[ Upstream commit add1511c38166cf1036765f8c4aa939f0275a799 ]
Move a rarely called function call site out of the hot path.
This is an exceptionally small improvement because the compiler
inlines most of the functions that nfsd_cache_lookup() calls.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 0f29ce32fbc56cfdb304eec8a4deb920ccfd89c3 ]
Force the compiler to skip unneeded initialization for cases that
don't need those values. For example, NFSv4 COMPOUND operations are
RC_NOCACHE.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 378a6109dd142a678f629b740f558365150f60f9 ]
Clean up: The details of finding the right hash bucket are exactly
the same in both nfsd_cache_lookup() and nfsd_cache_update().
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 7578b2f628db27281d3165af0aa862311883a858 ]
Commit 7142b98d9fd7 ("nfsd: Clean up drc cache in preparation for
global spinlock elimination"), billed as a clean-up, added
be32_to_cpu() to the DRC hash function without explanation. That
commit removed two comments that state that byte-swapping in the
hash function is unnecessary without explaining whether there was
a need for that change.
On some Intel CPUs, the swab32 instruction is known to cause a CPU
pipeline stall. be32_to_cpu() does not add extra randomness, since
the hash multiplication is done /before/ shifting to the high-order
bits of the result.
As a micro-optimization, remove the unnecessary transform from the
DRC hash function.
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 8847ecc9274a14114385d1cb4030326baa0766eb ]
DRC bucket pruning is done by nfsd_cache_lookup(), which is part of
every NFSv2 and NFSv3 dispatch (ie, it's done while the client is
waiting).
I added a trace_printk() in prune_bucket() to see just how long
it takes to prune. Here are two ends of the spectrum:
prune_bucket: Scanned 1 and freed 0 in 90 ns, 62 entries remaining
prune_bucket: Scanned 2 and freed 1 in 716 ns, 63 entries remaining
...
prune_bucket: Scanned 75 and freed 74 in 34149 ns, 1 entries remaining
Pruning latency is noticeable on fast transports with fast storage.
By noticeable, I mean that the latency measured here in the worst
case is the same order of magnitude as the round trip time for
cached server operations.
We could do something like moving expired entries to an expired list
and then free them later instead of freeing them right in
prune_bucket(). But simply limiting the number of entries that can
be pruned by a lookup is simple and retains more entries in the
cache, making the DRC somewhat more effective.
Comparison with a 70/30 fio 8KB 12 thread direct I/O test:
Before:
write: IOPS=61.6k, BW=481MiB/s (505MB/s)(14.1GiB/30001msec); 0 zone resets
WRITE:
1848726 ops (30%)
avg bytes sent per op: 8340 avg bytes received per op: 136
backlog wait: 0.635158 RTT: 0.128525 total execute time: 0.827242 (milliseconds)
After:
write: IOPS=63.0k, BW=492MiB/s (516MB/s)(14.4GiB/30001msec); 0 zone resets
WRITE:
1891144 ops (30%)
avg bytes sent per op: 8340 avg bytes received per op: 136
backlog wait: 0.616114 RTT: 0.126842 total execute time: 0.805348 (milliseconds)
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit e567b98ce9a4b35b63c364d24828a9e5cd7a8179 ]
nfsd stats counters can be updated by concurrent nfsd threads without any
protection.
Convert some nfsd_stats and nfsd_net struct members to use percpu counters.
The longest_chain* members of struct nfsd_net remain unprotected.
Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
