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_start(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         if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 237                         "btree node with incorrect min_key at btree %s level %u:\n"
 238                         "  prev %s\n"
 239                         "  cur %s",
 240                         bch2_btree_ids[b->c.btree_id], b->c.level,
 241                         buf1,
 242                         (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key)), buf2))) {
 243                 if (prev &&
 244                     bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 245                     BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 246                         if (bkey_cmp(prev->data->min_key,
 247                                      cur->data->min_key) <= 0)
 248                                 return DROP_PREV_NODE;
 249
 250                         ret = set_node_max(c, prev,
 251                                            bpos_predecessor(cur->data->min_key));
 252                 } else {
 253                         if (bkey_cmp(expected_start, b->data->max_key) >= 0)
 254                                 return DROP_THIS_NODE;
 255
 256                         ret = set_node_min(c, cur, expected_start);
 257                 }
 258                 if (ret)
 259                         return ret;
 260         }
 261 fsck_err:
 262         return ret;
 263 }
 264
 265 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 266                                  struct btree *child)
 267 {
 268         char buf1[200], buf2[200];
 269         int ret = 0;
 270
 271         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 272                         "btree node with incorrect max_key at btree %s level %u:\n"
 273                         "  %s\n"
 274                         "  expected %s",
 275                         bch2_btree_ids[b->c.btree_id], b->c.level,
 276                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
 277                         (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
 278                 ret = set_node_max(c, child, b->key.k.p);
 279                 if (ret)
 280                         return ret;
 281         }
 282 fsck_err:
 283         return ret;
 284 }
 285
 286 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 287 {
 288         struct btree_and_journal_iter iter;
 289         struct bkey_s_c k;
 290         struct bkey_buf prev_k, cur_k;
 291         struct btree *prev = NULL, *cur = NULL;
 292         bool have_child, dropped_children = false;
 293         char buf[200];
 294         int ret = 0;
 295
 296         if (!b->c.level)
 297                 return 0;
 298 again:
 299         prev = NULL;
 300         have_child = dropped_children = false;
 301         bch2_bkey_buf_init(&prev_k);
 302         bch2_bkey_buf_init(&cur_k);
 303         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 304
 305         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 306                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 307                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 308
 309                 bch2_btree_and_journal_iter_advance(&iter);
 310                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 311
 312                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 313                                         b->c.btree_id, b->c.level - 1,
 314                                         false);
 315                 ret = PTR_ERR_OR_ZERO(cur);
 316
 317                 if (mustfix_fsck_err_on(ret == -EIO, c,
 318                                 "Unreadable btree node at btree %s level %u:\n"
 319                                 "  %s",
 320                                 bch2_btree_ids[b->c.btree_id],
 321                                 b->c.level - 1,
 322                                 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
 323                         bch2_btree_node_evict(c, cur_k.k);
 324                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 325                                                       b->c.level, cur_k.k->k.p);
 326                         if (ret)
 327                                 break;
 328                         continue;
 329                 }
 330
 331                 if (ret) {
 332                         bch_err(c, "%s: error %i getting btree node",
 333                                 __func__, ret);
 334                         break;
 335                 }
 336
 337                 ret = btree_repair_node_start(c, b, prev, cur);
 338
 339                 if (ret == DROP_THIS_NODE) {
 340                         six_unlock_read(&cur->c.lock);
 341                         bch2_btree_node_evict(c, cur_k.k);
 342                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 343                                                       b->c.level, cur_k.k->k.p);
 344                         if (ret)
 345                                 break;
 346                         continue;
 347                 }
 348
 349                 if (prev)
 350                         six_unlock_read(&prev->c.lock);
 351                 prev = NULL;
 352
 353                 if (ret == DROP_PREV_NODE) {
 354                         bch2_btree_node_evict(c, prev_k.k);
 355                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 356                                                       b->c.level, prev_k.k->k.p);
 357                         if (ret)
 358                                 break;
 359
 360                         bch2_btree_and_journal_iter_exit(&iter);
 361                         bch2_bkey_buf_exit(&prev_k, c);
 362                         bch2_bkey_buf_exit(&cur_k, c);
 363                         goto again;
 364                 } else if (ret)
 365                         break;
 366
 367                 prev = cur;
 368                 cur = NULL;
 369                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 370         }
 371
 372         if (!ret && !IS_ERR_OR_NULL(prev)) {
 373                 BUG_ON(cur);
 374                 ret = btree_repair_node_end(c, b, prev);
 375         }
 376
 377         if (!IS_ERR_OR_NULL(prev))
 378                 six_unlock_read(&prev->c.lock);
 379         prev = NULL;
 380         if (!IS_ERR_OR_NULL(cur))
 381                 six_unlock_read(&cur->c.lock);
 382         cur = NULL;
 383
 384         if (ret)
 385                 goto err;
 386
 387         bch2_btree_and_journal_iter_exit(&iter);
 388         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 389
 390         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 391                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 392                 bch2_btree_and_journal_iter_advance(&iter);
 393
 394                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 395                                         b->c.btree_id, b->c.level - 1,
 396                                         false);
 397                 ret = PTR_ERR_OR_ZERO(cur);
 398
 399                 if (ret) {
 400                         bch_err(c, "%s: error %i getting btree node",
 401                                 __func__, ret);
 402                         goto err;
 403                 }
 404
 405                 ret = bch2_btree_repair_topology_recurse(c, cur);
 406                 six_unlock_read(&cur->c.lock);
 407                 cur = NULL;
 408
 409                 if (ret == DROP_THIS_NODE) {
 410                         bch2_btree_node_evict(c, cur_k.k);
 411                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 412                                                       b->c.level, cur_k.k->k.p);
 413                         dropped_children = true;
 414                 }
 415
 416                 if (ret)
 417                         goto err;
 418
 419                 have_child = true;
 420         }
 421
 422         if (mustfix_fsck_err_on(!have_child, c,
 423                         "empty interior btree node at btree %s level %u\n"
 424                         "  %s",
 425                         bch2_btree_ids[b->c.btree_id],
 426                         b->c.level,
 427                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
 428                 ret = DROP_THIS_NODE;
 429 err:
 430 fsck_err:
 431         if (!IS_ERR_OR_NULL(prev))
 432                 six_unlock_read(&prev->c.lock);
 433         if (!IS_ERR_OR_NULL(cur))
 434                 six_unlock_read(&cur->c.lock);
 435
 436         bch2_btree_and_journal_iter_exit(&iter);
 437         bch2_bkey_buf_exit(&prev_k, c);
 438         bch2_bkey_buf_exit(&cur_k, c);
 439
 440         if (!ret && dropped_children)
 441                 goto again;
 442
 443         return ret;
 444 }
 445
 446 static int bch2_repair_topology(struct bch_fs *c)
 447 {
 448         struct btree *b;
 449         unsigned i;
 450         int ret = 0;
 451
 452         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 453                 b = c->btree_roots[i].b;
 454                 if (btree_node_fake(b))
 455                         continue;
 456
 457                 six_lock_read(&b->c.lock, NULL, NULL);
 458                 ret = bch2_btree_repair_topology_recurse(c, b);
 459                 six_unlock_read(&b->c.lock);
 460
 461                 if (ret == DROP_THIS_NODE) {
 462                         bch_err(c, "empty btree root - repair unimplemented");
 463                         ret = FSCK_ERR_EXIT;
 464                 }
 465         }
 466
 467         return ret;
 468 }
 469
 470 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 471                                unsigned level, bool is_root,
 472                                struct bkey_s_c *k)
 473 {
 474         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 475         const union bch_extent_entry *entry;
 476         struct extent_ptr_decoded p = { 0 };
 477         bool do_update = false;
 478         char buf[200];
 479         int ret = 0;
 480
 481         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 482                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 483                 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
 484                 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
 485                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 486
 487                 if (fsck_err_on(g->mark.data_type &&
 488                                 g->mark.data_type != data_type, c,
 489                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 490                                 "while marking %s",
 491                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 492                                 bch2_data_types[g->mark.data_type],
 493                                 bch2_data_types[data_type],
 494                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 495                         if (data_type == BCH_DATA_btree) {
 496                                 g2->_mark.data_type = g->_mark.data_type = data_type;
 497                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 498                         } else {
 499                                 do_update = true;
 500                         }
 501                 }
 502
 503                 if (fsck_err_on(!g->gen_valid, c,
 504                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 505                                 "while marking %s",
 506                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 507                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 508                                 p.ptr.gen,
 509                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 510                         if (!p.ptr.cached) {
 511                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 512                                 g2->gen_valid   = g->gen_valid          = true;
 513                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 514                         } else {
 515                                 do_update = true;
 516                         }
 517                 }
 518
 519                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 520                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 521                                 "while marking %s",
 522                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 523                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 524                                 p.ptr.gen, g->mark.gen,
 525                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 526                         if (!p.ptr.cached) {
 527                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 528                                 g2->gen_valid   = g->gen_valid  = true;
 529                                 g2->_mark.data_type             = 0;
 530                                 g2->_mark.dirty_sectors         = 0;
 531                                 g2->_mark.cached_sectors        = 0;
 532                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 533                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 534                         } else {
 535                                 do_update = true;
 536                         }
 537                 }
 538
 539                 if (fsck_err_on(!p.ptr.cached &&
 540                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 541                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 542                                 "while marking %s",
 543                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 544                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 545                                 p.ptr.gen, g->mark.gen,
 546                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 547                         do_update = true;
 548
 549                 if (p.has_ec) {
 550                         struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
 551
 552                         if (fsck_err_on(!m || !m->alive, c,
 553                                         "pointer to nonexistent stripe %llu\n"
 554                                         "while marking %s",
 555                                         (u64) p.ec.idx,
 556                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 557                                 do_update = true;
 558
 559                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 560                                         "pointer does not match stripe %llu\n"
 561                                         "while marking %s",
 562                                         (u64) p.ec.idx,
 563                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 564                                 do_update = true;
 565                 }
 566         }
 567
 568         if (do_update) {
 569                 struct bkey_ptrs ptrs;
 570                 union bch_extent_entry *entry;
 571                 struct bch_extent_ptr *ptr;
 572                 struct bkey_i *new;
 573
 574                 if (is_root) {
 575                         bch_err(c, "cannot update btree roots yet");
 576                         return -EINVAL;
 577                 }
 578
 579                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 580                 if (!new) {
 581                         bch_err(c, "%s: error allocating new key", __func__);
 582                         return -ENOMEM;
 583                 }
 584
 585                 bkey_reassemble(new, *k);
 586
 587                 if (level) {
 588                         /*
 589                          * We don't want to drop btree node pointers - if the
 590                          * btree node isn't there anymore, the read path will
 591                          * sort it out:
 592                          */
 593                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 594                         bkey_for_each_ptr(ptrs, ptr) {
 595                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 596                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 597
 598                                 ptr->gen = g->mark.gen;
 599                         }
 600                 } else {
 601                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 602                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 603                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 604                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 605
 606                                 (ptr->cached &&
 607                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 608                                 (!ptr->cached &&
