Patch: A different fix for the "out of memory" problem

[Steven Grimm]

The 1.1.13 prerelease has code to automatically preallocate a slab in 
each slab class to avoid spurious "out of memory" errors when some slab 
classes aren't in use by the time memcached reaches its memory limit. 
Among other changes we've made to memcached 1.1.12 (patches 
forthcoming!) we decided to solve the same problem in a different way: 
if needed, we let memcached exceed its configured memory limit so it can 
allocate at least one slab in each slab class. This is only on an 
as-needed basis; if all slab classes are already in use by the time the 
memory limit is hit, this doesn't eat any extra memory.

The disadvantage is that memcached no longer adheres as strictly to its 
memory limit. The advantage is that if you don't ever need a slab of a 
given class, you don't waste memory preallocating it. A lot of our 
memcached servers are dedicated to fixed-size items, so we only ever 
need a single slab class on those machines -- there's no point setting 
aside megabytes of memory for "set" requests that can never happen.

We don't view the disadvantage as a big deal, in part because 
memcached's memory usage isn't strictly lower than the configured limit 
to begin with: additional memory is already used by the stack, allocated 
memory other than the slabs, TCP buffers and so forth.

Anyway, here's the patch, which is about a tenth the size of the 
explanation.

-Steven Grimm
Facebook


--- ../memcached-1.1.12/slabs.c 2006-03-24 23:17:18.000000000 -0800
+++ slabs.c     2006-05-02 14:16:08.000000000 -0700
@@ -103,7 +103,7 @@
    int len = POWER_BLOCK;
    char *ptr;

-    if (mem_limit && mem_malloced + len > mem_limit)
+    if (mem_limit && mem_malloced + len > mem_limit && p->slabs > 0)
        return 0;

    if (! grow_slab_list(id)) return 0;