mm/list_lru: dont make them memcg aware if kmem is disabled
Subsystems (3)
PATCH 1/1 mm/list_lru: dont make them memcg aware if kmem is disabled
When alloc_super() allocates list_lrus for dentries and inodes they are made memcg aware if KMEM is compiled in, we should also check if kmem was disabled at runtime.
This overhead is about 32 bytes extra per possible nodes per caller of list_lru_init()
--- include/linux/memcontrol.h | 6 ++++++ mm/list_lru.c | 9 +++++++++ mm/memcontrol.c | 9 +++++++++ 3 files changed, 24 insertions(+)
include/linux/memcontrol.h +6 -0
@@ -1636,6 +1636,7 @@ static inline bool memcg_kmem_enabled(void)return static_branch_likely (&memcg_kmem_enabled_key); } +extern bool cgroup_kmem_disabled (void );static inline int memcg_kmem_charge_page (struct page *page,gfp_t gfp,int order) {
@@ -1691,6 +1692,11 @@ static inline bool memcg_kmem_enabled(void)return false; } +static inline bool cgroup_kmem_disabled (void ) +{ +return false; +} +static inline int memcg_cache_id (struct mem_cgroup *memcg) {return -1 ;
mm/list_lru.c +9 -0
@@ -439,6 +439,15 @@ static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware) {int i; +/* +* Override memcg_aware here, otherwise callers +* will allocate memcg aware lru's and that wastes +* memory and the amount can scale with the number of +* nodes and caller contexts. +*/ +if (cgroup_kmem_disabled ()) + memcg_aware = false; + lru->memcg_aware = memcg_aware;if (!memcg_aware)
mm/memcontrol.c +9 -0
@@ -99,6 +99,15 @@ static bool do_memsw_account(void)return !cgroup_subsys_on_dfl (memory_cgrp_subsys) && !cgroup_memory_noswap; } +/* +* NOTE: This is not the logical opposite of what +* memcg_kmem_enabled() does +*/ +bool cgroup_kmem_disabled (void ) +{ +return cgroup_memory_nokmem; +} +#define THRESHOLDS_EVENTS_TARGET 128#define SOFTLIMIT_EVENTS_TARGET 1024
Signature · 1 line
2.25.1
Quoted 8 lines · Balbir Singh
When
alloc_super()allocateslist_lrusfor dentries and inodes they are made memcg aware if KMEM is compiled in, we should also check if kmem was disabled at runtime.This overhead is about 32 bytes extra per possible nodes per caller of
list_lru_init()Signed-off-by Balbir Singh <[email protected]>
I'd rather export cgroup_memory_nokmem and make cgroup_kmem_disabled() inline, put it next to memcg_kmem_enabled() and explain in comments what each means.
And ideally, the current memcg_kmem_enabled() should be named e.g. memcg_kmem_active(), and then the new cgroup_kmem_disabled() could be named memcg_kmem_enabled(). But that's churn and potential future backport hazard, so dunno.
Quoted 71 lines
--- include/linux/memcontrol.h | 6 ++++++ mm/list_lru.c | 9 +++++++++ mm/memcontrol.c | 9 +++++++++ 3 files changed, 24 insertions(+)include/linux/memcontrol.h +6
-0@@ -1636,6 +1636,7 @@ static inline bool memcg_kmem_enabled(void)return static_branch_likely (&memcg_kmem_enabled_key); } +extern bool cgroup_kmem_disabled (void );static inline int memcg_kmem_charge_page (struct page *page,gfp_t gfp,int order) {@@ -1691,6 +1692,11 @@ static inline bool memcg_kmem_enabled(void)return false; } +static inline bool cgroup_kmem_disabled (void ) +{ +return false; +} +static inline int memcg_cache_id (struct mem_cgroup *memcg) {return -1 ;mm/list_lru.c +9
-0@@ -439,6 +439,15 @@ static int memcg_init_list_lru(struct list_lru *lru, bool memcg_aware) {int i; +/* +* Override memcg_aware here, otherwise callers +* will allocate memcg aware lru's and that wastes +* memory and the amount can scale with the number of +* nodes and caller contexts. +*/ +if (cgroup_kmem_disabled ()) + memcg_aware = false; + lru->memcg_aware = memcg_aware;if (!memcg_aware)mm/memcontrol.c +9
-0@@ -99,6 +99,15 @@ static bool do_memsw_account(void)return !cgroup_subsys_on_dfl (memory_cgrp_subsys) && !cgroup_memory_noswap; } +/* +* NOTE: This is not the logical opposite of what +* memcg_kmem_enabled() does +*/ +bool cgroup_kmem_disabled (void ) +{ +return cgroup_memory_nokmem; +} +#define THRESHOLDS_EVENTS_TARGET 128#define SOFTLIMIT_EVENTS_TARGET 1024
Quoted 10 lines · Vlastimil Babka
Quoted 8 lines · Balbir Singh
When
alloc_super()allocateslist_lrusfor dentries and inodes they are made memcg aware if KMEM is compiled in, we should also check if kmem was disabled at runtime.This overhead is about 32 bytes extra per possible nodes per caller of
list_lru_init()Signed-off-by Balbir Singh <[email protected]>
Quoted 8 lines
I'd rather export
cgroup_memory_nokmemand makecgroup_kmem_disabled()inline, put it next tomemcg_kmem_enabled()and explain in comments what each means.And ideally, the current
memcg_kmem_enabled()should be named e.g.memcg_kmem_active(), and then the newcgroup_kmem_disabled()could be namedmemcg_kmem_enabled(). But that's churn and potential future backport hazard, so dunno.
