mm, slab: Extend vm/drop_caches to shrink kmem slabs
Subsystems (7)
cover mm, slab: Extend vm/drop_caches to shrink kmem slabs
The purpose of this patchset is to allow system administrators to have the ability to shrink all the kmem slabs in order to free up memory and get a more accurate picture of how many slab objects are actually being used.
Patch 1 adds a new memcg_iterate_all() that is used by the patch 2 to iterate on all the memory cgroups.
Waiman Long (2):
mm, memcontrol: Add memcg_iterate_all()
mm, slab: Extend vm/drop_caches to shrink kmem slabs
Documentation/sysctl/vm.txt | 11 ++++++++-- fs/drop_caches.c | 4 ++++ include/linux/memcontrol.h | 3 +++ include/linux/slab.h | 1 + kernel/sysctl.c | 4 ++-- mm/memcontrol.c | 13 +++++++++++ mm/slab_common.c | 44 +++++++++++++++++++++++++++++++++++++ 7 files changed, 76 insertions(+), 4 deletions(-)
Signature · 1 line
2.18.1
PATCH 1/2 mm, memcontrol: Add memcg_iterate_all()
Add a memcg_iterate_all() function for iterating all the available memory cgroups and call the given callback function for each of the memory cgruops.
--- include/linux/memcontrol.h | 3 +++ mm/memcontrol.c | 13 +++++++++++++ 2 files changed, 16 insertions(+)
include/linux/memcontrol.h +3 -0
@@ -1268,6 +1268,9 @@ static inline bool mem_cgroup_under_socket_pressure(struct mem_cgroup *memcg)struct kmem_cache *memcg_kmem_get_cache (struct kmem_cache *cachep);void memcg_kmem_put_cache (struct kmem_cache *cachep); +extern void memcg_iterate_all (void (*callback)(struct mem_cgroup *memcg, +void *arg),void *arg); +#ifdef CONFIG_MEMCG_KMEM int __memcg_kmem_charge (struct page *page,gfp_t gfp,int order);void __memcg_kmem_uncharge (struct page *page,int order);
mm/memcontrol.c +13 -0
@@ -443,6 +443,19 @@ static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)static void memcg_free_shrinker_maps (struct mem_cgroup *memcg) { }#endif /* CONFIG_MEMCG_KMEM */ +/* +* Iterate all the memory cgroups and call the given callback function +* for each of the memory cgroups. +*/ +void memcg_iterate_all (void (*callback)(struct mem_cgroup *memcg,void *arg), +void *arg) +{ +struct mem_cgroup *memcg; + +for_each_mem_cgroup (memcg) +callback (memcg, arg); +} +/** * mem_cgroup_css_from_page - css of the memcg associated with a page * @page: page of interest
Signature · 1 line
2.18.1
Michal Hocko
maintainerRe: [PATCH 1/2] mm, memcontrol: Add memcg_iterate_all()
Quoted 3 lines · Waiman Long
Add a
memcg_iterate_all()function for iterating all the available memory cgroups and call the given callback function for each of the memory cgruops.
Why is a trivial wrapper any better than open coded usage of the iterator?
Quoted 46 lines
Signed-off-by Waiman Long <[email protected]>--- include/linux/memcontrol.h | 3 +++ mm/memcontrol.c | 13 +++++++++++++ 2 files changed, 16 insertions(+)include/linux/memcontrol.h +3
-0@@ -1268,6 +1268,9 @@ static inline bool mem_cgroup_under_socket_pressure(struct mem_cgroup *memcg)struct kmem_cache *memcg_kmem_get_cache (struct kmem_cache *cachep);void memcg_kmem_put_cache (struct kmem_cache *cachep); +extern void memcg_iterate_all (void (*callback)(struct mem_cgroup *memcg, +void *arg),void *arg); +#ifdef CONFIG_MEMCG_KMEM int __memcg_kmem_charge (struct page *page,gfp_t gfp,int order);void __memcg_kmem_uncharge (struct page *page,int order);mm/memcontrol.c +13
-0@@ -443,6 +443,19 @@ static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)static void memcg_free_shrinker_maps (struct mem_cgroup *memcg) { }#endif /* CONFIG_MEMCG_KMEM */ +/* +* Iterate all the memory cgroups and call the given callback function +* for each of the memory cgroups. +*/ +void memcg_iterate_all (void (*callback)(struct mem_cgroup *memcg,void *arg), +void *arg) +{ +struct mem_cgroup *memcg; + +for_each_mem_cgroup (memcg) +callback (memcg, arg); +} +/** * mem_cgroup_css_from_page - css of the memcg associated with a page * @page: page of interest Signature · 1 line
2.18.1
Signature · 2 lines
Michal Hocko SUSE Labs
Re: [PATCH 1/2] mm, memcontrol: Add memcg_iterate_all()
Quoted 6 lines · Michal Hocko
Quoted 3 lines · Waiman Long
Add a
memcg_iterate_all()function for iterating all the available memory cgroups and call the given callback function for each of the memory cgruops.Why is a trivial wrapper any better than open coded usage of the iterator?
