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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_key_cache.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_gc.h"
  11 #include "buckets.h"
  12 #include "buckets_waiting_for_journal.h"
  13 #include "clock.h"
  14 #include "debug.h"
  15 #include "ec.h"
  16 #include "error.h"
  17 #include "lru.h"
  18 #include "recovery.h"
  19 #include "varint.h"
  20
  21 #include <linux/kthread.h>
  22 #include <linux/math64.h>
  23 #include <linux/random.h>
  24 #include <linux/rculist.h>
  25 #include <linux/rcupdate.h>
  26 #include <linux/sched/task.h>
  27 #include <linux/sort.h>
  28 #include <trace/events/bcachefs.h>
  29
  30 /* Persistent alloc info: */
  31
  32 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  33 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  34         BCH_ALLOC_FIELDS_V1()
  35 #undef x
  36 };
  37
  38 struct bkey_alloc_unpacked {
  39         u64             journal_seq;
  40         u64             bucket;
  41         u8              dev;
  42         u8              gen;
  43         u8              oldest_gen;
  44         u8              data_type;
  45         bool            need_discard:1;
  46         bool            need_inc_gen:1;
  47 #define x(_name, _bits) u##_bits _name;
  48         BCH_ALLOC_FIELDS_V2()
  49 #undef  x
  50 };
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  83                                       unsigned field, u64 v)
  84 {
  85         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  86
  87         if (!v)
  88                 return;
  89
  90         a->v.fields |= 1 << field;
  91
  92         switch (bytes) {
  93         case 1:
  94                 *((u8 *) *p) = v;
  95                 break;
  96         case 2:
  97                 *((__le16 *) *p) = cpu_to_le16(v);
  98                 break;
  99         case 4:
 100                 *((__le32 *) *p) = cpu_to_le32(v);
 101                 break;
 102         case 8:
 103                 *((__le64 *) *p) = cpu_to_le64(v);
 104                 break;
 105         default:
 106                 BUG();
 107         }
 108
 109         *p += bytes;
 110 }
 111
 112 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 113                                  struct bkey_s_c k)
 114 {
 115         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 116         const void *d = in->data;
 117         unsigned idx = 0;
 118
 119         out->gen = in->gen;
 120
 121 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 122         BCH_ALLOC_FIELDS_V1()
 123 #undef  x
 124 }
 125
 126 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 127                                 struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 130         const u8 *in = a.v->data;
 131         const u8 *end = bkey_val_end(a);
 132         unsigned fieldnr = 0;
 133         int ret;
 134         u64 v;
 135
 136         out->gen        = a.v->gen;
 137         out->oldest_gen = a.v->oldest_gen;
 138         out->data_type  = a.v->data_type;
 139
 140 #define x(_name, _bits)                                                 \
 141         if (fieldnr < a.v->nr_fields) {                                 \
 142                 ret = bch2_varint_decode_fast(in, end, &v);             \
 143                 if (ret < 0)                                            \
 144                         return ret;                                     \
 145                 in += ret;                                              \
 146         } else {                                                        \
 147                 v = 0;                                                  \
 148         }                                                               \
 149         out->_name = v;                                                 \
 150         if (v != out->_name)                                            \
 151                 return -1;                                              \
 152         fieldnr++;
 153
 154         BCH_ALLOC_FIELDS_V2()
 155 #undef  x
 156         return 0;
 157 }
 158
 159 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 160                                 struct bkey_s_c k)
 161 {
 162         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 163         const u8 *in = a.v->data;
 164         const u8 *end = bkey_val_end(a);
 165         unsigned fieldnr = 0;
 166         int ret;
 167         u64 v;
 168
 169         out->gen        = a.v->gen;
 170         out->oldest_gen = a.v->oldest_gen;
 171         out->data_type  = a.v->data_type;
 172         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 173         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 174         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 175
 176 #define x(_name, _bits)                                                 \
 177         if (fieldnr < a.v->nr_fields) {                                 \
 178                 ret = bch2_varint_decode_fast(in, end, &v);             \
 179                 if (ret < 0)                                            \
 180                         return ret;                                     \
 181                 in += ret;                                              \
 182         } else {                                                        \
 183                 v = 0;                                                  \
 184         }                                                               \
 185         out->_name = v;                                                 \
 186         if (v != out->_name)                                            \
 187                 return -1;                                              \
 188         fieldnr++;
 189
 190         BCH_ALLOC_FIELDS_V2()
 191 #undef  x
 192         return 0;
 193 }
 194
 195 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 196 {
 197         struct bkey_alloc_unpacked ret = {
 198                 .dev    = k.k->p.inode,
 199                 .bucket = k.k->p.offset,
 200                 .gen    = 0,
 201         };
 202
 203         switch (k.k->type) {
 204         case KEY_TYPE_alloc:
 205                 bch2_alloc_unpack_v1(&ret, k);
 206                 break;
 207         case KEY_TYPE_alloc_v2:
 208                 bch2_alloc_unpack_v2(&ret, k);
 209                 break;
 210         case KEY_TYPE_alloc_v3:
 211                 bch2_alloc_unpack_v3(&ret, k);
 212                 break;
 213         }
 214
 215         return ret;
 216 }
 217
 218 void bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 219 {
 220         if (k.k->type == KEY_TYPE_alloc_v4) {
 221                 *out = *bkey_s_c_to_alloc_v4(k).v;
 222         } else {
 223                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 224
 225                 *out = (struct bch_alloc_v4) {
