[PATCH 01/18] vinyl: rework internal quota API
Konstantin Osipov
kostja at tarantool.org
Mon Aug 20 14:07:21 MSK 2018
* Vladimir Davydov <vdavydov.dev at gmail.com> [18/08/16 23:03]:
> + *
> + * Usage pattern:
> + *
> + * size_t reserved = <estimate>;
> + * if (vy_quota_try_use(q, reserved, timeout) != 0)
> + * return -1;
> + * <allocate memory>
> + * size_t used = <actually allocated>;
> + * vy_quota_commit_use(q, reserved, used);
How is this different from vy_quota_use(<estimate>); followed by
vy_quota_release(<estimate> - <actually allocated>)?
If vy_quota_commit_use() is actually release *or* force-use,
depending on the sign of the result (<estimate> - <actually
allocated>) then the new api is actually less clear than the old
one.
If you would like to introduce a new call which would either
release quota or force-use it, then this call should be called
vy_quota_adjust() or something like that, and the old call
(vy_quota_use()) should be left intact. The new names are imho
less clear.
> + * because we may not yield after we start inserting statements
> + * into a space so we estimate the allocation size and wait for
> + * quota before committing statements. At the same time, we
> + * cannot precisely estimate the size of memory we are going to
> + * consume so we adjust the quota after the allocation.
> + *
> + * The size of memory allocated while committing a transaction
> + * may be greater than an estimate, because insertion of a
> + * statement into an in-memory index can trigger allocation
> + * of a new index extent. This should not normally result in a
> + * noticeable breach in the memory limit, because most memory
> + * is occupied by statements, but we need to adjust the quota
> + * accordingly after the allocation in this case.
> + *
> + * The actual memory allocation size may also be less than an
> + * estimate if the space has multiple indexes, because statements
> + * are stored in the common memory level, which isn't taken into
> + * account while estimating the size of a memory allocation.
> */
> static inline int
> -vy_quota_use(struct vy_quota *q, size_t size, double timeout)
> +vy_quota_try_use(struct vy_quota *q, size_t size, double timeout)
> {
> double start_time = ev_monotonic_now(loop());
try_use() suggests that the call quickly fails if it can't succeed
(e.g. trylock()).
> double deadline = start_time + timeout;
> @@ -178,6 +208,27 @@ vy_quota_use(struct vy_quota *q, size_t size, double timeout)
> }
>
> /**
> + * Adjust quota after allocating memory.
interestingly you use the verb adjust in the comment yourself.
> + *
> + * @reserved: size of quota reserved by vy_quota_try_use().
> + * @used: size of memory actually allocated.
> + *
> + * See also vy_quota_try_use().
> + */
> +static inline void
> +vy_quota_commit_use(struct vy_quota *q, size_t reserved, size_t used)
> +{
> + if (reserved > used) {
> + size_t excess = reserved - used;
> + assert(q->used >= excess);
> + q->used -= excess;
> + fiber_cond_broadcast(&q->cond);
> + }
> + if (reserved < used)
> + vy_quota_force_use(q, used - reserved);
> +}
--
Konstantin Osipov, Moscow, Russia, +7 903 626 22 32
http://tarantool.io - www.twitter.com/kostja_osipov
More information about the Tarantool-patches
mailing list