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         return 0;
  68 trans_commit:
  69         return  bch2_trans_update(trans, iter, &wb->k, 0) ?:
  70                 bch2_trans_commit(trans, NULL, NULL,
  71                                   commit_flags|
  72                                   BTREE_INSERT_NOFAIL|
  73                                   BTREE_INSERT_JOURNAL_RECLAIM);
  74 }
  75
  76 static union btree_write_buffer_state btree_write_buffer_switch(struct btree_write_buffer *wb)
  77 {
  78         union btree_write_buffer_state old, new;
  79         u64 v = READ_ONCE(wb->state.v);
  80
  81         do {
  82                 old.v = new.v = v;
  83
  84                 new.nr = 0;
  85                 new.idx++;
  86         } while ((v = atomic64_cmpxchg_acquire(&wb->state.counter, old.v, new.v)) != old.v);
  87
  88         while (old.idx == 0 ? wb->state.ref0 : wb->state.ref1)
  89                 cpu_relax();
  90
  91         return old;
  92 }
  93
  94 int __bch2_btree_write_buffer_flush(struct btree_trans *trans, unsigned commit_flags,
  95                                     bool locked)
  96 {
  97         struct bch_fs *c = trans->c;
  98         struct journal *j = &c->journal;
  99         struct btree_write_buffer *wb = &c->btree_write_buffer;
 100         struct journal_entry_pin pin;
 101         struct btree_write_buffered_key *i, *dst, *keys;
 102         struct btree_iter iter = { NULL };
 103         size_t nr = 0, skipped = 0, fast = 0;
 104         bool write_locked = false;
 105         union btree_write_buffer_state s;
 106         int ret = 0;
 107
 108         memset(&pin, 0, sizeof(pin));
 109
 110         if (!locked && !mutex_trylock(&wb->flush_lock))
 111                 return 0;
 112
 113         bch2_journal_pin_copy(j, &pin, &wb->journal_pin, NULL);
 114         bch2_journal_pin_drop(j, &wb->journal_pin);
 115
 116         s = btree_write_buffer_switch(wb);
 117         keys = wb->keys[s.idx];
 118         nr = s.nr;
 119
 120         /*
 121          * We first sort so that we can detect and skip redundant updates, and
 122          * then we attempt to flush in sorted btree order, as this is most
 123          * efficient.
 124          *
 125          * However, since we're not flushing in the order they appear in the
 126          * journal we won't be able to drop our journal pin until everything is
 127          * flushed - which means this could deadlock the journal, if we weren't
 128          * passing BTREE_INSERT_JORUNAL_RECLAIM. This causes the update to fail
 129          * if it would block taking a journal reservation.
 130          *
 131          * If that happens, we sort them by the order they appeared in the
 132          * journal - after dropping redundant entries - and then restart
 133          * flushing, this time dropping journal pins as we go.
 134          */
 135
 136         sort(keys, nr, sizeof(keys[0]),
 137              btree_write_buffered_key_cmp, NULL);
 138
 139         for (i = keys; i < keys + nr; i++) {
 140                 if (i + 1 < keys + nr &&
 141                     i[0].btree == i[1].btree &&
 142                     bpos_eq(i[0].k.k.p, i[1].k.k.p)) {
 143                         skipped++;
 144                         continue;
 145                 }
 146
 147                 if (write_locked &&
 148                     (iter.path->btree_id != i->btree ||
 149                      bpos_gt(i->k.k.p, iter.path->l[0].b->key.k.p))) {
 150                         bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 151                         write_locked = false;
 152                 }
 153
 154                 if (!iter.path || iter.path->btree_id != i->btree) {
 155                         bch2_trans_iter_exit(trans, &iter);
 156                         bch2_trans_iter_init(trans, &iter, i->btree, i->k.k.p, BTREE_ITER_INTENT);
 157                 }
 158
 159                 bch2_btree_iter_set_pos(&iter, i->k.k.p);
 160                 iter.path->preserve = false;
 161
 162                 do {
 163                         ret = bch2_btree_write_buffer_flush_one(trans, &iter, i,
 164                                                 commit_flags, &write_locked, &fast);
 165                         if (!write_locked)
 166                                 bch2_trans_begin(trans);
 167                 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
 168
 169                 if (ret)
 170                         break;
 171         }
 172
 173         if (write_locked)
 174                 bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 175         bch2_trans_iter_exit(trans, &iter);
 176
 177         trace_write_buffer_flush(trans, nr, skipped, fast, wb->size);
 178
 179         if (ret == -BCH_ERR_journal_reclaim_would_deadlock)
 180                 goto slowpath;
 181
 182         bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret));
 183 out:
 184         bch2_journal_pin_drop(j, &pin);
 185         mutex_unlock(&wb->flush_lock);
 186         return ret;
 187 slowpath:
 188         trace_write_buffer_flush_slowpath(trans, i - keys, nr);
 189
 190         dst = keys;
