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_key_cache.h"
  13 #include "btree_locking.h"
  14 #include "btree_update_interior.h"
  15 #include "btree_io.h"
  16 #include "btree_gc.h"
  17 #include "buckets.h"
  18 #include "clock.h"
  19 #include "debug.h"
  20 #include "ec.h"
  21 #include "error.h"
  22 #include "extents.h"
  23 #include "journal.h"
  24 #include "keylist.h"
  25 #include "move.h"
  26 #include "recovery.h"
  27 #include "reflink.h"
  28 #include "replicas.h"
  29 #include "super-io.h"
  30
  31 #include <linux/slab.h>
  32 #include <linux/bitops.h>
  33 #include <linux/freezer.h>
  34 #include <linux/kthread.h>
  35 #include <linux/preempt.h>
  36 #include <linux/rcupdate.h>
  37 #include <linux/sched/task.h>
  38 #include <trace/events/bcachefs.h>
  39
  40 #define DROP_THIS_NODE          10
  41 #define DROP_PREV_NODE          11
  42
  43 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  44 {
  45         preempt_disable();
  46         write_seqcount_begin(&c->gc_pos_lock);
  47         c->gc_pos = new_pos;
  48         write_seqcount_end(&c->gc_pos_lock);
  49         preempt_enable();
  50 }
  51
  52 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  53 {
  54         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  55         __gc_pos_set(c, new_pos);
  56 }
  57
  58 /*
  59  * Missing: if an interior btree node is empty, we need to do something -
  60  * perhaps just kill it
  61  */
  62 static int bch2_gc_check_topology(struct bch_fs *c,
  63                                   struct btree *b,
  64                                   struct bkey_buf *prev,
  65                                   struct bkey_buf cur,
  66                                   bool is_last)
  67 {
  68         struct bpos node_start  = b->data->min_key;
  69         struct bpos node_end    = b->data->max_key;
  70         struct bpos expected_start = bkey_deleted(&prev->k->k)
  71                 ? node_start
  72                 : bpos_successor(prev->k->k.p);
  73         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
  74         int ret = 0;
  75
  76         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  77                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  78
  79                 if (bpos_cmp(expected_start, bp->v.min_key)) {
  80                         bch2_topology_error(c);
  81
  82                         if (bkey_deleted(&prev->k->k)) {
  83                                 prt_printf(&buf1, "start of node: ");
  84                                 bch2_bpos_to_text(&buf1, node_start);
  85                         } else {
  86                                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(prev->k));
  87                         }
  88                         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(cur.k));
  89
  90                         if (__fsck_err(c,
  91                                   FSCK_CAN_FIX|
  92                                   FSCK_CAN_IGNORE|
  93                                   FSCK_NO_RATELIMIT,
  94                                   "btree node with incorrect min_key at btree %s level %u:\n"
  95                                   "  prev %s\n"
  96                                   "  cur %s",
  97                                   bch2_btree_ids[b->c.btree_id], b->c.level,
  98                                   buf1.buf, buf2.buf) &&
  99                             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 100                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 101                                 ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 102                                 goto err;
 103                         } else {
 104                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 105                         }
 106                 }
 107         }
 108
 109         if (is_last && bpos_cmp(cur.k->k.p, node_end)) {
 110                 bch2_topology_error(c);
 111
 112                 printbuf_reset(&buf1);
 113                 printbuf_reset(&buf2);
 114
 115                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(cur.k));
 116                 bch2_bpos_to_text(&buf2, node_end);
 117
 118                 if (__fsck_err(c,
 119                           FSCK_CAN_FIX|
 120                           FSCK_CAN_IGNORE|
 121                           FSCK_NO_RATELIMIT,
 122                           "btree node with incorrect max_key at btree %s level %u:\n"
 123                           "  %s\n"
 124                           "  expected %s",
 125                           bch2_btree_ids[b->c.btree_id], b->c.level,
 126                           buf1.buf, buf2.buf) &&
 127                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 128                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 129                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 130                         goto err;
 131                 } else {
 132                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 133                 }
 134         }
 135
 136         bch2_bkey_buf_copy(prev, c, cur.k);
 137 err:
 138 fsck_err:
 139         printbuf_exit(&buf2);
 140         printbuf_exit(&buf1);
 141         return ret;
 142 }
 143
 144 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 145 {
 146         switch (b->key.k.type) {
 147         case KEY_TYPE_btree_ptr: {
 148                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 149
 150                 dst->k.p                = src->k.p;
 151                 dst->v.mem_ptr          = 0;
 152                 dst->v.seq              = b->data->keys.seq;
 153                 dst->v.sectors_written  = 0;
 154                 dst->v.flags            = 0;
 155                 dst->v.min_key          = b->data->min_key;
 156                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 157                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 158                 break;
 159         }
 160         case KEY_TYPE_btree_ptr_v2:
 161                 bkey_copy(&dst->k_i, &b->key);
 162                 break;
 163         default:
 164                 BUG();
 165         }
 166 }
 167
 168 static void bch2_btree_node_update_key_early(struct bch_fs *c,
 169                                              enum btree_id btree, unsigned level,
 170                                              struct bkey_s_c old, struct bkey_i *new)
 171 {
 172         struct btree *b;
 173         struct bkey_buf tmp;
 174         int ret;
 175
 176         bch2_bkey_buf_init(&tmp);
 177         bch2_bkey_buf_reassemble(&tmp, c, old);
 178
 179         b = bch2_btree_node_get_noiter(c, tmp.k, btree, level, true);
 180         if (!IS_ERR_OR_NULL(b)) {
 181                 mutex_lock(&c->btree_cache.lock);
 182
 183                 bch2_btree_node_hash_remove(&c->btree_cache, b);
 184
 185                 bkey_copy(&b->key, new);
 186                 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 187                 BUG_ON(ret);
 188
 189                 mutex_unlock(&c->btree_cache.lock);
 190                 six_unlock_read(&b->c.lock);
 191         }
 192
 193         bch2_bkey_buf_exit(&tmp, c);
 194 }
 195
 196 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 197 {
 198         struct bkey_i_btree_ptr_v2 *new;
 199         int ret;
 200
 201         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 202         if (!new)
 203                 return -ENOMEM;
 204
 205         btree_ptr_to_v2(b, new);
 206         b->data->min_key        = new_min;
 207         new->v.min_key          = new_min;
 208         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 209
 210         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 211         if (ret) {
 212                 kfree(new);
 213                 return ret;
 214         }
 215
 216         bch2_btree_node_drop_keys_outside_node(b);
 217         bkey_copy(&b->key, &new->k_i);
 218         return 0;
 219 }
 220
 221 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 222 {
 223         struct bkey_i_btree_ptr_v2 *new;
 224         int ret;
 225
 226         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 227         if (ret)
 228                 return ret;
 229
 230         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 231         if (!new)
 232                 return -ENOMEM;
 233
 234         btree_ptr_to_v2(b, new);
 235         b->data->max_key        = new_max;
 236         new->k.p                = new_max;
 237         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 238
 239         ret = bch2_journal_key_insert_take(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 240         if (ret) {
 241                 kfree(new);
 242                 return ret;
 243         }
 244
 245         bch2_btree_node_drop_keys_outside_node(b);
 246
 247         mutex_lock(&c->btree_cache.lock);
 248         bch2_btree_node_hash_remove(&c->btree_cache, b);
 249
 250         bkey_copy(&b->key, &new->k_i);
 251         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 252         BUG_ON(ret);
 253         mutex_unlock(&c->btree_cache.lock);
 254         return 0;
 255 }
 256
 257 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 258                                         struct btree *prev, struct btree *cur)
 259 {
 260         struct bpos expected_start = !prev
 261                 ? b->data->min_key
 262                 : bpos_successor(prev->key.k.p);
 263         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 264         int ret = 0;
 265
 266         if (!prev) {
 267                 prt_printf(&buf1, "start of node: ");
 268                 bch2_bpos_to_text(&buf1, b->data->min_key);
 269         } else {
 270                 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&prev->key));
 271         }
 272
 273         bch2_bkey_val_to_text(&buf2, c, bkey_i_to_s_c(&cur->key));
 274
 275         if (prev &&
 276             bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 277             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 278                 /* cur overwrites prev: */
 279
 280                 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
 281                                                  cur->data->min_key) >= 0, c,
 282                                 "btree node overwritten by next node at btree %s level %u:\n"
 283                                 "  node %s\n"
 284                                 "  next %s",
