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