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 void bch2_btree_node_update_key_early(struct bch_fs *c,
 159                                              enum btree_id btree, unsigned level,
 160                                              struct bkey_s_c old, struct bkey_i *new)
 161 {
 162         struct btree *b;
 163         struct bkey_buf tmp;
 164         int ret;
 165
 166         bch2_bkey_buf_init(&tmp);
 167         bch2_bkey_buf_reassemble(&tmp, c, old);
 168
 169         b = bch2_btree_node_get_noiter(c, tmp.k, btree, level, true);
 170         if (!IS_ERR_OR_NULL(b)) {
 171                 mutex_lock(&c->btree_cache.lock);
 172
 173                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 174
 175                 bkey_copy(&b->key, new);
 176                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 177                 BUG_ON(ret);
 178
 179                 mutex_unlock(&c->btree_cache.lock);
 180                 six_unlock_read(&b->c.lock);
 181         }
 182
 183         bch2_bkey_buf_exit(&tmp, c);
 184 }
 185
 186 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 187 {
 188         struct bkey_i_btree_ptr_v2 *new;
 189         int ret;
 190
 191         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 192         if (!new)
 193                 return -ENOMEM;
 194
 195         btree_ptr_to_v2(b, new);
 196         b->data->min_key        = new_min;
 197         new->v.min_key          = new_min;
 198         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 199
 200         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 201         if (ret) {
 202                 kfree(new);
 203                 return ret;
 204         }
 205
 206         bch2_btree_node_drop_keys_outside_node(b);
 207
 208         return 0;
 209 }
 210
 211 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 212 {
 213         struct bkey_i_btree_ptr_v2 *new;
 214         int ret;
 215
 216         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 217         if (ret)
 218                 return ret;
 219
 220         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 221         if (!new)
 222                 return -ENOMEM;
 223
 224         btree_ptr_to_v2(b, new);
 225         b->data->max_key        = new_max;
 226         new->k.p                = new_max;
 227         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 228
 229         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 230         if (ret) {
 231                 kfree(new);
 232                 return ret;
 233         }
 234
 235         bch2_btree_node_drop_keys_outside_node(b);
 236
 237         mutex_lock(&c->btree_cache.lock);
 238         bch2_btree_node_hash_remove(&c->btree_cache, b);
 239
 240         bkey_copy(&b->key, &new->k_i);
 241         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 242         BUG_ON(ret);
 243         mutex_unlock(&c->btree_cache.lock);
 244         return 0;
 245 }
 246
 247 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 248                                         struct btree *prev, struct btree *cur)
 249 {
 250         struct bpos expected_start = !prev
 251                 ? b->data->min_key
 252                 : bpos_successor(prev->key.k.p);
 253         char buf1[200], buf2[200];
 254         int ret = 0;
 255
 256         if (!prev) {
 257                 struct printbuf out = PBUF(buf1);
 258                 pr_buf(&out, "start of node: ");
 259                 bch2_bpos_to_text(&out, b->data->min_key);
 260         } else {
 261                 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
 262         }
 263
 264         bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key));
 265
 266         if (prev &&
 267             bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 268             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 269                 /* cur overwrites prev: */
 270
 271                 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
 272                                                  cur->data->min_key) >= 0, c,
 273                                 "btree node overwritten by next node at btree %s level %u:\n"
 274                                 "  node %s\n"
 275                                 "  next %s",
 276                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 277                                 buf1, buf2))
 278                         return DROP_PREV_NODE;
 279
 280                 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
 281                                                  bpos_predecessor(cur->data->min_key)), c,
 282                                 "btree node with incorrect max_key at btree %s level %u:\n"
 283                                 "  node %s\n"
 284                                 "  next %s",
 285                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 286                                 buf1, buf2))
 287                         ret = set_node_max(c, prev,
 288                                            bpos_predecessor(cur->data->min_key));
 289         } else {
 290                 /* prev overwrites cur: */
 291
 292                 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
 293                                                  cur->data->max_key) >= 0, c,
 294                                 "btree node overwritten by prev node at btree %s level %u:\n"
 295                                 "  prev %s\n"
 296                                 "  node %s",
 297                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 298                                 buf1, buf2))
 299                         return DROP_THIS_NODE;
 300
 301                 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 302                                 "btree node with incorrect min_key at btree %s level %u:\n"
 303                                 "  prev %s\n"
 304                                 "  node %s",
 305                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 306                                 buf1, buf2))
 307                     ret = set_node_min(c, cur, expected_start);
 308         }
 309 fsck_err:
 310         return ret;
 311 }
 312
 313 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 314                                  struct btree *child)
 315 {
 316         char buf1[200], buf2[200];
 317         int ret = 0;
 318
 319         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 320                         "btree node with incorrect max_key at btree %s level %u:\n"
 321                         "  %s\n"
 322                         "  expected %s",
 323                         bch2_btree_ids[b->c.btree_id], b->c.level,
 324                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
 325                         (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
 326                 ret = set_node_max(c, child, b->key.k.p);
 327                 if (ret)
 328                         return ret;
 329         }
 330 fsck_err:
 331         return ret;
 332 }
 333
 334 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 335 {
 336         struct btree_and_journal_iter iter;
 337         struct bkey_s_c k;
 338         struct bkey_buf prev_k, cur_k;
 339         struct btree *prev = NULL, *cur = NULL;
 340         bool have_child, dropped_children = false;
 341         char buf[200];
 342         int ret = 0;
 343
 344         if (!b->c.level)
 345                 return 0;
 346 again:
 347         prev = NULL;
 348         have_child = dropped_children = false;
 349         bch2_bkey_buf_init(&prev_k);
 350         bch2_bkey_buf_init(&cur_k);
 351         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 352
 353         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 354                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 355                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 356
 357                 bch2_btree_and_journal_iter_advance(&iter);
 358                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 359
 360                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 361                                         b->c.btree_id, b->c.level - 1,
 362                                         false);
 363                 ret = PTR_ERR_OR_ZERO(cur);
 364
 365                 if (mustfix_fsck_err_on(ret == -EIO, c,
 366                                 "Unreadable btree node at btree %s level %u:\n"
 367                                 "  %s",
 368                                 bch2_btree_ids[b->c.btree_id],
 369                                 b->c.level - 1,
 370                                 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
 371                         bch2_btree_node_evict(c, cur_k.k);
 372                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 373                                                       b->c.level, cur_k.k->k.p);
 374                         if (ret)
 375                                 break;
