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(trans, 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 int __bch2_btree_write_buffer_flush(struct btree_trans *trans, unsigned commit_flags,
 107                                     bool locked)
 108 {
 109         struct bch_fs *c = trans->c;
 110         struct journal *j = &c->journal;
 111         struct btree_write_buffer *wb = &c->btree_write_buffer;
 112         struct journal_entry_pin pin;
 113         struct btree_write_buffered_key *i, *keys;
 114         struct btree_iter iter = { NULL };
 115         size_t nr = 0, skipped = 0, fast = 0, slowpath = 0;
 116         bool write_locked = false;
 117         union btree_write_buffer_state s;
 118         int ret = 0;
 119
 120         memset(&pin, 0, sizeof(pin));
 121
 122         if (!locked && !mutex_trylock(&wb->flush_lock))
 123                 return 0;
 124
 125         bch2_journal_pin_copy(j, &pin, &wb->journal_pin, NULL);
 126         bch2_journal_pin_drop(j, &wb->journal_pin);
 127
 128         s = btree_write_buffer_switch(wb);
 129         keys = wb->keys[s.idx];
 130         nr = s.nr;
 131
 132         if (race_fault())
 133                 goto slowpath;
 134
 135         /*
 136          * We first sort so that we can detect and skip redundant updates, and
 137          * then we attempt to flush in sorted btree order, as this is most
 138          * efficient.
 139          *
 140          * However, since we're not flushing in the order they appear in the
 141          * journal we won't be able to drop our journal pin until everything is
 142          * flushed - which means this could deadlock the journal if we weren't
 143          * passing BTREE_INSERT_JOURNAL_RECLAIM. This causes the update to fail
 144          * if it would block taking a journal reservation.
 145          *
 146          * If that happens, simply skip the key so we can optimistically insert
 147          * as many keys as possible in the fast path.
 148          */
 149         sort(keys, nr, sizeof(keys[0]),
 150              btree_write_buffered_key_cmp, NULL);
 151
 152         for (i = keys; i < keys + nr; i++) {
 153                 if (i + 1 < keys + nr &&
 154                     i[0].btree == i[1].btree &&
 155                     bpos_eq(i[0].k.k.p, i[1].k.k.p)) {
 156                         skipped++;
 157                         i->journal_seq = 0;
 158                         continue;
 159                 }
 160
 161                 if (write_locked &&
 162                     (iter.path->btree_id != i->btree ||
 163                      bpos_gt(i->k.k.p, iter.path->l[0].b->key.k.p))) {
 164                         bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 165                         write_locked = false;
 166                 }
 167
 168                 if (!iter.path || iter.path->btree_id != i->btree) {
 169                         bch2_trans_iter_exit(trans, &iter);
 170                         bch2_trans_iter_init(trans, &iter, i->btree, i->k.k.p, BTREE_ITER_INTENT);
 171                 }
 172
 173                 bch2_btree_iter_set_pos(&iter, i->k.k.p);
 174                 iter.path->preserve = false;
 175
 176                 do {
 177                         ret = bch2_btree_write_buffer_flush_one(trans, &iter, i,
 178                                                 commit_flags, &write_locked, &fast);
 179                         if (!write_locked)
 180                                 bch2_trans_begin(trans);
 181                 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
 182
 183                 if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
 184                         slowpath++;
 185                         continue;
 186                 }
 187                 if (ret)
 188                         break;
 189
 190                 i->journal_seq = 0;
 191         }
 192
 193         if (write_locked)
 194                 bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 195         bch2_trans_iter_exit(trans, &iter);
 196
 197         trace_write_buffer_flush(trans, nr, skipped, fast, wb->size);
 198
 199         if (slowpath)
 200                 goto slowpath;
 201
 202         bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret));
 203 out:
 204         bch2_journal_pin_drop(j, &pin);
 205         mutex_unlock(&wb->flush_lock);
 206         return ret;
 207 slowpath:
 208         trace_write_buffer_flush_slowpath(trans, i - keys, nr);
 209
 210         /*
 211          * Now sort the rest by journal seq and bump the journal pin as we go.
 212          * The slowpath zapped the seq of keys that were successfully flushed so
 213          * we can skip those here.
