From: Liu Shixin <[email protected]>
The page on pcplist could be used, but not counted into memory free or
avaliable, and pcp_free is only showed by show_mem(). Since commit
d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
significant decrease in the display of free memory, with a large number
of cpus and nodes, the number of pages in the percpu list can be very
large, so it is better to let user to know the pcp count.
Signed-off-by: Liu Shixin <[email protected]>
Signed-off-by: Kefeng Wang <[email protected]>
---
drivers/base/node.c | 14 +++++++++++++-
fs/proc/meminfo.c | 9 +++++++++
2 files changed, 22 insertions(+), 1 deletion(-)
diff --git a/drivers/base/node.c b/drivers/base/node.c
index eb0f43784c2b..846864e45db6 100644
--- a/drivers/base/node.c
+++ b/drivers/base/node.c
@@ -375,6 +375,9 @@ static ssize_t node_read_meminfo(struct device *dev,
struct sysinfo i;
unsigned long sreclaimable, sunreclaimable;
unsigned long swapcached = 0;
+ unsigned long free_pcp = 0;
+ struct zone *zone;
+ int cpu;
si_meminfo_node(&i, nid);
sreclaimable = node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B);
@@ -382,9 +385,17 @@ static ssize_t node_read_meminfo(struct device *dev,
#ifdef CONFIG_SWAP
swapcached = node_page_state_pages(pgdat, NR_SWAPCACHE);
#endif
+ for_each_populated_zone(zone) {
+ if (zone_to_nid(zone) != nid)
+ continue;
+ for_each_online_cpu(cpu)
+ free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
+ }
+
len = sysfs_emit_at(buf, len,
"Node %d MemTotal: %8lu kB\n"
"Node %d MemFree: %8lu kB\n"
+ "Node %d PcpFree: %8lu kB\n"
"Node %d MemUsed: %8lu kB\n"
"Node %d SwapCached: %8lu kB\n"
"Node %d Active: %8lu kB\n"
@@ -397,7 +408,8 @@ static ssize_t node_read_meminfo(struct device *dev,
"Node %d Mlocked: %8lu kB\n",
nid, K(i.totalram),
nid, K(i.freeram),
- nid, K(i.totalram - i.freeram),
+ nid, K(free_pcp),
+ nid, K(i.totalram - i.freeram - free_pcp),
nid, K(swapcached),
nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
node_page_state(pgdat, NR_ACTIVE_FILE)),
diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c
index 6e89f0e2fd20..672c784dfc8a 100644
--- a/fs/proc/meminfo.c
+++ b/fs/proc/meminfo.c
@@ -38,6 +38,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
unsigned long pages[NR_LRU_LISTS];
unsigned long sreclaimable, sunreclaim;
int lru;
+ unsigned long free_pcp = 0;
+ struct zone *zone;
+ int cpu;
si_meminfo(&i);
si_swapinfo(&i);
@@ -55,8 +58,14 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
sreclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B);
sunreclaim = global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B);
+ for_each_populated_zone(zone) {
+ for_each_online_cpu(cpu)
+ free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
+ }
+
show_val_kb(m, "MemTotal: ", i.totalram);
show_val_kb(m, "MemFree: ", i.freeram);
+ show_val_kb(m, "PcpFree: ", free_pcp);
show_val_kb(m, "MemAvailable: ", available);
show_val_kb(m, "Buffers: ", i.bufferram);
show_val_kb(m, "Cached: ", cached);
--
2.35.3
On 2022/8/16 16:44, Kefeng Wang wrote:
> From: Liu Shixin <[email protected]>
>
> The page on pcplist could be used, but not counted into memory free or
> avaliable, and pcp_free is only showed by show_mem(). Since commit
> d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
> significant decrease in the display of free memory, with a large number
> of cpus and nodes, the number of pages in the percpu list can be very
> large, so it is better to let user to know the pcp count.
Add more experts according to commit d8a759b57035，
any advice would be much appreciated，thanks.
