]> git.ozlabs.org Git - ccan/blobdiff - ccan/tdb2/free.c
tdb2: limit coalescing based on how successful we are.
[ccan] / ccan / tdb2 / free.c
index c03680cfd943557d883daeee50b5ede41b707e75..be9b18455342a76ee316f98eff4363ade141849e 100644 (file)
@@ -109,7 +109,7 @@ static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
        tdb_off_t off, prev = 0, first;
        struct tdb_free_record r;
 
-       first = off = tdb_read_off(tdb, b_off);
+       first = off = (tdb_read_off(tdb, b_off) & TDB_OFF_MASK);
        while (off != 0) {
                tdb_read_convert(tdb, off, &r, sizeof(r));
                if (frec_magic(&r) != TDB_FREE_MAGIC)
@@ -150,17 +150,21 @@ static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
 
        /* If prev->next == 0, we were head: update bucket to point to next. */
        if (prev_next == 0) {
-#ifdef CCAN_TDB2_DEBUG
-               if (tdb_read_off(tdb, b_off) != r_off) {
+               /* We must preserve upper bits. */
+               head = tdb_read_off(tdb, b_off);
+               if (TDB_OFF_IS_ERR(head))
+                       return head;
+
+               if ((head & TDB_OFF_MASK) != r_off) {
                        return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
                                          "remove_from_list:"
                                          " %llu head %llu on list %llu",
                                          (long long)r_off,
-                                         (long long)tdb_read_off(tdb, b_off),
+                                         (long long)head,
                                          (long long)b_off);
                }
-#endif
-               ecode = tdb_write_off(tdb, b_off, r->next);
+               head = ((head & ~TDB_OFF_MASK) | r->next);
+               ecode = tdb_write_off(tdb, b_off, head);
                if (ecode != TDB_SUCCESS)
                        return ecode;
        } else {
@@ -175,6 +179,7 @@ static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
                head = tdb_read_off(tdb, b_off);
                if (TDB_OFF_IS_ERR(head))
                        return head;
+               head &= TDB_OFF_MASK;
                off = head + offsetof(struct tdb_free_record, magic_and_prev);
        } else {
                /* off = &r->next->prev */
@@ -195,26 +200,29 @@ static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
        return tdb_write_off(tdb, off, r->magic_and_prev);
 }
 
-/* Enqueue in this free bucket. */
+/* Enqueue in this free bucket: sets coalesce if we've added 128
+ * entries to it. */
 static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
                                      tdb_off_t b_off,
                                      tdb_off_t off,
-                                     tdb_len_t len)
+                                     tdb_len_t len,
+                                     bool *coalesce)
 {
        struct tdb_free_record new;
        enum TDB_ERROR ecode;
-       tdb_off_t prev;
+       tdb_off_t prev, head;
        uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
 
+       head = tdb_read_off(tdb, b_off);
+       if (TDB_OFF_IS_ERR(head))
+               return head;
+
        /* We only need to set ftable_and_len; rest is set in enqueue_in_free */
        new.ftable_and_len = ((uint64_t)tdb->ftable << (64 - TDB_OFF_UPPER_STEAL))
                | len;
 
        /* new->next = head. */
-       new.next = tdb_read_off(tdb, b_off);
-       if (TDB_OFF_IS_ERR(new.next)) {
-               return new.next;
-       }
+       new.next = (head & TDB_OFF_MASK);
 
        /* First element?  Prev points to ourselves. */
        if (!new.next) {
@@ -255,64 +263,23 @@ static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
                }
 #endif
        }
-       /* head = new */
-       ecode = tdb_write_off(tdb, b_off, off);
-       if (ecode != TDB_SUCCESS) {
-               return ecode;
-       }
-
-       return tdb_write_convert(tdb, off, &new, sizeof(new));
-}
-
-/* List need not be locked. */
-enum TDB_ERROR add_free_record(struct tdb_context *tdb,
-                              tdb_off_t off, tdb_len_t len_with_header)
-{
-       tdb_off_t b_off;
-       tdb_len_t len;
-       enum TDB_ERROR ecode;
-
-       assert(len_with_header >= sizeof(struct tdb_free_record));
 
-       len = len_with_header - sizeof(struct tdb_used_record);
+       /* Update enqueue count, but don't set high bit: see TDB_OFF_IS_ERR */
+       if (*coalesce)
+               head += (1ULL << (64 - TDB_OFF_UPPER_STEAL));
+       head &= ~(TDB_OFF_MASK | (1ULL << 63));
+       head |= off;
 
