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 "replicas.h"
  27 #include "super-io.h"
  28
  29 #include <linux/slab.h>
  30 #include <linux/bitops.h>
  31 #include <linux/freezer.h>
  32 #include <linux/kthread.h>
  33 #include <linux/preempt.h>
  34 #include <linux/rcupdate.h>
  35 #include <linux/sched/task.h>
  36 #include <trace/events/bcachefs.h>
  37
  38 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  39 {
  40         preempt_disable();
  41         write_seqcount_begin(&c->gc_pos_lock);
  42         c->gc_pos = new_pos;
  43         write_seqcount_end(&c->gc_pos_lock);
  44         preempt_enable();
  45 }
  46
  47 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  48 {
  49         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  50         __gc_pos_set(c, new_pos);
  51 }
  52
  53 /*
  54  * Missing: if an interior btree node is empty, we need to do something -
  55  * perhaps just kill it
  56  */
  57 static int bch2_gc_check_topology(struct bch_fs *c,
  58                                   struct btree *b,
  59                                   struct bkey_buf *prev,
  60                                   struct bkey_buf cur,
  61                                   bool is_last)
  62 {
  63         struct bpos node_start  = b->data->min_key;
  64         struct bpos node_end    = b->data->max_key;
  65         struct bpos expected_start = bkey_deleted(&prev->k->k)
  66                 ? node_start
  67                 : bpos_successor(prev->k->k.p);
  68         char buf1[200], buf2[200];
  69         bool update_min = false;
  70         bool update_max = false;
  71         int ret = 0;
  72
  73         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  74                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  75
  76                 if (bkey_deleted(&prev->k->k)) {
  77                         struct printbuf out = PBUF(buf1);
  78                         pr_buf(&out, "start of node: ");
  79                         bch2_bpos_to_text(&out, node_start);
  80                 } else {
  81                         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
  82                 }
  83
  84                 if (fsck_err_on(bpos_cmp(expected_start, bp->v.min_key), c,
  85                                 "btree node with incorrect min_key at btree %s level %u:\n"
  86                                 "  prev %s\n"
  87                                 "  cur %s",
  88                                 bch2_btree_ids[b->c.btree_id], b->c.level,
  89                                 buf1,
  90                                 (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2)))
  91                         update_min = true;
  92         }
  93
  94         if (fsck_err_on(is_last &&
  95                         bpos_cmp(cur.k->k.p, node_end), c,
  96                         "btree node with incorrect max_key at btree %s level %u:\n"
  97                         "  %s\n"
  98                         "  expected %s",
  99                         bch2_btree_ids[b->c.btree_id], b->c.level,
 100                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
 101                         (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2)))
 102                 update_max = true;
 103
 104         bch2_bkey_buf_copy(prev, c, cur.k);
 105
 106         if (update_min || update_max) {
 107                 struct bkey_i *new;
 108                 struct bkey_i_btree_ptr_v2 *bp = NULL;
 109                 struct btree *n;
 110
 111                 if (update_max) {
 112                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 113                                                       b->c.level, cur.k->k.p);
 114                         if (ret)
 115                                 return ret;
 116                 }
 117
 118                 new = kmalloc(bkey_bytes(&cur.k->k), GFP_KERNEL);
 119                 if (!new) {
 120                         bch_err(c, "%s: error allocating new key", __func__);
 121                         return -ENOMEM;
 122                 }
 123
 124                 bkey_copy(new, cur.k);
 125
 126                 if (new->k.type == KEY_TYPE_btree_ptr_v2)
 127                         bp = bkey_i_to_btree_ptr_v2(new);
 128
 129                 if (update_min)
 130                         bp->v.min_key = expected_start;
 131                 if (update_max)
 132                         new->k.p = node_end;
 133                 if (bp)
 134                         SET_BTREE_PTR_RANGE_UPDATED(&bp->v, true);
 135
 136                 ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level, new);
 137                 if (ret) {
 138                         kfree(new);
 139                         return ret;
 140                 }
 141
 142                 n = bch2_btree_node_get_noiter(c, cur.k, b->c.btree_id,
 143                                                b->c.level - 1, true);
 144                 if (n) {
 145                         mutex_lock(&c->btree_cache.lock);
 146                         bch2_btree_node_hash_remove(&c->btree_cache, n);
 147
 148                         bkey_copy(&n->key, new);
 149                         if (update_min)
 150                                 n->data->min_key = expected_start;
 151                         if (update_max)
 152                                 n->data->max_key = node_end;
 153
 154                         ret = __bch2_btree_node_hash_insert(&c->btree_cache, n);
 155                         BUG_ON(ret);
 156                         mutex_unlock(&c->btree_cache.lock);
 157                         six_unlock_read(&n->c.lock);
 158                 }
 159         }
 160 fsck_err:
 161         return ret;
 162 }
 163
 164 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 165                                unsigned level, bool is_root,
 166                                struct bkey_s_c *k)
 167 {
 168         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 169         const union bch_extent_entry *entry;
 170         struct extent_ptr_decoded p = { 0 };
 171         bool do_update = false;
 172         int ret = 0;
 173
 174         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 175                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 176                 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
 177                 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
 178
 179                 if (fsck_err_on(!g->gen_valid, c,
 180                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree",
 181                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 182                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 183                                 p.ptr.gen)) {
 184                         if (p.ptr.cached) {
 185                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 186                                 g2->gen_valid   = g->gen_valid          = true;
 187                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 188                         } else {
 189                                 do_update = true;
 190                         }
 191                 }
 192
 193                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 194                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u",
 195                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 196                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 197                                 p.ptr.gen, g->mark.gen)) {
 198                         if (p.ptr.cached) {
 199                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 200                                 g2->gen_valid   = g->gen_valid  = true;
 201                                 g2->_mark.data_type             = 0;
 202                                 g2->_mark.dirty_sectors         = 0;
 203                                 g2->_mark.cached_sectors        = 0;
 204                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 205                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 206                         } else {
 207                                 do_update = true;
 208                         }
 209                 }
 210
 211                 if (fsck_err_on(!p.ptr.cached &&