 226                         .journal_seq            = u.journal_seq,
 227                         .flags                  = u.need_discard,
 228                         .gen                    = u.gen,
 229                         .oldest_gen             = u.oldest_gen,
 230                         .data_type              = u.data_type,
 231                         .stripe_redundancy      = u.stripe_redundancy,
 232                         .dirty_sectors          = u.dirty_sectors,
 233                         .cached_sectors         = u.cached_sectors,
 234                         .io_time[READ]          = u.read_time,
 235                         .io_time[WRITE]         = u.write_time,
 236                         .stripe                 = u.stripe,
 237                 };
 238         }
 239 }
 240
 241 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 242 {
 243         struct bkey_i_alloc_v4 *ret;
 244
 245         if (k.k->type == KEY_TYPE_alloc_v4) {
 246                 ret = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 247                 if (!IS_ERR(ret))
 248                         bkey_reassemble(&ret->k_i, k);
 249         } else {
 250                 ret = bch2_trans_kmalloc(trans, sizeof(*ret));
 251                 if (!IS_ERR(ret)) {
 252                         bkey_alloc_v4_init(&ret->k_i);
 253                         ret->k.p = k.k->p;
 254                         bch2_alloc_to_v4(k, &ret->v);
 255                 }
 256         }
 257         return ret;
 258 }
 259
 260 struct bkey_i_alloc_v4 *
 261 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 262                               struct bpos pos)
 263 {
 264         struct bkey_s_c k;
 265         struct bkey_i_alloc_v4 *a;
 266         int ret;
 267
 268         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 269                              BTREE_ITER_WITH_UPDATES|
 270                              BTREE_ITER_CACHED|
 271                              BTREE_ITER_INTENT);
 272         k = bch2_btree_iter_peek_slot(iter);
 273         ret = bkey_err(k);
 274         if (ret) {
 275                 bch2_trans_iter_exit(trans, iter);
 276                 return ERR_PTR(ret);
 277         }
 278
 279         a = bch2_alloc_to_v4_mut(trans, k);
 280         if (IS_ERR(a))
 281                 bch2_trans_iter_exit(trans, iter);
 282         return a;
 283 }
 284
 285 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 286 {
 287         unsigned i, bytes = offsetof(struct bch_alloc, data);
 288
 289         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 290                 if (a->fields & (1 << i))
 291                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 292
 293         return DIV_ROUND_UP(bytes, sizeof(u64));
 294 }
 295
 296 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 297                           int rw, struct printbuf *err)
 298 {
 299         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 300
 301         /* allow for unknown fields */
 302         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 303                 pr_buf(err, "incorrect value size (%zu < %u)",
 304                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 305                 return -EINVAL;
 306         }
 307
 308         return 0;
 309 }
 310
 311 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 312                           int rw, struct printbuf *err)
 313 {
 314         struct bkey_alloc_unpacked u;
 315
 316         if (bch2_alloc_unpack_v2(&u, k)) {
 317                 pr_buf(err, "unpack error");
 318                 return -EINVAL;
 319         }
 320
 321         return 0;
 322 }
 323
 324 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 325                           int rw, struct printbuf *err)
 326 {
 327         struct bkey_alloc_unpacked u;
 328
 329         if (bch2_alloc_unpack_v3(&u, k)) {
 330                 pr_buf(err, "unpack error");
 331                 return -EINVAL;
 332         }
 333
 334         return 0;
 335 }
 336
 337 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 338                           int rw, struct printbuf *err)
 339 {
 340         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 341
 342         if (bkey_val_bytes(k.k) != sizeof(struct bch_alloc_v4)) {
 343                 pr_buf(err, "bad val size (%zu != %zu)",
 344                        bkey_val_bytes(k.k), sizeof(struct bch_alloc_v4));
 345                 return -EINVAL;
 346         }
 347
 348         if (rw == WRITE) {
 349                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 350                         pr_buf(err, "invalid data type (got %u should be %u)",
 351                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 352                         return -EINVAL;
 353                 }
 354
 355                 switch (a.v->data_type) {
 356                 case BCH_DATA_free:
 357                 case BCH_DATA_need_gc_gens:
 358                 case BCH_DATA_need_discard:
 359                         if (a.v->dirty_sectors ||
 360                             a.v->cached_sectors ||
 361                             a.v->stripe) {
 362                                 pr_buf(err, "empty data type free but have data");
 363                                 return -EINVAL;
 364                         }
 365                         break;
 366                 case BCH_DATA_sb:
 367                 case BCH_DATA_journal:
 368                 case BCH_DATA_btree:
 369                 case BCH_DATA_user:
 370                 case BCH_DATA_parity:
 371                         if (!a.v->dirty_sectors) {
 372                                 pr_buf(err, "data_type %s but dirty_sectors==0",
 373                                        bch2_data_types[a.v->data_type]);
 374                                 return -EINVAL;
 375                         }
 376                         break;
 377                 case BCH_DATA_cached:
 378                         if (!a.v->cached_sectors ||
 379                             a.v->dirty_sectors ||
 380                             a.v->stripe) {
 381                                 pr_buf(err, "data type inconsistency");
 382                                 return -EINVAL;
 383                         }
 384
 385                         if (!a.v->io_time[READ]) {
 386                                 pr_buf(err, "cached bucket with read_time == 0");
 387                                 return -EINVAL;
 388                         }
 389                         break;
 390                 case BCH_DATA_stripe:
 391                         if (!a.v->stripe) {
 392                                 pr_buf(err, "data_type %s but stripe==0",
 393                                        bch2_data_types[a.v->data_type]);
 394                                 return -EINVAL;
 395                         }
 396                         break;