 376                         continue;
 377                 }
 378
 379                 if (ret) {
 380                         bch_err(c, "%s: error %i getting btree node",
 381                                 __func__, ret);
 382                         break;
 383                 }
 384
 385                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 386
 387                 if (ret == DROP_THIS_NODE) {
 388                         six_unlock_read(&cur->c.lock);
 389                         bch2_btree_node_evict(c, cur_k.k);
 390                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 391                                                       b->c.level, cur_k.k->k.p);
 392                         if (ret)
 393                                 break;
 394                         continue;
 395                 }
 396
 397                 if (prev)
 398                         six_unlock_read(&prev->c.lock);
 399                 prev = NULL;
 400
 401                 if (ret == DROP_PREV_NODE) {
 402                         bch2_btree_node_evict(c, prev_k.k);
 403                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 404                                                       b->c.level, prev_k.k->k.p);
 405                         if (ret)
 406                                 break;
 407
 408                         bch2_btree_and_journal_iter_exit(&iter);
 409                         bch2_bkey_buf_exit(&prev_k, c);
 410                         bch2_bkey_buf_exit(&cur_k, c);
 411                         goto again;
 412                 } else if (ret)
 413                         break;
 414
 415                 prev = cur;
 416                 cur = NULL;
 417                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 418         }
 419
 420         if (!ret && !IS_ERR_OR_NULL(prev)) {
 421                 BUG_ON(cur);
 422                 ret = btree_repair_node_end(c, b, prev);
 423         }
 424
 425         if (!IS_ERR_OR_NULL(prev))
 426                 six_unlock_read(&prev->c.lock);
 427         prev = NULL;
 428         if (!IS_ERR_OR_NULL(cur))
 429                 six_unlock_read(&cur->c.lock);
 430         cur = NULL;
 431
 432         if (ret)
 433                 goto err;
 434
 435         bch2_btree_and_journal_iter_exit(&iter);
 436         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 437
 438         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 439                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 440                 bch2_btree_and_journal_iter_advance(&iter);
 441
 442                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 443                                         b->c.btree_id, b->c.level - 1,
 444                                         false);
 445                 ret = PTR_ERR_OR_ZERO(cur);
 446
 447                 if (ret) {
 448                         bch_err(c, "%s: error %i getting btree node",
 449                                 __func__, ret);
 450                         goto err;
 451                 }
 452
 453                 ret = bch2_btree_repair_topology_recurse(c, cur);
 454                 six_unlock_read(&cur->c.lock);
 455                 cur = NULL;
 456
 457                 if (ret == DROP_THIS_NODE) {
 458                         bch2_btree_node_evict(c, cur_k.k);
 459                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 460                                                       b->c.level, cur_k.k->k.p);
 461                         dropped_children = true;
 462                 }
 463
 464                 if (ret)
 465                         goto err;
 466
 467                 have_child = true;
 468         }
 469
 470         if (mustfix_fsck_err_on(!have_child, c,
 471                         "empty interior btree node at btree %s level %u\n"
 472                         "  %s",
 473                         bch2_btree_ids[b->c.btree_id],
 474                         b->c.level,
 475                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
 476                 ret = DROP_THIS_NODE;
 477 err:
 478 fsck_err:
 479         if (!IS_ERR_OR_NULL(prev))
 480                 six_unlock_read(&prev->c.lock);
 481         if (!IS_ERR_OR_NULL(cur))
 482                 six_unlock_read(&cur->c.lock);
 483
 484         bch2_btree_and_journal_iter_exit(&iter);
 485         bch2_bkey_buf_exit(&prev_k, c);
 486         bch2_bkey_buf_exit(&cur_k, c);
 487
 488         if (!ret && dropped_children)
 489                 goto again;
 490
 491         return ret;
 492 }
 493
 494 static int bch2_repair_topology(struct bch_fs *c)
 495 {
 496         struct btree *b;
 497         unsigned i;
 498         int ret = 0;
 499
 500         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 501                 b = c->btree_roots[i].b;
 502                 if (btree_node_fake(b))
 503                         continue;
 504
 505                 six_lock_read(&b->c.lock, NULL, NULL);
 506                 ret = bch2_btree_repair_topology_recurse(c, b);
 507                 six_unlock_read(&b->c.lock);
 508
 509                 if (ret == DROP_THIS_NODE) {
 510                         bch_err(c, "empty btree root - repair unimplemented");
 511                         ret = FSCK_ERR_EXIT;
 512                 }
 513         }
 514
 515         return ret;
 516 }
 517
 518 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 519                                unsigned level, bool is_root,
 520                                struct bkey_s_c *k)
 521 {
 522         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 523         const union bch_extent_entry *entry;
 524         struct extent_ptr_decoded p = { 0 };
 525         bool do_update = false;
 526         char buf[200];
 527         int ret = 0;
 528
 529         /*
 530          * XXX
 531          * use check_bucket_ref here
 532          */
 533         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 534                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 535                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 536                 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr);
 537                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 538
 539                 if (fsck_err_on(!g->gen_valid, c,
 540                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 541                                 "while marking %s",
 542                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 543                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 544                                 p.ptr.gen,
 545                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 546                         if (!p.ptr.cached) {
 547                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 548                                 g2->gen_valid   = g->gen_valid          = true;
 549                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 550                         } else {
 551                                 do_update = true;
 552                         }
 553                 }
 554
 555                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 556                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 557                                 "while marking %s",
 558                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 559                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 560                                 p.ptr.gen, g->mark.gen,
 561                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 562                         if (!p.ptr.cached) {
 563                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 564                                 g2->gen_valid   = g->gen_valid  = true;
 565                                 g2->_mark.data_type             = 0;
 566                                 g2->_mark.dirty_sectors         = 0;
 567                                 g2->_mark.cached_sectors        = 0;
 568                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 569                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 570                         } else {
 571                                 do_update = true;
 572                         }
 573                 }
 574
 575                 if (fsck_err_on(gen_cmp(g->mark.gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
 576                                 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 577                                 "while marking %s",
 578                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->mark.gen,
 579                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 580                                 p.ptr.gen,
 581                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 582                         do_update = true;
 583
 584                 if (fsck_err_on(!p.ptr.cached &&
 585                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 586                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 587                                 "while marking %s",
 588                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 589                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 590                                 p.ptr.gen, g->mark.gen,
 591                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 592                         do_update = true;
 593
 594                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->mark.gen)
 595                         continue;
 596
 597                 if (fsck_err_on(g->mark.data_type &&
 598                                 g->mark.data_type != data_type, c,
 599                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 600                                 "while marking %s",
 601                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 602                                 bch2_data_types[g->mark.data_type],
 603                                 bch2_data_types[data_type],
 604                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 605                         if (data_type == BCH_DATA_btree) {
 606                                 g2->_mark.data_type     = g->_mark.data_type    = data_type;
 607                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 608                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 609                         } else {
 610                                 do_update = true;
 611                         }
 612                 }
 613
 614                 if (p.has_ec) {
 615                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 616
 617                         if (fsck_err_on(!m || !m->alive, c,
 618                                         "pointer to nonexistent stripe %llu\n"
 619                                         "while marking %s",
 620                                         (u64) p.ec.idx,
 621                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 622                                 do_update = true;
 623
 624                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 625                                         "pointer does not match stripe %llu\n"
 626                                         "while marking %s",
 627                                         (u64) p.ec.idx,
 628                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 629                                 do_update = true;
 630                 }
 631         }
 632
 633         if (do_update) {
 634                 struct bkey_ptrs ptrs;
 635                 union bch_extent_entry *entry;
 636                 struct bch_extent_ptr *ptr;
 637                 struct bkey_i *new;
 638
 639                 if (is_root) {
 640                         bch_err(c, "cannot update btree roots yet");
 641                         return -EINVAL;
 642                 }
 643
 644                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 645                 if (!new) {
 646                         bch_err(c, "%s: error allocating new key", __func__);
 647                         return -ENOMEM;
 648                 }
 649
 650                 bkey_reassemble(new, *k);
 651
 652                 if (level) {
 653                         /*
 654                          * We don't want to drop btree node pointers - if the
 655                          * btree node isn't there anymore, the read path will
 656                          * sort it out:
 657                          */
 658                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 659                         bkey_for_each_ptr(ptrs, ptr) {
 660                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 661                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 662
 663                                 ptr->gen = g->mark.gen;
 664                         }
 665                 } else {
 666                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 667                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 668                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 669                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 670
 671                                 (ptr->cached &&
 672                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 673                                 (!ptr->cached &&
 674                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 675                                 gen_cmp(g->mark.gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 676                                 (g->mark.data_type &&
 677                                  g->mark.data_type != data_type);
 678                         }));
 679 again:
 680                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 681                         bkey_extent_entry_for_each(ptrs, entry) {
 682                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 683                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 684                                                                         entry->stripe_ptr.idx);
 685                                         union bch_extent_entry *next_ptr;
 686
 687                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 688                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 689                                                         goto found;
 690                                         next_ptr = NULL;
 691 found:
 692                                         if (!next_ptr) {
 693                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 694                                                 continue;
 695                                         }
 696
 697                                         if (!m || !m->alive ||
 698                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 699                                                                        &next_ptr->ptr,
 700                                                                        m->sectors)) {
 701                                                 bch2_bkey_extent_entry_drop(new, entry);
 702                                                 goto again;
 703                                         }
 704                                 }
 705                         }
 706                 }
 707
 708                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 709                 if (ret) {
 710                         kfree(new);
 711                         return ret;
 712                 }
 713
 714                 if (level)
 715                         bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
 716
 717                 bch2_bkey_val_to_text(&PBUF(buf), c, *k);
 718                 bch_info(c, "updated %s", buf);
 719                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(new));
 720                 bch_info(c, "new key %s", buf);
 721                 *k = bkey_i_to_s_c(new);
 722         }
 723 fsck_err:
 724         return ret;
 725 }
 726
 727 /* marking of btree keys/nodes: */
 728
 729 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 730                             unsigned level, bool is_root,
 731                             struct bkey_s_c *k,
 732                             u8 *max_stale, bool initial)
 733 {
 734         struct bch_fs *c = trans->c;
 735         struct bkey_ptrs_c ptrs;
 736         const struct bch_extent_ptr *ptr;
 737         struct bkey deleted = KEY(0, 0, 0);
 738         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 739         unsigned flags =
 740                 BTREE_TRIGGER_GC|
 741                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 742         int ret = 0;
 743
 744         deleted.p = k->k->p;
 745
 746         if (initial) {
 747                 BUG_ON(bch2_journal_seq_verify &&
 748                        k->k->version.lo > journal_cur_seq(&c->journal));
 749
 750                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 751                 if (ret)
 752                         goto err;
 753
 754                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 755                                 "key version number higher than recorded: %llu > %llu",
 756                                 k->k->version.lo,
 757                                 atomic64_read(&c->key_version)))
 758                         atomic64_set(&c->key_version, k->k->version.lo);
 759         }
 760
 761         ptrs = bch2_bkey_ptrs_c(*k);
 762         bkey_for_each_ptr(ptrs, ptr) {
 763                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 764                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 765
 766                 if (gen_after(g->oldest_gen, ptr->gen))
 767                         g->oldest_gen = ptr->gen;
 768
 769                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 770         }
 771
 772         ret = bch2_mark_key(trans, old, *k, flags);
 773 fsck_err:
 774 err:
 775         if (ret)