 212                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 213                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u",
 214                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 215                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 216                                 p.ptr.gen, g->mark.gen))
 217                         do_update = true;
 218
 219                 if (p.has_ec) {
 220                         struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
 221
 222                         if (fsck_err_on(!m || !m->alive, c,
 223                                         "pointer to nonexistent stripe %llu",
 224                                         (u64) p.ec.idx))
 225                                 do_update = true;
 226
 227                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 228                                         "pointer does not match stripe %llu",
 229                                         (u64) p.ec.idx))
 230                                 do_update = true;
 231                 }
 232         }
 233
 234         if (do_update) {
 235                 struct bkey_ptrs ptrs;
 236                 union bch_extent_entry *entry;
 237                 struct bch_extent_ptr *ptr;
 238                 struct bkey_i *new;
 239
 240                 if (is_root) {
 241                         bch_err(c, "cannot update btree roots yet");
 242                         return -EINVAL;
 243                 }
 244
 245                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 246                 if (!new) {
 247                         bch_err(c, "%s: error allocating new key", __func__);
 248                         return -ENOMEM;
 249                 }
 250
 251                 bkey_reassemble(new, *k);
 252
 253                 bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 254                         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 255                         struct bucket *g = PTR_BUCKET(ca, ptr, true);
 256
 257                         (ptr->cached &&
 258                          (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 259                         (!ptr->cached &&
 260                          gen_cmp(ptr->gen, g->mark.gen) < 0);
 261                 }));
 262 again:
 263                 ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 264                 bkey_extent_entry_for_each(ptrs, entry) {
 265                         if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 266                                 struct stripe *m = genradix_ptr(&c->stripes[true],
 267                                                                 entry->stripe_ptr.idx);
 268                                 union bch_extent_entry *next_ptr;
 269
 270                                 bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 271                                         if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 272                                                 goto found;
 273                                 next_ptr = NULL;
 274 found:
 275                                 if (!next_ptr) {
 276                                         bch_err(c, "aieee, found stripe ptr with no data ptr");
 277                                         continue;
 278                                 }
 279
 280                                 if (!m || !m->alive ||
 281                                     !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 282                                                                &next_ptr->ptr,
 283                                                                m->sectors)) {
 284                                         bch2_bkey_extent_entry_drop(new, entry);
 285                                         goto again;
 286                                 }
 287                         }
 288                 }
 289
 290                 ret = bch2_journal_key_insert(c, btree_id, level, new);
 291                 if (ret)
 292                         kfree(new);
 293                 else
 294                         *k = bkey_i_to_s_c(new);
 295         }
 296 fsck_err:
 297         return ret;
 298 }
 299
 300 /* marking of btree keys/nodes: */
 301
 302 static int bch2_gc_mark_key(struct bch_fs *c, enum btree_id btree_id,
 303                             unsigned level, bool is_root,
 304                             struct bkey_s_c k,
 305                             u8 *max_stale, bool initial)
 306 {
 307         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 308         const struct bch_extent_ptr *ptr;
 309         unsigned flags =
 310                 BTREE_TRIGGER_GC|
 311                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 312         int ret = 0;
 313
 314         if (initial) {
 315                 BUG_ON(bch2_journal_seq_verify &&
 316                        k.k->version.lo > journal_cur_seq(&c->journal));
 317
 318                 if (fsck_err_on(k.k->version.lo > atomic64_read(&c->key_version), c,
 319                                 "key version number higher than recorded: %llu > %llu",
 320                                 k.k->version.lo,
 321                                 atomic64_read(&c->key_version)))
 322                         atomic64_set(&c->key_version, k.k->version.lo);
 323
 324                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 325                     fsck_err_on(!bch2_bkey_replicas_marked(c, k), c,
 326                                 "superblock not marked as containing replicas (type %u)",
 327                                 k.k->type)) {
 328                         ret = bch2_mark_bkey_replicas(c, k);
 329                         if (ret) {
 330                                 bch_err(c, "error marking bkey replicas: %i", ret);
 331                                 goto err;
 332                         }
 333                 }
 334
 335                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, &k);
 336         }
 337
 338         bkey_for_each_ptr(ptrs, ptr) {
 339                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 340                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 341
 342                 if (gen_after(g->oldest_gen, ptr->gen))
 343                         g->oldest_gen = ptr->gen;
 344
 345                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 346         }
 347
 348         bch2_mark_key(c, k, 0, k.k->size, NULL, 0, flags);
 349 fsck_err:
 350 err:
 351         if (ret)
 352                 bch_err(c, "%s: ret %i", __func__, ret);
 353         return ret;
 354 }
 355
 356 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
 357                               bool initial)
 358 {
 359         struct btree_node_iter iter;
 360         struct bkey unpacked;
 361         struct bkey_s_c k;
 362         struct bkey_buf prev, cur;
 363         int ret = 0;
 364
 365         *max_stale = 0;
 366
 367         if (!btree_node_type_needs_gc(btree_node_type(b)))
 368                 return 0;
 369
 370         bch2_btree_node_iter_init_from_start(&iter, b);
 371         bch2_bkey_buf_init(&prev);
 372         bch2_bkey_buf_init(&cur);
 373         bkey_init(&prev.k->k);
 374
 375         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 376                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 377                                        k, max_stale, initial);
 378                 if (ret)
 379                         break;
 380
 381                 bch2_btree_node_iter_advance(&iter, b);
 382
 383                 if (b->c.level) {
 384                         bch2_bkey_buf_reassemble(&cur, c, k);
 385
 386                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 387                                         bch2_btree_node_iter_end(&iter));
 388                         if (ret)
 389                                 break;
 390                 }
 391         }
 392
 393         bch2_bkey_buf_exit(&cur, c);
 394         bch2_bkey_buf_exit(&prev, c);
 395         return ret;
 396 }
 397
 398 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 399                          bool initial)
 400 {
 401         struct btree_trans trans;
 402         struct btree_iter *iter;
 403         struct btree *b;
 404         unsigned depth = bch2_expensive_debug_checks    ? 0
