Skip to contentSkip to conversation

Regression from 5.7.17 to 5.9.9 with memory.low cgroup constraints

DISCUSSION Bruno Prémont 3 participants 8 messages Nov 25, 2020 – Dec 10, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Hello,

On a production system I've encountered a rather harsh behavior from kernel in the context of memory cgroup (v2) after updating kernel from 5.7 series to 5.9 series.

It seems like kernel is reclaiming file cache but leaving inode cache (reclaimable slabs) alone in a way that the server ends up trashing and maxing out on IO to one of its disks instead of doing actual work.

My setup, server has 64G of RAM:
root

   + system        { min=0, low=128M, high=8G, max=8G }
     + base        { no specific constraints }
     + backup      { min=0, low=32M, high=2G, max=2G }
     + shell       { no specific constraints }
  + websrv         { min=0, low=4G, high=32G, max=32G }
  + website        { min=0, low=16G, high=40T, max=40T }
    + website1     { min=0, low=64M, high=2G, max=2G }
    + website2     { min=0, low=64M, high=2G, max=2G }
      ...
  + remote         { min=0, low=1G, high=14G, max=14G }
    + webuser1     { min=0, low=64M, high=2G, max=2G }
    + webuser2     { min=0, low=64M, high=2G, max=2G }
      ...

When the server was struggling I've had mostly IO on disk hosting system processes and some cache files of websrv processes. It seems that running backup does make the issue much more probable.

The processes in websrv are the most impacted by the trashing and this is the one with lots of disk cache and inode cache assigned to it. (note a helper running in websrv cgroup scan whole file system hierarchy once per hour and this keeps inode cache pretty filled. Dropping just file cache (about 10G) did not unlock situation but dropping reclaimable slabs (inode cache, about 30G) got the system back running.

Some metrics I have collected during a trashing period (metrics collected at about 5min interval) - I don't have ful memory.stat unfortunately:

system/memory.min              0              = 0
system/memory.low              134217728      = 134217728
system/memory.high             8589934592     = 8589934592
system/memory.max              8589934592     = 8589934592
system/memory.pressure
    some avg10=54.41 avg60=59.28 avg300=69.46 total=7347640237
    full avg10=27.45 avg60=22.19 avg300=29.28 total=3287847481
  ->
    some avg10=77.25 avg60=73.24 avg300=69.63 total=7619662740
    full avg10=23.04 avg60=25.26 avg300=27.97 total=3401421903
system/memory.current          262533120      < 263929856
system/memory.events.local
    low                        5399469        = 5399469
    high                       0              = 0
    max                        112303         = 112303
    oom                        0              = 0
    oom_kill                   0              = 0
system/base/memory.min         0              = 0
system/base/memory.low         0              = 0
system/base/memory.high        max            = max
system/base/memory.max         max            = max
system/base/memory.pressure
    some avg10=18.89 avg60=20.34 avg300=24.95 total=5156816349
    full avg10=10.90 avg60=8.50 avg300=11.68 total=2253916169
  ->
    some avg10=33.82 avg60=32.26 avg300=26.95 total=5258381824
    full avg10=12.51 avg60=13.01 avg300=12.05 total=2301375471
system/base/memory.current     31363072       < 32243712
system/base/memory.events.local
    low                        0              = 0
    high                       0              = 0
    max                        0              = 0
    oom                        0              = 0
    oom_kill                   0              = 0
system/backup/memory.min       0              = 0
system/backup/memory.low       33554432       = 33554432
system/backup/memory.high      2147483648     = 2147483648
system/backup/memory.max       2147483648     = 2147483648
system/backup/memory.pressure
    some avg10=41.73 avg60=45.97 avg300=56.27 total=3385780085
    full avg10=21.78 avg60=18.15 avg300=25.35 total=1571263731
  ->
    some avg10=60.27 avg60=55.44 avg300=54.37 total=3599850643
    full avg10=19.52 avg60=20.91 avg300=23.58 total=1667430954
system/backup/memory.current  222130176       < 222543872
system/backup/memory.events.local
    low                       5446            = 5446
    high                      0               = 0
    max                       0               = 0
    oom                       0               = 0
    oom_kill                  0               = 0
system/shell/memory.min       0               = 0
system/shell/memory.low       0               = 0
system/shell/memory.high      max             = max
system/shell/memory.max       max             = max
system/shell/memory.pressure
    some avg10=0.00 avg60=0.12 avg300=0.25 total=1348427661
    full avg10=0.00 avg60=0.04 avg300=0.06 total=493582108
  ->
    some avg10=0.00 avg60=0.00 avg300=0.06 total=1348516773
    full avg10=0.00 avg60=0.00 avg300=0.00 total=493591500
system/shell/memory.current  8814592          < 8888320
system/shell/memory.events.local
    low                      0                = 0
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0
website/memory.min           0                = 0
website/memory.low           17179869184      = 17179869184
website/memory.high          45131717672960   = 45131717672960
website/memory.max           45131717672960   = 45131717672960
website/memory.pressure
    some avg10=0.00 avg60=0.00 avg300=0.00 total=415009408
    full avg10=0.00 avg60=0.00 avg300=0.00 total=201868483
  ->
    some avg10=0.00 avg60=0.00 avg300=0.00 total=415009408
    full avg10=0.00 avg60=0.00 avg300=0.00 total=201868483
website/memory.current       11811520512      > 11456942080
website/memory.events.local
    low                      11372142         < 11377350
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0
remote/memory.min            0
remote/memory.low            1073741824
remote/memory.high           15032385536
remote/memory.max            15032385536
remote/memory.pressure
    some avg10=0.00 avg60=0.25 avg300=0.50 total=2017364408
    full avg10=0.00 avg60=0.00 avg300=0.01 total=738071296
  ->
remote/memory.current        84439040         > 81797120
remote/memory.events.local
    low                      11372142         < 11377350
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0
websrv/memory.min            0                = 0
websrv/memory.low            4294967296       = 4294967296
websrv/memory.high           34359738368      = 34359738368
websrv/memory.max            34426847232      = 34426847232
websrv/memory.pressure
    some avg10=40.38 avg60=62.58 avg300=68.83 total=7760096704
    full avg10=7.80 avg60=10.78 avg300=12.64 total=2254679370
  ->
    some avg10=89.97 avg60=83.78 avg300=72.99 total=8040513640
    full avg10=11.46 avg60=11.49 avg300=11.47 total=2300116237
websrv/memory.current        18421673984      < 18421936128
websrv/memory.events.local
    low                      0                = 0
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0

Is there something important I'm missing in my setup that could prevent things from starving?

Did memory.low meaning change between 5.7 and 5.9? From behavior it feels as if inodes are not accounted to cgroup at all and kernel pushes cgroups down to their memory.low by killing file cache if there is not enough free memory to hold all promises (and not only when a cgroup tries to use up to its promised amount of memory).
As system was trashing as much with 10G of file cache dropped (completely unused memory) as with it in use.

I will try to create a test-case for it to reproduce it on a test machine an be able to verify a fix or eventually bisect to triggering patch though it this all rings a bell, please tell!

Note until I have a test-case I'm reluctant to just wait [on production system] for next occurrence (usually at unpractical times) to gather some more metrics.

Regards,
Bruno

#1 lore raw

Hi,
thanks for the detailed report.

On Wed 25-11-20 12:39:56, Bruno Prémont wrote:
[...]

Quoted 1 line

Did memory.low meaning change between 5.7 and 5.9?

The latest semantic change in the low limit protection semantic was introduced in 5.7 (recursive protection) but it requires an explicit enablinig.

Quoted 5 lines

From behavior it
feels as if inodes are not accounted to cgroup at all and kernel pushes cgroups down to their memory.low by killing file cache if there is not enough free memory to hold all promises (and not only when a cgroup tries to use up to its promised amount of memory).

Your counters indeed show that the low protection has been breached, most likely because the reclaim couldn't make any progress. Considering that this is the case for all/most of your cgroups it suggests that the memory pressure was global rather than limit imposed. In fact even top level cgroups got reclaimed below the low limit.

This suggests that this is not likely to be memcg specific. It is more likely that this is a general memory reclaim regression for your workload. There were larger changes in that area. Be it lru balancing based on cost model by Johannes or working set tracking for anonymous pages by Joonsoo. Maybe even more. Both of them can influence page cache reclaim but you are suggesting that slab accounted memory is not reclaimed properly. I am not sure sure there were considerable changes there. Would it be possible to collect /prov/vmstat as well?

Signature · 2 lines
Michal Hocko
SUSE Labs
#2 lore raw

Hi Michal,

Quoted 10 lines · On Wed, 25 Nov 2020 14:37:40 +0100 Michal Hocko…

Hi,
thanks for the detailed report.

On Wed 25-11-20 12:39:56, Bruno Prémont wrote:
[...]

Quoted 1 line

Did memory.low meaning change between 5.7 and 5.9?

The latest semantic change in the low limit protection semantic was introduced in 5.7 (recursive protection) but it requires an explicit enablinig.

No specific mount options set for v2 cgroup, so not active.

Quoted 11 lines
Quoted 5 lines

From behavior it
feels as if inodes are not accounted to cgroup at all and kernel pushes cgroups down to their memory.low by killing file cache if there is not enough free memory to hold all promises (and not only when a cgroup tries to use up to its promised amount of memory).

Your counters indeed show that the low protection has been breached, most likely because the reclaim couldn't make any progress. Considering that this is the case for all/most of your cgroups it suggests that the memory pressure was global rather than limit imposed. In fact even top level cgroups got reclaimed below the low limit.

Note that the "original" counters we partially triggered by a first event where I had one cgroup (websrv) of the with a rather very high memory.low (16G or even 32G) which caused counters everywhere to increase.

So before the last trashing during which the values were collected the event counters and current looked as follows:

system/memory.pressure
some avg10=0.04 avg60=0.28 avg300=0.12 total=5844917510 full avg10=0.04 avg60=0.26 avg300=0.11 total=2439353404
system/memory.current
96432128
system/memory.events.local

  low      5399469   (unchanged)
  high     0
  max      112303    (unchanged)
  oom      0

oom_kill 0

system/base/memory.pressure
some avg10=0.04 avg60=0.28 avg300=0.12 total=4589562039 full avg10=0.04 avg60=0.28 avg300=0.12 total=1926984197
system/base/memory.current
59305984
system/base/memory.events.local

  low      0   (unchanged)
  high     0
  max      0   (unchanged)
  oom      0

oom_kill 0

system/backup/memory.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=2123293649 full avg10=0.00 avg60=0.00 avg300=0.00 total=815450446
system/backup/memory.current
32444416
system/backup/memory.events.local

  low      5446   (unchanged)
  high     0
  max      0
  oom      0

oom_kill 0

system/shell/memory.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=1345965660 full avg10=0.00 avg60=0.00 avg300=0.00 total=492812915
system/shell/memory.current
4571136
system/shell/memory.events.local

  low      0
  high     0
  max      0
  oom      0

oom_kill 0

website/memory.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=415008878 full avg10=0.00 avg60=0.00 avg300=0.00 total=201868483
website/memory.current
12104380416
website/memory.events.local

  low      11264569  (during trashing: 11372142 then 11377350)
  high     0
  max      0
  oom      0

oom_kill 0

remote/memory.pressure
some avg10=0.00 avg60=0.00 avg300=0.00 total=2005130126 full avg10=0.00 avg60=0.00 avg300=0.00 total=735366752
remote/memory.current
116330496
remote/memory.events.local

  low      11264569  (during trashing: 11372142 then 11377350)
  high     0
  max      0
  oom      0

oom_kill 0

websrv/memory.pressure
some avg10=0.02 avg60=0.11 avg300=0.03 total=6650355162 full avg10=0.02 avg60=0.11 avg300=0.03 total=2034584579
websrv/memory.current
18483359744
websrv/memory.events.local

  low      0
  high     0
  max      0
  oom      0

oom_kill 0

Quoted 7 lines

This suggests that this is not likely to be memcg specific. It is more likely that this is a general memory reclaim regression for your workload. There were larger changes in that area. Be it lru balancing based on cost model by Johannes or working set tracking for anonymous pages by Joonsoo. Maybe even more. Both of them can influence page cache reclaim but you are suggesting that slab accounted memory is not reclaimed properly.

That is my impression, yes. No idea though if memcg can influence the way reclaim tries to perform its work or if slab_reclaimable not associated to any (child) cg would somehow be excluded from reclaim.

