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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "bkey_buf.h"
   6 #include "btree_update.h"
   7 #include "buckets.h"
   8 #include "data_update.h"
   9 #include "ec.h"
  10 #include "extents.h"
  11 #include "io.h"
  12 #include "keylist.h"
  13 #include "move.h"
  14 #include "subvolume.h"
  15
  16 #include <trace/events/bcachefs.h>
  17
  18 static int insert_snapshot_whiteouts(struct btree_trans *trans,
  19                                      enum btree_id id,
  20                                      struct bpos old_pos,
  21                                      struct bpos new_pos)
  22 {
  23         struct bch_fs *c = trans->c;
  24         struct btree_iter iter, update_iter;
  25         struct bkey_s_c k;
  26         snapshot_id_list s;
  27         int ret;
  28
  29         if (!btree_type_has_snapshots(id))
  30                 return 0;
  31
  32         darray_init(&s);
  33
  34         if (!bkey_cmp(old_pos, new_pos))
  35                 return 0;
  36
  37         if (!snapshot_t(c, old_pos.snapshot)->children[0])
  38                 return 0;
  39
  40         bch2_trans_iter_init(trans, &iter, id, old_pos,
  41                              BTREE_ITER_NOT_EXTENTS|
  42                              BTREE_ITER_ALL_SNAPSHOTS);
  43         while (1) {
  44                 k = bch2_btree_iter_prev(&iter);
  45                 ret = bkey_err(k);
  46                 if (ret)
  47                         break;
  48
  49                 if (bkey_cmp(old_pos, k.k->p))
  50                         break;
  51
  52                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot)) {
  53                         struct bkey_i *update;
  54
  55                         if (snapshot_list_has_ancestor(c, &s, k.k->p.snapshot))
  56                                 continue;
  57
  58                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
  59
  60                         ret = PTR_ERR_OR_ZERO(update);
  61                         if (ret)
  62                                 break;
  63
  64                         bkey_init(&update->k);
  65                         update->k.p = new_pos;
  66                         update->k.p.snapshot = k.k->p.snapshot;
  67
  68                         bch2_trans_iter_init(trans, &update_iter, id, update->k.p,
  69                                              BTREE_ITER_NOT_EXTENTS|
  70                                              BTREE_ITER_ALL_SNAPSHOTS|
  71                                              BTREE_ITER_INTENT);
  72                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
  73                                 bch2_trans_update(trans, &update_iter, update,
  74                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
  75                         bch2_trans_iter_exit(trans, &update_iter);
  76                         if (ret)
  77                                 break;
  78
  79                         ret = snapshot_list_add(c, &s, k.k->p.snapshot);
  80                         if (ret)
  81                                 break;
  82                 }
  83         }
  84         bch2_trans_iter_exit(trans, &iter);
  85         darray_exit(&s);
  86
  87         return ret;
  88 }
  89
  90 static void bch2_bkey_mark_dev_cached(struct bkey_s k, unsigned dev)
  91 {
  92         struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
  93         struct bch_extent_ptr *ptr;
  94
  95         bkey_for_each_ptr(ptrs, ptr)
  96                 if (ptr->dev == dev)
  97                         ptr->cached = true;
  98 }
  99
 100 static int __bch2_data_update_index_update(struct btree_trans *trans,
 101                                            struct bch_write_op *op)
 102 {
 103         struct bch_fs *c = op->c;
 104         struct btree_iter iter;
 105         struct data_update *m =
 106                 container_of(op, struct data_update, op);
 107         struct keylist *keys = &op->insert_keys;
 108         struct bkey_buf _new, _insert;
 109         int ret = 0;
 110
 111         bch2_bkey_buf_init(&_new);
 112         bch2_bkey_buf_init(&_insert);
 113         bch2_bkey_buf_realloc(&_insert, c, U8_MAX);
 114
 115         bch2_trans_iter_init(trans, &iter, m->btree_id,
 116                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 117                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 118
 119         while (1) {
 120                 struct bkey_s_c k;
 121                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 122                 struct bkey_i *insert;
 123                 struct bkey_i_extent *new;
 124                 const union bch_extent_entry *entry;
 125                 struct extent_ptr_decoded p;
 126                 struct bpos next_pos;
 127                 bool did_work = false;
 128                 bool should_check_enospc;
 129                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 130                 unsigned i;
 131
 132                 bch2_trans_begin(trans);
 133
