
1687
1688         /* Indicates that gc is no longer in progress: */
1689         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1690
1691         bch2_gc_free(c);
1692         percpu_up_write(&c->mark_lock);
1693
1694         up_write(&c->gc_lock);
1695
1696         trace_gc_end(c);
1697         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1698
1699         /*
1700          * Wake up allocator in case it was waiting for buckets
1701          * because of not being able to inc gens
1702          */
1703         for_each_member_device(ca, c, i)
1704                 bch2_wake_allocator(ca);
1705
1706         /*
1707          * At startup, allocations can happen directly instead of via the
1708          * allocator thread - issue wakeup in case they blocked on gc_lock:
1709          */
1710         closure_wake_up(&c->freelist_wait);
1711         return ret;
1712 }
1713
1714 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1715 {
1716         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1717         const struct bch_extent_ptr *ptr;
1718
1719         percpu_down_read(&c->mark_lock);
1720         bkey_for_each_ptr(ptrs, ptr) {
1721                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1722                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1723
1724                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1725                         percpu_up_read(&c->mark_lock);
1726                         return true;
1727                 }
1728         }
1729
1730         bkey_for_each_ptr(ptrs, ptr) {
1731                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1732                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1733
1734                 if (gen_after(g->gc_gen, ptr->gen))
1735                         g->gc_gen = ptr->gen;
1736         }
1737         percpu_up_read(&c->mark_lock);
1738
1739         return false;
1740 }
1741
1742 /*
1743  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1744  * node pointers currently never have cached pointers that can become stale:
1745  */
1746 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1747 {
1748         struct btree_trans trans;
1749         struct btree_iter iter;
1750         struct bkey_s_c k;
1751         struct bkey_buf sk;
1752         int ret = 0, commit_err = 0;
1753
1754         bch2_bkey_buf_init(&sk);
1755         bch2_trans_init(&trans, c, 0, 0);
1756
1757         bch2_trans_iter_init(&trans, &iter, btree_id, POS_MIN,
1758                              BTREE_ITER_PREFETCH|
1759                              BTREE_ITER_NOT_EXTENTS|
1760                              BTREE_ITER_ALL_SNAPSHOTS);
1761
1762         while ((bch2_trans_begin(&trans),
1763                 k = bch2_btree_iter_peek(&iter)).k) {
1764                 ret = bkey_err(k);
1765
1766                 if (ret == -EINTR)
1767                         continue;
1768                 if (ret)
1769                         break;
1770
1771                 c->gc_gens_pos = iter.pos;
1772
1773                 if (gc_btree_gens_key(c, k) && !commit_err) {
1774                         bch2_bkey_buf_reassemble(&sk, c, k);
1775                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1776
1777                         commit_err =
1778                                 bch2_trans_update(&trans, &iter, sk.k, 0) ?:
1779                                 bch2_trans_commit(&trans, NULL, NULL,
1780                                                        BTREE_INSERT_NOWAIT|
1781                                                        BTREE_INSERT_NOFAIL);
1782                         if (commit_err == -EINTR) {
1783                                 commit_err = 0;
1784                                 continue;
1785                         }
1786                 }
1787
1788                 bch2_btree_iter_advance(&iter);
1789         }
1790         bch2_trans_iter_exit(&trans, &iter);
1791
1792         bch2_trans_exit(&trans);
1793         bch2_bkey_buf_exit(&sk, c);
1794
1795         return ret;
1796 }
1797
1798 int bch2_gc_gens(struct bch_fs *c)
1799 {
1800         struct bch_dev *ca;
1801         struct bucket_array *buckets;
1802         struct bucket *g;
1803         unsigned i;
1804         int ret;
1805
1806         /*
1807          * Ideally we would be using state_lock and not gc_lock here, but that