Quoted 2 lines

I am not sure sure there were considerable changes there. Would it be possible to collect /prov/vmstat as well?

I will have a look at gathering memory.stat and /proc/vmstat at next opportunity.
Will first try with a test system with not too much memory and lots of files to reproduce about 50% of memory usage by slab_reclaimable and see how far I get.

Thanks,
Bruno

#3 lore raw
In reply to Bruno Prémont
Quoted 185 lines · Bruno Prémont

Hello,

On a production system I've encountered a rather harsh behavior from kernel in the context of memory cgroup (v2) after updating kernel from 5.7 series to 5.9 series.

It seems like kernel is reclaiming file cache but leaving inode cache (reclaimable slabs) alone in a way that the server ends up trashing and maxing out on IO to one of its disks instead of doing actual work.

My setup, server has 64G of RAM:
root

   + system        { min=0, low=128M, high=8G, max=8G }
     + base        { no specific constraints }
     + backup      { min=0, low=32M, high=2G, max=2G }
     + shell       { no specific constraints }
  + websrv         { min=0, low=4G, high=32G, max=32G }
  + website        { min=0, low=16G, high=40T, max=40T }
    + website1     { min=0, low=64M, high=2G, max=2G }
    + website2     { min=0, low=64M, high=2G, max=2G }
      ...
  + remote         { min=0, low=1G, high=14G, max=14G }
    + webuser1     { min=0, low=64M, high=2G, max=2G }
    + webuser2     { min=0, low=64M, high=2G, max=2G }
      ...

When the server was struggling I've had mostly IO on disk hosting system processes and some cache files of websrv processes. It seems that running backup does make the issue much more probable.

The processes in websrv are the most impacted by the trashing and this is the one with lots of disk cache and inode cache assigned to it. (note a helper running in websrv cgroup scan whole file system hierarchy once per hour and this keeps inode cache pretty filled. Dropping just file cache (about 10G) did not unlock situation but dropping reclaimable slabs (inode cache, about 30G) got the system back running.

Some metrics I have collected during a trashing period (metrics collected at about 5min interval) - I don't have ful memory.stat unfortunately:

system/memory.min              0              = 0
system/memory.low              134217728      = 134217728
system/memory.high             8589934592     = 8589934592
system/memory.max              8589934592     = 8589934592
system/memory.pressure
    some avg10=54.41 avg60=59.28 avg300=69.46 total=7347640237
    full avg10=27.45 avg60=22.19 avg300=29.28 total=3287847481
  ->
    some avg10=77.25 avg60=73.24 avg300=69.63 total=7619662740
    full avg10=23.04 avg60=25.26 avg300=27.97 total=3401421903
system/memory.current          262533120      < 263929856
system/memory.events.local
    low                        5399469        = 5399469
    high                       0              = 0
    max                        112303         = 112303
    oom                        0              = 0
    oom_kill                   0              = 0
system/base/memory.min         0              = 0
system/base/memory.low         0              = 0
system/base/memory.high        max            = max
system/base/memory.max         max            = max
system/base/memory.pressure
    some avg10=18.89 avg60=20.34 avg300=24.95 total=5156816349
    full avg10=10.90 avg60=8.50 avg300=11.68 total=2253916169
  ->
    some avg10=33.82 avg60=32.26 avg300=26.95 total=5258381824
    full avg10=12.51 avg60=13.01 avg300=12.05 total=2301375471
system/base/memory.current     31363072       < 32243712
system/base/memory.events.local
    low                        0              = 0
    high                       0              = 0
    max                        0              = 0
    oom                        0              = 0
    oom_kill                   0              = 0
system/backup/memory.min       0              = 0
system/backup/memory.low       33554432       = 33554432
system/backup/memory.high      2147483648     = 2147483648
system/backup/memory.max       2147483648     = 2147483648
system/backup/memory.pressure
    some avg10=41.73 avg60=45.97 avg300=56.27 total=3385780085
    full avg10=21.78 avg60=18.15 avg300=25.35 total=1571263731
  ->
    some avg10=60.27 avg60=55.44 avg300=54.37 total=3599850643
    full avg10=19.52 avg60=20.91 avg300=23.58 total=1667430954
system/backup/memory.current  222130176       < 222543872
system/backup/memory.events.local
    low                       5446            = 5446
    high                      0               = 0
    max                       0               = 0
    oom                       0               = 0
    oom_kill                  0               = 0
system/shell/memory.min       0               = 0
system/shell/memory.low       0               = 0
system/shell/memory.high      max             = max
system/shell/memory.max       max             = max
system/shell/memory.pressure
    some avg10=0.00 avg60=0.12 avg300=0.25 total=1348427661
    full avg10=0.00 avg60=0.04 avg300=0.06 total=493582108
  ->
    some avg10=0.00 avg60=0.00 avg300=0.06 total=1348516773
    full avg10=0.00 avg60=0.00 avg300=0.00 total=493591500
system/shell/memory.current  8814592          < 8888320
system/shell/memory.events.local
    low                      0                = 0
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0
website/memory.min           0                = 0
website/memory.low           17179869184      = 17179869184
website/memory.high          45131717672960   = 45131717672960
website/memory.max           45131717672960   = 45131717672960
website/memory.pressure
    some avg10=0.00 avg60=0.00 avg300=0.00 total=415009408
    full avg10=0.00 avg60=0.00 avg300=0.00 total=201868483
  ->
    some avg10=0.00 avg60=0.00 avg300=0.00 total=415009408
    full avg10=0.00 avg60=0.00 avg300=0.00 total=201868483
website/memory.current       11811520512      > 11456942080
website/memory.events.local
    low                      11372142         < 11377350
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0
remote/memory.min            0
remote/memory.low            1073741824
remote/memory.high           15032385536
remote/memory.max            15032385536
remote/memory.pressure
    some avg10=0.00 avg60=0.25 avg300=0.50 total=2017364408
    full avg10=0.00 avg60=0.00 avg300=0.01 total=738071296
  ->
remote/memory.current        84439040         > 81797120
remote/memory.events.local
    low                      11372142         < 11377350
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0
websrv/memory.min            0                = 0
websrv/memory.low            4294967296       = 4294967296
websrv/memory.high           34359738368      = 34359738368
websrv/memory.max            34426847232      = 34426847232
websrv/memory.pressure
    some avg10=40.38 avg60=62.58 avg300=68.83 total=7760096704
    full avg10=7.80 avg60=10.78 avg300=12.64 total=2254679370
  ->
    some avg10=89.97 avg60=83.78 avg300=72.99 total=8040513640
    full avg10=11.46 avg60=11.49 avg300=11.47 total=2300116237
websrv/memory.current        18421673984      < 18421936128
websrv/memory.events.local
    low                      0                = 0
    high                     0                = 0
    max                      0                = 0
    oom                      0                = 0
    oom_kill                 0                = 0

Is there something important I'm missing in my setup that could prevent things from starving?

