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 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  40 {
  41         preempt_disable();
  42         write_seqcount_begin(&c->gc_pos_lock);
  43         c->gc_pos = new_pos;
  44         write_seqcount_end(&c->gc_pos_lock);
  45         preempt_enable();
  46 }
  47
  48 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  49 {
  50         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  51         __gc_pos_set(c, new_pos);
  52 }
  53
  54 /*
  55  * Missing: if an interior btree node is empty, we need to do something -
  56  * perhaps just kill it
  57  */
  58 static int bch2_gc_check_topology(struct bch_fs *c,
  59                                   struct btree *b,
  60                                   struct bkey_buf *prev,
  61                                   struct bkey_buf cur,
  62                                   bool is_last)
  63 {
  64         struct bpos node_start  = b->data->min_key;
  65         struct bpos node_end    = b->data->max_key;
  66         struct bpos expected_start = bkey_deleted(&prev->k->k)
  67                 ? node_start
  68                 : bpos_successor(prev->k->k.p);
  69         char buf1[200], buf2[200];
  70         int ret = 0;
  71
  72         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  73                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  74
  75                 if (bkey_deleted(&prev->k->k)) {
  76                         struct printbuf out = PBUF(buf1);
  77                         pr_buf(&out, "start of node: ");
  78                         bch2_bpos_to_text(&out, node_start);
  79                 } else {
  80                         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
  81                 }
  82
  83                 if (bpos_cmp(expected_start, bp->v.min_key)) {
  84                         bch2_topology_error(c);
  85
  86                         if (fsck_err(c, "btree node with incorrect min_key at btree %s level %u:\n"
  87                                      "  prev %s\n"
  88                                      "  cur %s",
  89                                      bch2_btree_ids[b->c.btree_id], b->c.level,
  90                                      buf1,
  91                                      (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2))) {
  92                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
  93                                 return FSCK_ERR_START_TOPOLOGY_REPAIR;
  94                         } else {
  95                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
  96                         }
  97                 }
  98         }
  99
 100         if (is_last && bpos_cmp(cur.k->k.p, node_end)) {
 101                 bch2_topology_error(c);
 102
 103                 if (fsck_err(c, "btree node with incorrect max_key at btree %s level %u:\n"
 104                              "  %s\n"
 105                              "  expected %s",
 106                              bch2_btree_ids[b->c.btree_id], b->c.level,
 107                              (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
 108                              (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2))) {
 109                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 110                         return FSCK_ERR_START_TOPOLOGY_REPAIR;
 111                 } else {
 112                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 113                 }
 114         }
 115
 116         bch2_bkey_buf_copy(prev, c, cur.k);
 117 fsck_err:
 118         return ret;
 119 }
 120
 121 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 122 {
 123         switch (b->key.k.type) {
 124         case KEY_TYPE_btree_ptr: {
 125                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 126
 127                 dst->k.p                = src->k.p;
 128                 dst->v.mem_ptr          = 0;
 129                 dst->v.seq              = b->data->keys.seq;
 130                 dst->v.sectors_written  = 0;
 131                 dst->v.flags            = 0;
 132                 dst->v.min_key          = b->data->min_key;
 133                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 134                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 135                 break;
 136         }
 137         case KEY_TYPE_btree_ptr_v2:
 138                 bkey_copy(&dst->k_i, &b->key);
 139                 break;
 140         default:
 141                 BUG();
 142         }
 143 }
 144
 145 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 146 {
 147         struct bkey_i_btree_ptr_v2 *new;
 148         int ret;
 149
 150         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 151         if (!new)
 152                 return -ENOMEM;
 153
 154         btree_ptr_to_v2(b, new);
 155         b->data->min_key        = new_min;
 156         new->v.min_key          = new_min;
 157         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 158
 159         ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 160         if (ret) {
 161                 kfree(new);
 162                 return ret;
 163         }
 164
 165         bch2_btree_node_drop_keys_outside_node(b);
 166
 167         return 0;
 168 }
 169
 170 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 171 {
 172         struct bkey_i_btree_ptr_v2 *new;
 173         int ret;
 174
 175         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 176         if (ret)
 177                 return ret;
 178
 179         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 180         if (!new)
 181                 return -ENOMEM;
 182
 183         btree_ptr_to_v2(b, new);
 184         b->data->max_key        = new_max;
 185         new->k.p                = new_max;
 186         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 187
 188         ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 189         if (ret) {
 190                 kfree(new);
 191                 return ret;
 192         }
 193
 194         bch2_btree_node_drop_keys_outside_node(b);
 195
 196         mutex_lock(&c->btree_cache.lock);
 197         bch2_btree_node_hash_remove(&c->btree_cache, b);
 198
 199         bkey_copy(&b->key, &new->k_i);
 200         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 201         BUG_ON(ret);
 202         mutex_unlock(&c->btree_cache.lock);
 203         return 0;
 204 }
 205
 206 static int btree_repair_node_start(struct bch_fs *c, struct btree *b,
 207                                    struct btree *prev, struct btree *cur)
 208 {
 209         struct bpos expected_start = !prev
 210                 ? b->data->min_key
 211                 : bpos_successor(prev->key.k.p);
 212         char buf1[200], buf2[200];
 213         int ret = 0;
 214
 215         if (!prev) {
 216                 struct printbuf out = PBUF(buf1);
 217                 pr_buf(&out, "start of node: ");
 218                 bch2_bpos_to_text(&out, b->data->min_key);
 219         } else {
 220                 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
 221         }
 222
 223         if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 224                         "btree node with incorrect min_key at btree %s level %u:\n"
 225                         "  prev %s\n"
 226                         "  cur %s",
 227                         bch2_btree_ids[b->c.btree_id], b->c.level,
 228                         buf1,
 229                         (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key)), buf2))) {
 230                 if (prev &&
 231                     bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 232                     BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data))
 233                         ret = set_node_max(c, prev,
 234                                 bpos_predecessor(cur->data->min_key));
 235                 else
 236                         ret = set_node_min(c, cur, expected_start);
 237                 if (ret)
 238                         return ret;
 239         }
 240 fsck_err:
 241         return ret;
 242 }
 243
 244 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 245                                  struct btree *child)
 246 {
 247         char buf1[200], buf2[200];
 248         int ret = 0;
 249
 250         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 251                         "btree node with incorrect max_key at btree %s level %u:\n"
 252                         "  %s\n"
 253                         "  expected %s",
 254                         bch2_btree_ids[b->c.btree_id], b->c.level,
 255                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
 256                         (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
 257                 ret = set_node_max(c, child, b->key.k.p);
 258                 if (ret)
 259                         return ret;
 260         }
 261 fsck_err:
 262         return ret;
 263 }
 264
