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                                 pr_buf(&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
 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                 pr_buf(&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;
 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                                 "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                         if (ret)
 399                                 break;
 400                         continue;
 401                 }
 402
 403                 if (ret) {
 404                         bch_err(c, "%s: error %i getting btree node",
 405                                 __func__, ret);
 406                         break;
 407                 }
 408
 409                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 410
 411                 if (ret == DROP_THIS_NODE) {
 412                         six_unlock_read(&cur->c.lock);
 413                         bch2_btree_node_evict(c, cur_k.k);
 414                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 415                                                       b->c.level, cur_k.k->k.p);
 416                         if (ret)
 417                                 break;
 418                         continue;
 419                 }
 420
 421                 if (prev)
 422                         six_unlock_read(&prev->c.lock);
 423                 prev = NULL;
 424
 425                 if (ret == DROP_PREV_NODE) {
 426                         bch2_btree_node_evict(c, prev_k.k);
 427                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 428                                                       b->c.level, prev_k.k->k.p);
 429                         if (ret)
 430                                 break;
 431
 432                         bch2_btree_and_journal_iter_exit(&iter);
 433                         bch2_bkey_buf_exit(&prev_k, c);
 434                         bch2_bkey_buf_exit(&cur_k, c);
 435                         goto again;
 436                 } else if (ret)
 437                         break;
 438
 439                 prev = cur;
 440                 cur = NULL;
 441                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 442         }
 443
 444         if (!ret && !IS_ERR_OR_NULL(prev)) {
 445                 BUG_ON(cur);
 446                 ret = btree_repair_node_end(c, b, prev);
 447         }
 448
 449         if (!IS_ERR_OR_NULL(prev))
 450                 six_unlock_read(&prev->c.lock);
 451         prev = NULL;
 452         if (!IS_ERR_OR_NULL(cur))
 453                 six_unlock_read(&cur->c.lock);
 454         cur = NULL;
 455
 456         if (ret)
 457                 goto err;
 458
 459         bch2_btree_and_journal_iter_exit(&iter);
 460         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 461
 462         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 463                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 464                 bch2_btree_and_journal_iter_advance(&iter);
 465
 466                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 467                                         b->c.btree_id, b->c.level - 1,
 468                                         false);
 469                 ret = PTR_ERR_OR_ZERO(cur);
 470
 471                 if (ret) {
 472                         bch_err(c, "%s: error %i getting btree node",
 473                                 __func__, ret);
 474                         goto err;
 475                 }
 476
 477                 ret = bch2_btree_repair_topology_recurse(c, cur);
 478                 six_unlock_read(&cur->c.lock);
 479                 cur = NULL;
 480
 481                 if (ret == DROP_THIS_NODE) {
 482                         bch2_btree_node_evict(c, cur_k.k);
 483                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 484                                                       b->c.level, cur_k.k->k.p);
 485                         dropped_children = true;
 486                 }
 487
 488                 if (ret)
 489                         goto err;
 490
 491                 have_child = true;
 492         }
 493
 494         printbuf_reset(&buf);
 495         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 496
 497         if (mustfix_fsck_err_on(!have_child, c,
 498                         "empty interior btree node at btree %s level %u\n"
 499                         "  %s",
 500                         bch2_btree_ids[b->c.btree_id],
 501                         b->c.level, buf.buf))
 502                 ret = DROP_THIS_NODE;
 503 err:
 504 fsck_err:
 505         if (!IS_ERR_OR_NULL(prev))
 506                 six_unlock_read(&prev->c.lock);
 507         if (!IS_ERR_OR_NULL(cur))
 508                 six_unlock_read(&cur->c.lock);
 509
 510         bch2_btree_and_journal_iter_exit(&iter);
 511         bch2_bkey_buf_exit(&prev_k, c);
 512         bch2_bkey_buf_exit(&cur_k, c);
 513
 514         if (!ret && dropped_children)
 515                 goto again;
 516
 517         printbuf_exit(&buf);
 518         return ret;
 519 }
 520
 521 static int bch2_repair_topology(struct bch_fs *c)
 522 {
 523         struct btree *b;
 524         unsigned i;
 525         int ret = 0;
 526
 527         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 528                 b = c->btree_roots[i].b;
 529                 if (btree_node_fake(b))
 530                         continue;
 531
 532                 six_lock_read(&b->c.lock, NULL, NULL);
 533                 ret = bch2_btree_repair_topology_recurse(c, b);
 534                 six_unlock_read(&b->c.lock);
 535
 536                 if (ret == DROP_THIS_NODE) {
 537                         bch_err(c, "empty btree root - repair unimplemented");
 538                         ret = FSCK_ERR_EXIT;
 539                 }
 540         }
 541
 542         return ret;
 543 }
 544
 545 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 546                                unsigned level, bool is_root,
 547                                struct bkey_s_c *k)
 548 {
 549         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 550         const union bch_extent_entry *entry;
 551         struct extent_ptr_decoded p = { 0 };
 552         bool do_update = false;
 553         struct printbuf buf = PRINTBUF;
 554         int ret = 0;
 555
 556         /*
 557          * XXX
 558          * use check_bucket_ref here
 559          */
 560         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 561                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 562                 struct bucket *g = PTR_GC_BUCKET(ca, &p.ptr);
 563                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 564
 565                 if (fsck_err_on(!g->gen_valid, c,
 566                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 567                                 "while marking %s",
 568                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 569                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 570                                 p.ptr.gen,
 571                                 (printbuf_reset(&buf),
 572                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 573                         if (!p.ptr.cached) {
 574                                 g->gen_valid            = true;
 575                                 g->gen                  = p.ptr.gen;
 576                         } else {
 577                                 do_update = true;
 578                         }
 579                 }
 580
 581                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->gen) > 0, c,
 582                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 583                                 "while marking %s",
 584                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 585                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 586                                 p.ptr.gen, g->gen,
 587                                 (printbuf_reset(&buf),
 588                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 589                         if (!p.ptr.cached) {
 590                                 g->gen_valid            = true;
 591                                 g->gen                  = p.ptr.gen;
 592                                 g->data_type            = 0;
 593                                 g->dirty_sectors        = 0;
 594                                 g->cached_sectors       = 0;
 595                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 596                         } else {
 597                                 do_update = true;
 598                         }
 599                 }
 600
 601                 if (fsck_err_on(gen_cmp(g->gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
 602                                 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 603                                 "while marking %s",
