| Message ID | 20220907043537.3457014-4-shakeelb@google.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | memcg: reduce memory overhead of memory cgroups | expand |
On Wed, Sep 07, 2022 at 04:35:37AM +0000, Shakeel Butt wrote: > The struct memcg_vmstats and struct memcg_vmstats_percpu contains two > arrays each for events of size NR_VM_EVENT_ITEMS which can be as large > as 110. However the memcg v1 only uses 4 of those while memcg v2 uses > 15. The union of both is 17. On a 64 bit system, we are wasting > approximately ((110 - 17) * 8 * 2) * (nr_cpus + 1) bytes which is > significant on large machines. > > This patch reduces the size of the given structures by adding one > indirection and only stores array of events which are actually used by > the memcg code. With this patch, the size of memcg_vmstats has reduced > from 2544 bytes to 1056 bytes while the size of memcg_vmstats_percpu has > reduced from 2568 bytes to 1080 bytes. This is pretty impressive! Thank you, Shakeel! Acked-by: Roman Gushchin <roman.gushchin@linux.dev> for the series. Thanks!
On Tue, Sep 6, 2022 at 9:36 PM Shakeel Butt <shakeelb@google.com> wrote: > [...] > > static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event) > { > long x = 0; > int cpu; > + int index = memcg_events_index(event); > + > + if (index < 0) > + return 0; > > for_each_possible_cpu(cpu) > x += per_cpu(memcg->vmstats_percpu->events[event], cpu); Andrew, can you please replace 'event' in the above line with 'index'? I had this correct in the original single patch but messed up while breaking up that patch into three patches for easier review.
On Wed, 7 Sep 2022 19:35:10 -0700 Shakeel Butt <shakeelb@google.com> wrote: > On Tue, Sep 6, 2022 at 9:36 PM Shakeel Butt <shakeelb@google.com> wrote: > > > [...] > > > > static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event) > > { > > long x = 0; > > int cpu; > > + int index = memcg_events_index(event); > > + > > + if (index < 0) > > + return 0; > > > > for_each_possible_cpu(cpu) > > x += per_cpu(memcg->vmstats_percpu->events[event], cpu); > > Andrew, can you please replace 'event' in the above line with 'index'? > I had this correct in the original single patch but messed up while > breaking up that patch into three patches for easier review. No probs. From: Andrew Morton <akpm@linux-foundation.org> Subject: memcg-reduce-size-of-memcg-vmstats-structures-fix Date: Thu Sep 8 03:35:53 PM PDT 2022 fix memcg_events_local() array index, per Shakeel Link: https://lkml.kernel.org/r/CALvZod70Mvxr+Nzb6k0yiU2RFYjTD=0NFhKK-Eyp+5ejd1PSFw@mail.gmail.com Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Michal Hocko <mhocko@kernel.org> Cc: Muchun Song <songmuchun@bytedance.com> Cc: Roman Gushchin <roman.gushchin@linux.dev> Cc: Shakeel Butt <shakeelb@google.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> --- mm/memcontrol.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) --- a/mm/memcontrol.c~memcg-reduce-size-of-memcg-vmstats-structures-fix +++ a/mm/memcontrol.c @@ -921,7 +921,7 @@ static unsigned long memcg_events_local( return 0; for_each_possible_cpu(cpu) - x += per_cpu(memcg->vmstats_percpu->events[event], cpu); + x += per_cpu(memcg->vmstats_percpu->events[index], cpu); return x; }
Hello. On Wed, Sep 07, 2022 at 04:35:37AM +0000, Shakeel Butt <shakeelb@google.com> wrote: > /* Subset of vm_event_item to report for memcg event stats */ > static const unsigned int memcg_vm_event_stat[] = { > + PGPGIN, > + PGPGOUT, > PGSCAN_KSWAPD, > PGSCAN_DIRECT, > PGSTEAL_KSWAPD, What about adding a dummy entry at the beginning like: static const unsigned int memcg_vm_event_stat[] = { + NR_VM_EVENT_ITEMS, + PGPGIN, + PGPGOUT, PGSCAN_KSWAPD, PGSCAN_DIRECT, > @@ -692,14 +694,30 @@ static const unsigned int memcg_vm_event_stat[] = { > #endif > }; > > +#define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat) > +static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly; > + > +static void init_memcg_events(void) > +{ > + int i; > + > + for (i = 0; i < NR_MEMCG_EVENTS; ++i) > + mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1; Start such loops from i = 1, save i to the