[haiku-commits] Re: r35601 - in haiku/trunk: headers/private/kernel/slab src/system/kernel/cache src/system/kernel/slab src/system/kernel/vm
- From: Ingo Weinhold <ingo_weinhold@xxxxxx>
- To: haiku-commits@xxxxxxxxxxxxx
- Date: Wed, 24 Feb 2010 17:27:58 +0100
On 2010-02-24 at 15:43:20 [+0100], axeld@xxxxxxxxxxxxxxxx wrote:
> Author: axeld
> Date: 2010-02-24 15:43:20 +0100 (Wed, 24 Feb 2010)
> New Revision: 35601
> Changeset: http://dev.haiku-os.org/changeset/35601/haiku
[...]
> static bool
> -exchange_with_empty(object_depot* depot, DepotMagazine*& magazine)
> +exchange_with_empty(object_depot* depot, DepotMagazine*& magazine,
> + DepotMagazine*& freeMagazine)
> {
> + ASSERT(magazine == NULL || magazine->IsFull());
> +
> SpinLocker _(depot->inner_lock);
>
> if (depot->empty == NULL)
> @@ -137,9 +139,12 @@
>
> depot->empty_count--;
>
> - if (magazine) {
> - _push(depot->full, magazine);
> - depot->full_count++;
> + if (magazine != NULL) {
> + if (depot->full_count < depot->max_count) {
> + _push(depot->full, magazine);
> + depot->full_count++;
> + } else
> + freeMagazine = magazine;
A disadvantage of this implementation is that a relatively recently used
magazine is discarded while older ones are kept, which from a caching
perspective is definitely suboptimal. Have you by any chance tested how that
affects performance?
> Modified: haiku/trunk/src/system/kernel/slab/Slab.cpp
> ===================================================================
> --- haiku/trunk/src/system/kernel/slab/Slab.cpp 2010-02-24 14:10:17 UTC
> (rev 35600)
> +++ haiku/trunk/src/system/kernel/slab/Slab.cpp 2010-02-24 14:43:20 UTC
> (rev 35601)
> @@ -413,6 +413,11 @@
> if (cache->reclaimer)
> cache->reclaimer(cache->cookie, level);
>
> + if ((cache->flags & CACHE_NO_DEPOT) == 0) {
> + object_depot_make_empty(&cache->depot, 0);
> + // TODO: what flags?
The apropriate alloc/free flags. I suppose CACHE_DONT_WAIT_FOR_MEMORY could
be specified to prevent the somewhat pathological situation that the low
memory handler has to wait for memory (when the object depot for the
magazines object cache wants to allocate a magazine and has to go to
MemoryManager; or when a HashedObjectCache's hash table is resized). I guess
it would be better to introduce a special flag CACHE_DONT_ALLOCATE for this
case, though.
CU, Ingo
Other related posts:
