git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_gc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  * Copyright (C) 2014 Datera Inc.
   5  */
   6
   7 #include "bcachefs.h"
   8 #include "alloc_background.h"
   9 #include "alloc_foreground.h"
  10 #include "bkey_methods.h"
  11 #include "bkey_buf.h"
  12 #include "btree_locking.h"
  13 #include "btree_update_interior.h"
  14 #include "btree_io.h"
  15 #include "btree_gc.h"
  16 #include "buckets.h"
  17 #include "clock.h"
  18 #include "debug.h"
  19 #include "ec.h"
  20 #include "error.h"
  21 #include "extents.h"
  22 #include "journal.h"
  23 #include "keylist.h"
  24 #include "move.h"
  25 #include "recovery.h"
  26 #include "reflink.h"
  27 #include "replicas.h"
  28 #include "super-io.h"
  29
  30 #include <linux/slab.h>
  31 #include <linux/bitops.h>
  32 #include <linux/freezer.h>
  33 #include <linux/kthread.h>
  34 #include <linux/preempt.h>
  35 #include <linux/rcupdate.h>
  36 #include <linux/sched/task.h>
  37 #include <trace/events/bcachefs.h>
  38
  39 #define DROP_THIS_NODE          10
  40 #define DROP_PREV_NODE          11
  41
  42 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  43 {
  44         preempt_disable();
  45         write_seqcount_begin(&c->gc_pos_lock);
  46         c->gc_pos = new_pos;
  47         write_seqcount_end(&c->gc_pos_lock);
  48         preempt_enable();
  49 }
  50
  51 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  52 {
  53         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  54         __gc_pos_set(c, new_pos);
  55 }
  56
  57 /*
  58  * Missing: if an interior btree node is empty, we need to do something -
  59  * perhaps just kill it
  60  */
  61 static int bch2_gc_check_topology(struct bch_fs *c,
  62                                   struct btree *b,
  63                                   struct bkey_buf *prev,
  64                                   struct bkey_buf cur,
  65                                   bool is_last)
  66 {
  67         struct bpos node_start  = b->data->min_key;
  68         struct bpos node_end    = b->data->max_key;
  69         struct bpos expected_start = bkey_deleted(&prev->k->k)
  70                 ? node_start
  71                 : bpos_successor(prev->k->k.p);
  72         char buf1[200], buf2[200];
  73         int ret = 0;
  74
  75         if (cur.k->k.type == KEY_TYPE_btree_ptr_v2) {
  76                 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(cur.k);
  77
  78                 if (bkey_deleted(&prev->k->k)) {
  79                         struct printbuf out = PBUF(buf1);
  80                         pr_buf(&out, "start of node: ");
  81                         bch2_bpos_to_text(&out, node_start);
  82                 } else {
  83                         bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev->k));
  84                 }
  85
  86                 if (bpos_cmp(expected_start, bp->v.min_key)) {
  87                         bch2_topology_error(c);
  88
  89                         if (__fsck_err(c,
  90                                   FSCK_CAN_FIX|
  91                                   FSCK_CAN_IGNORE|
  92                                   FSCK_NO_RATELIMIT,
  93                                   "btree node with incorrect min_key at btree %s level %u:\n"
  94                                   "  prev %s\n"
  95                                   "  cur %s",
  96                                   bch2_btree_ids[b->c.btree_id], b->c.level,
  97                                   buf1,
  98                                   (bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(cur.k)), buf2)) &&
  99                             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 100                                 bch_info(c, "Halting mark and sweep to start topology repair pass");
 101                                 return FSCK_ERR_START_TOPOLOGY_REPAIR;
 102                         } else {
 103                                 set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 104                         }
 105                 }
 106         }
 107
 108         if (is_last && bpos_cmp(cur.k->k.p, node_end)) {
 109                 bch2_topology_error(c);
 110
 111                 if (__fsck_err(c,
 112                           FSCK_CAN_FIX|
 113                           FSCK_CAN_IGNORE|
 114                           FSCK_NO_RATELIMIT,
 115                           "btree node with incorrect max_key at btree %s level %u:\n"
 116                           "  %s\n"
 117                           "  expected %s",
 118                           bch2_btree_ids[b->c.btree_id], b->c.level,
 119                           (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(cur.k)), buf1),
 120                           (bch2_bpos_to_text(&PBUF(buf2), node_end), buf2)) &&
 121                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 122                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 123                         return FSCK_ERR_START_TOPOLOGY_REPAIR;
 124                 } else {
 125                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 126                 }
 127         }
 128
 129         bch2_bkey_buf_copy(prev, c, cur.k);
 130 fsck_err:
 131         return ret;
 132 }
 133
 134 static void btree_ptr_to_v2(struct btree *b, struct bkey_i_btree_ptr_v2 *dst)
 135 {
 136         switch (b->key.k.type) {
 137         case KEY_TYPE_btree_ptr: {
 138                 struct bkey_i_btree_ptr *src = bkey_i_to_btree_ptr(&b->key);
 139
 140                 dst->k.p                = src->k.p;
 141                 dst->v.mem_ptr          = 0;
 142                 dst->v.seq              = b->data->keys.seq;
 143                 dst->v.sectors_written  = 0;
 144                 dst->v.flags            = 0;
 145                 dst->v.min_key          = b->data->min_key;
 146                 set_bkey_val_bytes(&dst->k, sizeof(dst->v) + bkey_val_bytes(&src->k));
 147                 memcpy(dst->v.start, src->v.start, bkey_val_bytes(&src->k));
 148                 break;
 149         }
 150         case KEY_TYPE_btree_ptr_v2:
 151                 bkey_copy(&dst->k_i, &b->key);
 152                 break;
 153         default:
 154                 BUG();
 155         }
 156 }
 157
 158 static int set_node_min(struct bch_fs *c, struct btree *b, struct bpos new_min)
 159 {
 160         struct bkey_i_btree_ptr_v2 *new;
 161         int ret;
 162
 163         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 164         if (!new)
 165                 return -ENOMEM;
 166
 167         btree_ptr_to_v2(b, new);
 168         b->data->min_key        = new_min;
 169         new->v.min_key          = new_min;
 170         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 171
 172         ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 173         if (ret) {
 174                 kfree(new);
 175                 return ret;
 176         }
 177
 178         bch2_btree_node_drop_keys_outside_node(b);
 179
 180         return 0;
 181 }
 182
 183 static int set_node_max(struct bch_fs *c, struct btree *b, struct bpos new_max)
 184 {
 185         struct bkey_i_btree_ptr_v2 *new;
 186         int ret;
 187
 188         ret = bch2_journal_key_delete(c, b->c.btree_id, b->c.level + 1, b->key.k.p);
 189         if (ret)
 190                 return ret;
 191
 192         new = kmalloc(BKEY_BTREE_PTR_U64s_MAX * sizeof(u64), GFP_KERNEL);
 193         if (!new)
 194                 return -ENOMEM;
 195
 196         btree_ptr_to_v2(b, new);
 197         b->data->max_key        = new_max;
 198         new->k.p                = new_max;
 199         SET_BTREE_PTR_RANGE_UPDATED(&new->v, true);
 200
 201         ret = bch2_journal_key_insert(c, b->c.btree_id, b->c.level + 1, &new->k_i);
 202         if (ret) {
 203                 kfree(new);
 204                 return ret;
 205         }
 206
 207         bch2_btree_node_drop_keys_outside_node(b);
 208
 209         mutex_lock(&c->btree_cache.lock);
 210         bch2_btree_node_hash_remove(&c->btree_cache, b);
 211
 212         bkey_copy(&b->key, &new->k_i);
 213         ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 214         BUG_ON(ret);
 215         mutex_unlock(&c->btree_cache.lock);
 216         return 0;
 217 }
 218
 219 static int btree_repair_node_boundaries(struct bch_fs *c, struct btree *b,
 220                                         struct btree *prev, struct btree *cur)
 221 {
 222         struct bpos expected_start = !prev
 223                 ? b->data->min_key
 224                 : bpos_successor(prev->key.k.p);
 225         char buf1[200], buf2[200];
 226         int ret = 0;
 227
 228         if (!prev) {
 229                 struct printbuf out = PBUF(buf1);
 230                 pr_buf(&out, "start of node: ");
 231                 bch2_bpos_to_text(&out, b->data->min_key);
 232         } else {
 233                 bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&prev->key));
 234         }
 235
 236         bch2_bkey_val_to_text(&PBUF(buf2), c, bkey_i_to_s_c(&cur->key));
 237
 238         if (prev &&
 239             bpos_cmp(expected_start, cur->data->min_key) > 0 &&
 240             BTREE_NODE_SEQ(cur->data) > BTREE_NODE_SEQ(prev->data)) {
 241                 /* cur overwrites prev: */
 242
 243                 if (mustfix_fsck_err_on(bpos_cmp(prev->data->min_key,
 244                                                  cur->data->min_key) >= 0, c,
 245                                 "btree node overwritten by next node at btree %s level %u:\n"
 246                                 "  node %s\n"
 247                                 "  next %s",
 248                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 249                                 buf1, buf2))