 405                 : !btree_node_type_needs_gc(btree_id)   ? 1
 406                 : 0;
 407         u8 max_stale = 0;
 408         int ret = 0;
 409
 410         bch2_trans_init(&trans, c, 0, 0);
 411
 412         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 413
 414         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 415                               0, depth, BTREE_ITER_PREFETCH, b) {
 416                 bch2_verify_btree_nr_keys(b);
 417
 418                 gc_pos_set(c, gc_pos_btree_node(b));
 419
 420                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 421                 if (ret)
 422                         break;
 423
 424                 if (!initial) {
 425                         if (max_stale > 64)
 426                                 bch2_btree_node_rewrite(c, iter,
 427                                                 b->data->keys.seq,
 428                                                 BTREE_INSERT_NOWAIT|
 429                                                 BTREE_INSERT_GC_LOCK_HELD);
 430                         else if (!bch2_btree_gc_rewrite_disabled &&
 431                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 432                                 bch2_btree_node_rewrite(c, iter,
 433                                                 b->data->keys.seq,
 434                                                 BTREE_INSERT_NOWAIT|
 435                                                 BTREE_INSERT_GC_LOCK_HELD);
 436                 }
 437
 438                 bch2_trans_cond_resched(&trans);
 439         }
 440         bch2_trans_iter_put(&trans, iter);
 441
 442         ret = bch2_trans_exit(&trans) ?: ret;
 443         if (ret)
 444                 return ret;
 445
 446         mutex_lock(&c->btree_root_lock);
 447         b = c->btree_roots[btree_id].b;
 448         if (!btree_node_fake(b))
 449                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 450                                        bkey_i_to_s_c(&b->key),
 451                                        &max_stale, initial);
 452         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 453         mutex_unlock(&c->btree_root_lock);
 454
 455         return ret;
 456 }
 457
 458 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
 459                                       unsigned target_depth)
 460 {
 461         struct btree_and_journal_iter iter;
 462         struct bkey_s_c k;
 463         struct bkey_buf cur, prev;
 464         u8 max_stale = 0;
 465         int ret = 0;
 466
 467         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 468         bch2_bkey_buf_init(&prev);
 469         bch2_bkey_buf_init(&cur);
 470         bkey_init(&prev.k->k);
 471
 472         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 473                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 474                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 475
 476                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 477                                        k, &max_stale, true);
 478                 if (ret) {
 479                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 480                         break;
 481                 }
 482
 483                 if (b->c.level) {
 484                         bch2_bkey_buf_reassemble(&cur, c, k);
 485                         k = bkey_i_to_s_c(cur.k);
 486
 487                         bch2_btree_and_journal_iter_advance(&iter);
 488
 489                         ret = bch2_gc_check_topology(c, b,
 490                                         &prev, cur,
 491                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 492                         if (ret)
 493                                 break;
 494                 } else {
 495                         bch2_btree_and_journal_iter_advance(&iter);
 496                 }
 497         }
 498
 499         if (b->c.level > target_depth) {
 500                 bch2_btree_and_journal_iter_exit(&iter);
 501                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 502
 503                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 504                         struct btree *child;
 505
 506                         bch2_bkey_buf_reassemble(&cur, c, k);
 507                         bch2_btree_and_journal_iter_advance(&iter);
 508
 509                         child = bch2_btree_node_get_noiter(c, cur.k,
 510                                                 b->c.btree_id, b->c.level - 1,
 511                                                 false);
 512                         ret = PTR_ERR_OR_ZERO(child);
 513
 514                         if (fsck_err_on(ret == -EIO, c,
 515                                         "unreadable btree node")) {
 516                                 ret = bch2_journal_key_delete(c, b->c.btree_id,
 517                                                               b->c.level, cur.k->k.p);
 518                                 if (ret)
 519                                         return ret;
 520
 521                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 522                                 continue;
 523                         }
 524
 525                         if (ret) {
 526                                 bch_err(c, "%s: error %i getting btree node",
 527                                         __func__, ret);
 528                                 break;
 529                         }
 530
 531                         ret = bch2_gc_btree_init_recurse(c, child,
 532                                                          target_depth);
 533                         six_unlock_read(&child->c.lock);
 534
 535                         if (ret)
 536                                 break;
 537                 }
 538         }
 539 fsck_err:
 540         bch2_bkey_buf_exit(&cur, c);
 541         bch2_bkey_buf_exit(&prev, c);
 542         bch2_btree_and_journal_iter_exit(&iter);
 543         return ret;
 544 }
 545
 546 static int bch2_gc_btree_init(struct bch_fs *c,
 547                               enum btree_id btree_id)
 548 {
 549         struct btree *b;
 550         unsigned target_depth = bch2_expensive_debug_checks     ? 0
 551                 : !btree_node_type_needs_gc(btree_id)           ? 1
 552                 : 0;
 553         u8 max_stale = 0;
 554         char buf[100];
 555         int ret = 0;
 556
 557         b = c->btree_roots[btree_id].b;
 558
 559         if (btree_node_fake(b))
 560                 return 0;
 561
 562         six_lock_read(&b->c.lock, NULL, NULL);
 563         if (fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 564                         "btree root with incorrect min_key: %s",
 565                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 566                 BUG();
 567         }
 568
 569         if (fsck_err_on(bpos_cmp(b->data->max_key, POS_MAX), c,
 570                         "btree root with incorrect max_key: %s",
 571                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 572                 BUG();
 573         }
 574
 575         if (b->c.level >= target_depth)
 576                 ret = bch2_gc_btree_init_recurse(c, b, target_depth);
 577
 578         if (!ret)
 579                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 580                                        bkey_i_to_s_c(&b->key),
 581                                        &max_stale, true);
 582 fsck_err:
 583         six_unlock_read(&b->c.lock);
 584
 585         if (ret)
 586                 bch_err(c, "%s: ret %i", __func__, ret);
 587         return ret;
 588 }
 589
 590 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 591 {
 592         return  (int) btree_id_to_gc_phase(l) -
 593                 (int) btree_id_to_gc_phase(r);
 594 }
 595
 596 static int bch2_gc_btrees(struct bch_fs *c, bool initial)
 597 {
 598         enum btree_id ids[BTREE_ID_NR];
 599         unsigned i;
 600
 601         for (i = 0; i < BTREE_ID_NR; i++)
 602                 ids[i] = i;
 603         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
 604
 605         for (i = 0; i < BTREE_ID_NR; i++) {
 606                 enum btree_id id = ids[i];
 607                 int ret = initial
 608                         ? bch2_gc_btree_init(c, id)
 609                         : bch2_gc_btree(c, id, initial);