 397                 }
 398         }
 399
 400         return 0;
 401 }
 402
 403 void bch2_alloc_v4_swab(struct bkey_s k)
 404 {
 405         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 406
 407         a->journal_seq          = swab64(a->journal_seq);
 408         a->flags                = swab32(a->flags);
 409         a->dirty_sectors        = swab32(a->dirty_sectors);
 410         a->cached_sectors       = swab32(a->cached_sectors);
 411         a->io_time[0]           = swab64(a->io_time[0]);
 412         a->io_time[1]           = swab64(a->io_time[1]);
 413         a->stripe               = swab32(a->stripe);
 414         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 415 }
 416
 417 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 418 {
 419         struct bch_alloc_v4 a;
 420
 421         bch2_alloc_to_v4(k, &a);
 422
 423         pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu need_discard %llu need_inc_gen %llu",
 424                a.gen, a.oldest_gen, bch2_data_types[a.data_type],
 425                a.journal_seq,
 426                BCH_ALLOC_V4_NEED_DISCARD(&a),
 427                BCH_ALLOC_V4_NEED_INC_GEN(&a));
 428         pr_buf(out, " dirty_sectors %u",        a.dirty_sectors);
 429         pr_buf(out, " cached_sectors %u",       a.cached_sectors);
 430         pr_buf(out, " stripe %u",               a.stripe);
 431         pr_buf(out, " stripe_redundancy %u",    a.stripe_redundancy);
 432         pr_buf(out, " read_time %llu",          a.io_time[READ]);
 433         pr_buf(out, " write_time %llu",         a.io_time[WRITE]);
 434 }
 435
 436 int bch2_alloc_read(struct bch_fs *c)
 437 {
 438         struct btree_trans trans;
 439         struct btree_iter iter;
 440         struct bkey_s_c k;
 441         struct bch_alloc_v4 a;
 442         struct bch_dev *ca;
 443         int ret;
 444
 445         bch2_trans_init(&trans, c, 0, 0);
 446
 447         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 448                            BTREE_ITER_PREFETCH, k, ret) {
 449                 /*
 450                  * Not a fsck error because this is checked/repaired by
 451                  * bch2_check_alloc_key() which runs later:
 452                  */
 453                 if (!bch2_dev_bucket_exists(c, k.k->p))
 454                         continue;
 455
 456                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 457                 bch2_alloc_to_v4(k, &a);
 458
 459                 *bucket_gen(ca, k.k->p.offset) = a.gen;
 460         }
 461         bch2_trans_iter_exit(&trans, &iter);
 462
 463         bch2_trans_exit(&trans);
 464
 465         if (ret)
 466                 bch_err(c, "error reading alloc info: %i", ret);
 467
 468         return ret;
 469 }
 470
 471 /* Free space/discard btree: */
 472
 473 static int bch2_bucket_do_index(struct btree_trans *trans,
 474                                 struct bkey_s_c alloc_k,
 475                                 const struct bch_alloc_v4 *a,
 476                                 bool set)
 477 {
 478         struct bch_fs *c = trans->c;
 479         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 480         struct btree_iter iter;
 481         struct bkey_s_c old;
 482         struct bkey_i *k;
 483         enum btree_id btree;
 484         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 485         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 486         struct printbuf buf = PRINTBUF;
 487         int ret;
 488
 489         if (a->data_type != BCH_DATA_free &&
 490             a->data_type != BCH_DATA_need_discard)
 491                 return 0;
 492
 493         k = bch2_trans_kmalloc(trans, sizeof(*k));
 494         if (IS_ERR(k))
 495                 return PTR_ERR(k);
 496
 497         bkey_init(&k->k);
 498         k->k.type = new_type;
 499
 500         switch (a->data_type) {
 501         case BCH_DATA_free:
 502                 btree = BTREE_ID_freespace;
 503                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 504                 bch2_key_resize(&k->k, 1);
 505                 break;
 506         case BCH_DATA_need_discard:
 507                 btree = BTREE_ID_need_discard;
 508                 k->k.p = alloc_k.k->p;
 509                 break;
 510         default:
 511                 return 0;
 512         }
 513
 514         bch2_trans_iter_init(trans, &iter, btree,
 515                              bkey_start_pos(&k->k),
 516                              BTREE_ITER_INTENT);
 517         old = bch2_btree_iter_peek_slot(&iter);
 518         ret = bkey_err(old);
 519         if (ret)
 520                 goto err;
 521
 522         if (ca->mi.freespace_initialized &&
 523             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 524                         "incorrect key when %s %s btree (got %s should be %s)\n"
 525                         "  for %s",
 526                         set ? "setting" : "clearing",
 527                         bch2_btree_ids[btree],
 528                         bch2_bkey_types[old.k->type],
 529                         bch2_bkey_types[old_type],
 530                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 531                 ret = -EIO;
 532                 goto err;
 533         }
 534
 535         ret = bch2_trans_update(trans, &iter, k, 0);
 536 err:
 537         bch2_trans_iter_exit(trans, &iter);
 538         printbuf_exit(&buf);
 539         return ret;
 540 }
 541
 542 int bch2_trans_mark_alloc(struct btree_trans *trans,
 543                           struct bkey_s_c old, struct bkey_i *new,
 544                           unsigned flags)
 545 {
 546         struct bch_fs *c = trans->c;
 547         struct bch_alloc_v4 old_a, *new_a;
 548         u64 old_lru, new_lru;
 549         int ret = 0;
 550
 551         /*
 552          * Deletion only happens in the device removal path, with
 553          * BTREE_TRIGGER_NORUN:
 554          */
 555         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 556
 557         bch2_alloc_to_v4(old, &old_a);
 558         new_a = &bkey_i_to_alloc_v4(new)->v;
 559
 560         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 561
 562         if (new_a->dirty_sectors > old_a.dirty_sectors ||
 563             new_a->cached_sectors > old_a.cached_sectors) {
 564                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 565                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 566                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 567                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 568         }
 569
 570         if (data_type_is_empty(new_a->data_type) &&
 571             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 572             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 573                 new_a->gen++;
 574                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 575         }