 250                         return DROP_PREV_NODE;
 251
 252                 if (mustfix_fsck_err_on(bpos_cmp(prev->key.k.p,
 253                                                  bpos_predecessor(cur->data->min_key)), c,
 254                                 "btree node with incorrect max_key at btree %s level %u:\n"
 255                                 "  node %s\n"
 256                                 "  next %s",
 257                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 258                                 buf1, buf2))
 259                         ret = set_node_max(c, prev,
 260                                            bpos_predecessor(cur->data->min_key));
 261         } else {
 262                 /* prev overwrites cur: */
 263
 264                 if (mustfix_fsck_err_on(bpos_cmp(expected_start,
 265                                                  cur->data->max_key) >= 0, c,
 266                                 "btree node overwritten by prev node at btree %s level %u:\n"
 267                                 "  prev %s\n"
 268                                 "  node %s",
 269                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 270                                 buf1, buf2))
 271                         return DROP_THIS_NODE;
 272
 273                 if (mustfix_fsck_err_on(bpos_cmp(expected_start, cur->data->min_key), c,
 274                                 "btree node with incorrect min_key at btree %s level %u:\n"
 275                                 "  prev %s\n"
 276                                 "  node %s",
 277                                 bch2_btree_ids[b->c.btree_id], b->c.level,
 278                                 buf1, buf2))
 279                     ret = set_node_min(c, cur, expected_start);
 280         }
 281 fsck_err:
 282         return ret;
 283 }
 284
 285 static int btree_repair_node_end(struct bch_fs *c, struct btree *b,
 286                                  struct btree *child)
 287 {
 288         char buf1[200], buf2[200];
 289         int ret = 0;
 290
 291         if (mustfix_fsck_err_on(bpos_cmp(child->key.k.p, b->key.k.p), c,
 292                         "btree node with incorrect max_key at btree %s level %u:\n"
 293                         "  %s\n"
 294                         "  expected %s",
 295                         bch2_btree_ids[b->c.btree_id], b->c.level,
 296                         (bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(&child->key)), buf1),
 297                         (bch2_bpos_to_text(&PBUF(buf2), b->key.k.p), buf2))) {
 298                 ret = set_node_max(c, child, b->key.k.p);
 299                 if (ret)
 300                         return ret;
 301         }
 302 fsck_err:
 303         return ret;
 304 }
 305
 306 static int bch2_btree_repair_topology_recurse(struct bch_fs *c, struct btree *b)
 307 {
 308         struct btree_and_journal_iter iter;
 309         struct bkey_s_c k;
 310         struct bkey_buf prev_k, cur_k;
 311         struct btree *prev = NULL, *cur = NULL;
 312         bool have_child, dropped_children = false;
 313         char buf[200];
 314         int ret = 0;
 315
 316         if (!b->c.level)
 317                 return 0;
 318 again:
 319         prev = NULL;
 320         have_child = dropped_children = false;
 321         bch2_bkey_buf_init(&prev_k);
 322         bch2_bkey_buf_init(&cur_k);
 323         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 324
 325         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 326                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 327                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 328
 329                 bch2_btree_and_journal_iter_advance(&iter);
 330                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 331
 332                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 333                                         b->c.btree_id, b->c.level - 1,
 334                                         false);
 335                 ret = PTR_ERR_OR_ZERO(cur);
 336
 337                 if (mustfix_fsck_err_on(ret == -EIO, c,
 338                                 "Unreadable btree node at btree %s level %u:\n"
 339                                 "  %s",
 340                                 bch2_btree_ids[b->c.btree_id],
 341                                 b->c.level - 1,
 342                                 (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur_k.k)), buf))) {
 343                         bch2_btree_node_evict(c, cur_k.k);
 344                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 345                                                       b->c.level, cur_k.k->k.p);
 346                         if (ret)
 347                                 break;
 348                         continue;
 349                 }
 350
 351                 if (ret) {
 352                         bch_err(c, "%s: error %i getting btree node",
 353                                 __func__, ret);
 354                         break;
 355                 }
 356
 357                 ret = btree_repair_node_boundaries(c, b, prev, cur);
 358
 359                 if (ret == DROP_THIS_NODE) {
 360                         six_unlock_read(&cur->c.lock);
 361                         bch2_btree_node_evict(c, cur_k.k);
 362                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 363                                                       b->c.level, cur_k.k->k.p);
 364                         if (ret)
 365                                 break;
 366                         continue;
 367                 }
 368
 369                 if (prev)
 370                         six_unlock_read(&prev->c.lock);
 371                 prev = NULL;
 372
 373                 if (ret == DROP_PREV_NODE) {
 374                         bch2_btree_node_evict(c, prev_k.k);
 375                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 376                                                       b->c.level, prev_k.k->k.p);
 377                         if (ret)
 378                                 break;
 379
 380                         bch2_btree_and_journal_iter_exit(&iter);
 381                         bch2_bkey_buf_exit(&prev_k, c);
 382                         bch2_bkey_buf_exit(&cur_k, c);
 383                         goto again;
 384                 } else if (ret)
 385                         break;
 386
 387                 prev = cur;
 388                 cur = NULL;
 389                 bch2_bkey_buf_copy(&prev_k, c, cur_k.k);
 390         }
 391
 392         if (!ret && !IS_ERR_OR_NULL(prev)) {
 393                 BUG_ON(cur);
 394                 ret = btree_repair_node_end(c, b, prev);
 395         }
 396
 397         if (!IS_ERR_OR_NULL(prev))
 398                 six_unlock_read(&prev->c.lock);
 399         prev = NULL;
 400         if (!IS_ERR_OR_NULL(cur))
 401                 six_unlock_read(&cur->c.lock);
 402         cur = NULL;
 403
 404         if (ret)
 405                 goto err;
 406
 407         bch2_btree_and_journal_iter_exit(&iter);
 408         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 409
 410         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 411                 bch2_bkey_buf_reassemble(&cur_k, c, k);
 412                 bch2_btree_and_journal_iter_advance(&iter);
 413
 414                 cur = bch2_btree_node_get_noiter(c, cur_k.k,
 415                                         b->c.btree_id, b->c.level - 1,
 416                                         false);
 417                 ret = PTR_ERR_OR_ZERO(cur);
 418
 419                 if (ret) {
 420                         bch_err(c, "%s: error %i getting btree node",
 421                                 __func__, ret);
 422                         goto err;
 423                 }
 424
 425                 ret = bch2_btree_repair_topology_recurse(c, cur);
 426                 six_unlock_read(&cur->c.lock);
 427                 cur = NULL;
 428
 429                 if (ret == DROP_THIS_NODE) {
 430                         bch2_btree_node_evict(c, cur_k.k);
 431                         ret = bch2_journal_key_delete(c, b->c.btree_id,
 432                                                       b->c.level, cur_k.k->k.p);
 433                         dropped_children = true;
 434                 }
 435
 436                 if (ret)
 437                         goto err;
 438
 439                 have_child = true;
 440         }
 441
 442         if (mustfix_fsck_err_on(!have_child, c,
 443                         "empty interior btree node at btree %s level %u\n"
 444                         "  %s",
 445                         bch2_btree_ids[b->c.btree_id],
 446                         b->c.level,
 447                         (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(&b->key)), buf)))
 448                 ret = DROP_THIS_NODE;
 449 err:
 450 fsck_err:
 451         if (!IS_ERR_OR_NULL(prev))
 452                 six_unlock_read(&prev->c.lock);
 453         if (!IS_ERR_OR_NULL(cur))
 454                 six_unlock_read(&cur->c.lock);
 455
 456         bch2_btree_and_journal_iter_exit(&iter);
 457         bch2_bkey_buf_exit(&prev_k, c);
 458         bch2_bkey_buf_exit(&cur_k, c);
 459
 460         if (!ret && dropped_children)
 461                 goto again;
 462
 463         return ret;
 464 }
 465
 466 static int bch2_repair_topology(struct bch_fs *c)
 467 {
 468         struct btree *b;
 469         unsigned i;
 470         int ret = 0;
 471
 472         for (i = 0; i < BTREE_ID_NR && !ret; i++) {
 473                 b = c->btree_roots[i].b;
 474                 if (btree_node_fake(b))
 475                         continue;
 476
 477                 six_lock_read(&b->c.lock, NULL, NULL);
 478                 ret = bch2_btree_repair_topology_recurse(c, b);