 610                 if (ret) {
 611                         bch_err(c, "%s: ret %i", __func__, ret);
 612                         return ret;
 613                 }
 614         }
 615
 616         return 0;
 617 }
 618
 619 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 620                                   u64 start, u64 end,
 621                                   enum bch_data_type type,
 622                                   unsigned flags)
 623 {
 624         u64 b = sector_to_bucket(ca, start);
 625
 626         do {
 627                 unsigned sectors =
 628                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
 629
 630                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
 631                                           gc_phase(GC_PHASE_SB), flags);
 632                 b++;
 633                 start += sectors;
 634         } while (start < end);
 635 }
 636
 637 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 638                               unsigned flags)
 639 {
 640         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 641         unsigned i;
 642         u64 b;
 643
 644         /*
 645          * This conditional is kind of gross, but we may be called from the
 646          * device add path, before the new device has actually been added to the
 647          * running filesystem:
 648          */
 649         if (c) {
 650                 lockdep_assert_held(&c->sb_lock);
 651                 percpu_down_read(&c->mark_lock);
 652         }
 653
 654         for (i = 0; i < layout->nr_superblocks; i++) {
 655                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 656
 657                 if (offset == BCH_SB_SECTOR)
 658                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 659                                               BCH_DATA_sb, flags);
 660
 661                 mark_metadata_sectors(c, ca, offset,
 662                                       offset + (1 << layout->sb_max_size_bits),
 663                                       BCH_DATA_sb, flags);
 664         }
 665
 666         for (i = 0; i < ca->journal.nr; i++) {
 667                 b = ca->journal.buckets[i];
 668                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
 669                                           ca->mi.bucket_size,
 670                                           gc_phase(GC_PHASE_SB), flags);
 671         }
 672
 673         if (c)
 674                 percpu_up_read(&c->mark_lock);
 675 }
 676
 677 static void bch2_mark_superblocks(struct bch_fs *c)
 678 {
 679         struct bch_dev *ca;
 680         unsigned i;
 681
 682         mutex_lock(&c->sb_lock);
 683         gc_pos_set(c, gc_phase(GC_PHASE_SB));
 684
 685         for_each_online_member(ca, c, i)
 686                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
 687         mutex_unlock(&c->sb_lock);
 688 }
 689
 690 #if 0
 691 /* Also see bch2_pending_btree_node_free_insert_done() */
 692 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 693 {
 694         struct btree_update *as;
 695         struct pending_btree_node_free *d;
 696
 697         mutex_lock(&c->btree_interior_update_lock);
 698         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 699
 700         for_each_pending_btree_node_free(c, as, d)
 701                 if (d->index_update_done)
 702                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
 703                                       0, 0, NULL, 0,
 704                                       BTREE_TRIGGER_GC);
 705
 706         mutex_unlock(&c->btree_interior_update_lock);
 707 }
 708 #endif
 709
 710 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 711 {
 712         struct bch_dev *ca;
 713         struct open_bucket *ob;
 714         size_t i, j, iter;
 715         unsigned ci;
 716
 717         percpu_down_read(&c->mark_lock);
 718
 719         spin_lock(&c->freelist_lock);
 720         gc_pos_set(c, gc_pos_alloc(c, NULL));
 721
 722         for_each_member_device(ca, c, ci) {
 723                 fifo_for_each_entry(i, &ca->free_inc, iter)
 724                         bch2_mark_alloc_bucket(c, ca, i, true,
 725                                                gc_pos_alloc(c, NULL),
 726                                                BTREE_TRIGGER_GC);
 727
 728
 729
 730                 for (j = 0; j < RESERVE_NR; j++)
 731                         fifo_for_each_entry(i, &ca->free[j], iter)
 732                                 bch2_mark_alloc_bucket(c, ca, i, true,
 733                                                        gc_pos_alloc(c, NULL),
 734                                                        BTREE_TRIGGER_GC);
 735         }
 736
 737         spin_unlock(&c->freelist_lock);
 738
 739         for (ob = c->open_buckets;
 740              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 741              ob++) {
 742                 spin_lock(&ob->lock);
 743                 if (ob->valid) {
 744                         gc_pos_set(c, gc_pos_alloc(c, ob));
 745                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 746                         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
 747                                                gc_pos_alloc(c, ob),
 748                                                BTREE_TRIGGER_GC);
 749                 }
 750                 spin_unlock(&ob->lock);
 751         }
 752
 753         percpu_up_read(&c->mark_lock);
 754 }
 755
 756 static void bch2_gc_free(struct bch_fs *c)
 757 {
 758         struct bch_dev *ca;
 759         unsigned i;
 760
 761         genradix_free(&c->stripes[1]);
 762
 763         for_each_member_device(ca, c, i) {
 764                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
 765                         sizeof(struct bucket_array) +
 766                         ca->mi.nbuckets * sizeof(struct bucket));
 767                 ca->buckets[1] = NULL;
 768
 769                 free_percpu(ca->usage_gc);
 770                 ca->usage_gc = NULL;
 771         }
 772
 773         free_percpu(c->usage_gc);
 774         c->usage_gc = NULL;
 775 }
 776
 777 static int bch2_gc_done(struct bch_fs *c,
 778                         bool initial)
 779 {
 780         struct bch_dev *ca;
 781         bool verify = (!initial ||
 782                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
 783         unsigned i, dev;
 784         int ret = 0;
 785
 786 #define copy_field(_f, _msg, ...)                                       \
 787         if (dst->_f != src->_f) {                                       \
 788                 if (verify)                                             \
 789                         fsck_err(c, _msg ": got %llu, should be %llu"   \
 790                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
 791                 dst->_f = src->_f;                                      \
 792                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 793         }
 794 #define copy_stripe_field(_f, _msg, ...)                                \
 795         if (dst->_f != src->_f) {                                       \
 796                 if (verify)                                             \
 797                         fsck_err(c, "stripe %zu has wrong "_msg         \
 798                                 ": got %u, should be %u",               \
 799                                 iter.pos, ##__VA_ARGS__,                \
 800                                 dst->_f, src->_f);                      \
 801                 dst->_f = src->_f;                                      \
 802                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 803         }
 804 #define copy_bucket_field(_f)                                           \
 805         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
 806                 if (verify)                                             \
 807                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
 808                                 ": got %u, should be %u", i, b,         \
 809                                 dst->b[b].mark.gen,                     \
 810                                 bch2_data_types[dst->b[b].mark.data_type],\
 811                                 dst->b[b].mark._f, src->b[b].mark._f);  \
