aboutsummaryrefslogtreecommitdiffstats
path: root/mm
diff options
context:
space:
mode:
authorEric Dumazet <dada1@cosmosbay.com>2008-03-04 17:28:35 -0500
committerLinus Torvalds <torvalds@woody.linux-foundation.org>2008-03-04 19:35:11 -0500
commitbe852795e1c8d3829ddf3cb1ce806113611fa555 (patch)
tree3671dbd5ba64f11650f93fdd9178666bcbb84f7e /mm
parente3892296de632e3f9299d9fabe0c746740004891 (diff)
alloc_percpu() fails to allocate percpu data
Some oprofile results obtained while using tbench on a 2x2 cpu machine were very surprising. For example, loopback_xmit() function was using high number of cpu cycles to perform the statistic updates, supposed to be real cheap since they use percpu data pcpu_lstats = netdev_priv(dev); lb_stats = per_cpu_ptr(pcpu_lstats, smp_processor_id()); lb_stats->packets++; /* HERE : serious contention */ lb_stats->bytes += skb->len; struct pcpu_lstats is a small structure containing two longs. It appears that on my 32bits platform, alloc_percpu(8) allocates a single cache line, instead of giving to each cpu a separate cache line. Using the following patch gave me impressive boost in various benchmarks ( 6 % in tbench) (all percpu_counters hit this bug too) Long term fix (ie >= 2.6.26) would be to let each CPU allocate their own block of memory, so that we dont need to roudup sizes to L1_CACHE_BYTES, or merging the SGI stuff of course... Note : SLUB vs SLAB is important here to *show* the improvement, since they dont have the same minimum allocation sizes (8 bytes vs 32 bytes). This could very well explain regressions some guys reported when they switched to SLUB. Signed-off-by: Eric Dumazet <dada1@cosmosbay.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'mm')
-rw-r--r--mm/allocpercpu.c15
1 files changed, 14 insertions, 1 deletions
diff --git a/mm/allocpercpu.c b/mm/allocpercpu.c
index 7e58322b7134..b0012e27fea8 100644
--- a/mm/allocpercpu.c
+++ b/mm/allocpercpu.c
@@ -6,6 +6,10 @@
6#include <linux/mm.h> 6#include <linux/mm.h>
7#include <linux/module.h> 7#include <linux/module.h>
8 8
9#ifndef cache_line_size
10#define cache_line_size() L1_CACHE_BYTES
11#endif
12
9/** 13/**
10 * percpu_depopulate - depopulate per-cpu data for given cpu 14 * percpu_depopulate - depopulate per-cpu data for given cpu
11 * @__pdata: per-cpu data to depopulate 15 * @__pdata: per-cpu data to depopulate
@@ -52,6 +56,11 @@ void *percpu_populate(void *__pdata, size_t size, gfp_t gfp, int cpu)
52 struct percpu_data *pdata = __percpu_disguise(__pdata); 56 struct percpu_data *pdata = __percpu_disguise(__pdata);
53 int node = cpu_to_node(cpu); 57 int node = cpu_to_node(cpu);
54 58
59 /*
60 * We should make sure each CPU gets private memory.
61 */
62 size = roundup(size, cache_line_size());
63
55 BUG_ON(pdata->ptrs[cpu]); 64 BUG_ON(pdata->ptrs[cpu]);
56 if (node_online(node)) 65 if (node_online(node))
57 pdata->ptrs[cpu] = kmalloc_node(size, gfp|__GFP_ZERO, node); 66 pdata->ptrs[cpu] = kmalloc_node(size, gfp|__GFP_ZERO, node);
@@ -98,7 +107,11 @@ EXPORT_SYMBOL_GPL(__percpu_populate_mask);
98 */ 107 */
99void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask) 108void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask)
100{ 109{
101 void *pdata = kzalloc(nr_cpu_ids * sizeof(void *), gfp); 110 /*
111 * We allocate whole cache lines to avoid false sharing
112 */
113 size_t sz = roundup(nr_cpu_ids * sizeof(void *), cache_line_size());
114 void *pdata = kzalloc(sz, gfp);
102 void *__pdata = __percpu_disguise(pdata); 115 void *__pdata = __percpu_disguise(pdata);
103 116
104 if (unlikely(!pdata)) 117 if (unlikely(!pdata))