 265 #define DROP_THIS_NODE          10
 266
 267 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 268 {
 269         struct btree_and_journal_iter iter;
 270         struct bkey_s_c k;
 271         struct bkey_buf tmp;
 272         struct btree *prev = NULL, *cur = NULL;
 273         bool have_child, dropped_children = false;
 274         char buf[200];
 275         int ret = 0;
 276
 277         if (!b->c.level)
 278                 return 0;
 279 again:
 280         have_child = dropped_children = false;
 281         bch2_bkey_buf_init(&tmp);
 282         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 283
 284         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 285                 bch2_btree_and_journal_iter_advance(&iter);
 286                 bch2_bkey_buf_reassemble(&tmp, c, k);
 287
 288                 cur = bch2_btree_node_get_noiter(c, tmp.k,
 289                                         b->c.btree_id, b->c.level - 1,
 290                                         false);
 291                 ret = PTR_ERR_OR_ZERO(cur);
 292
 293                 if (mustfix_fsck_err_on(ret == -EIO, c,
 294                                 "Unreadable btree node at btree %s level %u:\n"
 295                                 "  %s",
 296                                 bch2_btree_ids[b->c.btree_id],
 297                                 b->c.level - 1,
 298                                 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(tmp.k)), buf))) {
 299                         bch2_btree_node_evict(c, tmp.k);
 300                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 301                                                       b->c.level, tmp.k->k.p);
 302                         if (ret)
 303                                 goto err;
 304                         continue;
 305                 }
 306
 307                 if (ret) {
 308                         bch_err(c, "%s: error %i getting btree node",
 309                                 __func__, ret);
 310                         break;
 311                 }
 312
 313                 ret = btree_repair_node_start(c, b, prev, cur);
 314                 if (prev)
 315                         six_unlock_read(&prev->c.lock);
 316                 prev = cur;
 317                 cur = NULL;
 318
 319                 if (ret)
 320                         break;
 321         }
 322
 323         if (!ret && !IS_ERR_OR_NULL(prev)) {
 324                 BUG_ON(cur);
 325                 ret = btree_repair_node_end(c, b, prev);
 326         }
 327
 328         if (!IS_ERR_OR_NULL(prev))
 329                 six_unlock_read(&prev->c.lock);
 330         prev = NULL;
 331         if (!IS_ERR_OR_NULL(cur))
 332                 six_unlock_read(&cur->c.lock);
 333         cur = NULL;
 334
 335         if (ret)
 336                 goto err;
 337
 338         bch2_btree_and_journal_iter_exit(&iter);
 339         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 340
 341         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 342                 bch2_bkey_buf_reassemble(&tmp, c, k);
 343                 bch2_btree_and_journal_iter_advance(&iter);
 344
 345                 cur = bch2_btree_node_get_noiter(c, tmp.k,
 346                                         b->c.btree_id, b->c.level - 1,
 347                                         false);
 348                 ret = PTR_ERR_OR_ZERO(cur);
 349
 350                 if (ret) {
 351                         bch_err(c, "%s: error %i getting btree node",
 352                                 __func__, ret);
 353                         goto err;
 354                 }
 355
 356                 ret = bch2_btree_repair_topology_recurse(c, cur);
 357                 six_unlock_read(&cur->c.lock);
 358                 cur = NULL;
 359
 360                 if (ret == DROP_THIS_NODE) {
 361                         bch2_btree_node_evict(c, tmp.k);
 362                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 363                                                       b->c.level, tmp.k->k.p);
 364                         dropped_children = true;
 365                 }
 366
 367                 if (ret)
 368                         goto err;
 369
 370                 have_child = true;
 371         }
 372
 373         if (mustfix_fsck_err_on(!have_child, c,
 374                         "empty interior btree node at btree %s level %u\n"
 375                         "  %s",
 376                         bch2_btree_ids[b->c.btree_id],
 377                         b->c.level,
 378                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
 379                 ret = DROP_THIS_NODE;
 380 err:
 381 fsck_err:
 382         if (!IS_ERR_OR_NULL(prev))
 383                 six_unlock_read(&prev->c.lock);
 384         if (!IS_ERR_OR_NULL(cur))
 385                 six_unlock_read(&cur->c.lock);
 386
 387         bch2_btree_and_journal_iter_exit(&iter);
 388         bch2_bkey_buf_exit(&tmp, c);
 389
 390         if (!ret && dropped_children)
 391                 goto again;
 392
 393         return ret;
 394 }
 395
 396 static int bch2_repair_topology(struct bch_fs *c)
 397 {
 398         struct btree *b;
 399         unsigned i;
 400         int ret = 0;
 401
 402         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 403                 b = c->btree_roots[i].b;
 404                 if (btree_node_fake(b))
 405                         continue;
 406
 407                 six_lock_read(&b->c.lock, NULL, NULL);
 408                 ret = bch2_btree_repair_topology_recurse(c, b);
 409                 six_unlock_read(&b->c.lock);
 410
 411                 if (ret == DROP_THIS_NODE) {
 412                         bch_err(c, "empty btree root - repair unimplemented");
 413                         ret = FSCK_ERR_EXIT;
 414                 }
 415         }
 416
 417         return ret;
 418 }
 419
 420 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 421                                unsigned level, bool is_root,
 422                                struct bkey_s_c *k)
 423 {
 424         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 425         const union bch_extent_entry *entry;
 426         struct extent_ptr_decoded p = { 0 };
 427         bool do_update = false;
 428         char buf[200];
 429         int ret = 0;
 430
 431         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 432                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 433                 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
 434                 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
 435                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 436
 437                 if (fsck_err_on(g->mark.data_type &&
 438                                 g->mark.data_type != data_type, c,
 439                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 440                                 "while marking %s",
 441                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 442                                 bch2_data_types[g->mark.data_type],
 443                                 bch2_data_types[data_type],
 444                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 445                         if (data_type == BCH_DATA_btree) {
 446                                 g2->_mark.data_type = g->_mark.data_type = data_type;
 447                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 448                         } else {
 449                                 do_update = true;
 450                         }
 451                 }
 452
 453                 if (fsck_err_on(!g->gen_valid, c,
 454                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 455                                 "while marking %s",
 456                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 457                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 458                                 p.ptr.gen,
 459                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 460                         if (!p.ptr.cached) {
 461                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 462                                 g2->gen_valid   = g->gen_valid          = true;
 463                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 464                         } else {
 465                                 do_update = true;
 466                         }
 467                 }
 468
 469                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 470                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 471                                 "while marking %s",
 472                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 473                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 474                                 p.ptr.gen, g->mark.gen,
 475                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 476                         if (!p.ptr.cached) {