 134                 k = bch2_btree_iter_peek_slot(&iter);
 135                 ret = bkey_err(k);
 136                 if (ret)
 137                         goto err;
 138
 139                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 140
 141                 if (!bch2_extents_match(k, old))
 142                         goto nomatch;
 143
 144                 bkey_reassemble(_insert.k, k);
 145                 insert = _insert.k;
 146
 147                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 148                 new = bkey_i_to_extent(_new.k);
 149                 bch2_cut_front(iter.pos, &new->k_i);
 150
 151                 bch2_cut_front(iter.pos,        insert);
 152                 bch2_cut_back(new->k.p,         insert);
 153                 bch2_cut_back(insert->k.p,      &new->k_i);
 154
 155                 /*
 156                  * @old: extent that we read from
 157                  * @insert: key that we're going to update, initialized from
 158                  * extent currently in btree - same as @old unless we raced with
 159                  * other updates
 160                  * @new: extent with new pointers that we'll be adding to @insert
 161                  *
 162                  * Fist, drop rewrite_ptrs from @new:
 163                  */
 164                 i = 0;
 165                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
 166                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 167                             bch2_extent_has_ptr(old, p, bkey_i_to_s_c(insert))) {
 168                                 /*
 169                                  * If we're going to be adding a pointer to the
 170                                  * same device, we have to drop the old one -
 171                                  * otherwise, we can just mark it cached:
 172                                  */
 173                                 if (bch2_bkey_has_device(bkey_i_to_s_c(&new->k_i), p.ptr.dev))
 174                                         bch2_bkey_drop_device_noerror(bkey_i_to_s(insert), p.ptr.dev);
 175                                 else
 176                                         bch2_bkey_mark_dev_cached(bkey_i_to_s(insert), p.ptr.dev);
 177                         }
 178                         i++;
 179                 }
 180
 181
 182                 /* Add new ptrs: */
 183                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 184                         if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
 185                                 /*
 186                                  * raced with another move op? extent already
 187                                  * has a pointer to the device we just wrote
 188                                  * data to
 189                                  */
 190                                 continue;
 191                         }
 192
 193                         bch2_extent_ptr_decoded_append(insert, &p);
 194                         did_work = true;
 195                 }
 196
 197                 if (!did_work)
 198                         goto nomatch;
 199
 200                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 201                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 202
 203                 ret = bch2_sum_sector_overwrites(trans, &iter, insert,
 204                                                  &should_check_enospc,
 205                                                  &i_sectors_delta,
 206                                                  &disk_sectors_delta);
 207                 if (ret)
 208                         goto err;
 209
 210                 if (disk_sectors_delta > (s64) op->res.sectors) {
 211                         ret = bch2_disk_reservation_add(c, &op->res,
 212                                                 disk_sectors_delta - op->res.sectors,
 213                                                 !should_check_enospc
 214                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 215                         if (ret)
 216                                 goto out;
 217                 }
 218
 219                 next_pos = insert->k.p;
 220
 221                 ret   = insert_snapshot_whiteouts(trans, m->btree_id,
 222                                                   k.k->p, insert->k.p) ?:
 223                         bch2_trans_update(trans, &iter, insert,
 224                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 225                         bch2_trans_commit(trans, &op->res,
 226                                 NULL,
 227                                 BTREE_INSERT_NOFAIL|
 228                                 m->data_opts.btree_insert_flags);
 229                 if (!ret) {
 230                         bch2_btree_iter_set_pos(&iter, next_pos);
 231
 232                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 233                         trace_move_extent_finish(&new->k);
 234                 }
 235 err:
 236                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 237                         ret = 0;