Yes, I am happy with whatever approach works to fast track the patches
Andrew, thoughts/comments?
Balbir Singh.
Andrew Morton
maintainerQuoted 15 lines · On Fri, 27 Nov 2020 20:21:11 +1100 Balbir Singh <[email protected]> wrote:
Quoted 14 lines · Vlastimil Babka
Quoted 8 lines · Balbir Singh
When
alloc_super()allocateslist_lrusfor dentries and inodes they are made memcg aware if KMEM is compiled in, we should also check if kmem was disabled at runtime.This overhead is about 32 bytes extra per possible nodes per caller of
list_lru_init()Signed-off-by Balbir Singh <[email protected]>I'd rather export
cgroup_memory_nokmemand makecgroup_kmem_disabled()inline, put it next tomemcg_kmem_enabled()and explain in comments what each means.
Quoted 9 lines
Quoted 4 lines
And ideally, the current
memcg_kmem_enabled()should be named e.g.memcg_kmem_active(), and then the newcgroup_kmem_disabled()could be namedmemcg_kmem_enabled(). But that's churn and potential future backport hazard, so dunno.Yes, I am happy with whatever approach works to fast track the patches
Andrew, thoughts/comments?
Your original changelog doesn't make the case that the patch should be fast tracked, so it looks like there's missing information. Please don't miss information ;)
If we're looking for a backportable quickfix for the not-yet-really-explained problem then yes, Vlastimil's suggestion (as a high-priority patch and a separate low-priority patch) sounds good.
It is rather a twisty maze of identifiers, so please do comment everything well.
Quoted 22 lines · Andrew Morton
Quoted 20 lines · On Fri, 27 Nov 2020 20:21:11 +1100 Balbir Singh <[email protected]> wrote:
Quoted 18 lines · Vlastimil Babka
Quoted 8 lines · Balbir Singh
When
alloc_super()allocateslist_lrusfor dentries and inodes they are made memcg aware if KMEM is compiled in, we should also check if kmem was disabled at runtime.This overhead is about 32 bytes extra per possible nodes per caller of
list_lru_init()Signed-off-by Balbir Singh <[email protected]>I'd rather export
cgroup_memory_nokmemand makecgroup_kmem_disabled()inline, put it next tomemcg_kmem_enabled()and explain in comments what each means.And ideally, the current
memcg_kmem_enabled()should be named e.g.memcg_kmem_active(), and then the newcgroup_kmem_disabled()could be namedmemcg_kmem_enabled(). But that's churn and potential future backport hazard, so dunno.
Quoted 9 lines
Quoted 4 lines
Yes, I am happy with whatever approach works to fast track the patches
Andrew, thoughts/comments?
Your original changelog doesn't make the case that the patch should be fast tracked, so it looks like there's missing information. Please don't miss information ;)
I just wanted to get the patches done :) Nothing to be seriously backported. I can refactor the code, I've seen cases where even with kmem disabled on pressure, kvmalloc leads to wasting memory for allocating memcg aware list_lrus.
With older kernels (5.4 and before) where, creation of a new namespace always caused these allocations to occur via pid_ns_prepare_proc(), the issue is quite obviously exposed - the waster is a function of
wasted_memory = sizeof (list_lru_memcg+4) * number of nodes * number of memcgs with kmem enabled ( per clone/container ) -- (1)
That's not too bad today, but I am afraid it's useless memory being allocated and can become larger if the number of possible nodes increase.
Quoted 6 lines
If we're looking for a backportable quickfix for the not-yet-really-explained problem then yes, Vlastimil's suggestion (as a high-priority patch and a separate low-priority patch) sounds good.
It is rather a twisty maze of identifiers, so please do comment everything well.
Yes, I am going to refactor things out a bit and resend the patches
Balbir Singh.