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->mark.data_type &&
 508                                 g->mark.data_type != data_type, c,
 509                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 510                                 "while marking %s",
 511                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 512                                 bch2_data_types[g->mark.data_type],
 513                                 bch2_data_types[data_type],
 514                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 515                         if (data_type == BCH_DATA_btree) {
 516                                 g2->_mark.data_type = g->_mark.data_type = data_type;
 517                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 518                         } else {
 519                                 do_update = true;
 520                         }
 521                 }
 522
 523                 if (fsck_err_on(!g->gen_valid, c,
 524                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 525                                 "while marking %s",
 526                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 527                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 528                                 p.ptr.gen,
 529                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 530                         if (!p.ptr.cached) {
 531                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 532                                 g2->gen_valid   = g->gen_valid          = true;
 533                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 534                         } else {
 535                                 do_update = true;
 536                         }
 537                 }
 538
 539                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 540                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 541                                 "while marking %s",
 542                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 543                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 544                                 p.ptr.gen, g->mark.gen,
 545                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 546                         if (!p.ptr.cached) {
 547                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 548                                 g2->gen_valid   = g->gen_valid  = true;
 549                                 g2->_mark.data_type             = 0;
 550                                 g2->_mark.dirty_sectors         = 0;
 551                                 g2->_mark.cached_sectors        = 0;
 552                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 553                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 554                         } else {
 555                                 do_update = true;
 556                         }
 557                 }
 558
 559                 if (fsck_err_on(!p.ptr.cached &&
 560                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 561                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 562                                 "while marking %s",
 563                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 564                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 565                                 p.ptr.gen, g->mark.gen,
 566                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 567                         do_update = true;
 568
 569                 if (p.has_ec) {
 570                         struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
 571
 572                         if (fsck_err_on(!m || !m->alive, c,
 573                                         "pointer to nonexistent stripe %llu\n"
 574                                         "while marking %s",
 575                                         (u64) p.ec.idx,
 576                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 577                                 do_update = true;
 578
 579                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 580                                         "pointer does not match stripe %llu\n"
 581                                         "while marking %s",
 582                                         (u64) p.ec.idx,
 583                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 584                                 do_update = true;
 585                 }
 586         }
 587
 588         if (do_update) {
 589                 struct bkey_ptrs ptrs;
 590                 union bch_extent_entry *entry;
 591                 struct bch_extent_ptr *ptr;
 592                 struct bkey_i *new;
 593
 594                 if (is_root) {
 595                         bch_err(c, "cannot update btree roots yet");
 596                         return -EINVAL;
 597                 }
 598
 599                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 600                 if (!new) {
 601                         bch_err(c, "%s: error allocating new key", __func__);
 602                         return -ENOMEM;
 603                 }
 604
 605                 bkey_reassemble(new, *k);
 606
 607                 if (level) {
 608                         /*
 609                          * We don't want to drop btree node pointers - if the
 610                          * btree node isn't there anymore, the read path will
 611                          * sort it out:
 612                          */
 613                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 614                         bkey_for_each_ptr(ptrs, ptr) {
 615                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 616                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 617
 618                                 ptr->gen = g->mark.gen;
 619                         }
 620                 } else {
 621                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 622                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 623                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 624                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 625
 626                                 (ptr->cached &&
 627                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 628                                 (!ptr->cached &&
 629                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 630                                 (g->mark.data_type &&
 631                                  g->mark.data_type != data_type);
 632                         }));
 633 again:
 634                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 635                         bkey_extent_entry_for_each(ptrs, entry) {
 636                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 637                                         struct stripe *m = genradix_ptr(&c->stripes[true],
 638                                                                         entry->stripe_ptr.idx);
 639                                         union bch_extent_entry *next_ptr;
 640
 641                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 642                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 643                                                         goto found;
 644                                         next_ptr = NULL;
 645 found:
 646                                         if (!next_ptr) {
 647                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 648                                                 continue;
 649                                         }
 650
 651                                         if (!m || !m->alive ||
 652                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 653                                                                        &next_ptr->ptr,
 654                                                                        m->sectors)) {
 655                                                 bch2_bkey_extent_entry_drop(new, entry);
 656                                                 goto again;
 657                                         }
 658                                 }
 659                         }
 660                 }
 661
 662                 ret = bch2_journal_key_insert(c, btree_id, level, new);
 663                 if (ret)
 664                         kfree(new);
 665                 else
 666                         *k = bkey_i_to_s_c(new);
 667         }
 668 fsck_err:
 669         return ret;
 670 }
 671
 672 /* marking of btree keys/nodes: */
 673
 674 static int bch2_gc_mark_key(struct bch_fs *c, enum btree_id btree_id,
 675                             unsigned level, bool is_root,
 676                             struct bkey_s_c *k,
 677                             u8 *max_stale, bool initial)
 678 {
 679         struct bkey_ptrs_c ptrs;
 680         const struct bch_extent_ptr *ptr;
 681         unsigned flags =
 682                 BTREE_TRIGGER_INSERT|
 683                 BTREE_TRIGGER_GC|
 684                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 685         int ret = 0;
 686
 687         if (initial) {
 688                 BUG_ON(bch2_journal_seq_verify &&
 689                        k->k->version.lo > journal_cur_seq(&c->journal));
 690
 691                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 692                 if (ret)
 693                         goto err;
 694
 695                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 696                                 "key version number higher than recorded: %llu > %llu",
 697                                 k->k->version.lo,
 698                                 atomic64_read(&c->key_version)))
 699                         atomic64_set(&c->key_version, k->k->version.lo);
 700
 701                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 702                     fsck_err_on(!bch2_bkey_replicas_marked(c, *k), c,
 703                                 "superblock not marked as containing replicas (type %u)",
 704                                 k->k->type)) {
 705                         ret = bch2_mark_bkey_replicas(c, *k);
 706                         if (ret) {
 707                                 bch_err(c, "error marking bkey replicas: %i", ret);
 708                                 goto err;
 709                         }
 710                 }
 711         }
 712
 713         ptrs = bch2_bkey_ptrs_c(*k);
 714         bkey_for_each_ptr(ptrs, ptr) {
 715                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 716                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 717
 718                 if (gen_after(g->oldest_gen, ptr->gen))
 719                         g->oldest_gen = ptr->gen;
 720
 721                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 722         }
 723
 724         bch2_mark_key(c, *k, flags);
 725 fsck_err:
 726 err:
 727         if (ret)
 728                 bch_err(c, "%s: ret %i", __func__, ret);
 729         return ret;
 730 }
 731
 732 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
 733                               bool initial)
 734 {
 735         struct btree_node_iter iter;
 736         struct bkey unpacked;
 737         struct bkey_s_c k;
 738         struct bkey_buf prev, cur;
 739         int ret = 0;
 740
 741         *max_stale = 0;
 742
 743         if (!btree_node_type_needs_gc(btree_node_type(b)))
 744                 return 0;
 745
 746         bch2_btree_node_iter_init_from_start(&iter, b);
 747         bch2_bkey_buf_init(&prev);
 748         bch2_bkey_buf_init(&cur);
 749         bkey_init(&prev.k->k);
 750
 751         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 752                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 753                                        &k, max_stale, initial);
 754                 if (ret)
 755                         break;
 756
 757                 bch2_btree_node_iter_advance(&iter, b);
 758
 759                 if (b->c.level) {
 760                         bch2_bkey_buf_reassemble(&cur, c, k);
 761
 762                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 763                                         bch2_btree_node_iter_end(&iter));
 764                         if (ret)
 765                                 break;
 766                 }
 767         }
 768
 769         bch2_bkey_buf_exit(&cur, c);
 770         bch2_bkey_buf_exit(&prev, c);
 771         return ret;
 772 }
 773
 774 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 775                          bool initial, bool metadata_only)
 776 {
 777         struct btree_trans trans;
 778         struct btree_iter iter;
 779         struct btree *b;
 780         unsigned depth = metadata_only                  ? 1
 781                 : bch2_expensive_debug_checks           ? 0
 782                 : !btree_node_type_needs_gc(btree_id)   ? 1
 783                 : 0;
 784         u8 max_stale = 0;
 785         int ret = 0;
 786
 787         bch2_trans_init(&trans, c, 0, 0);
 788
 789         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 790
 791         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 792                               0, depth, BTREE_ITER_PREFETCH, b) {
 793                 bch2_verify_btree_nr_keys(b);
 794
 795                 gc_pos_set(c, gc_pos_btree_node(b));
 796
 797                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 798                 if (ret)
 799                         break;
 800
 801                 if (!initial) {
 802                         if (max_stale > 64)
 803                                 bch2_btree_node_rewrite(&trans, &iter,
 804                                                 b->data->keys.seq,
 805                                                 BTREE_INSERT_NOWAIT|
 806                                                 BTREE_INSERT_GC_LOCK_HELD);
 807                         else if (!bch2_btree_gc_rewrite_disabled &&
 808                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 809                                 bch2_btree_node_rewrite(&trans, &iter,
 810                                                 b->data->keys.seq,
 811                                                 BTREE_INSERT_NOWAIT|
 812                                                 BTREE_INSERT_GC_LOCK_HELD);
 813                 }
 814
 815                 bch2_trans_cond_resched(&trans);
 816         }
 817         bch2_trans_iter_exit(&trans, &iter);
 818
 819         ret = bch2_trans_exit(&trans) ?: ret;
 820         if (ret)
 821                 return ret;
 822
 823         mutex_lock(&c->btree_root_lock);
 824         b = c->btree_roots[btree_id].b;
 825         if (!btree_node_fake(b)) {
 826                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 827