 776                 bch_err(c, "%s: ret %i", __func__, ret);
 777         return ret;
 778 }
 779
 780 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, u8 *max_stale,
 781                               bool initial)
 782 {
 783         struct bch_fs *c = trans->c;
 784         struct btree_node_iter iter;
 785         struct bkey unpacked;
 786         struct bkey_s_c k;
 787         struct bkey_buf prev, cur;
 788         int ret = 0;
 789
 790         *max_stale = 0;
 791
 792         if (!btree_node_type_needs_gc(btree_node_type(b)))
 793                 return 0;
 794
 795         bch2_btree_node_iter_init_from_start(&iter, b);
 796         bch2_bkey_buf_init(&prev);
 797         bch2_bkey_buf_init(&cur);
 798         bkey_init(&prev.k->k);
 799
 800         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 801                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 802                                        &k, max_stale, initial);
 803                 if (ret)
 804                         break;
 805
 806                 bch2_btree_node_iter_advance(&iter, b);
 807
 808                 if (b->c.level) {
 809                         bch2_bkey_buf_reassemble(&cur, c, k);
 810
 811                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 812                                         bch2_btree_node_iter_end(&iter));
 813                         if (ret)
 814                                 break;
 815                 }
 816         }
 817
 818         bch2_bkey_buf_exit(&cur, c);
 819         bch2_bkey_buf_exit(&prev, c);
 820         return ret;
 821 }
 822
 823 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 824                          bool initial, bool metadata_only)
 825 {
 826         struct bch_fs *c = trans->c;
 827         struct btree_iter iter;
 828         struct btree *b;
 829         unsigned depth = metadata_only                  ? 1
 830                 : bch2_expensive_debug_checks           ? 0
 831                 : !btree_node_type_needs_gc(btree_id)   ? 1
 832                 : 0;
 833         u8 max_stale = 0;
 834         int ret = 0;
 835
 836         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 837
 838         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 839                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 840                 bch2_verify_btree_nr_keys(b);
 841
 842                 gc_pos_set(c, gc_pos_btree_node(b));
 843
 844                 ret = btree_gc_mark_node(trans, b, &max_stale, initial);
 845                 if (ret)
 846                         break;
 847
 848                 if (!initial) {
 849                         if (max_stale > 64)
 850                                 bch2_btree_node_rewrite(trans, &iter, b,
 851                                                 BTREE_INSERT_NOWAIT|
 852                                                 BTREE_INSERT_GC_LOCK_HELD);
 853                         else if (!bch2_btree_gc_rewrite_disabled &&
 854                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 855                                 bch2_btree_node_rewrite(trans, &iter,
 856                                                 b, BTREE_INSERT_NOWAIT|
 857                                                 BTREE_INSERT_GC_LOCK_HELD);
 858                 }
 859         }
 860         bch2_trans_iter_exit(trans, &iter);
 861
 862         if (ret)
 863                 return ret;
 864
 865         mutex_lock(&c->btree_root_lock);
 866         b = c->btree_roots[btree_id].b;
 867         if (!btree_node_fake(b)) {
 868                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 869
 870                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
 871                                        &k, &max_stale, initial);
 872         }
 873         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 874         mutex_unlock(&c->btree_root_lock);
 875
 876         return ret;
 877 }
 878
 879 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 880                                       unsigned target_depth)
 881 {
 882         struct bch_fs *c = trans->c;
 883         struct btree_and_journal_iter iter;
 884         struct bkey_s_c k;
 885         struct bkey_buf cur, prev;
 886         u8 max_stale = 0;
 887         char buf[200];
 888         int ret = 0;
 889
 890         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 891         bch2_bkey_buf_init(&prev);
 892         bch2_bkey_buf_init(&cur);
 893         bkey_init(&prev.k->k);
 894
 895         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 896                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 897                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 898
 899                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 900                                        &k, &max_stale, true);
 901                 if (ret) {
 902                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 903                         goto fsck_err;
 904                 }
 905
 906                 if (b->c.level) {
 907                         bch2_bkey_buf_reassemble(&cur, c, k);
 908                         k = bkey_i_to_s_c(cur.k);
 909
 910                         bch2_btree_and_journal_iter_advance(&iter);
 911
 912                         ret = bch2_gc_check_topology(c, b,
 913                                         &prev, cur,
 914                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 915                         if (ret)
 916                                 goto fsck_err;
 917                 } else {
 918                         bch2_btree_and_journal_iter_advance(&iter);
 919                 }
 920         }
 921
 922         if (b->c.level > target_depth) {
 923                 bch2_btree_and_journal_iter_exit(&iter);
 924                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 925
 926                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 927                         struct btree *child;
 928
 929                         bch2_bkey_buf_reassemble(&cur, c, k);
 930                         bch2_btree_and_journal_iter_advance(&iter);
 931
 932                         child = bch2_btree_node_get_noiter(c, cur.k,
 933                                                 b->c.btree_id, b->c.level - 1,
 934                                                 false);
 935                         ret = PTR_ERR_OR_ZERO(child);
 936
 937                         if (ret == -EIO) {
 938                                 bch2_topology_error(c);
 939
 940                                 if (__fsck_err(c,
 941                                           FSCK_CAN_FIX|
 942                                           FSCK_CAN_IGNORE|
 943                                           FSCK_NO_RATELIMIT,
 944                                           "Unreadable btree node at btree %s level %u:\n"
 945                                           "  %s",
 946                                           bch2_btree_ids[b->c.btree_id],
 947                                           b->c.level - 1,
 948                                           (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
 949                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 950                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 951                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 952                                         goto fsck_err;
 953                                 } else {
 954                                         /* Continue marking when opted to not
 955                                          * fix the error: */
 956                                         ret = 0;
 957                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 958                                         continue;
 959                                 }
 960                         } else if (ret) {
 961                                 bch_err(c, "%s: error %i getting btree node",
 962                                         __func__, ret);
 963                                 break;
 964                         }
 965
 966                         ret = bch2_gc_btree_init_recurse(trans, child,
 967                                                          target_depth);
 968                         six_unlock_read(&child->c.lock);
 969
 970                         if (ret)
 971                                 break;
 972                 }
 973         }
 974 fsck_err:
 975         bch2_bkey_buf_exit(&cur, c);
 976         bch2_bkey_buf_exit(&prev, c);
 977         bch2_btree_and_journal_iter_exit(&iter);
 978         return ret;
 979 }
 980
 981 static int bch2_gc_btree_init(struct btree_trans *trans,
 982                               enum btree_id btree_id,
 983                               bool metadata_only)
 984 {
 985         struct bch_fs *c = trans->c;
 986         struct btree *b;
 987         unsigned target_depth = metadata_only           ? 1
 988                 : bch2_expensive_debug_checks           ? 0