 479                 six_unlock_read(&b->c.lock);
 480
 481                 if (ret == DROP_THIS_NODE) {
 482                         bch_err(c, "empty btree root - repair unimplemented");
 483                         ret = FSCK_ERR_EXIT;
 484                 }
 485         }
 486
 487         return ret;
 488 }
 489
 490 static int bch2_check_fix_ptrs(struct bch_fs *c, enum btree_id btree_id,
 491                                unsigned level, bool is_root,
 492                                struct bkey_s_c *k)
 493 {
 494         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(*k);
 495         const union bch_extent_entry *entry;
 496         struct extent_ptr_decoded p = { 0 };
 497         bool do_update = false;
 498         char buf[200];
 499         int ret = 0;
 500
 501         /*
 502          * XXX
 503          * use check_bucket_ref here
 504          */
 505         bkey_for_each_ptr_decode(k->k, ptrs, p, entry) {
 506                 struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 507                 struct bucket *g = PTR_BUCKET(ca, &p.ptr, true);
 508                 struct bucket *g2 = PTR_BUCKET(ca, &p.ptr, false);
 509                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, &entry->ptr);
 510
 511                 if (fsck_err_on(!g->gen_valid, c,
 512                                 "bucket %u:%zu data type %s ptr gen %u missing in alloc btree\n"
 513                                 "while marking %s",
 514                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 515                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 516                                 p.ptr.gen,
 517                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 518                         if (!p.ptr.cached) {
 519                                 g2->_mark.gen   = g->_mark.gen          = p.ptr.gen;
 520                                 g2->gen_valid   = g->gen_valid          = true;
 521                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 522                         } else {
 523                                 do_update = true;
 524                         }
 525                 }
 526
 527                 if (fsck_err_on(data_type == BCH_DATA_btree &&
 528                                 g->mark.gen != p.ptr.gen, c,
 529                                 "bucket %u:%zu data type %s has metadata but wrong gen: %u != %u\n"
 530                                 "while marking %s",
 531                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 532                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 533                                 p.ptr.gen, g->mark.gen,
 534                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 535                         g2->_mark.data_type     = g->_mark.data_type    = data_type;
 536                         g2->gen_valid           = g->gen_valid          = true;
 537                         set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 538                 }
 539
 540                 if (fsck_err_on(gen_cmp(p.ptr.gen, g->mark.gen) > 0, c,
 541                                 "bucket %u:%zu data type %s ptr gen in the future: %u > %u\n"
 542                                 "while marking %s",
 543                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 544                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 545                                 p.ptr.gen, g->mark.gen,
 546                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 547                         if (!p.ptr.cached) {
 548                                 g2->_mark.gen   = g->_mark.gen  = p.ptr.gen;
 549                                 g2->gen_valid   = g->gen_valid  = true;
 550                                 g2->_mark.data_type             = 0;
 551                                 g2->_mark.dirty_sectors         = 0;
 552                                 g2->_mark.cached_sectors        = 0;
 553                                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
 554                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 555                         } else {
 556                                 do_update = true;
 557                         }
 558                 }
 559
 560                 if (fsck_err_on(gen_cmp(g->mark.gen, p.ptr.gen) > BUCKET_GC_GEN_MAX, c,
 561                                 "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 562                                 "while marking %s",
 563                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr), g->mark.gen,
 564                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 565                                 p.ptr.gen,
 566                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 567                         do_update = true;
 568
 569                 if (fsck_err_on(!p.ptr.cached &&
 570                                 gen_cmp(p.ptr.gen, g->mark.gen) < 0, c,
 571                                 "bucket %u:%zu data type %s stale dirty ptr: %u < %u\n"
 572                                 "while marking %s",
 573                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 574                                 bch2_data_types[ptr_data_type(k->k, &p.ptr)],
 575                                 p.ptr.gen, g->mark.gen,
 576                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 577                         do_update = true;
 578
 579                 if (p.ptr.gen != g->mark.gen)
 580                         continue;
 581
 582                 if (fsck_err_on(g->mark.data_type &&
 583                                 g->mark.data_type != data_type, c,
 584                                 "bucket %u:%zu different types of data in same bucket: %s, %s\n"
 585                                 "while marking %s",
 586                                 p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 587                                 bch2_data_types[g->mark.data_type],
 588                                 bch2_data_types[data_type],
 589                                 (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf))) {
 590                         if (data_type == BCH_DATA_btree) {
 591                                 g2->_mark.data_type     = g->_mark.data_type    = data_type;
 592                                 g2->gen_valid           = g->gen_valid          = true;
 593                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 594                         } else {
 595                                 do_update = true;
 596                         }
 597                 }
 598
 599                 if (p.has_ec) {
 600                         struct stripe *m = genradix_ptr(&c->stripes[true], p.ec.idx);
 601
 602                         if (fsck_err_on(!m || !m->alive, c,
 603                                         "pointer to nonexistent stripe %llu\n"
 604                                         "while marking %s",
 605                                         (u64) p.ec.idx,
 606                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 607                                 do_update = true;
 608
 609                         if (fsck_err_on(!bch2_ptr_matches_stripe_m(m, p), c,
 610                                         "pointer does not match stripe %llu\n"
 611                                         "while marking %s",
 612                                         (u64) p.ec.idx,
 613                                         (bch2_bkey_val_to_text(&PBUF(buf), c, *k), buf)))
 614                                 do_update = true;
 615                 }
 616         }
 617
 618         if (do_update) {
 619                 struct bkey_ptrs ptrs;
 620                 union bch_extent_entry *entry;
 621                 struct bch_extent_ptr *ptr;
 622                 struct bkey_i *new;
 623
 624                 if (is_root) {
 625                         bch_err(c, "cannot update btree roots yet");
 626                         return -EINVAL;
 627                 }
 628
 629                 new = kmalloc(bkey_bytes(k->k), GFP_KERNEL);
 630                 if (!new) {
 631                         bch_err(c, "%s: error allocating new key", __func__);
 632                         return -ENOMEM;
 633                 }
 634
 635                 bkey_reassemble(new, *k);
 636
 637                 if (level) {
 638                         /*
 639                          * We don't want to drop btree node pointers - if the
 640                          * btree node isn't there anymore, the read path will
 641                          * sort it out:
 642                          */
 643                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 644                         bkey_for_each_ptr(ptrs, ptr) {
 645                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 646                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 647
 648                                 ptr->gen = g->mark.gen;
 649                         }
 650                 } else {
 651                         bch2_bkey_drop_ptrs(bkey_i_to_s(new), ptr, ({
 652                                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 653                                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 654                                 enum bch_data_type data_type = bch2_bkey_ptr_data_type(*k, ptr);
 655
 656                                 (ptr->cached &&
 657                                  (!g->gen_valid || gen_cmp(ptr->gen, g->mark.gen) > 0)) ||
 658                                 (!ptr->cached &&
 659                                  gen_cmp(ptr->gen, g->mark.gen) < 0) ||
 660                                 gen_cmp(g->mark.gen, ptr->gen) > BUCKET_GC_GEN_MAX ||