Because the iterator is only defined within memcontrol.c. So an alternative may be to put the iterator into a header file that can be used by others. Will take a look at that.
Cheers,
Longman
Michal Hocko
maintainerRe: [PATCH 1/2] mm, memcontrol: Add memcg_iterate_all()
Quoted 11 lines · Waiman Long
Quoted 6 lines · Michal Hocko
Quoted 3 lines · Waiman Long
Add a
memcg_iterate_all()function for iterating all the available memory cgroups and call the given callback function for each of the memory cgruops.Why is a trivial wrapper any better than open coded usage of the iterator?
Because the iterator is only defined within
memcontrol.c. So an alternative may be to put the iterator into a header file that can be used by others. Will take a look at that.
That would be preferred.
Thanks!
Signature · 2 lines
Michal Hocko SUSE Labs
PATCH 2/2 mm, slab: Extend vm/drop_caches to shrink kmem slabs
In reply to Waiman LongWith the slub memory allocator, the numbers of active slab objects reported in /proc/slabinfo are not real because they include objects that are held by the per-cpu slab structures whether they are actually used or not. The problem gets worse the more CPUs a system have. For instance, looking at the reported number of active task_struct objects, one will wonder where all the missing tasks gone.
I know it is hard and costly to get a real count of active objects. So I am not advocating for that. Instead, this patch extends the /proc/sys/vm/drop_caches sysctl parameter by using a new bit (bit 3) to shrink all the kmem slabs which will flush out all the slabs in the per-cpu structures and give a more accurate view of how much memory are really used up by the active slab objects. This is a costly operation, of course, but it gives a way to have a clearer picture of the actual number of slab objects used, if the need arises.
The upper range of the drop_caches sysctl parameter is increased to 15 to allow all possible combinations of the lowest 4 bits.
On a 2-socket 64-core 256-thread ARM64 system with 64k page size after a parallel kernel build, the amount of memory occupied by slabs before and after echoing to drop_caches were:
# grep task_struct /proc/slabinfo task_struct 48376 48434 4288 61 4 : tunables 0 0 0 : slabdata 794 794 0 # grep "^S[lRU]" /proc/meminfo Slab: 3419072 kB SReclaimable: 354688 kB SUnreclaim: 3064384 kB # echo 3 > /proc/sys/vm/drop_caches # grep "^S[lRU]" /proc/meminfo Slab: 3351680 kB SReclaimable: 316096 kB SUnreclaim: 3035584 kB # echo 8 > /proc/sys/vm/drop_caches # grep "^S[lRU]" /proc/meminfo Slab: 1008192 kB SReclaimable: 126912 kB SUnreclaim: 881280 kB # grep task_struct /proc/slabinfo task_struct 2601 6588 4288 61 4 : tunables 0 0 0 : slabdata 108 108 0
Shrinking the slabs saves more than 2GB of memory in this case. This new feature certainly fulfills the promise of dropping caches.
Unlike counting objects in the per-node caches done by /proc/slabinfo which is rather light weight, iterating all the per-cpu caches and shrinking them is much more heavy weight.
For this particular instance, the time taken to shrinks all the root caches was about 30.2ms. There were 73 memory cgroup and the longest time taken for shrinking the largest one was about 16.4ms. The total shrinking time was about 101ms.
Because of the potential long time to shrinks all the caches, the slab_mutex was taken multiple times - once for all the root caches and once for each memory cgroup. This is to reduce the slab_mutex hold time to minimize impact to other running applications that may need to acquire the mutex.
The slab shrinking feature is only available when CONFIG_MEMCG_KMEM is defined as the code need to access slab_root_caches to iterate all the root caches.
--- Documentation/sysctl/vm.txt | 11 ++++++++-- fs/drop_caches.c | 4 ++++ include/linux/slab.h | 1 + kernel/sysctl.c | 4 ++-- mm/slab_common.c | 44 +++++++++++++++++++++++++++++++++++++ 5 files changed, 60 insertions(+), 4 deletions(-)
Documentation/sysctl/vm.txt +9 -2
@@ -207,8 +207,8 @@ Setting this to zero disables periodic writeback altogether. drop_caches Writing to this will cause the kernel to drop clean caches, as well as -reclaimable slab objects like dentries and inodes. Once dropped, their -memory becomes free. +reclaimable slab objects like dentries and inodes. It can also be used +to shrink the slabs. Once dropped, their memory becomes free. To free pagecache: echo 1 > /proc/sys/vm/drop_caches
@@ -216,6 +216,8 @@ To free reclaimable slab objects (includes dentries and inodes): echo 2 > /proc/sys/vm/drop_caches To free slab objects and pagecache: echo 3 > /proc/sys/vm/drop_caches +To shrink the slabs: + echo 8 > /proc/sys/vm/drop_caches This is a non-destructive operation and will not free any dirty objects. To increase the number of objects freed by this operation, the user may run
@@ -223,6 +225,11 @@ To increase the number of objects freed by this operation, the user may run number of dirty objects on the system and create more candidates to be dropped. +Shrinking the slabs can reduce the memory footprint used by the slabs. +It also makes the number of active objects reported in /proc/slabinfo +more representative of the actual number of objects used for the slub +memory allocator. + This file is not a means to control the growth of the various kernel caches (inodes, dentries, pagecache, etc...) These objects are automatically reclaimed by the kernel when memory is needed elsewhere on the system.
