On most large NFS servers, there are very few per-export stats
entries and each one will be used on multiple CPUs. This tends to
make them a contention point on high call rate workloads (which is
why the per-export and per-client stats are designed to be disabled
at runtime). Experiment showed that prefetching the stats entry had
a small positive benefit on this contention.
If I was to do this properly, the per-export stats objects would
be stored per-cpu and stitched together on demand for export to
userspace. However, that would make the reference counting and
pruning semantics....interesting.
Signed-off-by: Greg Banks <[email protected]>
---
fs/nfsd/stats.c | 11 +++++++++++
1 file changed, 11 insertions(+)
Index: bfields/fs/nfsd/stats.c
===================================================================
--- bfields.orig/fs/nfsd/stats.c
+++ bfields/fs/nfsd/stats.c
@@ -33,6 +33,7 @@
#include <linux/list.h>
#include <linux/swap.h>
#include <linux/log2.h>
+#include <linux/prefetch.h>
#include <linux/sunrpc/svc.h>
#include <linux/sunrpc/stats.h>
@@ -44,6 +45,15 @@
#define hentry_from_hnode(hn) \
hlist_entry((hn), nfsd_stats_hentry_t, se_node)
+static inline void nfsd_stats_prefetch(nfsd_stats_hentry_t *se)
+{
+ unsigned int i;
+
+ for (i = 0 ; i < sizeof(nfsd_stats_hentry_t) ; i += PREFETCH_STRIDE)
+ prefetch((char *)se+i);
+}
+
+
struct nfsd_stats nfsdstats;
struct svc_stat nfsd_svcstats = {
.program = &nfsd_program,
@@ -395,6 +405,7 @@ void nfsd_stats_update_op(struct svc_rqs
fh->fh_export != NULL &&
(se = fh->fh_export->ex_stats) != NULL &&
rqstp->rq_export_stats == NULL) {
+ nfsd_stats_prefetch(se);
/*
* We want the stats to survive fh_put() of the filehandle
* so we can update os_bytes_out and service time in
--
Greg