> Signed-off-by: Liu Shixin <[email protected]>
> Signed-off-by: Kefeng Wang <[email protected]>
> ---
> drivers/base/node.c | 14 +++++++++++++-
> fs/proc/meminfo.c | 9 +++++++++
> 2 files changed, 22 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/base/node.c b/drivers/base/node.c
> index eb0f43784c2b..846864e45db6 100644
> --- a/drivers/base/node.c
> +++ b/drivers/base/node.c
> @@ -375,6 +375,9 @@ static ssize_t node_read_meminfo(struct device *dev,
> struct sysinfo i;
> unsigned long sreclaimable, sunreclaimable;
> unsigned long swapcached = 0;
> + unsigned long free_pcp = 0;
> + struct zone *zone;
> + int cpu;
>
> si_meminfo_node(&i, nid);
> sreclaimable = node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B);
> @@ -382,9 +385,17 @@ static ssize_t node_read_meminfo(struct device *dev,
> #ifdef CONFIG_SWAP
> swapcached = node_page_state_pages(pgdat, NR_SWAPCACHE);
> #endif
> + for_each_populated_zone(zone) {
> + if (zone_to_nid(zone) != nid)
> + continue;
> + for_each_online_cpu(cpu)
> + free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
> + }
> +
> len = sysfs_emit_at(buf, len,
> "Node %d MemTotal: %8lu kB\n"
> "Node %d MemFree: %8lu kB\n"
> + "Node %d PcpFree: %8lu kB\n"
> "Node %d MemUsed: %8lu kB\n"
> "Node %d SwapCached: %8lu kB\n"
> "Node %d Active: %8lu kB\n"
> @@ -397,7 +408,8 @@ static ssize_t node_read_meminfo(struct device *dev,
> "Node %d Mlocked: %8lu kB\n",
> nid, K(i.totalram),
> nid, K(i.freeram),
> - nid, K(i.totalram - i.freeram),
> + nid, K(free_pcp),
> + nid, K(i.totalram - i.freeram - free_pcp),
> nid, K(swapcached),
> nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
> node_page_state(pgdat, NR_ACTIVE_FILE)),
> diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c
> index 6e89f0e2fd20..672c784dfc8a 100644
> --- a/fs/proc/meminfo.c
> +++ b/fs/proc/meminfo.c
> @@ -38,6 +38,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
> unsigned long pages[NR_LRU_LISTS];
> unsigned long sreclaimable, sunreclaim;
> int lru;
> + unsigned long free_pcp = 0;
> + struct zone *zone;
> + int cpu;
>
> si_meminfo(&i);
> si_swapinfo(&i);
> @@ -55,8 +58,14 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
> sreclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B);
> sunreclaim = global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B);
>
> + for_each_populated_zone(zone) {
> + for_each_online_cpu(cpu)
> + free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
> + }
> +
> show_val_kb(m, "MemTotal: ", i.totalram);
> show_val_kb(m, "MemFree: ", i.freeram);
> + show_val_kb(m, "PcpFree: ", free_pcp);
> show_val_kb(m, "MemAvailable: ", available);
> show_val_kb(m, "Buffers: ", i.bufferram);
> show_val_kb(m, "Cached: ", cached);
On 8/17/22 00:16, Kefeng Wang wrote:
> On 2022/8/16 16:44, Kefeng Wang wrote:
>> From: Liu Shixin <[email protected]>
>>
>> The page on pcplist could be used, but not counted into memory free or
>> avaliable, and pcp_free is only showed by show_mem(). Since commit
>> d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
>> significant decrease in the display of free memory, with a large number
>> of cpus and nodes, the number of pages in the percpu list can be very
>> large, so it is better to let user to know the pcp count.
> Add more experts according to commit d8a759b57035，
> any advice would be much appreciated，thanks.
Adding a new meminfo field seems like overkill. I'd just make this a
part of MemFree if anything.
Also, some actual data would be nice here. Like:
Before commit d8a759b57035, the maximum amount of pages in the
pcp lists was theoretically $FOO MB. After the patch, the lists
can hold $BAR MB. It has been observed to be $BAZ MB in
practice.