 609                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 610                                 (g->mark.data_type &&
 611                                  g->mark.data_type != data_type);
 612                         }));
 613 again:
 614                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 615                         bkey_extent_entry_for_each(ptrs, entry) {
 616                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 617                                         struct stripe *m = genradix_ptr(&c->stripes[true],
 618                                                                         entry->stripe_ptr.idx);
 619                                         union bch_extent_entry *next_ptr;
 620
 621                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 622                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 623                                                         goto found;
 624                                         next_ptr = NULL;
 625 found:
 626                                         if (!next_ptr) {
 627                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 628                                                 continue;
 629                                         }
 630
 631                                         if (!m || !m->alive ||
 632                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 633                                                                        &next_ptr->ptr,
 634                                                                        m->sectors)) {
 635                                                 bch2_bkey_extent_entry_drop(new, entry);
 636                                                 goto again;
 637                                         }
 638                                 }
 639                         }
 640                 }
 641
 642                 ret = bch2_journal_key_insert(c, btree_id, level, new);
 643                 if (ret)
 644                         kfree(new);
 645                 else
 646                         *k = bkey_i_to_s_c(new);
 647         }
 648 fsck_err:
 649         return ret;
 650 }
 651
 652 /* marking of btree keys/nodes: */
 653
 654 static int bch2_gc_mark_key(struct bch_fs *c, enum btree_id btree_id,
 655                             unsigned level, bool is_root,
 656                             struct bkey_s_c *k,
 657                             u8 *max_stale, bool initial)
 658 {
 659         struct bkey_ptrs_c ptrs;
 660         const struct bch_extent_ptr *ptr;
 661         unsigned flags =
 662                 BTREE_TRIGGER_INSERT|
 663                 BTREE_TRIGGER_GC|
 664                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 665         int ret = 0;
 666
 667         if (initial) {
 668                 BUG_ON(bch2_journal_seq_verify &&
 669                        k->k->version.lo > journal_cur_seq(&c->journal));
 670
 671                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 672                 if (ret)
 673                         goto err;
 674
 675                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 676                                 "key version number higher than recorded: %llu > %llu",
 677                                 k->k->version.lo,
 678                                 atomic64_read(&c->key_version)))
 679                         atomic64_set(&c->key_version, k->k->version.lo);
 680
 681                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 682                     fsck_err_on(!bch2_bkey_replicas_marked(c, *k), c,
 683                                 "superblock not marked as containing replicas (type %u)",
 684                                 k->k->type)) {
 685                         ret = bch2_mark_bkey_replicas(c, *k);
 686                         if (ret) {
 687                                 bch_err(c, "error marking bkey replicas: %i", ret);
 688                                 goto err;
 689                         }
 690                 }
 691         }
 692
 693         ptrs = bch2_bkey_ptrs_c(*k);
 694         bkey_for_each_ptr(ptrs, ptr) {
 695                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 696                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 697
 698                 if (gen_after(g->oldest_gen, ptr->gen))
 699                         g->oldest_gen = ptr->gen;
 700
 701                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 702         }
 703
 704         bch2_mark_key(c, *k, flags);
 705 fsck_err:
 706 err:
 707         if (ret)
 708                 bch_err(c, "%s: ret %i", __func__, ret);
 709         return ret;
 710 }
 711
 712 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
 713                               bool initial)
 714 {
 715         struct btree_node_iter iter;
 716         struct bkey unpacked;
 717         struct bkey_s_c k;
 718         struct bkey_buf prev, cur;
 719         int ret = 0;
 720
 721         *max_stale = 0;
 722
 723         if (!btree_node_type_needs_gc(btree_node_type(b)))
 724                 return 0;
 725
 726         bch2_btree_node_iter_init_from_start(&iter, b);
 727         bch2_bkey_buf_init(&prev);
 728         bch2_bkey_buf_init(&cur);
 729         bkey_init(&prev.k->k);
 730
 731         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 732                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 733                                        &k, max_stale, initial);
 734                 if (ret)
 735                         break;
 736
 737                 bch2_btree_node_iter_advance(&iter, b);
 738
 739                 if (b->c.level) {
 740                         bch2_bkey_buf_reassemble(&cur, c, k);
 741
 742                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 743                                         bch2_btree_node_iter_end(&iter));
 744                         if (ret)
 745                                 break;
 746                 }
 747         }
 748
 749         bch2_bkey_buf_exit(&cur, c);
 750         bch2_bkey_buf_exit(&prev, c);
 751         return ret;
 752 }
 753
 754 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 755                          bool initial, bool metadata_only)
 756 {
 757         struct btree_trans trans;
 758         struct btree_iter *iter;
 759         struct btree *b;
 760         unsigned depth = metadata_only                  ? 1
 761                 : bch2_expensive_debug_checks           ? 0
 762                 : !btree_node_type_needs_gc(btree_id)   ? 1
 763                 : 0;
 764         u8 max_stale = 0;
 765         int ret = 0;
 766
 767         bch2_trans_init(&trans, c, 0, 0);
 768
 769         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 770
 771         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 772                               0, depth, BTREE_ITER_PREFETCH, b) {
 773                 bch2_verify_btree_nr_keys(b);
 774
 775                 gc_pos_set(c, gc_pos_btree_node(b));
 776
 777                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 778                 if (ret)
 779                         break;
 780
 781                 if (!initial) {
 782                         if (max_stale > 64)
 783                                 bch2_btree_node_rewrite(c, iter,
 784                                                 b->data->keys.seq,
 785                                                 BTREE_INSERT_NOWAIT|
 786                                                 BTREE_INSERT_GC_LOCK_HELD);
 787                         else if (!bch2_btree_gc_rewrite_disabled &&
 788                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 789                                 bch2_btree_node_rewrite(c, iter,
 790                                                 b->data->keys.seq,
 791                                                 BTREE_INSERT_NOWAIT|
 792                                                 BTREE_INSERT_GC_LOCK_HELD);
 793                 }
 794
 795                 bch2_trans_cond_resched(&trans);
 796         }
 797         bch2_trans_iter_put(&trans, iter);
 798
 799         ret = bch2_trans_exit(&trans) ?: ret;
 800         if (ret)
 801                 return ret;
 802
 803         mutex_lock(&c->btree_root_lock);
 804         b = c->btree_roots[btree_id].b;
 805         if (!btree_node_fake(b)) {