 477                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 478                                 g2->gen_valid   = g->gen_valid  = true;
 479                                 g2->_mark.data_type             = 0;
 480                                 g2->_mark.dirty_sectors         = 0;
 481                                 g2->_mark.cached_sectors        = 0;
 482                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 483                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 484                         } else {
 485                                 do_update = true;
 486                         }
 487                 }
 488
 489                 if (fsck_err_on(!p.ptr.cached &&
 490                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 491                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 492                                 "while marking %s",
 493                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 494                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 495                                 p.ptr.gen, g->mark.gen,
 496                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 497                         do_update = true;
 498
 499                 if (p.has_ec) {
 500                         struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
 501
 502                         if (fsck_err_on(!m || !m->alive, c,
 503                                         "pointer to nonexistent stripe %llu\n"
 504                                         "while marking %s",
 505                                         (u64) p.ec.idx,
 506                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 507                                 do_update = true;
 508
 509                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 510                                         "pointer does not match stripe %llu\n"
 511                                         "while marking %s",
 512                                         (u64) p.ec.idx,
 513                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 514                                 do_update = true;
 515                 }
 516         }
 517
 518         if (do_update) {
 519                 struct bkey_ptrs ptrs;
 520                 union bch_extent_entry *entry;
 521                 struct bch_extent_ptr *ptr;
 522                 struct bkey_i *new;
 523
 524                 if (is_root) {
 525                         bch_err(c, "cannot update btree roots yet");
 526                         return -EINVAL;
 527                 }
 528
 529                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 530                 if (!new) {
 531                         bch_err(c, "%s: error allocating new key", __func__);
 532                         return -ENOMEM;
 533                 }
 534
 535                 bkey_reassemble(new, *k);
 536
 537                 if (level) {
 538                         /*
 539                          * We don't want to drop btree node pointers - if the
 540                          * btree node isn't there anymore, the read path will
 541                          * sort it out:
 542                          */
 543                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 544                         bkey_for_each_ptr(ptrs, ptr) {
 545                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 546                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 547
 548                                 ptr->gen = g->mark.gen;
 549                         }
 550                 } else {
 551                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 552                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 553                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 554                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 555
 556                                 (ptr->cached &&
 557                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 558                                 (!ptr->cached &&
 559                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 560                                 (g->mark.data_type &&
 561                                  g->mark.data_type != data_type);
 562                         }));
 563 again:
 564                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 565                         bkey_extent_entry_for_each(ptrs, entry) {
 566                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 567                                         struct stripe *m = genradix_ptr(&c->stripes[true],
 568                                                                         entry->stripe_ptr.idx);
 569                                         union bch_extent_entry *next_ptr;
 570
 571                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 572                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 573                                                         goto found;
 574                                         next_ptr = NULL;
 575 found:
 576                                         if (!next_ptr) {
 577                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 578                                                 continue;
 579                                         }
 580
 581                                         if (!m || !m->alive ||
 582                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 583                                                                        &next_ptr->ptr,
 584                                                                        m->sectors)) {
 585                                                 bch2_bkey_extent_entry_drop(new, entry);
 586                                                 goto again;
 587                                         }
 588                                 }
 589                         }
 590                 }
 591
 592                 ret = bch2_journal_key_insert(c, btree_id, level, new);
 593                 if (ret)
 594                         kfree(new);
 595                 else
 596                         *k = bkey_i_to_s_c(new);
 597         }
 598 fsck_err:
 599         return ret;
 600 }
 601
 602 /* marking of btree keys/nodes: */
 603
 604 static int bch2_gc_mark_key(struct bch_fs *c, enum btree_id btree_id,
 605                             unsigned level, bool is_root,
 606                             struct bkey_s_c *k,
 607                             u8 *max_stale, bool initial)
 608 {
 609         struct bkey_ptrs_c ptrs;
 610         const struct bch_extent_ptr *ptr;
 611         unsigned flags =
 612                 BTREE_TRIGGER_GC|
 613                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 614         int ret = 0;
 615
 616         if (initial) {
 617                 BUG_ON(bch2_journal_seq_verify &&
 618                        k->k->version.lo > journal_cur_seq(&c->journal));
 619
 620                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 621                 if (ret)
 622                         goto err;
 623
 624                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 625                                 "key version number higher than recorded: %llu > %llu",
 626                                 k->k->version.lo,
 627                                 atomic64_read(&c->key_version)))
 628                         atomic64_set(&c->key_version, k->k->version.lo);
 629
 630                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 631                     fsck_err_on(!bch2_bkey_replicas_marked(c, *k), c,
 632                                 "superblock not marked as containing replicas (type %u)",
 633                                 k->k->type)) {
 634                         ret = bch2_mark_bkey_replicas(c, *k);
 635                         if (ret) {
 636                                 bch_err(c, "error marking bkey replicas: %i", ret);
 637                                 goto err;
 638                         }
 639                 }
 640         }
 641
 642         ptrs = bch2_bkey_ptrs_c(*k);
 643         bkey_for_each_ptr(ptrs, ptr) {
 644                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 645                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 646
 647                 if (gen_after(g->oldest_gen, ptr->gen))
 648                         g->oldest_gen = ptr->gen;
 649
 650                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 651         }
 652
 653         bch2_mark_key(c, *k, 0, k->k->size, NULL, 0, flags);
 654 fsck_err:
 655 err:
 656         if (ret)
 657                 bch_err(c, "%s: ret %i", __func__, ret);
 658         return ret;
 659 }
 660
 661 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
 662                               bool initial)
 663 {
 664         struct btree_node_iter iter;
 665         struct bkey unpacked;
 666         struct bkey_s_c k;
 667         struct bkey_buf prev, cur;