 238                 if (ret)
 239                         break;
 240 next:
 241                 while (bkey_cmp(iter.pos, bch2_keylist_front(keys)->k.p) >= 0) {
 242                         bch2_keylist_pop_front(keys);
 243                         if (bch2_keylist_empty(keys))
 244                                 goto out;
 245                 }
 246                 continue;
 247 nomatch:
 248                 if (m->ctxt) {
 249                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 250                         atomic64_inc(&m->ctxt->stats->keys_raced);
 251                         atomic64_add(k.k->p.offset - iter.pos.offset,
 252                                      &m->ctxt->stats->sectors_raced);
 253                 }
 254
 255                 this_cpu_add(c->counters[BCH_COUNTER_move_extent_race], new->k.size);
 256                 trace_move_extent_race(&new->k);
 257
 258                 bch2_btree_iter_advance(&iter);
 259                 goto next;
 260         }
 261 out:
 262         bch2_trans_iter_exit(trans, &iter);
 263         bch2_bkey_buf_exit(&_insert, c);
 264         bch2_bkey_buf_exit(&_new, c);
 265         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 266         return ret;
 267 }
 268
 269 int bch2_data_update_index_update(struct bch_write_op *op)
 270 {
 271         struct bch_fs *c = op->c;
 272         struct btree_trans trans;
 273         int ret;
 274
 275         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 276         ret = __bch2_data_update_index_update(&trans, op);
 277         bch2_trans_exit(&trans);
 278
 279         return ret;
 280 }
 281
 282 void bch2_data_update_read_done(struct data_update *m,
 283                                 struct bch_extent_crc_unpacked crc)
 284 {
 285         /* write bio must own pages: */
 286         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 287
 288         m->op.crc = crc;
 289         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 290
 291         closure_call(&m->op.cl, bch2_write, NULL, NULL);
 292 }
 293
 294 void bch2_data_update_exit(struct data_update *update)
 295 {
 296         struct bch_fs *c = update->op.c;
 297         struct bkey_ptrs_c ptrs =
 298                 bch2_bkey_ptrs_c(bkey_i_to_s_c(update->k.k));
 299         const struct bch_extent_ptr *ptr;
 300
 301         bkey_for_each_ptr(ptrs, ptr)
 302                 bch2_bucket_nocow_unlock(&c->nocow_locks,
 303                                        PTR_BUCKET_POS(c, ptr), 0);
 304
 305         bch2_bkey_buf_exit(&update->k, c);
 306         bch2_disk_reservation_put(c, &update->op.res);
 307         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 308 }
 309
 310 void bch2_update_unwritten_extent(struct btree_trans *trans,
 311                                   struct data_update *update)
 312 {
 313         struct bch_fs *c = update->op.c;
 314         struct bio *bio = &update->op.wbio.bio;
 315         struct bkey_i_extent *e;
 316         struct write_point *wp;
 317         struct bch_extent_ptr *ptr;
 318         struct closure cl;
 319         struct btree_iter iter;
 320         struct bkey_s_c k;
 321         int ret;
 322
 323         closure_init_stack(&cl);
 324         bch2_keylist_init(&update->op.insert_keys, update->op.inline_keys);
 325
 326         while (bio_sectors(bio)) {
 327                 unsigned sectors = bio_sectors(bio);
 328
 329                 bch2_trans_iter_init(trans, &iter, update->btree_id, update->op.pos,
 330                                      BTREE_ITER_SLOTS);
 331                 ret = lockrestart_do(trans, ({
 332                         k = bch2_btree_iter_peek_slot(&iter);
 333                         bkey_err(k);
 334                 }));
 335                 bch2_trans_iter_exit(trans, &iter);
 336
 337                 if (ret || !bch2_extents_match(k, bkey_i_to_s_c(update->k.k)))
 338                         break;
 339
 340                 e = bkey_extent_init(update->op.insert_keys.top);
 341                 e->k.p = update->op.pos;
 342
 343                 ret = bch2_alloc_sectors_start_trans(trans,
 344                                 update->op.target,
 345                                 false,
 346                                 update->op.write_point,
 347                                 &update->op.devs_have,
 348                                 update->op.nr_replicas,
 349                                 update->op.nr_replicas,
 350                                 update->op.alloc_reserve,
 351                                 0, &cl, &wp);
 352                 if (ret == -EAGAIN) {
 353                         bch2_trans_unlock(trans);
 354                         closure_sync(&cl);
 355                         continue;
 356                 }
 357
 358                 if (ret)
 359                         return;
 360
 361                 sectors = min(sectors, wp->sectors_free);
 362
 363                 bch2_key_resize(&e->k, sectors);
 364
 365                 bch2_open_bucket_get(c, wp, &update->op.open_buckets);