 989                 : !btree_node_type_needs_gc(btree_id)   ? 1
 990                 : 0;
 991         u8 max_stale = 0;
 992         char buf[100];
 993         int ret = 0;
 994
 995         b = c->btree_roots[btree_id].b;
 996
 997         if (btree_node_fake(b))
 998                 return 0;
 999
1000         six_lock_read(&b->c.lock, NULL, NULL);
1001         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1002                         "btree root with incorrect min_key: %s",
1003                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
1004                 bch_err(c, "repair unimplemented");
1005                 ret = FSCK_ERR_EXIT;
1006                 goto fsck_err;
1007         }
1008
1009         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1010                         "btree root with incorrect max_key: %s",
1011                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
1012                 bch_err(c, "repair unimplemented");
1013                 ret = FSCK_ERR_EXIT;
1014                 goto fsck_err;
1015         }
1016
1017         if (b->c.level >= target_depth)
1018                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1019
1020         if (!ret) {
1021                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1022
1023                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1024                                        &k, &max_stale, true);
1025         }
1026 fsck_err:
1027         six_unlock_read(&b->c.lock);
1028
1029         if (ret < 0)
1030                 bch_err(c, "%s: ret %i", __func__, ret);
1031         return ret;
1032 }
1033
1034 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1035 {
1036         return  (int) btree_id_to_gc_phase(l) -
1037                 (int) btree_id_to_gc_phase(r);
1038 }
1039
1040 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1041 {
1042         struct btree_trans trans;
1043         enum btree_id ids[BTREE_ID_NR];
1044         unsigned i;
1045         int ret = 0;
1046
1047         bch2_trans_init(&trans, c, 0, 0);
1048
1049         for (i = 0; i < BTREE_ID_NR; i++)
1050                 ids[i] = i;
1051         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1052
1053         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1054                 ret = initial
1055                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1056                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1057
1058         if (ret < 0)
1059                 bch_err(c, "%s: ret %i", __func__, ret);
1060
1061         bch2_trans_exit(&trans);
1062         return ret;
1063 }
1064
1065 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1066                                   u64 start, u64 end,
1067                                   enum bch_data_type type,
1068                                   unsigned flags)
1069 {
1070         u64 b = sector_to_bucket(ca, start);
1071
1072         do {
1073                 unsigned sectors =
1074                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1075
1076                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1077                                           gc_phase(GC_PHASE_SB), flags);
1078                 b++;
1079                 start += sectors;
1080         } while (start < end);
1081 }
1082
1083 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1084                                      unsigned flags)
1085 {
1086         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1087         unsigned i;
1088         u64 b;
1089
1090         for (i = 0; i < layout->nr_superblocks; i++) {
1091                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1092
1093                 if (offset == BCH_SB_SECTOR)
1094                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1095                                               BCH_DATA_sb, flags);
1096
1097                 mark_metadata_sectors(c, ca, offset,
1098                                       offset + (1 << layout->sb_max_size_bits),
1099                                       BCH_DATA_sb, flags);
1100         }
1101
1102         for (i = 0; i < ca->journal.nr; i++) {
1103                 b = ca->journal.buckets[i];
1104                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1105                                           ca->mi.bucket_size,
1106                                           gc_phase(GC_PHASE_SB), flags);
1107         }
1108 }
1109
1110 static void bch2_mark_superblocks(struct bch_fs *c)
1111 {
1112         struct bch_dev *ca;
1113         unsigned i;
1114
1115         mutex_lock(&c->sb_lock);
1116         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1117
1118         for_each_online_member(ca, c, i)
1119                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1120         mutex_unlock(&c->sb_lock);
1121 }
1122
1123 #if 0
1124 /* Also see bch2_pending_btree_node_free_insert_done() */
1125 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1126 {
1127         struct btree_update *as;
1128         struct pending_btree_node_free *d;
1129
1130         mutex_lock(&c->btree_interior_update_lock);
1131         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1132
1133         for_each_pending_btree_node_free(c, as, d)
1134                 if (d->index_update_done)
1135                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1136
1137         mutex_unlock(&c->btree_interior_update_lock);
1138 }
1139 #endif
1140
1141 static void bch2_gc_free(struct bch_fs *c)
1142 {
1143         struct bch_dev *ca;
1144         unsigned i;
1145
1146         genradix_free(&c->reflink_gc_table);
1147         genradix_free(&c->gc_stripes);
1148
1149         for_each_member_device(ca, c, i) {
1150                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1151                         sizeof(struct bucket_array) +
1152                         ca->mi.nbuckets * sizeof(struct bucket));
1153                 ca->buckets[1] = NULL;
1154
1155                 free_percpu(ca->usage_gc);
1156                 ca->usage_gc = NULL;
1157         }
1158
1159         free_percpu(c->usage_gc);
1160         c->usage_gc = NULL;
1161 }
1162
1163 static int bch2_gc_done(struct bch_fs *c,
1164                         bool initial, bool metadata_only)
1165 {
1166         struct bch_dev *ca = NULL;
1167         bool verify = !metadata_only && (!initial ||
1168                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1169         unsigned i, dev;
1170         int ret = 0;
1171
1172 #define copy_field(_f, _msg, ...)                                       \
1173         if (dst->_f != src->_f) {                                       \
1174                 if (verify)                                             \
1175                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1176                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1177                 dst->_f = src->_f;                                      \
1178                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1179         }
1180 #define copy_stripe_field(_f, _msg, ...)                                \
1181         if (dst->_f != src->_f) {                                       \
1182                 if (verify)                                             \
1183                         fsck_err(c, "stripe %zu has wrong "_msg         \
1184                                 ": got %u, should be %u",               \
1185                                 iter.pos, ##__VA_ARGS__,                \
1186                                 dst->_f, src->_f);                      \
1187                 dst->_f = src->_f;                                      \
1188                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1189         }
1190 #define copy_bucket_field(_f)                                           \
1191         if (dst->b[b]._f != src->b[b]._f) {                             \
1192                 if (verify)                                             \
1193                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
1194                                 ": got %u, should be %u", dev, b,       \
1195                                 dst->b[b].mark.gen,                     \
1196                                 bch2_data_types[dst->b[b].mark.data_type],\
1197                                 dst->b[b]._f, src->b[b]._f);            \
1198                 dst->b[b]._f = src->b[b]._f;                            \
1199                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1200         }
1201 #define copy_dev_field(_f, _msg, ...)                                   \
1202         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1203 #define copy_fs_field(_f, _msg, ...)                                    \
1204         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1205
1206         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1207                 bch2_fs_usage_acc_to_base(c, i);
1208
1209         for_each_member_device(ca, c, dev) {