Did memory.low meaning change between 5.7 and 5.9? From behavior it feels as if inodes are not accounted to cgroup at all and kernel pushes cgroups down to their memory.low by killing file cache if there is not enough free memory to hold all promises (and not only when a cgroup tries to use up to its promised amount of memory).
As system was trashing as much with 10G of file cache dropped (completely unused memory) as with it in use.

Quoted 7 lines

I will try to create a test-case for it to reproduce it on a test machine an be able to verify a fix or eventually bisect to triggering patch though it this all rings a bell, please tell!

Note until I have a test-case I'm reluctant to just wait [on production system] for next occurrence (usually at unpractical times) to gather some more metrics.

Hi Bruno!

Thank you for the report.

Can you, please, check if the following patch fixes the issue?

Thanks!

--

mm/slab.h +6 -2
index 6cc323f1313a..ef02b841bcd8 100644
@@ -290,7 +290,7 @@ static inline struct obj_cgroup *memcg_slab_pre_alloc_hook(struct kmem_cache *s,
 
        if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s))) {
                obj_cgroup_put(objcg);
-               return NULL;
+               return (struct obj_cgroup *)-1UL;
        }
 
        return objcg;
@@ -501,9 +501,13 @@ static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
                return NULL;
 
        if (memcg_kmem_enabled() &&
-           ((flags & __GFP_ACCOUNT) || (s->flags & SLAB_ACCOUNT)))
+           ((flags & __GFP_ACCOUNT) || (s->flags & SLAB_ACCOUNT))) {
                *objcgp = memcg_slab_pre_alloc_hook(s, size, flags);
 
+               if (unlikely(*objcgp == (struct obj_cgroup *)-1UL))
+                       return NULL;
+       }
+
        return s;
 }
 
#4 lore raw

Hello Roman,

Sorry for having taken so much time to reply, I've only had the opportunity to deploy the patch on Tuesday morning for testing and now two days later the trashing occurred again.

Quoted 17 lines
mm/slab.h +1 -1
index 6cc323f1313a..ef02b841bcd8 100644
@@ -290,7 +290,7 @@ static inline struct obj_cgroup *memcg_slab_pre_alloc_hook(struct kmem_cache *s,
 
        if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s))) {
                obj_cgroup_put(objcg);
-               return NULL;
+               return (struct obj_cgroup *)-1UL;
        }
 
        return objcg;
@@ -501,9 +501,13 @@ static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
                return NULL;
 
        if (memcg_kmem_enabled() &&

mm/slab.h static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, · line 512

Quoted 10 lines
mm/slab.h +5 -1
-           ((flags & __GFP_ACCOUNT) || (s->flags & SLAB_ACCOUNT)))
+           ((flags & __GFP_ACCOUNT) || (s->flags & SLAB_ACCOUNT))) {
                *objcgp = memcg_slab_pre_alloc_hook(s, size, flags);
 
+               if (unlikely(*objcgp == (struct obj_cgroup *)-1UL))
+                       return NULL;
+       }
+
        return s;
 }

Seems your proposed patch didn't really help.

Compared to initial occurrence I do now have some more details (all but /proc/slabinfo since boot) and according to /proc/slabinfo a good deal of reclaimable slabs seem to be dentries (and probably xfs_inode/xfs_ifork related to them) - not sure if those are assigned to cgroups or not-accounted and not seen as candidate for reclaim...

xfs_buf           444908 445068    448   36    4 : tunables    0    0    0 : slabdata  12363  12363      0
xfs_bui_item           0      0    232   35    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_bud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_cui_item           0      0    456   35    4 : tunables    0    0    0 : slabdata      0      0      0
xfs_cud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_rui_item           0      0    712   46    8 : tunables    0    0    0 : slabdata      0      0      0
xfs_rud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_icr                0    156    208   39    2 : tunables    0    0    0 : slabdata      4      4      0
xfs_ili           1223169 1535904    224   36    2 : tunables    0    0    0 : slabdata  42664  42664      0
xfs_inode         12851565 22081140   1088   30    8 : tunables    0    0    0 : slabdata 736038 736038      0
xfs_efi_item           0    280    456   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_efd_item           0    280    464   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_buf_item           7    216    296   27    2 : tunables    0    0    0 : slabdata      8      8      0
xf_trans               0    224    288   28    2 : tunables    0    0    0 : slabdata      8      8      0
xfs_ifork         12834992 46309928     72   56    1 : tunables    0    0    0 : slabdata 826963 826963      0
xfs_da_state           0    224    512   32    4 : tunables    0    0    0 : slabdata      7      7      0
xfs_btree_cur          0    224    256   32    2 : tunables    0    0    0 : slabdata      7      7      0
xfs_bmap_free_item      0    230     88   46    1 : tunables    0    0    0 : slabdata      5      5      0
xfs_log_ticket         4    296    216   37    2 : tunables    0    0    0 : slabdata      8      8      0
fat_inode_cache        0      0    744   44    8 : tunables    0    0    0 : slabdata      0      0      0
fat_cache              0      0     64   64    1 : tunables    0    0    0 : slabdata      0      0      0
mnt_cache            114    180    448   36    4 : tunables    0    0    0 : slabdata      5      5      0
filp                6228  15582    384   42    4 : tunables    0    0    0 : slabdata    371    371      0
inode_cache         6669  16016    608   26    4 : tunables    0    0    0 : slabdata    616    616      0

dentry 8092159 15642504 224 36 2 : tunables 0 0 0 : slabdata 434514 434514 0

The full collected details are available at
https://faramir-fj.hosting-restena.lu/cgmon-20201203.txt
(please take a copy as that file will not stay there forever)

A visual graph of memory evolution is available at
https://faramir-fj.hosting-restena.lu/system-memory-20201203.png
with reboot on Tuesday morning and steady increase of slabs starting Webnesday evening correlating with start of backup until trashing started at about 3:30 and the large drop in memory being me doing
echo 2 > /proc/sys/vm/drop_caches
which stopped the trashing as well.