 828                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 829                                        &k, &max_stale, initial);
 830         }
 831         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 832         mutex_unlock(&c->btree_root_lock);
 833
 834         return ret;
 835 }
 836
 837 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
 838                                       unsigned target_depth)
 839 {
 840         struct btree_and_journal_iter iter;
 841         struct bkey_s_c k;
 842         struct bkey_buf cur, prev;
 843         u8 max_stale = 0;
 844         char buf[200];
 845         int ret = 0;
 846
 847         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 848         bch2_bkey_buf_init(&prev);
 849         bch2_bkey_buf_init(&cur);
 850         bkey_init(&prev.k->k);
 851
 852         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 853                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 854                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 855
 856                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 857                                        &k, &max_stale, true);
 858                 if (ret) {
 859                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 860                         goto fsck_err;
 861                 }
 862
 863                 if (b->c.level) {
 864                         bch2_bkey_buf_reassemble(&cur, c, k);
 865                         k = bkey_i_to_s_c(cur.k);
 866
 867                         bch2_btree_and_journal_iter_advance(&iter);
 868
 869                         ret = bch2_gc_check_topology(c, b,
 870                                         &prev, cur,
 871                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 872                         if (ret)
 873                                 goto fsck_err;
 874                 } else {
 875                         bch2_btree_and_journal_iter_advance(&iter);
 876                 }
 877         }
 878
 879         if (b->c.level > target_depth) {
 880                 bch2_btree_and_journal_iter_exit(&iter);
 881                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 882
 883                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 884                         struct btree *child;
 885
 886                         bch2_bkey_buf_reassemble(&cur, c, k);
 887                         bch2_btree_and_journal_iter_advance(&iter);
 888
 889                         child = bch2_btree_node_get_noiter(c, cur.k,
 890                                                 b->c.btree_id, b->c.level - 1,
 891                                                 false);
 892                         ret = PTR_ERR_OR_ZERO(child);
 893
 894                         if (ret == -EIO) {
 895                                 bch2_topology_error(c);
 896
 897                                 if (__fsck_err(c,
 898                                           FSCK_CAN_FIX|
 899                                           FSCK_CAN_IGNORE|
 900                                           FSCK_NO_RATELIMIT,
 901                                           "Unreadable btree node at btree %s level %u:\n"
 902                                           "  %s",
 903                                           bch2_btree_ids[b->c.btree_id],
 904                                           b->c.level - 1,
 905                                           (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
 906                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 907                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 908                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 909                                         goto fsck_err;
 910                                 } else {
 911                                         /* Continue marking when opted to not
 912                                          * fix the error: */
 913                                         ret = 0;
 914                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 915                                         continue;
 916                                 }
 917                         } else if (ret) {
 918                                 bch_err(c, "%s: error %i getting btree node",
 919                                         __func__, ret);
 920                                 break;
 921                         }
 922
 923                         ret = bch2_gc_btree_init_recurse(c, child,
 924                                                          target_depth);
 925                         six_unlock_read(&child->c.lock);
 926
 927                         if (ret)
 928                                 break;
 929                 }
 930         }
 931 fsck_err:
 932         bch2_bkey_buf_exit(&cur, c);
 933         bch2_bkey_buf_exit(&prev, c);
 934         bch2_btree_and_journal_iter_exit(&iter);
 935         return ret;
 936 }
 937
 938 static int bch2_gc_btree_init(struct bch_fs *c,
 939                               enum btree_id btree_id,
 940                               bool metadata_only)
 941 {
 942         struct btree *b;
 943         unsigned target_depth = metadata_only           ? 1
 944                 : bch2_expensive_debug_checks           ? 0
 945                 : !btree_node_type_needs_gc(btree_id)   ? 1
 946                 : 0;
 947         u8 max_stale = 0;
 948         char buf[100];
 949         int ret = 0;
 950
 951         b = c->btree_roots[btree_id].b;
 952
 953         if (btree_node_fake(b))
 954                 return 0;
 955
 956         six_lock_read(&b->c.lock, NULL, NULL);
 957         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 958                         "btree root with incorrect min_key: %s",
 959                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 960                 bch_err(c, "repair unimplemented");
 961                 ret = FSCK_ERR_EXIT;
 962                 goto fsck_err;
 963         }
 964
 965         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
 966                         "btree root with incorrect max_key: %s",
 967                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 968                 bch_err(c, "repair unimplemented");
 969                 ret = FSCK_ERR_EXIT;
 970                 goto fsck_err;
 971         }
 972
 973         if (b->c.level >= target_depth)
 974                 ret = bch2_gc_btree_init_recurse(c, b, target_depth);
 975
 976         if (!ret) {
 977                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 978
 979                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 980                                        &k, &max_stale, true);
 981         }
 982 fsck_err:
 983         six_unlock_read(&b->c.lock);
 984
 985         if (ret < 0)
 986                 bch_err(c, "%s: ret %i", __func__, ret);
 987         return ret;
 988 }
 989
 990 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 991 {
 992         return  (int) btree_id_to_gc_phase(l) -
 993                 (int) btree_id_to_gc_phase(r);
 994 }
 995
 996 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
 997 {
 998         enum btree_id ids[BTREE_ID_NR];
 999         unsigned i;
1000         int ret = 0;
1001
1002         for (i = 0; i < BTREE_ID_NR; i++)
1003                 ids[i] = i;
1004         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1005
1006         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1007                 ret = initial