 604                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->gen,
 605                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 606                                 p.ptr.gen,
 607                                 (printbuf_reset(&buf),
 608                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 609                         do_update = true;
 610
 611                 if (fsck_err_on(!p.ptr.cached &&
 612                                 gen_cmp(p.ptr.gen, g->gen) < 0, c,
 613                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 614                                 "while marking %s",
 615                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 616                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 617                                 p.ptr.gen, g->gen,
 618                                 (printbuf_reset(&buf),
 619                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 620                         do_update = true;
 621
 622                 if (data_type != BCH_DATA_btree && p.ptr.gen != g->gen)
 623                         continue;
 624
 625                 if (fsck_err_on(g->data_type &&
 626                                 g->data_type != data_type, c,
 627                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 628                                 "while marking %s",
 629                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 630                                 bch2_data_types[g->data_type],
 631                                 bch2_data_types[data_type],
 632                                 (printbuf_reset(&buf),
 633                                  bch2_bkey_val_to_text(&buf, c, *k), buf.buf))) {
 634                         if (data_type == BCH_DATA_btree) {
 635                                 g->data_type    = data_type;
 636                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 637                         } else {
 638                                 do_update = true;
 639                         }
 640                 }
 641
 642                 if (p.has_ec) {
 643                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes, p.ec.idx);
 644
 645                         if (fsck_err_on(!m || !m->alive, c,
 646                                         "pointer to nonexistent stripe %llu\n"
 647                                         "while marking %s",
 648                                         (u64) p.ec.idx,
 649                                         (printbuf_reset(&buf),
 650                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 651                                 do_update = true;
 652
 653                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 654                                         "pointer does not match stripe %llu\n"
 655                                         "while marking %s",
 656                                         (u64) p.ec.idx,
 657                                         (printbuf_reset(&buf),
 658                                          bch2_bkey_val_to_text(&buf, c, *k), buf.buf)))
 659                                 do_update = true;
 660                 }
 661         }
 662
 663         if (do_update) {
 664                 struct bkey_ptrs ptrs;
 665                 union bch_extent_entry *entry;
 666                 struct bch_extent_ptr *ptr;
 667                 struct bkey_i *new;
 668
 669                 if (is_root) {
 670                         bch_err(c, "cannot update btree roots yet");
 671                         ret = -EINVAL;
 672                         goto err;
 673                 }
 674
 675                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 676                 if (!new) {
 677                         bch_err(c, "%s: error allocating new key", __func__);
 678                         ret = -ENOMEM;
 679                         goto err;
 680                 }
 681
 682                 bkey_reassemble(new, *k);
 683
 684                 if (level) {
 685                         /*
 686                          * We don't want to drop btree node pointers - if the
 687                          * btree node isn't there anymore, the read path will
 688                          * sort it out:
 689                          */
 690                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 691                         bkey_for_each_ptr(ptrs, ptr) {
 692                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 693                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 694
 695                                 ptr->gen = g->gen;
 696                         }
 697                 } else {
 698                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 699                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 700                                 struct bucket *g = PTR_GC_BUCKET(ca, ptr);
 701                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 702
 703                                 (ptr->cached &&
 704                                  (!g->gen_valid || gen_cmp(ptr->gen, g->gen) > 0)) ||
 705                                 (!ptr->cached &&
 706                                  gen_cmp(ptr->gen, g->gen) < 0) ||
 707                                 gen_cmp(g->gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 708                                 (g->data_type &&
 709                                  g->data_type != data_type);
 710                         }));
 711 again:
 712                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 713                         bkey_extent_entry_for_each(ptrs, entry) {
 714                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 715                                         struct gc_stripe *m = genradix_ptr(&c->gc_stripes,
 716                                                                         entry->stripe_ptr.idx);
 717                                         union bch_extent_entry *next_ptr;
 718
 719                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 720                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 721                                                         goto found;
 722                                         next_ptr = NULL;
 723 found:
 724                                         if (!next_ptr) {
 725                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 726                                                 continue;
 727                                         }
 728
 729                                         if (!m || !m->alive ||
 730                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 731                                                                        &next_ptr->ptr,
 732                                                                        m->sectors)) {
 733                                                 bch2_bkey_extent_entry_drop(new, entry);
 734                                                 goto again;
 735                                         }
 736                                 }
 737                         }
 738                 }
 739
 740                 ret = bch2_journal_key_insert_take(c, btree_id, level, new);
 741                 if (ret) {
 742                         kfree(new);
 743                         goto err;
 744                 }
 745
 746                 if (level)
 747                         bch2_btree_node_update_key_early(c, btree_id, level - 1, *k, new);
 748
 749                 printbuf_reset(&buf);
 750                 bch2_bkey_val_to_text(&buf, c, *k);
 751                 bch_info(c, "updated %s", buf.buf);
 752
 753                 printbuf_reset(&buf);
 754                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(new));
 755                 bch_info(c, "new key %s", buf.buf);
 756
 757                 *k = bkey_i_to_s_c(new);
 758         }
 759 err:
 760 fsck_err:
 761         printbuf_exit(&buf);
 762         return ret;
 763 }
 764
 765 /* marking of btree keys/nodes: */
 766
 767 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 768                             unsigned level, bool is_root,
 769                             struct bkey_s_c *k,
 770                             bool initial)
 771 {
 772         struct bch_fs *c = trans->c;
 773         struct bkey deleted = KEY(0, 0, 0);
 774         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 775         unsigned flags =
 776                 BTREE_TRIGGER_GC|
 777                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 778         int ret = 0;
 779
 780         deleted.p = k->k->p;
 781
 782         if (initial) {
 783                 BUG_ON(bch2_journal_seq_verify &&
 784                        k->k->version.lo > atomic64_read(&c->journal.seq));
 785
 786                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 787                 if (ret)
 788                         goto err;
 789
 790                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 791                                 "key version number higher than recorded: %llu > %llu",
 792                                 k->k->version.lo,
 793                                 atomic64_read(&c->key_version)))
 794                         atomic64_set(&c->key_version, k->k->version.lo);
 795         }
 796
 797         ret = __bch2_trans_do(trans, NULL, NULL, 0,
 798                         bch2_mark_key(trans, old, *k, flags));
 799 fsck_err:
 800 err:
 801         if (ret)
 802                 bch_err(c, "%s: ret %i", __func__, ret);
 803         return ret;
 804 }
 805
 806 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, bool initial)
 807 {
 808         struct bch_fs *c = trans->c;
 809         struct btree_node_iter iter;
 810         struct bkey unpacked;
 811         struct bkey_s_c k;
 812         struct bkey_buf prev, cur;
 813         int ret = 0;
 814
 815         if (!btree_node_type_needs_gc(btree_node_type(b)))
 816                 return 0;
 817
 818         bch2_btree_node_iter_init_from_start(&iter, b);
 819         bch2_bkey_buf_init(&prev);
 820         bch2_bkey_buf_init(&cur);
 821         bkey_init(&prev.k->k);
 822
 823         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 824                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 825                                        &k, initial);
 826                 if (ret)
 827                         break;
 828
 829                 bch2_btree_node_iter_advance(&iter, b);
 830
 831                 if (b->c.level) {
 832                         bch2_bkey_buf_reassemble(&cur, c, k);
 833
 834                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 835                                         bch2_btree_node_iter_end(&iter));
 836                         if (ret)
 837                                 break;
 838                 }
 839         }
 840
 841         bch2_bkey_buf_exit(&cur, c);
 842         bch2_bkey_buf_exit(&prev, c);
 843         return ret;
 844 }
 845
 846 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 847                          bool initial, bool metadata_only)
 848 {
 849         struct bch_fs *c = trans->c;
 850         struct btree_iter iter;
 851         struct btree *b;
 852         unsigned depth = metadata_only                  ? 1
 853                 : bch2_expensive_debug_checks           ? 0
 854                 : !btree_node_type_needs_gc(btree_id)   ? 1
 855                 : 0;
 856         int ret = 0;
 857
 858         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 859
 860         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 861                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 862                 bch2_verify_btree_nr_keys(b);
 863
 864                 gc_pos_set(c, gc_pos_btree_node(b));
 865
 866                 ret = btree_gc_mark_node(trans, b, initial);
 867                 if (ret)
 868                         break;
 869         }
 870         bch2_trans_iter_exit(trans, &iter);
 871
 872         if (ret)
 873                 return ret;
 874
 875         mutex_lock(&c->btree_root_lock);
 876         b = c->btree_roots[btree_id].b;
 877         if (!btree_node_fake(b)) {
 878                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 879
 880                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 881                                        true, &k, initial);
 882         }
 883         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 884         mutex_unlock(&c->btree_root_lock);
 885
 886         return ret;
 887 }
 888
 889 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 890                                       unsigned target_depth)
 891 {
 892         struct bch_fs *c = trans->c;
 893         struct btree_and_journal_iter iter;
 894         struct bkey_s_c k;
 895         struct bkey_buf cur, prev;
 896         struct printbuf buf = PRINTBUF;
 897         int ret = 0;
 898
 899         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 900         bch2_bkey_buf_init(&prev);
 901         bch2_bkey_buf_init(&cur);
 902         bkey_init(&prev.k->k);
 903
 904         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 905                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 906                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 907
 908                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level,
 909                                        false, &k, true);
 910                 if (ret) {
 911                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 912                         goto fsck_err;
 913                 }
 914
 915                 if (b->c.level) {
 916                         bch2_bkey_buf_reassemble(&cur, c, k);
 917                         k = bkey_i_to_s_c(cur.k);
 918
 919                         bch2_btree_and_journal_iter_advance(&iter);
 920
 921                         ret = bch2_gc_check_topology(c, b,
 922                                         &prev, cur,
 923                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 924                         if (ret)
 925                                 goto fsck_err;
 926                 } else {
 927                         bch2_btree_and_journal_iter_advance(&iter);
 928                 }
 929         }
 930
 931         if (b->c.level > target_depth) {
 932                 bch2_btree_and_journal_iter_exit(&iter);
 933                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 934
 935                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 936                         struct btree *child;
 937
 938                         bch2_bkey_buf_reassemble(&cur, c, k);
 939                         bch2_btree_and_journal_iter_advance(&iter);
 940
 941                         child = bch2_btree_node_get_noiter(c, cur.k,
 942                                                 b->c.btree_id, b->c.level - 1,
 943                                                 false);
 944                         ret = PTR_ERR_OR_ZERO(child);
 945
 946                         if (ret == -EIO) {
 947                                 bch2_topology_error(c);
 948
 949                                 if (__fsck_err(c,
 950                                           FSCK_CAN_FIX|
 951                                           FSCK_CAN_IGNORE|
 952                                           FSCK_NO_RATELIMIT,
 953                                           "Unreadable btree node at btree %s level %u:\n"
 954                                           "  %s",
 955                                           bch2_btree_ids[b->c.btree_id],
 956                                           b->c.level - 1,
 957                                           (printbuf_reset(&buf),
 958                                            bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur.k)), buf.buf)) &&
 959                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 960                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 961                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 962                                         goto fsck_err;
 963                                 } else {
 964                                         /* Continue marking when opted to not
 965                                          * fix the error: */
 966                                         ret = 0;
 967                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 968                                         continue;
 969                                 }
 970                         } else if (ret) {
 971                                 bch_err(c, "%s: error %i getting btree node",
 972                                         __func__, ret);
 973                                 break;
 974                         }
 975
 976                         ret = bch2_gc_btree_init_recurse(trans, child,
 977                                                          target_depth);
 978                         six_unlock_read(&child->c.lock);
 979
 980                         if (ret)
 981                                 break;
 982                 }
 983         }
 984 fsck_err:
 985         bch2_bkey_buf_exit(&cur, c);
 986         bch2_bkey_buf_exit(&prev, c);
 987         bch2_btree_and_journal_iter_exit(&iter);
 988         printbuf_exit(&buf);
 989         return ret;
 990 }
 991
 992 static int bch2_gc_btree_init(struct btree_trans *trans,
 993                               enum btree_id btree_id,
 994                               bool metadata_only)
 995 {
 996         struct bch_fs *c = trans->c;
 997         struct btree *b;
 998         unsigned target_depth = metadata_only           ? 1
 999                 : bch2_expensive_debug_checks           ? 0
1000                 : !btree_node_type_needs_gc(btree_id)   ? 1
