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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * io_misc.c - fallocate, fpunch, truncate:
   4  */
   5
   6 #include "bcachefs.h"
   7 #include "alloc_foreground.h"
   8 #include "bkey_buf.h"
   9 #include "btree_update.h"
  10 #include "buckets.h"
  11 #include "clock.h"
  12 #include "error.h"
  13 #include "extents.h"
  14 #include "extent_update.h"
  15 #include "inode.h"
  16 #include "io_misc.h"
  17 #include "io_write.h"
  18 #include "logged_ops.h"
  19 #include "rebalance.h"
  20 #include "subvolume.h"
  21
  22 /* Overwrites whatever was present with zeroes: */
  23 int bch2_extent_fallocate(struct btree_trans *trans,
  24                           subvol_inum inum,
  25                           struct btree_iter *iter,
  26                           u64 sectors,
  27                           struct bch_io_opts opts,
  28                           s64 *i_sectors_delta,
  29                           struct write_point_specifier write_point)
  30 {
  31         struct bch_fs *c = trans->c;
  32         struct disk_reservation disk_res = { 0 };
  33         struct closure cl;
  34         struct open_buckets open_buckets = { 0 };
  35         struct bkey_s_c k;
  36         struct bkey_buf old, new;
  37         unsigned sectors_allocated = 0, new_replicas;
  38         bool unwritten = opts.nocow &&
  39             c->sb.version >= bcachefs_metadata_version_unwritten_extents;
  40         int ret;
  41
  42         bch2_bkey_buf_init(&old);
  43         bch2_bkey_buf_init(&new);
  44         closure_init_stack(&cl);
  45
  46         k = bch2_btree_iter_peek_slot(iter);
  47         ret = bkey_err(k);
  48         if (ret)
  49                 return ret;
  50
  51         sectors = min_t(u64, sectors, k.k->p.offset - iter->pos.offset);
  52         new_replicas = max(0, (int) opts.data_replicas -
  53                            (int) bch2_bkey_nr_ptrs_fully_allocated(k));
  54
  55         /*
  56          * Get a disk reservation before (in the nocow case) calling
  57          * into the allocator:
  58          */
  59         ret = bch2_disk_reservation_get(c, &disk_res, sectors, new_replicas, 0);
  60         if (unlikely(ret))
  61                 goto err;
  62
  63         bch2_bkey_buf_reassemble(&old, c, k);
  64
  65         if (!unwritten) {
  66                 struct bkey_i_reservation *reservation;
  67
  68                 bch2_bkey_buf_realloc(&new, c, sizeof(*reservation) / sizeof(u64));
  69                 reservation = bkey_reservation_init(new.k);
  70                 reservation->k.p = iter->pos;
  71                 bch2_key_resize(&reservation->k, sectors);
  72                 reservation->v.nr_replicas = opts.data_replicas;
  73         } else {
  74                 struct bkey_i_extent *e;
  75                 struct bch_devs_list devs_have;
  76                 struct write_point *wp;
  77                 struct bch_extent_ptr *ptr;
  78
  79                 devs_have.nr = 0;
  80
  81                 bch2_bkey_buf_realloc(&new, c, BKEY_EXTENT_U64s_MAX);
  82
  83                 e = bkey_extent_init(new.k);
  84                 e->k.p = iter->pos;
  85
  86                 ret = bch2_alloc_sectors_start_trans(trans,
  87                                 opts.foreground_target,
  88                                 false,
  89                                 write_point,
  90                                 &devs_have,
  91                                 opts.data_replicas,
  92                                 opts.data_replicas,
  93                                 BCH_WATERMARK_normal, 0, &cl, &wp);
  94                 if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
  95                         ret = -BCH_ERR_transaction_restart_nested;
  96                 if (ret)
  97                         goto err;
  98
  99                 sectors = min_t(u64, sectors, wp->sectors_free);
 100                 sectors_allocated = sectors;
 101
 102                 bch2_key_resize(&e->k, sectors);
 103
 104                 bch2_open_bucket_get(c, wp, &open_buckets);
 105                 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, sectors, false);