1008                         ? bch2_gc_btree_init(c, ids[i], metadata_only)
1009                         : bch2_gc_btree(c, ids[i], initial, metadata_only);
1010
1011         if (ret < 0)
1012                 bch_err(c, "%s: ret %i", __func__, ret);
1013         return ret;
1014 }
1015
1016 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1017                                   u64 start, u64 end,
1018                                   enum bch_data_type type,
1019                                   unsigned flags)
1020 {
1021         u64 b = sector_to_bucket(ca, start);
1022
1023         do {
1024                 unsigned sectors =
1025                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1026
1027                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1028                                           gc_phase(GC_PHASE_SB), flags);
1029                 b++;
1030                 start += sectors;
1031         } while (start < end);
1032 }
1033
1034 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1035                               unsigned flags)
1036 {
1037         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1038         unsigned i;
1039         u64 b;
1040
1041         /*
1042          * This conditional is kind of gross, but we may be called from the
1043          * device add path, before the new device has actually been added to the
1044          * running filesystem:
1045          */
1046         if (c) {
1047                 lockdep_assert_held(&c->sb_lock);
1048                 percpu_down_read(&c->mark_lock);
1049         }
1050
1051         for (i = 0; i < layout->nr_superblocks; i++) {
1052                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1053
1054                 if (offset == BCH_SB_SECTOR)
1055                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1056                                               BCH_DATA_sb, flags);
1057
1058                 mark_metadata_sectors(c, ca, offset,
1059                                       offset + (1 << layout->sb_max_size_bits),
1060                                       BCH_DATA_sb, flags);
1061         }
1062
1063         for (i = 0; i < ca->journal.nr; i++) {
1064                 b = ca->journal.buckets[i];
1065                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1066                                           ca->mi.bucket_size,
1067                                           gc_phase(GC_PHASE_SB), flags);
1068         }
1069
1070         if (c)
1071                 percpu_up_read(&c->mark_lock);
1072 }
1073
1074 static void bch2_mark_superblocks(struct bch_fs *c)
1075 {
1076         struct bch_dev *ca;
1077         unsigned i;
1078
1079         mutex_lock(&c->sb_lock);
1080         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1081
1082         for_each_online_member(ca, c, i)
1083                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1084         mutex_unlock(&c->sb_lock);
1085 }
1086
1087 #if 0
1088 /* Also see bch2_pending_btree_node_free_insert_done() */
1089 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1090 {
1091         struct btree_update *as;
1092         struct pending_btree_node_free *d;
1093
1094         mutex_lock(&c->btree_interior_update_lock);
1095         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1096
1097         for_each_pending_btree_node_free(c, as, d)
1098                 if (d->index_update_done)
1099                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
1100                                       BTREE_TRIGGER_INSERT|BTREE_TRIGGER_GC);
1101
1102         mutex_unlock(&c->btree_interior_update_lock);
1103 }
1104 #endif
1105
1106 static void bch2_gc_free(struct bch_fs *c)
1107 {
1108         struct bch_dev *ca;
1109         unsigned i;
1110
1111         genradix_free(&c->stripes[1]);
1112
1113         for_each_member_device(ca, c, i) {
1114                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1115                         sizeof(struct bucket_array) +
1116                         ca->mi.nbuckets * sizeof(struct bucket));
1117                 ca->buckets[1] = NULL;
1118
1119                 free_percpu(ca->usage_gc);
1120                 ca->usage_gc = NULL;
1121         }
1122
1123         free_percpu(c->usage_gc);
1124         c->usage_gc = NULL;
1125 }
1126
1127 static int bch2_gc_done(struct bch_fs *c,
1128                         bool initial, bool metadata_only)
1129 {
1130         struct bch_dev *ca = NULL;
1131         bool verify = !metadata_only && (!initial ||
1132                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1133         unsigned i, dev;
1134         int ret = 0;
1135
1136 #define copy_field(_f, _msg, ...)                                       \
1137         if (dst->_f != src->_f) {                                       \
1138                 if (verify)                                             \
1139                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1140                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1141                 dst->_f = src->_f;                                      \
1142                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1143         }
1144 #define copy_stripe_field(_f, _msg, ...)                                \
1145         if (dst->_f != src->_f) {                                       \
1146                 if (verify)                                             \
1147                         fsck_err(c, "stripe %zu has wrong "_msg         \
1148                                 ": got %u, should be %u",               \
1149                                 iter.pos, ##__VA_ARGS__,                \
1150                                 dst->_f, src->_f);                      \
1151                 dst->_f = src->_f;                                      \
1152                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1153         }
1154 #define copy_bucket_field(_f)                                           \
1155         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
1156                 if (verify)                                             \
1157                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
1158                                 ": got %u, should be %u", dev, b,       \
1159                                 dst->b[b].mark.gen,                     \
1160                                 bch2_data_types[dst->b[b].mark.data_type],\
1161                                 dst->b[b].mark._f, src->b[b].mark._f);  \
1162                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
1163                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1164         }
1165 #define copy_dev_field(_f, _msg, ...)                                   \
1166         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1167 #define copy_fs_field(_f, _msg, ...)                                    \
1168         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1169
1170         if (!metadata_only) {
1171                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
1172                 struct stripe *dst, *src;
1173
1174                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
1175                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
1176
1177                         if (dst->alive          != src->alive ||
1178                             dst->sectors        != src->sectors ||
1179                             dst->algorithm      != src->algorithm ||
1180                             dst->nr_blocks      != src->nr_blocks ||
1181                             dst->nr_redundant   != src->nr_redundant) {
1182                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
1183                                 ret = -EINVAL;
1184                                 goto fsck_err;
1185                         }
1186
1187                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
1188                                 copy_stripe_field(block_sectors[i],
1189                                                   "block_sectors[%u]", i);
1190
1191                         dst->blocks_nonempty = 0;
1192                         for (i = 0; i < dst->nr_blocks; i++)
1193                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
1194
1195                         genradix_iter_advance(&iter, &c->stripes[1]);
1196                 }
1197         }
1198
1199         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1200                 bch2_fs_usage_acc_to_base(c, i);
1201
1202         for_each_member_device(ca, c, dev) {
1203                 struct bucket_array *dst = __bucket_array(ca, 0);
1204                 struct bucket_array *src = __bucket_array(ca, 1);
1205                 size_t b;
1206
1207                 for (b = 0; b < src->nbuckets; b++) {
1208                         copy_bucket_field(gen);
1209                         copy_bucket_field(data_type);
1210                         copy_bucket_field(stripe);
1211                         copy_bucket_field(dirty_sectors);
1212                         copy_bucket_field(cached_sectors);
1213
1214                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
1215                 }
1216
1217                 {
1218                         struct bch_dev_usage *dst = ca->usage_base;
1219                         struct bch_dev_usage *src = (void *)
1220                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1221                                                      dev_usage_u64s());
1222
1223                         copy_dev_field(buckets_ec,              "buckets_ec");
1224                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1225
1226                         for (i = 0; i < BCH_DATA_NR; i++) {
1227                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1228                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1229                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1230                         }
1231                 }
1232         };
1233
1234         {
1235                 unsigned nr = fs_usage_u64s(c);
1236                 struct bch_fs_usage *dst = c->usage_base;
1237                 struct bch_fs_usage *src = (void *)
1238                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1239
1240                 copy_fs_field(hidden,           "hidden");
1241                 copy_fs_field(btree,            "btree");
1242
1243                 if (!metadata_only) {
1244                         copy_fs_field(data,     "data");
1245                         copy_fs_field(cached,   "cached");
1246                         copy_fs_field(reserved, "reserved");
1247                         copy_fs_field(nr_inodes,"nr_inodes");
1248
1249                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1250                                 copy_fs_field(persistent_reserved[i],
1251                                               "persistent_reserved[%i]", i);
1252                 }
1253
1254                 for (i = 0; i < c->replicas.nr; i++) {
1255                         struct bch_replicas_entry *e =
1256                                 cpu_replicas_entry(&c->replicas, i);
1257                         char buf[80];
1258
1259                         if (metadata_only &&
1260                             (e->data_type == BCH_DATA_user ||
1261                              e->data_type == BCH_DATA_cached))
1262                                 continue;
1263
1264                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1265
1266                         copy_fs_field(replicas[i], "%s", buf);
1267                 }
1268         }
1269
1270 #undef copy_fs_field
1271 #undef copy_dev_field
1272 #undef copy_bucket_field
1273 #undef copy_stripe_field
1274 #undef copy_field
1275 fsck_err:
1276         if (ca)
1277                 percpu_ref_put(&ca->ref);
1278         if (ret)
1279                 bch_err(c, "%s: ret %i", __func__, ret);
1280         return ret;
1281 }
1282
1283 static int bch2_gc_start(struct bch_fs *c,
1284                          bool metadata_only)
1285 {
1286         struct bch_dev *ca = NULL;
1287         unsigned i;
1288         int ret;
1289
1290         BUG_ON(c->usage_gc);
1291
1292         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1293                                          sizeof(u64), GFP_KERNEL);
1294         if (!c->usage_gc) {
1295                 bch_err(c, "error allocating c->usage_gc");
1296                 return -ENOMEM;
1297         }
1298
1299         for_each_member_device(ca, c, i) {
1300                 BUG_ON(ca->buckets[1]);
1301                 BUG_ON(ca->usage_gc);
1302
1303                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1304                                 ca->mi.nbuckets * sizeof(struct bucket),
1305                                 GFP_KERNEL|__GFP_ZERO);
1306                 if (!ca->buckets[1]) {
1307                         percpu_ref_put(&ca->ref);
1308                         bch_err(c, "error allocating ca->buckets[gc]");
1309                         return -ENOMEM;
1310                 }
1311
1312                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1313                 if (!ca->usage_gc) {
1314                         bch_err(c, "error allocating ca->usage_gc");
1315                         percpu_ref_put(&ca->ref);
1316                         return -ENOMEM;
1317                 }
1318         }
1319
1320         ret = bch2_ec_mem_alloc(c, true);
1321         if (ret) {
1322                 bch_err(c, "error allocating ec gc mem");
1323                 return ret;
1324         }
1325
1326         percpu_down_write(&c->mark_lock);
1327
1328         /*
1329          * indicate to stripe code that we need to allocate for the gc stripes
1330          * radix tree, too
1331          */
1332         gc_pos_set(c, gc_phase(GC_PHASE_START));
1333
1334         for_each_member_device(ca, c, i) {
1335                 struct bucket_array *dst = __bucket_array(ca, 1);
1336                 struct bucket_array *src = __bucket_array(ca, 0);
1337                 size_t b;
1338
1339                 dst->first_bucket       = src->first_bucket;
1340                 dst->nbuckets           = src->nbuckets;
1341
1342                 for (b = 0; b < src->nbuckets; b++) {
1343                         struct bucket *d = &dst->b[b];
1344                         struct bucket *s = &src->b[b];
1345
1346                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1347                         d->gen_valid = s->gen_valid;
1348
1349                         if (metadata_only &&
1350                             (s->mark.data_type == BCH_DATA_user ||
1351                              s->mark.data_type == BCH_DATA_cached))
1352                                 d->_mark = s->mark;
1353                 }
1354         };
1355
1356         percpu_up_write(&c->mark_lock);
1357
1358         return 0;