This was all on a system with $X memory NUMA nodes and $Y CPUs.
On 2022/8/19 5:07, Dave Hansen wrote:
> On 8/17/22 00:16, Kefeng Wang wrote:
>> On 2022/8/16 16:44, Kefeng Wang wrote:
>>> From: Liu Shixin <[email protected]>
>>>
>>> The page on pcplist could be used, but not counted into memory free or
>>> avaliable, and pcp_free is only showed by show_mem(). Since commit
>>> d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
>>> significant decrease in the display of free memory, with a large number
>>> of cpus and nodes, the number of pages in the percpu list can be very
>>> large, so it is better to let user to know the pcp count.
>> Add more experts according to commit d8a759b57035，
>> any advice would be much appreciated，thanks.
> Adding a new meminfo field seems like overkill. I'd just make this a
> part of MemFree if anything.
I like this way too.
> Also, some actual data would be nice here. Like:
>
> Before commit d8a759b57035, the maximum amount of pages in the
> pcp lists was theoretically $FOO MB. After the patch, the lists
> can hold $BAR MB. It has been observed to be $BAZ MB in
> practice.
>
> This was all on a system with $X memory NUMA nodes and $Y CPUs.
>
> .
Same question in [1], will repost with update, thanks.
[1]
https://lore.kernel.org/linux-mm/[email protected]/t/#m50cf15911d9d203bd97238512fa2ae9ba1bd9e1e
On Tue, Aug 16, 2022 at 05:24:07PM +0800, Kefeng Wang wrote:
>
> On 2022/8/16 16:48, huang ying wrote:
> > On Tue, Aug 16, 2022 at 4:38 PM Kefeng Wang <[email protected]> wrote:
> > > From: Liu Shixin <[email protected]>
> > >
> > > The page on pcplist could be used, but not counted into memory free or
> > > avaliable, and pcp_free is only showed by show_mem(). Since commit
> > > d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
> > > significant decrease in the display of free memory, with a large number
> > > of cpus and nodes, the number of pages in the percpu list can be very
> > > large, so it is better to let user to know the pcp count.
> > Can you show some data?
>
> 80M+ with 72cpus/2node
>
80M+ for a 2 node system doesn't sound like a significant number.
> >
> > Another choice is to count PCP free pages in MemFree. Is that OK for
> > your use case too?
>
> Yes, the user will make policy according to MemFree, we think count PCP free
> pages
>
> in MemFree is better, but don't know whether it is right way.
>
Is there a real problem where user makes a sub-optimal policy due to the
not accounted 80M+ free memory?
Counting PCP pages as free seems natural, since they are indeed free
pages. One concern is, there might be much more calls of
__mod_zone_freepage_state() if you do free page counting for PCP pages,
not sure if that would hurt performance. Also, you will need to
differentiate in __free_one_page() whether counting for free pages are
still needed since some pages are freed through PCP(and thus already
counted) while some are not.
BTW, since commit b92ca18e8ca59("mm/page_alloc: disassociate the pcp->high
from pcp->batch"), pcp size is no longer associated with batch size. Is
it that you are testing on an older kernel?