1001                 : 0;
1002         struct printbuf buf = PRINTBUF;
1003         int ret = 0;
1004
1005         b = c->btree_roots[btree_id].b;
1006
1007         if (btree_node_fake(b))
1008                 return 0;
1009
1010         six_lock_read(&b->c.lock, NULL, NULL);
1011         printbuf_reset(&buf);
1012         bch2_bpos_to_text(&buf, b->data->min_key);
1013         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
1014                         "btree root with incorrect min_key: %s", buf.buf)) {
1015                 bch_err(c, "repair unimplemented");
1016                 ret = FSCK_ERR_EXIT;
1017                 goto fsck_err;
1018         }
1019
1020         printbuf_reset(&buf);
1021         bch2_bpos_to_text(&buf, b->data->max_key);
1022         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
1023                         "btree root with incorrect max_key: %s", buf.buf)) {
1024                 bch_err(c, "repair unimplemented");
1025                 ret = FSCK_ERR_EXIT;
1026                 goto fsck_err;
1027         }
1028
1029         if (b->c.level >= target_depth)
1030                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
1031
1032         if (!ret) {
1033                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1034
1035                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1036                                        &k, true);
1037         }
1038 fsck_err:
1039         six_unlock_read(&b->c.lock);
1040
1041         if (ret < 0)
1042                 bch_err(c, "%s: ret %i", __func__, ret);
1043         printbuf_exit(&buf);
1044         return ret;
1045 }
1046
1047 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1048 {
1049         return  (int) btree_id_to_gc_phase(l) -
1050                 (int) btree_id_to_gc_phase(r);
1051 }
1052
1053 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1054 {
1055         struct btree_trans trans;
1056         enum btree_id ids[BTREE_ID_NR];
1057         unsigned i;
1058         int ret = 0;
1059
1060         bch2_trans_init(&trans, c, 0, 0);
1061
1062         if (initial)
1063                 trans.is_initial_gc = true;
1064
1065         for (i = 0; i < BTREE_ID_NR; i++)
1066                 ids[i] = i;
1067         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1068
1069         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1070                 ret = initial
1071                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1072                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1073
1074         if (ret < 0)
1075                 bch_err(c, "%s: ret %i", __func__, ret);
1076
1077         bch2_trans_exit(&trans);
1078         return ret;
1079 }
1080
1081 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1082                                   u64 start, u64 end,
1083                                   enum bch_data_type type,
1084                                   unsigned flags)
1085 {
1086         u64 b = sector_to_bucket(ca, start);
1087
1088         do {
1089                 unsigned sectors =
1090                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1091
1092                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1093                                           gc_phase(GC_PHASE_SB), flags);
1094                 b++;
1095                 start += sectors;
1096         } while (start < end);
1097 }
1098
1099 static void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1100                                      unsigned flags)
1101 {
1102         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1103         unsigned i;
1104         u64 b;
1105
1106         for (i = 0; i < layout->nr_superblocks; i++) {
1107                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1108
1109                 if (offset == BCH_SB_SECTOR)
1110                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1111                                               BCH_DATA_sb, flags);
1112
1113                 mark_metadata_sectors(c, ca, offset,
1114                                       offset + (1 << layout->sb_max_size_bits),
1115                                       BCH_DATA_sb, flags);
1116         }
1117
1118         for (i = 0; i < ca->journal.nr; i++) {
1119                 b = ca->journal.buckets[i];
1120                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1121                                           ca->mi.bucket_size,
1122                                           gc_phase(GC_PHASE_SB), flags);
1123         }
1124 }
1125
1126 static void bch2_mark_superblocks(struct bch_fs *c)
1127 {
1128         struct bch_dev *ca;
1129         unsigned i;
1130
1131         mutex_lock(&c->sb_lock);
1132         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1133
1134         for_each_online_member(ca, c, i)
1135                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1136         mutex_unlock(&c->sb_lock);
1137 }
1138
1139 #if 0
1140 /* Also see bch2_pending_btree_node_free_insert_done() */
1141 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1142 {
1143         struct btree_update *as;
1144         struct pending_btree_node_free *d;
1145
1146         mutex_lock(&c->btree_interior_update_lock);
1147         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1148
1149         for_each_pending_btree_node_free(c, as, d)
1150                 if (d->index_update_done)
1151                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1152
1153         mutex_unlock(&c->btree_interior_update_lock);
1154 }
1155 #endif
1156
1157 static void bch2_gc_free(struct bch_fs *c)
1158 {
1159         struct bch_dev *ca;
1160         unsigned i;
1161
1162         genradix_free(&c->reflink_gc_table);
1163         genradix_free(&c->gc_stripes);
1164
1165         for_each_member_device(ca, c, i) {
1166                 kvpfree(rcu_dereference_protected(ca->buckets_gc, 1),
1167                         sizeof(struct bucket_array) +
1168                         ca->mi.nbuckets * sizeof(struct bucket));
1169                 ca->buckets_gc = NULL;
1170
1171                 free_percpu(ca->usage_gc);
1172                 ca->usage_gc = NULL;
1173         }
1174
1175         free_percpu(c->usage_gc);
1176         c->usage_gc = NULL;
1177 }
1178
1179 static int bch2_gc_done(struct bch_fs *c,
1180                         bool initial, bool metadata_only)
1181 {
1182         struct bch_dev *ca = NULL;
1183         struct printbuf buf = PRINTBUF;
1184         bool verify = !metadata_only && (!initial ||
1185                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1186         unsigned i, dev;
1187         int ret = 0;
1188
1189         percpu_down_write(&c->mark_lock);
1190
1191 #define copy_field(_f, _msg, ...)                                       \
1192         if (dst->_f != src->_f) {                                       \
1193                 if (verify)                                             \
1194                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1195                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1196                 dst->_f = src->_f;                                      \
1197         }
1198 #define copy_stripe_field(_f, _msg, ...)                                \
1199         if (dst->_f != src->_f) {                                       \
1200                 if (verify)                                             \
1201                         fsck_err(c, "stripe %zu has wrong "_msg         \
1202                                 ": got %u, should be %u",               \
1203                                 iter.pos, ##__VA_ARGS__,                \
1204                                 dst->_f, src->_f);                      \
1205                 dst->_f = src->_f;                                      \
1206         }
1207 #define copy_dev_field(_f, _msg, ...)                                   \
1208         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1209 #define copy_fs_field(_f, _msg, ...)                                    \
1210         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1211
1212         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1213                 bch2_fs_usage_acc_to_base(c, i);
1214
1215         for_each_member_device(ca, c, dev) {
1216                 struct bch_dev_usage *dst = ca->usage_base;
1217                 struct bch_dev_usage *src = (void *)
1218                         bch2_acc_percpu_u64s((void *) ca->usage_gc,
1219                                              dev_usage_u64s());
1220
1221                 copy_dev_field(buckets_ec,              "buckets_ec");
1222                 copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1223
1224                 for (i = 0; i < BCH_DATA_NR; i++) {
1225                         copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1226                         copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1227                         copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1228                 }