 661                                 (g->mark.data_type &&
 662                                  g->mark.data_type != data_type);
 663                         }));
 664 again:
 665                         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
 666                         bkey_extent_entry_for_each(ptrs, entry) {
 667                                 if (extent_entry_type(entry) == BCH_EXTENT_ENTRY_stripe_ptr) {
 668                                         struct stripe *m = genradix_ptr(&c->stripes[true],
 669                                                                         entry->stripe_ptr.idx);
 670                                         union bch_extent_entry *next_ptr;
 671
 672                                         bkey_extent_entry_for_each_from(ptrs, next_ptr, entry)
 673                                                 if (extent_entry_type(next_ptr) == BCH_EXTENT_ENTRY_ptr)
 674                                                         goto found;
 675                                         next_ptr = NULL;
 676 found:
 677                                         if (!next_ptr) {
 678                                                 bch_err(c, "aieee, found stripe ptr with no data ptr");
 679                                                 continue;
 680                                         }
 681
 682                                         if (!m || !m->alive ||
 683                                             !__bch2_ptr_matches_stripe(&m->ptrs[entry->stripe_ptr.block],
 684                                                                        &next_ptr->ptr,
 685                                                                        m->sectors)) {
 686                                                 bch2_bkey_extent_entry_drop(new, entry);
 687                                                 goto again;
 688                                         }
 689                                 }
 690                         }
 691                 }
 692
 693                 ret = bch2_journal_key_insert(c, btree_id, level, new);
 694                 if (ret)
 695                         kfree(new);
 696                 else
 697                         *k = bkey_i_to_s_c(new);
 698         }
 699 fsck_err:
 700         return ret;
 701 }
 702
 703 /* marking of btree keys/nodes: */
 704
 705 static int bch2_gc_mark_key(struct btree_trans *trans, enum btree_id btree_id,
 706                             unsigned level, bool is_root,
 707                             struct bkey_s_c *k,
 708                             u8 *max_stale, bool initial)
 709 {
 710         struct bch_fs *c = trans->c;
 711         struct bkey_ptrs_c ptrs;
 712         const struct bch_extent_ptr *ptr;
 713         struct bkey deleted = KEY(0, 0, 0);
 714         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
 715         unsigned flags =
 716                 BTREE_TRIGGER_GC|
 717                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 718         int ret = 0;
 719
 720         deleted.p = k->k->p;
 721
 722         if (initial) {
 723                 BUG_ON(bch2_journal_seq_verify &&
 724                        k->k->version.lo > journal_cur_seq(&c->journal));
 725
 726                 ret = bch2_check_fix_ptrs(c, btree_id, level, is_root, k);
 727                 if (ret)
 728                         goto err;
 729
 730                 if (fsck_err_on(k->k->version.lo > atomic64_read(&c->key_version), c,
 731                                 "key version number higher than recorded: %llu > %llu",
 732                                 k->k->version.lo,
 733                                 atomic64_read(&c->key_version)))
 734                         atomic64_set(&c->key_version, k->k->version.lo);
 735         }
 736
 737         ptrs = bch2_bkey_ptrs_c(*k);
 738         bkey_for_each_ptr(ptrs, ptr) {
 739                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 740                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 741
 742                 if (gen_after(g->oldest_gen, ptr->gen))
 743                         g->oldest_gen = ptr->gen;
 744
 745                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 746         }
 747
 748         ret = bch2_mark_key(trans, old, *k, flags);
 749 fsck_err:
 750 err:
 751         if (ret)
 752                 bch_err(c, "%s: ret %i", __func__, ret);
 753         return ret;
 754 }
 755
 756 static int btree_gc_mark_node(struct btree_trans *trans, struct btree *b, u8 *max_stale,
 757                               bool initial)
 758 {
 759         struct bch_fs *c = trans->c;
 760         struct btree_node_iter iter;
 761         struct bkey unpacked;
 762         struct bkey_s_c k;
 763         struct bkey_buf prev, cur;
 764         int ret = 0;
 765
 766         *max_stale = 0;
 767
 768         if (!btree_node_type_needs_gc(btree_node_type(b)))
 769                 return 0;
 770
 771         bch2_btree_node_iter_init_from_start(&iter, b);
 772         bch2_bkey_buf_init(&prev);
 773         bch2_bkey_buf_init(&cur);
 774         bkey_init(&prev.k->k);
 775
 776         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 777                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 778                                        &k, max_stale, initial);
 779                 if (ret)
 780                         break;
 781
 782                 bch2_btree_node_iter_advance(&iter, b);
 783
 784                 if (b->c.level) {
 785                         bch2_bkey_buf_reassemble(&cur, c, k);
 786
 787                         ret = bch2_gc_check_topology(c, b, &prev, cur,
 788                                         bch2_btree_node_iter_end(&iter));
 789                         if (ret)
 790                                 break;
 791                 }
 792         }
 793
 794         bch2_bkey_buf_exit(&cur, c);
 795         bch2_bkey_buf_exit(&prev, c);
 796         return ret;
 797 }
 798
 799 static int bch2_gc_btree(struct btree_trans *trans, enum btree_id btree_id,
 800                          bool initial, bool metadata_only)
 801 {
 802         struct bch_fs *c = trans->c;
 803         struct btree_iter iter;
 804         struct btree *b;
 805         unsigned depth = metadata_only                  ? 1
 806                 : bch2_expensive_debug_checks           ? 0
 807                 : !btree_node_type_needs_gc(btree_id)   ? 1
 808                 : 0;
 809         u8 max_stale = 0;
 810         int ret = 0;
 811
 812         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 813
 814         __for_each_btree_node(trans, iter, btree_id, POS_MIN,
 815                               0, depth, BTREE_ITER_PREFETCH, b, ret) {
 816                 bch2_verify_btree_nr_keys(b);
 817
 818                 gc_pos_set(c, gc_pos_btree_node(b));
 819
 820                 ret = btree_gc_mark_node(trans, b, &max_stale, initial);
 821                 if (ret)
 822                         break;
 823
 824                 if (!initial) {
 825                         if (max_stale > 64)
 826                                 bch2_btree_node_rewrite(trans, &iter, b,
 827                                                 BTREE_INSERT_NOWAIT|
 828                                                 BTREE_INSERT_GC_LOCK_HELD);
 829                         else if (!bch2_btree_gc_rewrite_disabled &&
 830                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 831                                 bch2_btree_node_rewrite(trans, &iter,
 832                                                 b, BTREE_INSERT_NOWAIT|
 833                                                 BTREE_INSERT_GC_LOCK_HELD);
 834                 }
 835         }
 836         bch2_trans_iter_exit(trans, &iter);
 837
 838         if (ret)
 839                 return ret;
 840
 841         mutex_lock(&c->btree_root_lock);
 842         b = c->btree_roots[btree_id].b;
 843         if (!btree_node_fake(b)) {
 844                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 845
 846                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
 847                                        &k, &max_stale, initial);
 848         }
 849         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 850         mutex_unlock(&c->btree_root_lock);
 851
 852         return ret;
 853 }
 854
 855 static int bch2_gc_btree_init_recurse(struct btree_trans *trans, struct btree *b,
 856                                       unsigned target_depth)
 857 {
 858         struct bch_fs *c = trans->c;
 859         struct btree_and_journal_iter iter;
 860         struct bkey_s_c k;
 861         struct bkey_buf cur, prev;
 862         u8 max_stale = 0;
 863         char buf[200];
 864         int ret = 0;
 865
 866         bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 867         bch2_bkey_buf_init(&prev);
 868         bch2_bkey_buf_init(&cur);
 869         bkey_init(&prev.k->k);
 870
 871         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 872                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 873                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 874
 875                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, false,
 876                                        &k, &max_stale, true);
 877                 if (ret) {
 878                         bch_err(c, "%s: error %i from bch2_gc_mark_key", __func__, ret);
 879                         goto fsck_err;
 880                 }
 881
 882                 if (b->c.level) {
 883                         bch2_bkey_buf_reassemble(&cur, c, k);
 884                         k = bkey_i_to_s_c(cur.k);
 885
 886                         bch2_btree_and_journal_iter_advance(&iter);
 887
 888                         ret = bch2_gc_check_topology(c, b,
