[Tarantool-patches] [PATCH] quota: add is_enabled field
Vladislav Shpilevoy
v.shpilevoy at tarantool.org
Tue Dec 15 02:41:51 MSK 2020
Thanks for the patch!
See 5 comments below.
On 11.12.2020 16:37, Ilya Kosarev wrote:
> By default the quota is enabled. If it is set to false, quota_use will
> allow to overuse the total available memory limit.
> In case of disabled quota smalloc() allocates (and frees) large slabs
> using malloc so that the quota will be able to shrink back after those
> slabs are freed. Test introduced.
>
> Part of tarantool/tarantool#3807
> ---
> Branch: https://github.com/tarantool/small/tree/i.kosarev/gh-3807-safe-alloc-on-truncation
> Issue: https://github.com/tarantool/tarantool/issues/3807
>
> small/quota.h | 21 ++++++++++++++++++++-
> small/slab_cache.c | 6 ++++++
> small/slab_cache.h | 3 +++
> small/small.c | 10 ++++++----
> test/small_alloc.c | 32 ++++++++++++++++++++++++++++++++
> 5 files changed, 67 insertions(+), 5 deletions(-)
>
> diff --git a/small/quota.h b/small/quota.h
> index 3d3b4f0..c71ee5d 100644
> --- a/small/quota.h
> +++ b/small/quota.h
> @@ -57,6 +57,12 @@ struct quota {
> * QUOTA_UNIT_SIZE.
> */
> uint64_t value;
> + /**
> + * By default the quota is enabled. If it is set to
> + * false, quota_use will allow to overuse the total
> + * available memory limit.
> + */
> + bool is_enabled;
1. Quota is accessed from many threads. This is why the 'value' is
updated using atomic cmpxchg operations. Here it seems you update
the quota is_enabled from tx thread, but the flag is used by
all threads.
Perhaps we could make the flag atomic as well. But I don't know
how often is the quota accessed, and if we can add more atomic
operations on each use().
I suggest you to ask Alexander. Since he is with us now, he should
be able to tell if we are going to the right direction.
> };
>
> /**
> diff --git a/small/slab_cache.c b/small/slab_cache.c
> index 2ba56c5..b3b7004 100644
> --- a/small/slab_cache.c
> +++ b/small/slab_cache.c
> @@ -462,3 +462,9 @@ slab_cache_check(struct slab_cache *cache)
> return;
> abort();
> }
> +
> +bool
> +slab_cache_quota_enabled(struct slab_cache *cache)
2. For checks we always include 'is' as a part of the name.
Also 'cache' can be declared 'const' here.
> +{
> + return cache->arena->quota->is_enabled;
> +}
> diff --git a/small/small.c b/small/small.c
> index 48085fb..154a0e9 100644
> --- a/small/small.c
> +++ b/small/small.c
> @@ -228,7 +228,7 @@ smalloc(struct small_alloc *alloc, size_t size)
> struct mempool *pool;
> int idx = (size - 1) >> STEP_SIZE_LB;
> idx = (idx > (int) alloc->step_pool0_step_count) ? idx - alloc->step_pool0_step_count : 0;
> - if (idx < STEP_POOL_MAX) {
> + if (idx < STEP_POOL_MAX && slab_cache_quota_enabled(alloc->cache)) {
3. These checks are super expensive - you dereference 3! pointers
for each such check. Why do you even needed? Why isn't it enough
to patch the quota?
> /* Allocate in a stepped pool. */
> pool = &alloc->step_pools[idx];
> assert(size <= pool->objsize &&
> diff --git a/test/small_alloc.c b/test/small_alloc.c
> index 421442a..1563cc9 100644
> --- a/test/small_alloc.c
> +++ b/test/small_alloc.c
> @@ -128,6 +128,37 @@ small_alloc_large(void)
> footer();
> }
>
> +static void
> +small_quota_release(void)
> +{
> + header();
> +
> + size_t total, used;
> + size_t alloc_size = 1000;
> + quota_get_total_and_used("a, &total, &used);
> + size_t amount = (total - used) / alloc_size;
> +
> + int **a = calloc(amount, sizeof(int *));
> + int i = 0;
> + while (quota_use("a, 1) >= 0) {
> + quota_disable("a);
> + a[i++] = smalloc(&alloc, alloc_size);
> + quota_enable("a);
> + }
> +
> + quota_get_total_and_used("a, &total, &used);
> + fail_unless((int)total - (int)used < 0);
4. Why so complex? Why can't you simply write 'total < used'?
The same below.
> +
> + quota_disable("a);
> + smfree(&alloc, a[0], alloc_size);
5. So you did 'amount' allocations, but freed only the
first one. Why?
> + quota_enable("a);
> + free(a);
> +
> + quota_get_total_and_used("a, &total, &used);
> + fail_unless((int)total - (int)used > 0);
> + footer();
> +}
More information about the Tarantool-patches
mailing list