fs/drop_caches.c +4 -0
@@ -9,6 +9,7 @@#include <linux/writeback.h> #include <linux/sysctl.h> #include <linux/gfp.h> +#include <linux/slab.h> #include "internal.h" /* A global variable is a bit ugly, but it keeps the code simple */
@@ -65,6 +66,9 @@ int drop_caches_sysctl_handler(struct ctl_table *table, int write,drop_slab ();count_vm_event (DROP_SLAB ); } +if (sysctl_drop_caches &8 ) { +kmem_cache_shrink_all (); + }if (!stfu) { pr_info("%s (%d): drop_caches: %d\n", current->comm, task_pid_nr(current),
include/linux/slab.h +1 -0
@@ -149,6 +149,7 @@ struct kmem_cache *kmem_cache_create_usercopy(const char *name, void (*ctor)(void *));void kmem_cache_destroy (struct kmem_cache *);int kmem_cache_shrink (struct kmem_cache *); +void kmem_cache_shrink_all (void );void memcg_create_kmem_cache (struct mem_cgroup *,struct kmem_cache *);void memcg_deactivate_kmem_caches (struct mem_cgroup *);
kernel/sysctl.c +2 -2
@@ -129,7 +129,7 @@ static int __maybe_unused neg_one = -1;static int zero;static int __maybe_unused one =1 ;static int __maybe_unused two =2 ; -static int __maybe_unused four =4 ; +static int __maybe_unused fifteen =15 ;static unsigned long zero_ul;static unsigned long one_ul =1 ;static unsigned long long_max =LONG_MAX ;
@@ -1455,7 +1455,7 @@ static struct ctl_table vm_table[] = { .mode = 0644, .proc_handler = drop_caches_sysctl_handler, .extra1 = &one, - .extra2 = &four, + .extra2 = &fifteen, },#ifdef CONFIG_COMPACTION {
mm/slab_common.c +44 -0
50 lines not on this page · raw message
Signature · 1 line
2.18.1
Quoted 172 lines · Waiman Long
With the slub memory allocator, the numbers of active slab objects reported in
/proc/slabinfoare not real because they include objects that are held by the per-cpu slab structures whether they are actually used or not. The problem gets worse the more CPUs a system have. For instance, looking at the reported number of activetask_structobjects, one will wonder where all the missing tasks gone.I know it is hard and costly to get a real count of active objects. So I am not advocating for that. Instead, this patch extends the
/proc/sys/vm/drop_cachessysctl parameter by using a new bit (bit 3) to shrink all the kmem slabs which will flush out all the slabs in the per-cpu structures and give a more accurate view of how much memory are really used up by the active slab objects. This is a costly operation, of course, but it gives a way to have a clearer picture of the actual number of slab objects used, if the need arises.The upper range of the
drop_cachessysctl parameter is increased to 15 to allow all possible combinations of the lowest 4 bits.On a 2-socket 64-core 256-thread ARM64 system with 64k page size after a parallel kernel build, the amount of memory occupied by slabs before and after echoing to
drop_cacheswere:# grep task_struct /proc/slabinfo task_struct 48376 48434 4288 61 4 : tunables 0 0 0 : slabdata 794 794 0 # grep "^S[lRU]" /proc/meminfo Slab: 3419072 kB SReclaimable: 354688 kB SUnreclaim: 3064384 kB # echo 3 > /proc/sys/vm/drop_caches # grep "^S[lRU]" /proc/meminfo Slab: 3351680 kB SReclaimable: 316096 kB SUnreclaim: 3035584 kB # echo 8 > /proc/sys/vm/drop_caches # grep "^S[lRU]" /proc/meminfo Slab: 1008192 kB SReclaimable: 126912 kB SUnreclaim: 881280 kB # grep task_struct /proc/slabinfo task_struct 2601 6588 4288 61 4 : tunables 0 0 0 : slabdata 108 108 0Shrinking the slabs saves more than 2GB of memory in this case. This new feature certainly fulfills the promise of dropping caches.