1210                 struct bucket_array *dst = __bucket_array(ca, 0);
1211                 struct bucket_array *src = __bucket_array(ca, 1);
1212                 size_t b;
1213
1214                 for (b = 0; b < src->nbuckets; b++) {
1215                         copy_bucket_field(_mark.gen);
1216                         copy_bucket_field(_mark.data_type);
1217                         copy_bucket_field(_mark.stripe);
1218                         copy_bucket_field(_mark.dirty_sectors);
1219                         copy_bucket_field(_mark.cached_sectors);
1220                         copy_bucket_field(stripe_redundancy);
1221                         copy_bucket_field(stripe);
1222
1223                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
1224                 }
1225
1226                 {
1227                         struct bch_dev_usage *dst = ca->usage_base;
1228                         struct bch_dev_usage *src = (void *)
1229                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1230                                                      dev_usage_u64s());
1231
1232                         copy_dev_field(buckets_ec,              "buckets_ec");
1233                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1234
1235                         for (i = 0; i < BCH_DATA_NR; i++) {
1236                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1237                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1238                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1239                         }
1240                 }
1241         };
1242
1243         {
1244                 unsigned nr = fs_usage_u64s(c);
1245                 struct bch_fs_usage *dst = c->usage_base;
1246                 struct bch_fs_usage *src = (void *)
1247                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1248
1249                 copy_fs_field(hidden,           "hidden");
1250                 copy_fs_field(btree,            "btree");
1251
1252                 if (!metadata_only) {
1253                         copy_fs_field(data,     "data");
1254                         copy_fs_field(cached,   "cached");
1255                         copy_fs_field(reserved, "reserved");
1256                         copy_fs_field(nr_inodes,"nr_inodes");
1257
1258                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1259                                 copy_fs_field(persistent_reserved[i],
1260                                               "persistent_reserved[%i]", i);
1261                 }
1262
1263                 for (i = 0; i < c->replicas.nr; i++) {
1264                         struct bch_replicas_entry *e =
1265                                 cpu_replicas_entry(&c->replicas, i);
1266                         char buf[80];
1267
1268                         if (metadata_only &&
1269                             (e->data_type == BCH_DATA_user ||
1270                              e->data_type == BCH_DATA_cached))
1271                                 continue;
1272
1273                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1274
1275                         copy_fs_field(replicas[i], "%s", buf);
1276                 }
1277         }
1278
1279 #undef copy_fs_field
1280 #undef copy_dev_field
1281 #undef copy_bucket_field
1282 #undef copy_stripe_field
1283 #undef copy_field
1284 fsck_err:
1285         if (ca)
1286                 percpu_ref_put(&ca->ref);
1287         if (ret)
1288                 bch_err(c, "%s: ret %i", __func__, ret);
1289         return ret;
1290 }
1291
1292 static int bch2_gc_start(struct bch_fs *c,
1293                          bool metadata_only)
1294 {
1295         struct bch_dev *ca = NULL;
1296         unsigned i;
1297
1298         BUG_ON(c->usage_gc);
1299
1300         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1301                                          sizeof(u64), GFP_KERNEL);
1302         if (!c->usage_gc) {
1303                 bch_err(c, "error allocating c->usage_gc");
1304                 return -ENOMEM;
1305         }
1306
1307         for_each_member_device(ca, c, i) {
1308                 BUG_ON(ca->buckets[1]);
1309                 BUG_ON(ca->usage_gc);
1310
1311                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1312                                 ca->mi.nbuckets * sizeof(struct bucket),
1313                                 GFP_KERNEL|__GFP_ZERO);
1314                 if (!ca->buckets[1]) {
1315                         percpu_ref_put(&ca->ref);
1316                         bch_err(c, "error allocating ca->buckets[gc]");
1317                         return -ENOMEM;
1318                 }
1319
1320                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1321                 if (!ca->usage_gc) {
1322                         bch_err(c, "error allocating ca->usage_gc");
1323                         percpu_ref_put(&ca->ref);
1324                         return -ENOMEM;
1325                 }
1326         }
1327
1328         percpu_down_write(&c->mark_lock);
1329
1330         for_each_member_device(ca, c, i) {
1331                 struct bucket_array *dst = __bucket_array(ca, 1);
1332                 struct bucket_array *src = __bucket_array(ca, 0);
1333                 size_t b;
1334
1335                 dst->first_bucket       = src->first_bucket;
1336                 dst->nbuckets           = src->nbuckets;
1337
1338                 for (b = 0; b < src->nbuckets; b++) {
1339                         struct bucket *d = &dst->b[b];
1340                         struct bucket *s = &src->b[b];
1341
1342                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1343                         d->gen_valid = s->gen_valid;
1344
1345                         if (metadata_only &&
1346                             (s->mark.data_type == BCH_DATA_user ||
1347                              s->mark.data_type == BCH_DATA_cached))
1348                                 d->_mark = s->mark;
1349                 }
1350         };
1351
1352         percpu_up_write(&c->mark_lock);
1353
1354         return 0;
1355 }
1356
1357 static void bch2_gc_alloc_reset(struct bch_fs *c, bool initial, bool metadata_only)
1358 {
1359         struct bch_dev *ca;
1360         unsigned i;
1361
1362         for_each_member_device(ca, c, i) {
1363                 struct bucket_array *buckets = __bucket_array(ca, true);
1364                 struct bucket *g;
1365
1366                 for_each_bucket(g, buckets) {
1367                         if (metadata_only &&
1368                             (g->mark.data_type == BCH_DATA_user ||
1369                              g->mark.data_type == BCH_DATA_cached ||
1370                              g->mark.data_type == BCH_DATA_parity))
1371                                 continue;
1372                         g->_mark.dirty_sectors = 0;
1373                         g->_mark.cached_sectors = 0;
1374                 }
1375         };
1376 }
1377
1378 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1379                                 bool metadata_only)
1380 {
1381         struct btree_trans trans;
1382         struct btree_iter iter;
1383         struct bkey_s_c k;
1384         struct reflink_gc *r;
1385         size_t idx = 0;
1386         char buf[200];
1387         int ret = 0;
1388
1389         if (metadata_only)
1390                 return 0;
1391
1392         bch2_trans_init(&trans, c, 0, 0);
1393
1394         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1395                            BTREE_ITER_PREFETCH, k, ret) {
1396                 const __le64 *refcount = bkey_refcount_c(k);
1397
1398                 if (!refcount)
1399                         continue;
1400
1401                 r = genradix_ptr(&c->reflink_gc_table, idx++);
1402                 if (!r ||
1403                     r->offset != k.k->p.offset ||
1404                     r->size != k.k->size) {
1405                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1406                         ret = -EINVAL;
1407                         break;
1408                 }
1409
1410                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1411                                 "reflink key has wrong refcount:\n"
1412                                 "  %s\n"
1413                                 "  should be %u",
1414                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1415                                 r->refcount)) {
1416                         struct bkey_i *new;
1417
1418                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1419                         if (!new) {
1420                                 ret = -ENOMEM;
1421                                 break;
1422                         }
1423
1424                         bkey_reassemble(new, k);
1425
1426                         if (!r->refcount)
1427                                 new->k.type = KEY_TYPE_deleted;
1428                         else
1429                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1430
1431                         ret = initial
1432                                ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, new)
1433                                : __bch2_trans_do(&trans, NULL, NULL, 0,
1434                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1435                         kfree(new);
1436
1437                         if (ret)
1438                                 break;
1439                 }
1440         }
1441 fsck_err:
1442         bch2_trans_iter_exit(&trans, &iter);
1443         c->reflink_gc_nr = 0;
1444         bch2_trans_exit(&trans);
1445         return ret;
1446 }
1447
1448 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1449                                  bool metadata_only)
1450 {
1451         struct btree_trans trans;
1452         struct btree_iter iter;
1453         struct bkey_s_c k;
1454         struct reflink_gc *r;
1455         int ret = 0;
1456
1457         if (metadata_only)
1458                 return 0;
1459
1460         bch2_trans_init(&trans, c, 0, 0);
1461         c->reflink_gc_nr = 0;
1462
1463         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1464                            BTREE_ITER_PREFETCH, k, ret) {
1465                 const __le64 *refcount = bkey_refcount_c(k);
1466
1467                 if (!refcount)
1468                         continue;
1469
1470                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1471                                        GFP_KERNEL);
1472                 if (!r) {
1473                         ret = -ENOMEM;
1474                         break;
1475                 }
1476
1477                 r->offset       = k.k->p.offset;
1478                 r->size         = k.k->size;
1479                 r->refcount     = 0;
1480         }
1481         bch2_trans_iter_exit(&trans, &iter);
1482
1483         bch2_trans_exit(&trans);
1484         return ret;
1485 }
1486
1487 static void bch2_gc_reflink_reset(struct bch_fs *c, bool initial,
1488                                   bool metadata_only)
1489 {
1490         struct genradix_iter iter;
1491         struct reflink_gc *r;
1492
1493         genradix_for_each(&c->reflink_gc_table, iter, r)
1494                 r->refcount = 0;
1495 }
1496
1497 static int bch2_gc_stripes_done(struct bch_fs *c, bool initial,
1498                                 bool metadata_only)
1499 {
1500         struct btree_trans trans;
1501         struct btree_iter iter;
1502         struct bkey_s_c k;
1503         struct gc_stripe *m;
1504         const struct bch_stripe *s;
1505         char buf[200];
1506         unsigned i;
1507         int ret = 0;
1508
1509         if (metadata_only)
1510                 return 0;
1511
1512         bch2_trans_init(&trans, c, 0, 0);
1513
1514         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1515                            BTREE_ITER_PREFETCH, k, ret) {
1516                 if (k.k->type != KEY_TYPE_stripe)
1517                         continue;
1518
1519                 s = bkey_s_c_to_stripe(k).v;
1520                 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1521
1522                 for (i = 0; i < s->nr_blocks; i++)
1523                         if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1524                                 goto inconsistent;
1525                 continue;
1526 inconsistent:
1527                 if (fsck_err_on(true, c,
1528                                 "stripe has wrong block sector count %u:\n"
1529                                 "  %s\n"
1530                                 "  should be %u", i,
1531                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1532                                 m ? m->block_sectors[i] : 0)) {
1533                         struct bkey_i_stripe *new;
1534
1535                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1536                         if (!new) {
1537                                 ret = -ENOMEM;
1538                                 break;
1539                         }
1540
1541                         bkey_reassemble(&new->k_i, k);
1542
1543                         for (i = 0; i < new->v.nr_blocks; i++)
1544                                 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1545
1546                         ret = initial
1547                                 ? bch2_journal_key_insert(c, BTREE_ID_stripes, 0, &new->k_i)
1548                                 : __bch2_trans_do(&trans, NULL, NULL, 0,
1549                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1550                         kfree(new);
1551                 }
1552         }
1553 fsck_err:
1554         bch2_trans_iter_exit(&trans, &iter);
1555
1556         bch2_trans_exit(&trans);
1557         return ret;
1558 }
1559
1560 static void bch2_gc_stripes_reset(struct bch_fs *c, bool initial,
1561                                 bool metadata_only)
1562 {
1563         genradix_free(&c->gc_stripes);
1564 }
1565
1566 /**
1567  * bch2_gc - walk _all_ references to buckets, and recompute them:
1568  *
1569  * Order matters here:
1570  *  - Concurrent GC relies on the fact that we have a total ordering for
1571  *    everything that GC walks - see  gc_will_visit_node(),
1572  *    gc_will_visit_root()
1573  *
1574  *  - also, references move around in the course of index updates and
1575  *    various other crap: everything needs to agree on the ordering
1576  *    references are allowed to move around in - e.g., we're allowed to
1577  *    start with a reference owned by an open_bucket (the allocator) and
1578  *    move it to the btree, but not the reverse.
1579  *
1580  *    This is necessary to ensure that gc doesn't miss references that
1581  *    move around - if references move backwards in the ordering GC
1582  *    uses, GC could skip past them
1583  */
1584 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1585 {
1586         struct bch_dev *ca;
1587         u64 start_time = local_clock();
1588         unsigned i, iter = 0;
1589         int ret;
1590
1591         lockdep_assert_held(&c->state_lock);
1592         trace_gc_start(c);
1593
1594         down_write(&c->gc_lock);
1595
1596         /* flush interior btree updates: */
1597         closure_wait_event(&c->btree_interior_update_wait,
1598                            !bch2_btree_interior_updates_nr_pending(c));
1599
1600         ret   = bch2_gc_start(c, metadata_only) ?:
1601                 bch2_gc_reflink_start(c, initial, metadata_only);
1602         if (ret)
1603                 goto out;
1604 again:
1605         gc_pos_set(c, gc_phase(GC_PHASE_START));
1606
1607         bch2_mark_superblocks(c);
1608
1609         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1610             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1611             c->opts.fix_errors != FSCK_OPT_NO) {
1612                 bch_info(c, "starting topology repair pass");
1613                 ret = bch2_repair_topology(c);
1614                 if (ret)
1615                         goto out;
1616                 bch_info(c, "topology repair pass done");
1617
1618                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1619         }
1620
1621         ret = bch2_gc_btrees(c, initial, metadata_only);
1622
1623         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1624             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1625             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1626                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1627                 ret = 0;
1628         }
1629
1630         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1631                 ret = FSCK_ERR_EXIT;
1632
1633         if (ret)
1634                 goto out;
1635
1636 #if 0
1637         bch2_mark_pending_btree_node_frees(c);
1638 #endif
1639         c->gc_count++;
1640
1641         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1642             (!iter && bch2_test_restart_gc)) {
1643                 if (iter++ > 2) {
1644                         bch_info(c, "Unable to fix bucket gens, looping");
1645                         ret = -EINVAL;
1646                         goto out;
1647                 }
1648
1649                 /*
1650                  * XXX: make sure gens we fixed got saved
1651                  */
1652                 bch_info(c, "Second GC pass needed, restarting:");
1653                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1654                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1655
1656                 bch2_gc_stripes_reset(c, initial, metadata_only);
1657                 bch2_gc_alloc_reset(c, initial, metadata_only);
1658                 bch2_gc_reflink_reset(c, initial, metadata_only);
1659
1660                 /* flush fsck errors, reset counters */
1661                 bch2_flush_fsck_errs(c);
1662                 goto again;
1663         }
1664 out:
1665         if (!ret) {
1666                 bch2_journal_block(&c->journal);
1667
1668                 percpu_down_write(&c->mark_lock);
1669                 ret   = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1670                         bch2_gc_stripes_done(c, initial, metadata_only) ?:
1671                         bch2_gc_done(c, initial, metadata_only);
1672
1673                 bch2_journal_unblock(&c->journal);
1674         } else {
1675                 percpu_down_write(&c->mark_lock);
1676         }
1677
1678         /* Indicates that gc is no longer in progress: */
1679         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1680
1681         bch2_gc_free(c);
1682         percpu_up_write(&c->mark_lock);
1683
1684         up_write(&c->gc_lock);
1685
1686         trace_gc_end(c);
1687         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1688
1689         /*
1690          * Wake up allocator in case it was waiting for buckets
1691          * because of not being able to inc gens
1692          */
1693         for_each_member_device(ca, c, i)
1694                 bch2_wake_allocator(ca);
1695
1696         /*
1697          * At startup, allocations can happen directly instead of via the
1698          * allocator thread - issue wakeup in case they blocked on gc_lock:
1699          */
1700         closure_wake_up(&c->freelist_wait);
1701         return ret;
1702 }
1703
1704 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1705 {
1706         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1707         const struct bch_extent_ptr *ptr;
1708
1709         percpu_down_read(&c->mark_lock);
1710         bkey_for_each_ptr(ptrs, ptr) {
1711                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1712                 struct bucket *g = PTR_BUCKET(ca, ptr);
1713
1714                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1715                         percpu_up_read(&c->mark_lock);
1716                         return true;
1717                 }
1718         }
1719
1720         bkey_for_each_ptr(ptrs, ptr) {
1721                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1722                 struct bucket *g = PTR_BUCKET(ca, ptr);
1723
1724                 if (gen_after(g->gc_gen, ptr->gen))
1725                         g->gc_gen = ptr->gen;
1726         }
1727         percpu_up_read(&c->mark_lock);
1728
1729         return false;
1730 }
1731
1732 /*
1733  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1734  * node pointers currently never have cached pointers that can become stale:
1735  */
1736 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1737 {
1738         struct btree_trans trans;
1739         struct btree_iter iter;
1740         struct bkey_s_c k;
1741         struct bkey_buf sk;
1742         int ret = 0, commit_err = 0;
1743
1744         bch2_bkey_buf_init(&sk);
1745         bch2_trans_init(&trans, c, 0, 0);
1746
1747         bch2_trans_iter_init(&trans, &iter, btree_id, POS_MIN,
1748                              BTREE_ITER_PREFETCH|
1749                              BTREE_ITER_NOT_EXTENTS|
1750                              BTREE_ITER_ALL_SNAPSHOTS);
1751
1752         while ((bch2_trans_begin(&trans),
1753                 k = bch2_btree_iter_peek(&iter)).k) {
1754                 ret = bkey_err(k);
1755
1756                 if (ret == -EINTR)
1757                         continue;
1758                 if (ret)
1759                         break;
1760
1761                 c->gc_gens_pos = iter.pos;
1762
1763                 if (gc_btree_gens_key(c, k) && !commit_err) {
1764                         bch2_bkey_buf_reassemble(&sk, c, k);
1765                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1766
1767                         commit_err =
1768                                 bch2_trans_update(&trans, &iter, sk.k, 0) ?:
1769                                 bch2_trans_commit(&trans, NULL, NULL,
1770                                                        BTREE_INSERT_NOWAIT|
1771                                                        BTREE_INSERT_NOFAIL);
1772                         if (commit_err == -EINTR) {
1773                                 commit_err = 0;
1774                                 continue;
1775                         }
1776                 }
1777
1778                 bch2_btree_iter_advance(&iter);
1779         }
1780         bch2_trans_iter_exit(&trans, &iter);
1781
1782         bch2_trans_exit(&trans);
1783         bch2_bkey_buf_exit(&sk, c);
1784
1785         return ret;
1786 }
1787
1788 int bch2_gc_gens(struct bch_fs *c)
1789 {
1790         struct bch_dev *ca;
1791         struct bucket_array *buckets;
1792         struct bucket *g;
1793         u64 start_time = local_clock();
1794         unsigned i;
1795         int ret;
1796
1797         /*
1798          * Ideally we would be using state_lock and not gc_lock here, but that
1799          * introduces a deadlock in the RO path - we currently take the state
1800          * lock at the start of going RO, thus the gc thread may get stuck:
1801          */
1802         down_read(&c->gc_lock);
1803
1804         for_each_member_device(ca, c, i) {
1805                 down_read(&ca->bucket_lock);
1806                 buckets = bucket_array(ca);
1807
1808                 for_each_bucket(g, buckets)
1809                         g->gc_gen = g->mark.gen;
1810                 up_read(&ca->bucket_lock);
1811         }
1812
1813         for (i = 0; i < BTREE_ID_NR; i++)
1814                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1815                         c->gc_gens_btree = i;
1816                         c->gc_gens_pos = POS_MIN;
1817                         ret = bch2_gc_btree_gens(c, i);
1818                         if (ret) {
1819                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1820                                 goto err;
1821                         }
1822                 }
1823
1824         for_each_member_device(ca, c, i) {
1825                 down_read(&ca->bucket_lock);
1826                 buckets = bucket_array(ca);
1827
1828                 for_each_bucket(g, buckets)
1829                         g->oldest_gen = g->gc_gen;
1830                 up_read(&ca->bucket_lock);
1831         }
1832
1833         c->gc_gens_btree        = 0;
1834         c->gc_gens_pos          = POS_MIN;
1835
1836         c->gc_count++;
1837
1838         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1839 err:
1840         up_read(&c->gc_lock);
1841         return ret;
1842 }
1843
1844 static int bch2_gc_thread(void *arg)
1845 {
1846         struct bch_fs *c = arg;
1847         struct io_clock *clock = &c->io_clock[WRITE];
1848         unsigned long last = atomic64_read(&clock->now);
1849         unsigned last_kick = atomic_read(&c->kick_gc);
1850         int ret;
1851
1852         set_freezable();
1853
1854         while (1) {
1855                 while (1) {
1856                         set_current_state(TASK_INTERRUPTIBLE);
1857
1858                         if (kthread_should_stop()) {
1859                                 __set_current_state(TASK_RUNNING);
1860                                 return 0;
1861                         }
1862
1863                         if (atomic_read(&c->kick_gc) != last_kick)
1864                                 break;
1865
1866                         if (c->btree_gc_periodic) {
1867                                 unsigned long next = last + c->capacity / 16;
1868
1869                                 if (atomic64_read(&clock->now) >= next)
1870                                         break;
1871
1872                                 bch2_io_clock_schedule_timeout(clock, next);
1873                         } else {
1874                                 schedule();
1875                         }
1876
1877                         try_to_freeze();
1878                 }
1879                 __set_current_state(TASK_RUNNING);
1880
1881                 last = atomic64_read(&clock->now);
1882                 last_kick = atomic_read(&c->kick_gc);
1883
1884                 /*
1885                  * Full gc is currently incompatible with btree key cache:
1886                  */
1887 #if 0
1888                 ret = bch2_gc(c, false, false);
1889 #else
1890                 ret = bch2_gc_gens(c);
1891 #endif
1892                 if (ret < 0)
1893                         bch_err(c, "btree gc failed: %i", ret);
1894
1895                 debug_check_no_locks_held();
1896         }
1897
1898         return 0;
1899 }
1900
1901 void bch2_gc_thread_stop(struct bch_fs *c)
1902 {
1903         struct task_struct *p;
1904
1905         p = c->gc_thread;
1906         c->gc_thread = NULL;
1907
1908         if (p) {
1909                 kthread_stop(p);
1910                 put_task_struct(p);
1911         }
1912 }
1913
1914 int bch2_gc_thread_start(struct bch_fs *c)
1915 {
1916         struct task_struct *p;
1917
1918         if (c->gc_thread)
1919                 return 0;
1920
1921         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1922         if (IS_ERR(p)) {
1923                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1924                 return PTR_ERR(p);
1925         }
1926
1927         get_task_struct(p);
1928         c->gc_thread = p;
1929         wake_up_process(p);
1930         return 0;
1931 }