 812                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
 813                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 814         }
 815 #define copy_dev_field(_f, _msg, ...)                                   \
 816         copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
 817 #define copy_fs_field(_f, _msg, ...)                                    \
 818         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
 819
 820         {
 821                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
 822                 struct stripe *dst, *src;
 823
 824                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
 825                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
 826
 827                         if (dst->alive          != src->alive ||
 828                             dst->sectors        != src->sectors ||
 829                             dst->algorithm      != src->algorithm ||
 830                             dst->nr_blocks      != src->nr_blocks ||
 831                             dst->nr_redundant   != src->nr_redundant) {
 832                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
 833                                 ret = -EINVAL;
 834                                 goto fsck_err;
 835                         }
 836
 837                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
 838                                 copy_stripe_field(block_sectors[i],
 839                                                   "block_sectors[%u]", i);
 840
 841                         dst->blocks_nonempty = 0;
 842                         for (i = 0; i < dst->nr_blocks; i++)
 843                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
 844
 845                         genradix_iter_advance(&iter, &c->stripes[1]);
 846                 }
 847         }
 848
 849         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 850                 bch2_fs_usage_acc_to_base(c, i);
 851
 852         for_each_member_device(ca, c, dev) {
 853                 struct bucket_array *dst = __bucket_array(ca, 0);
 854                 struct bucket_array *src = __bucket_array(ca, 1);
 855                 size_t b;
 856
 857                 for (b = 0; b < src->nbuckets; b++) {
 858                         copy_bucket_field(gen);
 859                         copy_bucket_field(data_type);
 860                         copy_bucket_field(owned_by_allocator);
 861                         copy_bucket_field(stripe);
 862                         copy_bucket_field(dirty_sectors);
 863                         copy_bucket_field(cached_sectors);
 864
 865                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
 866                 }
 867
 868                 {
 869                         struct bch_dev_usage *dst = ca->usage_base;
 870                         struct bch_dev_usage *src = (void *)
 871                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
 872                                                      dev_usage_u64s());
 873
 874                         copy_dev_field(buckets_ec,              "buckets_ec");
 875                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
 876
 877                         for (i = 0; i < BCH_DATA_NR; i++) {
 878                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
 879                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
 880                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
 881                         }
 882                 }
 883         };
 884
 885         {
 886                 unsigned nr = fs_usage_u64s(c);
 887                 struct bch_fs_usage *dst = c->usage_base;
 888                 struct bch_fs_usage *src = (void *)
 889                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
 890
 891                 copy_fs_field(hidden,           "hidden");
 892                 copy_fs_field(btree,            "btree");
 893                 copy_fs_field(data,     "data");
 894                 copy_fs_field(cached,   "cached");
 895                 copy_fs_field(reserved, "reserved");
 896                 copy_fs_field(nr_inodes,"nr_inodes");
 897
 898                 for (i = 0; i < BCH_REPLICAS_MAX; i++)
 899                         copy_fs_field(persistent_reserved[i],
 900                                       "persistent_reserved[%i]", i);
 901
 902                 for (i = 0; i < c->replicas.nr; i++) {
 903                         struct bch_replicas_entry *e =
 904                                 cpu_replicas_entry(&c->replicas, i);
 905                         char buf[80];
 906
 907                         bch2_replicas_entry_to_text(&PBUF(buf), e);
 908
 909                         copy_fs_field(replicas[i], "%s", buf);
 910                 }
 911         }
 912
 913 #undef copy_fs_field
 914 #undef copy_dev_field
 915 #undef copy_bucket_field
 916 #undef copy_stripe_field
 917 #undef copy_field
 918 fsck_err:
 919         if (ret)
 920                 bch_err(c, "%s: ret %i", __func__, ret);
 921         return ret;
 922 }
 923
 924 static int bch2_gc_start(struct bch_fs *c)
 925 {
 926         struct bch_dev *ca;
 927         unsigned i;
 928         int ret;
 929
 930         BUG_ON(c->usage_gc);
 931
 932         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
 933                                          sizeof(u64), GFP_KERNEL);
 934         if (!c->usage_gc) {
 935                 bch_err(c, "error allocating c->usage_gc");
 936                 return -ENOMEM;
 937         }
 938
 939         for_each_member_device(ca, c, i) {
 940                 BUG_ON(ca->buckets[1]);
 941                 BUG_ON(ca->usage_gc);
 942
 943                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
 944                                 ca->mi.nbuckets * sizeof(struct bucket),
 945                                 GFP_KERNEL|__GFP_ZERO);
 946                 if (!ca->buckets[1]) {
 947                         percpu_ref_put(&ca->ref);
 948                         bch_err(c, "error allocating ca->buckets[gc]");
 949                         return -ENOMEM;
 950                 }
 951
 952                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
 953                 if (!ca->usage_gc) {
 954                         bch_err(c, "error allocating ca->usage_gc");
 955                         percpu_ref_put(&ca->ref);
 956                         return -ENOMEM;
 957                 }
 958         }
 959
 960         ret = bch2_ec_mem_alloc(c, true);
 961         if (ret) {
 962                 bch_err(c, "error allocating ec gc mem");
 963                 return ret;
 964         }
 965
 966         percpu_down_write(&c->mark_lock);
 967
 968         /*
 969          * indicate to stripe code that we need to allocate for the gc stripes
 970          * radix tree, too
 971          */
 972         gc_pos_set(c, gc_phase(GC_PHASE_START));
 973
 974         for_each_member_device(ca, c, i) {
 975                 struct bucket_array *dst = __bucket_array(ca, 1);
 976                 struct bucket_array *src = __bucket_array(ca, 0);
 977                 size_t b;
 978
 979                 dst->first_bucket       = src->first_bucket;
 980                 dst->nbuckets           = src->nbuckets;
 981
 982                 for (b = 0; b < src->nbuckets; b++) {
 983                         struct bucket *d = &dst->b[b];
 984                         struct bucket *s = &src->b[b];
 985
 986                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
 987                         d->gen_valid = s->gen_valid;
 988                 }
 989         };
 990
 991         percpu_up_write(&c->mark_lock);
 992
 993         return 0;
 994 }
 995
 996 /**
 997  * bch2_gc - walk _all_ references to buckets, and recompute them:
 998  *
 999  * Order matters here:
1000  *  - Concurrent GC relies on the fact that we have a total ordering for
1001  *    everything that GC walks - see  gc_will_visit_node(),
1002  *    gc_will_visit_root()
1003  *
1004  *  - also, references move around in the course of index updates and
1005  *    various other crap: everything needs to agree on the ordering
1006  *    references are allowed to move around in - e.g., we're allowed to
1007  *    start with a reference owned by an open_bucket (the allocator) and
1008  *    move it to the btree, but not the reverse.
