This patch is onto mmotm Jan/15.
=
From: KAMEZAWA Hiroyuki <[email protected]>
When per-cpu counter for memcg was implemneted, dynamic percpu allocator
was not very good. But now, we have good one and useful macros.
This patch replaces memcg's private percpu counter implementation with
generic dynamic percpu allocator and macros.
The benefits are
- We can remove private implementation.
- The counters will be NUMA-aware. (Current one is not...)
- This patch reduces sizeof(struct mem_cgroup). Then,
struct mem_cgroup may be fit in page size on small config.
By this, size of text is reduced.
[Before]
[kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
text data bss dec hex filename
24373 2528 4132 31033 7939 mm/memcontrol.o
[After]
[kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
text data bss dec hex filename
23913 2528 4132 30573 776d mm/memcontrol.o
This includes no functional changes.
Signed-off-by: KAMEZAWA Hiroyuki <[email protected]>
---
mm/memcontrol.c | 184 +++++++++++++++++++-------------------------------------
1 file changed, 63 insertions(+), 121 deletions(-)
Index: mmotm-2.6.33-Jan15/mm/memcontrol.c
===================================================================
--- mmotm-2.6.33-Jan15.orig/mm/memcontrol.c
+++ mmotm-2.6.33-Jan15/mm/memcontrol.c
@@ -89,54 +89,8 @@ enum mem_cgroup_stat_index {
struct mem_cgroup_stat_cpu {
s64 count[MEM_CGROUP_STAT_NSTATS];
-} ____cacheline_aligned_in_smp;
-
-struct mem_cgroup_stat {
- struct mem_cgroup_stat_cpu cpustat[0];
};
-static inline void
-__mem_cgroup_stat_set_safe(struct mem_cgroup_stat_cpu *stat,
- enum mem_cgroup_stat_index idx, s64 val)
-{
- stat->count[idx] = val;
-}
-
-static inline s64
-__mem_cgroup_stat_read_local(struct mem_cgroup_stat_cpu *stat,
- enum mem_cgroup_stat_index idx)
-{
- return stat->count[idx];
-}
-
-/*
- * For accounting under irq disable, no need for increment preempt count.
- */
-static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
- enum mem_cgroup_stat_index idx, int val)
-{
- stat->count[idx] += val;
-}
-
-static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
- enum mem_cgroup_stat_index idx)
-{
- int cpu;
- s64 ret = 0;
- for_each_possible_cpu(cpu)
- ret += stat->cpustat[cpu].count[idx];
- return ret;
-}
-
-static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
-{
- s64 ret;
-
- ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
- ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
- return ret;
-}
-
/*
* per-zone information in memory controller.
*/
@@ -270,9 +224,9 @@ struct mem_cgroup {
unsigned long move_charge_at_immigrate;
/*
- * statistics. This must be placed at the end of memcg.
+ * percpu counter.