Against what does memcg attempt reclaim when it tries to satisfy a CG's low limit? Only against siblings or also against root or not-accounted? How does it take into account slabs where evictable entries will cause unevictable entries to be freed as well?

Quoted 7 lines
Quoted 7 lines

My setup, server has 64G of RAM:
root

   + system        { min=0, low=128M, high=8G, max=8G }
     + base        { no specific constraints }
     + backup      { min=0, low=32M, high=2G, max=2G }
     + shell       { no specific constraints }
  + websrv         { min=0, low=4G, high=32G, max=32G }
Quoted 8 lines
Quoted 8 lines
  + website        { min=0, low=16G, high=40T, max=40T }
    + website1     { min=0, low=64M, high=2G, max=2G }
    + website2     { min=0, low=64M, high=2G, max=2G }
      ...
  + remote         { min=0, low=1G, high=14G, max=14G }
    + webuser1     { min=0, low=64M, high=2G, max=2G }
    + webuser2     { min=0, low=64M, high=2G, max=2G }
      ...

Also interesting is that backup which is forced into 2G (system/backup CG) causes amount of slabs assigned to websrv CG to increase until that CG has almost only slab entries assigned to it to fill 16G, like file cache being reclaimed but not slab entries even if there is almost no file cache left and tons of slabs.
What I'm also surprised is the so much memory remains completely unused (instead of being used for file caches).

According to the documentation if I didn't get it wrong any limits of child CGs (e.g. webuser1...) are applied up to what their parent's limits allow. Thus, if looking at e.g. remote -> webuser1... even if I have 1000 webuserN they wont "reserve" 65G for themselves via memory.low limit when their parent sets memory.low to 1G? Or does this depend on on CG mount options (memory_recursiveprot)?

Regards,
Bruno

#5 lore raw
Quoted 27 lines · Bruno Prémont

Hello Roman,

Sorry for having taken so much time to reply, I've only had the opportunity to deploy the patch on Tuesday morning for testing and now two days later the trashing occurred again.

Quoted 21 lines
mm/slab.h +2 -2
index 6cc323f1313a..ef02b841bcd8 100644
@@ -290,7 +290,7 @@ static inline struct obj_cgroup *memcg_slab_pre_alloc_hook(struct kmem_cache *s,
 
        if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s))) {
                obj_cgroup_put(objcg);
-               return NULL;
+               return (struct obj_cgroup *)-1UL;
        }
 
        return objcg;
@@ -501,9 +501,13 @@ static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
                return NULL;
 
        if (memcg_kmem_enabled() &&
-           ((flags & __GFP_ACCOUNT) || (s->flags & SLAB_ACCOUNT)))
+           ((flags & __GFP_ACCOUNT) || (s->flags & SLAB_ACCOUNT))) {
                *objcgp = memcg_slab_pre_alloc_hook(s, size, flags);
 
Quoted 8 lines

mm/slab.h static inline struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, · line 512

Quoted 6 lines
mm/slab.h +4 -0
+               if (unlikely(*objcgp == (struct obj_cgroup *)-1UL))
+                       return NULL;
+       }
+
        return s;
 }

Seems your proposed patch didn't really help.

Anyway, thank you for testing! Actually your report helped me to reveal and fix this problem, so thank you!

In the meantime Yang Shi discovered a problem related slab shrinkers, which is to some extent similar to what you describe: under certain conditions large amounts of slab memory can be completely excluded from the reclaim process.

Can you, please, check if his fix will solve your problem?
Here is the final version: https://www.spinics.net/lists/stable/msg430601.html .

Quoted 50 lines

Compared to initial occurrence I do now have some more details (all but /proc/slabinfo since boot) and according to /proc/slabinfo a good deal of reclaimable slabs seem to be dentries (and probably xfs_inode/xfs_ifork related to them) - not sure if those are assigned to cgroups or not-accounted and not seen as candidate for reclaim...

xfs_buf           444908 445068    448   36    4 : tunables    0    0    0 : slabdata  12363  12363      0
xfs_bui_item           0      0    232   35    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_bud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_cui_item           0      0    456   35    4 : tunables    0    0    0 : slabdata      0      0      0
xfs_cud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_rui_item           0      0    712   46    8 : tunables    0    0    0 : slabdata      0      0      0
xfs_rud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_icr                0    156    208   39    2 : tunables    0    0    0 : slabdata      4      4      0
xfs_ili           1223169 1535904    224   36    2 : tunables    0    0    0 : slabdata  42664  42664      0
xfs_inode         12851565 22081140   1088   30    8 : tunables    0    0    0 : slabdata 736038 736038      0
xfs_efi_item           0    280    456   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_efd_item           0    280    464   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_buf_item           7    216    296   27    2 : tunables    0    0    0 : slabdata      8      8      0
xf_trans               0    224    288   28    2 : tunables    0    0    0 : slabdata      8      8      0
xfs_ifork         12834992 46309928     72   56    1 : tunables    0    0    0 : slabdata 826963 826963      0
xfs_da_state           0    224    512   32    4 : tunables    0    0    0 : slabdata      7      7      0
xfs_btree_cur          0    224    256   32    2 : tunables    0    0    0 : slabdata      7      7      0
xfs_bmap_free_item      0    230     88   46    1 : tunables    0    0    0 : slabdata      5      5      0
xfs_log_ticket         4    296    216   37    2 : tunables    0    0    0 : slabdata      8      8      0
fat_inode_cache        0      0    744   44    8 : tunables    0    0    0 : slabdata      0      0      0
fat_cache              0      0     64   64    1 : tunables    0    0    0 : slabdata      0      0      0
mnt_cache            114    180    448   36    4 : tunables    0    0    0 : slabdata      5      5      0
filp                6228  15582    384   42    4 : tunables    0    0    0 : slabdata    371    371      0
inode_cache         6669  16016    608   26    4 : tunables    0    0    0 : slabdata    616    616      0

dentry 8092159 15642504 224 36 2 : tunables 0 0 0 : slabdata 434514 434514 0

The full collected details are available at
https://faramir-fj.hosting-restena.lu/cgmon-20201203.txt
(please take a copy as that file will not stay there forever)

A visual graph of memory evolution is available at
https://faramir-fj.hosting-restena.lu/system-memory-20201203.png
with reboot on Tuesday morning and steady increase of slabs starting Webnesday evening correlating with start of backup until trashing started at about 3:30 and the large drop in memory being me doing
echo 2 > /proc/sys/vm/drop_caches
which stopped the trashing as well.

