From: Dave Chinner <dchinner@xxxxxxxxxx>
THe libxfs cache has a lot of false scalability about it. I can't
get lookups to scale past about one and half CPUs, with one of the
key problems being a preponderance of global locks.
Like just after doing a hash lookup, that is careful only to take
the hash chain lock, it takes a global cache lock to update the
cache hit statistic. Scalable? Not at all.
The node priority stuff is protected by the object locks (i.e the
buffer lock) and so it doesn't need locks, either.
This doesn't do very much to improve scalability, but it's a small
start.
Signed-off-by: Dave Chinner <dchinner@xxxxxxxxxx>
---
libxfs/cache.c | 6 ------
1 file changed, 6 deletions(-)
diff --git a/libxfs/cache.c b/libxfs/cache.c
index dc69689..9f7f8f4 100644
--- a/libxfs/cache.c
+++ b/libxfs/cache.c
@@ -391,9 +391,7 @@ cache_node_get(
pthread_mutex_unlock(&node->cn_mutex);
pthread_mutex_unlock(&hash->ch_mutex);
- pthread_mutex_lock(&cache->c_mutex);
cache->c_hits++;
- pthread_mutex_unlock(&cache->c_mutex);
*nodep = node;
return 0;
@@ -482,10 +480,8 @@ cache_node_set_priority(
else if (priority > CACHE_MAX_PRIORITY)
priority = CACHE_MAX_PRIORITY;
- pthread_mutex_lock(&node->cn_mutex);
ASSERT(node->cn_count > 0);
node->cn_priority = priority;
- pthread_mutex_unlock(&node->cn_mutex);
}
int
@@ -494,9 +490,7 @@ cache_node_get_priority(
{
int priority;
- pthread_mutex_lock(&node->cn_mutex);
priority = node->cn_priority;
- pthread_mutex_unlock(&node->cn_mutex);
return priority;
}
--
1.8.4.rc3
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