1009  *
1010  *    This is necessary to ensure that gc doesn't miss references that
1011  *    move around - if references move backwards in the ordering GC
1012  *    uses, GC could skip past them
1013  */
1014 int bch2_gc(struct bch_fs *c, bool initial)
1015 {
1016         struct bch_dev *ca;
1017         u64 start_time = local_clock();
1018         unsigned i, iter = 0;
1019         int ret;
1020
1021         lockdep_assert_held(&c->state_lock);
1022         trace_gc_start(c);
1023
1024         down_write(&c->gc_lock);
1025
1026         /* flush interior btree updates: */
1027         closure_wait_event(&c->btree_interior_update_wait,
1028                            !bch2_btree_interior_updates_nr_pending(c));
1029 again:
1030         ret = bch2_gc_start(c);
1031         if (ret)
1032                 goto out;
1033
1034         bch2_mark_superblocks(c);
1035
1036         ret = bch2_gc_btrees(c, initial);
1037         if (ret)
1038                 goto out;
1039
1040 #if 0
1041         bch2_mark_pending_btree_node_frees(c);
1042 #endif
1043         bch2_mark_allocator_buckets(c);
1044
1045         c->gc_count++;
1046
1047         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1048             (!iter && bch2_test_restart_gc)) {
1049                 /*
1050                  * XXX: make sure gens we fixed got saved
1051                  */
1052                 if (iter++ <= 2) {
1053                         bch_info(c, "Second GC pass needed, restarting:");
1054                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1055                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1056
1057                         percpu_down_write(&c->mark_lock);
1058                         bch2_gc_free(c);
1059                         percpu_up_write(&c->mark_lock);
1060                         /* flush fsck errors, reset counters */
1061                         bch2_flush_fsck_errs(c);
1062
1063                         goto again;
1064                 }
1065
1066                 bch_info(c, "Unable to fix bucket gens, looping");
1067                 ret = -EINVAL;
1068         }
1069 out:
1070         if (!ret) {
1071                 bch2_journal_block(&c->journal);
1072
1073                 percpu_down_write(&c->mark_lock);
1074                 ret = bch2_gc_done(c, initial);
1075
1076                 bch2_journal_unblock(&c->journal);
1077         } else {
1078                 percpu_down_write(&c->mark_lock);
1079         }
1080
1081         /* Indicates that gc is no longer in progress: */
1082         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1083
1084         bch2_gc_free(c);
1085         percpu_up_write(&c->mark_lock);
1086
1087         up_write(&c->gc_lock);
1088
1089         trace_gc_end(c);
1090         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1091
1092         /*
1093          * Wake up allocator in case it was waiting for buckets
1094          * because of not being able to inc gens
1095          */
1096         for_each_member_device(ca, c, i)
1097                 bch2_wake_allocator(ca);
1098
1099         /*
1100          * At startup, allocations can happen directly instead of via the
1101          * allocator thread - issue wakeup in case they blocked on gc_lock:
1102          */
1103         closure_wake_up(&c->freelist_wait);
1104         return ret;
1105 }
1106
1107 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1108 {
1109         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1110         const struct bch_extent_ptr *ptr;
1111
1112         percpu_down_read(&c->mark_lock);
1113         bkey_for_each_ptr(ptrs, ptr) {
1114                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1115                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1116
1117                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1118                         percpu_up_read(&c->mark_lock);
1119                         return true;
1120                 }
1121         }
1122
1123         bkey_for_each_ptr(ptrs, ptr) {
1124                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1125                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1126
1127                 if (gen_after(g->gc_gen, ptr->gen))
1128                         g->gc_gen = ptr->gen;
1129         }
1130         percpu_up_read(&c->mark_lock);
1131
1132         return false;
1133 }
1134
1135 /*
1136  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1137  * node pointers currently never have cached pointers that can become stale:
1138  */
1139 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1140 {
1141         struct btree_trans trans;
1142         struct btree_iter *iter;
1143         struct bkey_s_c k;
1144         struct bkey_buf sk;
1145         int ret = 0;
1146
1147         bch2_bkey_buf_init(&sk);
1148         bch2_trans_init(&trans, c, 0, 0);
1149
1150         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
1151                                    BTREE_ITER_PREFETCH|
1152                                    BTREE_ITER_NOT_EXTENTS|
1153                                    BTREE_ITER_ALL_SNAPSHOTS);
1154
1155         while ((k = bch2_btree_iter_peek(iter)).k &&
1156                !(ret = bkey_err(k))) {
1157                 if (gc_btree_gens_key(c, k)) {
1158                         bch2_bkey_buf_reassemble(&sk, c, k);
1159                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1160
1161                         bch2_trans_update(&trans, iter, sk.k, 0);
1162
1163                         ret = bch2_trans_commit(&trans, NULL, NULL,
1164                                                 BTREE_INSERT_NOFAIL);
1165                         if (ret == -EINTR)
1166                                 continue;
1167                         if (ret) {
1168                                 break;
1169                         }
1170                 }
1171
1172                 bch2_btree_iter_advance(iter);
1173         }
1174         bch2_trans_iter_put(&trans, iter);
1175
1176         bch2_trans_exit(&trans);
1177         bch2_bkey_buf_exit(&sk, c);
1178
1179         return ret;
1180 }
1181
1182 int bch2_gc_gens(struct bch_fs *c)
1183 {
1184         struct bch_dev *ca;
1185         struct bucket_array *buckets;
1186         struct bucket *g;
1187         unsigned i;
1188         int ret;
1189
1190         /*
1191          * Ideally we would be using state_lock and not gc_lock here, but that
1192          * introduces a deadlock in the RO path - we currently take the state
1193          * lock at the start of going RO, thus the gc thread may get stuck:
1194          */
1195         down_read(&c->gc_lock);
1196
1197         for_each_member_device(ca, c, i) {
1198                 down_read(&ca->bucket_lock);
1199                 buckets = bucket_array(ca);
1200
1201                 for_each_bucket(g, buckets)
1202                         g->gc_gen = g->mark.gen;
1203                 up_read(&ca->bucket_lock);
1204         }
1205
1206         for (i = 0; i < BTREE_ID_NR; i++)
1207                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1208                         ret = bch2_gc_btree_gens(c, i);
1209                         if (ret) {
1210                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1211                                 goto err;
1212                         }
1213                 }
1214
1215         for_each_member_device(ca, c, i) {
1216                 down_read(&ca->bucket_lock);
1217                 buckets = bucket_array(ca);
1218
1219                 for_each_bucket(g, buckets)
1220                         g->oldest_gen = g->gc_gen;
1221                 up_read(&ca->bucket_lock);
1222         }
1223
1224         c->gc_count++;
1225 err:
1226         up_read(&c->gc_lock);
1227         return ret;
1228 }
1229
1230 /* Btree coalescing */
1231
1232 static void recalc_packed_keys(struct btree *b)
1233 {
1234         struct bset *i = btree_bset_first(b);
1235         struct bkey_packed *k;
1236
1237         memset(&b->nr, 0, sizeof(b->nr));
1238
1239         BUG_ON(b->nsets != 1);
1240
1241         vstruct_for_each(i, k)
1242                 btree_keys_account_key_add(&b->nr, 0, k);
1243 }
1244
1245 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
1246                                 struct btree *old_nodes[GC_MERGE_NODES])
1247 {
1248         struct btree *parent = btree_node_parent(iter, old_nodes[0]);
1249         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
1250         unsigned blocks = btree_blocks(c) * 2 / 3;
1251         struct btree *new_nodes[GC_MERGE_NODES];
1252         struct btree_update *as;
1253         struct keylist keylist;
1254         struct bkey_format_state format_state;
1255         struct bkey_format new_format;
1256
1257         memset(new_nodes, 0, sizeof(new_nodes));
1258         bch2_keylist_init(&keylist, NULL);
1259
1260         /* Count keys that are not deleted */
1261         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
1262                 u64s += old_nodes[i]->nr.live_u64s;
1263
1264         nr_old_nodes = nr_new_nodes = i;
1265
1266         /* Check if all keys in @old_nodes could fit in one fewer node */
1267         if (nr_old_nodes <= 1 ||
1268             __vstruct_blocks(struct btree_node, c->block_bits,
1269                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
1270                 return;
1271
1272         /* Find a format that all keys in @old_nodes can pack into */
1273         bch2_bkey_format_init(&format_state);
1274
1275         /*
1276          * XXX: this won't correctly take it account the new min/max keys:
1277          */
1278         for (i = 0; i < nr_old_nodes; i++)
1279                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
1280
1281         new_format = bch2_bkey_format_done(&format_state);
1282
1283         /* Check if repacking would make any nodes too big to fit */
1284         for (i = 0; i < nr_old_nodes; i++)
1285                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
1286                         trace_btree_gc_coalesce_fail(c,
1287                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
1288                         return;
1289                 }
1290
1291         if (bch2_keylist_realloc(&keylist, NULL, 0,
1292                         BKEY_BTREE_PTR_U64s_MAX * nr_old_nodes)) {
1293                 trace_btree_gc_coalesce_fail(c,
1294                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
1295                 return;
1296         }
1297
1298         as = bch2_btree_update_start(iter, old_nodes[0]->c.level,
1299                         btree_update_reserve_required(c, parent) + nr_old_nodes,
1300                         BTREE_INSERT_NOFAIL|
1301                         BTREE_INSERT_USE_RESERVE);
1302         if (IS_ERR(as)) {
1303                 trace_btree_gc_coalesce_fail(c,
1304                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
1305                 bch2_keylist_free(&keylist, NULL);
1306                 return;
1307         }
1308
1309         trace_btree_gc_coalesce(c, old_nodes[0]);
1310
1311         for (i = 0; i < nr_old_nodes; i++)
1312                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
1313
1314         /* Repack everything with @new_format and sort down to one bset */
1315         for (i = 0; i < nr_old_nodes; i++)
1316                 new_nodes[i] =
1317                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
1318                                                             new_format);
1319
1320         /*
1321          * Conceptually we concatenate the nodes together and slice them
1322          * up at different boundaries.
