On 29/07/11 16:30, Eric Dumazet wrote:
Le vendredi 29 juillet 2011 à 16:05 +1000, Greg Banks a écrit :
On 29/07/11 15:32, NeilBrown wrote:
I seem to remember coming to the conclusion that Jeff eventually
addressed this problem...am I misremembering or did something regress?
Currently, all nfsd kthreads use memory for their kernel stack and
various initial data from a _single_ node, even if you use
sunrpc.pool_mode=pernode (or percpu)
That's just plain broken and I'm very pleased to see you fix it.
I was just surprised that it was still broken and wondering how that
happened. Looking at ToT I see that because I dropped the ball in 2008,
Jeff's patches didn't address the problem. In ToT
svc_pool_map_set_cpumask() is called *after* kthread_create() and
applies to the child thread, *after* it's stack has been allocated on
the wrong node. In the working SGI code, svc_pool_map_set_cpumask() is
called by the parent node on itself *before* calling kernel_thread() or
doing any of the data structure allocations, thus ensuring that
everything gets allocated using the default memory allocation policy,
which on SGI NFS servers was globally tuned to be "node-local".
With my patch, we make sure each thread gets its stack from its local
node.
Check commit 94dcf29a11b3d20a (kthread: use kthread_create_on_node()) to
see how this strategy already was adopted for ksoftirqd, kworker,
migration, and pktgend kthreads.
Ah, I see. It's unfortunate that the kthread_create() API ends up being
passed a CPU number but that's only used to format the name and not for
sensible things :(
--
Greg.
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html