1359 }
1360
1361 static int bch2_gc_reflink_done_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1362 {
1363         struct reflink_gc *r;
1364         const __le64 *refcount = bkey_refcount_c(k);
1365         char buf[200];
1366         int ret = 0;
1367
1368         if (!refcount)
1369                 return 0;
1370
1371         r = genradix_ptr(&c->reflink_gc_table, c->reflink_gc_idx++);
1372         if (!r)
1373                 return -ENOMEM;
1374
1375         if (!r ||
1376             r->offset != k.k->p.offset ||
1377             r->size != k.k->size) {
1378                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1379                 return -EINVAL;
1380         }
1381
1382         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1383                         "reflink key has wrong refcount:\n"
1384                         "  %s\n"
1385                         "  should be %u",
1386                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1387                         r->refcount)) {
1388                 struct bkey_i *new;
1389
1390                 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1391                 if (!new) {
1392                         ret = -ENOMEM;
1393                         goto fsck_err;
1394                 }
1395
1396                 bkey_reassemble(new, k);
1397
1398                 if (!r->refcount) {
1399                         new->k.type = KEY_TYPE_deleted;
1400                         new->k.size = 0;
1401                 } else {
1402                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1403                 }
1404
1405                 ret = bch2_journal_key_insert(c, BTREE_ID_reflink, 0, new);
1406                 if (ret)
1407                         kfree(new);
1408         }
1409 fsck_err:
1410         return ret;
1411 }
1412
1413 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1414                                 bool metadata_only)
1415 {
1416         struct btree_trans trans;
1417         struct btree_iter iter;
1418         struct bkey_s_c k;
1419         struct reflink_gc *r;
1420         size_t idx = 0;
1421         char buf[200];
1422         int ret = 0;
1423
1424         if (metadata_only)
1425                 return 0;
1426
1427         if (initial) {
1428                 c->reflink_gc_idx = 0;
1429
1430                 ret = bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1431                                 bch2_gc_reflink_done_initial_fn);
1432                 goto out;
1433         }
1434
1435         bch2_trans_init(&trans, c, 0, 0);
1436
1437         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1438                            BTREE_ITER_PREFETCH, k, ret) {
1439                 const __le64 *refcount = bkey_refcount_c(k);
1440
1441                 if (!refcount)
1442                         continue;
1443
1444                 r = genradix_ptr(&c->reflink_gc_table, idx);
1445                 if (!r ||
1446                     r->offset != k.k->p.offset ||
1447                     r->size != k.k->size) {
1448                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1449                         ret = -EINVAL;
1450                         break;
1451                 }
1452
1453                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1454                                 "reflink key has wrong refcount:\n"
1455                                 "  %s\n"
1456                                 "  should be %u",
1457                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1458                                 r->refcount)) {
1459                         struct bkey_i *new;
1460
1461                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1462                         if (!new) {
1463                                 ret = -ENOMEM;
1464                                 break;
1465                         }
1466
1467                         bkey_reassemble(new, k);
1468
1469                         if (!r->refcount)
1470                                 new->k.type = KEY_TYPE_deleted;
1471                         else
1472                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1473
1474                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1475                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1476                         kfree(new);
1477
1478                         if (ret)
1479                                 break;
1480                 }
1481         }
1482 fsck_err:
1483         bch2_trans_iter_exit(&trans, &iter);
1484         bch2_trans_exit(&trans);
1485 out:
1486         genradix_free(&c->reflink_gc_table);
1487         c->reflink_gc_nr = 0;
1488         return ret;
1489 }
1490
1491 static int bch2_gc_reflink_start_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1492 {
1493
1494         struct reflink_gc *r;
1495         const __le64 *refcount = bkey_refcount_c(k);
1496
1497         if (!refcount)
1498                 return 0;
1499
1500         r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1501                                GFP_KERNEL);
1502         if (!r)
1503                 return -ENOMEM;
1504
1505         r->offset       = k.k->p.offset;
1506         r->size         = k.k->size;
1507         r->refcount     = 0;
1508         return 0;
1509 }
1510
1511 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1512                                  bool metadata_only)
1513 {
1514         struct btree_trans trans;
1515         struct btree_iter iter;
1516         struct bkey_s_c k;
1517         struct reflink_gc *r;
1518         int ret;
1519
1520         if (metadata_only)
1521                 return 0;
1522
1523         genradix_free(&c->reflink_gc_table);
1524         c->reflink_gc_nr = 0;
1525
1526         if (initial)
1527                 return bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1528                                 bch2_gc_reflink_start_initial_fn);
1529
1530         bch2_trans_init(&trans, c, 0, 0);
1531
1532         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1533                            BTREE_ITER_PREFETCH, k, ret) {
1534                 const __le64 *refcount = bkey_refcount_c(k);
1535
1536                 if (!refcount)
1537                         continue;
1538
1539                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1540                                        GFP_KERNEL);
1541                 if (!r) {
1542                         ret = -ENOMEM;
1543                         break;
1544                 }
1545
1546                 r->offset       = k.k->p.offset;
1547                 r->size         = k.k->size;
1548                 r->refcount     = 0;
1549         }
1550         bch2_trans_iter_exit(&trans, &iter);
1551
1552         bch2_trans_exit(&trans);
1553         return 0;
1554 }
1555
1556 /**
1557  * bch2_gc - walk _all_ references to buckets, and recompute them:
1558  *
1559  * Order matters here:
1560  *  - Concurrent GC relies on the fact that we have a total ordering for
1561  *    everything that GC walks - see  gc_will_visit_node(),
1562  *    gc_will_visit_root()
1563  *
1564  *  - also, references move around in the course of index updates and
1565  *    various other crap: everything needs to agree on the ordering
1566  *    references are allowed to move around in - e.g., we're allowed to
1567  *    start with a reference owned by an open_bucket (the allocator) and
1568  *    move it to the btree, but not the reverse.
