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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_locking.h"
   5 #include "btree_update.h"
   6 #include "btree_update_interior.h"
   7 #include "btree_write_buffer.h"
   8 #include "error.h"
   9 #include "journal.h"
  10 #include "journal_reclaim.h"
  11
  12 #include <linux/sort.h>
  13
  14 static int btree_write_buffered_key_cmp(const void *_l, const void *_r)
  15 {
  16         const struct btree_write_buffered_key *l = _l;
  17         const struct btree_write_buffered_key *r = _r;
  18
  19         return  cmp_int(l->btree, r->btree) ?:
  20                 bpos_cmp(l->k.k.p, r->k.k.p) ?:
  21                 cmp_int(l->journal_seq, r->journal_seq) ?:
  22                 cmp_int(l->journal_offset, r->journal_offset);
  23 }
  24
  25 static int btree_write_buffered_journal_cmp(const void *_l, const void *_r)
  26 {
  27         const struct btree_write_buffered_key *l = _l;
  28         const struct btree_write_buffered_key *r = _r;
  29
  30         return  cmp_int(l->journal_seq, r->journal_seq);
  31 }
  32
  33 static int bch2_btree_write_buffer_flush_one(struct btree_trans *trans,
  34                                              struct btree_iter *iter,
  35                                              struct btree_write_buffered_key *wb,
  36                                              unsigned commit_flags,
  37                                              bool *write_locked,
  38                                              size_t *fast)
  39 {
  40         struct bch_fs *c = trans->c;
  41         struct btree_path *path;
  42         int ret;
  43
  44         ret = bch2_btree_iter_traverse(iter);
  45         if (ret)
  46                 return ret;
  47
  48         path = iter->path;
  49
  50         if (!*write_locked) {
  51                 ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
  52                 if (ret)
  53                         return ret;
  54
  55                 bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
  56                 *write_locked = true;
  57         }
  58
  59         if (!bch2_btree_node_insert_fits(c, path->l[0].b, wb->k.k.u64s)) {
  60                 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
  61                 *write_locked = false;
  62                 goto trans_commit;
  63         }
  64
  65         bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
  66         (*fast)++;
  67
  68         if (path->ref > 1) {
  69                 /*
  70                  * We can't clone a path that has write locks: if the path is
  71                  * shared, unlock before set_pos(), traverse():
  72                  */
  73                 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
  74                 *write_locked = false;
  75         }
  76         return 0;
  77 trans_commit:
  78         return  bch2_trans_update_seq(trans, wb->journal_seq, iter, &wb->k, 0) ?:
  79                 bch2_trans_commit(trans, NULL, NULL,
  80                                   commit_flags|
  81                                   BTREE_INSERT_NOCHECK_RW|
  82                                   BTREE_INSERT_NOFAIL|
  83                                   BTREE_INSERT_JOURNAL_RECLAIM);
  84 }
  85
  86 static union btree_write_buffer_state btree_write_buffer_switch(struct btree_write_buffer *wb)
  87 {
  88         union btree_write_buffer_state old, new;
  89         u64 v = READ_ONCE(wb->state.v);
  90
  91         do {
  92                 old.v = new.v = v;
  93
  94                 new.nr = 0;
  95                 new.idx++;
  96         } while ((v = atomic64_cmpxchg_acquire(&wb->state.counter, old.v, new.v)) != old.v);
  97
  98         while (old.idx == 0 ? wb->state.ref0 : wb->state.ref1)
  99                 cpu_relax();
 100
 101         smp_mb();
 102
 103         return old;
 104 }
 105
 106 /*
 107  * Update a btree with a write buffered key using the journal seq of the
 108  * original write buffer insert.
 109  *
 110  * It is not safe to rejournal the key once it has been inserted into the write
 111  * buffer because that may break recovery ordering. For example, the key may
 112  * have already been modified in the active write buffer in a seq that comes
 113  * before the current transaction. If we were to journal this key again and
 114  * crash, recovery would process updates in the wrong order.