-       b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
-       ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
+       ecode = tdb_write_off(tdb, b_off, head);
        if (ecode != TDB_SUCCESS) {
                return ecode;
        }
 
-       ecode = enqueue_in_free(tdb, b_off, off, len);
-       check_list(tdb, b_off);
-       tdb_unlock_free_bucket(tdb, b_off);
-       return ecode;
-}
-
-static size_t adjust_size(size_t keylen, size_t datalen)
-{
-       size_t size = keylen + datalen;
-
-       if (size < TDB_MIN_DATA_LEN)
-               size = TDB_MIN_DATA_LEN;
-
-       /* Round to next uint64_t boundary. */
-       return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
-}
-
-/* If we have enough left over to be useful, split that off. */
-static size_t record_leftover(size_t keylen, size_t datalen,
-                             bool want_extra, size_t total_len)
-{
-       ssize_t leftover;
-
-       if (want_extra)
-               datalen += datalen / 2;
-       leftover = total_len - adjust_size(keylen, datalen);
-
-       if (leftover < (ssize_t)sizeof(struct tdb_free_record))
-               return 0;
+       /* It's time to coalesce if counter wrapped. */
+       if (*coalesce)
+               *coalesce = ((head & ~TDB_OFF_MASK) == 0);
 
-       return leftover;
+       return tdb_write_convert(tdb, off, &new, sizeof(new));
 }
 
 static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
@@ -333,10 +300,12 @@ static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
        return off;
 }
 
-/* Note: we unlock the current bucket if we coalesce or fail. */
-static tdb_bool_err coalesce(struct tdb_context *tdb,
-                            tdb_off_t off, tdb_off_t b_off,
-                            tdb_len_t data_len)
+/* Note: we unlock the current bucket if fail (-ve), or coalesce (+ve) and
+ * need to blatt the *protect record (which is set to an error). */
+static tdb_len_t coalesce(struct tdb_context *tdb,
+                         tdb_off_t off, tdb_off_t b_off,
+                         tdb_len_t data_len,
+                         tdb_off_t *protect)
 {
        tdb_off_t end;
        struct tdb_free_record rec;
@@ -400,6 +369,10 @@ static tdb_bool_err coalesce(struct tdb_context *tdb,
                        break;
                }
 
+               /* Did we just mess up a record you were hoping to use? */
+               if (end == *protect)
+                       *protect = TDB_ERR_NOEXIST;
+
                ecode = remove_from_list(tdb, nb_off, end, &rec);
                check_list(tdb, nb_off);
                if (ecode != TDB_SUCCESS) {
@@ -414,7 +387,11 @@ static tdb_bool_err coalesce(struct tdb_context *tdb,
 
        /* Didn't find any adjacent free? */
        if (end == off + sizeof(struct tdb_used_record) + data_len)
-               return false;
+               return 0;
+
+       /* Before we expand, check this isn't one you wanted protected? */
+       if (off == *protect)
+               *protect = TDB_ERR_EXISTS;
 
        /* OK, expand initial record */
        ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
@@ -435,32 +412,228 @@ static tdb_bool_err coalesce(struct tdb_context *tdb,
                goto err;
        }
 
-       /* We have to drop this to avoid deadlocks, so make sure record
-        * doesn't get coalesced by someone else! */
-       rec.ftable_and_len = (TDB_FTABLE_NONE << (64 - TDB_OFF_UPPER_STEAL))
-               | (end - off - sizeof(struct tdb_used_record));
-       ecode = tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
-                                                 ftable_and_len),
-                             rec.ftable_and_len);
+       /* Try locking violation first.  We don't allow coalesce recursion! */
+       ecode = add_free_record(tdb, off, end - off, TDB_LOCK_NOWAIT, false);
        if (ecode != TDB_SUCCESS) {
-               goto err;
-       }
+               /* Need to drop lock.  Can't rely on anything stable. */
+               *protect = TDB_ERR_CORRUPT;
+
+               /* We have to drop this to avoid deadlocks, so make sure record
+                * doesn't get coalesced by someone else! */
+               rec.ftable_and_len = (TDB_FTABLE_NONE
+                                     << (64 - TDB_OFF_UPPER_STEAL))
+                       | (end - off - sizeof(struct tdb_used_record));
+               ecode = tdb_write_off(tdb,
+                                     off + offsetof(struct tdb_free_record,
+                                                    ftable_and_len),
+                                     rec.ftable_and_len);
+               if (ecode != TDB_SUCCESS) {
+                       goto err;
+               }
+
+               tdb_unlock_free_bucket(tdb, b_off);
 
+               ecode = add_free_record(tdb, off, end - off, TDB_LOCK_WAIT,
+                                       false);
+               if (ecode != TDB_SUCCESS) {
+                       return ecode;
+               }
+       } else if (TDB_OFF_IS_ERR(*protect)) {
+               /* For simplicity, we always drop lock if they can't continue */
+               tdb_unlock_free_bucket(tdb, b_off);
+       }
        tdb->stats.alloc_coalesce_succeeded++;
+
+       /* Return usable length. */
+       return end - off - sizeof(struct tdb_used_record);
+
+err:
+       /* To unify error paths, we *always* unlock bucket on error. */
        tdb_unlock_free_bucket(tdb, b_off);
+       return ecode;
+}
+
+/* List is locked: we unlock it. */
+static enum TDB_ERROR coalesce_list(struct tdb_context *tdb,
+                                   tdb_off_t ftable_off,
+                                   tdb_off_t b_off,
+                                   unsigned int limit)
+{
+       enum TDB_ERROR ecode;
+       tdb_off_t off;
 
-       ecode = add_free_record(tdb, off, end - off);
+       off = tdb_read_off(tdb, b_off);
+       if (TDB_OFF_IS_ERR(off)) {
+               ecode = off;
+               goto unlock_err;
+       }
+       /* A little bit of paranoia: counter should be 0. */
+       off &= TDB_OFF_MASK;
+
+       while (off && limit--) {
+               struct tdb_free_record rec;
+               tdb_len_t coal;
+               tdb_off_t next;
+
+               ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+
+               next = rec.next;
+               coal = coalesce(tdb, off, b_off, frec_len(&rec), &next);
+               if (TDB_OFF_IS_ERR(coal)) {
+                       /* This has already unlocked on error. */
+                       return coal;
+               }
+               if (TDB_OFF_IS_ERR(next)) {
+                       /* Coalescing had to unlock, so stop. */
+                       return TDB_SUCCESS;
+               }
+               /* Keep going if we're doing well... */
+               limit += size_to_bucket(coal / 16 + TDB_MIN_DATA_LEN);
+               off = next;
+       }
+
+       /* Now, move those elements to the tail of the list so we get something
+        * else next time. */
+       if (off) {
+               struct tdb_free_record oldhrec, newhrec, oldtrec, newtrec;
+               tdb_off_t oldhoff, oldtoff, newtoff;
+
+               /* The record we were up to is the new head. */
+               ecode = tdb_read_convert(tdb, off, &newhrec, sizeof(newhrec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+
+               /* Get the new tail. */
+               newtoff = frec_prev(&newhrec);
+               ecode = tdb_read_convert(tdb, newtoff, &newtrec,
+                                        sizeof(newtrec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+
+               /* Get the old head. */
+               oldhoff = tdb_read_off(tdb, b_off);
+               if (TDB_OFF_IS_ERR(oldhoff)) {
+                       ecode = oldhoff;
+                       goto unlock_err;
+               }
+
+               /* This could happen if they all coalesced away. */
+               if (oldhoff == off)
+                       goto out;
+
+               ecode = tdb_read_convert(tdb, oldhoff, &oldhrec,
+                                        sizeof(oldhrec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+
+               /* Get the old tail. */
+               oldtoff = frec_prev(&oldhrec);
+               ecode = tdb_read_convert(tdb, oldtoff, &oldtrec,
+                                        sizeof(oldtrec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+
+               /* Old tail's next points to old head. */
+               oldtrec.next = oldhoff;
+
+               /* Old head's prev points to old tail. */
+               oldhrec.magic_and_prev
+                       = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL))
+                       | oldtoff;
+
+               /* New tail's next is 0. */
+               newtrec.next = 0;
+
+               /* Write out the modified versions. */
+               ecode = tdb_write_convert(tdb, oldtoff, &oldtrec,
+                                         sizeof(oldtrec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+
+               ecode = tdb_write_convert(tdb, oldhoff, &oldhrec,
+                                         sizeof(oldhrec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+
+               ecode = tdb_write_convert(tdb, newtoff, &newtrec,
+                                         sizeof(newtrec));
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+               
+               /* And finally link in new head. */
+               ecode = tdb_write_off(tdb, b_off, off);
+               if (ecode != TDB_SUCCESS)
+                       goto unlock_err;
+       }
+out:
+       tdb_unlock_free_bucket(tdb, b_off);
+       return TDB_SUCCESS;
+
+unlock_err:
+       tdb_unlock_free_bucket(tdb, b_off);
+       return ecode;
+}
+
+/* List must not be locked if coalesce_ok is set. */
+enum TDB_ERROR add_free_record(struct tdb_context *tdb,
+                              tdb_off_t off, tdb_len_t len_with_header,
+                              enum tdb_lock_flags waitflag,
+                              bool coalesce)
+{
+       tdb_off_t b_off;
+       tdb_len_t len;
+       enum TDB_ERROR ecode;
+
+       assert(len_with_header >= sizeof(struct tdb_free_record));
+
+       len = len_with_header - sizeof(struct tdb_used_record);
+
+       b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
+       ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
        if (ecode != TDB_SUCCESS) {
                return ecode;
        }
-       return true;
 
-err:
-       /* To unify error paths, we *always* unlock bucket on error. */
-       tdb_unlock_free_bucket(tdb, b_off);
+       ecode = enqueue_in_free(tdb, b_off, off, len, &coalesce);
+       check_list(tdb, b_off);
+
+       /* Coalescing unlocks free list. */
+       if (!ecode && coalesce)
+               ecode = coalesce_list(tdb, tdb->ftable_off, b_off, 2);
+       else
+               tdb_unlock_free_bucket(tdb, b_off);
        return ecode;
 }
 
+static size_t adjust_size(size_t keylen, size_t datalen)
+{
+       size_t size = keylen + datalen;
+
+       if (size < TDB_MIN_DATA_LEN)
+               size = TDB_MIN_DATA_LEN;
+
+       /* Round to next uint64_t boundary. */
+       return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
+}
+
+/* If we have enough left over to be useful, split that off. */
+static size_t record_leftover(size_t keylen, size_t datalen,
+                             bool want_extra, size_t total_len)
+{
+       ssize_t leftover;
+
+       if (want_extra)
+               datalen += datalen / 2;
+       leftover = total_len - adjust_size(keylen, datalen);
+
+       if (leftover < (ssize_t)sizeof(struct tdb_free_record))
+               return 0;
+
+       return leftover;
+}
+
 /* We need size bytes to put our key and data in. */
 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
                                tdb_off_t ftable_off,
@@ -477,7 +650,6 @@ static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
        enum TDB_ERROR ecode;
 
        tdb->stats.allocs++;
-again:
        b_off = bucket_off(ftable_off, bucket);
 
        /* FIXME: Try non-blocking wait first, to measure contention. */
@@ -503,12 +675,12 @@ again:
                ecode = off;
                goto unlock_err;
        }
+       off &= TDB_OFF_MASK;
 
        while (off) {
                const struct tdb_free_record *r;
                tdb_len_t len;
                tdb_off_t next;
-               int coal;
 
                r = tdb_access_read(tdb, off, sizeof(*r), true);
                if (TDB_PTR_IS_ERR(r)) {
@@ -541,17 +713,6 @@ again:
                next = r->next;
                len = frec_len(r);
                tdb_access_release(tdb, r);
-
-               /* Since we're going slow anyway, try coalescing here. */
-               coal = coalesce(tdb, off, b_off, len);
-               if (coal == 1) {
-                       /* This has unlocked list, restart. */
-                       goto again;
-               }
-               if (coal < 0) {
-                       /* This has already unlocked on error. */
-                       return coal;
-               }
                off = next;
        }
 
@@ -600,7 +761,7 @@ again:
                        ecode = add_free_record(tdb,
                                                best_off + sizeof(rec)
                                                + frec_len(&best) - leftover,
-                                               leftover);
+                                               leftover, TDB_LOCK_WAIT, false);
                        if (ecode != TDB_SUCCESS) {
                                best_off = ecode;
                        }
@@ -774,7 +935,7 @@ static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
        tdb_unlock_expand(tdb, F_WRLCK);
 
        tdb->stats.expands++;
-       return add_free_record(tdb, old_size, wanted);
+       return add_free_record(tdb, old_size, wanted, TDB_LOCK_WAIT, true);
 }
 
 /* This won't fail: it will expand the database if it has to. */