*/
- struct mem_cgroup_stat stat;
+ struct mem_cgroup_stat_cpu *stat;
};
/* Stuffs for move charges at task migration. */
@@ -441,19 +395,14 @@ mem_cgroup_remove_exceeded(struct mem_cg
static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
{
bool ret = false;
- int cpu;
s64 val;
- struct mem_cgroup_stat_cpu *cpustat;
- cpu = get_cpu();
- cpustat = &mem->stat.cpustat[cpu];
- val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_SOFTLIMIT);
+ val = this_cpu_read(mem->stat->count[MEM_CGROUP_STAT_SOFTLIMIT]);
if (unlikely(val < 0)) {
- __mem_cgroup_stat_set_safe(cpustat, MEM_CGROUP_STAT_SOFTLIMIT,
+ this_cpu_write(mem->stat->count[MEM_CGROUP_STAT_SOFTLIMIT],
SOFTLIMIT_EVENTS_THRESH);
ret = true;
}
- put_cpu();
return ret;
}
@@ -549,17 +498,31 @@ mem_cgroup_largest_soft_limit_node(struc
return mz;
}
+static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
+ enum mem_cgroup_stat_index idx)
+{
+ int cpu;
+ s64 val = 0;
+
+ for_each_possible_cpu(cpu)
+ val += per_cpu(mem->stat->count[idx], cpu);
+ return val;
+}
+
+static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
+{
+ s64 ret;
+
+ ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
+ ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
+ return ret;
+}
+
static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
bool charge)
{
int val = (charge) ? 1 : -1;
- struct mem_cgroup_stat *stat = &mem->stat;
- struct mem_cgroup_stat_cpu *cpustat;
- int cpu = get_cpu();
-
- cpustat = &stat->cpustat[cpu];
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_SWAPOUT, val);
- put_cpu();
+ this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
}
static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
@@ -567,26 +530,22 @@ static void mem_cgroup_charge_statistics
bool charge)
{
int val = (charge) ? 1 : -1;
- struct mem_cgroup_stat *stat = &mem->stat;
- struct mem_cgroup_stat_cpu *cpustat;
- int cpu = get_cpu();
- cpustat = &stat->cpustat[cpu];
+ preempt_disable();
+
if (PageCgroupCache(pc))
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
+ __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
else
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
+ __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
if (charge)
- __mem_cgroup_stat_add_safe(cpustat,
- MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
+ __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
else
- __mem_cgroup_stat_add_safe(cpustat,
- MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_SOFTLIMIT, -1);
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_THRESHOLDS, -1);
+ __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
+ __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_SOFTLIMIT]);
+ __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_THRESHOLDS]);
- put_cpu();
+ preempt_enable();
}
static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
@@ -1244,7 +1203,7 @@ static int mem_cgroup_hierarchical_recla
}
}
}
- if (!mem_cgroup_local_usage(&victim->stat)) {
+ if (!mem_cgroup_local_usage(victim)) {
/* this cgroup's local usage == 0 */
css_put(&victim->css);
continue;
@@ -1310,9 +1269,6 @@ static void record_last_oom(struct mem_c
void mem_cgroup_update_file_mapped(struct page *page, int val)
{
struct mem_cgroup *mem;
- struct mem_cgroup_stat *stat;
- struct mem_cgroup_stat_cpu *cpustat;
- int cpu;
struct page_cgroup *pc;
pc = lookup_page_cgroup(page);
@@ -1328,13 +1284,10 @@ void mem_cgroup_update_file_mapped(struc
goto done;
/*
- * Preemption is already disabled, we don't need get_cpu()
+ * Preemption is already disabled. We can use __this_cpu_xxx
*/
- cpu = smp_processor_id();
- stat = &mem->stat;
- cpustat = &stat->cpustat[cpu];
+ __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED], val);
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED, val);
done:
unlock_page_cgroup(pc);
}
@@ -1761,9 +1714,6 @@ static void __mem_cgroup_move_account(st
struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
{
struct page *page;
- int cpu;
- struct mem_cgroup_stat *stat;
- struct mem_cgroup_stat_cpu *cpustat;
VM_BUG_ON(from == to);
VM_BUG_ON(PageLRU(pc->page));
@@ -1773,18 +1723,11 @@ static void __mem_cgroup_move_account(st
page = pc->page;
if (page_mapped(page) && !PageAnon(page)) {