 285                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 286                                 buf1.buf, buf2.buf)) {
 287                         ret = DROP_PREV_NODE;
 288                         goto out;
 289                 }
 290
 291                 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
 292                                                  bpos_predecessor(cur->data->min_key)), c,
 293                                 "btree node with incorrect max_key at btree %s level %u:\n"
 294                                 "  node %s\n"
 295                                 "  next %s",
 296                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 297                                 buf1.buf, buf2.buf))
 298                         ret = set_node_max(c, prev,
 299                                            bpos_predecessor(cur->data->min_key));
 300         } else {
 301                 /* prev overwrites cur: */
 302
 303                 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
 304                                                  cur->data->max_key) >= 0, c,
 305                                 "btree node overwritten by prev node at btree %s level %u:\n"
 306                                 "  prev %s\n"
 307                                 "  node %s",
 308                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 309                                 buf1.buf, buf2.buf)) {
 310                         ret = DROP_THIS_NODE;
 311                         goto out;
 312                 }
 313
 314                 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 315                                 "btree node with incorrect min_key at btree %s level %u:\n"
 316                                 "  prev %s\n"
 317                                 "  node %s",
 318                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 319                                 buf1.buf, buf2.buf))
 320                     ret = set_node_min(c, cur, expected_start);
 321         }
 322 out:
 323 fsck_err:
 324         printbuf_exit(&buf2);
 325         printbuf_exit(&buf1);
 326         return ret;
 327 }
 328
 329 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 330                                  struct btree *child)
 331 {
 332         struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
 333         int ret = 0;
 334
 335         bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&child->key));
 336         bch2_bpos_to_text(&buf2, b->key.k.p);
 337
 338         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 339                         "btree node with incorrect max_key at btree %s level %u:\n"
 340                         "  %s\n"
 341                         "  expected %s",
 342                         bch2_btree_ids[b->c.btree_id], b->c.level,
 343                         buf1.buf, buf2.buf)) {
 344                 ret = set_node_max(c, child, b->key.k.p);
 345                 if (ret)
 346                         goto err;
 347         }
 348 err:
 349 fsck_err:
 350         printbuf_exit(&buf2);
 351         printbuf_exit(&buf1);
 352         return ret;
 353 }
 354
 355 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 356 {
 357         struct btree_and_journal_iter iter;
 358         struct bkey_s_c k;
 359         struct bkey_buf prev_k, cur_k;
 360         struct btree *prev = NULL, *cur = NULL;
 361         bool have_child, dropped_children = false;
 362         struct printbuf buf = PRINTBUF;
 363         int ret = 0;
 364
 365         if (!b->c.level)
 366                 return 0;
 367 again:
 368         prev = NULL;
 369         have_child = dropped_children = false;
 370         bch2_bkey_buf_init(&prev_k);
 371         bch2_bkey_buf_init(&cur_k);
 372         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 373
 374         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 375                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 376                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 377
 378                 bch2_btree_and_journal_iter_advance(&iter);
 379                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 380
 381                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 382                                         b->c.btree_id, b->c.level - 1,
 383                                         false);
 384                 ret = PTR_ERR_OR_ZERO(cur);
 385
 386                 printbuf_reset(&buf);
 387                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 388
 389                 if (mustfix_fsck_err_on(ret == -EIO, c,
 390                                 "Topology repair: unreadable btree node at btree %s level %u:\n"
 391                                 "  %s",
 392                                 bch2_btree_ids[b->c.btree_id],
 393                                 b->c.level - 1,
 394                                 buf.buf)) {
 395                         bch2_btree_node_evict(c, cur_k.k);
 396                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 397                                                       b->c.level, cur_k.k->k.p);
 398                         cur = NULL;
 399                         if (ret)
 400                                 break;
 401                         continue;
 402                 }
 403
 404                 if (ret) {
 405                         bch_err(c, "%s: error %i getting btree node",
 406                                 __func__, ret);
 407                         break;
 408                 }
 409
 410                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 411
 412                 if (ret == DROP_THIS_NODE) {
 413                         six_unlock_read(&cur->c.lock);
 414                         bch2_btree_node_evict(c, cur_k.k);
 415                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 416                                                       b->c.level, cur_k.k->k.p);
 417                         cur = NULL;
 418                         if (ret)
 419                                 break;
 420                         continue;
 421                 }
 422
 423                 if (prev)
 424                         six_unlock_read(&prev->c.lock);
 425                 prev = NULL;
 426
 427                 if (ret == DROP_PREV_NODE) {
 428                         bch2_btree_node_evict(c, prev_k.k);
 429                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 430                                                       b->c.level, prev_k.k->k.p);
 431                         if (ret)
 432                                 break;
 433
 434                         bch2_btree_and_journal_iter_exit(&iter);
 435                         bch2_bkey_buf_exit(&prev_k, c);
 436                         bch2_bkey_buf_exit(&cur_k, c);
 437                         goto again;
 438                 } else if (ret)
 439                         break;
 440
 441                 prev = cur;
 442                 cur = NULL;
 443                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 444         }
 445
 446         if (!ret && !IS_ERR_OR_NULL(prev)) {
 447                 BUG_ON(cur);
 448                 ret = btree_repair_node_end(c, b, prev);
 449         }
 450
 451         if (!IS_ERR_OR_NULL(prev))
 452                 six_unlock_read(&prev->c.lock);
 453         prev = NULL;
 454         if (!IS_ERR_OR_NULL(cur))
 455                 six_unlock_read(&cur->c.lock);
 456         cur = NULL;
 457
 458         if (ret)
 459                 goto err;
 460
 461         bch2_btree_and_journal_iter_exit(&iter);
 462         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 463
 464         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 465                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 466                 bch2_btree_and_journal_iter_advance(&iter);
 467
 468                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 469                                         b->c.btree_id, b->c.level - 1,
 470                                         false);
 471                 ret = PTR_ERR_OR_ZERO(cur);
 472
 473                 if (ret) {
 474                         bch_err(c, "%s: error %i getting btree node",
 475                                 __func__, ret);
 476                         goto err;
 477                 }
 478
 479                 ret = bch2_btree_repair_topology_recurse(c, cur);
 480                 six_unlock_read(&cur->c.lock);
 481                 cur = NULL;
 482
 483                 if (ret == DROP_THIS_NODE) {
 484                         bch2_btree_node_evict(c, cur_k.k);
 485                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 486                                                       b->c.level, cur_k.k->k.p);
 487                         dropped_children = true;
 488                 }
 489
 490                 if (ret)
 491                         goto err;
 492
 493                 have_child = true;
 494         }
 495
 496         printbuf_reset(&buf);
 497         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 498
 499         if (mustfix_fsck_err_on(!have_child, c,
 500                         "empty interior btree node at btree %s level %u\n"
 501                         "  %s",
 502                         bch2_btree_ids[b->c.btree_id],
 503                         b->c.level, buf.buf))
 504                 ret = DROP_THIS_NODE;
 505 err:
 506 fsck_err:
 507         if (!IS_ERR_OR_NULL(prev))
 508                 six_unlock_read(&prev->c.lock);
 509         if (!IS_ERR_OR_NULL(cur))
 510                 six_unlock_read(&cur->c.lock);
 511
 512         bch2_btree_and_journal_iter_exit(&iter);
 513         bch2_bkey_buf_exit(&prev_k, c);
 514         bch2_bkey_buf_exit(&cur_k, c);
 515
 516         if (!ret && dropped_children)
 517                 goto again;
 518
 519         printbuf_exit(&buf);
 520         return ret;
 521 }
 522
 523 static int bch2_repair_topology(struct bch_fs *c)
 524 {
 525         struct btree *b;
 526         unsigned i;
 527         int ret = 0;
 528
 529         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 530                 b = c->btree_roots[i].b;
 531                 if (btree_node_fake(b))
 532                         continue;
 533
 534                 six_lock_read(&b->c.lock, NULL, NULL);
 535                 ret = bch2_btree_repair_topology_recurse(c, b);
 536                 six_unlock_read(&b->c.lock);
 537
 538                 if (ret == DROP_THIS_NODE) {