 889                                         &prev, cur,
 890                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 891                         if (ret)
 892                                 goto fsck_err;
 893                 } else {
 894                         bch2_btree_and_journal_iter_advance(&iter);
 895                 }
 896         }
 897
 898         if (b->c.level > target_depth) {
 899                 bch2_btree_and_journal_iter_exit(&iter);
 900                 bch2_btree_and_journal_iter_init_node_iter(&iter, c, b);
 901
 902                 while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 903                         struct btree *child;
 904
 905                         bch2_bkey_buf_reassemble(&cur, c, k);
 906                         bch2_btree_and_journal_iter_advance(&iter);
 907
 908                         child = bch2_btree_node_get_noiter(c, cur.k,
 909                                                 b->c.btree_id, b->c.level - 1,
 910                                                 false);
 911                         ret = PTR_ERR_OR_ZERO(child);
 912
 913                         if (ret == -EIO) {
 914                                 bch2_topology_error(c);
 915
 916                                 if (__fsck_err(c,
 917                                           FSCK_CAN_FIX|
 918                                           FSCK_CAN_IGNORE|
 919                                           FSCK_NO_RATELIMIT,
 920                                           "Unreadable btree node at btree %s level %u:\n"
 921                                           "  %s",
 922                                           bch2_btree_ids[b->c.btree_id],
 923                                           b->c.level - 1,
 924                                           (bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(cur.k)), buf)) &&
 925                                     !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags)) {
 926                                         ret = FSCK_ERR_START_TOPOLOGY_REPAIR;
 927                                         bch_info(c, "Halting mark and sweep to start topology repair pass");
 928                                         goto fsck_err;
 929                                 } else {
 930                                         /* Continue marking when opted to not
 931                                          * fix the error: */
 932                                         ret = 0;
 933                                         set_bit(BCH_FS_INITIAL_GC_UNFIXED, &c->flags);
 934                                         continue;
 935                                 }
 936                         } else if (ret) {
 937                                 bch_err(c, "%s: error %i getting btree node",
 938                                         __func__, ret);
 939                                 break;
 940                         }
 941
 942                         ret = bch2_gc_btree_init_recurse(trans, child,
 943                                                          target_depth);
 944                         six_unlock_read(&child->c.lock);
 945
 946                         if (ret)
 947                                 break;
 948                 }
 949         }
 950 fsck_err:
 951         bch2_bkey_buf_exit(&cur, c);
 952         bch2_bkey_buf_exit(&prev, c);
 953         bch2_btree_and_journal_iter_exit(&iter);
 954         return ret;
 955 }
 956
 957 static int bch2_gc_btree_init(struct btree_trans *trans,
 958                               enum btree_id btree_id,
 959                               bool metadata_only)
 960 {
 961         struct bch_fs *c = trans->c;
 962         struct btree *b;
 963         unsigned target_depth = metadata_only           ? 1
 964                 : bch2_expensive_debug_checks           ? 0
 965                 : !btree_node_type_needs_gc(btree_id)   ? 1
 966                 : 0;
 967         u8 max_stale = 0;
 968         char buf[100];
 969         int ret = 0;
 970
 971         b = c->btree_roots[btree_id].b;
 972
 973         if (btree_node_fake(b))
 974                 return 0;
 975
 976         six_lock_read(&b->c.lock, NULL, NULL);
 977         if (mustfix_fsck_err_on(bpos_cmp(b->data->min_key, POS_MIN), c,
 978                         "btree root with incorrect min_key: %s",
 979                         (bch2_bpos_to_text(&PBUF(buf), b->data->min_key), buf))) {
 980                 bch_err(c, "repair unimplemented");
 981                 ret = FSCK_ERR_EXIT;
 982                 goto fsck_err;
 983         }
 984
 985         if (mustfix_fsck_err_on(bpos_cmp(b->data->max_key, SPOS_MAX), c,
 986                         "btree root with incorrect max_key: %s",
 987                         (bch2_bpos_to_text(&PBUF(buf), b->data->max_key), buf))) {
 988                 bch_err(c, "repair unimplemented");
 989                 ret = FSCK_ERR_EXIT;
 990                 goto fsck_err;
 991         }
 992
 993         if (b->c.level >= target_depth)
 994                 ret = bch2_gc_btree_init_recurse(trans, b, target_depth);
 995
 996         if (!ret) {
 997                 struct bkey_s_c k = bkey_i_to_s_c(&b->key);
 998
 999                 ret = bch2_gc_mark_key(trans, b->c.btree_id, b->c.level, true,
1000                                        &k, &max_stale, true);
1001         }
1002 fsck_err:
1003         six_unlock_read(&b->c.lock);
1004
1005         if (ret < 0)
1006                 bch_err(c, "%s: ret %i", __func__, ret);
1007         return ret;
1008 }
1009
1010 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
1011 {
1012         return  (int) btree_id_to_gc_phase(l) -
1013                 (int) btree_id_to_gc_phase(r);
1014 }
1015
1016 static int bch2_gc_btrees(struct bch_fs *c, bool initial, bool metadata_only)
1017 {
1018         struct btree_trans trans;
1019         enum btree_id ids[BTREE_ID_NR];
1020         unsigned i;
1021         int ret = 0;
1022
1023         bch2_trans_init(&trans, c, 0, 0);
1024
1025         for (i = 0; i < BTREE_ID_NR; i++)
1026                 ids[i] = i;
1027         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
1028
1029         for (i = 0; i < BTREE_ID_NR && !ret; i++)
1030                 ret = initial
1031                         ? bch2_gc_btree_init(&trans, ids[i], metadata_only)
1032                         : bch2_gc_btree(&trans, ids[i], initial, metadata_only);
1033
1034         if (ret < 0)
1035                 bch_err(c, "%s: ret %i", __func__, ret);
1036
1037         bch2_trans_exit(&trans);
1038         return ret;
1039 }
1040
1041 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
1042                                   u64 start, u64 end,
1043                                   enum bch_data_type type,
1044                                   unsigned flags)
1045 {
1046         u64 b = sector_to_bucket(ca, start);
1047
1048         do {
1049                 unsigned sectors =
1050                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1051
1052                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
1053                                           gc_phase(GC_PHASE_SB), flags);
1054                 b++;
1055                 start += sectors;
1056         } while (start < end);
1057 }
1058
1059 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
1060                               unsigned flags)
1061 {
1062         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1063         unsigned i;
1064         u64 b;
1065
1066         /*
1067          * This conditional is kind of gross, but we may be called from the
1068          * device add path, before the new device has actually been added to the
1069          * running filesystem:
1070          */
1071         if (c) {
1072                 lockdep_assert_held(&c->sb_lock);
1073                 percpu_down_read(&c->mark_lock);
1074         }
1075
1076         for (i = 0; i < layout->nr_superblocks; i++) {
1077                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1078
1079                 if (offset == BCH_SB_SECTOR)
1080                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
1081                                               BCH_DATA_sb, flags);
1082
1083                 mark_metadata_sectors(c, ca, offset,
1084                                       offset + (1 << layout->sb_max_size_bits),
1085                                       BCH_DATA_sb, flags);
1086         }
1087
1088         for (i = 0; i < ca->journal.nr; i++) {
1089                 b = ca->journal.buckets[i];
1090                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
1091                                           ca->mi.bucket_size,
1092                                           gc_phase(GC_PHASE_SB), flags);
1093         }
1094
1095         if (c)
1096                 percpu_up_read(&c->mark_lock);
1097 }
1098
1099 static void bch2_mark_superblocks(struct bch_fs *c)
1100 {
1101         struct bch_dev *ca;
1102         unsigned i;
1103
1104         mutex_lock(&c->sb_lock);
1105         gc_pos_set(c, gc_phase(GC_PHASE_SB));
1106
1107         for_each_online_member(ca, c, i)
1108                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
1109         mutex_unlock(&c->sb_lock);
1110 }
1111
1112 #if 0
1113 /* Also see bch2_pending_btree_node_free_insert_done() */
1114 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
1115 {
1116         struct btree_update *as;
1117         struct pending_btree_node_free *d;
1118
1119         mutex_lock(&c->btree_interior_update_lock);
1120         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
1121
1122         for_each_pending_btree_node_free(c, as, d)
1123                 if (d->index_update_done)
1124                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), BTREE_TRIGGER_GC);