 806                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 807
 808                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 809                                        &k, &max_stale, initial);
 810         }
 811         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 812         mutex_unlock(&c->btree_root_lock);
 813
 814         return ret;
 815 }
 816
 817 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
 818                                       unsigned target_depth)
 819 {
 820         struct btree_and_journal_iter iter;
 821         struct bkey_s_c k;
 822         struct bkey_buf cur, prev;
 823         u8 max_stale = 0;
 824         char buf[200];
 825         int ret = 0;
 826
 827         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 828         bch2_bkey_buf_init(&prev);
 829         bch2_bkey_buf_init(&cur);
 830         bkey_init(&prev.k->k);
 831
 832         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 833                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 834                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 835
 836                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 837                                        &k, &max_stale, true);
 838                 if (ret) {
 839                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 840                         goto fsck_err;
 841                 }
 842
 843                 if (b->c.level) {
 844                         bch2_bkey_buf_reassemble(&cur, c, k);
 845                         k = bkey_i_to_s_c(cur.k);
 846
 847                         bch2_btree_and_journal_iter_advance(&iter);
 848
 849                         ret = bch2_gc_check_topology(c, b,
 850                                         &prev, cur,
 851                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 852                         if (ret)
 853                                 goto fsck_err;
 854                 } else {
 855                         bch2_btree_and_journal_iter_advance(&iter);
 856                 }
 857         }
 858
 859         if (b->c.level > target_depth) {
 860                 bch2_btree_and_journal_iter_exit(&iter);
 861                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 862
 863                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 864                         struct btree *child;
 865
 866                         bch2_bkey_buf_reassemble(&cur, c, k);
 867                         bch2_btree_and_journal_iter_advance(&iter);
 868
 869                         child = bch2_btree_node_get_noiter(c, cur.k,
 870                                                 b->c.btree_id, b->c.level - 1,
 871                                                 false);
 872                         ret = PTR_ERR_OR_ZERO(child);
 873
 874                         if (ret == -EIO) {
 875                                 bch2_topology_error(c);
 876
 877                                 if (__fsck_err(c,
 878                                           FSCK_CAN_FIX|
 879                                           FSCK_CAN_IGNORE|
 880                                           FSCK_NO_RATELIMIT,
 881                                           "Unreadable btree node at btree %s level %u:\n"
 882                                           "  %s",
 883                                           bch2_btree_ids[b->c.btree_id],
 884                                           b->c.level - 1,
 885                                           (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
 886                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 887                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 888                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 889                                         goto fsck_err;
 890                                 } else {
 891                                         /* Continue marking when opted to not
 892                                          * fix the error: */
 893                                         ret = 0;
 894                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 895                                         continue;
 896                                 }
 897                         } else if (ret) {
 898                                 bch_err(c, "%s: error %i getting btree node",
 899                                         __func__, ret);
 900                                 break;
 901                         }
 902
 903                         ret = bch2_gc_btree_init_recurse(c, child,
 904                                                          target_depth);
 905                         six_unlock_read(&child->c.lock);
 906
 907                         if (ret)
 908                                 break;
 909                 }
 910         }
 911 fsck_err:
 912         bch2_bkey_buf_exit(&cur, c);
 913         bch2_bkey_buf_exit(&prev, c);
 914         bch2_btree_and_journal_iter_exit(&iter);
 915         return ret;
 916 }
 917
 918 static int bch2_gc_btree_init(struct bch_fs *c,
 919                               enum btree_id btree_id,
 920                               bool metadata_only)
 921 {
 922         struct btree *b;
 923         unsigned target_depth = metadata_only           ? 1
 924                 : bch2_expensive_debug_checks           ? 0
 925                 : !btree_node_type_needs_gc(btree_id)   ? 1
 926                 : 0;
 927         u8 max_stale = 0;
 928         char buf[100];
 929         int ret = 0;
 930
 931         b = c->btree_roots[btree_id].b;
 932
 933         if (btree_node_fake(b))
 934                 return 0;
 935
 936         six_lock_read(&b->c.lock, NULL, NULL);
 937         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 938                         "btree root with incorrect min_key: %s",
 939                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 940                 bch_err(c, "repair unimplemented");
 941                 ret = FSCK_ERR_EXIT;
 942                 goto fsck_err;
 943         }
 944
 945         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, POS_MAX), c,
 946                         "btree root with incorrect max_key: %s",
 947                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 948                 bch_err(c, "repair unimplemented");
 949                 ret = FSCK_ERR_EXIT;
 950                 goto fsck_err;
 951         }
 952
 953         if (b->c.level >= target_depth)
 954                 ret = bch2_gc_btree_init_recurse(c, b, target_depth);
 955
 956         if (!ret) {
 957                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 958
 959                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 960                                        &k, &max_stale, true);
 961         }
 962 fsck_err:
 963         six_unlock_read(&b->c.lock);
 964
 965         if (ret < 0)
 966                 bch_err(c, "%s: ret %i", __func__, ret);
 967         return ret;
 968 }
 969
 970 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 971 {
 972         return  (int) btree_id_to_gc_phase(l) -
 973                 (int) btree_id_to_gc_phase(r);
 974 }
 975
 976 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
 977 {
 978         enum btree_id ids[BTREE_ID_NR];
 979         unsigned i;
 980         int ret = 0;
 981
 982         for (i = 0; i < BTREE_ID_NR; i++)
 983                 ids[i] = i;
 984         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
 985
 986         for (i = 0; i < BTREE_ID_NR && !ret; i++)
 987                 ret = initial
 988                         ? bch2_gc_btree_init(c, ids[i], metadata_only)
 989                         : bch2_gc_btree(c, ids[i], initial, metadata_only);
 990
 991         if (ret < 0)
 992                 bch_err(c, "%s: ret %i", __func__, ret);
 993         return ret;
 994 }
 995
 996 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 997                                   u64 start, u64 end,
 998                                   enum bch_data_type type,
 999                                   unsigned flags)
1000 {
1001         u64 b = sector_to_bucket(ca, start);
1002
1003         do {
1004                 unsigned sectors =
1005                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1006
1007                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1008                                           gc_phase(GC_PHASE_SB), flags);