Thanks,
Aaron
> > > Signed-off-by: Liu Shixin <[email protected]>
> > > Signed-off-by: Kefeng Wang <[email protected]>
> > > ---
> > > drivers/base/node.c | 14 +++++++++++++-
> > > fs/proc/meminfo.c | 9 +++++++++
> > > 2 files changed, 22 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/drivers/base/node.c b/drivers/base/node.c
> > > index eb0f43784c2b..846864e45db6 100644
> > > --- a/drivers/base/node.c
> > > +++ b/drivers/base/node.c
> > > @@ -375,6 +375,9 @@ static ssize_t node_read_meminfo(struct device *dev,
> > > struct sysinfo i;
> > > unsigned long sreclaimable, sunreclaimable;
> > > unsigned long swapcached = 0;
> > > + unsigned long free_pcp = 0;
> > > + struct zone *zone;
> > > + int cpu;
> > >
> > > si_meminfo_node(&i, nid);
> > > sreclaimable = node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B);
> > > @@ -382,9 +385,17 @@ static ssize_t node_read_meminfo(struct device *dev,
> > > #ifdef CONFIG_SWAP
> > > swapcached = node_page_state_pages(pgdat, NR_SWAPCACHE);
> > > #endif
> > > + for_each_populated_zone(zone) {
> > > + if (zone_to_nid(zone) != nid)
> > > + continue;
> > > + for_each_online_cpu(cpu)
> > > + free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
> > > + }
> > > +
> > > len = sysfs_emit_at(buf, len,
> > > "Node %d MemTotal: %8lu kB\n"
> > > "Node %d MemFree: %8lu kB\n"
> > > + "Node %d PcpFree: %8lu kB\n"
> > > "Node %d MemUsed: %8lu kB\n"
> > > "Node %d SwapCached: %8lu kB\n"
> > > "Node %d Active: %8lu kB\n"
> > > @@ -397,7 +408,8 @@ static ssize_t node_read_meminfo(struct device *dev,
> > > "Node %d Mlocked: %8lu kB\n",
> > > nid, K(i.totalram),
> > > nid, K(i.freeram),
> > > - nid, K(i.totalram - i.freeram),
> > > + nid, K(free_pcp),
> > > + nid, K(i.totalram - i.freeram - free_pcp),
> > > nid, K(swapcached),
> > > nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
> > > node_page_state(pgdat, NR_ACTIVE_FILE)),
> > > diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c
> > > index 6e89f0e2fd20..672c784dfc8a 100644
> > > --- a/fs/proc/meminfo.c
> > > +++ b/fs/proc/meminfo.c
> > > @@ -38,6 +38,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
> > > unsigned long pages[NR_LRU_LISTS];
> > > unsigned long sreclaimable, sunreclaim;
> > > int lru;
> > > + unsigned long free_pcp = 0;
> > > + struct zone *zone;
> > > + int cpu;
> > >
> > > si_meminfo(&i);
> > > si_swapinfo(&i);
> > > @@ -55,8 +58,14 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
> > > sreclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B);
> > > sunreclaim = global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B);
> > >
> > > + for_each_populated_zone(zone) {
> > > + for_each_online_cpu(cpu)
> > > + free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
> > > + }
> > > +
> > > show_val_kb(m, "MemTotal: ", i.totalram);
> > > show_val_kb(m, "MemFree: ", i.freeram);
> > > + show_val_kb(m, "PcpFree: ", free_pcp);
> > > show_val_kb(m, "MemAvailable: ", available);
> > > show_val_kb(m, "Buffers: ", i.bufferram);
> > > show_val_kb(m, "Cached: ", cached);
> > > --
> > > 2.35.3
> > >
> > >
> > .
On 8/19/2022 5:53 PM, Liu Shixin wrote:
>
>
> On 2022/8/19 15:40, Aaron Lu wrote:
>> On Tue, Aug 16, 2022 at 05:24:07PM +0800, Kefeng Wang wrote:
>>> On 2022/8/16 16:48, huang ying wrote:
>>>> On Tue, Aug 16, 2022 at 4:38 PM Kefeng Wang <[email protected]> wrote:
>>>>> From: Liu Shixin <[email protected]>
>>>>>
>>>>> The page on pcplist could be used, but not counted into memory free or
>>>>> avaliable, and pcp_free is only showed by show_mem(). Since commit
>>>>> d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
>>>>> significant decrease in the display of free memory, with a large number
>>>>> of cpus and nodes, the number of pages in the percpu list can be very
>>>>> large, so it is better to let user to know the pcp count.
>>>> Can you show some data?
>>> 80M+ with 72cpus/2node
>>>
>> 80M+ for a 2 node system doesn't sound like a significant number.
>>
>>>> Another choice is to count PCP free pages in MemFree. Is that OK for
>>>> your use case too?