 668         int ret = 0;
 669
 670         *max_stale = 0;
 671
 672         if (!btree_node_type_needs_gc(btree_node_type(b)))
 673                 return 0;
 674
 675         bch2_btree_node_iter_init_from_start(&iter, b);
 676         bch2_bkey_buf_init(&prev);
 677         bch2_bkey_buf_init(&cur);
 678         bkey_init(&prev.k->k);
 679
 680         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 681                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 682                                        &k, max_stale, initial);
 683                 if (ret)
 684                         break;
 685
 686                 bch2_btree_node_iter_advance(&iter, b);
 687
 688                 if (b->c.level) {
 689                         bch2_bkey_buf_reassemble(&cur, c, k);
 690
 691                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 692                                         bch2_btree_node_iter_end(&iter));
 693                         if (ret)
 694                                 break;
 695                 }
 696         }
 697
 698         bch2_bkey_buf_exit(&cur, c);
 699         bch2_bkey_buf_exit(&prev, c);
 700         return ret;
 701 }
 702
 703 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 704                          bool initial, bool metadata_only)
 705 {
 706         struct btree_trans trans;
 707         struct btree_iter *iter;
 708         struct btree *b;
 709         unsigned depth = metadata_only                  ? 1
 710                 : bch2_expensive_debug_checks           ? 0
 711                 : !btree_node_type_needs_gc(btree_id)   ? 1
 712                 : 0;
 713         u8 max_stale = 0;
 714         int ret = 0;
 715
 716         bch2_trans_init(&trans, c, 0, 0);
 717
 718         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 719
 720         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 721                               0, depth, BTREE_ITER_PREFETCH, b) {
 722                 bch2_verify_btree_nr_keys(b);
 723
 724                 gc_pos_set(c, gc_pos_btree_node(b));
 725
 726                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 727                 if (ret)
 728                         break;
 729
 730                 if (!initial) {
 731                         if (max_stale > 64)
 732                                 bch2_btree_node_rewrite(c, iter,
 733                                                 b->data->keys.seq,
 734                                                 BTREE_INSERT_NOWAIT|
 735                                                 BTREE_INSERT_GC_LOCK_HELD);
 736                         else if (!bch2_btree_gc_rewrite_disabled &&
 737                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 738                                 bch2_btree_node_rewrite(c, iter,
 739                                                 b->data->keys.seq,
 740                                                 BTREE_INSERT_NOWAIT|
 741                                                 BTREE_INSERT_GC_LOCK_HELD);
 742                 }
 743
 744                 bch2_trans_cond_resched(&trans);
 745         }
 746         bch2_trans_iter_put(&trans, iter);
 747
 748         ret = bch2_trans_exit(&trans) ?: ret;
 749         if (ret)
 750                 return ret;
 751
 752         mutex_lock(&c->btree_root_lock);
 753         b = c->btree_roots[btree_id].b;
 754         if (!btree_node_fake(b)) {
 755                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 756
 757                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 758                                        &k, &max_stale, initial);
 759         }
 760         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 761         mutex_unlock(&c->btree_root_lock);
 762
 763         return ret;
 764 }
 765
 766 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
 767                                       unsigned target_depth)
 768 {
 769         struct btree_and_journal_iter iter;
 770         struct bkey_s_c k;
 771         struct bkey_buf cur, prev;
 772         u8 max_stale = 0;
 773         char buf[200];
 774         int ret = 0;
 775
 776         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 777         bch2_bkey_buf_init(&prev);
 778         bch2_bkey_buf_init(&cur);
 779         bkey_init(&prev.k->k);
 780
 781         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 782                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 783                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 784
 785                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, false,
 786                                        &k, &max_stale, true);
 787                 if (ret) {
 788                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 789                         goto fsck_err;
 790                 }
 791
 792                 if (b->c.level) {
 793                         bch2_bkey_buf_reassemble(&cur, c, k);
 794                         k = bkey_i_to_s_c(cur.k);
 795
 796                         bch2_btree_and_journal_iter_advance(&iter);
 797
 798                         ret = bch2_gc_check_topology(c, b,
 799                                         &prev, cur,
 800                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 801                         if (ret)
 802                                 goto fsck_err;
 803                 } else {
 804                         bch2_btree_and_journal_iter_advance(&iter);
 805                 }
 806         }
 807
 808         if (b->c.level > target_depth) {
 809                 bch2_btree_and_journal_iter_exit(&iter);
 810                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 811
 812                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 813                         struct btree *child;
 814
 815                         bch2_bkey_buf_reassemble(&cur, c, k);
 816                         bch2_btree_and_journal_iter_advance(&iter);
 817
 818                         child = bch2_btree_node_get_noiter(c, cur.k,
 819                                                 b->c.btree_id, b->c.level - 1,
 820                                                 false);
 821                         ret = PTR_ERR_OR_ZERO(child);
 822
 823                         if (ret == -EIO) {
 824                                 bch2_topology_error(c);
 825
 826                                 if (fsck_err(c, "Unreadable btree node at btree %s level %u:\n"
 827                                         "  %s",
 828                                         bch2_btree_ids[b->c.btree_id],
 829                                         b->c.level - 1,
 830                                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf))) {
 831                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 832                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 833                                         goto fsck_err;
 834                                 } else {
 835                                         /* Continue marking when opted to not
 836                                          * fix the error: */
 837                                         ret = 0;
 838                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 839                                         continue;
 840                                 }
 841                         } else if (ret) {
 842                                 bch_err(c, "%s: error %i getting btree node",
 843                                         __func__, ret);
 844                                 break;
 845                         }
 846
 847                         ret = bch2_gc_btree_init_recurse(c, child,
 848                                                          target_depth);
 849                         six_unlock_read(&child->c.lock);
 850
 851                         if (ret)
 852                                 break;
 853                 }
 854         }
 855 fsck_err:
 856         bch2_bkey_buf_exit(&cur, c);
 857         bch2_bkey_buf_exit(&prev, c);
 858         bch2_btree_and_journal_iter_exit(&iter);
 859         return ret;
 860 }
 861
 862 static int bch2_gc_btree_init(struct bch_fs *c,
 863                               enum btree_id btree_id,
 864                               bool metadata_only)
 865 {
 866         struct btree *b;
 867         unsigned target_depth = metadata_only           ? 1
 868                 : bch2_expensive_debug_checks           ? 0
 869                 : !btree_node_type_needs_gc(btree_id)   ? 1
 870                 : 0;
 871         u8 max_stale = 0;
 872         char buf[100];
 873         int ret = 0;
 874
 875         b = c->btree_roots[btree_id].b;
 876
 877         if (btree_node_fake(b))
 878                 return 0;
 879
 880         six_lock_read(&b->c.lock, NULL, NULL);
 881         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 882                         "btree root with incorrect min_key: %s",
 883                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 884                 bch_err(c, "repair unimplemented");
 885                 ret = FSCK_ERR_EXIT;
 886                 goto fsck_err;
 887         }
 888
 889         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, POS_MAX), c,
 890                         "btree root with incorrect max_key: %s",
 891                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 892                 bch_err(c, "repair unimplemented");
 893                 ret = FSCK_ERR_EXIT;
 894                 goto fsck_err;
 895         }
 896
 897         if (b->c.level >= target_depth)
 898                 ret = bch2_gc_btree_init_recurse(c, b, target_depth);
 899
 900         if (!ret) {
 901                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 902
 903                 ret = bch2_gc_mark_key(c, b->c.btree_id, b->c.level, true,
 904                                        &k, &max_stale, true);
 905         }
 906 fsck_err:
 907         six_unlock_read(&b->c.lock);
 908
 909         if (ret < 0)
 910                 bch_err(c, "%s: ret %i", __func__, ret);
 911         return ret;
 912 }
 913
 914 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 915 {
 916         return  (int) btree_id_to_gc_phase(l) -
 917                 (int) btree_id_to_gc_phase(r);
 918 }
 919
 920 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
 921 {
 922         enum btree_id ids[BTREE_ID_NR];
 923         unsigned i;
 924         int ret = 0;
 925
 926         for (i = 0; i < BTREE_ID_NR; i++)
 927                 ids[i] = i;
 928         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
 929
 930         for (i = 0; i < BTREE_ID_NR && !ret; i++)
 931                 ret = initial
 932                         ? bch2_gc_btree_init(c, ids[i], metadata_only)
 933                         : bch2_gc_btree(c, ids[i], initial, metadata_only);
 934
 935         if (ret < 0)
 936                 bch_err(c, "%s: ret %i", __func__, ret);
 937         return ret;
 938 }
 939
 940 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 941                                   u64 start, u64 end,
 942                                   enum bch_data_type type,
 943                                   unsigned flags)
 944 {
 945         u64 b = sector_to_bucket(ca, start);
 946
 947         do {
 948                 unsigned sectors =
 949                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
 950
 951                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
 952                                           gc_phase(GC_PHASE_SB), flags);
 953                 b++;
 954                 start += sectors;
 955         } while (start < end);
 956 }
 957
 958 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 959                               unsigned flags)
 960 {
 961         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 962         unsigned i;
 963         u64 b;
 964
 965         /*
 966          * This conditional is kind of gross, but we may be called from the
 967          * device add path, before the new device has actually been added to the
 968          * running filesystem:
 969          */
 970         if (c) {
 971                 lockdep_assert_held(&c->sb_lock);
 972                 percpu_down_read(&c->mark_lock);
 973         }
 974
 975         for (i = 0; i < layout->nr_superblocks; i++) {
 976                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 977
 978                 if (offset == BCH_SB_SECTOR)
 979                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 980                                               BCH_DATA_sb, flags);
 981
 982                 mark_metadata_sectors(c, ca, offset,
 983                                       offset + (1 << layout->sb_max_size_bits),
 984                                       BCH_DATA_sb, flags);
 985         }
 986
 987         for (i = 0; i < ca->journal.nr; i++) {
 988                 b = ca->journal.buckets[i];
 989                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
 990                                           ca->mi.bucket_size,
 991                                           gc_phase(GC_PHASE_SB), flags);
 992         }
 993
 994         if (c)
 995                 percpu_up_read(&c->mark_lock);
 996 }
 997
 998 static void bch2_mark_superblocks(struct bch_fs *c)
 999 {
1000         struct bch_dev *ca;
1001         unsigned i;
1002
1003         mutex_lock(&c->sb_lock);
1004         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1005
1006         for_each_online_member(ca, c, i)
1007                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1008         mutex_unlock(&c->sb_lock);
1009 }
1010
1011 #if 0
1012 /* Also see bch2_pending_btree_node_free_insert_done() */
1013 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1014 {
1015         struct btree_update *as;
1016         struct pending_btree_node_free *d;
1017
1018         mutex_lock(&c->btree_interior_update_lock);
1019         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1020
1021         for_each_pending_btree_node_free(c, as, d)
1022                 if (d->index_update_done)
1023                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
1024                                       0, 0, NULL, 0,
1025                                       BTREE_TRIGGER_GC);
1026
1027         mutex_unlock(&c->btree_interior_update_lock);
1028 }
1029 #endif
1030
1031 static void bch2_gc_free(struct bch_fs *c)
1032 {
1033         struct bch_dev *ca;
1034         unsigned i;
1035
1036         genradix_free(&c->stripes[1]);
1037
1038         for_each_member_device(ca, c, i) {
1039                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1040                         sizeof(struct bucket_array) +
1041                         ca->mi.nbuckets * sizeof(struct bucket));
1042                 ca->buckets[1] = NULL;
1043
1044                 free_percpu(ca->usage_gc);
1045                 ca->usage_gc = NULL;
1046         }
1047
1048         free_percpu(c->usage_gc);
1049         c->usage_gc = NULL;
1050 }
1051
1052 static int bch2_gc_done(struct bch_fs *c,
1053                         bool initial, bool metadata_only)
1054 {
1055         struct bch_dev *ca = NULL;
1056         bool verify = !metadata_only && (!initial ||
1057                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1058         unsigned i, dev;
1059         int ret = 0;
1060
1061 #define copy_field(_f, _msg, ...)                                       \
1062         if (dst->_f != src->_f) {                                       \
1063                 if (verify)                                             \
1064                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1065                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1066                 dst->_f = src->_f;                                      \
1067                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1068         }
1069 #define copy_stripe_field(_f, _msg, ...)                                \
1070         if (dst->_f != src->_f) {                                       \
1071                 if (verify)                                             \
1072                         fsck_err(c, "stripe %zu has wrong "_msg         \
1073                                 ": got %u, should be %u",               \
1074                                 iter.pos, ##__VA_ARGS__,                \
1075                                 dst->_f, src->_f);                      \
1076                 dst->_f = src->_f;                                      \
1077                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1078         }
1079 #define copy_bucket_field(_f)                                           \
1080         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
1081                 if (verify)                                             \
1082                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
1083                                 ": got %u, should be %u", dev, b,       \
1084                                 dst->b[b].mark.gen,                     \
1085                                 bch2_data_types[dst->b[b].mark.data_type],\
1086                                 dst->b[b].mark._f, src->b[b].mark._f);  \
1087                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
1088                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1089         }
1090 #define copy_dev_field(_f, _msg, ...)                                   \
1091         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1092 #define copy_fs_field(_f, _msg, ...)                                    \
1093         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1094
1095         if (!metadata_only) {
1096                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
1097                 struct stripe *dst, *src;
1098
1099                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
1100                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
1101
1102                         if (dst->alive          != src->alive ||
1103                             dst->sectors        != src->sectors ||
1104                             dst->algorithm      != src->algorithm ||
1105                             dst->nr_blocks      != src->nr_blocks ||
1106                             dst->nr_redundant   != src->nr_redundant) {
1107                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
1108                                 ret = -EINVAL;
1109                                 goto fsck_err;
1110                         }
1111
1112                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
1113                                 copy_stripe_field(block_sectors[i],
1114                                                   "block_sectors[%u]", i);
1115
1116                         dst->blocks_nonempty = 0;
1117                         for (i = 0; i < dst->nr_blocks; i++)
1118                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
1119
1120                         genradix_iter_advance(&iter, &c->stripes[1]);
1121                 }
1122         }
1123
1124         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1125                 bch2_fs_usage_acc_to_base(c, i);
1126
1127         for_each_member_device(ca, c, dev) {
1128                 struct bucket_array *dst = __bucket_array(ca, 0);
1129                 struct bucket_array *src = __bucket_array(ca, 1);
1130                 size_t b;
1131
1132                 for (b = 0; b < src->nbuckets; b++) {
1133                         copy_bucket_field(gen);
1134                         copy_bucket_field(data_type);
1135                         copy_bucket_field(stripe);
1136                         copy_bucket_field(dirty_sectors);
1137                         copy_bucket_field(cached_sectors);
1138
1139                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
1140                 }
1141
1142                 {
1143                         struct bch_dev_usage *dst = ca->usage_base;
1144                         struct bch_dev_usage *src = (void *)
1145                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1146                                                      dev_usage_u64s());
1147
1148                         copy_dev_field(buckets_ec,              "buckets_ec");
1149                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1150
1151                         for (i = 0; i < BCH_DATA_NR; i++) {
1152                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1153                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1154                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1155                         }
1156                 }
1157         };
1158
1159         {
1160                 unsigned nr = fs_usage_u64s(c);
1161                 struct bch_fs_usage *dst = c->usage_base;
1162                 struct bch_fs_usage *src = (void *)
1163                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1164
1165                 copy_fs_field(hidden,           "hidden");
1166                 copy_fs_field(btree,            "btree");
1167
1168                 if (!metadata_only) {
1169                         copy_fs_field(data,     "data");
1170                         copy_fs_field(cached,   "cached");
1171                         copy_fs_field(reserved, "reserved");
1172                         copy_fs_field(nr_inodes,"nr_inodes");
1173
1174                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1175                                 copy_fs_field(persistent_reserved[i],
1176                                               "persistent_reserved[%i]", i);
1177                 }
1178
1179                 for (i = 0; i < c->replicas.nr; i++) {
1180                         struct bch_replicas_entry *e =
1181                                 cpu_replicas_entry(&c->replicas, i);
1182                         char buf[80];
1183
1184                         if (metadata_only &&
1185                             (e->data_type == BCH_DATA_user ||
1186                              e->data_type == BCH_DATA_cached))
1187                                 continue;
1188
1189                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1190
1191                         copy_fs_field(replicas[i], "%s", buf);
1192                 }
1193         }
1194
1195 #undef copy_fs_field
1196 #undef copy_dev_field
1197 #undef copy_bucket_field
1198 #undef copy_stripe_field
1199 #undef copy_field
1200 fsck_err:
1201         if (ca)
1202                 percpu_ref_put(&ca->ref);
1203         if (ret)
1204                 bch_err(c, "%s: ret %i", __func__, ret);
1205         return ret;
1206 }
1207
1208 static int bch2_gc_start(struct bch_fs *c,
1209                          bool metadata_only)
1210 {
1211         struct bch_dev *ca = NULL;
1212         unsigned i;
1213         int ret;
1214
1215         BUG_ON(c->usage_gc);
1216
1217         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1218                                          sizeof(u64), GFP_KERNEL);
1219         if (!c->usage_gc) {
1220                 bch_err(c, "error allocating c->usage_gc");
1221                 return -ENOMEM;
1222         }
1223
1224         for_each_member_device(ca, c, i) {
1225                 BUG_ON(ca->buckets[1]);
1226                 BUG_ON(ca->usage_gc);
1227
1228                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1229                                 ca->mi.nbuckets * sizeof(struct bucket),
1230                                 GFP_KERNEL|__GFP_ZERO);
1231                 if (!ca->buckets[1]) {
1232                         percpu_ref_put(&ca->ref);
1233                         bch_err(c, "error allocating ca->buckets[gc]");
1234                         return -ENOMEM;
1235                 }
1236
1237                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1238                 if (!ca->usage_gc) {
1239                         bch_err(c, "error allocating ca->usage_gc");
1240                         percpu_ref_put(&ca->ref);
1241                         return -ENOMEM;
1242                 }
1243         }
1244
1245         ret = bch2_ec_mem_alloc(c, true);
1246         if (ret) {
1247                 bch_err(c, "error allocating ec gc mem");
1248                 return ret;
1249         }
1250
1251         percpu_down_write(&c->mark_lock);
1252
1253         /*
1254          * indicate to stripe code that we need to allocate for the gc stripes
1255          * radix tree, too
1256          */
1257         gc_pos_set(c, gc_phase(GC_PHASE_START));
1258
1259         for_each_member_device(ca, c, i) {
1260                 struct bucket_array *dst = __bucket_array(ca, 1);
1261                 struct bucket_array *src = __bucket_array(ca, 0);
1262                 size_t b;
1263
1264                 dst->first_bucket       = src->first_bucket;
1265                 dst->nbuckets           = src->nbuckets;
1266
1267                 for (b = 0; b < src->nbuckets; b++) {
1268                         struct bucket *d = &dst->b[b];
1269                         struct bucket *s = &src->b[b];
1270
1271                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1272                         d->gen_valid = s->gen_valid;
1273
1274                         if (metadata_only &&
1275                             (s->mark.data_type == BCH_DATA_user ||
1276                              s->mark.data_type == BCH_DATA_cached))
1277                                 d->_mark = s->mark;
1278                 }
1279         };
1280
1281         percpu_up_write(&c->mark_lock);
1282
1283         return 0;
1284 }
1285
1286 static int bch2_gc_reflink_done_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1287 {
1288         struct reflink_gc *r;
1289         const __le64 *refcount = bkey_refcount_c(k);
1290         char buf[200];
1291         int ret = 0;
1292
1293         if (!refcount)
1294                 return 0;
1295
1296         r = genradix_ptr(&c->reflink_gc_table, c->reflink_gc_idx++);
1297         if (!r)
1298                 return -ENOMEM;
1299
1300         if (!r ||
1301             r->offset != k.k->p.offset ||
1302             r->size != k.k->size) {
1303                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1304                 return -EINVAL;
1305         }
1306
1307         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1308                         "reflink key has wrong refcount:\n"
1309                         "  %s\n"
1310                         "  should be %u",
1311                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1312                         r->refcount)) {
1313                 struct bkey_i *new;
1314
1315                 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1316                 if (!new) {
1317                         ret = -ENOMEM;
1318                         goto fsck_err;
1319                 }
1320
1321                 bkey_reassemble(new, k);