1229         };
1230
1231         {
1232                 unsigned nr = fs_usage_u64s(c);
1233                 struct bch_fs_usage *dst = c->usage_base;
1234                 struct bch_fs_usage *src = (void *)
1235                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1236
1237                 copy_fs_field(hidden,           "hidden");
1238                 copy_fs_field(btree,            "btree");
1239
1240                 if (!metadata_only) {
1241                         copy_fs_field(data,     "data");
1242                         copy_fs_field(cached,   "cached");
1243                         copy_fs_field(reserved, "reserved");
1244                         copy_fs_field(nr_inodes,"nr_inodes");
1245
1246                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1247                                 copy_fs_field(persistent_reserved[i],
1248                                               "persistent_reserved[%i]", i);
1249                 }
1250
1251                 for (i = 0; i < c->replicas.nr; i++) {
1252                         struct bch_replicas_entry *e =
1253                                 cpu_replicas_entry(&c->replicas, i);
1254
1255                         if (metadata_only &&
1256                             (e->data_type == BCH_DATA_user ||
1257                              e->data_type == BCH_DATA_cached))
1258                                 continue;
1259
1260                         printbuf_reset(&buf);
1261                         bch2_replicas_entry_to_text(&buf, e);
1262
1263                         copy_fs_field(replicas[i], "%s", buf.buf);
1264                 }
1265         }
1266
1267 #undef copy_fs_field
1268 #undef copy_dev_field
1269 #undef copy_stripe_field
1270 #undef copy_field
1271 fsck_err:
1272         if (ca)
1273                 percpu_ref_put(&ca->ref);
1274         if (ret)
1275                 bch_err(c, "%s: ret %i", __func__, ret);
1276
1277         percpu_up_write(&c->mark_lock);
1278         printbuf_exit(&buf);
1279         return ret;
1280 }
1281
1282 static int bch2_gc_start(struct bch_fs *c,
1283                          bool metadata_only)
1284 {
1285         struct bch_dev *ca = NULL;
1286         unsigned i;
1287
1288         BUG_ON(c->usage_gc);
1289
1290         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1291                                          sizeof(u64), GFP_KERNEL);
1292         if (!c->usage_gc) {
1293                 bch_err(c, "error allocating c->usage_gc");
1294                 return -ENOMEM;
1295         }
1296
1297         for_each_member_device(ca, c, i) {
1298                 BUG_ON(ca->buckets_gc);
1299                 BUG_ON(ca->usage_gc);
1300
1301                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1302                 if (!ca->usage_gc) {
1303                         bch_err(c, "error allocating ca->usage_gc");
1304                         percpu_ref_put(&ca->ref);
1305                         return -ENOMEM;
1306                 }
1307         }
1308
1309         return 0;
1310 }
1311
1312 static int bch2_alloc_write_key(struct btree_trans *trans,
1313                                 struct btree_iter *iter,
1314                                 bool metadata_only)
1315 {
1316         struct bch_fs *c = trans->c;
1317         struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
1318         struct bucket gc;
1319         struct bkey_s_c k;
1320         struct bkey_alloc_unpacked old_u, new_u;
1321         struct bkey_alloc_buf *a;
1322         int ret;
1323
1324         k = bch2_btree_iter_peek_slot(iter);
1325         ret = bkey_err(k);
1326         if (ret)
1327                 return ret;
1328
1329         old_u = new_u = bch2_alloc_unpack(k);
1330
1331         percpu_down_read(&c->mark_lock);
1332         gc = *gc_bucket(ca, iter->pos.offset);
1333         percpu_up_read(&c->mark_lock);
1334
1335         if (metadata_only &&
1336             gc.data_type != BCH_DATA_sb &&
1337             gc.data_type != BCH_DATA_journal &&
1338             gc.data_type != BCH_DATA_btree)
1339                 return 0;
1340
1341         if (gen_after(old_u.gen, gc.gen))
1342                 return 0;
1343
1344 #define copy_bucket_field(_f)                                           \
1345         if (fsck_err_on(new_u._f != gc._f, c,                           \
1346                         "bucket %llu:%llu gen %u data type %s has wrong " #_f   \
1347                         ": got %u, should be %u",                       \
1348                         iter->pos.inode, iter->pos.offset,              \
1349                         gc.gen,                                         \
1350                         bch2_data_types[gc.data_type],                  \
1351                         new_u._f, gc._f))                               \
1352                 new_u._f = gc._f;                                       \
1353
1354         copy_bucket_field(gen);
1355         copy_bucket_field(data_type);
1356         copy_bucket_field(stripe);
1357         copy_bucket_field(dirty_sectors);
1358         copy_bucket_field(cached_sectors);
1359         copy_bucket_field(stripe_redundancy);
1360         copy_bucket_field(stripe);
1361 #undef copy_bucket_field
1362
1363         if (!bkey_alloc_unpacked_cmp(old_u, new_u))
1364                 return 0;
1365
1366         a = bch2_alloc_pack(trans, new_u);
1367         if (IS_ERR(a))
1368                 return PTR_ERR(a);
1369
1370         ret = bch2_trans_update(trans, iter, &a->k, BTREE_TRIGGER_NORUN);
1371 fsck_err:
1372         return ret;
1373 }
1374
1375 static int bch2_gc_alloc_done(struct bch_fs *c, bool metadata_only)
1376 {
1377         struct btree_trans trans;
1378         struct btree_iter iter;
1379         struct bkey_s_c k;
1380         struct bch_dev *ca;
1381         unsigned i;
1382         int ret = 0;
1383
1384         bch2_trans_init(&trans, c, 0, 0);
1385
1386         for_each_member_device(ca, c, i) {
1387                 for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1388                                    POS(ca->dev_idx, ca->mi.first_bucket),
1389                                    BTREE_ITER_SLOTS|
1390                                    BTREE_ITER_PREFETCH, k, ret) {
1391                         if (bkey_cmp(iter.pos, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
1392                                 break;
1393
1394                         ret = __bch2_trans_do(&trans, NULL, NULL,
1395                                               BTREE_INSERT_LAZY_RW,
1396                                         bch2_alloc_write_key(&trans, &iter,
1397                                                              metadata_only));
1398                         if (ret)
1399                                 break;
1400                 }
1401                 bch2_trans_iter_exit(&trans, &iter);
1402
1403                 if (ret) {
1404                         bch_err(c, "error writing alloc info: %i", ret);
1405                         percpu_ref_put(&ca->ref);
1406                         break;
1407                 }
1408         }
1409
1410         bch2_trans_exit(&trans);
1411         return ret;
1412 }
1413
1414 static int bch2_gc_alloc_start(struct bch_fs *c, bool metadata_only)
1415 {
1416         struct bch_dev *ca;
1417         struct btree_trans trans;
1418         struct btree_iter iter;
1419         struct bkey_s_c k;
1420         struct bucket *g;
1421         struct bkey_alloc_unpacked u;
1422         unsigned i;
1423         int ret;
1424
1425         for_each_member_device(ca, c, i) {
1426                 struct bucket_array *buckets = kvpmalloc(sizeof(struct bucket_array) +
1427                                 ca->mi.nbuckets * sizeof(struct bucket),
1428                                 GFP_KERNEL|__GFP_ZERO);
1429                 if (!buckets) {
1430                         percpu_ref_put(&ca->ref);
1431                         bch_err(c, "error allocating ca->buckets[gc]");
1432                         return -ENOMEM;
1433                 }
1434
1435                 buckets->first_bucket   = ca->mi.first_bucket;
1436                 buckets->nbuckets       = ca->mi.nbuckets;
1437                 rcu_assign_pointer(ca->buckets_gc, buckets);
1438         };
1439
1440         bch2_trans_init(&trans, c, 0, 0);
1441
1442         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1443                            BTREE_ITER_PREFETCH, k, ret) {
1444                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
1445                 g = gc_bucket(ca, k.k->p.offset);
1446                 u = bch2_alloc_unpack(k);
1447
1448                 g->gen_valid    = 1;
1449                 g->gen          = u.gen;
1450
1451                 if (metadata_only &&
1452                     (u.data_type == BCH_DATA_user ||
1453                      u.data_type == BCH_DATA_cached ||