- cpu = smp_processor_id();
- /* Update mapped_file data for mem_cgroup "from" */
- stat = &from->stat;
- cpustat = &stat->cpustat[cpu];
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED,
- -1);
-
- /* Update mapped_file data for mem_cgroup "to" */
- stat = &to->stat;
- cpustat = &stat->cpustat[cpu];
- __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_FILE_MAPPED,
- 1);
+ /* Update mapped_file data for mem_cgroup */
+ preempt_disable();
+ __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
+ __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
+ preempt_enable();
}
mem_cgroup_charge_statistics(from, pc, false);
if (uncharge)
@@ -2885,7 +2828,7 @@ static int
mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
{
struct mem_cgroup_idx_data *d = data;
- d->val += mem_cgroup_read_stat(&mem->stat, d->idx);
+ d->val += mem_cgroup_read_stat(mem, d->idx);
return 0;
}
@@ -3126,18 +3069,18 @@ static int mem_cgroup_get_local_stat(str
s64 val;
/* per cpu stat */
- val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
+ val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
s->stat[MCS_CACHE] += val * PAGE_SIZE;
- val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
+ val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
s->stat[MCS_RSS] += val * PAGE_SIZE;
- val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_FILE_MAPPED);
+ val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
- val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
+ val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
s->stat[MCS_PGPGIN] += val;
- val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
+ val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
s->stat[MCS_PGPGOUT] += val;
if (do_swap_account) {
- val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_SWAPOUT);
+ val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
s->stat[MCS_SWAP] += val * PAGE_SIZE;
}
@@ -3287,19 +3230,14 @@ static int mem_cgroup_swappiness_write(s
static bool mem_cgroup_threshold_check(struct mem_cgroup *mem)
{
bool ret = false;
- int cpu;
s64 val;
- struct mem_cgroup_stat_cpu *cpustat;
- cpu = get_cpu();
- cpustat = &mem->stat.cpustat[cpu];
- val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_THRESHOLDS);
+ val = this_cpu_read(mem->stat->count[MEM_CGROUP_STAT_THRESHOLDS]);
if (unlikely(val < 0)) {
- __mem_cgroup_stat_set_safe(cpustat, MEM_CGROUP_STAT_THRESHOLDS,
+ this_cpu_write(mem->stat->count[MEM_CGROUP_STAT_THRESHOLDS],
THRESHOLDS_EVENTS_THRESH);
ret = true;
}
- put_cpu();
return ret;
}
@@ -3687,17 +3625,12 @@ static void free_mem_cgroup_per_zone_inf
kfree(mem->info.nodeinfo[node]);
}
-static int mem_cgroup_size(void)
-{
- int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
- return sizeof(struct mem_cgroup) + cpustat_size;
-}
-
static struct mem_cgroup *mem_cgroup_alloc(void)
{
struct mem_cgroup *mem;
- int size = mem_cgroup_size();
+ int size = sizeof(struct mem_cgroup);
+ /* Can be very big if MAX_NUMNODES is very big */
if (size < PAGE_SIZE)
mem = kmalloc(size, GFP_KERNEL);
else
@@ -3705,6 +3638,14 @@ static struct mem_cgroup *mem_cgroup_all
if (mem)
memset(mem, 0, size);
+ mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
+ if (!mem->stat) {
+ if (size < PAGE_SIZE)
+ kfree(mem);
+ else
+ vfree(mem);
+ mem = NULL;
+ }
return mem;
}
@@ -3729,7 +3670,8 @@ static void __mem_cgroup_free(struct mem
for_each_node_state(node, N_POSSIBLE)
free_mem_cgroup_per_zone_info(mem, node);
- if (mem_cgroup_size() < PAGE_SIZE)
+ free_percpu(mem->stat);
+ if (sizeof(struct mem_cgroup) < PAGE_SIZE)
kfree(mem);
else
vfree(mem);
On Wednesday 20 January 2010 12:48 PM, KAMEZAWA Hiroyuki wrote:
> This patch is onto mmotm Jan/15.
> =
> From: KAMEZAWA Hiroyuki <[email protected]>
>
> When per-cpu counter for memcg was implemneted, dynamic percpu allocator
> was not very good. But now, we have good one and useful macros.
> This patch replaces memcg's private percpu counter implementation with
> generic dynamic percpu allocator and macros.
>
> The benefits are
> - We can remove private implementation.
> - The counters will be NUMA-aware. (Current one is not...)
> - This patch reduces sizeof(struct mem_cgroup). Then,
> struct mem_cgroup may be fit in page size on small config.
