On Wed, Feb 03, 2010 at 05:31:30PM +1100, NeilBrown wrote:
> Thanks to Bruce challenging me to justify the complexity of my
> previous version of this I have managed to simplify it significantly.
And thanks for persisting! As you know it's also been a worry of mine
that the v4 compound behavior isn't correct with respect to deferrals,
so it will be reassuring to have this sorted out. Maybe be a day or two
before I get to take a careful look.
--b.
>
> I have changed the rules for sunrpc_caches so that items that have
> expired get removed at the earliest opportunity even if they are still
> referenced, and in particular so that sunrpc_cache_lookup never
> returns an expired item.
> This means that cache_check doesn't need to check for "expired" any
> more and so only initiates an upcall for items that are not VALID.
>
> This means that when the upcall is responded to, it will always be
> exactly that item that is updated - never a different item with the
> same key. So there is no longer any need to repeat the lookup.
>
> The last 3 patches in this series are simply "cleanups" that I
> happened across while mucking about in the code. The should have zero
> change in functionality, and if you don't think they are cleanups
> (second last is now questionable), feel free to ignore them.
>
> I have tested this to ensure that it doesn't completely break things,
> and to ensure that it fixes the problem(*) but I haven't hammered on
> it very hard.
>
> (*)
> The problem is exhibited by sending a stream of writes to the NFS
> server and then occasionally flushing the export cache (exportfs -f).
> The problem manifests by a write not getting a reply and the client
> having to retransmit.
> It is 'fixed' if there are no retransmit delays.
> The follow generates the required writes and shows the delays.
>
> Without the patch I get delays of 60 seconds with TCP and 5 seconds
> with UDP.
>
> NeilBrown
>
> /*
> * write to NFS server and report delays exceeding 1 second.
> */
>
> #define _GNU_SOURCE
>
> #include <sys/time.h>
> #include <stdio.h>
> #include <sys/fcntl.h>
> #include <malloc.h>
> #include <memory.h>
>
> main(int argc, char *argv[])
> {
> int usec;
> //int fd = open(argv[1], O_WRONLY|O_DIRECT|O_CREAT, 0666);
> //int fd = open(argv[1], O_WRONLY|O_SYNC|O_CREAT, 0666);
> int fd = open(argv[1], O_WRONLY|O_CREAT, 0666);
> char *buf;
> struct timeval tv1, tv2;
>
> posix_memalign(&buf, 4096, 409600);
> memset(buf, 0x5a, 409600);
>
> while(1) {
> gettimeofday(&tv1, NULL);
> write(fd, buf, 409600);
> gettimeofday(&tv2, NULL);
> usec = (tv2.tv_sec*1000000 + tv2.tv_usec) -
> (tv1.tv_sec*1000000 + tv1.tv_usec);
> if (usec > 1000000)
> printf(" %d\n", usec/1000000);
> else
> printf(".");
> fflush(stdout);
> }
> }
>
>
> ---
>
> NeilBrown (9):
> sunrpc: don't keep expired entries in the auth caches.
> sunrpc/cache: factor out cache_is_expired
> sunrpc: never return expired entries in sunrpc_cache_lookup
> sunrpc/cache: allow threads to block while waiting for cache update.
> nfsd/idmap: drop special request deferal in favour of improved default.
> sunrpc: close connection when a request is irretrievably lost.
> nfsd: factor out hash functions for export caches.
> svcauth_gss: replace a trivial 'switch' with an 'if'
> sunrpc/cache: change deferred-request hash table to use hlist.
>
>
> fs/nfsd/export.c | 40 ++++++++------
> fs/nfsd/nfs4idmap.c | 105 ++++--------------------------------
> include/linux/sunrpc/cache.h | 5 +-
> include/linux/sunrpc/svcauth.h | 10 ++-
> net/sunrpc/auth_gss/svcauth_gss.c | 51 ++++++++---------
> net/sunrpc/cache.c | 109 ++++++++++++++++++++++++++-----------
> net/sunrpc/svc.c | 3 +
> net/sunrpc/svc_xprt.c | 11 ++++
> net/sunrpc/svcauth_unix.c | 11 +++-
> 9 files changed, 169 insertions(+), 176 deletions(-)
>
>
--
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