1454                      u.data_type == BCH_DATA_parity)) {
1455                         g->data_type            = u.data_type;
1456                         g->dirty_sectors        = u.dirty_sectors;
1457                         g->cached_sectors       = u.cached_sectors;
1458                         g->stripe               = u.stripe;
1459                         g->stripe_redundancy    = u.stripe_redundancy;
1460                 }
1461         }
1462         bch2_trans_iter_exit(&trans, &iter);
1463
1464         bch2_trans_exit(&trans);
1465
1466         if (ret)
1467                 bch_err(c, "error reading alloc info at gc start: %i", ret);
1468
1469         return ret;
1470 }
1471
1472 static void bch2_gc_alloc_reset(struct bch_fs *c, bool metadata_only)
1473 {
1474         struct bch_dev *ca;
1475         unsigned i;
1476
1477         for_each_member_device(ca, c, i) {
1478                 struct bucket_array *buckets = gc_bucket_array(ca);
1479                 struct bucket *g;
1480
1481                 for_each_bucket(g, buckets) {
1482                         if (metadata_only &&
1483                             (g->data_type == BCH_DATA_user ||
1484                              g->data_type == BCH_DATA_cached ||
1485                              g->data_type == BCH_DATA_parity))
1486                                 continue;
1487                         g->dirty_sectors = 0;
1488                         g->cached_sectors = 0;
1489                 }
1490         };
1491 }
1492
1493 static int bch2_gc_reflink_done(struct bch_fs *c, bool metadata_only)
1494 {
1495         struct btree_trans trans;
1496         struct btree_iter iter;
1497         struct bkey_s_c k;
1498         struct reflink_gc *r;
1499         size_t idx = 0;
1500         struct printbuf buf = PRINTBUF;
1501         int ret = 0;
1502
1503         if (metadata_only)
1504                 return 0;
1505
1506         bch2_trans_init(&trans, c, 0, 0);
1507
1508         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1509                            BTREE_ITER_PREFETCH, k, ret) {
1510                 const __le64 *refcount = bkey_refcount_c(k);
1511
1512                 if (!refcount)
1513                         continue;
1514
1515                 r = genradix_ptr(&c->reflink_gc_table, idx++);
1516                 if (!r ||
1517                     r->offset != k.k->p.offset ||
1518                     r->size != k.k->size) {
1519                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1520                         ret = -EINVAL;
1521                         break;
1522                 }
1523
1524                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1525                                 "reflink key has wrong refcount:\n"
1526                                 "  %s\n"
1527                                 "  should be %u",
1528                                 (printbuf_reset(&buf),
1529                                  bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1530                                 r->refcount)) {
1531                         struct bkey_i *new;
1532
1533                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1534                         if (!new) {
1535                                 ret = -ENOMEM;
1536                                 break;
1537                         }
1538
1539                         bkey_reassemble(new, k);
1540
1541                         if (!r->refcount)
1542                                 new->k.type = KEY_TYPE_deleted;
1543                         else
1544                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1545
1546                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1547                                 __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1548                         kfree(new);
1549
1550                         if (ret)
1551                                 break;
1552                 }
1553         }
1554 fsck_err:
1555         bch2_trans_iter_exit(&trans, &iter);
1556         c->reflink_gc_nr = 0;
1557         bch2_trans_exit(&trans);
1558         printbuf_exit(&buf);
1559         return ret;
1560 }
1561
1562 static int bch2_gc_reflink_start(struct bch_fs *c,
1563                                  bool metadata_only)
1564 {
1565         struct btree_trans trans;
1566         struct btree_iter iter;
1567         struct bkey_s_c k;
1568         struct reflink_gc *r;
1569         int ret = 0;
1570
1571         if (metadata_only)
1572                 return 0;
1573
1574         bch2_trans_init(&trans, c, 0, 0);
1575         c->reflink_gc_nr = 0;
1576
1577         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1578                            BTREE_ITER_PREFETCH, k, ret) {
1579                 const __le64 *refcount = bkey_refcount_c(k);
1580
1581                 if (!refcount)
1582                         continue;
1583
1584                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1585                                        GFP_KERNEL);
1586                 if (!r) {
1587                         ret = -ENOMEM;
1588                         break;
1589                 }
1590
1591                 r->offset       = k.k->p.offset;
1592                 r->size         = k.k->size;
1593                 r->refcount     = 0;
1594         }
1595         bch2_trans_iter_exit(&trans, &iter);
1596
1597         bch2_trans_exit(&trans);
1598         return ret;
1599 }
1600
1601 static void bch2_gc_reflink_reset(struct bch_fs *c, bool metadata_only)
1602 {
1603         struct genradix_iter iter;
1604         struct reflink_gc *r;
1605
1606         genradix_for_each(&c->reflink_gc_table, iter, r)
1607                 r->refcount = 0;
1608 }
1609
1610 static int bch2_gc_stripes_done(struct bch_fs *c, bool metadata_only)
1611 {
1612         struct btree_trans trans;
1613         struct btree_iter iter;
1614         struct bkey_s_c k;
1615         struct gc_stripe *m;
1616         const struct bch_stripe *s;
1617         struct printbuf buf = PRINTBUF;
1618         unsigned i;
1619         int ret = 0;
1620
1621         if (metadata_only)
1622                 return 0;
1623
1624         bch2_trans_init(&trans, c, 0, 0);
1625
1626         for_each_btree_key(&trans, iter, BTREE_ID_stripes, POS_MIN,
1627                            BTREE_ITER_PREFETCH, k, ret) {
1628                 if (k.k->type != KEY_TYPE_stripe)
1629                         continue;
1630
1631                 s = bkey_s_c_to_stripe(k).v;
1632                 m = genradix_ptr(&c->gc_stripes, k.k->p.offset);
1633
1634                 for (i = 0; i < s->nr_blocks; i++)
1635                         if (stripe_blockcount_get(s, i) != (m ? m->block_sectors[i] : 0))
1636                                 goto inconsistent;
1637                 continue;
1638 inconsistent:
1639                 if (fsck_err_on(true, c,
1640                                 "stripe has wrong block sector count %u:\n"
1641                                 "  %s\n"
1642                                 "  should be %u", i,
1643                                 (printbuf_reset(&buf),
1644                                  bch2_bkey_val_to_text(&buf, c, k), buf.buf),
1645                                 m ? m->block_sectors[i] : 0)) {
1646                         struct bkey_i_stripe *new;
1647
1648                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1649                         if (!new) {
1650                                 ret = -ENOMEM;
1651                                 break;
1652                         }
1653
1654                         bkey_reassemble(&new->k_i, k);
1655
1656                         for (i = 0; i < new->v.nr_blocks; i++)
1657                                 stripe_blockcount_set(&new->v, i, m ? m->block_sectors[i] : 0);
1658
1659                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1660                                 __bch2_btree_insert(&trans, BTREE_ID_reflink, &new->k_i));
1661                         kfree(new);
1662                 }
1663         }
1664 fsck_err:
1665         bch2_trans_iter_exit(&trans, &iter);
1666
1667         bch2_trans_exit(&trans);
1668
1669         printbuf_exit(&buf);
1670         return ret;
1671 }
1672
1673 static void bch2_gc_stripes_reset(struct bch_fs *c, bool metadata_only)
1674 {
1675         genradix_free(&c->gc_stripes);
1676 }
1677
1678 /**
1679  * bch2_gc - walk _all_ references to buckets, and recompute them:
1680  *
1681  * Order matters here:
1682  *  - Concurrent GC relies on the fact that we have a total ordering for
1683  *    everything that GC walks - see  gc_will_visit_node(),
1684  *    gc_will_visit_root()
1685  *
1686  *  - also, references move around in the course of index updates and
1687  *    various other crap: everything needs to agree on the ordering
1688  *    references are allowed to move around in - e.g., we're allowed to
1689  *    start with a reference owned by an open_bucket (the allocator) and
1690  *    move it to the btree, but not the reverse.
1691  *
1692  *    This is necessary to ensure that gc doesn't miss references that
1693  *    move around - if references move backwards in the ordering GC
1694  *    uses, GC could skip past them
1695  */
1696 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1697 {
1698         u64 start_time = local_clock();
1699         unsigned iter = 0;
1700         int ret;
1701
1702         lockdep_assert_held(&c->state_lock);
1703         trace_gc_start(c);
1704
1705         down_write(&c->gc_lock);
1706
1707         /* flush interior btree updates: */
1708         closure_wait_event(&c->btree_interior_update_wait,
1709                            !bch2_btree_interior_updates_nr_pending(c));
1710
1711         ret   = bch2_gc_start(c, metadata_only) ?:
1712                 bch2_gc_alloc_start(c, metadata_only) ?:
1713                 bch2_gc_reflink_start(c, metadata_only);
1714         if (ret)
1715                 goto out;
1716 again:
1717         gc_pos_set(c, gc_phase(GC_PHASE_START));
1718
1719         bch2_mark_superblocks(c);
1720
1721         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1722             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1723             c->opts.fix_errors != FSCK_OPT_NO) {
1724                 bch_info(c, "starting topology repair pass");
1725                 ret = bch2_repair_topology(c);
1726                 if (ret)
1727                         goto out;
1728                 bch_info(c, "topology repair pass done");
1729
1730                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1731         }
1732
1733         ret = bch2_gc_btrees(c, initial, metadata_only);
1734
1735         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1736             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1737             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1738                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1739                 ret = 0;
1740         }
1741
1742         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1743                 ret = FSCK_ERR_EXIT;
1744
1745         if (ret)
1746                 goto out;
1747
1748 #if 0
1749         bch2_mark_pending_btree_node_frees(c);
1750 #endif
1751         c->gc_count++;
1752
1753         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1754             (!iter && bch2_test_restart_gc)) {
1755                 if (iter++ > 2) {
1756                         bch_info(c, "Unable to fix bucket gens, looping");
1757                         ret = -EINVAL;
1758                         goto out;
1759                 }
1760
1761                 /*
1762                  * XXX: make sure gens we fixed got saved
1763                  */
1764                 bch_info(c, "Second GC pass needed, restarting:");
1765                 clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1766                 __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1767
1768                 bch2_gc_stripes_reset(c, metadata_only);
1769                 bch2_gc_alloc_reset(c, metadata_only);
1770                 bch2_gc_reflink_reset(c, metadata_only);
1771
1772                 /* flush fsck errors, reset counters */
1773                 bch2_flush_fsck_errs(c);
1774                 goto again;
1775         }
1776 out:
1777         if (!ret) {
1778                 bch2_journal_block(&c->journal);
1779
1780                 ret   = bch2_gc_stripes_done(c, metadata_only) ?:
1781                         bch2_gc_reflink_done(c, metadata_only) ?:
1782                         bch2_gc_alloc_done(c, metadata_only) ?:
1783                         bch2_gc_done(c, initial, metadata_only);
1784
1785                 bch2_journal_unblock(&c->journal);
1786         }
1787
1788         percpu_down_write(&c->mark_lock);
1789         /* Indicates that gc is no longer in progress: */
1790         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1791
1792         bch2_gc_free(c);
1793         percpu_up_write(&c->mark_lock);
1794
1795         up_write(&c->gc_lock);
1796
1797         trace_gc_end(c);
1798         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1799
1800         /*
1801          * At startup, allocations can happen directly instead of via the
1802          * allocator thread - issue wakeup in case they blocked on gc_lock:
1803          */
1804         closure_wake_up(&c->freelist_wait);
1805         return ret;
1806 }
1807
1808 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1809 {
1810         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1811         const struct bch_extent_ptr *ptr;
1812
1813         percpu_down_read(&c->mark_lock);
1814         bkey_for_each_ptr(ptrs, ptr) {
1815                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1816
1817                 if (ptr_stale(ca, ptr) > 16) {
1818                         percpu_up_read(&c->mark_lock);
1819                         return true;
1820                 }
1821         }
1822
1823         bkey_for_each_ptr(ptrs, ptr) {
1824                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1825                 u8 *gen = &ca->oldest_gen[PTR_BUCKET_NR(ca, ptr)];
1826
1827                 if (gen_after(*gen, ptr->gen))
1828                         *gen = ptr->gen;
1829         }
1830         percpu_up_read(&c->mark_lock);
1831
1832         return false;
1833 }
1834
1835 /*
1836  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1837  * node pointers currently never have cached pointers that can become stale:
1838  */
1839 static int bch2_gc_btree_gens(struct btree_trans *trans, enum btree_id btree_id)
1840 {
1841         struct bch_fs *c = trans->c;
1842         struct btree_iter iter;
1843         struct bkey_s_c k;
1844         struct bkey_buf sk;
1845         int ret = 0, commit_err = 0;
1846
1847         bch2_bkey_buf_init(&sk);
1848
1849         bch2_trans_iter_init(trans, &iter, btree_id, POS_MIN,
1850                              BTREE_ITER_PREFETCH|
1851                              BTREE_ITER_NOT_EXTENTS|
1852                              BTREE_ITER_ALL_SNAPSHOTS);
1853
1854         while ((bch2_trans_begin(trans),
1855                 k = bch2_btree_iter_peek(&iter)).k) {
1856                 ret = bkey_err(k);
1857
1858                 if (ret == -EINTR)
1859                         continue;
1860                 if (ret)
1861                         break;
1862
1863                 c->gc_gens_pos = iter.pos;
1864
1865                 if (gc_btree_gens_key(c, k) && !commit_err) {
1866                         bch2_bkey_buf_reassemble(&sk, c, k);
1867                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1868
1869                         commit_err =
1870                                 bch2_trans_update(trans, &iter, sk.k, 0) ?:
1871                                 bch2_trans_commit(trans, NULL, NULL,
1872                                                   BTREE_INSERT_NOWAIT|
1873                                                   BTREE_INSERT_NOFAIL);
1874                         if (commit_err == -EINTR) {
1875                                 commit_err = 0;
1876                                 continue;
1877                         }
1878                 }
1879
1880                 bch2_btree_iter_advance(&iter);
1881         }
1882         bch2_trans_iter_exit(trans, &iter);
1883
1884         bch2_bkey_buf_exit(&sk, c);
1885
1886         return ret;
1887 }
1888
1889 static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_iter *iter)
1890 {
1891         struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
1892         struct bkey_s_c k;
1893         struct bkey_alloc_unpacked u;
1894         int ret;
1895
1896         k = bch2_btree_iter_peek_slot(iter);
1897         ret = bkey_err(k);
1898         if (ret)
1899                 return ret;
1900
1901         u = bch2_alloc_unpack(k);
1902
1903         if (u.oldest_gen == ca->oldest_gen[iter->pos.offset])
1904                 return 0;
1905
1906         u.oldest_gen = ca->oldest_gen[iter->pos.offset];
1907
1908         return bch2_alloc_write(trans, iter, &u, 0);
1909 }
1910
1911 int bch2_gc_gens(struct bch_fs *c)
1912 {
1913         struct btree_trans trans;
1914         struct btree_iter iter;
1915         struct bkey_s_c k;
1916         struct bch_dev *ca;
1917         u64 b, start_time = local_clock();
1918         unsigned i;
1919         int ret;
1920
1921         /*
1922          * Ideally we would be using state_lock and not gc_lock here, but that
1923          * introduces a deadlock in the RO path - we currently take the state
1924          * lock at the start of going RO, thus the gc thread may get stuck:
1925          */
1926         if (!mutex_trylock(&c->gc_gens_lock))
1927                 return 0;
1928
1929         down_read(&c->gc_lock);
1930         bch2_trans_init(&trans, c, 0, 0);
1931
1932         for_each_member_device(ca, c, i) {
1933                 struct bucket_gens *gens;
1934
1935                 BUG_ON(ca->oldest_gen);
1936
1937                 ca->oldest_gen = kvmalloc(ca->mi.nbuckets, GFP_KERNEL);
1938                 if (!ca->oldest_gen) {
1939                         percpu_ref_put(&ca->ref);
1940                         ret = -ENOMEM;
1941                         goto err;
1942                 }
1943
1944                 gens = bucket_gens(ca);
1945
1946                 for (b = gens->first_bucket;
1947                      b < gens->nbuckets; b++)
1948                         ca->oldest_gen[b] = gens->b[b];
1949         }
1950
1951         for (i = 0; i < BTREE_ID_NR; i++)
1952                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1953                         c->gc_gens_btree = i;
1954                         c->gc_gens_pos = POS_MIN;
1955                         ret = bch2_gc_btree_gens(&trans, i);
1956                         if (ret) {
1957                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1958                                 goto err;
1959                         }
1960                 }
1961
1962         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1963                            BTREE_ITER_PREFETCH, k, ret) {
1964                 ret = __bch2_trans_do(&trans, NULL, NULL,
1965                                       BTREE_INSERT_NOFAIL,
1966                                 bch2_alloc_write_oldest_gen(&trans, &iter));
1967                 if (ret) {
1968                         bch_err(c, "error writing oldest_gen: %i", ret);
1969                         break;
1970                 }
1971         }
1972         bch2_trans_iter_exit(&trans, &iter);
1973
1974         c->gc_gens_btree        = 0;
1975         c->gc_gens_pos          = POS_MIN;
1976
1977         c->gc_count++;
1978
1979         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1980 err:
1981         for_each_member_device(ca, c, i) {
1982                 kvfree(ca->oldest_gen);
1983                 ca->oldest_gen = NULL;
1984         }
1985
1986         bch2_trans_exit(&trans);
1987         up_read(&c->gc_lock);
1988         mutex_unlock(&c->gc_gens_lock);
1989         return ret;
1990 }
1991
1992 static int bch2_gc_thread(void *arg)
1993 {
1994         struct bch_fs *c = arg;
1995         struct io_clock *clock = &c->io_clock[WRITE];
1996         unsigned long last = atomic64_read(&clock->now);
1997         unsigned last_kick = atomic_read(&c->kick_gc);
1998         int ret;
1999
2000         set_freezable();
2001
2002         while (1) {
2003                 while (1) {
2004                         set_current_state(TASK_INTERRUPTIBLE);
2005
2006                         if (kthread_should_stop()) {
2007                                 __set_current_state(TASK_RUNNING);
2008                                 return 0;
2009                         }
2010
2011                         if (atomic_read(&c->kick_gc) != last_kick)
2012                                 break;
2013
2014                         if (c->btree_gc_periodic) {
2015                                 unsigned long next = last + c->capacity / 16;
2016
2017                                 if (atomic64_read(&clock->now) >= next)
2018                                         break;
2019
2020                                 bch2_io_clock_schedule_timeout(clock, next);
2021                         } else {
2022                                 schedule();
2023                         }
2024
2025                         try_to_freeze();
2026                 }
2027                 __set_current_state(TASK_RUNNING);
2028
2029                 last = atomic64_read(&clock->now);
2030                 last_kick = atomic_read(&c->kick_gc);
2031
2032                 /*
2033                  * Full gc is currently incompatible with btree key cache:
2034                  */
2035 #if 0
2036                 ret = bch2_gc(c, false, false);
2037 #else
2038                 ret = bch2_gc_gens(c);
2039 #endif
2040                 if (ret < 0)
2041                         bch_err(c, "btree gc failed: %i", ret);
2042
2043                 debug_check_no_locks_held();
2044         }
2045
2046         return 0;
2047 }
2048
2049 void bch2_gc_thread_stop(struct bch_fs *c)
2050 {
2051         struct task_struct *p;
2052
2053         p = c->gc_thread;
2054         c->gc_thread = NULL;
2055
2056         if (p) {
2057                 kthread_stop(p);
2058                 put_task_struct(p);
2059         }
2060 }
2061
2062 int bch2_gc_thread_start(struct bch_fs *c)
2063 {
2064         struct task_struct *p;
2065
2066         if (c->gc_thread)
2067                 return 0;
2068
2069         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
2070         if (IS_ERR(p)) {
2071                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
2072                 return PTR_ERR(p);
2073         }
2074
2075         get_task_struct(p);
2076         c->gc_thread = p;
2077         wake_up_process(p);
2078         return 0;
2079 }