>
> By this, size of text is reduced.
> [Before]
> [kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
> text data bss dec hex filename
> 24373 2528 4132 31033 7939 mm/memcontrol.o
> [After]
> [kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
> text data bss dec hex filename
> 23913 2528 4132 30573 776d mm/memcontrol.o
>
> This includes no functional changes.
>
> Signed-off-by: KAMEZAWA Hiroyuki <[email protected]>
Before review, could you please post parallel pagefault data on a large
system, since root now uses these per cpu counters and its overhead is
now dependent on these counters. Also the data read from root cgroup is
also dependent on these, could you make sure that is not broken.
Balbir
On Wed, 20 Jan 2010 15:07:52 +0530
Balbir Singh <[email protected]> wrote:
> On Wednesday 20 January 2010 12:48 PM, KAMEZAWA Hiroyuki wrote:
> > This patch is onto mmotm Jan/15.
> > =
> > From: KAMEZAWA Hiroyuki <[email protected]>
> >
> > When per-cpu counter for memcg was implemneted, dynamic percpu allocator
> > was not very good. But now, we have good one and useful macros.
> > This patch replaces memcg's private percpu counter implementation with
> > generic dynamic percpu allocator and macros.
> >
> > The benefits are
> > - We can remove private implementation.
> > - The counters will be NUMA-aware. (Current one is not...)
> > - This patch reduces sizeof(struct mem_cgroup). Then,
> > struct mem_cgroup may be fit in page size on small config.
> >
> > By this, size of text is reduced.
> > [Before]
> > [kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
> > text data bss dec hex filename
> > 24373 2528 4132 31033 7939 mm/memcontrol.o
> > [After]
> > [kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
> > text data bss dec hex filename
> > 23913 2528 4132 30573 776d mm/memcontrol.o
> >
> > This includes no functional changes.
> >
> > Signed-off-by: KAMEZAWA Hiroyuki <[email protected]>
>
>
> Before review, could you please post parallel pagefault data on a large
> system, since root now uses these per cpu counters and its overhead is
> now dependent on these counters. Also the data read from root cgroup is
> also dependent on these, could you make sure that is not broken.
>
No number difference before/after patch on my SMP quick test.
But I don't have NUMA. Could you test on NUMA ?
I'll measure again tomorrow if I have machine time.
Thanks,
-Kame
On Wednesday 20 January 2010 03:17 PM, KAMEZAWA Hiroyuki wrote:
> On Wed, 20 Jan 2010 15:07:52 +0530
> Balbir Singh <[email protected]> wrote:
>
>> On Wednesday 20 January 2010 12:48 PM, KAMEZAWA Hiroyuki wrote:
>>> This patch is onto mmotm Jan/15.
>>> =
>>> From: KAMEZAWA Hiroyuki <[email protected]>
>>>
>>> When per-cpu counter for memcg was implemneted, dynamic percpu allocator
>>> was not very good. But now, we have good one and useful macros.
>>> This patch replaces memcg's private percpu counter implementation with
>>> generic dynamic percpu allocator and macros.
>>>
>>> The benefits are
>>> - We can remove private implementation.
>>> - The counters will be NUMA-aware. (Current one is not...)
>>> - This patch reduces sizeof(struct mem_cgroup). Then,
>>> struct mem_cgroup may be fit in page size on small config.
>>>
>>> By this, size of text is reduced.
>>> [Before]
>>> [kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
>>> text data bss dec hex filename
>>> 24373 2528 4132 31033 7939 mm/memcontrol.o
>>> [After]
>>> [kamezawa@bluextal mmotm-2.6.33-Jan15]$ size mm/memcontrol.o
>>> text data bss dec hex filename
>>> 23913 2528 4132 30573 776d mm/memcontrol.o
>>>
>>> This includes no functional changes.