1808          * introduces a deadlock in the RO path - we currently take the state
1809          * lock at the start of going RO, thus the gc thread may get stuck:
1810          */
1811         down_read(&c->gc_lock);
1812
1813         for_each_member_device(ca, c, i) {
1814                 down_read(&ca->bucket_lock);
1815                 buckets = bucket_array(ca);
1816
1817                 for_each_bucket(g, buckets)
1818                         g->gc_gen = g->mark.gen;
1819                 up_read(&ca->bucket_lock);
1820         }
1821
1822         for (i = 0; i < BTREE_ID_NR; i++)
1823                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1824                         c->gc_gens_btree = i;
1825                         c->gc_gens_pos = POS_MIN;
1826                         ret = bch2_gc_btree_gens(c, i);
1827                         if (ret) {
1828                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1829                                 goto err;
1830                         }
1831                 }
1832
1833         for_each_member_device(ca, c, i) {
1834                 down_read(&ca->bucket_lock);
1835                 buckets = bucket_array(ca);
1836
1837                 for_each_bucket(g, buckets)
1838                         g->oldest_gen = g->gc_gen;
1839                 up_read(&ca->bucket_lock);
1840         }
1841
1842         c->gc_gens_btree        = 0;
1843         c->gc_gens_pos          = POS_MIN;
1844
1845         c->gc_count++;
1846 err:
1847         up_read(&c->gc_lock);
1848         return ret;
1849 }
1850
1851 static int bch2_gc_thread(void *arg)
1852 {
1853         struct bch_fs *c = arg;
1854         struct io_clock *clock = &c->io_clock[WRITE];
1855         unsigned long last = atomic64_read(&clock->now);
1856         unsigned last_kick = atomic_read(&c->kick_gc);
1857         int ret;
1858
1859         set_freezable();
1860
1861         while (1) {
1862                 while (1) {
1863                         set_current_state(TASK_INTERRUPTIBLE);
1864
1865                         if (kthread_should_stop()) {
1866                                 __set_current_state(TASK_RUNNING);
1867                                 return 0;
1868                         }
1869
1870                         if (atomic_read(&c->kick_gc) != last_kick)
1871                                 break;
1872
1873                         if (c->btree_gc_periodic) {
1874                                 unsigned long next = last + c->capacity / 16;
1875
1876                                 if (atomic64_read(&clock->now) >= next)
1877                                         break;
1878
1879                                 bch2_io_clock_schedule_timeout(clock, next);
1880                         } else {
1881                                 schedule();
1882                         }
1883
1884                         try_to_freeze();
1885                 }
1886                 __set_current_state(TASK_RUNNING);
1887
1888                 last = atomic64_read(&clock->now);
1889                 last_kick = atomic_read(&c->kick_gc);
1890
1891                 /*
1892                  * Full gc is currently incompatible with btree key cache:
1893                  */
1894 #if 0
1895                 ret = bch2_gc(c, false, false);
1896 #else
1897                 ret = bch2_gc_gens(c);
1898 #endif
1899                 if (ret < 0)
1900                         bch_err(c, "btree gc failed: %i", ret);
1901
1902                 debug_check_no_locks_held();
1903         }
1904
1905         return 0;
1906 }
1907
1908 void bch2_gc_thread_stop(struct bch_fs *c)
1909 {
1910         struct task_struct *p;
1911
1912         p = c->gc_thread;
1913         c->gc_thread = NULL;
1914
1915         if (p) {
1916                 kthread_stop(p);
1917                 put_task_struct(p);
1918         }
1919 }
1920
1921 int bch2_gc_thread_start(struct bch_fs *c)
1922 {
1923         struct task_struct *p;
1924
1925         if (c->gc_thread)
1926                 return 0;
1927
1928         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1929         if (IS_ERR(p)) {
1930                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1931                 return PTR_ERR(p);
1932         }
1933
1934         get_task_struct(p);
1935         c->gc_thread = p;
1936         wake_up_process(p);
1937         return 0;
1938 }