Unlike counting objects in the per-node caches done by
/proc/slabinfowhich is rather light weight, iterating all the per-cpu caches and shrinking them is much more heavy weight.For this particular instance, the time taken to shrinks all the root caches was about 30.2ms. There were 73 memory cgroup and the longest time taken for shrinking the largest one was about 16.4ms. The total shrinking time was about 101ms.
Because of the potential long time to shrinks all the caches, the
slab_mutexwas taken multiple times - once for all the root caches and once for each memory cgroup. This is to reduce theslab_mutexhold time to minimize impact to other running applications that may need to acquire the mutex.The slab shrinking feature is only available when
CONFIG_MEMCG_KMEMis defined as the code need to accessslab_root_cachesto iterate all the root caches.Signed-off-by Waiman Long <[email protected]>--- Documentation/sysctl/vm.txt | 11 ++++++++-- fs/drop_caches.c | 4 ++++ include/linux/slab.h | 1 + kernel/sysctl.c | 4 ++-- mm/slab_common.c | 44 +++++++++++++++++++++++++++++++++++++ 5 files changed, 60 insertions(+), 4 deletions(-)Documentation/sysctl/vm.txt +9
-2@@ -207,8 +207,8 @@ Setting this to zero disables periodic writeback altogether. drop_caches Writing to this will cause the kernel to drop clean caches, as well as -reclaimable slab objects like dentries and inodes. Once dropped, their -memory becomes free. +reclaimable slab objects like dentries and inodes. It can also be used +to shrink the slabs. Once dropped, their memory becomes free. To free pagecache: echo 1 > /proc/sys/vm/drop_caches@@ -216,6 +216,8 @@ To free reclaimable slab objects (includes dentries and inodes): echo 2 > /proc/sys/vm/drop_caches To free slab objects and pagecache: echo 3 > /proc/sys/vm/drop_caches +To shrink the slabs: + echo 8 > /proc/sys/vm/drop_caches This is a non-destructive operation and will not free any dirty objects. To increase the number of objects freed by this operation, the user may run@@ -223,6 +225,11 @@ To increase the number of objects freed by this operation, the user may run number of dirty objects on the system and create more candidates to be dropped. +Shrinking the slabs can reduce the memory footprint used by the slabs. +It also makes the number of active objects reported in /proc/slabinfo +more representative of the actual number of objects used for the slub +memory allocator. + This file is not a means to control the growth of the various kernel caches (inodes, dentries, pagecache, etc...) These objects are automatically reclaimed by the kernel when memory is needed elsewhere on the system.fs/drop_caches.c +4
-0@@ -9,6 +9,7 @@#include <linux/writeback.h> #include <linux/sysctl.h> #include <linux/gfp.h> +#include <linux/slab.h> #include "internal.h" /* A global variable is a bit ugly, but it keeps the code simple */ @@ -65,6 +66,9 @@ int drop_caches_sysctl_handler(struct ctl_table *table, int write,drop_slab ();count_vm_event (DROP_SLAB ); } +if (sysctl_drop_caches &8 ) { +kmem_cache_shrink_all (); + }if (!stfu) { pr_info("%s (%d): drop_caches: %d\n", current->comm, task_pid_nr(current),include/linux/slab.h +1
-0@@ -149,6 +149,7 @@ struct kmem_cache *kmem_cache_create_usercopy(const char *name, void (*ctor)(void *));void kmem_cache_destroy (struct kmem_cache *);int kmem_cache_shrink (struct kmem_cache *); +void kmem_cache_shrink_all (void );void memcg_create_kmem_cache (struct mem_cgroup *,struct kmem_cache *);void memcg_deactivate_kmem_caches (struct mem_cgroup *);kernel/sysctl.c +2
-2@@ -129,7 +129,7 @@ static int __maybe_unused neg_one = -1;static int zero;static int __maybe_unused one =1 ;static int __maybe_unused two =2 ; -static int __maybe_unused four =4 ; +static int __maybe_unused fifteen =15 ;static unsigned long zero_ul;static unsigned long one_ul =1 ;static unsigned long long_max =LONG_MAX ;@@ -1455,7 +1455,7 @@ static struct ctl_table vm_table[] = { .mode = 0644, .proc_handler = drop_caches_sysctl_handler, .extra1 = &one, - .extra2 = &four, + .extra2 = &fifteen, },#ifdef CONFIG_COMPACTION {mm/slab_common.c +0
-0
mm/slab_common.c int kmem_cache_shrink(struct kmem_cache *cachep) · line 957
Quoted 2 lines
mm/slab_common.c +0
-0@@ -956,6 +956,50 @@ int kmem_cache_shrink(struct kmem_cache *cachep) }EXPORT_SYMBOL (kmem_cache_shrink);
Hi Waiman!