 576
 577         if (old_a.data_type != new_a->data_type ||
 578             (new_a->data_type == BCH_DATA_free &&
 579              alloc_freespace_genbits(old_a) != alloc_freespace_genbits(*new_a))) {
 580                 ret =   bch2_bucket_do_index(trans, old, &old_a, false) ?:
 581                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 582                 if (ret)
 583                         return ret;
 584         }
 585
 586         if (new_a->data_type == BCH_DATA_cached &&
 587             !new_a->io_time[READ])
 588                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 589
 590
 591         old_lru = alloc_lru_idx(old_a);
 592         new_lru = alloc_lru_idx(*new_a);
 593
 594         if (old_lru != new_lru) {
 595                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 596                                       old_lru, &new_lru, old);
 597                 if (ret)
 598                         return ret;
 599
 600                 if (new_a->data_type == BCH_DATA_cached)
 601                         new_a->io_time[READ] = new_lru;
 602         }
 603
 604         return 0;
 605 }
 606
 607 static int bch2_check_alloc_key(struct btree_trans *trans,
 608                                 struct btree_iter *alloc_iter)
 609 {
 610         struct bch_fs *c = trans->c;
 611         struct bch_dev *ca;
 612         struct btree_iter discard_iter, freespace_iter;
 613         struct bch_alloc_v4 a;
 614         unsigned discard_key_type, freespace_key_type;
 615         struct bkey_s_c alloc_k, k;
 616         struct printbuf buf = PRINTBUF;
 617         struct printbuf buf2 = PRINTBUF;
 618         int ret;
 619
 620         alloc_k = bch2_btree_iter_peek(alloc_iter);
 621         if (!alloc_k.k)
 622                 return 0;
 623
 624         ret = bkey_err(alloc_k);
 625         if (ret)
 626                 return ret;
 627
 628         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
 629                         "alloc key for invalid device:bucket %llu:%llu",
 630                         alloc_k.k->p.inode, alloc_k.k->p.offset))
 631                 return bch2_btree_delete_at(trans, alloc_iter, 0);
 632
 633         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 634         if (!ca->mi.freespace_initialized)
 635                 return 0;
 636
 637         bch2_alloc_to_v4(alloc_k, &a);
 638
 639         discard_key_type = a.data_type == BCH_DATA_need_discard
 640                 ? KEY_TYPE_set : 0;
 641         freespace_key_type = a.data_type == BCH_DATA_free
 642                 ? KEY_TYPE_set : 0;
 643
 644         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard,
 645                              alloc_k.k->p, 0);
 646         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace,
 647                              alloc_freespace_pos(alloc_k.k->p, a), 0);
 648
 649         k = bch2_btree_iter_peek_slot(&discard_iter);
 650         ret = bkey_err(k);
 651         if (ret)
 652                 goto err;
 653
 654         if (fsck_err_on(k.k->type != discard_key_type, c,
 655                         "incorrect key in need_discard btree (got %s should be %s)\n"
 656                         "  %s",
 657                         bch2_bkey_types[k.k->type],
 658                         bch2_bkey_types[discard_key_type],
 659                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 660                 struct bkey_i *update =
 661                         bch2_trans_kmalloc(trans, sizeof(*update));
 662
 663                 ret = PTR_ERR_OR_ZERO(update);
 664                 if (ret)
 665                         goto err;
 666
 667                 bkey_init(&update->k);
 668                 update->k.type  = discard_key_type;
 669                 update->k.p     = discard_iter.pos;
 670
 671                 ret = bch2_trans_update(trans, &discard_iter, update, 0);
 672                 if (ret)
 673                         goto err;
 674         }
 675
 676         k = bch2_btree_iter_peek_slot(&freespace_iter);
 677         ret = bkey_err(k);
 678         if (ret)
 679                 goto err;
 680
 681         if (fsck_err_on(k.k->type != freespace_key_type, c,
 682                         "incorrect key in freespace btree (got %s should be %s)\n"
 683                         "  %s",
 684                         bch2_bkey_types[k.k->type],
 685                         bch2_bkey_types[freespace_key_type],
 686                         (printbuf_reset(&buf),
 687                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 688                 struct bkey_i *update =
 689                         bch2_trans_kmalloc(trans, sizeof(*update));
 690
 691                 ret = PTR_ERR_OR_ZERO(update);
 692                 if (ret)
 693                         goto err;
 694
 695                 bkey_init(&update->k);
 696                 update->k.type  = freespace_key_type;
 697                 update->k.p     = freespace_iter.pos;
 698                 bch2_key_resize(&update->k, 1);
 699
 700                 ret = bch2_trans_update(trans, &freespace_iter, update, 0);
 701                 if (ret)
 702                         goto err;
 703         }
 704 err:
 705 fsck_err:
 706         bch2_trans_iter_exit(trans, &freespace_iter);
 707         bch2_trans_iter_exit(trans, &discard_iter);
 708         printbuf_exit(&buf2);
 709         printbuf_exit(&buf);
 710         return ret;
 711 }
 712
 713 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
 714                                             struct btree_iter *iter)
 715 {
 716         struct bch_fs *c = trans->c;
 717         struct btree_iter alloc_iter;
 718         struct bkey_s_c k, freespace_k;
 719         struct bch_alloc_v4 a;
 720         u64 genbits;
 721         struct bpos pos;
 722         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
 723                 ? BCH_DATA_need_discard
 724                 : BCH_DATA_free;
 725         struct printbuf buf = PRINTBUF;
 726         int ret;
 727
 728         freespace_k = bch2_btree_iter_peek(iter);
 729         if (!freespace_k.k)
 730                 return 1;
 731
 732         ret = bkey_err(freespace_k);
 733         if (ret)
 734                 return ret;
 735
 736         pos = iter->pos;
 737         pos.offset &= ~(~0ULL << 56);
 738         genbits = iter->pos.offset & (~0ULL << 56);
 739
 740         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
 741
 742         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
 743                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
 744                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
 745                 goto delete;
 746
 747         k = bch2_btree_iter_peek_slot(&alloc_iter);
 748         ret = bkey_err(k);
 749         if (ret)
 750                 goto err;
 751