>>>
>>> Signed-off-by: KAMEZAWA Hiroyuki <[email protected]>
>>
>>
>> Before review, could you please post parallel pagefault data on a large
>> system, since root now uses these per cpu counters and its overhead is
>> now dependent on these counters. Also the data read from root cgroup is
>> also dependent on these, could you make sure that is not broken.
>>
> No number difference before/after patch on my SMP quick test.
> But I don't have NUMA. Could you test on NUMA ?
>
> I'll measure again tomorrow if I have machine time.
I'll do the same as well if possible.
Balbir
On Wed, 20 Jan 2010 15:07:52 +0530
Balbir Singh <[email protected]> wrote:
> > This includes no functional changes.
> >
> > Signed-off-by: KAMEZAWA Hiroyuki <[email protected]>
>
>
> Before review, could you please post parallel pagefault data on a large
> system, since root now uses these per cpu counters and its overhead is
> now dependent on these counters. Also the data read from root cgroup is
> also dependent on these, could you make sure that is not broken.
>
Hmm, I rewrote test program for avoidng mmap_sem. This version does fork()
instead of pthread_create() and meausre parallel-process page fault speed.
[Before patch]
[root@bluextal memory]# /root/bin/perf stat -e page-faults,cache-misses --repeat 5 ./multi-fault-fork 8
Performance counter stats for './multi-fault-fork 8' (5 runs):
45256919 page-faults ( +- 0.851% )
602230144 cache-misses ( +- 0.187% )
61.020533723 seconds time elapsed ( +- 0.002%
[After patch]
[root@bluextal memory]# /root/bin/perf stat -e page-faults,cache-misses --repeat 5 ./multi-fault-fork 8
Performance counter stats for './multi-fault-fork 8' (5 runs):
46007166 page-faults ( +- 0.339% )
599553505 cache-misses ( +- 0.298% )
61.020937843 seconds time elapsed ( +- 0.004% )
slightly improved ? But this test program does some extreme behavior and
you can't see difference in real-world applications, I think.
So, I guess this is in error-range in famous (not small) benchmarks.
Thanks,
-Kame
On Thursday 21 January 2010 07:37 AM, KAMEZAWA Hiroyuki wrote:
> On Wed, 20 Jan 2010 15:07:52 +0530
> Balbir Singh <[email protected]> wrote:
>
>>> This includes no functional changes.
>>>
>>> Signed-off-by: KAMEZAWA Hiroyuki <[email protected]>
>>
>>
>> Before review, could you please post parallel pagefault data on a large
>> system, since root now uses these per cpu counters and its overhead is
>> now dependent on these counters. Also the data read from root cgroup is
>> also dependent on these, could you make sure that is not broken.
>>
> Hmm, I rewrote test program for avoidng mmap_sem. This version does fork()
> instead of pthread_create() and meausre parallel-process page fault speed.
>
> [Before patch]
> [root@bluextal memory]# /root/bin/perf stat -e page-faults,cache-misses --repeat 5 ./multi-fault-fork 8
>
> Performance counter stats for './multi-fault-fork 8' (5 runs):
>
> 45256919 page-faults ( +- 0.851% )
> 602230144 cache-misses ( +- 0.187% )
>
> 61.020533723 seconds time elapsed ( +- 0.002%
>
> [After patch]
> [root@bluextal memory]# /root/bin/perf stat -e page-faults,cache-misses --repeat 5 ./multi-fault-fork 8
>
> Performance counter stats for './multi-fault-fork 8' (5 runs):
>
> 46007166 page-faults ( +- 0.339% )
> 599553505 cache-misses ( +- 0.298% )
>
> 61.020937843 seconds time elapsed ( +- 0.004% )
>
> slightly improved ? But this test program does some extreme behavior and
> you can't see difference in real-world applications, I think.
> So, I guess this is in error-range in famous (not small) benchmarks.
Looks good, please give me a couple of days to test, I'll revert back
with numbers and review.
--
Three Cheers,
Balbir Singh