 366                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 367                 bch2_alloc_sectors_done(c, wp);
 368
 369                 bio_advance(bio, sectors << 9);
 370                 update->op.pos.offset += sectors;
 371
 372                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 373                         ptr->unwritten = true;
 374                 bch2_keylist_push(&update->op.insert_keys);
 375
 376                 ret = __bch2_data_update_index_update(trans, &update->op);
 377
 378                 bch2_open_buckets_put(c, &update->op.open_buckets);
 379
 380                 if (ret)
 381                         break;
 382         }
 383
 384         if ((atomic_read(&cl.remaining) & CLOSURE_REMAINING_MASK) != 1) {
 385                 bch2_trans_unlock(trans);
 386                 closure_sync(&cl);
 387         }
 388 }
 389
 390 int bch2_data_update_init(struct bch_fs *c, struct data_update *m,
 391                           struct write_point_specifier wp,
 392                           struct bch_io_opts io_opts,
 393                           struct data_update_opts data_opts,
 394                           enum btree_id btree_id,
 395                           struct bkey_s_c k)
 396 {
 397         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 398         const union bch_extent_entry *entry;
 399         struct extent_ptr_decoded p;
 400         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 401         int ret;
 402
 403         bch2_bkey_buf_init(&m->k);
 404         bch2_bkey_buf_reassemble(&m->k, c, k);
 405         m->btree_id     = btree_id;
 406         m->data_opts    = data_opts;
 407
 408         bch2_write_op_init(&m->op, c, io_opts);
 409         m->op.pos       = bkey_start_pos(k.k);
 410         m->op.version   = k.k->version;
 411         m->op.target    = data_opts.target;
 412         m->op.write_point = wp;
 413         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 414                 BCH_WRITE_PAGES_OWNED|
 415                 BCH_WRITE_DATA_ENCODED|
 416                 BCH_WRITE_MOVE|
 417                 m->data_opts.write_flags;
 418         m->op.compression_type =
 419                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 420                                              io_opts.compression];
 421         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 422                 m->op.alloc_reserve = RESERVE_movinggc;
 423
 424         i = 0;
 425         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 426                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 427                     p.ptr.cached)
 428                         BUG();
 429
 430                 if (!((1U << i) & m->data_opts.rewrite_ptrs))
 431                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 432
 433                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 434                     crc_is_compressed(p.crc))
 435                         reserve_sectors += k.k->size;
 436
 437                 /*
 438                  * op->csum_type is normally initialized from the fs/file's
 439                  * current options - but if an extent is encrypted, we require
 440                  * that it stays encrypted:
 441                  */
 442                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 443                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 444                         m->op.csum_type = p.crc.csum_type;
 445                 }
 446
 447                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 448                         m->op.incompressible = true;
 449
 450                 i++;
 451
 452                 bch2_bucket_nocow_lock(&c->nocow_locks,
 453                                        PTR_BUCKET_POS(c, &p.ptr), 0);
 454         }
 455
 456         if (reserve_sectors) {
 457                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 458                                 m->data_opts.extra_replicas
 459                                 ? 0
 460                                 : BCH_DISK_RESERVATION_NOFAIL);
 461                 if (ret)
 462                         return ret;
 463         }
 464
 465         m->op.nr_replicas = m->op.nr_replicas_required =
 466                 hweight32(m->data_opts.rewrite_ptrs) + m->data_opts.extra_replicas;
 467
 468         BUG_ON(!m->op.nr_replicas);
 469
 470         /* Special handling required: */
 471         if (bkey_extent_is_unwritten(k))
 472                 return -BCH_ERR_unwritten_extent_update;
 473         return 0;
 474 }
 475
 476 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 477 {
 478         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 479         const struct bch_extent_ptr *ptr;
 480         unsigned i = 0;
 481
 482         bkey_for_each_ptr(ptrs, ptr) {
 483                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 484                         opts->kill_ptrs |= 1U << i;
 485                         opts->rewrite_ptrs ^= 1U << i;
 486                 }
 487
 488                 i++;
 489         }
 490 }