1569  *
1570  *    This is necessary to ensure that gc doesn't miss references that
1571  *    move around - if references move backwards in the ordering GC
1572  *    uses, GC could skip past them
1573  */
1574 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1575 {
1576         struct bch_dev *ca;
1577         u64 start_time = local_clock();
1578         unsigned i, iter = 0;
1579         int ret;
1580
1581         lockdep_assert_held(&c->state_lock);
1582         trace_gc_start(c);
1583
1584         down_write(&c->gc_lock);
1585
1586         /* flush interior btree updates: */
1587         closure_wait_event(&c->btree_interior_update_wait,
1588                            !bch2_btree_interior_updates_nr_pending(c));
1589 again:
1590         ret   = bch2_gc_start(c, metadata_only) ?:
1591                 bch2_gc_reflink_start(c, initial, metadata_only);
1592         if (ret)
1593                 goto out;
1594
1595         bch2_mark_superblocks(c);
1596
1597         if (test_bit(BCH_FS_TOPOLOGY_ERROR, &c->flags) &&
1598             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1599             c->opts.fix_errors != FSCK_OPT_NO) {
1600                 bch_info(c, "starting topology repair pass");
1601                 ret = bch2_repair_topology(c);
1602                 if (ret)
1603                         goto out;
1604                 bch_info(c, "topology repair pass done");
1605
1606                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1607         }
1608
1609         ret = bch2_gc_btrees(c, initial, metadata_only);
1610
1611         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1612             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1613             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1614                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1615                 ret = 0;
1616         }
1617
1618         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1619                 ret = FSCK_ERR_EXIT;
1620
1621         if (ret)
1622                 goto out;
1623
1624 #if 0
1625         bch2_mark_pending_btree_node_frees(c);
1626 #endif
1627         c->gc_count++;
1628
1629         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1630             (!iter && bch2_test_restart_gc)) {
1631                 /*
1632                  * XXX: make sure gens we fixed got saved
1633                  */
1634                 if (iter++ <= 2) {
1635                         bch_info(c, "Second GC pass needed, restarting:");
1636                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1637                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1638
1639                         percpu_down_write(&c->mark_lock);
1640                         bch2_gc_free(c);
1641                         percpu_up_write(&c->mark_lock);
1642                         /* flush fsck errors, reset counters */
1643                         bch2_flush_fsck_errs(c);
1644
1645                         goto again;
1646                 }
1647
1648                 bch_info(c, "Unable to fix bucket gens, looping");
1649                 ret = -EINVAL;
1650         }
1651 out:
1652         if (!ret) {
1653                 bch2_journal_block(&c->journal);
1654
1655                 percpu_down_write(&c->mark_lock);
1656                 ret   = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1657                         bch2_gc_done(c, initial, metadata_only);
1658
1659                 bch2_journal_unblock(&c->journal);
1660         } else {
1661                 percpu_down_write(&c->mark_lock);
1662         }
1663
1664         /* Indicates that gc is no longer in progress: */
1665         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1666
1667         bch2_gc_free(c);
1668         percpu_up_write(&c->mark_lock);
1669
1670         up_write(&c->gc_lock);
1671
1672         trace_gc_end(c);
1673         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1674
1675         /*
1676          * Wake up allocator in case it was waiting for buckets
1677          * because of not being able to inc gens
1678          */
1679         for_each_member_device(ca, c, i)
1680                 bch2_wake_allocator(ca);
1681
1682         /*
1683          * At startup, allocations can happen directly instead of via the
1684          * allocator thread - issue wakeup in case they blocked on gc_lock:
1685          */
1686         closure_wake_up(&c->freelist_wait);
1687         return ret;
1688 }
1689
1690 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1691 {
1692         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1693         const struct bch_extent_ptr *ptr;
1694
1695         percpu_down_read(&c->mark_lock);
1696         bkey_for_each_ptr(ptrs, ptr) {
1697                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1698                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1699
1700                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1701                         percpu_up_read(&c->mark_lock);
1702                         return true;
1703                 }
1704         }
1705
1706         bkey_for_each_ptr(ptrs, ptr) {
1707                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1708                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1709
1710                 if (gen_after(g->gc_gen, ptr->gen))
1711                         g->gc_gen = ptr->gen;
1712         }
1713         percpu_up_read(&c->mark_lock);
1714
1715         return false;
1716 }
1717
1718 /*
1719  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1720  * node pointers currently never have cached pointers that can become stale:
1721  */
1722 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1723 {
1724         struct btree_trans trans;
1725         struct btree_iter iter;
1726         struct bkey_s_c k;
1727         struct bkey_buf sk;
1728         int ret = 0, commit_err = 0;
1729
1730         bch2_bkey_buf_init(&sk);
1731         bch2_trans_init(&trans, c, 0, 0);
1732
1733         bch2_trans_iter_init(&trans, &iter, btree_id, POS_MIN,
1734                              BTREE_ITER_PREFETCH|