 752         bch2_alloc_to_v4(k, &a);
 753
 754         if (fsck_err_on(a.data_type != state ||
 755                         (state == BCH_DATA_free &&
 756                          genbits != alloc_freespace_genbits(a)), c,
 757                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
 758                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 759                         bch2_btree_ids[iter->btree_id],
 760                         a.data_type == state,
 761                         genbits >> 56, alloc_freespace_genbits(a) >> 56))
 762                 goto delete;
 763 out:
 764 err:
 765 fsck_err:
 766         bch2_trans_iter_exit(trans, &alloc_iter);
 767         printbuf_exit(&buf);
 768         return ret;
 769 delete:
 770         ret = bch2_btree_delete_extent_at(trans, iter,
 771                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
 772         goto out;
 773 }
 774
 775 int bch2_check_alloc_info(struct bch_fs *c)
 776 {
 777         struct btree_trans trans;
 778         struct btree_iter iter;
 779         struct bkey_s_c k;
 780         int ret = 0;
 781
 782         bch2_trans_init(&trans, c, 0, 0);
 783
 784         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 785                            BTREE_ITER_PREFETCH, k, ret) {
 786                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 787                         bch2_check_alloc_key(&trans, &iter));
 788                 if (ret)
 789                         break;
 790         }
 791         bch2_trans_iter_exit(&trans, &iter);
 792
 793         if (ret)
 794                 goto err;
 795
 796         bch2_trans_iter_init(&trans, &iter, BTREE_ID_need_discard, POS_MIN,
 797                              BTREE_ITER_PREFETCH);
 798         while (1) {
 799                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 800                         bch2_check_discard_freespace_key(&trans, &iter));
 801                 if (ret)
 802                         break;
 803
 804                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 805         }
 806         bch2_trans_iter_exit(&trans, &iter);
 807
 808         if (ret)
 809                 goto err;
 810
 811         bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
 812                              BTREE_ITER_PREFETCH);
 813         while (1) {
 814                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 815                         bch2_check_discard_freespace_key(&trans, &iter));
 816                 if (ret)
 817                         break;
 818
 819                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 820         }
 821         bch2_trans_iter_exit(&trans, &iter);
 822 err:
 823         bch2_trans_exit(&trans);
 824         return ret < 0 ? ret : 0;
 825 }
 826
 827 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
 828                                        struct btree_iter *alloc_iter)
 829 {
 830         struct bch_fs *c = trans->c;
 831         struct btree_iter lru_iter;
 832         struct bch_alloc_v4 a;
 833         struct bkey_s_c alloc_k, k;
 834         struct printbuf buf = PRINTBUF;
 835         struct printbuf buf2 = PRINTBUF;
 836         int ret;
 837
 838         alloc_k = bch2_btree_iter_peek(alloc_iter);
 839         if (!alloc_k.k)
 840                 return 0;
 841
 842         ret = bkey_err(alloc_k);
 843         if (ret)
 844                 return ret;
 845
 846         bch2_alloc_to_v4(alloc_k, &a);
 847
 848         if (a.data_type != BCH_DATA_cached)
 849                 return 0;
 850
 851         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 852                              POS(alloc_k.k->p.inode, a.io_time[READ]), 0);
 853
 854         k = bch2_btree_iter_peek_slot(&lru_iter);
 855         ret = bkey_err(k);
 856         if (ret)
 857                 goto err;
 858
 859         if (fsck_err_on(!a.io_time[READ], c,
 860                         "cached bucket with read_time 0\n"
 861                         "  %s",
 862                 (printbuf_reset(&buf),
 863                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
 864             fsck_err_on(k.k->type != KEY_TYPE_lru ||
 865                         le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
 866                         "incorrect/missing lru entry\n"
 867                         "  %s\n"
 868                         "  %s",
 869                         (printbuf_reset(&buf),
 870                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
 871                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
 872                 u64 read_time = a.io_time[READ];
 873
 874                 if (!a.io_time[READ])
 875                         a.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
 876
 877                 ret = bch2_lru_set(trans,
 878                                    alloc_k.k->p.inode,
 879                                    alloc_k.k->p.offset,
 880                                    &a.io_time[READ]);
 881                 if (ret)
 882                         goto err;
 883
 884                 if (a.io_time[READ] != read_time) {
 885                         struct bkey_i_alloc_v4 *a_mut =
 886                                 bch2_alloc_to_v4_mut(trans, alloc_k);
 887                         ret = PTR_ERR_OR_ZERO(a_mut);
 888                         if (ret)
 889                                 goto err;
 890
 891                         a_mut->v.io_time[READ] = a.io_time[READ];
 892                         ret = bch2_trans_update(trans, alloc_iter,
 893                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
 894                         if (ret)
 895                                 goto err;
 896                 }
 897         }
 898 err:
 899 fsck_err:
 900         bch2_trans_iter_exit(trans, &lru_iter);
 901         printbuf_exit(&buf2);
 902         printbuf_exit(&buf);
 903         return ret;
 904 }
 905
 906 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
 907 {
 908         struct btree_trans trans;
 909         struct btree_iter iter;
 910         struct bkey_s_c k;
 911         int ret = 0;
 912
 913         bch2_trans_init(&trans, c, 0, 0);
 914
 915         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 916                            BTREE_ITER_PREFETCH, k, ret) {
 917                 ret = __bch2_trans_do(&trans, NULL, NULL,
 918                                       BTREE_INSERT_NOFAIL|
 919                                       BTREE_INSERT_LAZY_RW,
 920                         bch2_check_alloc_to_lru_ref(&trans, &iter));
 921                 if (ret)
 922                         break;
 923         }
 924         bch2_trans_iter_exit(&trans, &iter);
 925
 926         bch2_trans_exit(&trans);
 927         return ret < 0 ? ret : 0;
 928 }
 929
 930 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