>>> Yes, the user will make policy according to MemFree, we think count PCP free
>>> pages
>>>
>>> in MemFree is better, but don't know whether it is right way.
>>>
>> Is there a real problem where user makes a sub-optimal policy due to the
>> not accounted 80M+ free memory?
> I need to explain that 80M+ is the increased after patch d8a759b57035. Actually in my test,
> the pcplist is about 114M after system startup, and in high load scenarios, the pcplist memory
> can reach 300M+.
> The downstream has sensed the memory change after the kernel is updated, which has an
> actual impact on them. That's why I sent this patch to ask if should count this
> part of memory.
>
>> Counting PCP pages as free seems natural, since they are indeed free
>> pages. One concern is, there might be much more calls of
>> __mod_zone_freepage_state() if you do free page counting for PCP pages,
>> not sure if that would hurt performance. Also, you will need to
>> differentiate in __free_one_page() whether counting for free pages are
>> still needed since some pages are freed through PCP(and thus already
>> counted) while some are not.
> I prefer to add this part of memory into free only when calculating MemFree and
> MemAvailable, without modifying other statistics to avoid directly hurt performance
> or cause other performance problems. How about this way?
>
That sounds workable to me.
Thanks,
Aaron
>> BTW, since commit b92ca18e8ca59("mm/page_alloc: disassociate the pcp->high
>> from pcp->batch"), pcp size is no longer associated with batch size. Is
>> it that you are testing on an older kernel?
> I met the problem on stable-5.10. I think this patch can't fix the problem I met. Futher,
> this series of patch make pcp->high to be greater. So the problem becomes even more
> acute in mainline.
>
> Thanks,
>>
>> Thanks,
>> Aaron
>>
>>>>> Signed-off-by: Liu Shixin <[email protected]>
>>>>> Signed-off-by: Kefeng Wang <[email protected]>
>>>>> ---
>>>>> drivers/base/node.c | 14 +++++++++++++-
>>>>> fs/proc/meminfo.c | 9 +++++++++
>>>>> 2 files changed, 22 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/drivers/base/node.c b/drivers/base/node.c
>>>>> index eb0f43784c2b..846864e45db6 100644
>>>>> --- a/drivers/base/node.c
>>>>> +++ b/drivers/base/node.c
>>>>> @@ -375,6 +375,9 @@ static ssize_t node_read_meminfo(struct device *dev,
>>>>> struct sysinfo i;
>>>>> unsigned long sreclaimable, sunreclaimable;
>>>>> unsigned long swapcached = 0;
>>>>> + unsigned long free_pcp = 0;
>>>>> + struct zone *zone;
>>>>> + int cpu;
>>>>>
>>>>> si_meminfo_node(&i, nid);
>>>>> sreclaimable = node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B);
>>>>> @@ -382,9 +385,17 @@ static ssize_t node_read_meminfo(struct device *dev,
>>>>> #ifdef CONFIG_SWAP
>>>>> swapcached = node_page_state_pages(pgdat, NR_SWAPCACHE);
>>>>> #endif
>>>>> + for_each_populated_zone(zone) {
>>>>> + if (zone_to_nid(zone) != nid)
>>>>> + continue;
>>>>> + for_each_online_cpu(cpu)
>>>>> + free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
>>>>> + }
>>>>> +
>>>>> len = sysfs_emit_at(buf, len,
>>>>> "Node %d MemTotal: %8lu kB\n"
>>>>> "Node %d MemFree: %8lu kB\n"
>>>>> + "Node %d PcpFree: %8lu kB\n"
>>>>> "Node %d MemUsed: %8lu kB\n"
>>>>> "Node %d SwapCached: %8lu kB\n"
>>>>> "Node %d Active: %8lu kB\n"
>>>>> @@ -397,7 +408,8 @@ static ssize_t node_read_meminfo(struct device *dev,
>>>>> "Node %d Mlocked: %8lu kB\n",
>>>>> nid, K(i.totalram),
>>>>> nid, K(i.freeram),
>>>>> - nid, K(i.totalram - i.freeram),
>>>>> + nid, K(free_pcp),