1322
1323                 if (!r->refcount) {
1324                         new->k.type = KEY_TYPE_deleted;
1325                         new->k.size = 0;
1326                 } else {
1327                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1328                 }
1329
1330                 ret = bch2_journal_key_insert(c, BTREE_ID_reflink, 0, new);
1331                 if (ret)
1332                         kfree(new);
1333         }
1334 fsck_err:
1335         return ret;
1336 }
1337
1338 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1339                                 bool metadata_only)
1340 {
1341         struct btree_trans trans;
1342         struct btree_iter *iter;
1343         struct bkey_s_c k;
1344         struct reflink_gc *r;
1345         size_t idx = 0;
1346         char buf[200];
1347         int ret = 0;
1348
1349         if (metadata_only)
1350                 return 0;
1351
1352         if (initial) {
1353                 c->reflink_gc_idx = 0;
1354
1355                 ret = bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1356                                 bch2_gc_reflink_done_initial_fn);
1357                 goto out;
1358         }
1359
1360         bch2_trans_init(&trans, c, 0, 0);
1361
1362         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1363                            BTREE_ITER_PREFETCH, k, ret) {
1364                 const __le64 *refcount = bkey_refcount_c(k);
1365
1366                 if (!refcount)
1367                         continue;
1368
1369                 r = genradix_ptr(&c->reflink_gc_table, idx);
1370                 if (!r ||
1371                     r->offset != k.k->p.offset ||
1372                     r->size != k.k->size) {
1373                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1374                         ret = -EINVAL;
1375                         break;
1376                 }
1377
1378                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1379                                 "reflink key has wrong refcount:\n"
1380                                 "  %s\n"
1381                                 "  should be %u",
1382                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1383                                 r->refcount)) {
1384                         struct bkey_i *new;
1385
1386                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1387                         if (!new) {
1388                                 ret = -ENOMEM;
1389                                 break;
1390                         }
1391
1392                         bkey_reassemble(new, k);
1393
1394                         if (!r->refcount)
1395                                 new->k.type = KEY_TYPE_deleted;
1396                         else
1397                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1398
1399                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1400                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1401                         kfree(new);
1402
1403                         if (ret)
1404                                 break;
1405                 }
1406         }
1407 fsck_err:
1408         bch2_trans_iter_put(&trans, iter);
1409         bch2_trans_exit(&trans);
1410 out:
1411         genradix_free(&c->reflink_gc_table);
1412         c->reflink_gc_nr = 0;
1413         return ret;
1414 }
1415
1416 static int bch2_gc_reflink_start_initial_fn(struct bch_fs *c, struct bkey_s_c k)
1417 {
1418
1419         struct reflink_gc *r;
1420         const __le64 *refcount = bkey_refcount_c(k);
1421
1422         if (!refcount)
1423                 return 0;
1424
1425         r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1426                                GFP_KERNEL);
1427         if (!r)
1428                 return -ENOMEM;
1429
1430         r->offset       = k.k->p.offset;
1431         r->size         = k.k->size;
1432         r->refcount     = 0;
1433         return 0;
1434 }
1435
1436 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1437                                  bool metadata_only)
1438 {
1439         struct btree_trans trans;
1440         struct btree_iter *iter;
1441         struct bkey_s_c k;
1442         struct reflink_gc *r;
1443         int ret;
1444
1445         if (metadata_only)
1446                 return 0;
1447
1448         genradix_free(&c->reflink_gc_table);
1449         c->reflink_gc_nr = 0;
1450
1451         if (initial)
1452                 return bch2_btree_and_journal_walk(c, BTREE_ID_reflink,
1453                                 bch2_gc_reflink_start_initial_fn);
1454
1455         bch2_trans_init(&trans, c, 0, 0);
1456
1457         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1458                            BTREE_ITER_PREFETCH, k, ret) {
1459                 const __le64 *refcount = bkey_refcount_c(k);
1460
1461                 if (!refcount)
1462                         continue;
1463
1464                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1465                                        GFP_KERNEL);
1466                 if (!r) {
1467                         ret = -ENOMEM;
1468                         break;
1469                 }
1470
1471                 r->offset       = k.k->p.offset;
1472                 r->size         = k.k->size;
1473                 r->refcount     = 0;
1474         }
1475         bch2_trans_iter_put(&trans, iter);
1476
1477         bch2_trans_exit(&trans);
1478         return 0;
1479 }
1480
1481 /**
1482  * bch2_gc - walk _all_ references to buckets, and recompute them:
1483  *
1484  * Order matters here:
1485  *  - Concurrent GC relies on the fact that we have a total ordering for
1486  *    everything that GC walks - see  gc_will_visit_node(),
1487  *    gc_will_visit_root()
1488  *
1489  *  - also, references move around in the course of index updates and
1490  *    various other crap: everything needs to agree on the ordering
1491  *    references are allowed to move around in - e.g., we're allowed to
1492  *    start with a reference owned by an open_bucket (the allocator) and
1493  *    move it to the btree, but not the reverse.
1494  *
1495  *    This is necessary to ensure that gc doesn't miss references that
1496  *    move around - if references move backwards in the ordering GC
1497  *    uses, GC could skip past them
1498  */
1499 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1500 {
1501         struct bch_dev *ca;
1502         u64 start_time = local_clock();
1503         unsigned i, iter = 0;
1504         int ret;
1505
1506         lockdep_assert_held(&c->state_lock);
1507         trace_gc_start(c);
1508
1509         down_write(&c->gc_lock);
1510
1511         /* flush interior btree updates: */
1512         closure_wait_event(&c->btree_interior_update_wait,
1513                            !bch2_btree_interior_updates_nr_pending(c));
1514 again:
1515         ret   = bch2_gc_start(c, metadata_only) ?:
1516                 bch2_gc_reflink_start(c, initial, metadata_only);
1517         if (ret)
1518                 goto out;
1519
1520         bch2_mark_superblocks(c);
1521
1522         if (test_bit(BCH_FS_TOPOLOGY_ERROR, &c->flags) &&
1523             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1524             c->opts.fix_errors != FSCK_OPT_NO) {
1525                 bch_info(c, "starting topology repair pass");
1526                 ret = bch2_repair_topology(c);
1527                 if (ret)
1528                         goto out;
1529                 bch_info(c, "topology repair pass done");
1530         }
1531
1532         ret = bch2_gc_btrees(c, initial, metadata_only);
1533
1534         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1535             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1536                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1537                 ret = 0;
1538         }
1539
1540         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1541                 ret = FSCK_ERR_EXIT;
1542
1543         if (ret)
1544                 goto out;
1545
1546 #if 0
1547         bch2_mark_pending_btree_node_frees(c);
1548 #endif
1549         c->gc_count++;
1550
1551         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1552             (!iter && bch2_test_restart_gc)) {
1553                 /*
1554                  * XXX: make sure gens we fixed got saved
1555                  */
1556                 if (iter++ <= 2) {
1557                         bch_info(c, "Second GC pass needed, restarting:");
1558                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1559                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1560
1561                         percpu_down_write(&c->mark_lock);
1562                         bch2_gc_free(c);
1563                         percpu_up_write(&c->mark_lock);
1564                         /* flush fsck errors, reset counters */
1565                         bch2_flush_fsck_errs(c);
1566
1567                         goto again;
1568                 }
1569