1323          */
1324         for (i = nr_new_nodes - 1; i > 0; --i) {
1325                 struct btree *n1 = new_nodes[i];
1326                 struct btree *n2 = new_nodes[i - 1];
1327
1328                 struct bset *s1 = btree_bset_first(n1);
1329                 struct bset *s2 = btree_bset_first(n2);
1330                 struct bkey_packed *k, *last = NULL;
1331
1332                 /* Calculate how many keys from @n2 we could fit inside @n1 */
1333                 u64s = 0;
1334
1335                 for (k = s2->start;
1336                      k < vstruct_last(s2) &&
1337                      vstruct_blocks_plus(n1->data, c->block_bits,
1338                                          u64s + k->u64s) <= blocks;
1339                      k = bkey_next(k)) {
1340                         last = k;
1341                         u64s += k->u64s;
1342                 }
1343
1344                 if (u64s == le16_to_cpu(s2->u64s)) {
1345                         /* n2 fits entirely in n1 */
1346                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
1347
1348                         memcpy_u64s(vstruct_last(s1),
1349                                     s2->start,
1350                                     le16_to_cpu(s2->u64s));
1351                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
1352
1353                         set_btree_bset_end(n1, n1->set);
1354
1355                         six_unlock_write(&n2->c.lock);
1356                         bch2_btree_node_free_never_inserted(c, n2);
1357                         six_unlock_intent(&n2->c.lock);
1358
1359                         memmove(new_nodes + i - 1,
1360                                 new_nodes + i,
1361                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
1362                         new_nodes[--nr_new_nodes] = NULL;
1363                 } else if (u64s) {
1364                         /* move part of n2 into n1 */
1365                         n1->key.k.p = n1->data->max_key =
1366                                 bkey_unpack_pos(n1, last);
1367
1368                         n2->data->min_key = bpos_successor(n1->data->max_key);
1369
1370                         memcpy_u64s(vstruct_last(s1),
1371                                     s2->start, u64s);
1372                         le16_add_cpu(&s1->u64s, u64s);
1373
1374                         memmove(s2->start,
1375                                 vstruct_idx(s2, u64s),
1376                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
1377                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
1378
1379                         set_btree_bset_end(n1, n1->set);
1380                         set_btree_bset_end(n2, n2->set);
1381                 }
1382         }
1383
1384         for (i = 0; i < nr_new_nodes; i++) {
1385                 struct btree *n = new_nodes[i];
1386
1387                 recalc_packed_keys(n);
1388                 btree_node_reset_sib_u64s(n);
1389
1390                 bch2_btree_build_aux_trees(n);
1391
1392                 bch2_btree_update_add_new_node(as, n);
1393                 six_unlock_write(&n->c.lock);
1394
1395                 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1396         }
1397
1398         /*
1399          * The keys for the old nodes get deleted. We don't want to insert keys
1400          * that compare equal to the keys for the new nodes we'll also be
1401          * inserting - we can't because keys on a keylist must be strictly
1402          * greater than the previous keys, and we also don't need to since the
1403          * key for the new node will serve the same purpose (overwriting the key
1404          * for the old node).
1405          */
1406         for (i = 0; i < nr_old_nodes; i++) {
1407                 struct bkey_i delete;
1408                 unsigned j;
1409
1410                 for (j = 0; j < nr_new_nodes; j++)
1411                         if (!bpos_cmp(old_nodes[i]->key.k.p,
1412                                       new_nodes[j]->key.k.p))
1413                                 goto next;
1414
1415                 bkey_init(&delete.k);
1416                 delete.k.p = old_nodes[i]->key.k.p;
1417                 bch2_keylist_add_in_order(&keylist, &delete);
1418 next:
1419                 i = i;
1420         }
1421
1422         /*
1423          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
1424          * does the lookup once and thus expects the keys to be in sorted order
1425          * so we have to make sure the new keys are correctly ordered with
1426          * respect to the deleted keys added in the previous loop
1427          */
1428         for (i = 0; i < nr_new_nodes; i++)
1429                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
1430
1431         /* Insert the newly coalesced nodes */
1432         bch2_btree_insert_node(as, parent, iter, &keylist, 0);
1433
1434         BUG_ON(!bch2_keylist_empty(&keylist));
1435
1436         BUG_ON(iter->l[old_nodes[0]->c.level].b != old_nodes[0]);
1437
1438         bch2_btree_iter_node_replace(iter, new_nodes[0]);
1439
1440         for (i = 0; i < nr_new_nodes; i++)
1441                 bch2_btree_update_get_open_buckets(as, new_nodes[i]);
1442
1443         /* Free the old nodes and update our sliding window */
1444         for (i = 0; i < nr_old_nodes; i++) {
1445                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1446
1447                 /*
1448                  * the index update might have triggered a split, in which case
1449                  * the nodes we coalesced - the new nodes we just created -
1450                  * might not be sibling nodes anymore - don't add them to the
1451                  * sliding window (except the first):
1452                  */
1453                 if (!i) {
1454                         old_nodes[i] = new_nodes[i];
1455                 } else {
1456                         old_nodes[i] = NULL;
1457                 }
1458         }
1459
1460         for (i = 0; i < nr_new_nodes; i++)
1461                 six_unlock_intent(&new_nodes[i]->c.lock);
1462
1463         bch2_btree_update_done(as);
1464         bch2_keylist_free(&keylist, NULL);
1465 }
1466
1467 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1468 {
1469         struct btree_trans trans;
1470         struct btree_iter *iter;
1471         struct btree *b;
1472         bool kthread = (current->flags & PF_KTHREAD) != 0;
1473         unsigned i;
1474         int ret = 0;
1475
1476         /* Sliding window of adjacent btree nodes */
1477         struct btree *merge[GC_MERGE_NODES];
1478         u32 lock_seq[GC_MERGE_NODES];
1479
1480         bch2_trans_init(&trans, c, 0, 0);
1481
1482         /*
1483          * XXX: We don't have a good way of positively matching on sibling nodes
1484          * that have the same parent - this code works by handling the cases
1485          * where they might not have the same parent, and is thus fragile. Ugh.