 539                         bch_err(c, "empty btree root - repair unimplemented");
 540                         ret = FSCK_ERR_EXIT;
 541                 }
 542         }
 543
 544         return ret;
 545 }
 546
 547 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 548                                unsigned level, bool is_root,
 549                                struct bkey_s_c *k)
 550 {
 551         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 552         const union bch_extent_entry *entry;
 553         struct extent_ptr_decoded p = { 0 };
 554         bool do_update = false;
 555         struct printbuf buf = PRINTBUF;
 556         int ret = 0;
 557
 558         /*
 559          * XXX
 560          * use check_bucket_ref here
 561          */
 562         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 563                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 564                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 565                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 566
 567                 if (fsck_err_on(!g->gen_valid, c,
 568                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 569                                 "while marking %s",
 570                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 571                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 572                                 p.ptr.gen,
 573                                 (printbuf_reset(&buf),
 574                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 575                         if (!p.ptr.cached) {
 576                                 g->gen_valid            = true;
 577                                 g->gen                  = p.ptr.gen;
 578                         } else {
 579                                 do_update = true;
 580                         }
 581                 }
 582
 583                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->gen) > 0, c,
 584                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 585                                 "while marking %s",
 586                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 587                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 588                                 p.ptr.gen, g->gen,
 589                                 (printbuf_reset(&buf),
 590                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 591                         if (!p.ptr.cached) {
 592                                 g->gen_valid            = true;
 593                                 g->gen                  = p.ptr.gen;
 594                                 g->data_type            = 0;
 595                                 g->dirty_sectors        = 0;
 596                                 g->cached_sectors       = 0;
 597                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 598                         } else {
 599                                 do_update = true;
 600                         }
 601                 }
 602
 603                 if (fsck_err_on(gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
 604                                 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 605                                 "while marking %s",
 606                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 607                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 608                                 p.ptr.gen,
 609                                 (printbuf_reset(&buf),
 610                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 611                         do_update = true;
 612
 613                 if (fsck_err_on(!p.ptr.cached &&
 614                                 gen_cmp(p.ptr.gen, g->gen) < 0, c,
 615                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 616                                 "while marking %s",
 617                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 618                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 619                                 p.ptr.gen, g->gen,
 620                                 (printbuf_reset(&buf),
 621                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 622                         do_update = true;
 623
 624                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 625                         continue;
 626
 627                 if (fsck_err_on(g->data_type &&
 628                                 g->data_type != data_type, c,
 629                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 630                                 "while marking %s",
 631                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 632                                 bch2_data_types[g->data_type],
 633                                 bch2_data_types[data_type],
 634                                 (printbuf_reset(&buf),
 635                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 636                         if (data_type == BCH_DATA_btree) {
 637                                 g->data_type    = data_type;
 638                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 639                         } else {
 640                                 do_update = true;
 641                         }
 642                 }
 643
 644                 if (p.has_ec) {
 645                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 646
 647                         if (fsck_err_on(!m || !m->alive, c,
 648                                         "pointer to nonexistent stripe %llu\n"
 649                                         "while marking %s",
 650                                         (u64) p.ec.idx,
 651                                         (printbuf_reset(&buf),
 652                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 653                                 do_update = true;
 654
 655                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 656                                         "pointer does not match stripe %llu\n"
 657                                         "while marking %s",
 658                                         (u64) p.ec.idx,
 659                                         (printbuf_reset(&buf),
 660                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 661                                 do_update = true;
 662                 }
 663         }
 664
 665         if (do_update) {
 666                 struct bkey_ptrs ptrs;
 667                 union bch_extent_entry *entry;
 668                 struct bch_extent_ptr *ptr;
 669                 struct bkey_i *new;
 670
 671                 if (is_root) {
 672                         bch_err(c, "cannot update btree roots yet");
 673                         ret = -EINVAL;
 674                         goto err;
 675                 }
 676
 677                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 678                 if (!new) {
 679                         bch_err(c, "%s: error allocating new key", __func__);
 680                         ret = -ENOMEM;
 681                         goto err;
 682                 }
 683
 684                 bkey_reassemble(new, *k);
 685
 686                 if (level) {
 687                         /*
 688                          * We don't want to drop btree node pointers - if the
 689                          * btree node isn't there anymore, the read path will
 690                          * sort it out:
 691                          */
 692                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 693                         bkey_for_each_ptr(ptrs, ptr) {
 694                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 695                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 696
 697                                 ptr->gen = g->gen;
 698                         }
 699                 } else {
 700                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 701                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 702                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 703                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 704
 705                                 (ptr->cached &&
 706                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 707                                 (!ptr->cached &&
 708                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 709                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 710                                 (g->data_type &&
 711                                  g->data_type != data_type);
 712                         }));
 713 again:
 714                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 715                         bkey_extent_entry_for_each(ptrs, entry) {
 716                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 717                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 718                                                                         entry->stripe_ptr.idx);
 719                                         union bch_extent_entry *next_ptr;
 720
 721                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 722                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 723                                                         goto found;
 724                                         next_ptr = NULL;
 725 found:
 726                                         if (!next_ptr) {
 727                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 728                                                 continue;
 729                                         }
 730
 731                                         if (!m || !m->alive ||
 732                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 733                                                                        &next_ptr->ptr,
 734                                                                        m->sectors)) {
 735                                                 bch2_bkey_extent_entry_drop(new, entry);
 736                                                 goto again;
 737                                         }
 738                                 }
 739                         }
 740                 }
 741
 742                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 743                 if (ret) {
 744                         kfree(new);
 745                         goto err;
 746                 }
 747
 748                 if (level)
 749                         bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
 750
 751                 printbuf_reset(&buf);
 752                 bch2_bkey_val_to_text(&buf, c, *k);
 753                 bch_info(c, "updated %s", buf.buf);
 754
 755                 printbuf_reset(&buf);
 756                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 757                 bch_info(c, "new key %s", buf.buf);
 758
 759                 *k = bkey_i_to_s_c(new);
 760         }
 761 err:
 762 fsck_err:
 763         printbuf_exit(&buf);
 764         return ret;
 765 }
 766
 767 /* marking of btree keys/nodes: */
 768
 769 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 770                             unsigned level, bool is_root,
 771                             struct bkey_s_c *k,
 772                             bool initial)
 773 {
 774         struct bch_fs *c = trans->c;
 775         struct bkey deleted = KEY(0, 0, 0);
 776         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 777         unsigned flags =
 778                 BTREE_TRIGGER_GC|
 779                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 780         int ret = 0;
 781
 782         deleted.p = k->k->p;
 783
 784         if (initial) {
 785                 BUG_ON(bch2_journal_seq_verify &&
 786                        k->k->version.lo > atomic64_read(&c->journal.seq));
 787
 788                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 789                 if (ret)
 790                         goto err;
 791
 792                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 793                                 "key version number higher than recorded: %llu > %llu",
 794                                 k->k->version.lo,
 795                                 atomic64_read(&c->key_version)))
 796                         atomic64_set(&c->key_version, k->k->version.lo);
 797         }
 798
 799         ret = __bch2_trans_do(trans, NULL, NULL, 0,
 800                         bch2_mark_key(trans, old, *k, flags));
 801 fsck_err:
 802 err:
 803         if (ret)
 804                 bch_err(c, "%s: ret %i", __func__, ret);
 805         return ret;
 806 }
 807
 808 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 809 {
 810         struct bch_fs *c = trans->c;
 811         struct btree_node_iter iter;
 812         struct bkey unpacked;
 813         struct bkey_s_c k;
 814         struct bkey_buf prev, cur;
 815         int ret = 0;
 816
 817         if (!btree_node_type_needs_gc(btree_node_type(b)))
 818                 return 0;
 819
 820         bch2_btree_node_iter_init_from_start(&iter, b);
 821         bch2_bkey_buf_init(&prev);
 822         bch2_bkey_buf_init(&cur);
 823         bkey_init(&prev.k->k);
 824
 825         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 826                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 827                                        &k, initial);
 828                 if (ret)
 829                         break;
 830
 831                 bch2_btree_node_iter_advance(&iter, b);
 832
 833                 if (b->c.level) {
 834                         bch2_bkey_buf_reassemble(&cur, c, k);
 835
 836                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 837                                         bch2_btree_node_iter_end(&iter));
 838                         if (ret)
 839                                 break;
 840                 }
 841         }
 842
 843         bch2_bkey_buf_exit(&cur, c);
 844         bch2_bkey_buf_exit(&prev, c);
 845         return ret;
 846 }
 847
 848 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 849                          bool initial, bool metadata_only)
 850 {
 851         struct bch_fs *c = trans->c;
 852         struct btree_iter iter;
 853         struct btree *b;
 854         unsigned depth = metadata_only ? 1 : 0;
 855         int ret = 0;
 856
 857         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 858
 859         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 860                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 861                 bch2_verify_btree_nr_keys(b);
 862
 863                 gc_pos_set(c, gc_pos_btree_node(b));
 864
 865                 ret = btree_gc_mark_node(trans, b, initial);
 866                 if (ret)
 867                         break;
 868         }
 869         bch2_trans_iter_exit(trans, &iter);
 870
 871         if (ret)
 872                 return ret;
 873
 874         mutex_lock(&c->btree_root_lock);
 875         b = c->btree_roots[btree_id].b;
 876         if (!btree_node_fake(b)) {
 877                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 878
 879                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 880                                        true, &k, initial);
 881         }
 882         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 883         mutex_unlock(&c->btree_root_lock);
 884
 885         return ret;
 886 }
 887
 888 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 889                                       unsigned target_depth)
 890 {
 891         struct bch_fs *c = trans->c;
 892         struct btree_and_journal_iter iter;
 893         struct bkey_s_c k;
 894         struct bkey_buf cur, prev;
 895         struct printbuf buf = PRINTBUF;
 896         int ret = 0;
 897
 898         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 899         bch2_bkey_buf_init(&prev);
 900         bch2_bkey_buf_init(&cur);
 901         bkey_init(&prev.k->k);
 902
 903         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 904                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 905                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 906
 907                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 908                                        false, &k, true);
 909                 if (ret) {
 910                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 911                         goto fsck_err;
 912                 }
 913
 914                 if (b->c.level) {
 915                         bch2_bkey_buf_reassemble(&cur, c, k);
 916                         k = bkey_i_to_s_c(cur.k);
 917
 918                         bch2_btree_and_journal_iter_advance(&iter);
 919
 920                         ret = bch2_gc_check_topology(c, b,
 921                                         &prev, cur,
 922                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 923                         if (ret)
 924                                 goto fsck_err;
 925                 } else {
 926                         bch2_btree_and_journal_iter_advance(&iter);
 927                 }
 928         }
 929
 930         if (b->c.level > target_depth) {
 931                 bch2_btree_and_journal_iter_exit(&iter);
 932                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 933
 934                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 935                         struct btree *child;
 936
 937                         bch2_bkey_buf_reassemble(&cur, c, k);
 938                         bch2_btree_and_journal_iter_advance(&iter);
 939
 940                         child = bch2_btree_node_get_noiter(c, cur.k,
 941                                                 b->c.btree_id, b->c.level - 1,
 942                                                 false);
 943                         ret = PTR_ERR_OR_ZERO(child);
 944
 945                         if (ret == -EIO) {
 946                                 bch2_topology_error(c);
 947
 948                                 if (__fsck_err(c,
 949                                           FSCK_CAN_FIX|
 950                                           FSCK_CAN_IGNORE|
 951                                           FSCK_NO_RATELIMIT,
 952                                           "Unreadable btree node at btree %s level %u:\n"
 953                                           "  %s",
 954                                           bch2_btree_ids[b->c.btree_id],
 955                                           b->c.level - 1,
 956                                           (printbuf_reset(&buf),
 957                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 958                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 959                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 960                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 961                                         goto fsck_err;
 962                                 } else {
 963                                         /* Continue marking when opted to not
 964                                          * fix the error: */
 965                                         ret = 0;
 966                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 967                                         continue;
 968                                 }
 969                         } else if (ret) {
 970                                 bch_err(c, "%s: error %i getting btree node",
 971                                         __func__, ret);
 972                                 break;
 973                         }
 974
 975                         ret = bch2_gc_btree_init_recurse(trans, child,
 976                                                          target_depth);
 977                         six_unlock_read(&child->c.lock);
 978
 979                         if (ret)
 980                                 break;
 981                 }
 982         }
 983 fsck_err:
 984         bch2_bkey_buf_exit(&cur, c);
 985         bch2_bkey_buf_exit(&prev, c);
 986         bch2_btree_and_journal_iter_exit(&iter);
 987         printbuf_exit(&buf);
 988         return ret;
 989 }
 990
 991 static int bch2_gc_btree_init(struct btree_trans *trans,
 992                               enum btree_id btree_id,
 993                               bool metadata_only)
 994 {
 995         struct bch_fs *c = trans->c;
 996         struct btree *b;
 997         unsigned target_depth = metadata_only ? 1 : 0;
 998         struct printbuf buf = PRINTBUF;
 999         int ret = 0;
1000
1001         b = c->btree_roots[btree_id].b;
1002
1003         if (btree_node_fake(b))
1004                 return 0;
1005
1006         six_lock_read(&b->c.lock, NULL, NULL);
1007         printbuf_reset(&buf);
1008         bch2_bpos_to_text(&buf, b->data->min_key);
1009         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1010                         "btree root with incorrect min_key: %s", buf.buf)) {
1011                 bch_err(c, "repair unimplemented");
1012                 ret = FSCK_ERR_EXIT;
1013                 goto fsck_err;
1014         }
1015
1016         printbuf_reset(&buf);
1017         bch2_bpos_to_text(&buf, b->data->max_key);
1018         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1019                         "btree root with incorrect max_key: %s", buf.buf)) {
1020                 bch_err(c, "repair unimplemented");
1021                 ret = FSCK_ERR_EXIT;
1022                 goto fsck_err;
1023         }
1024
1025         if (b->c.level >= target_depth)
1026                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1027
1028         if (!ret) {
1029                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1030
1031                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1032                                        &k, true);
1033         }
1034 fsck_err:
1035         six_unlock_read(&b->c.lock);
1036
1037         if (ret < 0)
1038                 bch_err(c, "%s: ret %i", __func__, ret);
1039         printbuf_exit(&buf);
1040         return ret;
1041 }
1042
1043 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1044 {
1045         return  (int) btree_id_to_gc_phase(l) -
1046                 (int) btree_id_to_gc_phase(r);
1047 }
1048
1049 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1050 {
1051         struct btree_trans trans;
1052         enum btree_id ids[BTREE_ID_NR];
1053         unsigned i;
1054         int ret = 0;
1055
1056         bch2_trans_init(&trans, c, 0, 0);
1057
1058         if (initial)
1059                 trans.is_initial_gc = true;
1060
1061         for (i = 0; i < BTREE_ID_NR; i++)
1062                 ids[i] = i;
1063         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1064
1065         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1066                 ret = initial
1067                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1068                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1069
1070         if (ret < 0)
1071                 bch_err(c, "%s: ret %i", __func__, ret);
1072
1073         bch2_trans_exit(&trans);
1074         return ret;
1075 }
1076
1077 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1078                                   u64 start, u64 end,
1079                                   enum bch_data_type type,
1080                                   unsigned flags)
1081 {
1082         u64 b = sector_to_bucket(ca, start);
1083
1084         do {
1085                 unsigned sectors =
1086                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1087
1088                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1089                                           gc_phase(GC_PHASE_SB), flags);
1090                 b++;
1091                 start += sectors;
1092         } while (start < end);
1093 }
1094
1095 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1096                                      unsigned flags)
1097 {
1098         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1099         unsigned i;
1100         u64 b;
1101
1102         for (i = 0; i < layout->nr_superblocks; i++) {
1103                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1104
1105                 if (offset == BCH_SB_SECTOR)
1106                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1107                                               BCH_DATA_sb, flags);
1108
1109                 mark_metadata_sectors(c, ca, offset,
1110                                       offset + (1 << layout->sb_max_size_bits),
1111                                       BCH_DATA_sb, flags);
1112         }
1113
1114         for (i = 0; i < ca->journal.nr; i++) {
1115                 b = ca->journal.buckets[i];
1116                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1117                                           ca->mi.bucket_size,
1118                                           gc_phase(GC_PHASE_SB), flags);
1119         }
1120 }
1121
1122 static void bch2_mark_superblocks(struct bch_fs *c)
1123 {
1124         struct bch_dev *ca;
1125         unsigned i;
1126
1127         mutex_lock(&c->sb_lock);
1128         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1129
1130         for_each_online_member(ca, c, i)
1131                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1132         mutex_unlock(&c->sb_lock);
1133 }
1134
1135 #if 0
1136 /* Also see bch2_pending_btree_node_free_insert_done() */
1137 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1138 {
1139         struct btree_update *as;
1140         struct pending_btree_node_free *d;
1141
1142         mutex_lock(&c->btree_interior_update_lock);
1143         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1144
1145         for_each_pending_btree_node_free(c, as, d)
1146                 if (d->index_update_done)
1147                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1148
1149         mutex_unlock(&c->btree_interior_update_lock);
1150 }
1151 #endif
1152
1153 static void bch2_gc_free(struct bch_fs *c)
1154 {
1155         struct bch_dev *ca;
1156         unsigned i;
1157
1158         genradix_free(&c->reflink_gc_table);
1159         genradix_free(&c->gc_stripes);
1160
1161         for_each_member_device(ca, c, i) {
1162                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1163                         sizeof(struct bucket_array) +
1164                         ca->mi.nbuckets * sizeof(struct bucket));
1165                 ca->buckets_gc = NULL;
1166
1167                 free_percpu(ca->usage_gc);
1168                 ca->usage_gc = NULL;
1169         }
1170
1171         free_percpu(c->usage_gc);
1172         c->usage_gc = NULL;
1173 }
1174
1175 static int bch2_gc_done(struct bch_fs *c,
1176                         bool initial, bool metadata_only)
1177 {
1178         struct bch_dev *ca = NULL;
1179         struct printbuf buf = PRINTBUF;
1180         bool verify = !metadata_only && (!initial ||
1181                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1182         unsigned i, dev;
1183         int ret = 0;
1184
1185         percpu_down_write(&c->mark_lock);
1186
1187 #define copy_field(_f, _msg, ...)                                       \
1188         if (dst->_f != src->_f) {                                       \
1189                 if (verify)                                             \
1190                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1191                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1192                 dst->_f = src->_f;                                      \
1193         }
1194 #define copy_stripe_field(_f, _msg, ...)                                \
1195         if (dst->_f != src->_f) {                                       \
1196                 if (verify)                                             \
1197                         fsck_err(c, "stripe %zu has wrong "_msg         \
1198                                 ": got %u, should be %u",               \
1199                                 iter.pos, ##__VA_ARGS__,                \
1200                                 dst->_f, src->_f);                      \
1201                 dst->_f = src->_f;                                      \
1202         }
1203 #define copy_dev_field(_f, _msg, ...)                                   \
1204         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1205 #define copy_fs_field(_f, _msg, ...)                                    \
1206         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1207
1208         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1209                 bch2_fs_usage_acc_to_base(c, i);
1210
1211         for_each_member_device(ca, c, dev) {
1212                 struct bch_dev_usage *dst = ca->usage_base;
1213                 struct bch_dev_usage *src = (void *)
1214                         bch2_acc_percpu_u64s((void *) ca->usage_gc,
1215                                              dev_usage_u64s());
1216
1217                 copy_dev_field(buckets_ec,              "buckets_ec");
1218
1219                 for (i = 0; i < BCH_DATA_NR; i++) {
1220                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1221                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1222                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1223                 }
1224         };
1225
1226         {
1227                 unsigned nr = fs_usage_u64s(c);
1228                 struct bch_fs_usage *dst = c->usage_base;
1229                 struct bch_fs_usage *src = (void *)
1230                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1231
1232                 copy_fs_field(hidden,           "hidden");
1233                 copy_fs_field(btree,            "btree");
1234
1235                 if (!metadata_only) {
1236                         copy_fs_field(data,     "data");
1237                         copy_fs_field(cached,   "cached");
1238                         copy_fs_field(reserved, "reserved");
1239                         copy_fs_field(nr_inodes,"nr_inodes");
1240
1241                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1242                                 copy_fs_field(persistent_reserved[i],
1243                                               "persistent_reserved[%i]", i);
1244                 }
1245
1246                 for (i = 0; i < c->replicas.nr; i++) {
1247                         struct bch_replicas_entry *e =
1248                                 cpu_replicas_entry(&c->replicas, i);
1249
1250                         if (metadata_only &&
1251                             (e->data_type == BCH_DATA_user ||
1252                              e->data_type == BCH_DATA_cached))
1253                                 continue;
1254
1255                         printbuf_reset(&buf);
1256                         bch2_replicas_entry_to_text(&buf, e);
1257
1258                         copy_fs_field(replicas[i], "%s", buf.buf);
1259                 }
1260         }
1261
1262 #undef copy_fs_field
1263 #undef copy_dev_field
1264 #undef copy_stripe_field
1265 #undef copy_field
1266 fsck_err:
1267         if (ca)
1268                 percpu_ref_put(&ca->ref);
1269         if (ret)
1270                 bch_err(c, "%s: ret %i", __func__, ret);
1271
1272         percpu_up_write(&c->mark_lock);