1570                 bch_info(c, "Unable to fix bucket gens, looping");
1571                 ret = -EINVAL;
1572         }
1573 out:
1574         if (!ret) {
1575                 bch2_journal_block(&c->journal);
1576
1577                 percpu_down_write(&c->mark_lock);
1578                 ret   = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1579                         bch2_gc_done(c, initial, metadata_only);
1580
1581                 bch2_journal_unblock(&c->journal);
1582         } else {
1583                 percpu_down_write(&c->mark_lock);
1584         }
1585
1586         /* Indicates that gc is no longer in progress: */
1587         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1588
1589         bch2_gc_free(c);
1590         percpu_up_write(&c->mark_lock);
1591
1592         up_write(&c->gc_lock);
1593
1594         trace_gc_end(c);
1595         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1596
1597         /*
1598          * Wake up allocator in case it was waiting for buckets
1599          * because of not being able to inc gens
1600          */
1601         for_each_member_device(ca, c, i)
1602                 bch2_wake_allocator(ca);
1603
1604         /*
1605          * At startup, allocations can happen directly instead of via the
1606          * allocator thread - issue wakeup in case they blocked on gc_lock:
1607          */
1608         closure_wake_up(&c->freelist_wait);
1609         return ret;
1610 }
1611
1612 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1613 {
1614         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1615         const struct bch_extent_ptr *ptr;
1616
1617         percpu_down_read(&c->mark_lock);
1618         bkey_for_each_ptr(ptrs, ptr) {
1619                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1620                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1621
1622                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1623                         percpu_up_read(&c->mark_lock);
1624                         return true;
1625                 }
1626         }
1627
1628         bkey_for_each_ptr(ptrs, ptr) {
1629                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1630                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1631
1632                 if (gen_after(g->gc_gen, ptr->gen))
1633                         g->gc_gen = ptr->gen;
1634         }
1635         percpu_up_read(&c->mark_lock);
1636
1637         return false;
1638 }
1639
1640 /*
1641  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1642  * node pointers currently never have cached pointers that can become stale:
1643  */
1644 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1645 {
1646         struct btree_trans trans;
1647         struct btree_iter *iter;
1648         struct bkey_s_c k;
1649         struct bkey_buf sk;
1650         int ret = 0, commit_err = 0;
1651
1652         bch2_bkey_buf_init(&sk);
1653         bch2_trans_init(&trans, c, 0, 0);
1654
1655         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
1656                                    BTREE_ITER_PREFETCH|
1657                                    BTREE_ITER_NOT_EXTENTS|
1658                                    BTREE_ITER_ALL_SNAPSHOTS);
1659
1660         while ((k = bch2_btree_iter_peek(iter)).k &&
1661                !(ret = bkey_err(k))) {
1662                 c->gc_gens_pos = iter->pos;
1663
1664                 if (gc_btree_gens_key(c, k) && !commit_err) {
1665                         bch2_bkey_buf_reassemble(&sk, c, k);
1666                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1667
1668                         bch2_trans_update(&trans, iter, sk.k, 0);
1669
1670                         commit_err = bch2_trans_commit(&trans, NULL, NULL,
1671                                                        BTREE_INSERT_NOWAIT|
1672                                                        BTREE_INSERT_NOFAIL);
1673                         if (commit_err == -EINTR) {
1674                                 commit_err = 0;
1675                                 continue;
1676                         }
1677                 }
1678
1679                 bch2_btree_iter_advance(iter);
1680         }
1681         bch2_trans_iter_put(&trans, iter);
1682
1683         bch2_trans_exit(&trans);
1684         bch2_bkey_buf_exit(&sk, c);
1685
1686         return ret;
1687 }
1688
1689 int bch2_gc_gens(struct bch_fs *c)
1690 {
1691         struct bch_dev *ca;
1692         struct bucket_array *buckets;
1693         struct bucket *g;
1694         unsigned i;
1695         int ret;
1696
1697         /*
1698          * Ideally we would be using state_lock and not gc_lock here, but that
1699          * introduces a deadlock in the RO path - we currently take the state
1700          * lock at the start of going RO, thus the gc thread may get stuck:
1701          */
1702         down_read(&c->gc_lock);
1703
1704         for_each_member_device(ca, c, i) {
1705                 down_read(&ca->bucket_lock);
1706                 buckets = bucket_array(ca);
1707
1708                 for_each_bucket(g, buckets)
1709                         g->gc_gen = g->mark.gen;
1710                 up_read(&ca->bucket_lock);
1711         }
1712
1713         for (i = 0; i < BTREE_ID_NR; i++)
1714                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1715                         c->gc_gens_btree = i;
1716                         c->gc_gens_pos = POS_MIN;
1717                         ret = bch2_gc_btree_gens(c, i);
1718                         if (ret) {
1719                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1720                                 goto err;
1721                         }
1722                 }
1723
1724         for_each_member_device(ca, c, i) {
1725                 down_read(&ca->bucket_lock);
1726                 buckets = bucket_array(ca);
1727
1728                 for_each_bucket(g, buckets)
1729                         g->oldest_gen = g->gc_gen;
1730                 up_read(&ca->bucket_lock);
1731         }
1732
1733         c->gc_gens_btree        = 0;
1734         c->gc_gens_pos          = POS_MIN;
1735
1736         c->gc_count++;
1737 err:
1738         up_read(&c->gc_lock);
1739         return ret;
1740 }
1741
1742 static int bch2_gc_thread(void *arg)
1743 {
1744         struct bch_fs *c = arg;
1745         struct io_clock *clock = &c->io_clock[WRITE];
1746         unsigned long last = atomic64_read(&clock->now);
1747         unsigned last_kick = atomic_read(&c->kick_gc);
1748         int ret;
1749
1750         set_freezable();
1751
1752         while (1) {
1753                 while (1) {
1754                         set_current_state(TASK_INTERRUPTIBLE);
1755
1756                         if (kthread_should_stop()) {
1757                                 __set_current_state(TASK_RUNNING);
1758                                 return 0;
1759                         }
1760
1761                         if (atomic_read(&c->kick_gc) != last_kick)
1762                                 break;
1763
1764                         if (c->btree_gc_periodic) {
1765                                 unsigned long next = last + c->capacity / 16;
1766
1767                                 if (atomic64_read(&clock->now) >= next)
1768                                         break;
1769
1770                                 bch2_io_clock_schedule_timeout(clock, next);
1771                         } else {
1772                                 schedule();
1773                         }
1774
1775                         try_to_freeze();
1776                 }
1777                 __set_current_state(TASK_RUNNING);
1778
1779                 last = atomic64_read(&clock->now);
1780                 last_kick = atomic_read(&c->kick_gc);
1781
1782                 /*
1783                  * Full gc is currently incompatible with btree key cache:
1784                  */
1785 #if 0
1786                 ret = bch2_gc(c, false, false);
1787 #else
1788                 ret = bch2_gc_gens(c);
1789 #endif
1790                 if (ret < 0)
1791                         bch_err(c, "btree gc failed: %i", ret);
1792
1793                 debug_check_no_locks_held();
1794         }
1795
1796         return 0;
1797 }
1798
1799 void bch2_gc_thread_stop(struct bch_fs *c)
1800 {
1801         struct task_struct *p;
1802
1803         p = c->gc_thread;
1804         c->gc_thread = NULL;
1805
1806         if (p) {
1807                 kthread_stop(p);
1808                 put_task_struct(p);
1809         }
1810 }
1811
1812 int bch2_gc_thread_start(struct bch_fs *c)
1813 {
1814         struct task_struct *p;
1815
1816         if (c->gc_thread)
1817                 return 0;
1818
1819         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1820         if (IS_ERR(p)) {
1821                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1822                 return PTR_ERR(p);
1823         }
1824
1825         get_task_struct(p);
1826         c->gc_thread = p;
1827         wake_up_process(p);
1828         return 0;
1829 }