1009                 b++;
1010                 start += sectors;
1011         } while (start < end);
1012 }
1013
1014 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1015                               unsigned flags)
1016 {
1017         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1018         unsigned i;
1019         u64 b;
1020
1021         /*
1022          * This conditional is kind of gross, but we may be called from the
1023          * device add path, before the new device has actually been added to the
1024          * running filesystem:
1025          */
1026         if (c) {
1027                 lockdep_assert_held(&c->sb_lock);
1028                 percpu_down_read(&c->mark_lock);
1029         }
1030
1031         for (i = 0; i < layout->nr_superblocks; i++) {
1032                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1033
1034                 if (offset == BCH_SB_SECTOR)
1035                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1036                                               BCH_DATA_sb, flags);
1037
1038                 mark_metadata_sectors(c, ca, offset,
1039                                       offset + (1 << layout->sb_max_size_bits),
1040                                       BCH_DATA_sb, flags);
1041         }
1042
1043         for (i = 0; i < ca->journal.nr; i++) {
1044                 b = ca->journal.buckets[i];
1045                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1046                                           ca->mi.bucket_size,
1047                                           gc_phase(GC_PHASE_SB), flags);
1048         }
1049
1050         if (c)
1051                 percpu_up_read(&c->mark_lock);
1052 }
1053
1054 static void bch2_mark_superblocks(struct bch_fs *c)
1055 {
1056         struct bch_dev *ca;
1057         unsigned i;
1058
1059         mutex_lock(&c->sb_lock);
1060         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1061
1062         for_each_online_member(ca, c, i)
1063                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1064         mutex_unlock(&c->sb_lock);
1065 }
1066
1067 #if 0
1068 /* Also see bch2_pending_btree_node_free_insert_done() */
1069 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1070 {
1071         struct btree_update *as;
1072         struct pending_btree_node_free *d;
1073
1074         mutex_lock(&c->btree_interior_update_lock);
1075         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1076
1077         for_each_pending_btree_node_free(c, as, d)
1078                 if (d->index_update_done)
1079                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
1080                                       BTREE_TRIGGER_INSERT|BTREE_TRIGGER_GC);
1081
1082         mutex_unlock(&c->btree_interior_update_lock);
1083 }
1084 #endif
1085
1086 static void bch2_gc_free(struct bch_fs *c)
1087 {
1088         struct bch_dev *ca;
1089         unsigned i;
1090
1091         genradix_free(&c->stripes[1]);
1092
1093         for_each_member_device(ca, c, i) {
1094                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1095                         sizeof(struct bucket_array) +
1096                         ca->mi.nbuckets * sizeof(struct bucket));
1097                 ca->buckets[1] = NULL;
1098
1099                 free_percpu(ca->usage_gc);
1100                 ca->usage_gc = NULL;
1101         }
1102
1103         free_percpu(c->usage_gc);
1104         c->usage_gc = NULL;
1105 }
1106
1107 static int bch2_gc_done(struct bch_fs *c,
1108                         bool initial, bool metadata_only)
1109 {
1110         struct bch_dev *ca = NULL;
1111         bool verify = !metadata_only && (!initial ||
1112                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1113         unsigned i, dev;
1114         int ret = 0;
1115
1116 #define copy_field(_f, _msg, ...)                                       \
1117         if (dst->_f != src->_f) {                                       \
1118                 if (verify)                                             \
1119                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1120                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1121                 dst->_f = src->_f;                                      \
1122                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1123         }
1124 #define copy_stripe_field(_f, _msg, ...)                                \
1125         if (dst->_f != src->_f) {                                       \
1126                 if (verify)                                             \
1127                         fsck_err(c, "stripe %zu has wrong "_msg         \
1128                                 ": got %u, should be %u",               \
1129                                 iter.pos, ##__VA_ARGS__,                \
1130                                 dst->_f, src->_f);                      \
1131                 dst->_f = src->_f;                                      \
1132                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1133         }
1134 #define copy_bucket_field(_f)                                           \
1135         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
1136                 if (verify)                                             \
1137                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
1138                                 ": got %u, should be %u", dev, b,       \
1139                                 dst->b[b].mark.gen,                     \
1140                                 bch2_data_types[dst->b[b].mark.data_type],\
1141                                 dst->b[b].mark._f, src->b[b].mark._f);  \
1142                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
1143                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1144         }
1145 #define copy_dev_field(_f, _msg, ...)                                   \
1146         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1147 #define copy_fs_field(_f, _msg, ...)                                    \
1148         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1149
1150         if (!metadata_only) {
1151                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
1152                 struct stripe *dst, *src;
1153
1154                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
1155                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
1156
1157                         if (dst->alive          != src->alive ||
1158                             dst->sectors        != src->sectors ||
1159                             dst->algorithm      != src->algorithm ||
1160                             dst->nr_blocks      != src->nr_blocks ||
1161                             dst->nr_redundant   != src->nr_redundant) {
1162                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
1163                                 ret = -EINVAL;
1164                                 goto fsck_err;
1165                         }
1166
1167                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
1168                                 copy_stripe_field(block_sectors[i],
1169                                                   "block_sectors[%u]", i);
1170
1171                         dst->blocks_nonempty = 0;
1172                         for (i = 0; i < dst->nr_blocks; i++)
1173                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
1174
1175                         genradix_iter_advance(&iter, &c->stripes[1]);
1176                 }
1177         }
1178
1179         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1180                 bch2_fs_usage_acc_to_base(c, i);
1181
1182         for_each_member_device(ca, c, dev) {
1183                 struct bucket_array *dst = __bucket_array(ca, 0);
1184                 struct bucket_array *src = __bucket_array(ca, 1);
1185                 size_t b;
1186
1187                 for (b = 0; b < src->nbuckets; b++) {
1188                         copy_bucket_field(gen);
1189                         copy_bucket_field(data_type);
1190                         copy_bucket_field(stripe);
1191                         copy_bucket_field(dirty_sectors);
1192                         copy_bucket_field(cached_sectors);
1193
1194                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
1195                 }
1196