1273         printbuf_exit(&buf);
1274         return ret;
1275 }
1276
1277 static int bch2_gc_start(struct bch_fs *c,
1278                          bool metadata_only)
1279 {
1280         struct bch_dev *ca = NULL;
1281         unsigned i;
1282
1283         BUG_ON(c->usage_gc);
1284
1285         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1286                                          sizeof(u64), GFP_KERNEL);
1287         if (!c->usage_gc) {
1288                 bch_err(c, "error allocating c->usage_gc");
1289                 return -ENOMEM;
1290         }
1291
1292         for_each_member_device(ca, c, i) {
1293                 BUG_ON(ca->buckets_gc);
1294                 BUG_ON(ca->usage_gc);
1295
1296                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1297                 if (!ca->usage_gc) {
1298                         bch_err(c, "error allocating ca->usage_gc");
1299                         percpu_ref_put(&ca->ref);
1300                         return -ENOMEM;
1301                 }
1302
1303                 this_cpu_write(ca->usage_gc->d[BCH_DATA_free].buckets,
1304                                ca->mi.nbuckets - ca->mi.first_bucket);
1305         }
1306
1307         return 0;
1308 }
1309
1310 /* returns true if not equal */
1311 static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
1312                                      struct bch_alloc_v4 r)
1313 {
1314         return  l.gen != r.gen                          ||
1315                 l.oldest_gen != r.oldest_gen            ||
1316                 l.data_type != r.data_type              ||
1317                 l.dirty_sectors != r.dirty_sectors      ||
1318                 l.cached_sectors != r.cached_sectors     ||
1319                 l.stripe_redundancy != r.stripe_redundancy ||
1320                 l.stripe != r.stripe;
1321 }
1322
1323 static int bch2_alloc_write_key(struct btree_trans *trans,
1324                                 struct btree_iter *iter,
1325                                 bool metadata_only)
1326 {
1327         struct bch_fs *c = trans->c;
1328         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1329         struct bucket gc, *b;
1330         struct bkey_s_c k;
1331         struct bkey_i_alloc_v4 *a;
1332         struct bch_alloc_v4 old, new;
1333         enum bch_data_type type;
1334         int ret;
1335
1336         k = bch2_btree_iter_peek_slot(iter);
1337         ret = bkey_err(k);
1338         if (ret)
1339                 return ret;
1340
1341         bch2_alloc_to_v4(k, &old);
1342         new = old;
1343
1344         percpu_down_read(&c->mark_lock);
1345         b = gc_bucket(ca, iter->pos.offset);
1346
1347         /*
1348          * b->data_type doesn't yet include need_discard & need_gc_gen states -
1349          * fix that here:
1350          */
1351         type = __alloc_data_type(b->dirty_sectors,
1352                                  b->cached_sectors,
1353                                  b->stripe,
1354                                  old,
1355                                  b->data_type);
1356         if (b->data_type != type) {
1357                 struct bch_dev_usage *u;
1358
1359                 preempt_disable();
1360                 u = this_cpu_ptr(ca->usage_gc);
1361                 u->d[b->data_type].buckets--;
1362                 b->data_type = type;
1363                 u->d[b->data_type].buckets++;
1364                 preempt_enable();
1365         }
1366
1367         gc = *b;
1368         percpu_up_read(&c->mark_lock);
1369
1370         if (metadata_only &&
1371             gc.data_type != BCH_DATA_sb &&
1372             gc.data_type != BCH_DATA_journal &&
1373             gc.data_type != BCH_DATA_btree)
1374                 return 0;
1375
1376         if (gen_after(old.gen, gc.gen))
1377                 return 0;
1378
1379 #define copy_bucket_field(_f)                                           \
1380         if (fsck_err_on(new._f != gc._f, c,                             \
1381                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1382                         ": got %u, should be %u",                       \
1383                         iter->pos.inode, iter->pos.offset,              \
1384                         gc.gen,                                         \
1385                         bch2_data_types[gc.data_type],                  \
1386                         new._f, gc._f))                                 \
1387                 new._f = gc._f;                                         \
1388
1389         copy_bucket_field(gen);
1390         copy_bucket_field(data_type);
1391         copy_bucket_field(dirty_sectors);
1392         copy_bucket_field(cached_sectors);
1393         copy_bucket_field(stripe_redundancy);
1394         copy_bucket_field(stripe);
1395 #undef copy_bucket_field
1396
1397         if (!bch2_alloc_v4_cmp(old, new))
1398                 return 0;
1399
1400         a = bch2_alloc_to_v4_mut(trans, k);
1401         ret = PTR_ERR_OR_ZERO(a);
1402         if (ret)
1403                 return ret;
1404
1405         a->v = new;
1406
1407         /*
1408          * The trigger normally makes sure this is set, but we're not running
1409          * triggers:
1410          */
1411         if (a->v.data_type == BCH_DATA_cached && !a->v.io_time[READ])
1412                 a->v.io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
1413
1414         ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
1415 fsck_err:
1416         return ret;
1417 }
1418
1419 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1420 {
1421         struct btree_trans trans;
1422         struct btree_iter iter;
1423         struct bkey_s_c k;
1424         struct bch_dev *ca;
1425         unsigned i;
1426         int ret = 0;
1427
1428         bch2_trans_init(&trans, c, 0, 0);
1429
1430         for_each_member_device(ca, c, i) {
1431                 for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1432                                    POS(ca->dev_idx, ca->mi.first_bucket),
1433                                    BTREE_ITER_SLOTS|
1434                                    BTREE_ITER_PREFETCH, k, ret) {
1435                         if (bkey_cmp(iter.pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1436                                 break;
1437
1438                         ret = __bch2_trans_do(&trans, NULL, NULL,
1439                                               BTREE_INSERT_LAZY_RW,
1440                                         bch2_alloc_write_key(&trans, &iter,
1441                                                              metadata_only));
1442                         if (ret)
1443                                 break;
1444                 }
1445                 bch2_trans_iter_exit(&trans, &iter);
1446
1447                 if (ret) {
1448                         bch_err(c, "error writing alloc info: %i", ret);
1449                         percpu_ref_put(&ca->ref);
1450                         break;
1451                 }
1452         }
1453
1454         bch2_trans_exit(&trans);
1455         return ret;
1456 }
1457
1458 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1459 {
1460         struct bch_dev *ca;
1461         struct btree_trans trans;
1462         struct btree_iter iter;
1463         struct bkey_s_c k;
1464         struct bucket *g;
1465         struct bch_alloc_v4 a;
1466         unsigned i;
1467         int ret;
1468
1469         for_each_member_device(ca, c, i) {
1470                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1471                                 ca->mi.nbuckets * sizeof(struct bucket),
1472                                 GFP_KERNEL|__GFP_ZERO);
1473                 if (!buckets) {
1474                         percpu_ref_put(&ca->ref);
1475                         bch_err(c, "error allocating ca->buckets[gc]");
1476                         return -ENOMEM;
1477                 }
1478
1479                 buckets->first_bucket   = ca->mi.first_bucket;
1480                 buckets->nbuckets       = ca->mi.nbuckets;
1481                 rcu_assign_pointer(ca->buckets_gc, buckets);
1482         };
1483
1484         bch2_trans_init(&trans, c, 0, 0);
1485
1486         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1487                            BTREE_ITER_PREFETCH, k, ret) {
1488                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1489                 g = gc_bucket(ca, k.k->p.offset);
1490
1491                 bch2_alloc_to_v4(k, &a);
1492
1493                 g->gen_valid    = 1;
1494                 g->gen          = a.gen;
1495
1496                 if (metadata_only &&
1497                     (a.data_type == BCH_DATA_user ||
1498                      a.data_type == BCH_DATA_cached ||
1499                      a.data_type == BCH_DATA_parity)) {
1500                         g->data_type            = a.data_type;
1501                         g->dirty_sectors        = a.dirty_sectors;
1502                         g->cached_sectors       = a.cached_sectors;
1503                         g->stripe               = a.stripe;
1504                         g->stripe_redundancy    = a.stripe_redundancy;
1505                 }
1506         }
1507         bch2_trans_iter_exit(&trans, &iter);
1508
1509         bch2_trans_exit(&trans);
1510
1511         if (ret)
1512                 bch_err(c, "error reading alloc info at gc start: %i", ret);
1513
1514         return ret;
1515 }
1516
1517 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1518 {
1519         struct bch_dev *ca;
1520         unsigned i;
1521
1522         for_each_member_device(ca, c, i) {
1523                 struct bucket_array *buckets = gc_bucket_array(ca);
1524                 struct bucket *g;
1525
1526                 for_each_bucket(g, buckets) {
1527                         if (metadata_only &&
1528                             (g->data_type == BCH_DATA_user ||
1529                              g->data_type == BCH_DATA_cached ||
1530                              g->data_type == BCH_DATA_parity))
1531                                 continue;
1532                         g->data_type = 0;
1533                         g->dirty_sectors = 0;
1534                         g->cached_sectors = 0;
1535                 }
1536         };
1537 }
1538
1539 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1540 {
1541         struct btree_trans trans;
1542         struct btree_iter iter;
1543         struct bkey_s_c k;
1544         struct reflink_gc *r;
1545         size_t idx = 0;
1546         struct printbuf buf = PRINTBUF;
1547         int ret = 0;
1548
1549         if (metadata_only)
1550                 return 0;
1551
1552         bch2_trans_init(&trans, c, 0, 0);
1553
1554         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1555                            BTREE_ITER_PREFETCH, k, ret) {
1556                 const __le64 *refcount = bkey_refcount_c(k);
1557
1558                 if (!refcount)
1559                         continue;
1560
1561                 r = genradix_ptr(&c->reflink_gc_table, idx++);
1562                 if (!r ||
1563                     r->offset != k.k->p.offset ||
1564                     r->size != k.k->size) {
1565                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1566                         ret = -EINVAL;
1567                         break;
1568                 }
1569
1570                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1571                                 "reflink key has wrong refcount:\n"
1572                                 "  %s\n"
1573                                 "  should be %u",
1574                                 (printbuf_reset(&buf),
1575                                  bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1576                                 r->refcount)) {
1577                         struct bkey_i *new;
1578
1579                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1580                         if (!new) {
1581                                 ret = -ENOMEM;
1582                                 break;
1583                         }
1584
1585                         bkey_reassemble(new, k);
1586
1587                         if (!r->refcount)
1588                                 new->k.type = KEY_TYPE_deleted;
1589                         else
1590                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1591
1592                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1593                                 __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1594                         kfree(new);
1595
1596                         if (ret)
1597                                 break;
1598                 }
1599         }
1600 fsck_err:
1601         bch2_trans_iter_exit(&trans, &iter);
1602         c->reflink_gc_nr = 0;
1603         bch2_trans_exit(&trans);
1604         printbuf_exit(&buf);
1605         return ret;
1606 }
1607
1608 static int bch2_gc_reflink_start(struct bch_fs *c,
1609                                  bool metadata_only)
1610 {
1611         struct btree_trans trans;
1612         struct btree_iter iter;
1613         struct bkey_s_c k;
1614         struct reflink_gc *r;
1615         int ret = 0;
1616
1617         if (metadata_only)
1618                 return 0;
1619
1620         bch2_trans_init(&trans, c, 0, 0);
1621         c->reflink_gc_nr = 0;
1622
1623         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1624                            BTREE_ITER_PREFETCH, k, ret) {
1625                 const __le64 *refcount = bkey_refcount_c(k);
1626
1627                 if (!refcount)
1628                         continue;
1629
1630                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1631                                        GFP_KERNEL);
1632                 if (!r) {
1633                         ret = -ENOMEM;
1634                         break;
1635                 }
1636
1637                 r->offset       = k.k->p.offset;
1638                 r->size         = k.k->size;
1639                 r->refcount     = 0;
1640         }
1641         bch2_trans_iter_exit(&trans, &iter);
1642
1643         bch2_trans_exit(&trans);
1644         return ret;
1645 }
1646
1647 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1648 {
1649         struct genradix_iter iter;
1650         struct reflink_gc *r;
1651
1652         genradix_for_each(&c->reflink_gc_table, iter, r)
1653                 r->refcount = 0;
1654 }
1655
1656 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1657 {
1658         struct btree_trans trans;
1659         struct btree_iter iter;
1660         struct bkey_s_c k;
1661         struct gc_stripe *m;
1662         const struct bch_stripe *s;
1663         struct printbuf buf = PRINTBUF;
1664         unsigned i;
1665         int ret = 0;
1666
1667         if (metadata_only)
1668                 return 0;
1669
1670         bch2_trans_init(&trans, c, 0, 0);
1671
1672         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1673                            BTREE_ITER_PREFETCH, k, ret) {
1674                 if (k.k->type != KEY_TYPE_stripe)
1675                         continue;
1676
1677                 s = bkey_s_c_to_stripe(k).v;
1678                 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1679
1680                 for (i = 0; i < s->nr_blocks; i++)
1681                         if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1682                                 goto inconsistent;
1683                 continue;
1684 inconsistent:
1685                 if (fsck_err_on(true, c,
1686                                 "stripe has wrong block sector count %u:\n"
1687                                 "  %s\n"
1688                                 "  should be %u", i,
1689                                 (printbuf_reset(&buf),
1690                                  bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1691                                 m ? m->block_sectors[i] : 0)) {
1692                         struct bkey_i_stripe *new;
1693
1694                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1695                         if (!new) {
1696                                 ret = -ENOMEM;
1697                                 break;
1698                         }
1699
1700                         bkey_reassemble(&new->k_i, k);
1701
1702                         for (i = 0; i < new->v.nr_blocks; i++)
1703                                 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1704
1705                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1706                                 __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1707                         kfree(new);
1708                 }
1709         }
1710 fsck_err:
1711         bch2_trans_iter_exit(&trans, &iter);
1712
1713         bch2_trans_exit(&trans);
1714
1715         printbuf_exit(&buf);
1716         return ret;
1717 }
1718
1719 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1720 {
1721         genradix_free(&c->gc_stripes);
1722 }
1723
1724 /**
1725  * bch2_gc - walk _all_ references to buckets, and recompute them:
1726  *
1727  * Order matters here:
1728  *  - Concurrent GC relies on the fact that we have a total ordering for
1729  *    everything that GC walks - see  gc_will_visit_node(),
1730  *    gc_will_visit_root()
1731  *
1732  *  - also, references move around in the course of index updates and
1733  *    various other crap: everything needs to agree on the ordering
1734  *    references are allowed to move around in - e.g., we're allowed to
1735  *    start with a reference owned by an open_bucket (the allocator) and
1736  *    move it to the btree, but not the reverse.
1737  *
1738  *    This is necessary to ensure that gc doesn't miss references that
1739  *    move around - if references move backwards in the ordering GC
1740  *    uses, GC could skip past them
1741  */
1742 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1743 {
1744         unsigned iter = 0;
1745         int ret;
1746
1747         lockdep_assert_held(&c->state_lock);
1748
1749         down_write(&c->gc_lock);
1750
1751         bch2_btree_interior_updates_flush(c);
1752
1753         ret   = bch2_gc_start(c, metadata_only) ?:
1754                 bch2_gc_alloc_start(c, metadata_only) ?:
1755                 bch2_gc_reflink_start(c, metadata_only);
1756         if (ret)
1757                 goto out;
1758 again:
1759         gc_pos_set(c, gc_phase(GC_PHASE_START));
1760
1761         bch2_mark_superblocks(c);
1762
1763         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1764             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1765             c->opts.fix_errors != FSCK_OPT_NO) {
1766                 bch_info(c, "Starting topology repair pass");
1767                 ret = bch2_repair_topology(c);
1768                 if (ret)
1769                         goto out;
1770                 bch_info(c, "Topology repair pass done");
1771
1772                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1773         }
1774
1775         ret = bch2_gc_btrees(c, initial, metadata_only);
1776
1777         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1778             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1779             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1780                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1781                 SET_BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb, true);
1782                 ret = 0;
1783         }
1784
1785         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1786                 ret = FSCK_ERR_EXIT;
1787
1788         if (ret)
1789                 goto out;
1790
1791 #if 0
1792         bch2_mark_pending_btree_node_frees(c);
1793 #endif
1794         c->gc_count++;
1795
1796         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1797             (!iter && bch2_test_restart_gc)) {
1798                 if (iter++ > 2) {
1799                         bch_info(c, "Unable to fix bucket gens, looping");
1800                         ret = -EINVAL;
1801                         goto out;
1802                 }
1803
1804                 /*
1805                  * XXX: make sure gens we fixed got saved
1806                  */
1807                 bch_info(c, "Second GC pass needed, restarting:");
1808                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1809                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1810
1811                 bch2_gc_stripes_reset(c, metadata_only);
1812                 bch2_gc_alloc_reset(c, metadata_only);
1813                 bch2_gc_reflink_reset(c, metadata_only);
1814
1815                 /* flush fsck errors, reset counters */
1816                 bch2_flush_fsck_errs(c);
1817                 goto again;
1818         }
1819 out:
1820         if (!ret) {
1821                 bch2_journal_block(&c->journal);
1822
1823                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1824                         bch2_gc_reflink_done(c, metadata_only) ?:
1825                         bch2_gc_alloc_done(c, metadata_only) ?:
1826                         bch2_gc_done(c, initial, metadata_only);
1827
1828                 bch2_journal_unblock(&c->journal);
1829         }
1830
1831         percpu_down_write(&c->mark_lock);
1832         /* Indicates that gc is no longer in progress: */
1833         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1834
1835         bch2_gc_free(c);
1836         percpu_up_write(&c->mark_lock);
1837
1838         up_write(&c->gc_lock);
1839
1840         /*
1841          * At startup, allocations can happen directly instead of via the
1842          * allocator thread - issue wakeup in case they blocked on gc_lock:
1843          */
1844         closure_wake_up(&c->freelist_wait);
1845         return ret;
1846 }
1847
1848 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1849 {
1850         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1851         const struct bch_extent_ptr *ptr;
1852
1853         percpu_down_read(&c->mark_lock);
1854         bkey_for_each_ptr(ptrs, ptr) {
1855                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1856
1857                 if (ptr_stale(ca, ptr) > 16) {
1858                         percpu_up_read(&c->mark_lock);
1859                         return true;
1860                 }
1861         }
1862
1863         bkey_for_each_ptr(ptrs, ptr) {
1864                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1865                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1866
1867                 if (gen_after(*gen, ptr->gen))
1868                         *gen = ptr->gen;
1869         }
1870         percpu_up_read(&c->mark_lock);
1871
1872         return false;
1873 }
1874
1875 /*
1876  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1877  * node pointers currently never have cached pointers that can become stale:
1878  */
1879 static int bch2_gc_btree_gens(struct btree_trans *trans, enum btree_id btree_id)
1880 {
1881         struct bch_fs *c = trans->c;
1882         struct btree_iter iter;
1883         struct bkey_s_c k;
1884         struct bkey_buf sk;
1885         int ret = 0, commit_err = 0;
1886
1887         bch2_bkey_buf_init(&sk);
1888
1889         bch2_trans_iter_init(trans, &iter, btree_id, POS_MIN,
1890                              BTREE_ITER_PREFETCH|
1891                              BTREE_ITER_NOT_EXTENTS|
1892                              BTREE_ITER_ALL_SNAPSHOTS);
1893
1894         while ((bch2_trans_begin(trans),
1895                 k = bch2_btree_iter_peek(&iter)).k) {
1896                 ret = bkey_err(k);
1897
1898                 if (ret == -EINTR)
1899                         continue;
1900                 if (ret)
1901                         break;
1902
1903                 c->gc_gens_pos = iter.pos;
1904
1905                 if (gc_btree_gens_key(c, k) && !commit_err) {
1906                         bch2_bkey_buf_reassemble(&sk, c, k);
1907                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1908
1909                         commit_err =
1910                                 bch2_trans_update(trans, &iter, sk.k, 0) ?:
1911                                 bch2_trans_commit(trans, NULL, NULL,
1912                                                   BTREE_INSERT_NOWAIT|
1913                                                   BTREE_INSERT_NOFAIL);
1914                         if (commit_err == -EINTR) {
1915                                 commit_err = 0;
1916                                 continue;
1917                         }
1918                 }
1919
1920                 bch2_btree_iter_advance(&iter);
1921         }
1922         bch2_trans_iter_exit(trans, &iter);
1923
1924         bch2_bkey_buf_exit(&sk, c);
1925
1926         return ret;
1927 }
1928
1929 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter)
1930 {
1931         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1932         struct bkey_s_c k;
1933         struct bch_alloc_v4 a;
1934         struct bkey_i_alloc_v4 *a_mut;
1935         int ret;
1936
1937         k = bch2_btree_iter_peek_slot(iter);
1938         ret = bkey_err(k);
1939         if (ret)
1940                 return ret;
1941
1942         bch2_alloc_to_v4(k, &a);
1943
1944         if (a.oldest_gen == ca->oldest_gen[iter->pos.offset])
1945                 return 0;
1946
1947         a_mut = bch2_alloc_to_v4_mut(trans, k);
1948         ret = PTR_ERR_OR_ZERO(a_mut);
1949         if (ret)
1950                 return ret;
1951
1952         a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
1953         a_mut->v.data_type = alloc_data_type(a_mut->v, a_mut->v.data_type);
1954
1955         return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
1956 }
1957
1958 int bch2_gc_gens(struct bch_fs *c)
1959 {
1960         struct btree_trans trans;
1961         struct btree_iter iter;
1962         struct bkey_s_c k;
1963         struct bch_dev *ca;
1964         u64 b, start_time = local_clock();
1965         unsigned i;
1966         int ret;
1967
1968         /*
1969          * Ideally we would be using state_lock and not gc_lock here, but that
1970          * introduces a deadlock in the RO path - we currently take the state
1971          * lock at the start of going RO, thus the gc thread may get stuck:
1972          */
1973         if (!mutex_trylock(&c->gc_gens_lock))
1974                 return 0;
1975
1976         trace_gc_gens_start(c);
1977         down_read(&c->gc_lock);
1978         bch2_trans_init(&trans, c, 0, 0);
1979
1980         for_each_member_device(ca, c, i) {
1981                 struct bucket_gens *gens;
1982
1983                 BUG_ON(ca->oldest_gen);
1984
1985                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1986                 if (!ca->oldest_gen) {
1987                         percpu_ref_put(&ca->ref);
1988                         ret = -ENOMEM;
1989                         goto err;
1990                 }
1991
1992                 gens = bucket_gens(ca);
1993
1994                 for (b = gens->first_bucket;
1995                      b < gens->nbuckets; b++)
1996                         ca->oldest_gen[b] = gens->b[b];
1997         }
1998
1999         for (i = 0; i < BTREE_ID_NR; i++)
2000                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
2001                         c->gc_gens_btree = i;
2002                         c->gc_gens_pos = POS_MIN;
2003                         ret = bch2_gc_btree_gens(&trans, i);
2004                         if (ret) {
2005                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
2006                                 goto err;
2007                         }
2008                 }
2009
2010         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
2011                            BTREE_ITER_PREFETCH, k, ret) {
2012                 ret = __bch2_trans_do(&trans, NULL, NULL,
2013                                       BTREE_INSERT_NOFAIL,
2014                                 bch2_alloc_write_oldest_gen(&trans, &iter));
2015                 if (ret) {
2016                         bch_err(c, "error writing oldest_gen: %i", ret);
2017                         break;
2018                 }
2019         }
2020         bch2_trans_iter_exit(&trans, &iter);
2021
2022         c->gc_gens_btree        = 0;
2023         c->gc_gens_pos          = POS_MIN;
2024
2025         c->gc_count++;
2026
2027         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
2028         trace_gc_gens_end(c);
2029 err:
2030         for_each_member_device(ca, c, i) {
2031                 kvfree(ca->oldest_gen);
2032                 ca->oldest_gen = NULL;
2033         }
2034
2035         bch2_trans_exit(&trans);
2036         up_read(&c->gc_lock);
2037         mutex_unlock(&c->gc_gens_lock);
2038         return ret;
2039 }
2040
2041 static int bch2_gc_thread(void *arg)
2042 {
2043         struct bch_fs *c = arg;
2044         struct io_clock *clock = &c->io_clock[WRITE];
2045         unsigned long last = atomic64_read(&clock->now);
2046         unsigned last_kick = atomic_read(&c->kick_gc);
2047         int ret;
2048
2049         set_freezable();
2050
2051         while (1) {
2052                 while (1) {
2053                         set_current_state(TASK_INTERRUPTIBLE);
2054
2055                         if (kthread_should_stop()) {
2056                                 __set_current_state(TASK_RUNNING);
2057                                 return 0;
2058                         }
2059
2060                         if (atomic_read(&c->kick_gc) != last_kick)
2061                                 break;
2062
2063                         if (c->btree_gc_periodic) {
2064                                 unsigned long next = last + c->capacity / 16;
2065
2066                                 if (atomic64_read(&clock->now) >= next)
2067                                         break;
2068
2069                                 bch2_io_clock_schedule_timeout(clock, next);
2070                         } else {
2071                                 schedule();
2072                         }
2073
2074                         try_to_freeze();
2075                 }
2076                 __set_current_state(TASK_RUNNING);
2077
2078                 last = atomic64_read(&clock->now);
2079                 last_kick = atomic_read(&c->kick_gc);
2080
2081                 /*
2082                  * Full gc is currently incompatible with btree key cache:
2083                  */
2084 #if 0
2085                 ret = bch2_gc(c, false, false);
2086 #else
2087                 ret = bch2_gc_gens(c);
2088 #endif
2089                 if (ret < 0)
2090                         bch_err(c, "btree gc failed: %i", ret);
2091
2092                 debug_check_no_locks_held();
2093         }
2094
2095         return 0;
2096 }
2097
2098 void bch2_gc_thread_stop(struct bch_fs *c)
2099 {
2100         struct task_struct *p;
2101
2102         p = c->gc_thread;
2103         c->gc_thread = NULL;
2104
2105         if (p) {
2106                 kthread_stop(p);
2107                 put_task_struct(p);
2108         }
2109 }
2110
2111 int bch2_gc_thread_start(struct bch_fs *c)
2112 {
2113         struct task_struct *p;
2114
2115         if (c->gc_thread)
2116                 return 0;
2117
2118         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2119         if (IS_ERR(p)) {
2120                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
2121                 return PTR_ERR(p);
2122         }
2123
2124         get_task_struct(p);
2125         c->gc_thread = p;
2126         wake_up_process(p);
2127         return 0;
2128 }