Quoted 4 lines

Against what does memcg attempt reclaim when it tries to satisfy a CG's low limit? Only against siblings or also against root or not-accounted? How does it take into account slabs where evictable entries will cause unevictable entries to be freed as well?

Low limits are working by excluding some portions of memory from the reclaim, not by adding a memory pressure to something else.

Quoted 25 lines
Quoted 16 lines
Quoted 16 lines

My setup, server has 64G of RAM:
root

   + system        { min=0, low=128M, high=8G, max=8G }
     + base        { no specific constraints }
     + backup      { min=0, low=32M, high=2G, max=2G }
     + shell       { no specific constraints }
  + websrv         { min=0, low=4G, high=32G, max=32G }
  + website        { min=0, low=16G, high=40T, max=40T }
    + website1     { min=0, low=64M, high=2G, max=2G }
    + website2     { min=0, low=64M, high=2G, max=2G }
      ...
  + remote         { min=0, low=1G, high=14G, max=14G }
    + webuser1     { min=0, low=64M, high=2G, max=2G }
    + webuser2     { min=0, low=64M, high=2G, max=2G }
      ...

Also interesting is that backup which is forced into 2G (system/backup CG) causes amount of slabs assigned to websrv CG to increase until that CG has almost only slab entries assigned to it to fill 16G, like file cache being reclaimed but not slab entries even if there is almost no file cache left and tons of slabs.
What I'm also surprised is the so much memory remains completely unused (instead of being used for file caches).

Quoted 6 lines

According to the documentation if I didn't get it wrong any limits of child CGs (e.g. webuser1...) are applied up to what their parent's limits allow. Thus, if looking at e.g. remote -> webuser1... even if I have 1000 webuserN they wont "reserve" 65G for themselves via memory.low limit when their parent sets memory.low to 1G? Or does this depend on on CG mount options (memory_recursiveprot)?

It does. What you're describing is the old (!memory_recursiveprot) behavior.

Thanks!

#6 lore raw
Quoted 6 lines · On Thu, 3 Dec 2020 12:55:59 -0800 Roman Gushchin…

In the meantime Yang Shi discovered a problem related slab shrinkers, which is to some extent similar to what you describe: under certain conditions large amounts of slab memory can be completely excluded from the reclaim process.

Can you, please, check if his fix will solve your problem?
Here is the final version: https://www.spinics.net/lists/stable/msg430601.html .

I've added that patch on top of yours but it seems not to completely help either.
With this patch is seems that such dentries might get reclaimed as a last resort instead of not at all.

I've added logs since current boot:
https://faramir-fj.hosting-restena.lu/cgmon-20201204.txt
https://faramir-fj.hosting-restena.lu/cgmon-20201205.txt
https://faramir-fj.hosting-restena.lu/cgmon-20201206.txt
with the memory evolution. Evolution started to degrade over past night where memory usage started to increase from 40G tending to full use but with only slabs growing (not file cache) and memory assigned to cgroups staying more or less constant - even root cgroup's memory stats seem not to list a great deal of used memory.

The only cgroup not "sufficiently" protected by memory.low (websrv) has seen its memory use somehow clamped to about 16G while it should be allowed to go up to 32G according to memory.high and of those 16G in use at time of writing it only had 100M of file cache left, all the rest being slabs.

As system now is using most of its memory I've bumped websrv CG's memory.low to 20G so it should stay protected some more (which after a few minutes showed its filecache growing again) with the aim of moving pressure out of leaf-cgroups to non-cg-assigned-memory.

Somehow this move seems to prove getting me some success.

I will report back later today or tomorrow with more details on the evolution with "no unused" memory. At least production service tends not to suffer (more than from storage response time).

I have the impression that memory reclaim now only looks at cgroup and if it can make some progress it will not bother looking anywhere else.

I also have the vague impression that distribution of my tasks on the two NUMA nodes somehow impacts when or how memory reclaim happens.

NUMA node0 CPU(s):     0-5,12-17
NUMA node1 CPU(s):     6-11,18-23

CPUs MEMs
system: 0-3 0
websrv: 8-11 0 (allowed mems=0-1 as of 2020-12-06 12:18
after increasing memory.low from 4G to 20G which did release quite some pression and allow)
website: 12-23 1
remote: 4-7 0
(assignment done using cpuset cgroup)

(seems NUMA distribution changed or I missed the non-linear node distribution of CPUs - cores versus hyperthreading as my Intent was to have website on 1 socket and the rest on the other socket. Memory is mapped as I planned it, but tasks not really)

So calculating:

  system: 128M..8G     mems=0    \
  remote: 1G..14G      mems=0     }-->   5G..54G
  websrv: 4G..32G      mems=0    /
  website: 16G..       mems=1

But with everything except websrv lying below its low limit I wonder why reclaim only hits the cgroups's file cache but still mostly ignores its slabs.

Bruno

Quoted 90 lines
Quoted 47 lines

Compared to initial occurrence I do now have some more details (all but /proc/slabinfo since boot) and according to /proc/slabinfo a good deal of reclaimable slabs seem to be dentries (and probably xfs_inode/xfs_ifork related to them) - not sure if those are assigned to cgroups or not-accounted and not seen as candidate for reclaim...