1197                 {
1198                         struct bch_dev_usage *dst = ca->usage_base;
1199                         struct bch_dev_usage *src = (void *)
1200                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1201                                                      dev_usage_u64s());
1202
1203                         copy_dev_field(buckets_ec,              "buckets_ec");
1204                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1205
1206                         for (i = 0; i < BCH_DATA_NR; i++) {
1207                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1208                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1209                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1210                         }
1211                 }
1212         };
1213
1214         {
1215                 unsigned nr = fs_usage_u64s(c);
1216                 struct bch_fs_usage *dst = c->usage_base;
1217                 struct bch_fs_usage *src = (void *)
1218                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1219
1220                 copy_fs_field(hidden,           "hidden");
1221                 copy_fs_field(btree,            "btree");
1222
1223                 if (!metadata_only) {
1224                         copy_fs_field(data,     "data");
1225                         copy_fs_field(cached,   "cached");
1226                         copy_fs_field(reserved, "reserved");
1227                         copy_fs_field(nr_inodes,"nr_inodes");
1228
1229                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1230                                 copy_fs_field(persistent_reserved[i],
1231                                               "persistent_reserved[%i]", i);
1232                 }
1233
1234                 for (i = 0; i < c->replicas.nr; i++) {
1235                         struct bch_replicas_entry *e =
1236                                 cpu_replicas_entry(&c->replicas, i);
1237                         char buf[80];
1238
1239                         if (metadata_only &&
1240                             (e->data_type == BCH_DATA_user ||
1241                              e->data_type == BCH_DATA_cached))
1242                                 continue;
1243
1244                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1245
1246                         copy_fs_field(replicas[i], "%s", buf);
1247                 }
1248         }
1249
1250 #undef copy_fs_field
1251 #undef copy_dev_field
1252 #undef copy_bucket_field
1253 #undef copy_stripe_field
1254 #undef copy_field
1255 fsck_err:
1256         if (ca)
1257                 percpu_ref_put(&ca->ref);
1258         if (ret)
1259                 bch_err(c, "%s: ret %i", __func__, ret);
1260         return ret;
1261 }
1262
1263 static int bch2_gc_start(struct bch_fs *c,
1264                          bool metadata_only)
1265 {
1266         struct bch_dev *ca = NULL;
1267         unsigned i;
1268         int ret;
1269
1270         BUG_ON(c->usage_gc);
1271
1272         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1273                                          sizeof(u64), GFP_KERNEL);
1274         if (!c->usage_gc) {
1275                 bch_err(c, "error allocating c->usage_gc");
1276                 return -ENOMEM;
1277         }
1278
1279         for_each_member_device(ca, c, i) {
1280                 BUG_ON(ca->buckets[1]);
1281                 BUG_ON(ca->usage_gc);
1282
1283                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1284                                 ca->mi.nbuckets * sizeof(struct bucket),
1285                                 GFP_KERNEL|__GFP_ZERO);
1286                 if (!ca->buckets[1]) {
1287                         percpu_ref_put(&ca->ref);
1288                         bch_err(c, "error allocating ca->buckets[gc]");
1289                         return -ENOMEM;
1290                 }
1291
1292                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1293                 if (!ca->usage_gc) {
1294                         bch_err(c, "error allocating ca->usage_gc");
1295                         percpu_ref_put(&ca->ref);
1296                         return -ENOMEM;
1297                 }
1298         }
1299
1300         ret = bch2_ec_mem_alloc(c, true);
1301         if (ret) {
1302                 bch_err(c, "error allocating ec gc mem");
1303                 return ret;
1304         }
1305
1306         percpu_down_write(&c->mark_lock);
1307
1308         /*
1309          * indicate to stripe code that we need to allocate for the gc stripes
1310          * radix tree, too
1311          */
1312         gc_pos_set(c, gc_phase(GC_PHASE_START));
1313
1314         for_each_member_device(ca, c, i) {
1315                 struct bucket_array *dst = __bucket_array(ca, 1);
1316                 struct bucket_array *src = __bucket_array(ca, 0);
1317                 size_t b;
1318
1319                 dst->first_bucket       = src->first_bucket;
1320                 dst->nbuckets           = src->nbuckets;
1321
1322                 for (b = 0; b < src->nbuckets; b++) {
1323                         struct bucket *d = &dst->b[b];
1324                         struct bucket *s = &src->b[b];
1325
1326                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1327                         d->gen_valid = s->gen_valid;
1328
1329                         if (metadata_only &&
1330                             (s->mark.data_type == BCH_DATA_user ||
1331                              s->mark.data_type == BCH_DATA_cached))
1332                                 d->_mark = s->mark;
1333                 }
1334         };
1335
1336         percpu_up_write(&c->mark_lock);
1337
1338         return 0;
1339 }
1340
1341 static int bch2_gc_reflink_done_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1342 {
1343         struct reflink_gc *r;
1344         const __le64 *refcount = bkey_refcount_c(k);
1345         char buf[200];
1346         int ret = 0;
1347
1348         if (!refcount)
1349                 return 0;
1350
1351         r = genradix_ptr(&c->reflink_gc_table, c->reflink_gc_idx++);
1352         if (!r)
1353                 return -ENOMEM;
1354
1355         if (!r ||
1356             r->offset != k.k->p.offset ||
1357             r->size != k.k->size) {
1358                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1359                 return -EINVAL;
1360         }
1361
1362         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1363                         "reflink key has wrong refcount:\n"
1364                         "  %s\n"
1365                         "  should be %u",
1366                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1367                         r->refcount)) {
1368                 struct bkey_i *new;
1369
1370                 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1371                 if (!new) {
1372                         ret = -ENOMEM;
1373                         goto fsck_err;
1374                 }
1375
1376                 bkey_reassemble(new, k);
1377
1378                 if (!r->refcount) {
1379                         new->k.type = KEY_TYPE_deleted;
1380                         new->k.size = 0;
1381                 } else {
1382                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1383                 }
1384
1385                 ret = bch2_journal_key_insert(c, BTREE_ID_reflink, 0, new);
1386                 if (ret)
1387                         kfree(new);
1388         }
1389 fsck_err:
1390         return ret;
1391 }
1392
1393 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1394                                 bool metadata_only)
1395 {
1396         struct btree_trans trans;
1397         struct btree_iter *iter;
1398         struct bkey_s_c k;
1399         struct reflink_gc *r;
1400         size_t idx = 0;
1401         char buf[200];
1402         int ret = 0;
1403
1404         if (metadata_only)
1405                 return 0;
1406
1407         if (initial) {
1408                 c->reflink_gc_idx = 0;
1409
1410                 ret = bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1411                                 bch2_gc_reflink_done_initial_fn);
1412                 goto out;
1413         }
1414
1415         bch2_trans_init(&trans, c, 0, 0);
1416
1417         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1418                            BTREE_ITER_PREFETCH, k, ret) {