 115  */
 116 static int
 117 btree_write_buffered_insert(struct btree_trans *trans,
 118                           struct btree_write_buffered_key *wb)
 119 {
 120         struct btree_iter iter;
 121         int ret;
 122
 123         bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
 124                              BTREE_ITER_CACHED|BTREE_ITER_INTENT);
 125
 126         ret   = bch2_btree_iter_traverse(&iter) ?:
 127                 bch2_trans_update_seq(trans, wb->journal_seq, &iter, &wb->k, 0);
 128         bch2_trans_iter_exit(trans, &iter);
 129         return ret;
 130 }
 131
 132 int __bch2_btree_write_buffer_flush(struct btree_trans *trans, unsigned commit_flags,
 133                                     bool locked)
 134 {
 135         struct bch_fs *c = trans->c;
 136         struct journal *j = &c->journal;
 137         struct btree_write_buffer *wb = &c->btree_write_buffer;
 138         struct journal_entry_pin pin;
 139         struct btree_write_buffered_key *i, *keys;
 140         struct btree_iter iter = { NULL };
 141         size_t nr = 0, skipped = 0, fast = 0, slowpath = 0;
 142         bool write_locked = false;
 143         union btree_write_buffer_state s;
 144         int ret = 0;
 145
 146         memset(&pin, 0, sizeof(pin));
 147
 148         if (!locked && !mutex_trylock(&wb->flush_lock))
 149                 return 0;
 150
 151         bch2_journal_pin_copy(j, &pin, &wb->journal_pin, NULL);
 152         bch2_journal_pin_drop(j, &wb->journal_pin);
 153
 154         s = btree_write_buffer_switch(wb);
 155         keys = wb->keys[s.idx];
 156         nr = s.nr;
 157
 158         if (race_fault())
 159                 goto slowpath;
 160
 161         /*
 162          * We first sort so that we can detect and skip redundant updates, and
 163          * then we attempt to flush in sorted btree order, as this is most
 164          * efficient.
 165          *
 166          * However, since we're not flushing in the order they appear in the
 167          * journal we won't be able to drop our journal pin until everything is
 168          * flushed - which means this could deadlock the journal if we weren't
 169          * passing BTREE_INSERT_JOURNAL_RECLAIM. This causes the update to fail
 170          * if it would block taking a journal reservation.
 171          *
 172          * If that happens, simply skip the key so we can optimistically insert
 173          * as many keys as possible in the fast path.
 174          */
 175         sort(keys, nr, sizeof(keys[0]),
 176              btree_write_buffered_key_cmp, NULL);
 177
 178         for (i = keys; i < keys + nr; i++) {
 179                 if (i + 1 < keys + nr &&
 180                     i[0].btree == i[1].btree &&
 181                     bpos_eq(i[0].k.k.p, i[1].k.k.p)) {
 182                         skipped++;
 183                         i->journal_seq = 0;
 184                         continue;
 185                 }
 186
 187                 if (write_locked &&
 188                     (iter.path->btree_id != i->btree ||
 189                      bpos_gt(i->k.k.p, iter.path->l[0].b->key.k.p))) {
 190                         bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 191                         write_locked = false;
 192                 }
 193
 194                 if (!iter.path || iter.path->btree_id != i->btree) {
 195                         bch2_trans_iter_exit(trans, &iter);
 196                         bch2_trans_iter_init(trans, &iter, i->btree, i->k.k.p, BTREE_ITER_INTENT);
 197                 }
 198
 199                 bch2_btree_iter_set_pos(&iter, i->k.k.p);
 200                 iter.path->preserve = false;
 201
 202                 do {
 203                         ret = bch2_btree_write_buffer_flush_one(trans, &iter, i,
 204                                                 commit_flags, &write_locked, &fast);
 205                         if (!write_locked)
 206                                 bch2_trans_begin(trans);
 207                 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
 208
 209                 if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
 210                         slowpath++;
 211                         continue;
 212                 }
 213                 if (ret)
 214                         break;
 215
 216                 i->journal_seq = 0;
 217         }
 218
 219         if (write_locked)
 220                 bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 221         bch2_trans_iter_exit(trans, &iter);
 222
 223         trace_write_buffer_flush(trans, nr, skipped, fast, wb->size);
 224
 225         if (slowpath)
 226                 goto slowpath;
 227
 228         bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret));
 229 out:
 230         bch2_journal_pin_drop(j, &pin);
 231         mutex_unlock(&wb->flush_lock);
 232         return ret;
 233 slowpath:
 234         trace_write_buffer_flush_slowpath(trans, i - keys, nr);
 235
 236         /*
 237          * Now sort the rest by journal seq and bump the journal pin as we go.
 238          * The slowpath zapped the seq of keys that were successfully flushed so
 239          * we can skip those here.
