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 bch_write_op *op)
 101 {
 102         struct bch_fs *c = op->c;
 103         struct btree_trans trans;
 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_init(&trans, c, BTREE_ITER_MAX, 1024);
 116
 117         bch2_trans_iter_init(&trans, &iter, m->btree_id,
 118                              bkey_start_pos(&bch2_keylist_front(keys)->k),
 119                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 120
 121         while (1) {
 122                 struct bkey_s_c k;
 123                 struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
 124                 struct bkey_i *insert;
 125                 struct bkey_i_extent *new;
 126                 const union bch_extent_entry *entry;
 127                 struct extent_ptr_decoded p;
 128                 struct bpos next_pos;
 129                 bool did_work = false;
 130                 bool should_check_enospc;
 131                 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 132                 unsigned i;
 133
 134                 bch2_trans_begin(&trans);
 135
 136                 k = bch2_btree_iter_peek_slot(&iter);
 137                 ret = bkey_err(k);
 138                 if (ret)
 139                         goto err;
 140
 141                 new = bkey_i_to_extent(bch2_keylist_front(keys));
 142
 143                 if (!bch2_extents_match(k, old))
 144                         goto nomatch;
 145
 146                 bkey_reassemble(_insert.k, k);
 147                 insert = _insert.k;
 148
 149                 bch2_bkey_buf_copy(&_new, c, bch2_keylist_front(keys));
 150                 new = bkey_i_to_extent(_new.k);
 151                 bch2_cut_front(iter.pos, &new->k_i);
 152
 153                 bch2_cut_front(iter.pos,        insert);
 154                 bch2_cut_back(new->k.p,         insert);
 155                 bch2_cut_back(insert->k.p,      &new->k_i);
 156
 157                 /*
 158                  * @old: extent that we read from
 159                  * @insert: key that we're going to update, initialized from
 160                  * extent currently in btree - same as @old unless we raced with
 161                  * other updates
 162                  * @new: extent with new pointers that we'll be adding to @insert
 163                  *
 164                  * Fist, drop rewrite_ptrs from @new:
 165                  */
 166                 i = 0;
 167                 bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
 168                         if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 169                             bch2_extent_has_ptr(old, p, bkey_i_to_s_c(insert))) {
 170                                 /*
 171                                  * If we're going to be adding a pointer to the
 172                                  * same device, we have to drop the old one -
 173                                  * otherwise, we can just mark it cached:
 174                                  */
 175                                 if (bch2_bkey_has_device(bkey_i_to_s_c(&new->k_i), p.ptr.dev))
 176                                         bch2_bkey_drop_device_noerror(bkey_i_to_s(insert), p.ptr.dev);
 177                                 else
 178                                         bch2_bkey_mark_dev_cached(bkey_i_to_s(insert), p.ptr.dev);
 179                         }
 180                         i++;
 181                 }
 182
 183
 184                 /* Add new ptrs: */
 185                 extent_for_each_ptr_decode(extent_i_to_s(new), p, entry) {
 186                         if (bch2_bkey_has_device(bkey_i_to_s_c(insert), p.ptr.dev)) {
 187                                 /*
 188                                  * raced with another move op? extent already
 189                                  * has a pointer to the device we just wrote
 190                                  * data to
 191                                  */
 192                                 continue;
 193                         }
 194
 195                         bch2_extent_ptr_decoded_append(insert, &p);
 196                         did_work = true;
 197                 }
 198
 199                 if (!did_work)
 200                         goto nomatch;
 201
 202                 bch2_bkey_narrow_crcs(insert, (struct bch_extent_crc_unpacked) { 0 });
 203                 bch2_extent_normalize(c, bkey_i_to_s(insert));
 204
 205                 ret = bch2_sum_sector_overwrites(&trans, &iter, insert,
 206                                                  &should_check_enospc,
 207                                                  &i_sectors_delta,
 208                                                  &disk_sectors_delta);
 209                 if (ret)
 210                         goto err;
 211
 212                 if (disk_sectors_delta > (s64) op->res.sectors) {
 213                         ret = bch2_disk_reservation_add(c, &op->res,
 214                                                 disk_sectors_delta - op->res.sectors,
 215                                                 !should_check_enospc
 216                                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 217                         if (ret)
 218                                 goto out;
 219                 }
 220
 221                 next_pos = insert->k.p;
 222
 223                 ret   = insert_snapshot_whiteouts(&trans, m->btree_id,
 224                                                   k.k->p, insert->k.p) ?:
 225                         bch2_trans_update(&trans, &iter, insert,
 226                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 227                         bch2_trans_commit(&trans, &op->res,
 228                                 op_journal_seq(op),
 229                                 BTREE_INSERT_NOFAIL|
 230                                 m->data_opts.btree_insert_flags);
 231                 if (!ret) {
 232                         bch2_btree_iter_set_pos(&iter, next_pos);
 233
 234                         this_cpu_add(c->counters[BCH_COUNTER_move_extent_finish], new->k.size);
 235                         trace_move_extent_finish(&new->k);
 236                 }
 237 err:
 238                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 239                         ret = 0;
 240                 if (ret)
 241                         break;
 242 next:
 243                 while (bkey_cmp(iter.pos, bch2_keylist_front(keys)->k.p) >= 0) {
 244                         bch2_keylist_pop_front(keys);
 245                         if (bch2_keylist_empty(keys))
 246                                 goto out;
 247                 }
 248                 continue;
 249 nomatch:
 250                 if (m->ctxt) {
 251                         BUG_ON(k.k->p.offset <= iter.pos.offset);
 252                         atomic64_inc(&m->ctxt->stats->keys_raced);
 253                         atomic64_add(k.k->p.offset - iter.pos.offset,
 254                                      &m->ctxt->stats->sectors_raced);
 255                 }
 256
 257                 this_cpu_add(c->counters[BCH_COUNTER_move_extent_race], new->k.size);
 258                 trace_move_extent_race(&new->k);
 259
 260                 bch2_btree_iter_advance(&iter);
 261                 goto next;
 262         }
 263 out:
 264         bch2_trans_iter_exit(&trans, &iter);
 265         bch2_trans_exit(&trans);
 266         bch2_bkey_buf_exit(&_insert, c);
 267         bch2_bkey_buf_exit(&_new, c);
 268         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 269         return ret;
 270 }
 271
 272 void bch2_data_update_read_done(struct data_update *m,
 273                                 struct bch_extent_crc_unpacked crc,
 274                                 struct closure *cl)
 275 {
 276         /* write bio must own pages: */
 277         BUG_ON(!m->op.wbio.bio.bi_vcnt);
 278
 279         m->op.crc = crc;
 280         m->op.wbio.bio.bi_iter.bi_size = crc.compressed_size << 9;
 281
 282         closure_call(&m->op.cl, bch2_write, NULL, cl);
 283 }
 284
 285 void bch2_data_update_exit(struct data_update *update)
 286 {
 287         struct bch_fs *c = update->op.c;
 288
 289         bch2_bkey_buf_exit(&update->k, c);
 290         bch2_disk_reservation_put(c, &update->op.res);
 291         bch2_bio_free_pages_pool(c, &update->op.wbio.bio);
 292 }
 293
 294 int bch2_data_update_init(struct bch_fs *c, struct data_update *m,
 295                           struct write_point_specifier wp,
 296                           struct bch_io_opts io_opts,
 297                           struct data_update_opts data_opts,
 298                           enum btree_id btree_id,
 299                           struct bkey_s_c k)
 300 {
 301         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 302         const union bch_extent_entry *entry;
 303         struct extent_ptr_decoded p;
 304         unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
 305         int ret;
 306
 307         bch2_bkey_buf_init(&m->k);
 308         bch2_bkey_buf_reassemble(&m->k, c, k);
 309         m->btree_id     = btree_id;
 310         m->data_opts    = data_opts;
 311
 312         bch2_write_op_init(&m->op, c, io_opts);
 313         m->op.pos       = bkey_start_pos(k.k);
 314         m->op.version   = k.k->version;
 315         m->op.target    = data_opts.target;
 316         m->op.write_point = wp;
 317         m->op.flags     |= BCH_WRITE_PAGES_STABLE|
 318                 BCH_WRITE_PAGES_OWNED|
 319                 BCH_WRITE_DATA_ENCODED|
 320                 BCH_WRITE_FROM_INTERNAL|
 321                 m->data_opts.write_flags;
 322         m->op.compression_type =
 323                 bch2_compression_opt_to_type[io_opts.background_compression ?:
 324                                              io_opts.compression];
 325         if (m->data_opts.btree_insert_flags & BTREE_INSERT_USE_RESERVE)
 326                 m->op.alloc_reserve = RESERVE_movinggc;
 327         m->op.index_update_fn   = bch2_data_update_index_update;
 328
 329         i = 0;
 330         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 331                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 332                     p.ptr.cached)
 333                         BUG();
 334
 335                 if (!((1U << i) & m->data_opts.rewrite_ptrs))
 336                         bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 337
 338                 if (((1U << i) & m->data_opts.rewrite_ptrs) &&
 339                     crc_is_compressed(p.crc))
 340                         reserve_sectors += k.k->size;
 341
 342                 /*
 343                  * op->csum_type is normally initialized from the fs/file's
 344                  * current options - but if an extent is encrypted, we require
 345                  * that it stays encrypted:
 346                  */
 347                 if (bch2_csum_type_is_encryption(p.crc.csum_type)) {
 348                         m->op.nonce     = p.crc.nonce + p.crc.offset;
 349                         m->op.csum_type = p.crc.csum_type;
 350                 }
 351
 352                 if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 353                         m->op.incompressible = true;
 354
 355                 i++;
 356         }
 357
 358         if (reserve_sectors) {
 359                 ret = bch2_disk_reservation_add(c, &m->op.res, reserve_sectors,
 360                                 m->data_opts.extra_replicas
 361                                 ? 0
 362                                 : BCH_DISK_RESERVATION_NOFAIL);
 363                 if (ret)
 364                         return ret;
 365         }
 366
 367         m->op.nr_replicas = m->op.nr_replicas_required =
 368                 hweight32(m->data_opts.rewrite_ptrs) + m->data_opts.extra_replicas;
 369
 370         BUG_ON(!m->op.nr_replicas);
 371         return 0;
 372 }
 373
 374 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 375 {
 376         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 377         const struct bch_extent_ptr *ptr;
 378         unsigned i = 0;
 379
 380         bkey_for_each_ptr(ptrs, ptr) {
 381                 if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
 382                         opts->kill_ptrs |= 1U << i;
 383                         opts->rewrite_ptrs ^= 1U << i;
 384                 }
 385
 386                 i++;
 387         }
 388 }