 191         for (; i < keys + nr; i++) {
 192                 if (i + 1 < keys + nr &&
 193                     i[0].btree == i[1].btree &&
 194                     bpos_eq(i[0].k.k.p, i[1].k.k.p))
 195                         continue;
 196
 197                 *dst = *i;
 198                 dst++;
 199         }
 200         nr = dst - keys;
 201
 202         sort(keys, nr, sizeof(keys[0]),
 203              btree_write_buffered_journal_cmp,
 204              NULL);
 205
 206         for (i = keys; i < keys + nr; i++) {
 207                 if (i->journal_seq > pin.seq) {
 208                         struct journal_entry_pin pin2;
 209
 210                         memset(&pin2, 0, sizeof(pin2));
 211
 212                         bch2_journal_pin_add(j, i->journal_seq, &pin2, NULL);
 213                         bch2_journal_pin_drop(j, &pin);
 214                         bch2_journal_pin_copy(j, &pin, &pin2, NULL);
 215                         bch2_journal_pin_drop(j, &pin2);
 216                 }
 217
 218                 ret = commit_do(trans, NULL, NULL,
 219                                 commit_flags|
 220                                 BTREE_INSERT_NOFAIL|
 221                                 BTREE_INSERT_JOURNAL_RECLAIM|
 222                                 JOURNAL_WATERMARK_reserved,
 223                                 __bch2_btree_insert(trans, i->btree, &i->k, 0));
 224                 if (bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret)))
 225                         break;
 226         }
 227
 228         goto out;
 229 }
 230
 231 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
 232 {
 233         bch2_trans_unlock(trans);
 234         mutex_lock(&trans->c->btree_write_buffer.flush_lock);
 235         return __bch2_btree_write_buffer_flush(trans, 0, true);
 236 }
 237
 238 int bch2_btree_write_buffer_flush(struct btree_trans *trans)
 239 {
 240         return __bch2_btree_write_buffer_flush(trans, 0, false);
 241 }
 242
 243 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
 244                                 struct journal_entry_pin *_pin, u64 seq)
 245 {
 246         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 247         struct btree_write_buffer *wb = &c->btree_write_buffer;
 248
 249         mutex_lock(&wb->flush_lock);
 250
 251         return bch2_trans_run(c,
 252                         __bch2_btree_write_buffer_flush(&trans, BTREE_INSERT_NOCHECK_RW, true));
 253 }
 254
 255 static inline u64 btree_write_buffer_ref(int idx)
 256 {
 257         return ((union btree_write_buffer_state) {
 258                 .ref0 = idx == 0,
 259                 .ref1 = idx == 1,
 260         }).v;
 261 }
 262
 263 int bch2_btree_insert_keys_write_buffer(struct btree_trans *trans)
 264 {
 265         struct bch_fs *c = trans->c;
 266         struct btree_write_buffer *wb = &c->btree_write_buffer;
 267         struct btree_write_buffered_key *i;
 268         union btree_write_buffer_state old, new;
 269         int ret = 0;
 270         u64 v;
 271
 272         trans_for_each_wb_update(trans, i) {
 273                 EBUG_ON(i->k.k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
 274
 275                 i->journal_seq          = trans->journal_res.seq;
 276                 i->journal_offset       = trans->journal_res.offset;
 277         }
 278
 279         preempt_disable();
 280         v = READ_ONCE(wb->state.v);
 281         do {
 282                 old.v = new.v = v;
 283
 284                 new.v += btree_write_buffer_ref(new.idx);
 285                 new.nr += trans->nr_wb_updates;
 286                 if (new.nr > wb->size) {
 287                         ret = -BCH_ERR_btree_insert_need_flush_buffer;
 288                         goto out;
 289                 }
 290         } while ((v = atomic64_cmpxchg_acquire(&wb->state.counter, old.v, new.v)) != old.v);
 291
 292         memcpy(wb->keys[new.idx] + old.nr,
 293                trans->wb_updates,
 294                sizeof(trans->wb_updates[0]) * trans->nr_wb_updates);
 295
 296         bch2_journal_pin_add(&c->journal, trans->journal_res.seq, &wb->journal_pin,
 297                              bch2_btree_write_buffer_journal_flush);
 298
 299         atomic64_sub_return_release(btree_write_buffer_ref(new.idx), &wb->state.counter);
 300 out:
 301         preempt_enable();
 302         return ret;
 303 }
 304
 305 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
 306 {
 307         struct btree_write_buffer *wb = &c->btree_write_buffer;
 308
 309         BUG_ON(wb->state.nr && !bch2_journal_error(&c->journal));
 310
 311         kvfree(wb->keys[1]);
 312         kvfree(wb->keys[0]);
 313 }
 314
 315 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
 316 {
 317         struct btree_write_buffer *wb = &c->btree_write_buffer;
 318
 319         mutex_init(&wb->flush_lock);
 320         wb->size = c->opts.btree_write_buffer_size;
 321
 322         wb->keys[0] = kvmalloc_array(wb->size, sizeof(*wb->keys[0]), GFP_KERNEL);
 323         wb->keys[1] = kvmalloc_array(wb->size, sizeof(*wb->keys[1]), GFP_KERNEL);
 324         if (!wb->keys[0] || !wb->keys[1])
 325                 return -ENOMEM;
 326
 327         return 0;
 328 }