Quoted 7 lines
mm/slab_common.c +6
-0+#ifdef CONFIG_MEMCG_KMEM +static void kmem_cache_shrink_memcg (struct mem_cgroup *memcg, +void __maybe_unused *arg) +{ +struct kmem_cache *s; +
mm/slab_common.c int kmem_cache_shrink(struct kmem_cache *cachep) · line 974
Quoted 10 lines
mm/slab_common.c +10
-0+if (memcg == root_mem_cgroup) +return ; +mutex_lock (&slab_mutex); +list_for_each_entry (s, &memcg->kmem_caches, + memcg_params.kmem_caches_node) { +kmem_cache_shrink (s); + } +mutex_unlock (&slab_mutex); +cond_resched (); +}
A couple of questions:
1) how about skipping already offlined kmem_caches? They are already shrunk, so you probably won't get much out of them. Or isn't it true?
2) what's your long-term vision here? do you think that we need to shrink kmem_caches periodically, depending on memory pressure? how a user will use this new sysctl?
What's the problem you're trying to solve in general?
Thanks!
Roman
Quoted 12 lines · Roman Gushchin
Quoted 12 lines
mm/slab_common.c +11
-0+#ifdef CONFIG_MEMCG_KMEM +static void kmem_cache_shrink_memcg (struct mem_cgroup *memcg, +void __maybe_unused *arg) +{ +struct kmem_cache *s; + +if (memcg == root_mem_cgroup) +return ; +mutex_lock (&slab_mutex); +list_for_each_entry (s, &memcg->kmem_caches, + memcg_params.kmem_caches_node) {
Quoted 8 lines
mm/slab_common.c int kmem_cache_shrink(struct kmem_cache *cachep) · line 974
Quoted 5 lines
mm/slab_common.c +5
-0+kmem_cache_shrink (s); + } +mutex_unlock (&slab_mutex); +cond_resched (); +}A couple of questions:
1) how about skipping already offlinedkmem_caches? They are already shrunk, so you probably won't get much out of them. Or isn't it true?
I have been thinking about that. This patch is based on the linux tree and so don't have an easy to find out if the kmem caches have been shrinked. Rebasing this on top of linux-next, I can use the SLAB_DEACTIVATED flag as a marker for skipping the shrink.
With all the latest patches, I am still seeing 121 out of a total of 726 memcg kmem caches (1/6) that are deactivated caches after system bootup one of the test systems. My system is still using cgroup v1 and so the number may be different in a v2 setup. The next step is probably to figure out why those deactivated caches are still there.
Quoted 3 lines
2) what's your long-term vision here? do you think that we need to shrink
kmem_cachesperiodically, depending on memory pressure? how a user will use this new sysctl?
Shrinking the kmem caches under extreme memory pressure can be one way to free up extra pages, but the effect will probably be temporary.
Quoted 1 line
What's the problem you're trying to solve in general?
At least for the slub allocator, shrinking the caches allow the number of active objects reported in slabinfo to be more accurate. In addition, this allow to know the real slab memory consumption. I have been working on a BZ about continuous memory leaks with a container based workloads. The ability to shrink caches allow us to get a more accurate memory consumption picture. Another alternative is to turn on slub_debug which will then disables all the per-cpu slabs.
Anyway, I think this can be useful to others that is why I posted the patch.
Cheers,
Longman
Quoted 27 lines · Waiman Long
Quoted 12 lines · Roman Gushchin
Quoted 12 lines
mm/slab_common.c +11
-0+#ifdef CONFIG_MEMCG_KMEM +static void kmem_cache_shrink_memcg (struct mem_cgroup *memcg, +void __maybe_unused *arg) +{ +struct kmem_cache *s; + +if (memcg == root_mem_cgroup) +return ; +mutex_lock (&slab_mutex); +list_for_each_entry (s, &memcg->kmem_caches, + memcg_params.kmem_caches_node) {Quoted 8 lines
mm/slab_common.c int kmem_cache_shrink(struct kmem_cache *cachep) · line 974
Quoted 5 lines
mm/slab_common.c +5
-0+kmem_cache_shrink (s); + } +mutex_unlock (&slab_mutex); +cond_resched (); +}A couple of questions:
1) how about skipping already offlinedkmem_caches? They are already shrunk, so you probably won't get much out of them. Or isn't it true?I have been thinking about that. This patch is based on the linux tree and so don't have an easy to find out if the kmem caches have been shrinked. Rebasing this on top of linux-next, I can use the
SLAB_DEACTIVATEDflag as a marker for skipping the shrink.
Quoted 5 lines
With all the latest patches, I am still seeing 121 out of a total of 726 memcg kmem caches (1/6) that are deactivated caches after system bootup one of the test systems. My system is still using cgroup v1 and so the number may be different in a v2 setup. The next step is probably to figure out why those deactivated caches are still there.
It's not a secret: these kmem_caches are holding objects, which are in use. It's a drawback of the current slab accounting implementation: every object holds a whole page and the corresponding kmem_cache. It's optimized for a large number of objects, which are created and destroyed within the life of the cgroup (e.g. task_structs), and it works worse for long-living objects like vfs cache.
