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