From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932743AbYDQBQ6 (ORCPT ); Wed, 16 Apr 2008 21:16:58 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1755446AbYDQBHu (ORCPT ); Wed, 16 Apr 2008 21:07:50 -0400 Received: from sous-sol.org ([216.99.217.87]:58880 "EHLO sous-sol.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1759232AbYDQBHs (ORCPT ); Wed, 16 Apr 2008 21:07:48 -0400 Message-Id: <20080417010331.416680812@sous-sol.org> References: <20080417010122.148289106@sous-sol.org> User-Agent: quilt/0.46-1 Date: Wed, 16 Apr 2008 18:01:39 -0700 From: Chris Wright To: linux-kernel@vger.kernel.org, stable@kernel.org Cc: Justin Forbes , Zwane Mwaikambo , "Theodore Ts'o" , Randy Dunlap , Dave Jones , Chuck Wolber , Chris Wedgwood , Michael Krufky , Chuck Ebbert , Domenico Andreoli , torvalds@linux-foundation.org, akpm@linux-foundation.org, alan@lxorguk.ukuu.org.uk, snitzer@gmail.com, dada1@cosmosbay.com, Peter Zijlstra Subject: alloc_percpu() fails to allocate percpu data Content-Disposition: inline; filename=alloc_percpu-fails-to-allocate-percpu-data.patch Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org -stable review patch. If anyone has any objections, please let us know. --------------------- From: Eric Dumazet upstream commit: be852795e1c8d3829ddf3cb1ce806113611fa555 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 Acked-by: Peter Zijlstra Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds Signed-off-by: Chris Wright --- mm/allocpercpu.c | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-) This is appropriate for 2.6.24.y --- a/mm/allocpercpu.c +++ b/mm/allocpercpu.c @@ -6,6 +6,10 @@ #include #include +#ifndef cache_line_size +#define cache_line_size() L1_CACHE_BYTES +#endif + /** * percpu_depopulate - depopulate per-cpu data for given cpu * @__pdata: per-cpu data to depopulate @@ -52,6 +56,11 @@ void *percpu_populate(void *__pdata, siz struct percpu_data *pdata = __percpu_disguise(__pdata); int node = cpu_to_node(cpu); + /* + * We should make sure each CPU gets private memory. + */ + size = roundup(size, cache_line_size()); + BUG_ON(pdata->ptrs[cpu]); if (node_online(node)) pdata->ptrs[cpu] = kmalloc_node(size, gfp|__GFP_ZERO, node); @@ -98,7 +107,11 @@ EXPORT_SYMBOL_GPL(__percpu_populate_mask */ void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask) { - void *pdata = kzalloc(nr_cpu_ids * sizeof(void *), gfp); + /* + * We allocate whole cache lines to avoid false sharing + */ + size_t sz = roundup(nr_cpu_ids * sizeof(void *), cache_line_size()); + void *pdata = kzalloc(sz, gfp); void *__pdata = __percpu_disguise(pdata); if (unlikely(!pdata)) --