1419                 const __le64 *refcount = bkey_refcount_c(k);
1420
1421                 if (!refcount)
1422                         continue;
1423
1424                 r = genradix_ptr(&c->reflink_gc_table, idx);
1425                 if (!r ||
1426                     r->offset != k.k->p.offset ||
1427                     r->size != k.k->size) {
1428                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1429                         ret = -EINVAL;
1430                         break;
1431                 }
1432
1433                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1434                                 "reflink key has wrong refcount:\n"
1435                                 "  %s\n"
1436                                 "  should be %u",
1437                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1438                                 r->refcount)) {
1439                         struct bkey_i *new;
1440
1441                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1442                         if (!new) {
1443                                 ret = -ENOMEM;
1444                                 break;
1445                         }
1446
1447                         bkey_reassemble(new, k);
1448
1449                         if (!r->refcount)
1450                                 new->k.type = KEY_TYPE_deleted;
1451                         else
1452                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1453
1454                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1455                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1456                         kfree(new);
1457
1458                         if (ret)
1459                                 break;
1460                 }
1461         }
1462 fsck_err:
1463         bch2_trans_iter_put(&trans, iter);
1464         bch2_trans_exit(&trans);
1465 out:
1466         genradix_free(&c->reflink_gc_table);
1467         c->reflink_gc_nr = 0;
1468         return ret;
1469 }
1470
1471 static int bch2_gc_reflink_start_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1472 {
1473
1474         struct reflink_gc *r;
1475         const __le64 *refcount = bkey_refcount_c(k);
1476
1477         if (!refcount)
1478                 return 0;
1479
1480         r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1481                                GFP_KERNEL);
1482         if (!r)
1483                 return -ENOMEM;
1484
1485         r->offset       = k.k->p.offset;
1486         r->size         = k.k->size;
1487         r->refcount     = 0;
1488         return 0;
1489 }
1490
1491 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1492                                  bool metadata_only)
1493 {
1494         struct btree_trans trans;
1495         struct btree_iter *iter;
1496         struct bkey_s_c k;
1497         struct reflink_gc *r;
1498         int ret;
1499
1500         if (metadata_only)
1501                 return 0;
1502
1503         genradix_free(&c->reflink_gc_table);
1504         c->reflink_gc_nr = 0;
1505
1506         if (initial)
1507                 return bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1508                                 bch2_gc_reflink_start_initial_fn);
1509
1510         bch2_trans_init(&trans, c, 0, 0);
1511
1512         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1513                            BTREE_ITER_PREFETCH, k, ret) {
1514                 const __le64 *refcount = bkey_refcount_c(k);
1515
1516                 if (!refcount)
1517                         continue;
1518
1519                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1520                                        GFP_KERNEL);
1521                 if (!r) {
1522                         ret = -ENOMEM;
1523                         break;
1524                 }
1525
1526                 r->offset       = k.k->p.offset;
1527                 r->size         = k.k->size;
1528                 r->refcount     = 0;
1529         }
1530         bch2_trans_iter_put(&trans, iter);
1531
1532         bch2_trans_exit(&trans);
1533         return 0;
1534 }
1535
1536 /**
1537  * bch2_gc - walk _all_ references to buckets, and recompute them:
1538  *
1539  * Order matters here:
1540  *  - Concurrent GC relies on the fact that we have a total ordering for
1541  *    everything that GC walks - see  gc_will_visit_node(),
1542  *    gc_will_visit_root()
1543  *
1544  *  - also, references move around in the course of index updates and
1545  *    various other crap: everything needs to agree on the ordering
1546  *    references are allowed to move around in - e.g., we're allowed to
1547  *    start with a reference owned by an open_bucket (the allocator) and
1548  *    move it to the btree, but not the reverse.
1549  *
1550  *    This is necessary to ensure that gc doesn't miss references that
1551  *    move around - if references move backwards in the ordering GC
1552  *    uses, GC could skip past them
1553  */
1554 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1555 {
1556         struct bch_dev *ca;
1557         u64 start_time = local_clock();
1558         unsigned i, iter = 0;
1559         int ret;
1560
1561         lockdep_assert_held(&c->state_lock);
1562         trace_gc_start(c);
1563
1564         down_write(&c->gc_lock);
1565
1566         /* flush interior btree updates: */
1567         closure_wait_event(&c->btree_interior_update_wait,
1568                            !bch2_btree_interior_updates_nr_pending(c));
1569 again:
1570         ret   = bch2_gc_start(c, metadata_only) ?:
1571                 bch2_gc_reflink_start(c, initial, metadata_only);
1572         if (ret)
1573                 goto out;
1574
1575         bch2_mark_superblocks(c);
1576
1577         if (test_bit(BCH_FS_TOPOLOGY_ERROR, &c->flags) &&
1578             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1579             c->opts.fix_errors != FSCK_OPT_NO) {
1580                 bch_info(c, "starting topology repair pass");
1581                 ret = bch2_repair_topology(c);
1582                 if (ret)
1583                         goto out;
1584                 bch_info(c, "topology repair pass done");
1585
1586                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1587         }
1588
1589         ret = bch2_gc_btrees(c, initial, metadata_only);
1590
1591         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1592             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1593             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1594                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1595                 ret = 0;
1596         }
1597
1598         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1599                 ret = FSCK_ERR_EXIT;
1600
1601         if (ret)
1602                 goto out;
1603
1604 #if 0
1605         bch2_mark_pending_btree_node_frees(c);
1606 #endif
1607         c->gc_count++;
1608
1609         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1610             (!iter && bch2_test_restart_gc)) {
1611                 /*
1612                  * XXX: make sure gens we fixed got saved
1613                  */
1614                 if (iter++ <= 2) {
1615                         bch_info(c, "Second GC pass needed, restarting:");
1616                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1617                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1618
1619                         percpu_down_write(&c->mark_lock);
1620                         bch2_gc_free(c);
1621                         percpu_up_write(&c->mark_lock);
1622                         /* flush fsck errors, reset counters */
1623                         bch2_flush_fsck_errs(c);
1624
1625                         goto again;
1626                 }
1627
1628                 bch_info(c, "Unable to fix bucket gens, looping");
1629                 ret = -EINVAL;
1630         }
1631 out:
1632         if (!ret) {
1633                 bch2_journal_block(&c->journal);
1634
1635                 percpu_down_write(&c->mark_lock);
1636                 ret   = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1637                         bch2_gc_done(c, initial, metadata_only);
1638
1639                 bch2_journal_unblock(&c->journal);
1640         } else {
1641                 percpu_down_write(&c->mark_lock);
1642         }
1643
1644         /* Indicates that gc is no longer in progress: */
1645         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1646
1647         bch2_gc_free(c);
1648         percpu_up_write(&c->mark_lock);
1649
1650         up_write(&c->gc_lock);
1651
1652         trace_gc_end(c);
1653         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1654
1655         /*
1656          * Wake up allocator in case it was waiting for buckets
1657          * because of not being able to inc gens
1658          */
1659         for_each_member_device(ca, c, i)
1660                 bch2_wake_allocator(ca);
1661
1662         /*
1663          * At startup, allocations can happen directly instead of via the
1664          * allocator thread - issue wakeup in case they blocked on gc_lock:
1665          */
1666         closure_wake_up(&c->freelist_wait);
1667         return ret;
1668 }
1669
1670 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1671 {
1672         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1673         const struct bch_extent_ptr *ptr;
1674
1675         percpu_down_read(&c->mark_lock);
1676         bkey_for_each_ptr(ptrs, ptr) {
1677                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1678                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1679
1680                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1681                         percpu_up_read(&c->mark_lock);
1682                         return true;
1683                 }
1684         }
1685
1686         bkey_for_each_ptr(ptrs, ptr) {
1687                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1688                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1689
1690                 if (gen_after(g->gc_gen, ptr->gen))
1691                         g->gc_gen = ptr->gen;
1692         }
1693         percpu_up_read(&c->mark_lock);
1694
1695         return false;
1696 }
1697
1698 /*
1699  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1700  * node pointers currently never have cached pointers that can become stale:
1701  */
1702 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1703 {
1704         struct btree_trans trans;
1705         struct btree_iter *iter;
1706         struct bkey_s_c k;
1707         struct bkey_buf sk;
1708         int ret = 0, commit_err = 0;
1709
1710         bch2_bkey_buf_init(&sk);
1711         bch2_trans_init(&trans, c, 0, 0);
1712
1713         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
1714                                    BTREE_ITER_PREFETCH|
1715                                    BTREE_ITER_NOT_EXTENTS|
1716                                    BTREE_ITER_ALL_SNAPSHOTS);
1717
1718         while ((k = bch2_btree_iter_peek(iter)).k &&
1719                !(ret = bkey_err(k))) {
1720                 c->gc_gens_pos = iter->pos;
1721
1722                 if (gc_btree_gens_key(c, k) && !commit_err) {
1723                         bch2_bkey_buf_reassemble(&sk, c, k);
1724                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1725
1726
1727                         commit_err =
1728                                 bch2_trans_update(&trans, iter, sk.k, 0) ?:
1729                                 bch2_trans_commit(&trans, NULL, NULL,
1730                                                        BTREE_INSERT_NOWAIT|
1731                                                        BTREE_INSERT_NOFAIL);
1732                         if (commit_err == -EINTR) {
1733                                 commit_err = 0;
1734                                 continue;
1735                         }
1736                 }
1737
1738                 bch2_btree_iter_advance(iter);
1739         }
1740         bch2_trans_iter_put(&trans, iter);
1741
1742         bch2_trans_exit(&trans);
1743         bch2_bkey_buf_exit(&sk, c);
1744
1745         return ret;
1746 }
1747
1748 int bch2_gc_gens(struct bch_fs *c)
1749 {
1750         struct bch_dev *ca;
1751         struct bucket_array *buckets;
1752         struct bucket *g;
1753         unsigned i;
1754         int ret;
1755
1756         /*
1757          * Ideally we would be using state_lock and not gc_lock here, but that
1758          * introduces a deadlock in the RO path - we currently take the state
1759          * lock at the start of going RO, thus the gc thread may get stuck:
1760          */
1761         down_read(&c->gc_lock);
1762
1763         for_each_member_device(ca, c, i) {
1764                 down_read(&ca->bucket_lock);
1765                 buckets = bucket_array(ca);
1766
1767                 for_each_bucket(g, buckets)
1768                         g->gc_gen = g->mark.gen;
1769                 up_read(&ca->bucket_lock);
1770         }
1771
1772         for (i = 0; i < BTREE_ID_NR; i++)
1773                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1774                         c->gc_gens_btree = i;
1775                         c->gc_gens_pos = POS_MIN;
1776                         ret = bch2_gc_btree_gens(c, i);
1777                         if (ret) {
1778                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1779                                 goto err;
1780                         }
1781                 }
1782
1783         for_each_member_device(ca, c, i) {
1784                 down_read(&ca->bucket_lock);
1785                 buckets = bucket_array(ca);
1786
1787                 for_each_bucket(g, buckets)
1788                         g->oldest_gen = g->gc_gen;
1789                 up_read(&ca->bucket_lock);
1790         }
1791
1792         c->gc_gens_btree        = 0;
1793         c->gc_gens_pos          = POS_MIN;
1794
1795         c->gc_count++;
1796 err:
1797         up_read(&c->gc_lock);
1798         return ret;
1799 }
1800
1801 static int bch2_gc_thread(void *arg)
1802 {
1803         struct bch_fs *c = arg;
1804         struct io_clock *clock = &c->io_clock[WRITE];
1805         unsigned long last = atomic64_read(&clock->now);
1806         unsigned last_kick = atomic_read(&c->kick_gc);
1807         int ret;
1808
1809         set_freezable();
1810
1811         while (1) {
1812                 while (1) {
1813                         set_current_state(TASK_INTERRUPTIBLE);
1814
1815                         if (kthread_should_stop()) {
1816                                 __set_current_state(TASK_RUNNING);
1817                                 return 0;
1818                         }
1819
1820                         if (atomic_read(&c->kick_gc) != last_kick)
1821                                 break;
1822
1823                         if (c->btree_gc_periodic) {
1824                                 unsigned long next = last + c->capacity / 16;
1825
1826                                 if (atomic64_read(&clock->now) >= next)
1827                                         break;
1828
1829                                 bch2_io_clock_schedule_timeout(clock, next);
1830                         } else {
1831                                 schedule();
1832                         }
1833
1834                         try_to_freeze();
1835                 }
1836                 __set_current_state(TASK_RUNNING);
1837
1838                 last = atomic64_read(&clock->now);
1839                 last_kick = atomic_read(&c->kick_gc);
1840
1841                 /*
1842                  * Full gc is currently incompatible with btree key cache:
1843                  */
1844 #if 0
1845                 ret = bch2_gc(c, false, false);
1846 #else
1847                 ret = bch2_gc_gens(c);
1848 #endif
1849                 if (ret < 0)
1850                         bch_err(c, "btree gc failed: %i", ret);
1851
1852                 debug_check_no_locks_held();
1853         }
1854
1855         return 0;
1856 }
1857
1858 void bch2_gc_thread_stop(struct bch_fs *c)
1859 {
1860         struct task_struct *p;
1861
1862         p = c->gc_thread;
1863         c->gc_thread = NULL;
1864
1865         if (p) {
1866                 kthread_stop(p);
1867                 put_task_struct(p);
1868         }
1869 }
1870
1871 int bch2_gc_thread_start(struct bch_fs *c)
1872 {
1873         struct task_struct *p;
1874
1875         if (c->gc_thread)
1876                 return 0;
1877
1878         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1879         if (IS_ERR(p)) {
1880                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1881                 return PTR_ERR(p);
1882         }
1883
1884         get_task_struct(p);
1885         c->gc_thread = p;
1886         wake_up_process(p);
1887         return 0;
1888 }