 106                 bch2_alloc_sectors_done(c, wp);
 107
 108                 extent_for_each_ptr(extent_i_to_s(e), ptr)
 109                         ptr->unwritten = true;
 110         }
 111
 112         ret = bch2_extent_update(trans, inum, iter, new.k, &disk_res,
 113                                  0, i_sectors_delta, true);
 114 err:
 115         if (!ret && sectors_allocated)
 116                 bch2_increment_clock(c, sectors_allocated, WRITE);
 117         if (should_print_err(ret))
 118                 bch_err_inum_offset_ratelimited(c,
 119                         inum.inum,
 120                         iter->pos.offset << 9,
 121                         "%s(): error: %s", __func__, bch2_err_str(ret));
 122
 123         bch2_open_buckets_put(c, &open_buckets);
 124         bch2_disk_reservation_put(c, &disk_res);
 125         bch2_bkey_buf_exit(&new, c);
 126         bch2_bkey_buf_exit(&old, c);
 127
 128         if (closure_nr_remaining(&cl) != 1) {
 129                 bch2_trans_unlock(trans);
 130                 closure_sync(&cl);
 131         }
 132
 133         return ret;
 134 }
 135
 136 /*
 137  * Returns -BCH_ERR_transacton_restart if we had to drop locks:
 138  */
 139 int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
 140                    subvol_inum inum, u64 end,
 141                    s64 *i_sectors_delta)
 142 {
 143         struct bch_fs *c        = trans->c;
 144         unsigned max_sectors    = KEY_SIZE_MAX & (~0 << c->block_bits);
 145         struct bpos end_pos = POS(inum.inum, end);
 146         struct bkey_s_c k;
 147         int ret = 0, ret2 = 0;
 148         u32 snapshot;
 149
 150         while (!ret ||
 151                bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 152                 struct disk_reservation disk_res =
 153                         bch2_disk_reservation_init(c, 0);
 154                 struct bkey_i delete;
 155
 156                 if (ret)
 157                         ret2 = ret;
 158
 159                 bch2_trans_begin(trans);
 160
 161                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 162                 if (ret)
 163                         continue;
 164
 165                 bch2_btree_iter_set_snapshot(iter, snapshot);
 166
 167                 /*
 168                  * peek_upto() doesn't have ideal semantics for extents:
 169                  */
 170                 k = bch2_btree_iter_peek_upto(iter, end_pos);
 171                 if (!k.k)
 172                         break;
 173
 174                 ret = bkey_err(k);
 175                 if (ret)
 176                         continue;
 177
 178                 bkey_init(&delete.k);
 179                 delete.k.p = iter->pos;
 180
 181                 /* create the biggest key we can */
 182                 bch2_key_resize(&delete.k, max_sectors);
 183                 bch2_cut_back(end_pos, &delete);
 184
 185                 ret = bch2_extent_update(trans, inum, iter, &delete,
 186                                 &disk_res, 0, i_sectors_delta, false);
 187                 bch2_disk_reservation_put(c, &disk_res);
 188         }
 189
 190         return ret ?: ret2;
 191 }
 192
 193 int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
 194                 s64 *i_sectors_delta)
 195 {
 196         struct btree_trans *trans = bch2_trans_get(c);
 197         struct btree_iter iter;
 198         int ret;
 199
 200         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
 201                              POS(inum.inum, start),
 202                              BTREE_ITER_INTENT);
 203
 204         ret = bch2_fpunch_at(trans, &iter, inum, end, i_sectors_delta);
 205
 206         bch2_trans_iter_exit(trans, &iter);
 207         bch2_trans_put(trans);
 208
 209         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 210                 ret = 0;
 211
 212         return ret;
 213 }
 214
 215 /* truncate: */
 216
 217 void bch2_logged_op_truncate_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 218 {
 219         struct bkey_s_c_logged_op_truncate op = bkey_s_c_to_logged_op_truncate(k);
 220
 221         prt_printf(out, "subvol=%u", le32_to_cpu(op.v->subvol));
