On Thu, 2008-10-23 at 14:36 +0200, Max Kellermann wrote:I am finally getting some time to look at my problem that I originally reported in October (SUNRPC problem with 2.6.26 and beyond), and I am seeing the same behavior as Max Kellermann when my machine slows as I described earlier. The system in question is currently running 2.6.27.7. Here is what I see when it is misbehaving:
On 2008/10/22 11:12, Max Kellermann <max@xxxxxxxxxxx> wrote:
after I was able to fix http://lkml.org/lkml/2008/10/17/147, theI have bisected the problem: 98a8e323 is the result ("SUNRPC: Add a
server which was already upgraded to 2.6.27.2 still gets very high
load. It is a web server with NFS file storage (NetApp), and while
the others in the cluster (kernel 2.6.25) have a load of 1-3, 2.6.27.2
gets 30-50.
I did an oprofile, with the following results (server just started,
load "only" 5-10):
87593 56.1116 (no location information) vmlinux
vmlinux rpcauth_lookup_credcache
16037 10.2732 auth_generic.c:0 vmlinux
vmlinux generic_match
6460 4.1382 (no location information) php4
php4 (no symbols)
2478 1.5874 (no location information) libc-2.7.so
libc-2.7.so (no symbols)
[...]
We havn't configured any special authentication method. It is a NFSv3
over UDP mount, but the kernel has NFSv4 and therefore KRB5 enabled.
Any ideas why rpcauth_lookup_credcache() goes overboard with CPU
usage?
helper rpcauth_lookup_generic_cred()"). 5c691044 is ok.
See the attached oprofile annotation data for both commits. I guess
that the function rpcauth_lookup_credcache() is waiting for a spinlock
too often and too long. Trond, any idea?
Can you add a '-v' to the rpc.gssd daemon startup line? I'd like to see
how often you are creating new gss contexts.
Harry: added you to Cc because your problem sounds similar.
Harry's problem is should be unrelated. afaik, he is seeing a problem
with userland RPC code, not kernel rpc code.
Trond