 931                                    struct bch_dev *ca, bool *discard_done)
 932 {
 933         struct bch_fs *c = trans->c;
 934         struct btree_iter iter;
 935         struct bkey_s_c k;
 936         struct bkey_i_alloc_v4 *a;
 937         struct printbuf buf = PRINTBUF;
 938         int ret;
 939
 940         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos,
 941                              BTREE_ITER_CACHED);
 942         k = bch2_btree_iter_peek_slot(&iter);
 943         ret = bkey_err(k);
 944         if (ret)
 945                 goto out;
 946
 947         a = bch2_alloc_to_v4_mut(trans, k);
 948         ret = PTR_ERR_OR_ZERO(a);
 949         if (ret)
 950                 goto out;
 951
 952         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
 953                 a->v.gen++;
 954                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
 955                 goto write;
 956         }
 957
 958         if (bch2_trans_inconsistent_on(a->v.journal_seq > c->journal.flushed_seq_ondisk, trans,
 959                         "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
 960                         "%s",
 961                         a->v.journal_seq,
 962                         c->journal.flushed_seq_ondisk,
 963                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 964                 ret = -EIO;
 965                 goto out;
 966         }
 967
 968         if (bch2_trans_inconsistent_on(a->v.data_type != BCH_DATA_need_discard, trans,
 969                         "bucket incorrectly set in need_discard btree\n"
 970                         "%s",
 971                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 972                 ret = -EIO;
 973                 goto out;
 974         }
 975
 976         if (!*discard_done && ca->mi.discard && !c->opts.nochanges) {
 977                 /*
 978                  * This works without any other locks because this is the only
 979                  * thread that removes items from the need_discard tree
 980                  */
 981                 bch2_trans_unlock(trans);
 982                 blkdev_issue_discard(ca->disk_sb.bdev,
 983                                      k.k->p.offset * ca->mi.bucket_size,
 984                                      ca->mi.bucket_size,
 985                                      GFP_KERNEL, 0);
 986                 *discard_done = true;
 987
 988                 ret = bch2_trans_relock(trans) ? 0 : -EINTR;
 989                 if (ret)
 990                         goto out;
 991         }
 992
 993         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
 994         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
 995 write:
 996         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 997 out:
 998         bch2_trans_iter_exit(trans, &iter);
 999         printbuf_exit(&buf);
1000         return ret;
1001 }
1002
1003 static void bch2_do_discards_work(struct work_struct *work)
1004 {
1005         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1006         struct bch_dev *ca = NULL;
1007         struct btree_trans trans;
1008         struct btree_iter iter;
1009         struct bkey_s_c k;
1010         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1011         int ret;
1012
1013         bch2_trans_init(&trans, c, 0, 0);
1014
1015         for_each_btree_key(&trans, iter, BTREE_ID_need_discard,
1016                            POS_MIN, 0, k, ret) {
1017                 bool discard_done = false;
1018
1019                 if (ca && k.k->p.inode != ca->dev_idx) {
1020                         percpu_ref_put(&ca->io_ref);
1021                         ca = NULL;
1022                 }
1023
1024                 if (!ca) {
1025                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1026                         if (!percpu_ref_tryget(&ca->io_ref)) {
1027                                 ca = NULL;
1028                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
1029                                 continue;
1030                         }
1031                 }
1032
1033                 seen++;
1034
1035                 if (bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset)) {
1036                         open++;
1037                         continue;
1038                 }
1039
1040                 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1041                                 c->journal.flushed_seq_ondisk,
1042                                 k.k->p.inode, k.k->p.offset)) {
1043                         need_journal_commit++;
1044                         continue;
1045                 }
1046
1047                 ret = __bch2_trans_do(&trans, NULL, NULL,
1048                                       BTREE_INSERT_USE_RESERVE|
1049                                       BTREE_INSERT_NOFAIL,
1050                                 bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done));
1051                 if (ret)
1052                         break;
1053
1054                 discarded++;
1055         }
1056         bch2_trans_iter_exit(&trans, &iter);
1057
1058         if (ca)
1059                 percpu_ref_put(&ca->io_ref);
1060
1061         bch2_trans_exit(&trans);
1062
1063         if (need_journal_commit * 2 > seen)
1064                 bch2_journal_flush_async(&c->journal, NULL);
1065
1066         percpu_ref_put(&c->writes);
1067
1068         trace_do_discards(c, seen, open, need_journal_commit, discarded, ret);
1069 }
1070
1071 void bch2_do_discards(struct bch_fs *c)
1072 {
1073         if (percpu_ref_tryget(&c->writes) &&
1074             !queue_work(system_long_wq, &c->discard_work))
1075                 percpu_ref_put(&c->writes);
1076 }
1077
1078 static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca)
1079 {
1080         struct bch_fs *c = trans->c;
1081         struct btree_iter lru_iter, alloc_iter = { NULL };
1082         struct bkey_s_c k;
1083         struct bkey_i_alloc_v4 *a;
1084         u64 bucket, idx;
1085         struct printbuf buf = PRINTBUF;
1086         int ret;
1087
1088         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1089                              POS(ca->dev_idx, 0), 0);
1090         k = bch2_btree_iter_peek(&lru_iter);
1091         ret = bkey_err(k);
1092         if (ret)
1093                 goto out;
1094
1095         if (!k.k || k.k->p.inode != ca->dev_idx)
1096                 goto out;
1097
1098         if (bch2_trans_inconsistent_on(k.k->type != KEY_TYPE_lru, trans,
1099                                        "non lru key in lru btree"))
1100                 goto out;
1101
1102         idx     = k.k->p.offset;
1103         bucket  = le64_to_cpu(bkey_s_c_to_lru(k).v->idx);
1104
1105         a = bch2_trans_start_alloc_update(trans, &alloc_iter,
1106                                           POS(ca->dev_idx, bucket));
1107         ret = PTR_ERR_OR_ZERO(a);
1108         if (ret)