1125
1126         mutex_unlock(&c->btree_interior_update_lock);
1127 }
1128 #endif
1129
1130 static void bch2_gc_free(struct bch_fs *c)
1131 {
1132         struct bch_dev *ca;
1133         unsigned i;
1134
1135         genradix_free(&c->stripes[1]);
1136
1137         for_each_member_device(ca, c, i) {
1138                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
1139                         sizeof(struct bucket_array) +
1140                         ca->mi.nbuckets * sizeof(struct bucket));
1141                 ca->buckets[1] = NULL;
1142
1143                 free_percpu(ca->usage_gc);
1144                 ca->usage_gc = NULL;
1145         }
1146
1147         free_percpu(c->usage_gc);
1148         c->usage_gc = NULL;
1149 }
1150
1151 static int bch2_gc_done(struct bch_fs *c,
1152                         bool initial, bool metadata_only)
1153 {
1154         struct bch_dev *ca = NULL;
1155         bool verify = !metadata_only && (!initial ||
1156                        (c->sb.compat & (1ULL << BCH_COMPAT_alloc_info)));
1157         unsigned i, dev;
1158         int ret = 0;
1159
1160 #define copy_field(_f, _msg, ...)                                       \
1161         if (dst->_f != src->_f) {                                       \
1162                 if (verify)                                             \
1163                         fsck_err(c, _msg ": got %llu, should be %llu"   \
1164                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
1165                 dst->_f = src->_f;                                      \
1166                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1167         }
1168 #define copy_stripe_field(_f, _msg, ...)                                \
1169         if (dst->_f != src->_f) {                                       \
1170                 if (verify)                                             \
1171                         fsck_err(c, "stripe %zu has wrong "_msg         \
1172                                 ": got %u, should be %u",               \
1173                                 iter.pos, ##__VA_ARGS__,                \
1174                                 dst->_f, src->_f);                      \
1175                 dst->_f = src->_f;                                      \
1176                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1177         }
1178 #define copy_bucket_field(_f)                                           \
1179         if (dst->b[b]._f != src->b[b]._f) {                             \
1180                 if (verify)                                             \
1181                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
1182                                 ": got %u, should be %u", dev, b,       \
1183                                 dst->b[b].mark.gen,                     \
1184                                 bch2_data_types[dst->b[b].mark.data_type],\
1185                                 dst->b[b]._f, src->b[b]._f);            \
1186                 dst->b[b]._f = src->b[b]._f;                            \
1187                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
1188         }
1189 #define copy_dev_field(_f, _msg, ...)                                   \
1190         copy_field(_f, "dev %u has wrong " _msg, dev, ##__VA_ARGS__)
1191 #define copy_fs_field(_f, _msg, ...)                                    \
1192         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
1193
1194         if (!metadata_only) {
1195                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
1196                 struct stripe *dst, *src;
1197
1198                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
1199                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
1200
1201                         if (dst->alive          != src->alive ||
1202                             dst->sectors        != src->sectors ||
1203                             dst->algorithm      != src->algorithm ||
1204                             dst->nr_blocks      != src->nr_blocks ||
1205                             dst->nr_redundant   != src->nr_redundant) {
1206                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
1207                                 ret = -EINVAL;
1208                                 goto fsck_err;
1209                         }
1210
1211                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
1212                                 copy_stripe_field(block_sectors[i],
1213                                                   "block_sectors[%u]", i);
1214
1215                         dst->blocks_nonempty = 0;
1216                         for (i = 0; i < dst->nr_blocks; i++)
1217                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
1218
1219                         genradix_iter_advance(&iter, &c->stripes[1]);
1220                 }
1221         }
1222
1223         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
1224                 bch2_fs_usage_acc_to_base(c, i);
1225
1226         for_each_member_device(ca, c, dev) {
1227                 struct bucket_array *dst = __bucket_array(ca, 0);
1228                 struct bucket_array *src = __bucket_array(ca, 1);
1229                 size_t b;
1230
1231                 for (b = 0; b < src->nbuckets; b++) {
1232                         copy_bucket_field(_mark.gen);
1233                         copy_bucket_field(_mark.data_type);
1234                         copy_bucket_field(_mark.stripe);
1235                         copy_bucket_field(_mark.dirty_sectors);
1236                         copy_bucket_field(_mark.cached_sectors);
1237                         copy_bucket_field(stripe_redundancy);
1238                         copy_bucket_field(stripe);
1239
1240                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
1241                 }
1242
1243                 {
1244                         struct bch_dev_usage *dst = ca->usage_base;
1245                         struct bch_dev_usage *src = (void *)
1246                                 bch2_acc_percpu_u64s((void *) ca->usage_gc,
1247                                                      dev_usage_u64s());
1248
1249                         copy_dev_field(buckets_ec,              "buckets_ec");
1250                         copy_dev_field(buckets_unavailable,     "buckets_unavailable");
1251
1252                         for (i = 0; i < BCH_DATA_NR; i++) {
1253                                 copy_dev_field(d[i].buckets,    "%s buckets", bch2_data_types[i]);
1254                                 copy_dev_field(d[i].sectors,    "%s sectors", bch2_data_types[i]);
1255                                 copy_dev_field(d[i].fragmented, "%s fragmented", bch2_data_types[i]);
1256                         }
1257                 }
1258         };
1259
1260         {
1261                 unsigned nr = fs_usage_u64s(c);
1262                 struct bch_fs_usage *dst = c->usage_base;
1263                 struct bch_fs_usage *src = (void *)
1264                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
1265
1266                 copy_fs_field(hidden,           "hidden");
1267                 copy_fs_field(btree,            "btree");
1268
1269                 if (!metadata_only) {
1270                         copy_fs_field(data,     "data");
1271                         copy_fs_field(cached,   "cached");
1272                         copy_fs_field(reserved, "reserved");
1273                         copy_fs_field(nr_inodes,"nr_inodes");
1274
1275                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
1276                                 copy_fs_field(persistent_reserved[i],
1277                                               "persistent_reserved[%i]", i);
1278                 }
1279
1280                 for (i = 0; i < c->replicas.nr; i++) {
1281                         struct bch_replicas_entry *e =
1282                                 cpu_replicas_entry(&c->replicas, i);
1283                         char buf[80];
1284
1285                         if (metadata_only &&
1286                             (e->data_type == BCH_DATA_user ||
1287                              e->data_type == BCH_DATA_cached))
1288                                 continue;
1289
1290                         bch2_replicas_entry_to_text(&PBUF(buf), e);
1291
1292                         copy_fs_field(replicas[i], "%s", buf);
1293                 }
1294         }
1295
1296 #undef copy_fs_field
1297 #undef copy_dev_field
1298 #undef copy_bucket_field
1299 #undef copy_stripe_field
1300 #undef copy_field
1301 fsck_err:
1302         if (ca)
1303                 percpu_ref_put(&ca->ref);
1304         if (ret)
1305                 bch_err(c, "%s: ret %i", __func__, ret);
1306         return ret;
1307 }
1308
1309 static int bch2_gc_start(struct bch_fs *c,
1310                          bool metadata_only)
1311 {
1312         struct bch_dev *ca = NULL;
1313         unsigned i;
1314         int ret;
1315
1316         BUG_ON(c->usage_gc);
1317
1318         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
1319                                          sizeof(u64), GFP_KERNEL);
1320         if (!c->usage_gc) {
1321                 bch_err(c, "error allocating c->usage_gc");
1322                 return -ENOMEM;
1323         }
1324
1325         for_each_member_device(ca, c, i) {
1326                 BUG_ON(ca->buckets[1]);
1327                 BUG_ON(ca->usage_gc);
1328
1329                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
1330                                 ca->mi.nbuckets * sizeof(struct bucket),
1331                                 GFP_KERNEL|__GFP_ZERO);
1332                 if (!ca->buckets[1]) {
1333                         percpu_ref_put(&ca->ref);
1334                         bch_err(c, "error allocating ca->buckets[gc]");