 222         prt_printf(out, " inum=%llu", le64_to_cpu(op.v->inum));
 223         prt_printf(out, " new_i_size=%llu", le64_to_cpu(op.v->new_i_size));
 224 }
 225
 226 static int truncate_set_isize(struct btree_trans *trans,
 227                               subvol_inum inum,
 228                               u64 new_i_size)
 229 {
 230         struct btree_iter iter = { NULL };
 231         struct bch_inode_unpacked inode_u;
 232         int ret;
 233
 234         ret   = bch2_inode_peek(trans, &iter, &inode_u, inum, BTREE_ITER_INTENT) ?:
 235                 (inode_u.bi_size = new_i_size, 0) ?:
 236                 bch2_inode_write(trans, &iter, &inode_u);
 237
 238         bch2_trans_iter_exit(trans, &iter);
 239         return ret;
 240 }
 241
 242 static int __bch2_resume_logged_op_truncate(struct btree_trans *trans,
 243                                             struct bkey_i *op_k,
 244                                             u64 *i_sectors_delta)
 245 {
 246         struct bch_fs *c = trans->c;
 247         struct btree_iter fpunch_iter;
 248         struct bkey_i_logged_op_truncate *op = bkey_i_to_logged_op_truncate(op_k);
 249         subvol_inum inum = { le32_to_cpu(op->v.subvol), le64_to_cpu(op->v.inum) };
 250         u64 new_i_size = le64_to_cpu(op->v.new_i_size);
 251         int ret;
 252
 253         ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 254                         truncate_set_isize(trans, inum, new_i_size));
 255         if (ret)
 256                 goto err;
 257
 258         bch2_trans_iter_init(trans, &fpunch_iter, BTREE_ID_extents,
 259                              POS(inum.inum, round_up(new_i_size, block_bytes(c)) >> 9),
 260                              BTREE_ITER_INTENT);
 261         ret = bch2_fpunch_at(trans, &fpunch_iter, inum, U64_MAX, i_sectors_delta);
 262         bch2_trans_iter_exit(trans, &fpunch_iter);
 263
 264         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 265                 ret = 0;
 266 err:
 267         bch2_logged_op_finish(trans, op_k);
 268         return ret;
 269 }
 270
 271 int bch2_resume_logged_op_truncate(struct btree_trans *trans, struct bkey_i *op_k)
 272 {
 273         return __bch2_resume_logged_op_truncate(trans, op_k, NULL);
 274 }
 275
 276 int bch2_truncate(struct bch_fs *c, subvol_inum inum, u64 new_i_size, u64 *i_sectors_delta)
 277 {
 278         struct bkey_i_logged_op_truncate op;
 279
 280         bkey_logged_op_truncate_init(&op.k_i);
 281         op.v.subvol     = cpu_to_le32(inum.subvol);
 282         op.v.inum       = cpu_to_le64(inum.inum);
 283         op.v.new_i_size = cpu_to_le64(new_i_size);
 284
 285         /*
 286          * Logged ops aren't atomic w.r.t. snapshot creation: creating a
 287          * snapshot while they're in progress, then crashing, will result in the
 288          * resume only proceeding in one of the snapshots
 289          */
 290         down_read(&c->snapshot_create_lock);
 291         int ret = bch2_trans_run(c,
 292                 bch2_logged_op_start(trans, &op.k_i) ?:
 293                 __bch2_resume_logged_op_truncate(trans, &op.k_i, i_sectors_delta));
 294         up_read(&c->snapshot_create_lock);
 295
 296         return ret;
 297 }
 298
 299 /* finsert/fcollapse: */
 300
 301 void bch2_logged_op_finsert_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 302 {
 303         struct bkey_s_c_logged_op_finsert op = bkey_s_c_to_logged_op_finsert(k);
 304
 305         prt_printf(out, "subvol=%u",            le32_to_cpu(op.v->subvol));
 306         prt_printf(out, " inum=%llu",           le64_to_cpu(op.v->inum));
 307         prt_printf(out, " dst_offset=%lli",     le64_to_cpu(op.v->dst_offset));
 308         prt_printf(out, " src_offset=%llu",     le64_to_cpu(op.v->src_offset));
 309 }
 310
 311 static int adjust_i_size(struct btree_trans *trans, subvol_inum inum, u64 offset, s64 len)
 312 {
 313         struct btree_iter iter;
 314         struct bch_inode_unpacked inode_u;
 315         int ret;
 316
 317         offset  <<= 9;
 318         len     <<= 9;
 319
 320         ret = bch2_inode_peek(trans, &iter, &inode_u, inum, BTREE_ITER_INTENT);
 321         if (ret)
 322                 return ret;
 323
 324         if (len > 0) {
 325                 if (MAX_LFS_FILESIZE - inode_u.bi_size < len) {
 326                         ret = -EFBIG;
 327                         goto err;
 328                 }
 329
 330                 if (offset >= inode_u.bi_size) {
 331                         ret = -EINVAL;
 332                         goto err;
 333                 }
 334         }
 335
 336         inode_u.bi_size += len;
 337         inode_u.bi_mtime = inode_u.bi_ctime = bch2_current_time(trans->c);
 338
 339         ret = bch2_inode_write(trans, &iter, &inode_u);
 340 err:
 341         bch2_trans_iter_exit(trans, &iter);
 342         return ret;
 343 }
 344
 345 static int __bch2_resume_logged_op_finsert(struct btree_trans *trans,
 346                                            struct bkey_i *op_k,
 347                                            u64 *i_sectors_delta)
 348 {
 349         struct bch_fs *c = trans->c;
 350         struct btree_iter iter;
 351         struct bkey_i_logged_op_finsert *op = bkey_i_to_logged_op_finsert(op_k);
 352         subvol_inum inum = { le32_to_cpu(op->v.subvol), le64_to_cpu(op->v.inum) };
 353         struct bch_io_opts opts;
 354         u64 dst_offset = le64_to_cpu(op->v.dst_offset);
 355         u64 src_offset = le64_to_cpu(op->v.src_offset);
 356         s64 shift = dst_offset - src_offset;
 357         u64 len = abs(shift);
 358         u64 pos = le64_to_cpu(op->v.pos);
 359         bool insert = shift > 0;
 360         int ret = 0;
 361
 362         ret = bch2_inum_opts_get(trans, inum, &opts);
 363         if (ret)
 364                 return ret;
 365
 366         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
 367                              POS(inum.inum, 0),
 368                              BTREE_ITER_INTENT);
 369
 370         switch (op->v.state) {
 371 case LOGGED_OP_FINSERT_start:
 372         op->v.state = LOGGED_OP_FINSERT_shift_extents;
 373
 374         if (insert) {
 375                 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 376                                 adjust_i_size(trans, inum, src_offset, len) ?:
 377                                 bch2_logged_op_update(trans, &op->k_i));
 378                 if (ret)
 379                         goto err;
 380         } else {
 381                 bch2_btree_iter_set_pos(&iter, POS(inum.inum, src_offset));
 382
 383                 ret = bch2_fpunch_at(trans, &iter, inum, src_offset + len, i_sectors_delta);
 384                 if (ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart))
 385                         goto err;
 386
 387                 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 388                                 bch2_logged_op_update(trans, &op->k_i));
 389         }
 390
 391         fallthrough;
 392 case LOGGED_OP_FINSERT_shift_extents:
 393         while (1) {
 394                 struct disk_reservation disk_res =
 395                         bch2_disk_reservation_init(c, 0);
 396                 struct bkey_i delete, *copy;
 397                 struct bkey_s_c k;
 398                 struct bpos src_pos = POS(inum.inum, src_offset);
 399                 u32 snapshot;
 400
 401                 bch2_trans_begin(trans);
 402
 403                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 404                 if (ret)
 405                         goto btree_err;
 406
 407                 bch2_btree_iter_set_snapshot(&iter, snapshot);
 408                 bch2_btree_iter_set_pos(&iter, SPOS(inum.inum, pos, snapshot));
 409
 410                 k = insert
 411                         ? bch2_btree_iter_peek_prev(&iter)
 412                         : bch2_btree_iter_peek_upto(&iter, POS(inum.inum, U64_MAX));
 413                 if ((ret = bkey_err(k)))
 414                         goto btree_err;
 415
 416                 if (!k.k ||
 417                     k.k->p.inode != inum.inum ||
 418                     bkey_le(k.k->p, POS(inum.inum, src_offset)))
 419                         break;
 420
 421                 copy = bch2_bkey_make_mut_noupdate(trans, k);
 422                 if ((ret = PTR_ERR_OR_ZERO(copy)))
 423                         goto btree_err;
 424