1486          *
1487          * Perhaps redo this to use multiple linked iterators?
1488          */
1489         memset(merge, 0, sizeof(merge));
1490
1491         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1492                               BTREE_MAX_DEPTH, 0,
1493                               BTREE_ITER_PREFETCH, b) {
1494                 memmove(merge + 1, merge,
1495                         sizeof(merge) - sizeof(merge[0]));
1496                 memmove(lock_seq + 1, lock_seq,
1497                         sizeof(lock_seq) - sizeof(lock_seq[0]));
1498
1499                 merge[0] = b;
1500
1501                 for (i = 1; i < GC_MERGE_NODES; i++) {
1502                         if (!merge[i] ||
1503                             !six_relock_intent(&merge[i]->c.lock, lock_seq[i]))
1504                                 break;
1505
1506                         if (merge[i]->c.level != merge[0]->c.level) {
1507                                 six_unlock_intent(&merge[i]->c.lock);
1508                                 break;
1509                         }
1510                 }
1511                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1512
1513                 bch2_coalesce_nodes(c, iter, merge);
1514
1515                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1516                         lock_seq[i] = merge[i]->c.lock.state.seq;
1517                         six_unlock_intent(&merge[i]->c.lock);
1518                 }
1519
1520                 lock_seq[0] = merge[0]->c.lock.state.seq;
1521
1522                 if (kthread && kthread_should_stop()) {
1523                         ret = -ESHUTDOWN;
1524                         break;
1525                 }
1526
1527                 bch2_trans_cond_resched(&trans);
1528
1529                 /*
1530                  * If the parent node wasn't relocked, it might have been split
1531                  * and the nodes in our sliding window might not have the same
1532                  * parent anymore - blow away the sliding window:
1533                  */
1534                 if (btree_iter_node(iter, iter->level + 1) &&
1535                     !btree_node_intent_locked(iter, iter->level + 1))
1536                         memset(merge + 1, 0,
1537                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1538         }
1539         bch2_trans_iter_put(&trans, iter);
1540
1541         return bch2_trans_exit(&trans) ?: ret;
1542 }
1543
1544 /**
1545  * bch_coalesce - coalesce adjacent nodes with low occupancy
1546  */
1547 void bch2_coalesce(struct bch_fs *c)
1548 {
1549         enum btree_id id;
1550
1551         down_read(&c->gc_lock);
1552         trace_gc_coalesce_start(c);
1553
1554         for (id = 0; id < BTREE_ID_NR; id++) {
1555                 int ret = c->btree_roots[id].b
1556                         ? bch2_coalesce_btree(c, id)
1557                         : 0;
1558
1559                 if (ret) {
1560                         if (ret != -ESHUTDOWN)
1561                                 bch_err(c, "btree coalescing failed: %d", ret);
1562                         return;
1563                 }
1564         }
1565
1566         trace_gc_coalesce_end(c);
1567         up_read(&c->gc_lock);
1568 }
1569
1570 static int bch2_gc_thread(void *arg)
1571 {
1572         struct bch_fs *c = arg;
1573         struct io_clock *clock = &c->io_clock[WRITE];
1574         unsigned long last = atomic64_read(&clock->now);
1575         unsigned last_kick = atomic_read(&c->kick_gc);
1576         int ret;
1577
1578         set_freezable();
1579
1580         while (1) {
1581                 while (1) {
1582                         set_current_state(TASK_INTERRUPTIBLE);
1583
1584                         if (kthread_should_stop()) {
1585                                 __set_current_state(TASK_RUNNING);
1586                                 return 0;
1587                         }
1588
1589                         if (atomic_read(&c->kick_gc) != last_kick)
1590                                 break;
1591
1592                         if (c->btree_gc_periodic) {
1593                                 unsigned long next = last + c->capacity / 16;
1594
1595                                 if (atomic64_read(&clock->now) >= next)
1596                                         break;
1597
1598                                 bch2_io_clock_schedule_timeout(clock, next);
1599                         } else {
1600                                 schedule();
1601                         }
1602
1603                         try_to_freeze();
1604                 }
1605                 __set_current_state(TASK_RUNNING);
1606
1607                 last = atomic64_read(&clock->now);
1608                 last_kick = atomic_read(&c->kick_gc);
1609
1610                 /*
1611                  * Full gc is currently incompatible with btree key cache:
1612                  */
1613 #if 0
1614                 ret = bch2_gc(c, false, false);
1615 #else
1616                 ret = bch2_gc_gens(c);
1617 #endif
1618                 if (ret < 0)
1619                         bch_err(c, "btree gc failed: %i", ret);
1620
1621                 debug_check_no_locks_held();
1622         }
1623
1624         return 0;
1625 }
1626
1627 void bch2_gc_thread_stop(struct bch_fs *c)
1628 {
1629         struct task_struct *p;
1630
1631         p = c->gc_thread;
1632         c->gc_thread = NULL;
1633
1634         if (p) {
1635                 kthread_stop(p);
1636                 put_task_struct(p);
1637         }
1638 }
1639
1640 int bch2_gc_thread_start(struct bch_fs *c)
1641 {
1642         struct task_struct *p;
1643
1644         if (c->gc_thread)
1645                 return 0;
1646
1647         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1648         if (IS_ERR(p)) {
1649                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1650                 return PTR_ERR(p);
1651         }
1652
1653         get_task_struct(p);
1654         c->gc_thread = p;
1655         wake_up_process(p);
1656         return 0;
1657 }