 240          */
 241         sort(keys, nr, sizeof(keys[0]),
 242              btree_write_buffered_journal_cmp,
 243              NULL);
 244
 245         commit_flags &= ~BCH_WATERMARK_MASK;
 246         commit_flags |= BCH_WATERMARK_reclaim;
 247
 248         for (i = keys; i < keys + nr; i++) {
 249                 if (!i->journal_seq)
 250                         continue;
 251
 252                 if (i->journal_seq > pin.seq) {
 253                         struct journal_entry_pin pin2;
 254
 255                         memset(&pin2, 0, sizeof(pin2));
 256
 257                         bch2_journal_pin_add(j, i->journal_seq, &pin2, NULL);
 258                         bch2_journal_pin_drop(j, &pin);
 259                         bch2_journal_pin_copy(j, &pin, &pin2, NULL);
 260                         bch2_journal_pin_drop(j, &pin2);
 261                 }
 262
 263                 ret = commit_do(trans, NULL, NULL,
 264                                 commit_flags|
 265                                 BTREE_INSERT_NOFAIL|
 266                                 BTREE_INSERT_JOURNAL_RECLAIM,
 267                                 btree_write_buffered_insert(trans, i));
 268                 if (bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret)))
 269                         break;
 270         }
 271
 272         goto out;
 273 }
 274
 275 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
 276 {
 277         bch2_trans_unlock(trans);
 278         mutex_lock(&trans->c->btree_write_buffer.flush_lock);
 279         return __bch2_btree_write_buffer_flush(trans, 0, true);
 280 }
 281
 282 int bch2_btree_write_buffer_flush(struct btree_trans *trans)
 283 {
 284         return __bch2_btree_write_buffer_flush(trans, 0, false);
 285 }
 286
 287 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
 288                                 struct journal_entry_pin *_pin, u64 seq)
 289 {
 290         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 291         struct btree_write_buffer *wb = &c->btree_write_buffer;
 292
 293         mutex_lock(&wb->flush_lock);
 294
 295         return bch2_trans_run(c,
 296                         __bch2_btree_write_buffer_flush(&trans, BTREE_INSERT_NOCHECK_RW, true));
 297 }
 298
 299 static inline u64 btree_write_buffer_ref(int idx)
 300 {
 301         return ((union btree_write_buffer_state) {
 302                 .ref0 = idx == 0,
 303                 .ref1 = idx == 1,
 304         }).v;
 305 }
 306
 307 int bch2_btree_insert_keys_write_buffer(struct btree_trans *trans)
 308 {
 309         struct bch_fs *c = trans->c;
 310         struct btree_write_buffer *wb = &c->btree_write_buffer;
 311         struct btree_write_buffered_key *i;
 312         union btree_write_buffer_state old, new;
 313         int ret = 0;
 314         u64 v;
 315
 316         trans_for_each_wb_update(trans, i) {
 317                 EBUG_ON(i->k.k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
 318
 319                 i->journal_seq          = trans->journal_res.seq;
 320                 i->journal_offset       = trans->journal_res.offset;
 321         }
 322
 323         preempt_disable();
 324         v = READ_ONCE(wb->state.v);
 325         do {
 326                 old.v = new.v = v;
 327
 328                 new.v += btree_write_buffer_ref(new.idx);
 329                 new.nr += trans->nr_wb_updates;
 330                 if (new.nr > wb->size) {
 331                         ret = -BCH_ERR_btree_insert_need_flush_buffer;
 332                         goto out;
 333                 }
 334         } while ((v = atomic64_cmpxchg_acquire(&wb->state.counter, old.v, new.v)) != old.v);
 335
 336         memcpy(wb->keys[new.idx] + old.nr,
 337                trans->wb_updates,
 338                sizeof(trans->wb_updates[0]) * trans->nr_wb_updates);
 339
 340         bch2_journal_pin_add(&c->journal, trans->journal_res.seq, &wb->journal_pin,
 341                              bch2_btree_write_buffer_journal_flush);
 342
 343         atomic64_sub_return_release(btree_write_buffer_ref(new.idx), &wb->state.counter);
 344 out:
 345         preempt_enable();
 346         return ret;
 347 }
 348
 349 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
 350 {
 351         struct btree_write_buffer *wb = &c->btree_write_buffer;
 352
 353         BUG_ON(wb->state.nr && !bch2_journal_error(&c->journal));
 354
 355         kvfree(wb->keys[1]);
 356         kvfree(wb->keys[0]);
 357 }
 358
 359 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
 360 {
 361         struct btree_write_buffer *wb = &c->btree_write_buffer;
 362
 363         mutex_init(&wb->flush_lock);
 364         wb->size = c->opts.btree_write_buffer_size;
 365
 366         wb->keys[0] = kvmalloc_array(wb->size, sizeof(*wb->keys[0]), GFP_KERNEL);
 367         wb->keys[1] = kvmalloc_array(wb->size, sizeof(*wb->keys[1]), GFP_KERNEL);
 368         if (!wb->keys[0] || !wb->keys[1])
 369                 return -BCH_ERR_ENOMEM_fs_btree_write_buffer_init;
 370
 371         return 0;
 372 }