xfs_buf           444908 445068    448   36    4 : tunables    0    0    0 : slabdata  12363  12363      0
xfs_bui_item           0      0    232   35    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_bud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_cui_item           0      0    456   35    4 : tunables    0    0    0 : slabdata      0      0      0
xfs_cud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_rui_item           0      0    712   46    8 : tunables    0    0    0 : slabdata      0      0      0
xfs_rud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_icr                0    156    208   39    2 : tunables    0    0    0 : slabdata      4      4      0
xfs_ili           1223169 1535904    224   36    2 : tunables    0    0    0 : slabdata  42664  42664      0
xfs_inode         12851565 22081140   1088   30    8 : tunables    0    0    0 : slabdata 736038 736038      0
xfs_efi_item           0    280    456   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_efd_item           0    280    464   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_buf_item           7    216    296   27    2 : tunables    0    0    0 : slabdata      8      8      0
xf_trans               0    224    288   28    2 : tunables    0    0    0 : slabdata      8      8      0
xfs_ifork         12834992 46309928     72   56    1 : tunables    0    0    0 : slabdata 826963 826963      0
xfs_da_state           0    224    512   32    4 : tunables    0    0    0 : slabdata      7      7      0
xfs_btree_cur          0    224    256   32    2 : tunables    0    0    0 : slabdata      7      7      0
xfs_bmap_free_item      0    230     88   46    1 : tunables    0    0    0 : slabdata      5      5      0
xfs_log_ticket         4    296    216   37    2 : tunables    0    0    0 : slabdata      8      8      0
fat_inode_cache        0      0    744   44    8 : tunables    0    0    0 : slabdata      0      0      0
fat_cache              0      0     64   64    1 : tunables    0    0    0 : slabdata      0      0      0
mnt_cache            114    180    448   36    4 : tunables    0    0    0 : slabdata      5      5      0
filp                6228  15582    384   42    4 : tunables    0    0    0 : slabdata    371    371      0
inode_cache         6669  16016    608   26    4 : tunables    0    0    0 : slabdata    616    616      0

dentry 8092159 15642504 224 36 2 : tunables 0 0 0 : slabdata 434514 434514 0

The full collected details are available at
https://faramir-fj.hosting-restena.lu/cgmon-20201203.txt
(please take a copy as that file will not stay there forever)

A visual graph of memory evolution is available at
https://faramir-fj.hosting-restena.lu/system-memory-20201203.png
with reboot on Tuesday morning and steady increase of slabs starting Webnesday evening correlating with start of backup until trashing started at about 3:30 and the large drop in memory being me doing
echo 2 > /proc/sys/vm/drop_caches
which stopped the trashing as well.

Quoted 4 lines

Against what does memcg attempt reclaim when it tries to satisfy a CG's low limit? Only against siblings or also against root or not-accounted? How does it take into account slabs where evictable entries will cause unevictable entries to be freed as well?

Low limits are working by excluding some portions of memory from the reclaim, not by adding a memory pressure to something else.

Quoted 25 lines
Quoted 16 lines
Quoted 16 lines

My setup, server has 64G of RAM:
root

   + system        { min=0, low=128M, high=8G, max=8G }
     + base        { no specific constraints }
     + backup      { min=0, low=32M, high=2G, max=2G }
     + shell       { no specific constraints }
  + websrv         { min=0, low=4G, high=32G, max=32G }
  + website        { min=0, low=16G, high=40T, max=40T }
    + website1     { min=0, low=64M, high=2G, max=2G }
    + website2     { min=0, low=64M, high=2G, max=2G }
      ...
  + remote         { min=0, low=1G, high=14G, max=14G }
    + webuser1     { min=0, low=64M, high=2G, max=2G }
    + webuser2     { min=0, low=64M, high=2G, max=2G }
      ...

Also interesting is that backup which is forced into 2G (system/backup CG) causes amount of slabs assigned to websrv CG to increase until that CG has almost only slab entries assigned to it to fill 16G, like file cache being reclaimed but not slab entries even if there is almost no file cache left and tons of slabs.
What I'm also surprised is the so much memory remains completely unused (instead of being used for file caches).

Quoted 6 lines

According to the documentation if I didn't get it wrong any limits of child CGs (e.g. webuser1...) are applied up to what their parent's limits allow. Thus, if looking at e.g. remote -> webuser1... even if I have 1000 webuserN they wont "reserve" 65G for themselves via memory.low limit when their parent sets memory.low to 1G? Or does this depend on on CG mount options (memory_recursiveprot)?

It does. What you're describing is the old (!memory_recursiveprot) behavior.

Thanks!

#7 lore raw

Hello All,

Since my last changes (allowing the websrv CG to use both NUMA memory areas) the system as a whole runs reasonably without trashing and also makes a way better use of memory.

As such it really seems that NUMA node restrictions (memory wise at least) do not properly interact with reclaim.

Is there a way to see how memory usage is on the different NUMA nodes?

I have the impressions some reclaim was ongoing because memory allocation was ask on one node which may have been "full" and there only file cache was being reclaimed (in cgroups where memory.low didn't protect it).

Thanks,
Bruno

Quoted 173 lines · On Sun, 6 Dec 2020 12:30:21 +0100 Bruno Prémont wrote:
Quoted 6 lines · On Thu, 3 Dec 2020 12:55:59 -0800 Roman Gushchin…

In the meantime Yang Shi discovered a problem related slab shrinkers, which is to some extent similar to what you describe: under certain conditions large amounts of slab memory can be completely excluded from the reclaim process.

Can you, please, check if his fix will solve your problem?
Here is the final version: https://www.spinics.net/lists/stable/msg430601.html .

I've added that patch on top of yours but it seems not to completely help either.
With this patch is seems that such dentries might get reclaimed as a last resort instead of not at all.

I've added logs since current boot:
https://faramir-fj.hosting-restena.lu/cgmon-20201204.txt
https://faramir-fj.hosting-restena.lu/cgmon-20201205.txt
https://faramir-fj.hosting-restena.lu/cgmon-20201206.txt
with the memory evolution. Evolution started to degrade over past night where memory usage started to increase from 40G tending to full use but with only slabs growing (not file cache) and memory assigned to cgroups staying more or less constant - even root cgroup's memory stats seem not to list a great deal of used memory.

The only cgroup not "sufficiently" protected by memory.low (websrv) has seen its memory use somehow clamped to about 16G while it should be allowed to go up to 32G according to memory.high and of those 16G in use at time of writing it only had 100M of file cache left, all the rest being slabs.

As system now is using most of its memory I've bumped websrv CG's memory.low to 20G so it should stay protected some more (which after a few minutes showed its filecache growing again) with the aim of moving pressure out of leaf-cgroups to non-cg-assigned-memory.

Somehow this move seems to prove getting me some success.

I will report back later today or tomorrow with more details on the evolution with "no unused" memory. At least production service tends not to suffer (more than from storage response time).