>>>>> + nid, K(i.totalram - i.freeram - free_pcp),
>>>>> nid, K(swapcached),
>>>>> nid, K(node_page_state(pgdat, NR_ACTIVE_ANON) +
>>>>> node_page_state(pgdat, NR_ACTIVE_FILE)),
>>>>> diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c
>>>>> index 6e89f0e2fd20..672c784dfc8a 100644
>>>>> --- a/fs/proc/meminfo.c
>>>>> +++ b/fs/proc/meminfo.c
>>>>> @@ -38,6 +38,9 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
>>>>> unsigned long pages[NR_LRU_LISTS];
>>>>> unsigned long sreclaimable, sunreclaim;
>>>>> int lru;
>>>>> + unsigned long free_pcp = 0;
>>>>> + struct zone *zone;
>>>>> + int cpu;
>>>>>
>>>>> si_meminfo(&i);
>>>>> si_swapinfo(&i);
>>>>> @@ -55,8 +58,14 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
>>>>> sreclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B);
>>>>> sunreclaim = global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B);
>>>>>
>>>>> + for_each_populated_zone(zone) {
>>>>> + for_each_online_cpu(cpu)
>>>>> + free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
>>>>> + }
>>>>> +
>>>>> show_val_kb(m, "MemTotal: ", i.totalram);
>>>>> show_val_kb(m, "MemFree: ", i.freeram);
>>>>> + show_val_kb(m, "PcpFree: ", free_pcp);
>>>>> show_val_kb(m, "MemAvailable: ", available);
>>>>> show_val_kb(m, "Buffers: ", i.bufferram);
>>>>> show_val_kb(m, "Cached: ", cached);
>>>>> --
>>>>> 2.35.3
>>>>>
>>>>>
>>>> .
>> .
>>
>
On Fri 19-08-22 17:53:27, Liu Shixin wrote:
>
>
> On 2022/8/19 15:40, Aaron Lu wrote:
> > On Tue, Aug 16, 2022 at 05:24:07PM +0800, Kefeng Wang wrote:
> >> On 2022/8/16 16:48, huang ying wrote:
> >>> On Tue, Aug 16, 2022 at 4:38 PM Kefeng Wang <[email protected]> wrote:
> >>>> From: Liu Shixin <[email protected]>
> >>>>
> >>>> The page on pcplist could be used, but not counted into memory free or
> >>>> avaliable, and pcp_free is only showed by show_mem(). Since commit
> >>>> d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
> >>>> significant decrease in the display of free memory, with a large number
> >>>> of cpus and nodes, the number of pages in the percpu list can be very
> >>>> large, so it is better to let user to know the pcp count.
> >>> Can you show some data?
> >> 80M+ with 72cpus/2node
I would expect that such system would have quite some memory as well and
80MB wouldn't be a really noticeable. What is that amount in %tage
> > 80M+ for a 2 node system doesn't sound like a significant number.
> >
> >>> Another choice is to count PCP free pages in MemFree. Is that OK for
> >>> your use case too?
> >> Yes, the user will make policy according to MemFree, we think count PCP free
> >> pages
> >>
> >> in MemFree is better, but don't know whether it is right way.
> >>
> > Is there a real problem where user makes a sub-optimal policy due to the
> > not accounted 80M+ free memory?
> I need to explain that 80M+ is the increased after patch d8a759b57035. Actually in my test,
> the pcplist is about 114M after system startup, and in high load scenarios, the pcplist memory
> can reach 300M+.
> The downstream has sensed the memory change after the kernel is updated, which has an
> actual impact on them. That's why I sent this patch to ask if should count this
> part of memory.
It would be really great to be more explicit about this. Because if this
is really runtime noticeable then we might need to consider an improved
tunning or way to manually configure the pcp batch sizes. Reporting the
amount on its own is unlikely going to help without being able to do
anything about that.
--
Michal Hocko
SUSE Labs