1735                              BTREE_ITER_NOT_EXTENTS|
1736                              BTREE_ITER_ALL_SNAPSHOTS);
1737
1738         while ((bch2_trans_begin(&trans),
1739                 k = bch2_btree_iter_peek(&iter)).k) {
1740                 ret = bkey_err(k);
1741
1742                 if (ret == -EINTR)
1743                         continue;
1744                 if (ret)
1745                         break;
1746
1747                 c->gc_gens_pos = iter.pos;
1748
1749                 if (gc_btree_gens_key(c, k) && !commit_err) {
1750                         bch2_bkey_buf_reassemble(&sk, c, k);
1751                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1752
1753
1754                         commit_err =
1755                                 bch2_trans_update(&trans, &iter, sk.k, 0) ?:
1756                                 bch2_trans_commit(&trans, NULL, NULL,
1757                                                        BTREE_INSERT_NOWAIT|
1758                                                        BTREE_INSERT_NOFAIL);
1759                         if (commit_err == -EINTR) {
1760                                 commit_err = 0;
1761                                 continue;
1762                         }
1763                 }
1764
1765                 bch2_btree_iter_advance(&iter);
1766         }
1767         bch2_trans_iter_exit(&trans, &iter);
1768
1769         bch2_trans_exit(&trans);
1770         bch2_bkey_buf_exit(&sk, c);
1771
1772         return ret;
1773 }
1774
1775 int bch2_gc_gens(struct bch_fs *c)
1776 {
1777         struct bch_dev *ca;
1778         struct bucket_array *buckets;
1779         struct bucket *g;
1780         unsigned i;
1781         int ret;
1782
1783         /*
1784          * Ideally we would be using state_lock and not gc_lock here, but that
1785          * introduces a deadlock in the RO path - we currently take the state
1786          * lock at the start of going RO, thus the gc thread may get stuck:
1787          */
1788         down_read(&c->gc_lock);
1789
1790         for_each_member_device(ca, c, i) {
1791                 down_read(&ca->bucket_lock);
1792                 buckets = bucket_array(ca);
1793
1794                 for_each_bucket(g, buckets)
1795                         g->gc_gen = g->mark.gen;
1796                 up_read(&ca->bucket_lock);
1797         }
1798
1799         for (i = 0; i < BTREE_ID_NR; i++)
1800                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1801                         c->gc_gens_btree = i;
1802                         c->gc_gens_pos = POS_MIN;
1803                         ret = bch2_gc_btree_gens(c, i);
1804                         if (ret) {
1805                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1806                                 goto err;
1807                         }
1808                 }
1809
1810         for_each_member_device(ca, c, i) {
1811                 down_read(&ca->bucket_lock);
1812                 buckets = bucket_array(ca);
1813
1814                 for_each_bucket(g, buckets)
1815                         g->oldest_gen = g->gc_gen;
1816                 up_read(&ca->bucket_lock);
1817         }
1818
1819         c->gc_gens_btree        = 0;
1820         c->gc_gens_pos          = POS_MIN;
1821
1822         c->gc_count++;
1823 err:
1824         up_read(&c->gc_lock);
1825         return ret;
1826 }
1827
1828 static int bch2_gc_thread(void *arg)
1829 {
1830         struct bch_fs *c = arg;
1831         struct io_clock *clock = &c->io_clock[WRITE];
1832         unsigned long last = atomic64_read(&clock->now);
1833         unsigned last_kick = atomic_read(&c->kick_gc);
1834         int ret;
1835
1836         set_freezable();
1837
1838         while (1) {
1839                 while (1) {
1840                         set_current_state(TASK_INTERRUPTIBLE);
1841
1842                         if (kthread_should_stop()) {
1843                                 __set_current_state(TASK_RUNNING);
1844                                 return 0;
1845                         }
1846
1847                         if (atomic_read(&c->kick_gc) != last_kick)
1848                                 break;
1849
1850                         if (c->btree_gc_periodic) {
1851                                 unsigned long next = last + c->capacity / 16;
1852
1853                                 if (atomic64_read(&clock->now) >= next)
1854                                         break;
1855
1856                                 bch2_io_clock_schedule_timeout(clock, next);
1857                         } else {
1858                                 schedule();
1859                         }
1860
1861                         try_to_freeze();
1862                 }
1863                 __set_current_state(TASK_RUNNING);
1864
1865                 last = atomic64_read(&clock->now);
1866                 last_kick = atomic_read(&c->kick_gc);
1867
1868                 /*
1869                  * Full gc is currently incompatible with btree key cache:
1870                  */
1871 #if 0
1872                 ret = bch2_gc(c, false, false);
1873 #else
1874                 ret = bch2_gc_gens(c);
1875 #endif
1876                 if (ret < 0)
1877                         bch_err(c, "btree gc failed: %i", ret);
1878
1879                 debug_check_no_locks_held();
1880         }
1881
1882         return 0;
1883 }
1884
1885 void bch2_gc_thread_stop(struct bch_fs *c)
1886 {
1887         struct task_struct *p;
1888
1889         p = c->gc_thread;
1890         c->gc_thread = NULL;
1891
1892         if (p) {
1893                 kthread_stop(p);
1894                 put_task_struct(p);
1895         }
1896 }
1897
1898 int bch2_gc_thread_start(struct bch_fs *c)
1899 {
1900         struct task_struct *p;
1901
1902         if (c->gc_thread)
1903                 return 0;
1904
1905         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1906         if (IS_ERR(p)) {
1907                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1908                 return PTR_ERR(p);
1909         }
1910
1911         get_task_struct(p);
1912         c->gc_thread = p;
1913         wake_up_process(p);
1914         return 0;
1915 }