I have the impression that memory reclaim now only looks at cgroup and if it can make some progress it will not bother looking anywhere else.

I also have the vague impression that distribution of my tasks on the two NUMA nodes somehow impacts when or how memory reclaim happens.

NUMA node0 CPU(s):     0-5,12-17
NUMA node1 CPU(s):     6-11,18-23

CPUs MEMs
system: 0-3 0
websrv: 8-11 0 (allowed mems=0-1 as of 2020-12-06 12:18
after increasing memory.low from 4G to 20G which did release quite some pression and allow)
website: 12-23 1
remote: 4-7 0
(assignment done using cpuset cgroup)

(seems NUMA distribution changed or I missed the non-linear node distribution of CPUs - cores versus hyperthreading as my Intent was to have website on 1 socket and the rest on the other socket. Memory is mapped as I planned it, but tasks not really)

So calculating:

  system: 128M..8G     mems=0    \
  remote: 1G..14G      mems=0     }-->   5G..54G
  websrv: 4G..32G      mems=0    /
  website: 16G..       mems=1

But with everything except websrv lying below its low limit I wonder why reclaim only hits the cgroups's file cache but still mostly ignores its slabs.

Bruno

Quoted 90 lines
Quoted 47 lines

Compared to initial occurrence I do now have some more details (all but /proc/slabinfo since boot) and according to /proc/slabinfo a good deal of reclaimable slabs seem to be dentries (and probably xfs_inode/xfs_ifork related to them) - not sure if those are assigned to cgroups or not-accounted and not seen as candidate for reclaim...

xfs_buf           444908 445068    448   36    4 : tunables    0    0    0 : slabdata  12363  12363      0
xfs_bui_item           0      0    232   35    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_bud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_cui_item           0      0    456   35    4 : tunables    0    0    0 : slabdata      0      0      0
xfs_cud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_rui_item           0      0    712   46    8 : tunables    0    0    0 : slabdata      0      0      0
xfs_rud_item           0      0    200   40    2 : tunables    0    0    0 : slabdata      0      0      0
xfs_icr                0    156    208   39    2 : tunables    0    0    0 : slabdata      4      4      0
xfs_ili           1223169 1535904    224   36    2 : tunables    0    0    0 : slabdata  42664  42664      0
xfs_inode         12851565 22081140   1088   30    8 : tunables    0    0    0 : slabdata 736038 736038      0
xfs_efi_item           0    280    456   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_efd_item           0    280    464   35    4 : tunables    0    0    0 : slabdata      8      8      0
xfs_buf_item           7    216    296   27    2 : tunables    0    0    0 : slabdata      8      8      0
xf_trans               0    224    288   28    2 : tunables    0    0    0 : slabdata      8      8      0
xfs_ifork         12834992 46309928     72   56    1 : tunables    0    0    0 : slabdata 826963 826963      0
xfs_da_state           0    224    512   32    4 : tunables    0    0    0 : slabdata      7      7      0
xfs_btree_cur          0    224    256   32    2 : tunables    0    0    0 : slabdata      7      7      0
xfs_bmap_free_item      0    230     88   46    1 : tunables    0    0    0 : slabdata      5      5      0
xfs_log_ticket         4    296    216   37    2 : tunables    0    0    0 : slabdata      8      8      0
fat_inode_cache        0      0    744   44    8 : tunables    0    0    0 : slabdata      0      0      0
fat_cache              0      0     64   64    1 : tunables    0    0    0 : slabdata      0      0      0
mnt_cache            114    180    448   36    4 : tunables    0    0    0 : slabdata      5      5      0
filp                6228  15582    384   42    4 : tunables    0    0    0 : slabdata    371    371      0
inode_cache         6669  16016    608   26    4 : tunables    0    0    0 : slabdata    616    616      0

dentry 8092159 15642504 224 36 2 : tunables 0 0 0 : slabdata 434514 434514 0

The full collected details are available at
https://faramir-fj.hosting-restena.lu/cgmon-20201203.txt
(please take a copy as that file will not stay there forever)

A visual graph of memory evolution is available at
https://faramir-fj.hosting-restena.lu/system-memory-20201203.png
with reboot on Tuesday morning and steady increase of slabs starting Webnesday evening correlating with start of backup until trashing started at about 3:30 and the large drop in memory being me doing
echo 2 > /proc/sys/vm/drop_caches
which stopped the trashing as well.

Quoted 4 lines

Against what does memcg attempt reclaim when it tries to satisfy a CG's low limit? Only against siblings or also against root or not-accounted? How does it take into account slabs where evictable entries will cause unevictable entries to be freed as well?

Low limits are working by excluding some portions of memory from the reclaim, not by adding a memory pressure to something else.

Quoted 25 lines
Quoted 16 lines
Quoted 16 lines

My setup, server has 64G of RAM:
root

   + system        { min=0, low=128M, high=8G, max=8G }
     + base        { no specific constraints }
     + backup      { min=0, low=32M, high=2G, max=2G }
     + shell       { no specific constraints }
  + websrv         { min=0, low=4G, high=32G, max=32G }
  + website        { min=0, low=16G, high=40T, max=40T }
    + website1     { min=0, low=64M, high=2G, max=2G }
    + website2     { min=0, low=64M, high=2G, max=2G }
      ...
  + remote         { min=0, low=1G, high=14G, max=14G }
    + webuser1     { min=0, low=64M, high=2G, max=2G }
    + webuser2     { min=0, low=64M, high=2G, max=2G }
      ...  

Also interesting is that backup which is forced into 2G (system/backup CG) causes amount of slabs assigned to websrv CG to increase until that CG has almost only slab entries assigned to it to fill 16G, like file cache being reclaimed but not slab entries even if there is almost no file cache left and tons of slabs.
What I'm also surprised is the so much memory remains completely unused (instead of being used for file caches).

Quoted 6 lines

According to the documentation if I didn't get it wrong any limits of child CGs (e.g. webuser1...) are applied up to what their parent's limits allow. Thus, if looking at e.g. remote -> webuser1... even if I have 1000 webuserN they wont "reserve" 65G for themselves via memory.low limit when their parent sets memory.low to 1G? Or does this depend on on CG mount options (memory_recursiveprot)?

It does. What you're describing is the old (!memory_recursiveprot) behavior.

Thanks!