Long-term I think we need a different implementation for long-living objects, so that objects belonging to different memory cgroups can share the same page and kmem_caches.
It's a fairly big change though.
Quoted 8 lines
Quoted 4 lines
2) what's your long-term vision here? do you think that we need to shrink
kmem_cachesperiodically, depending on memory pressure? how a user will use this new sysctl?Shrinking the kmem caches under extreme memory pressure can be one way to free up extra pages, but the effect will probably be temporary.
Quoted 1 line
What's the problem you're trying to solve in general?
Quoted 7 lines
At least for the slub allocator, shrinking the caches allow the number of active objects reported in slabinfo to be more accurate. In addition, this allow to know the real slab memory consumption. I have been working on a BZ about continuous memory leaks with a container based workloads. The ability to shrink caches allow us to get a more accurate memory consumption picture. Another alternative is to turn on
slub_debugwhich will then disables all the per-cpu slabs.
I see... I agree with Michal here, that extending drop_caches sysctl isn't the best idea. Isn't it possible to achieve the same effect using slub sysfs?
Thanks!
Quoted 13 lines · Roman Gushchin
Quoted 5 lines
Quoted 3 lines
2) what's your long-term vision here? do you think that we need to shrink
kmem_cachesperiodically, depending on memory pressure? how a user will use this new sysctl?Shrinking the kmem caches under extreme memory pressure can be one way to free up extra pages, but the effect will probably be temporary.
Quoted 8 lines
Quoted 1 line
What's the problem you're trying to solve in general?
At least for the slub allocator, shrinking the caches allow the number of active objects reported in slabinfo to be more accurate. In addition, this allow to know the real slab memory consumption. I have been working on a BZ about continuous memory leaks with a container based workloads. The ability to shrink caches allow us to get a more accurate memory consumption picture. Another alternative is to turn on
slub_debugwhich will then disables all the per-cpu slabs.
Quoted 2 lines
I see... I agree with Michal here, that extending
drop_cachessysctl isn't the best idea. Isn't it possible to achieve the same effect using slub sysfs?
Yes, using the slub sysfs interface can be a possible alternative.
Cheers,
Longman
Quoted 4 lines · Roman Gushchin
so that objects belonging to different memory cgroups can share the same page and
kmem_caches.It's a fairly big change though.
Could this be done at another level? Put a cgoup pointer into the corresponding structures and then go back to just a single kmen_cache for the system as a whole? You can still account them per cgroup and there will be no cleanup problem anymore. You could scan through a slab cache to remove the objects of a certain cgroup and then the fragmentation problem that cgroups create here will be handled by the slab allocators in the traditional way. The duplication of the kmem_cache was not designed into the allocators but bolted on later.
Quoted 7 lines · Christopher Lameter
Quoted 4 lines · Roman Gushchin
so that objects belonging to different memory cgroups can share the same page and
kmem_caches.It's a fairly big change though.
Quoted 10 lines
Could this be done at another level? Put a cgoup pointer into the corresponding structures and then go back to just a single
kmen_cachefor the system as a whole?
You can still account them per cgroup and there
will be no cleanup problem anymore. You could scan through a slab cache to remove the objects of a certain cgroup and then the fragmentation problem that cgroups create here will be handled by the slab allocators in the traditional way. The duplication of thekmem_cachewas not designed into the allocators but bolted on later.
Yeah, this is exactly what I'm talking about. Idk how big the performance penalty will be for small and short-living objects, it should be measured. But for long-living objects it will be much better for sure...
Thanks!
Quoted 8 lines · Christopher Lameter
Quoted 4 lines · Roman Gushchin
so that objects belonging to different memory cgroups can share the same page and
kmem_caches.It's a fairly big change though.
Quoted 8 lines
Could this be done at another level? Put a cgoup pointer into the corresponding structures and then go back to just a single
kmen_cachefor the system as a whole? You can still account them per cgroup and there will be no cleanup problem anymore. You could scan through a slab cache to remove the objects of a certain cgroup and then the fragmentation problem that cgroups create here will be handled by the slab allocators in the traditional way. The duplication of thekmem_cachewas not designed into the allocators but bolted on later.
I'm afraid this may bring in another problem for memcg page reclaim. When shrinking the slabs, the shrinker may end up scanning a very long list to find out the slabs for a specific memcg. Particularly for the count operation, it may have to scan the list from the beginning all the way down to the end. It may take unbounded time.
When I worked on THP deferred split shrinker problem, I used to do like this, but it turns out it may take milliseconds to count the objects on the list, but it may just need reclaim a few of them.
Quoted 1 line
Quoted 24 lines · Yang Shi
Quoted 8 lines · Christopher Lameter
Quoted 4 lines · Roman Gushchin
so that objects belonging to different memory cgroups can share the same page and
kmem_caches.It's a fairly big change though.
Quoted 8 lines
Could this be done at another level? Put a cgoup pointer into the corresponding structures and then go back to just a single
kmen_cachefor the system as a whole? You can still account them per cgroup and there will be no cleanup problem anymore. You could scan through a slab cache to remove the objects of a certain cgroup and then the fragmentation problem that cgroups create here will be handled by the slab allocators in the traditional way. The duplication of thekmem_cachewas not designed into the allocators but bolted on later.I'm afraid this may bring in another problem for memcg page reclaim. When shrinking the slabs, the shrinker may end up scanning a very long list to find out the slabs for a specific memcg. Particularly for the count operation, it may have to scan the list from the beginning all the way down to the end. It may take unbounded time.
Quoted 3 lines
When I worked on THP deferred split shrinker problem, I used to do like this, but it turns out it may take milliseconds to count the objects on the list, but it may just need reclaim a few of them.
I don't think the shrinker mechanism should be altered. Shrinker lists already contain individual objects, and I don't see any reasons, why these objects can't reside on a shared set of pages.
What we're discussing is that it's way too costly (under some conditions) to have many sets of kmem_caches, if each of them is containing only few objects.
Thanks!
Quoted 36 lines · Waiman Long
Quoted 12 lines · Roman Gushchin
Quoted 12 lines
mm/slab_common.c +11
-0+#ifdef CONFIG_MEMCG_KMEM +static void kmem_cache_shrink_memcg (struct mem_cgroup *memcg, +void __maybe_unused *arg) +{ +struct kmem_cache *s; + +if (memcg == root_mem_cgroup) +return ; +mutex_lock (&slab_mutex); +list_for_each_entry (s, &memcg->kmem_caches, + memcg_params.kmem_caches_node) {Quoted 8 lines
mm/slab_common.c int kmem_cache_shrink(struct kmem_cache *cachep) · line 974
Quoted 5 lines
mm/slab_common.c +5
-0+kmem_cache_shrink (s); + } +mutex_unlock (&slab_mutex); +cond_resched (); +}A couple of questions:
1) how about skipping already offlinedkmem_caches? They are already shrunk, so you probably won't get much out of them. Or isn't it true?I have been thinking about that. This patch is based on the linux tree and so don't have an easy to find out if the kmem caches have been shrinked. Rebasing this on top of linux-next, I can use the
SLAB_DEACTIVATEDflag as a marker for skipping the shrink.With all the latest patches, I am still seeing 121 out of a total of 726 memcg kmem caches (1/6) that are deactivated caches after system bootup one of the test systems. My system is still using cgroup v1 and so the number may be different in a v2 setup. The next step is probably to figure out why those deactivated caches are still there.
Quoted 3 lines
2) what's your long-term vision here? do you think that we need to shrink
kmem_cachesperiodically, depending on memory pressure? how a user will use this new sysctl?
Quoted 8 lines
Shrinking the kmem caches under extreme memory pressure can be one way to free up extra pages, but the effect will probably be temporary.
Quoted 1 line
What's the problem you're trying to solve in general?
At least for the slub allocator, shrinking the caches allow the number of active objects reported in slabinfo to be more accurate. In addition, this allow to know the real slab memory consumption. I have been working on a BZ about continuous memory leaks with a container based workloads.
So.. this is still a work around?
Quoted 3 lines
The ability to shrink caches allow us to get a more accurate memory consumption picture. Another alternative is to turn on
slub_debugwhich will then disables all the per-cpu slabs.
So this is a debugging mechanism?
Quoted 1 line
Anyway, I think this can be useful to others that is why I posted the patch.
Since this is debug stuff, please add this to /proc/sys/debug/ instead. That would reflect the intention, and would avoid the concern that folks in production would use these things.
Since we only have 2 users of /proc/sys/debug/ I am now wondering if would be best to add a new sysctl debug taint flag. This way bug reports with these stupid knobs can got to /dev/null inbox for bug reports.
Masami, /proc/sys/debug/kprobes-optimization is debug. Would you be OK to add the taint for it too?
Masoud, /proc/sys/debug/exception-trace seems to actually be enabled by default, and its goal seems to be to enable disabling it. So I don't think it would make sense to taint there.
So.. maybe we need something /proc/sys/taints/ or /proc/sys/debug/taints/ so its *very* clear this is by no way ever expected to be used in production.
May even be good to long term add a symlink for vm/drop_caches there as well?
Luis
Michal Hocko
maintainerQuoted 8 lines · Waiman Long
With the slub memory allocator, the numbers of active slab objects reported in
/proc/slabinfoare not real because they include objects that are held by the per-cpu slab structures whether they are actually used or not. The problem gets worse the more CPUs a system have. For instance, looking at the reported number of activetask_structobjects, one will wonder where all the missing tasks gone.I know it is hard and costly to get a real count of active objects.
What exactly is expensive? Why cannot slabinfo reduce the number of active objects by per-cpu cached objects?