table. > +} > + > +static inline int memcg_events_index(enum vm_event_item idx) > +{ > + return mem_cgroup_events_index[idx] - 1; > +} And the there'd be no need for the reverse transforms -1. I.e. it might be just a negligible micro-optimization but since the event updates are on some fast (albeit longer) paths, it may be worth sacrificing one of the saved 8Bs in favor of no arithmetics. What do you think about this? > static unsigned long memcg_events(struct mem_cgroup *memcg, int event) > { > - return READ_ONCE(memcg->vmstats->events[event]); > + int index = memcg_events_index(event); > + > + if (index < 0) > + return 0; As a bonus these undefined maps could use the zero at the dummy location without branch (slow paths though). > @@ -5477,7 +5511,7 @@ static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu) > parent->vmstats->state_pending[i] += delta; > } > > - for (i = 0; i < NR_VM_EVENT_ITEMS; i++) { > + for (i = 0; i < NR_MEMCG_EVENTS; i++) { I applaud this part :-) Michal
On Thu, Sep 8, 2022 at 5:23 PM Michal Koutný <mkoutny@suse.com> wrote: > > Hello. > > On Wed, Sep 07, 2022 at 04:35:37AM +0000, Shakeel Butt <shakeelb@google.com> wrote: > > /* Subset of vm_event_item to report for memcg event stats */ > > static const unsigned int memcg_vm_event_stat[] = { > > + PGPGIN, > > + PGPGOUT, > > PGSCAN_KSWAPD, > > PGSCAN_DIRECT, > > PGSTEAL_KSWAPD, > > What about adding a dummy entry at the beginning like: > > static const unsigned int memcg_vm_event_stat[] = { > + NR_VM_EVENT_ITEMS, > + PGPGIN, > + PGPGOUT, > PGSCAN_KSWAPD, > PGSCAN_DIRECT, > > > > @@ -692,14 +694,30 @@ static const unsigned int memcg_vm_event_stat[] = { > > #endif > > }; > > > > +#define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat) > > +static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly; > > + > > +static void init_memcg_events(void) > > +{ > > + int i; > > + > > + for (i = 0; i < NR_MEMCG_EVENTS; ++i) > > + mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1; > > Start such loops from i = 1, save i to the table. > > > +} > > + > > +static inline int memcg_events_index(enum vm_event_item idx) > > +{ > > + return mem_cgroup_events_index[idx] - 1; > > +} > > And the there'd be no need for the reverse transforms -1. > > I.e. it might be just a negligible micro-optimization but since the > event updates are on some fast (albeit longer) paths, it may be worth > sacrificing one of the saved 8Bs in favor of no arithmetics. > > What do you think about this? > > > static unsigned long memcg_events(struct mem_cgroup *memcg, int event) > > { > > - return READ_ONCE(memcg->vmstats->events[event]); > > + int index = memcg_events_index(event); > > + > > + if (index < 0) > > + return 0; > > As a bonus these undefined maps could use the zero at the dummy location > without branch (slow paths though). > > > > @@ -5477,7 +5511,7 @@ static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu) > > parent->vmstats->state_pending[i] += delta; > > } > > > > - for (i = 0; i < NR_VM_EVENT_ITEMS; i++) { > > + for (i = 0; i < NR_MEMCG_EVENTS; i++) { > > I applaud this part :-) > > Hi Michal, Thanks for taking a look. Let me get back to you on this later. I am at the moment rearranging struct mem_cgroup for better packing and will be running some benchmarks. Later I will see if your suggestion has any performance benefit or just more readable code then I will follow up. Shakeel
diff --git a/mm/memcontrol.c b/mm/memcontrol.c index d0ccc16ed416..a60012be6140 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -671,6 +671,8 @@ static void flush_memcg_stats_dwork(struct work_struct *w) /* Subset of vm_event_item to report for memcg event stats */ static const unsigned int memcg_vm_event_stat[] = { + PGPGIN, + PGPGOUT, PGSCAN_KSWAPD, PGSCAN_DIRECT, PGSTEAL_KSWAPD, @@ -692,14 +694,30 @@ static const unsigned int memcg_vm_event_stat[] = { #endif }; +#define NR_MEMCG_EVENTS ARRAY_SIZE(memcg_vm_event_stat) +static int mem_cgroup_events_index[NR_VM_EVENT_ITEMS] __read_mostly; + +static void init_memcg_events(void) +{ + int i; + + for (i = 0; i < NR_MEMCG_EVENTS; ++i) + mem_cgroup_events_index[memcg_vm_event_stat[i]] = i + 1; +} + +static inline int memcg_events_index(enum vm_event_item idx) +{ + return mem_cgroup_events_index[idx] - 1; +} + struct memcg_vmstats_percpu { /* Local (CPU and cgroup) page state & events */ long state[MEMCG_NR_STAT]; - unsigned long events[NR_VM_EVENT_ITEMS]; + unsigned long events[NR_MEMCG_EVENTS]; /* Delta calculation for lockless upward propagation */ long state_prev[MEMCG_NR_STAT]; - unsigned long events_prev[NR_VM_EVENT_ITEMS]; + unsigned long events_prev[NR_MEMCG_EVENTS]; /* Cgroup1: threshold notifications & softlimit tree updates */ unsigned long nr_page_events; @@ -709,11 +727,11 @@ struct memcg_vmstats_percpu { struct memcg_vmstats { /* Aggregated (CPU and subtree) page state & events */ long state[MEMCG_NR_STAT]; - unsigned long events[NR_VM_EVENT_ITEMS]; + unsigned long events[NR_MEMCG_EVENTS]; /* Pending child counts during tree propagation */ long state_pending[MEMCG_NR_STAT]; - unsigned long events_pending[NR_VM_EVENT_ITEMS]; + unsigned long events_pending[NR_MEMCG_EVENTS]; }; unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx) @@ -873,24 +891,34 @@ void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val) void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx, unsigned long count) { - if (mem_cgroup_disabled()) + int index = memcg_events_index(idx); + + if (mem_cgroup_disabled() || index < 0) return; memcg_stats_lock(); - __this_cpu_add(memcg->vmstats_percpu->events[idx], count); + __this_cpu_add(memcg->vmstats_percpu->events[index], count); memcg_rstat_updated(memcg, count); memcg_stats_unlock(); } static unsigned long memcg_events(struct mem_cgroup *memcg, int event) { - return READ_ONCE(memcg->vmstats->events[event]); + int index = memcg_events_index(event); + + if (index < 0) + return 0; + return READ_ONCE(memcg->vmstats->events[index]); } static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event) { long x = 0; int cpu; + int index = memcg_events_index(event); + + if (index < 0) + return 0; for_each_possible_cpu(cpu) x += per_cpu(memcg->vmstats_percpu->events[event], cpu); @@ -1564,10 +1592,15 @@ static void memory_stat_format(struct mem_cgroup *memcg, char *buf, int bufsize) memcg_events(memcg, PGSTEAL_KSWAPD) + memcg_events(memcg, PGSTEAL_DIRECT)); - for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) + for (i = 0; i < ARRAY_SIZE(memcg_vm_event_stat); i++) { + if (memcg_vm_event_stat[i] == PGPGIN || + memcg_vm_event_stat[i] == PGPGOUT) + continue; + seq_buf_printf(&s, "%s %lu\n", vm_event_name(memcg_vm_event_stat[i]), memcg_events(memcg, memcg_vm_event_stat[i])); + } /* The above should easily fit into one page */ WARN_ON_ONCE(seq_buf_has_overflowed(&s)); @@ -5309,6 +5342,7 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) page_counter_init(&memcg->kmem, &parent->kmem); page_counter_init(&memcg->tcpmem, &parent->tcpmem); } else { + init_memcg_events(); page_counter_init(&memcg->memory, NULL); page_counter_init(&memcg->swap, NULL); page_counter_init(&memcg->kmem, NULL); @@ -5477,7 +5511,7 @@ static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu) parent->vmstats->state_pending[i] += delta; } - for (i = 0; i < NR_VM_EVENT_ITEMS; i++) { + for (i = 0; i < NR_MEMCG_EVENTS; i++) { delta = memcg->vmstats->events_pending[i]; if (delta) memcg->vmstats->events_pending[i] = 0;
The struct memcg_vmstats and struct memcg_vmstats_percpu contains two arrays each for events of size NR_VM_EVENT_ITEMS which can be as large as 110. However the memcg v1 only uses 4 of those while memcg v2 uses 15. The union of both is 17. On a 64 bit system, we are wasting approximately ((110 - 17) * 8 * 2) * (nr_cpus + 1) bytes which is significant on large machines. This patch reduces the size of the given structures by adding one indirection and only stores array of events which are actually used by the memcg code. With this patch, the size of memcg_vmstats has reduced from 2544 bytes to 1056 bytes while the size of memcg_vmstats_percpu has reduced from 2568 bytes to 1080 bytes. Signed-off-by: Shakeel Butt <shakeelb@google.com> --- mm/memcontrol.c | 52 ++++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 43 insertions(+), 9 deletions(-)