1109                 goto out;
1110
1111         if (idx != alloc_lru_idx(a->v)) {
1112                 pr_buf(&buf, "alloc key does not point back to lru entry when invalidating bucket:\n  ");
1113
1114                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1115                 pr_buf(&buf, "\n  ");
1116                 bch2_bkey_val_to_text(&buf, c, k);
1117                 bch2_trans_inconsistent(trans, "%s", buf.buf);
1118                 ret = -EINVAL;
1119                 goto out;
1120         }
1121
1122         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1123         a->v.gen++;
1124         a->v.data_type          = 0;
1125         a->v.dirty_sectors      = 0;
1126         a->v.cached_sectors     = 0;
1127         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1128         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1129
1130         ret = bch2_trans_update(trans, &alloc_iter, &a->k_i,
1131                                 BTREE_TRIGGER_BUCKET_INVALIDATE);
1132 out:
1133         bch2_trans_iter_exit(trans, &alloc_iter);
1134         bch2_trans_iter_exit(trans, &lru_iter);
1135         printbuf_exit(&buf);
1136         return ret;
1137 }
1138
1139 static void bch2_do_invalidates_work(struct work_struct *work)
1140 {
1141         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1142         struct bch_dev *ca;
1143         struct btree_trans trans;
1144         unsigned i;
1145         int ret = 0;
1146
1147         bch2_trans_init(&trans, c, 0, 0);
1148
1149         for_each_member_device(ca, c, i) {
1150                 s64 nr_to_invalidate =
1151                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1152
1153                 while (!ret && nr_to_invalidate-- >= 0)
1154                         ret = __bch2_trans_do(&trans, NULL, NULL,
1155                                               BTREE_INSERT_USE_RESERVE|
1156                                               BTREE_INSERT_NOFAIL,
1157                                         invalidate_one_bucket(&trans, ca));
1158         }
1159
1160         bch2_trans_exit(&trans);
1161         percpu_ref_put(&c->writes);
1162 }
1163
1164 void bch2_do_invalidates(struct bch_fs *c)
1165 {
1166         if (percpu_ref_tryget(&c->writes))
1167                 queue_work(system_long_wq, &c->invalidate_work);
1168 }
1169
1170 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1171 {
1172         struct btree_trans trans;
1173         struct btree_iter iter;
1174         struct bkey_s_c k;
1175         struct bch_alloc_v4 a;
1176         struct bch_member *m;
1177         int ret;
1178
1179         bch2_trans_init(&trans, c, 0, 0);
1180
1181         for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1182                            POS(ca->dev_idx, ca->mi.first_bucket),
1183                            BTREE_ITER_SLOTS|
1184                            BTREE_ITER_PREFETCH, k, ret) {
1185                 if (iter.pos.offset >= ca->mi.nbuckets)
1186                         break;
1187
1188                 bch2_alloc_to_v4(k, &a);
1189                 ret = __bch2_trans_do(&trans, NULL, NULL,
1190                                       BTREE_INSERT_LAZY_RW,
1191                                  bch2_bucket_do_index(&trans, k, &a, true));
1192                 if (ret)
1193                         break;
1194         }
1195         bch2_trans_iter_exit(&trans, &iter);
1196
1197         bch2_trans_exit(&trans);
1198
1199         if (ret) {
1200                 bch_err(ca, "error initializing free space: %i", ret);
1201                 return ret;
1202         }
1203
1204         mutex_lock(&c->sb_lock);
1205         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1206         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1207         mutex_unlock(&c->sb_lock);
1208
1209         return ret;
1210 }
1211
1212 int bch2_fs_freespace_init(struct bch_fs *c)
1213 {
1214         struct bch_dev *ca;
1215         unsigned i;
1216         int ret = 0;
1217         bool doing_init = false;
1218
1219         /*
1220          * We can crash during the device add path, so we need to check this on
1221          * every mount:
1222          */
1223
1224         for_each_member_device(ca, c, i) {
1225                 if (ca->mi.freespace_initialized)
1226                         continue;
1227
1228                 if (!doing_init) {
1229                         bch_info(c, "initializing freespace");
1230                         doing_init = true;
1231                 }
1232
1233                 ret = bch2_dev_freespace_init(c, ca);
1234                 if (ret) {
1235                         percpu_ref_put(&ca->ref);
1236                         return ret;
1237                 }
1238         }
1239
1240         if (doing_init) {
1241                 mutex_lock(&c->sb_lock);
1242                 bch2_write_super(c);
1243                 mutex_unlock(&c->sb_lock);
1244
1245                 bch_verbose(c, "done initializing freespace");
1246         }
1247
1248         return ret;
1249 }
1250
1251 /* Bucket IO clocks: */
1252
1253 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1254                               size_t bucket_nr, int rw)
1255 {
1256         struct bch_fs *c = trans->c;
1257         struct btree_iter iter;
1258         struct bkey_i_alloc_v4 *a;
1259         u64 now;
1260         int ret = 0;
1261
1262         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1263         ret = PTR_ERR_OR_ZERO(a);
1264         if (ret)
1265                 return ret;
1266
1267         now = atomic64_read(&c->io_clock[rw].now);
1268         if (a->v.io_time[rw] == now)
1269                 goto out;
1270
1271         a->v.io_time[rw] = now;
1272
1273         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1274                 bch2_trans_commit(trans, NULL, NULL, 0);
1275 out:
1276         bch2_trans_iter_exit(trans, &iter);
1277         return ret;
1278 }
1279
1280 /* Startup/shutdown (ro/rw): */
1281
1282 void bch2_recalc_capacity(struct bch_fs *c)
1283 {
1284         struct bch_dev *ca;
1285         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1286         unsigned bucket_size_max = 0;
1287         unsigned long ra_pages = 0;
1288         unsigned i;
1289
1290         lockdep_assert_held(&c->state_lock);
1291
1292         for_each_online_member(ca, c, i) {
1293                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
1294
1295                 ra_pages += bdi->ra_pages;
1296         }
1297
1298         bch2_set_ra_pages(c, ra_pages);
1299
1300         for_each_rw_member(ca, c, i) {
1301                 u64 dev_reserve = 0;
1302
1303                 /*
1304                  * We need to reserve buckets (from the number
1305                  * of currently available buckets) against
1306                  * foreground writes so that mainly copygc can
1307                  * make forward progress.
1308                  *
1309                  * We need enough to refill the various reserves
1310                  * from scratch - copygc will use its entire
1311                  * reserve all at once, then run against when
1312                  * its reserve is refilled (from the formerly
1313                  * available buckets).
1314                  *
1315                  * This reserve is just used when considering if
1316                  * allocations for foreground writes must wait -
1317                  * not -ENOSPC calculations.
1318                  */
1319
1320                 dev_reserve += ca->nr_btree_reserve * 2;
1321                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
1322
1323                 dev_reserve += 1;       /* btree write point */
1324                 dev_reserve += 1;       /* copygc write point */
1325                 dev_reserve += 1;       /* rebalance write point */
1326
1327                 dev_reserve *= ca->mi.bucket_size;
1328
1329                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1330                                              ca->mi.first_bucket);
1331
1332                 reserved_sectors += dev_reserve * 2;
1333
1334                 bucket_size_max = max_t(unsigned, bucket_size_max,
1335                                         ca->mi.bucket_size);
1336         }
1337
1338         gc_reserve = c->opts.gc_reserve_bytes
1339                 ? c->opts.gc_reserve_bytes >> 9
1340                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1341
1342         reserved_sectors = max(gc_reserve, reserved_sectors);
1343
1344         reserved_sectors = min(reserved_sectors, capacity);
1345
1346         c->capacity = capacity - reserved_sectors;
1347
1348         c->bucket_size_max = bucket_size_max;
1349
1350         /* Wake up case someone was waiting for buckets */
1351         closure_wake_up(&c->freelist_wait);
1352 }
1353
1354 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1355 {
1356         struct open_bucket *ob;
1357         bool ret = false;
1358
1359         for (ob = c->open_buckets;
1360              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1361              ob++) {
1362                 spin_lock(&ob->lock);
1363                 if (ob->valid && !ob->on_partial_list &&
1364                     ob->dev == ca->dev_idx)
1365                         ret = true;
1366                 spin_unlock(&ob->lock);
1367         }
1368
1369         return ret;
1370 }
1371
1372 /* device goes ro: */
1373 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1374 {
1375         unsigned i;
1376
1377         /* First, remove device from allocation groups: */
1378
1379         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1380                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1381
1382         /*
1383          * Capacity is calculated based off of devices in allocation groups:
1384          */
1385         bch2_recalc_capacity(c);
1386
1387         /* Next, close write points that point to this device... */
1388         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1389                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1390
1391         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1392         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1393         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1394
1395         mutex_lock(&c->btree_reserve_cache_lock);
1396         while (c->btree_reserve_cache_nr) {
1397                 struct btree_alloc *a =
1398                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1399
1400                 bch2_open_buckets_put(c, &a->ob);
1401         }
1402         mutex_unlock(&c->btree_reserve_cache_lock);
1403
1404         while (1) {
1405                 struct open_bucket *ob;
1406
1407                 spin_lock(&c->freelist_lock);
1408                 if (!ca->open_buckets_partial_nr) {
1409                         spin_unlock(&c->freelist_lock);
1410                         break;
1411                 }
1412                 ob = c->open_buckets +
1413                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1414                 ob->on_partial_list = false;
1415                 spin_unlock(&c->freelist_lock);
1416
1417                 bch2_open_bucket_put(c, ob);
1418         }
1419
1420         bch2_ec_stop_dev(c, ca);
1421
1422         /*
1423          * Wake up threads that were blocked on allocation, so they can notice
1424          * the device can no longer be removed and the capacity has changed:
1425          */
1426         closure_wake_up(&c->freelist_wait);
1427
1428         /*
1429          * journal_res_get() can block waiting for free space in the journal -
1430          * it needs to notice there may not be devices to allocate from anymore:
1431          */
1432         wake_up(&c->journal.wait);
1433
1434         /* Now wait for any in flight writes: */
1435
1436         closure_wait_event(&c->open_buckets_wait,
1437                            !bch2_dev_has_open_write_point(c, ca));
1438 }
1439
1440 /* device goes rw: */
1441 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1442 {
1443         unsigned i;
1444
1445         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1446                 if (ca->mi.data_allowed & (1 << i))
1447                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1448 }
1449
1450 void bch2_fs_allocator_background_init(struct bch_fs *c)
1451 {
1452         spin_lock_init(&c->freelist_lock);
1453         INIT_WORK(&c->discard_work, bch2_do_discards_work);
1454         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
1455 }