 214          */
 215         sort(keys, nr, sizeof(keys[0]),
 216              btree_write_buffered_journal_cmp,
 217              NULL);
 218
 219         commit_flags &= ~BCH_WATERMARK_MASK;
 220         commit_flags |= BCH_WATERMARK_reclaim;
 221
 222         for (i = keys; i < keys + nr; i++) {
 223                 if (!i->journal_seq)
 224                         continue;
 225
 226                 if (i->journal_seq > pin.seq) {
 227                         struct journal_entry_pin pin2;
 228
 229                         memset(&pin2, 0, sizeof(pin2));
 230
 231                         bch2_journal_pin_add(j, i->journal_seq, &pin2, NULL);
 232                         bch2_journal_pin_drop(j, &pin);
 233                         bch2_journal_pin_copy(j, &pin, &pin2, NULL);
 234                         bch2_journal_pin_drop(j, &pin2);
 235                 }
 236
 237                 ret = commit_do(trans, NULL, NULL,
 238                                 commit_flags|
 239                                 BTREE_INSERT_NOFAIL|
 240                                 BTREE_INSERT_JOURNAL_RECLAIM,
 241                                 __bch2_btree_insert(trans, i->btree, &i->k, 0));
 242                 if (bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret)))
 243                         break;
 244         }
 245
 246         goto out;
 247 }
 248
 249 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
 250 {
 251         bch2_trans_unlock(trans);
 252         mutex_lock(&trans->c->btree_write_buffer.flush_lock);
 253         return __bch2_btree_write_buffer_flush(trans, 0, true);
 254 }
 255
 256 int bch2_btree_write_buffer_flush(struct btree_trans *trans)
 257 {
 258         return __bch2_btree_write_buffer_flush(trans, 0, false);
 259 }
 260
 261 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
 262                                 struct journal_entry_pin *_pin, u64 seq)
 263 {
 264         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 265         struct btree_write_buffer *wb = &c->btree_write_buffer;
 266
 267         mutex_lock(&wb->flush_lock);
 268
 269         return bch2_trans_run(c,
 270                         __bch2_btree_write_buffer_flush(&trans, BTREE_INSERT_NOCHECK_RW, true));
 271 }
 272
 273 static inline u64 btree_write_buffer_ref(int idx)
 274 {
 275         return ((union btree_write_buffer_state) {
 276                 .ref0 = idx == 0,
 277                 .ref1 = idx == 1,
 278         }).v;
 279 }
 280
 281 int bch2_btree_insert_keys_write_buffer(struct btree_trans *trans)
 282 {
 283         struct bch_fs *c = trans->c;
 284         struct btree_write_buffer *wb = &c->btree_write_buffer;
 285         struct btree_write_buffered_key *i;
 286         union btree_write_buffer_state old, new;
 287         int ret = 0;
 288         u64 v;
 289
 290         trans_for_each_wb_update(trans, i) {
 291                 EBUG_ON(i->k.k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
 292
 293                 i->journal_seq          = trans->journal_res.seq;
 294                 i->journal_offset       = trans->journal_res.offset;
 295         }
 296
 297         preempt_disable();
 298         v = READ_ONCE(wb->state.v);
 299         do {
 300                 old.v = new.v = v;
 301
 302                 new.v += btree_write_buffer_ref(new.idx);
 303                 new.nr += trans->nr_wb_updates;
 304                 if (new.nr > wb->size) {
 305                         ret = -BCH_ERR_btree_insert_need_flush_buffer;
 306                         goto out;
 307                 }
 308         } while ((v = atomic64_cmpxchg_acquire(&wb->state.counter, old.v, new.v)) != old.v);
 309
 310         memcpy(wb->keys[new.idx] + old.nr,
 311                trans->wb_updates,
 312                sizeof(trans->wb_updates[0]) * trans->nr_wb_updates);
 313
 314         bch2_journal_pin_add(&c->journal, trans->journal_res.seq, &wb->journal_pin,
 315                              bch2_btree_write_buffer_journal_flush);
 316
 317         atomic64_sub_return_release(btree_write_buffer_ref(new.idx), &wb->state.counter);
 318 out:
 319         preempt_enable();
 320         return ret;
 321 }
 322
 323 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
 324 {
 325         struct btree_write_buffer *wb = &c->btree_write_buffer;
 326
 327         BUG_ON(wb->state.nr && !bch2_journal_error(&c->journal));
 328
 329         kvfree(wb->keys[1]);
 330         kvfree(wb->keys[0]);
 331 }
 332
 333 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
 334 {
 335         struct btree_write_buffer *wb = &c->btree_write_buffer;
 336
 337         mutex_init(&wb->flush_lock);
 338         wb->size = c->opts.btree_write_buffer_size;
 339
 340         wb->keys[0] = kvmalloc_array(wb->size, sizeof(*wb->keys[0]), GFP_KERNEL);
 341         wb->keys[1] = kvmalloc_array(wb->size, sizeof(*wb->keys[1]), GFP_KERNEL);
 342         if (!wb->keys[0] || !wb->keys[1])
 343                 return -BCH_ERR_ENOMEM_fs_btree_write_buffer_init;
 344
 345         return 0;
 346 }