Quoted 8 lines
So
I am not advocating for that. Instead, this patch extends the/proc/sys/vm/drop_cachessysctl parameter by using a new bit (bit 3) to shrink all the kmem slabs which will flush out all the slabs in the per-cpu structures and give a more accurate view of how much memory are really used up by the active slab objects. This is a costly operation, of course, but it gives a way to have a clearer picture of the actual number of slab objects used, if the need arises.
drop_caches is a terrible interface. It destroys all the caching and people are just too easy in using it to solve any kind of problem they think they might have and cause others they might not see immediately. I am strongly discouraging anybody - except for some tests which really do want to see reproducible results without cache effects - from using this interface and therefore I am not really happy to paper over something that might be a real problem with yet another mode. If SLUB indeed caches too aggressively on large machines then this should be fixed.
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 8 lines · Michal Hocko
Quoted 7 lines · Waiman Long
With the slub memory allocator, the numbers of active slab objects reported in
/proc/slabinfoare not real because they include objects that are held by the per-cpu slab structures whether they are actually used or not. The problem gets worse the more CPUs a system have. For instance, looking at the reported number of activetask_structobjects, one will wonder where all the missing tasks gone.
Quoted 4 lines
Quoted 1 line
I know it is hard and costly to get a real count of active objects.
What exactly is expensive? Why cannot slabinfo reduce the number of active objects by per-cpu cached objects?
The number of cachelines that needs to be accessed in order to get an accurate count will be much higher if we need to iterate through all the per-cpu structures. In addition, accessing the per-cpu partial list will be racy.
Quoted 8 lines
Quoted 8 lines
So
I am not advocating for that. Instead, this patch extends the/proc/sys/vm/drop_cachessysctl parameter by using a new bit (bit 3) to shrink all the kmem slabs which will flush out all the slabs in the per-cpu structures and give a more accurate view of how much memory are really used up by the active slab objects. This is a costly operation, of course, but it gives a way to have a clearer picture of the actual number of slab objects used, if the need arises.
Quoted 10 lines
drop_cachesis a terrible interface. It destroys all the caching and people are just too easy in using it to solve any kind of problem they think they might have and cause others they might not see immediately. I am strongly discouraging anybody - except for some tests which really do want to see reproducible results without cache effects - from using this interface and therefore I am not really happy to paper over something that might be a real problem with yet another mode. If SLUB indeed caches too aggressively on large machines then this should be fixed.
OK, as explained in another thread, the main reason for doing this patch is to be able to do more accurate measurement of changes in kmem cache memory consumption. Yes, I do agree that drop_caches is not a general purpose interface that should be used lightly.
Cheers,
Longman
Michal Hocko
maintainerQuoted 9 lines · Waiman Long
Quoted 8 lines · Michal Hocko
Quoted 7 lines · Waiman Long
With the slub memory allocator, the numbers of active slab objects reported in
/proc/slabinfoare not real because they include objects that are held by the per-cpu slab structures whether they are actually used or not. The problem gets worse the more CPUs a system have. For instance, looking at the reported number of activetask_structobjects, one will wonder where all the missing tasks gone.
Quoted 8 lines
Quoted 4 lines
Quoted 1 line
I know it is hard and costly to get a real count of active objects.
What exactly is expensive? Why cannot slabinfo reduce the number of active objects by per-cpu cached objects?
The number of cachelines that needs to be accessed in order to get an accurate count will be much higher if we need to iterate through all the per-cpu structures. In addition, accessing the per-cpu partial list will be racy.
Why is all that a problem for a root only interface that should be used quite rarely (it is not something that you should be reading hundreds time per second, right)?
Signature · 2 lines
Michal Hocko SUSE Labs
Quoted 14 lines · Michal Hocko
Quoted 13 lines · Waiman Long
Quoted 12 lines · Michal Hocko
Quoted 8 lines · Waiman Long
With the slub memory allocator, the numbers of active slab objects reported in
/proc/slabinfoare not real because they include objects that are held by the per-cpu slab structures whether they are actually used or not. The problem gets worse the more CPUs a system have. For instance, looking at the reported number of activetask_structobjects, one will wonder where all the missing tasks gone.I know it is hard and costly to get a real count of active objects.
What exactly is expensive? Why cannot slabinfo reduce the number of active objects by per-cpu cached objects?
Quoted 7 lines
Quoted 4 lines
The number of cachelines that needs to be accessed in order to get an accurate count will be much higher if we need to iterate through all the per-cpu structures. In addition, accessing the per-cpu partial list will be racy.
Why is all that a problem for a root only interface that should be used quite rarely (it is not something that you should be reading hundreds time per second, right)?
That can be true. Anyway, I have posted a new patch to use the existing <slab>/shrink sysfs file to perform memcg cache shrinking as well. So I am not going to pursue this patch.
Thanks,
Longman