1335                         return -ENOMEM;
1336                 }
1337
1338                 ca->usage_gc = alloc_percpu(struct bch_dev_usage);
1339                 if (!ca->usage_gc) {
1340                         bch_err(c, "error allocating ca->usage_gc");
1341                         percpu_ref_put(&ca->ref);
1342                         return -ENOMEM;
1343                 }
1344         }
1345
1346         ret = bch2_ec_mem_alloc(c, true);
1347         if (ret) {
1348                 bch_err(c, "error allocating ec gc mem");
1349                 return ret;
1350         }
1351
1352         percpu_down_write(&c->mark_lock);
1353
1354         /*
1355          * indicate to stripe code that we need to allocate for the gc stripes
1356          * radix tree, too
1357          */
1358         gc_pos_set(c, gc_phase(GC_PHASE_START));
1359
1360         for_each_member_device(ca, c, i) {
1361                 struct bucket_array *dst = __bucket_array(ca, 1);
1362                 struct bucket_array *src = __bucket_array(ca, 0);
1363                 size_t b;
1364
1365                 dst->first_bucket       = src->first_bucket;
1366                 dst->nbuckets           = src->nbuckets;
1367
1368                 for (b = 0; b < src->nbuckets; b++) {
1369                         struct bucket *d = &dst->b[b];
1370                         struct bucket *s = &src->b[b];
1371
1372                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
1373                         d->gen_valid = s->gen_valid;
1374
1375                         if (metadata_only &&
1376                             (s->mark.data_type == BCH_DATA_user ||
1377                              s->mark.data_type == BCH_DATA_cached))
1378                                 d->_mark = s->mark;
1379                 }
1380         };
1381
1382         percpu_up_write(&c->mark_lock);
1383
1384         return 0;
1385 }
1386
1387 static int bch2_gc_reflink_done_initial_fn(struct btree_trans *trans,
1388                                            struct bkey_s_c k)
1389 {
1390         struct bch_fs *c = trans->c;
1391         struct reflink_gc *r;
1392         const __le64 *refcount = bkey_refcount_c(k);
1393         char buf[200];
1394         int ret = 0;
1395
1396         if (!refcount)
1397                 return 0;
1398
1399         r = genradix_ptr(&c->reflink_gc_table, c->reflink_gc_idx++);
1400         if (!r)
1401                 return -ENOMEM;
1402
1403         if (!r ||
1404             r->offset != k.k->p.offset ||
1405             r->size != k.k->size) {
1406                 bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1407                 return -EINVAL;
1408         }
1409
1410         if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1411                         "reflink key has wrong refcount:\n"
1412                         "  %s\n"
1413                         "  should be %u",
1414                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1415                         r->refcount)) {
1416                 struct bkey_i *new;
1417
1418                 new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1419                 if (!new) {
1420                         ret = -ENOMEM;
1421                         goto fsck_err;
1422                 }
1423
1424                 bkey_reassemble(new, k);
1425
1426                 if (!r->refcount) {
1427                         new->k.type = KEY_TYPE_deleted;
1428                         new->k.size = 0;
1429                 } else {
1430                         *bkey_refcount(new) = cpu_to_le64(r->refcount);
1431                 }
1432
1433                 ret = bch2_journal_key_insert(c, BTREE_ID_reflink, 0, new);
1434                 if (ret)
1435                         kfree(new);
1436         }
1437 fsck_err:
1438         return ret;
1439 }
1440
1441 static int bch2_gc_reflink_done(struct bch_fs *c, bool initial,
1442                                 bool metadata_only)
1443 {
1444         struct btree_trans trans;
1445         struct btree_iter iter;
1446         struct bkey_s_c k;
1447         struct reflink_gc *r;
1448         size_t idx = 0;
1449         char buf[200];
1450         int ret = 0;
1451
1452         if (metadata_only)
1453                 return 0;
1454
1455         bch2_trans_init(&trans, c, 0, 0);
1456
1457         if (initial) {
1458                 c->reflink_gc_idx = 0;
1459
1460                 ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_reflink,
1461                                 bch2_gc_reflink_done_initial_fn);
1462                 goto out;
1463         }
1464
1465         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1466                            BTREE_ITER_PREFETCH, k, ret) {
1467                 const __le64 *refcount = bkey_refcount_c(k);
1468
1469                 if (!refcount)
1470                         continue;
1471
1472                 r = genradix_ptr(&c->reflink_gc_table, idx);
1473                 if (!r ||
1474                     r->offset != k.k->p.offset ||
1475                     r->size != k.k->size) {
1476                         bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
1477                         ret = -EINVAL;
1478                         break;
1479                 }
1480
1481                 if (fsck_err_on(r->refcount != le64_to_cpu(*refcount), c,
1482                                 "reflink key has wrong refcount:\n"
1483                                 "  %s\n"
1484                                 "  should be %u",
1485                                 (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf),
1486                                 r->refcount)) {
1487                         struct bkey_i *new;
1488
1489                         new = kmalloc(bkey_bytes(k.k), GFP_KERNEL);
1490                         if (!new) {
1491                                 ret = -ENOMEM;
1492                                 break;
1493                         }
1494
1495                         bkey_reassemble(new, k);
1496
1497                         if (!r->refcount)
1498                                 new->k.type = KEY_TYPE_deleted;
1499                         else
1500                                 *bkey_refcount(new) = cpu_to_le64(r->refcount);
1501
1502                         ret = __bch2_trans_do(&trans, NULL, NULL, 0,
1503                                         __bch2_btree_insert(&trans, BTREE_ID_reflink, new));
1504                         kfree(new);
1505
1506                         if (ret)
1507                                 break;
1508                 }
1509         }
1510 fsck_err:
1511         bch2_trans_iter_exit(&trans, &iter);
1512 out:
1513         genradix_free(&c->reflink_gc_table);
1514         c->reflink_gc_nr = 0;
1515         bch2_trans_exit(&trans);
1516         return ret;
1517 }
1518
1519 static int bch2_gc_reflink_start_initial_fn(struct btree_trans *trans,
1520                                             struct bkey_s_c k)
1521 {
1522
1523         struct bch_fs *c = trans->c;
1524         struct reflink_gc *r;
1525         const __le64 *refcount = bkey_refcount_c(k);
1526
1527         if (!refcount)
1528                 return 0;
1529
1530         r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1531                                GFP_KERNEL);
1532         if (!r)
1533                 return -ENOMEM;
1534
1535         r->offset       = k.k->p.offset;
1536         r->size         = k.k->size;
1537         r->refcount     = 0;
1538         return 0;
1539 }
1540
1541 static int bch2_gc_reflink_start(struct bch_fs *c, bool initial,
1542                                  bool metadata_only)
1543 {
1544         struct btree_trans trans;
1545         struct btree_iter iter;
1546         struct bkey_s_c k;
1547         struct reflink_gc *r;
1548         int ret = 0;
1549
1550         if (metadata_only)
1551                 return 0;
1552
1553         bch2_trans_init(&trans, c, 0, 0);
1554         genradix_free(&c->reflink_gc_table);
1555         c->reflink_gc_nr = 0;
1556
1557         if (initial) {
1558                 ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_reflink,
1559                                                 bch2_gc_reflink_start_initial_fn);
1560                 goto out;
1561         }
1562
1563         for_each_btree_key(&trans, iter, BTREE_ID_reflink, POS_MIN,
1564                            BTREE_ITER_PREFETCH, k, ret) {
1565                 const __le64 *refcount = bkey_refcount_c(k);
1566
1567                 if (!refcount)
1568                         continue;
1569
1570                 r = genradix_ptr_alloc(&c->reflink_gc_table, c->reflink_gc_nr++,
1571                                        GFP_KERNEL);
1572                 if (!r) {
1573                         ret = -ENOMEM;
1574                         break;
1575                 }
1576
1577                 r->offset       = k.k->p.offset;
1578                 r->size         = k.k->size;
1579                 r->refcount     = 0;
1580         }
1581         bch2_trans_iter_exit(&trans, &iter);
1582 out:
1583         bch2_trans_exit(&trans);
1584         return ret;
1585 }
1586
1587 /**
1588  * bch2_gc - walk _all_ references to buckets, and recompute them:
1589  *
1590  * Order matters here:
1591  *  - Concurrent GC relies on the fact that we have a total ordering for
1592  *    everything that GC walks - see  gc_will_visit_node(),
1593  *    gc_will_visit_root()
1594  *
1595  *  - also, references move around in the course of index updates and
1596  *    various other crap: everything needs to agree on the ordering
1597  *    references are allowed to move around in - e.g., we're allowed to
1598  *    start with a reference owned by an open_bucket (the allocator) and
1599  *    move it to the btree, but not the reverse.
