Ah, I read the patch too fast.
In this snippet
[[
+/* Clear and release the given connection POOL.
+ */
+static void
+release_connection_pool(apr_pool_t *pool)
+{
+ svn_error_t *err;
+ svn_pool_clear(pool);
+
+ err = svn_mutex__lock(connection_pools_mutex);
+ if (err)
+ {
+ svn_error_clear(err);
+ svn_pool_destroy(pool);
+ }
+ else
+ {
+ APR_ARRAY_PUSH(connection_pools, apr_pool_t *) = pool;
+ svn_error_clear(svn_mutex__unlock(connection_pools_mutex,
+ SVN_NO_ERROR));
+ }
+}
+
]]
I automatically assumed the 'svn_pool_clear' was just a variable assignment,
because usually we have a white line before the first real code. instead of
just after the first line.
Maybe we should fix the whitespace here to follow our usual layout?
Bert
From: Stefan Fuhrmann [mailto:[email protected]]
Sent: zondag 15 september 2013 20:25
To: Bert Huijben
Cc: Subversion Development; [email protected]
Subject: Re: svn commit: r1523465 -
/subversion/trunk/subversion/svnserve/svnserve.c
On Sun, Sep 15, 2013 at 8:07 PM, Bert Huijben <[email protected]
<mailto:[email protected]> > wrote:
> -----Original Message-----
> From: [email protected] <mailto:[email protected]>
[mailto:[email protected] <mailto:[email protected]> ]
> Sent: zondag 15 september 2013 19:47
> To: [email protected] <mailto:[email protected]>
> Subject: svn commit: r1523465 -
> /subversion/trunk/subversion/svnserve/svnserve.c
>
> Author: stefan2
> Date: Sun Sep 15 17:46:36 2013
> New Revision: 1523465
>
> URL: http://svn.apache.org/r1523465
> Log:
> As it turns out, allocating memory from the OS in a multi-threaded
> environment is relatively costly. With APR pools, this happens
> every time we use a newly created root pool.
>
> Therefore, teach svnserve to recycle the connection pools, keeping
> those precious memory blocks allocated instead of disposing and
> re-allocating them.
Is this really the best way to do this?
Not sure. I'm open for suggestions.
Can't we create a subpool here? (Or do we also need multiple allocators,
etc.)
No. Those pools will be used concurrently by their
respective worker threads.
In the implementation I see that the existing pools are re-used, but they
are not *cleared* before re-use?
release_connection_pool() clears them.
Shouldn't we at least release the used memory (and thate) when handing back
the memory to the pool allocator?
-- Stefan^2.