 425                 if (insert &&
 426                     bkey_lt(bkey_start_pos(k.k), src_pos)) {
 427                         bch2_cut_front(src_pos, copy);
 428
 429                         /* Splitting compressed extent? */
 430                         bch2_disk_reservation_add(c, &disk_res,
 431                                         copy->k.size *
 432                                         bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(copy)),
 433                                         BCH_DISK_RESERVATION_NOFAIL);
 434                 }
 435
 436                 bkey_init(&delete.k);
 437                 delete.k.p = copy->k.p;
 438                 delete.k.p.snapshot = snapshot;
 439                 delete.k.size = copy->k.size;
 440
 441                 copy->k.p.offset += shift;
 442                 copy->k.p.snapshot = snapshot;
 443
 444                 op->v.pos = cpu_to_le64(insert ? bkey_start_offset(&delete.k) : delete.k.p.offset);
 445
 446                 ret =   bch2_bkey_set_needs_rebalance(c, copy,
 447                                         opts.background_target,
 448                                         opts.background_compression) ?:
 449                         bch2_btree_insert_trans(trans, BTREE_ID_extents, &delete, 0) ?:
 450                         bch2_btree_insert_trans(trans, BTREE_ID_extents, copy, 0) ?:
 451                         bch2_logged_op_update(trans, &op->k_i) ?:
 452                         bch2_trans_commit(trans, &disk_res, NULL, BCH_TRANS_COMMIT_no_enospc);
 453 btree_err:
 454                 bch2_disk_reservation_put(c, &disk_res);
 455
 456                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 457                         continue;
 458                 if (ret)
 459                         goto err;
 460
 461                 pos = le64_to_cpu(op->v.pos);
 462         }
 463
 464         op->v.state = LOGGED_OP_FINSERT_finish;
 465
 466         if (!insert) {
 467                 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 468                                 adjust_i_size(trans, inum, src_offset, shift) ?:
 469                                 bch2_logged_op_update(trans, &op->k_i));
 470         } else {
 471                 /* We need an inode update to update bi_journal_seq for fsync: */
 472                 ret = commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 473                                 adjust_i_size(trans, inum, 0, 0) ?:
 474                                 bch2_logged_op_update(trans, &op->k_i));
 475         }
 476
 477         break;
 478 case LOGGED_OP_FINSERT_finish:
 479         break;
 480         }
 481 err:
 482         bch2_logged_op_finish(trans, op_k);
 483         bch2_trans_iter_exit(trans, &iter);
 484         return ret;
 485 }
 486
 487 int bch2_resume_logged_op_finsert(struct btree_trans *trans, struct bkey_i *op_k)
 488 {
 489         return __bch2_resume_logged_op_finsert(trans, op_k, NULL);
 490 }
 491
 492 int bch2_fcollapse_finsert(struct bch_fs *c, subvol_inum inum,
 493                            u64 offset, u64 len, bool insert,
 494                            s64 *i_sectors_delta)
 495 {
 496         struct bkey_i_logged_op_finsert op;
 497         s64 shift = insert ? len : -len;
 498
 499         bkey_logged_op_finsert_init(&op.k_i);
 500         op.v.subvol     = cpu_to_le32(inum.subvol);
 501         op.v.inum       = cpu_to_le64(inum.inum);
 502         op.v.dst_offset = cpu_to_le64(offset + shift);
 503         op.v.src_offset = cpu_to_le64(offset);
 504         op.v.pos        = cpu_to_le64(insert ? U64_MAX : offset);
 505
 506         /*
 507          * Logged ops aren't atomic w.r.t. snapshot creation: creating a
 508          * snapshot while they're in progress, then crashing, will result in the
 509          * resume only proceeding in one of the snapshots
 510          */
 511         down_read(&c->snapshot_create_lock);
 512         int ret = bch2_trans_run(c,
 513                 bch2_logged_op_start(trans, &op.k_i) ?:
 514                 __bch2_resume_logged_op_finsert(trans, &op.k_i, i_sectors_delta));
 515         up_read(&c->snapshot_create_lock);
 516
 517         return ret;
 518 }