1600  *
1601  *    This is necessary to ensure that gc doesn't miss references that
1602  *    move around - if references move backwards in the ordering GC
1603  *    uses, GC could skip past them
1604  */
1605 int bch2_gc(struct bch_fs *c, bool initial, bool metadata_only)
1606 {
1607         struct bch_dev *ca;
1608         u64 start_time = local_clock();
1609         unsigned i, iter = 0;
1610         int ret;
1611
1612         lockdep_assert_held(&c->state_lock);
1613         trace_gc_start(c);
1614
1615         down_write(&c->gc_lock);
1616
1617         /* flush interior btree updates: */
1618         closure_wait_event(&c->btree_interior_update_wait,
1619                            !bch2_btree_interior_updates_nr_pending(c));
1620 again:
1621         ret   = bch2_gc_start(c, metadata_only) ?:
1622                 bch2_gc_reflink_start(c, initial, metadata_only);
1623         if (ret)
1624                 goto out;
1625
1626         bch2_mark_superblocks(c);
1627
1628         if (BCH_SB_HAS_TOPOLOGY_ERRORS(c->disk_sb.sb) &&
1629             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags) &&
1630             c->opts.fix_errors != FSCK_OPT_NO) {
1631                 bch_info(c, "starting topology repair pass");
1632                 ret = bch2_repair_topology(c);
1633                 if (ret)
1634                         goto out;
1635                 bch_info(c, "topology repair pass done");
1636
1637                 set_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags);
1638         }
1639
1640         ret = bch2_gc_btrees(c, initial, metadata_only);
1641
1642         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR &&
1643             !test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags) &&
1644             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {
1645                 set_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1646                 ret = 0;
1647         }
1648
1649         if (ret == FSCK_ERR_START_TOPOLOGY_REPAIR)
1650                 ret = FSCK_ERR_EXIT;
1651
1652         if (ret)
1653                 goto out;
1654
1655 #if 0
1656         bch2_mark_pending_btree_node_frees(c);
1657 #endif
1658         c->gc_count++;
1659
1660         if (test_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags) ||
1661             (!iter && bch2_test_restart_gc)) {
1662                 /*
1663                  * XXX: make sure gens we fixed got saved
1664                  */
1665                 if (iter++ <= 2) {
1666                         bch_info(c, "Second GC pass needed, restarting:");
1667                         clear_bit(BCH_FS_NEED_ANOTHER_GC, &c->flags);
1668                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1669
1670                         percpu_down_write(&c->mark_lock);
1671                         bch2_gc_free(c);
1672                         percpu_up_write(&c->mark_lock);
1673                         /* flush fsck errors, reset counters */
1674                         bch2_flush_fsck_errs(c);
1675
1676                         goto again;
1677                 }
1678
1679                 bch_info(c, "Unable to fix bucket gens, looping");
1680                 ret = -EINVAL;
1681         }
1682 out:
1683         if (!ret) {
1684                 bch2_journal_block(&c->journal);
1685
1686                 percpu_down_write(&c->mark_lock);
1687                 ret   = bch2_gc_reflink_done(c, initial, metadata_only) ?:
1688                         bch2_gc_done(c, initial, metadata_only);
1689
1690                 bch2_journal_unblock(&c->journal);
1691         } else {
1692                 percpu_down_write(&c->mark_lock);
1693         }
1694
1695         /* Indicates that gc is no longer in progress: */
1696         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
1697
1698         bch2_gc_free(c);
1699         percpu_up_write(&c->mark_lock);
1700
1701         up_write(&c->gc_lock);
1702
1703         trace_gc_end(c);
1704         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
1705
1706         /*
1707          * Wake up allocator in case it was waiting for buckets
1708          * because of not being able to inc gens
1709          */
1710         for_each_member_device(ca, c, i)
1711                 bch2_wake_allocator(ca);
1712
1713         /*
1714          * At startup, allocations can happen directly instead of via the
1715          * allocator thread - issue wakeup in case they blocked on gc_lock:
1716          */
1717         closure_wake_up(&c->freelist_wait);
1718         return ret;
1719 }
1720
1721 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
1722 {
1723         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1724         const struct bch_extent_ptr *ptr;
1725
1726         percpu_down_read(&c->mark_lock);
1727         bkey_for_each_ptr(ptrs, ptr) {
1728                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1729                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1730
1731                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
1732                         percpu_up_read(&c->mark_lock);
1733                         return true;
1734                 }
1735         }
1736
1737         bkey_for_each_ptr(ptrs, ptr) {
1738                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1739                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
1740
1741                 if (gen_after(g->gc_gen, ptr->gen))
1742                         g->gc_gen = ptr->gen;
1743         }
1744         percpu_up_read(&c->mark_lock);
1745
1746         return false;
1747 }
1748
1749 /*
1750  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
1751  * node pointers currently never have cached pointers that can become stale:
1752  */
1753 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
1754 {
1755         struct btree_trans trans;
1756         struct btree_iter iter;
1757         struct bkey_s_c k;
1758         struct bkey_buf sk;
1759         int ret = 0, commit_err = 0;
1760
1761         bch2_bkey_buf_init(&sk);
1762         bch2_trans_init(&trans, c, 0, 0);
1763
1764         bch2_trans_iter_init(&trans, &iter, btree_id, POS_MIN,
1765                              BTREE_ITER_PREFETCH|
1766                              BTREE_ITER_NOT_EXTENTS|
1767                              BTREE_ITER_ALL_SNAPSHOTS);
1768
1769         while ((bch2_trans_begin(&trans),
1770                 k = bch2_btree_iter_peek(&iter)).k) {
1771                 ret = bkey_err(k);
1772
1773                 if (ret == -EINTR)
1774                         continue;
1775                 if (ret)
1776                         break;
1777
1778                 c->gc_gens_pos = iter.pos;
1779
1780                 if (gc_btree_gens_key(c, k) && !commit_err) {
1781                         bch2_bkey_buf_reassemble(&sk, c, k);
1782                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
1783
1784                         commit_err =
1785                                 bch2_trans_update(&trans, &iter, sk.k, 0) ?:
1786                                 bch2_trans_commit(&trans, NULL, NULL,
1787                                                        BTREE_INSERT_NOWAIT|
1788                                                        BTREE_INSERT_NOFAIL);
1789                         if (commit_err == -EINTR) {
1790                                 commit_err = 0;
1791                                 continue;
1792                         }
1793                 }
1794
1795                 bch2_btree_iter_advance(&iter);
1796         }
1797         bch2_trans_iter_exit(&trans, &iter);
1798
1799         bch2_trans_exit(&trans);
1800         bch2_bkey_buf_exit(&sk, c);
1801
1802         return ret;
1803 }
1804
1805 int bch2_gc_gens(struct bch_fs *c)
1806 {
1807         struct bch_dev *ca;
1808         struct bucket_array *buckets;
1809         struct bucket *g;
1810         unsigned i;
1811         int ret;
1812
1813         /*
1814          * Ideally we would be using state_lock and not gc_lock here, but that
1815          * introduces a deadlock in the RO path - we currently take the state
1816          * lock at the start of going RO, thus the gc thread may get stuck:
1817          */
1818         down_read(&c->gc_lock);
1819
1820         for_each_member_device(ca, c, i) {
1821                 down_read(&ca->bucket_lock);
1822                 buckets = bucket_array(ca);
1823
1824                 for_each_bucket(g, buckets)
1825                         g->gc_gen = g->mark.gen;
1826                 up_read(&ca->bucket_lock);
1827         }
1828
1829         for (i = 0; i < BTREE_ID_NR; i++)
1830                 if ((1 << i) & BTREE_ID_HAS_PTRS) {
1831                         c->gc_gens_btree = i;
1832                         c->gc_gens_pos = POS_MIN;
1833                         ret = bch2_gc_btree_gens(c, i);
1834                         if (ret) {
1835                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
1836                                 goto err;
1837                         }
1838                 }
1839
1840         for_each_member_device(ca, c, i) {
1841                 down_read(&ca->bucket_lock);
1842                 buckets = bucket_array(ca);
1843
1844                 for_each_bucket(g, buckets)
1845                         g->oldest_gen = g->gc_gen;
1846                 up_read(&ca->bucket_lock);
1847         }
1848
1849         c->gc_gens_btree        = 0;
1850         c->gc_gens_pos          = POS_MIN;
1851
1852         c->gc_count++;
1853 err:
1854         up_read(&c->gc_lock);
1855         return ret;
1856 }
1857
1858 static int bch2_gc_thread(void *arg)
1859 {
1860         struct bch_fs *c = arg;
1861         struct io_clock *clock = &c->io_clock[WRITE];
1862         unsigned long last = atomic64_read(&clock->now);
1863         unsigned last_kick = atomic_read(&c->kick_gc);
1864         int ret;
1865
1866         set_freezable();
1867
1868         while (1) {
1869                 while (1) {
1870                         set_current_state(TASK_INTERRUPTIBLE);
1871
1872                         if (kthread_should_stop()) {
1873                                 __set_current_state(TASK_RUNNING);
1874                                 return 0;
1875                         }
1876
1877                         if (atomic_read(&c->kick_gc) != last_kick)
1878                                 break;
1879
1880                         if (c->btree_gc_periodic) {
1881                                 unsigned long next = last + c->capacity / 16;
1882
1883                                 if (atomic64_read(&clock->now) >= next)
1884                                         break;
1885
1886                                 bch2_io_clock_schedule_timeout(clock, next);
1887                         } else {
1888                                 schedule();
1889                         }
1890
1891                         try_to_freeze();
1892                 }
1893                 __set_current_state(TASK_RUNNING);
1894
1895                 last = atomic64_read(&clock->now);
1896                 last_kick = atomic_read(&c->kick_gc);
1897
1898                 /*
1899                  * Full gc is currently incompatible with btree key cache:
1900                  */
1901 #if 0
1902                 ret = bch2_gc(c, false, false);
1903 #else
1904                 ret = bch2_gc_gens(c);
1905 #endif
1906                 if (ret < 0)
1907                         bch_err(c, "btree gc failed: %i", ret);
1908
1909                 debug_check_no_locks_held();
1910         }
1911
1912         return 0;
1913 }
1914
1915 void bch2_gc_thread_stop(struct bch_fs *c)
1916 {
1917         struct task_struct *p;
1918
1919         p = c->gc_thread;
1920         c->gc_thread = NULL;
1921
1922         if (p) {
1923                 kthread_stop(p);
1924                 put_task_struct(p);
1925         }
1926 }
1927
1928 int bch2_gc_thread_start(struct bch_fs *c)
1929 {
1930         struct task_struct *p;
1931
1932         if (c->gc_thread)
1933                 return 0;
1934
1935         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1936         if (IS_ERR(p)) {
1937                 bch_err(c, "error creating gc thread: %li", PTR_ERR(p));
1938                 return PTR_ERR(p);
1939         }
1940
1941         get_task_struct(p);
1942         c->gc_thread = p;
1943         wake_up_process(p);
1944         return 0;
1945 }