Messages in this thread | | | Date | Mon, 13 Aug 2007 13:58:48 -0700 (PDT) | From | Christoph Lameter <> | Subject | Re: [patch] slab: revert "slab: fix alien cache handling" |
| |
On Mon, 13 Aug 2007, Siddha, Suresh B wrote:
> Can we revert git commit 3cdc0ed0cea50ea08dd146c1bbc82b1bcc2e1b80 ?
Only if you find another way to fix the bug that is addressed there.
> This is introducing a performance regression and nullifies the previous > commit.
The implementation of noalien was breaking slab NUMA locality so that BUG() statements were triggered. This fixed it.
> Looking at the 3cdc0ed0cea50ea08dd146c1bbc82b1bcc2e1b80 changelog, I can't find > enough info of why cache_free_alien() must be called regardless if we use alien > caches or not.
cache_free_alien has to be called for any remote object because the per cpu queue does only contain node local objects. The noalian option was putting remote objects on the per cpu queue!
If you contaminate the per cpu queue with foreign objects then NUMA locality no longer works. Objects from other nodes are considered to be on the local node etc. So first of all requests for memory from one node will be served with memory from another,.
Plus the locks for slab lists are per node. So when we final drain the objects from the per cpu queue to the slab the local listlock may be taken to free an object to a remote node. If another free is in progress on the remote node then this will lead to corruption of slab meta data.
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |