mm: reparent slab memory on cgroup removal
Subsystems (4)
v5 cover mm: reparent slab memory on cgroup removal
# Why do we need this?
We've noticed that the number of dying cgroups is steadily growing on most of our hosts in production. The following investigation revealed an issue in userspace memory reclaim code [1], accounting of kernel stacks [2], and also the mainreason: slab objects.
The underlying problem is quite simple: any page charged
to a cgroup holds a reference to it, so the cgroup can't be reclaimed unless all charged pages are gone. If a slab object is actively used by other cgroups, it won't be reclaimed, and will prevent the origin cgroup from being reclaimed.
Slab objects, and first of all vfs cache, is shared between cgroups, which are using the same underlying fs, and what's even more important, it's shared between multiple generations of the same workload. So if something is running periodically every time in a new cgroup (like how systemd works), we do accumulate multiple dying cgroups.
Strictly speaking pagecache isn't different here, but there is a key difference:
we disable protection and apply some extra pressure on LRUs of dying cgroups, and these LRUs contain all charged pages.
My experiments show that with the disabled kernel memory accounting the number of dying cgroups stabilizes at a relatively small number (~100, depends on memory pressure and cgroup creation rate), and with kernel memory accounting it grows pretty steadily up to several thousands.
Memory cgroups are quite complex and big objects (mostly due to percpu stats), so it leads to noticeable memory losses. Memory occupied by dying cgroups is measured in hundreds of megabytes. I've even seen a host with more than 100Gb of memory wasted for dying cgroups. It leads to a degradation of performance with the uptime, and generally limits the usage of cgroups.
My previous attempt [3] to fix the problem by applying extra pressure on slab shrinker lists caused a regressions with xfs and ext4, and has been reverted [4]. The following attempts to find the right balance [5, 6] were not successful.
So instead of trying to find a maybe non-existing balance, let's do reparent the accounted slabs to the parent cgroup on cgroup removal.
# Implementation approach
There is however a significant problem with reparenting of slab memory:
there is no list of charged pages. Some of them are in shrinker lists, but not all. Introducing of a new list is really not an option.
But fortunately there is a way forward: every slab page has a stable pointer to the corresponding kmem_cache. So the idea is to reparent kmem_caches instead of slab pages.
It's actually simpler and cheaper, but requires some underlying changes:
1) Make kmem_caches to hold a single reference to the memory cgroup, instead of a separate reference per every slab page.
2) Stop setting page->mem_cgroup pointer for memcg slab pages and use page->kmem_cache->memcg indirection instead. It's used only on slab page release, so it shouldn't be a big issue.
3) Introduce a refcounter for non-root slab caches. It's required to be able to destroy kmem_caches when they become empty and release the associated memory cgroup.
There is a bonus: currently we do release empty kmem_caches on cgroup removal, however all other are waiting for the releasing of the memory cgroup. These refactorings allow kmem_caches to be released as soon as they become inactive and free.
Some additional implementation details are provided in corresponding commit messages.
# Results
Below is the average number of dying cgroups on two groups of our production hosts. They do run some sort of web frontend workload, the memory pressure is moderate. As we can see, with the kernel memory reparenting the number stabilizes in 60s range; however with the original version it grows almost linearly and doesn't show any signs of plateauing. The difference in slab and percpu usage between patched and unpatched versions also grows linearly. In 7 days it exceeded 200Mb.
day 0 1 2 3 4 5 6 7
original 56 362 628 752 1070 1250 1490 1560
patched 23 46 51 55 60 57 67 69
mem diff(Mb) 22 74 123 152 164 182 214 241
# History
v5:
1) fixed a compilation warning around missing kmemcg_queue_cache_shutdown()
2) s/rcu_read_lock()/rcu_read_unlock() in memcg_kmem_get_cache()
v4:
1) removed excessive memcg != parent check in memcg_deactivate_kmem_caches()
2) fixed rcu_read_lock() usage in memcg_charge_slab()
3) fixed synchronization around dying flag in kmemcg_queue_cache_shutdown()
4) refreshed test results data
5) reworked PageTail() checks in memcg_from_slab_page()
6) added some comments in multiple places
v3:
1) reworked memcg kmem_cache search on allocation path
2) fixed /proc/kpagecgroup interface
v2:
1) switched to percpu kmem_cache refcounter
2) a reference to kmem_cache is held during the allocation
3) slabs stats are fixed for !MEMCG case (and the refactoring is separated into a standalone patch)
4) kmem_cache reparenting is performed from deactivatation context
v1:
https://lkml.org/lkml/2019/4/17/1095
# Links
[1]: commit 68600f623d69 ("mm: don't miss the last page because of round-off error")
[2]: commit 9b6f7e163cd0 ("mm: rework memcg kernel stack accounting") [3]: commit 172b06c32b94 ("mm: slowly shrink slabs with a relatively small number of objects")
[4]: commit a9a238e83fbb ("Revert "mm: slowly shrink slabs with a relatively small number of objects")
[5]: https://lkml.org/lkml/2019/1/28/1865
[6]: https://marc.info/?l=linux-mm&m=155064763626437&w=2
Roman Gushchin (7):
mm: postpone kmem_cache memcg pointer initialization to
memcg_link_cache()
mm: generalize postponed non-root kmem_cache deactivation
mm: introduce __memcg_kmem_uncharge_memcg()
mm: unify SLAB and SLUB page accounting
mm: rework non-root kmem_cache lifecycle management
mm: reparent slab memory on cgroup removal
mm: fix /proc/kpagecgroup interface for slab pages
include/linux/memcontrol.h | 10 +++ include/linux/slab.h | 13 +-- mm/memcontrol.c | 101 ++++++++++++++++------- mm/slab.c | 25 ++---- mm/slab.h | 137 ++++++++++++++++++++++++------- mm/slab_common.c | 162 +++++++++++++++++++++---------------- mm/slub.c | 36 ++------- 7 files changed, 299 insertions(+), 185 deletions(-)
Signature · 1 line
2.20.1
PATCH v5 1/7 mm: postpone kmem_cache memcg pointer initialization to memcg_link_cache()
Initialize kmem_cache->memcg_params.memcg pointer in memcg_link_cache() rather than in init_memcg_params().
Once kmem_cache will hold a reference to the memory cgroup, it will simplify the refcounting.
For non-root kmem_caches memcg_link_cache() is always called before the kmem_cache becomes visible to a user, so it's safe.
--- mm/slab.c | 2 +- mm/slab.h | 5 +++-- mm/slab_common.c | 14 +++++++------- mm/slub.c | 2 +- 4 files changed, 12 insertions(+), 11 deletions(-)
mm/slab.c +1 -1
@@ -1268,7 +1268,7 @@ void __init kmem_cache_init(void) nr_node_ids *sizeof (struct kmem_cache_node *), SLAB_HWCACHE_ALIGN, 0, 0);list_add (&kmem_cache->list, &slab_caches); -memcg_link_cache (kmem_cache); +memcg_link_cache (kmem_cache,NULL ); slab_state =PARTIAL ;/*
mm/slab.h +3 -2
@@ -289,7 +289,7 @@ static __always_inline void memcg_uncharge_slab(struct page *page, int order, }extern void slab_init_memcg_params (struct kmem_cache *); -extern void memcg_link_cache (struct kmem_cache *s); +extern void memcg_link_cache (struct kmem_cache *s,struct mem_cgroup *memcg);extern void slab_deactivate_memcg_cache_rcu_sched (struct kmem_cache *s,void (*deact_fn)(struct kmem_cache *));
@@ -344,7 +344,8 @@ static inline void slab_init_memcg_params(struct kmem_cache *s) { } -static inline void memcg_link_cache (struct kmem_cache *s) +static inline void memcg_link_cache (struct kmem_cache *s, +struct mem_cgroup *memcg) { }
mm/slab_common.c +7 -7
@@ -140,13 +140,12 @@ void slab_init_memcg_params(struct kmem_cache *s) }static int init_memcg_params (struct kmem_cache *s, -struct mem_cgroup *memcg,struct kmem_cache *root_cache) +struct kmem_cache *root_cache) {struct memcg_cache_array *arr;if (root_cache) { s->memcg_params.root_cache = root_cache; - s->memcg_params.memcg = memcg;INIT_LIST_HEAD (&s->memcg_params.children_node);INIT_LIST_HEAD (&s->memcg_params.kmem_caches_node);return 0 ;
@@ -221,11 +220,12 @@ int memcg_update_all_caches(int num_memcgs)return ret; } -void memcg_link_cache (struct kmem_cache *s) +void memcg_link_cache (struct kmem_cache *s,struct mem_cgroup *memcg) {if (is_root_cache (s)) {list_add (&s->root_caches_node, &slab_root_caches); }else { + s->memcg_params.memcg = memcg;list_add (&s->memcg_params.children_node, &s->memcg_params.root_cache->memcg_params.children); list_add(&s->memcg_params.kmem_caches_node,
@@ -244,7 +244,7 @@ static void memcg_unlink_cache(struct kmem_cache *s) }#else static inline int init_memcg_params (struct kmem_cache *s, -struct mem_cgroup *memcg,struct kmem_cache *root_cache) +struct kmem_cache *root_cache) {return 0 ; }
@@ -384,7 +384,7 @@ static struct kmem_cache *create_cache(const char *name, s->useroffset = useroffset; s->usersize = usersize; - err =init_memcg_params (s, memcg, root_cache); + err =init_memcg_params (s, root_cache);if (err) goto out_free_cache;
@@ -394,7 +394,7 @@ static struct kmem_cache *create_cache(const char *name, s->refcount =1 ;list_add (&s->list, &slab_caches); -memcg_link_cache (s); +memcg_link_cache (s, memcg); out:if (err)return ERR_PTR (err);
@@ -997,7 +997,7 @@ struct kmem_cache *__init create_kmalloc_cache(const char *name,create_boot_cache (s, name, size, flags, useroffset, usersize);list_add (&s->list, &slab_caches); -memcg_link_cache (s); +memcg_link_cache (s,NULL ); s->refcount =1 ;return s; }
mm/slub.c +1 -1
@@ -4215,7 +4215,7 @@ static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) }slab_init_memcg_params (s);list_add (&s->list, &slab_caches); -memcg_link_cache (s); +memcg_link_cache (s,NULL );return s; }
Signature · 1 line
2.20.1
Re: [PATCH v5 1/7] mm: postpone kmem_cache memcg pointer initialization to memcg_link_cache()
Quoted 11 lines · Roman Gushchin
Initialize
kmem_cache->memcg_params.memcgpointer inmemcg_link_cache()rather than ininit_memcg_params().Once
kmem_cachewill hold a reference to the memory cgroup, it will simplify the refcounting.For non-root
kmem_cachesmemcg_link_cache()is always called before thekmem_cachebecomes visible to a user, so it's safe.Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>
This one is a no-brainer.
Provided it's necessary for the rest of the series.
Re: [PATCH v5 1/7] mm: postpone kmem_cache memcg pointer initialization to memcg_link_cache()
In reply to Roman GushchinQuoted 11 lines · Roman Gushchin
Initialize
kmem_cache->memcg_params.memcgpointer inmemcg_link_cache()rather than ininit_memcg_params().Once
kmem_cachewill hold a reference to the memory cgroup, it will simplify the refcounting.For non-root
kmem_cachesmemcg_link_cache()is always called before thekmem_cachebecomes visible to a user, so it's safe.Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>
PATCH v5 2/7 mm: generalize postponed non-root kmem_cache deactivation
In reply to Roman GushchinCurrently SLUB uses a work scheduled after an RCU grace period to deactivate a non-root kmem_cache. This mechanism can be reused for kmem_caches reparenting, but requires some generalization.
Let's decouple all infrastructure (rcu callback, work callback) from the SLUB-specific code, so it can be used with SLAB as well.
Also, let's rename some functions to make the code look simpler. All SLAB/SLUB-specific functions start with "__". Remove "deact_" prefix from the corresponding struct fields.
Here is the graph of a new calling scheme:kmemcg_cache_deactivate()
__kmemcg_cache_deactivate() SLAB/SLUB-specific
kmemcg_schedule_work_after_rcu() rcu
kmemcg_after_rcu_workfn() work kmemcg_cache_deactivate_after_rcu()
__kmemcg_cache_deactivate_after_rcu() SLAB/SLUB-specific
instead of:
__kmemcg_cache_deactivate() SLAB/SLUB-specific
slab_deactivate_memcg_cache_rcu_sched() SLUB-only
kmemcg_deactivate_rcufn SLUB-only, rcu
kmemcg_deactivate_workfn SLUB-only, work
kmemcg_cache_deact_after_rcu() SLUB-only--- include/linux/slab.h | 6 ++--- mm/slab.c | 4 +++ mm/slab.h | 3 ++- mm/slab_common.c | 62 ++++++++++++++++++++------------------------ mm/slub.c | 8 +----- 5 files changed, 38 insertions(+), 45 deletions(-)
include/linux/slab.h +3 -3
@@ -642,10 +642,10 @@ struct memcg_cache_params {struct list_head children_node;struct list_head kmem_caches_node; -void (*deact_fn)(struct kmem_cache *); +void (*work_fn)(struct kmem_cache *);union { -struct rcu_head deact_rcu_head; -struct work_struct deact_work; +struct rcu_head rcu_head; +struct work_struct work; }; }; };
mm/slab.c +4 -0
@@ -2281,6 +2281,10 @@ void __kmemcg_cache_deactivate(struct kmem_cache *cachep) {__kmem_cache_shrink (cachep); } + +void __kmemcg_cache_deactivate_after_rcu (struct kmem_cache *s) +{ +}#endif int __kmem_cache_shutdown (struct kmem_cache *cachep)
mm/slab.h +2 -1
@@ -172,6 +172,7 @@ int __kmem_cache_shutdown(struct kmem_cache *);void __kmem_cache_release (struct kmem_cache *);int __kmem_cache_shrink (struct kmem_cache *);void __kmemcg_cache_deactivate (struct kmem_cache *s); +void __kmemcg_cache_deactivate_after_rcu (struct kmem_cache *s);void slab_kmem_cache_release (struct kmem_cache *);struct seq_file ;
@@ -291,7 +292,7 @@ static __always_inline void memcg_uncharge_slab(struct page *page, int order,extern void slab_init_memcg_params (struct kmem_cache *);extern void memcg_link_cache (struct kmem_cache *s,struct mem_cgroup *memcg);extern void slab_deactivate_memcg_cache_rcu_sched (struct kmem_cache *s, -void (*deact_fn)(struct kmem_cache *)); +void (*work_fn)(struct kmem_cache *));#else /* CONFIG_MEMCG_KMEM */
mm/slab_common.c +28 -34
108 lines not on this page · raw message
mm/slub.c +1 -7
@@ -4024,7 +4024,7 @@ int __kmem_cache_shrink(struct kmem_cache *s) }#ifdef CONFIG_MEMCG -static void kmemcg_cache_deact_after_rcu (struct kmem_cache *s) +void __kmemcg_cache_deactivate_after_rcu (struct kmem_cache *s) {/* * Called with all the locks held after a sched RCU grace period.
@@ -4050,12 +4050,6 @@ void __kmemcg_cache_deactivate(struct kmem_cache *s)*/ slub_set_cpu_partial (s,0 ); s->min_partial = 0; - -/* -* s->cpu_partial is checked locklessly (see put_cpu_partial), so -* we have to make sure the change is visible before shrinking. -*/ - slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu); }#endif /* CONFIG_MEMCG */
Signature · 1 line
2.20.1
Re: [PATCH v5 2/7] mm: generalize postponed non-root kmem_cache deactivation
Quoted 7 lines · Roman Gushchin
mm/slab_common.c +0
-0@@ -866,11 +859,12 @@ static void flush_memcg_workqueue(struct kmem_cache *s)mutex_unlock (&slab_mutex);
mm/slab_common.c static void flush_memcg_workqueue(struct kmem_cache *s) · line 867
Quoted 10 lines
mm/slab_common.c +4
-3/* -* SLUB deactivates the kmem_caches through call_rcu. Make +* SLAB and SLUB deactivate the kmem_caches through call_rcu. Make * sure all registered rcu callbacks have been invoked. */ -if (IS_ENABLED (CONFIG_SLUB )) -rcu_barrier (); +#ifndef CONFIG_SLOB +rcu_barrier (); +#endif
Nit: you don't need to check CONFIG_SLOB here as this code is under CONFIG_MEMCG_KMEM which depends on !SLOB.
Other than that, the patch looks good to me, provided using rcu for slab deactivation proves to be really necessary, although I'd split it in two patches - one doing renaming another refactoring - easier to review that way.
PATCH v5 3/7 mm: introduce __memcg_kmem_uncharge_memcg()
In reply to Roman GushchinLet's separate the page counter modification code out of __memcg_kmem_uncharge() in a way similar to what __memcg_kmem_charge() and __memcg_kmem_charge_memcg() work.
This will allow to reuse this code later using a new memcg_kmem_uncharge_memcg() wrapper, which calls __memcg_kmem_uncharge_memcg() if memcg_kmem_enabled() check is passed.
--- include/linux/memcontrol.h | 10 ++++++++++ mm/memcontrol.c | 25 +++++++++++++++++-------- 2 files changed, 27 insertions(+), 8 deletions(-)
include/linux/memcontrol.h +10 -0
@@ -1298,6 +1298,8 @@ int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order);void __memcg_kmem_uncharge (struct page *page,int order);int __memcg_kmem_charge_memcg (struct page *page,gfp_t gfp,int order,struct mem_cgroup *memcg); +void __memcg_kmem_uncharge_memcg (struct mem_cgroup *memcg, +unsigned int nr_pages);extern struct static_key_false memcg_kmem_enabled_key;extern struct workqueue_struct *memcg_kmem_cache_wq;
@@ -1339,6 +1341,14 @@ static inline int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp,return __memcg_kmem_charge_memcg (page, gfp, order, memcg);return 0 ; } + +static inline void memcg_kmem_uncharge_memcg (struct page *page,int order, +struct mem_cgroup *memcg) +{ +if (memcg_kmem_enabled ()) +__memcg_kmem_uncharge_memcg (memcg,1 << order); +} +/* * helper for accessing a memcg's index. It will be used as an index in the * child cache array in kmem_cache, and also to derive its name. This function
mm/memcontrol.c +17 -8
@@ -2750,6 +2750,22 @@ int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)css_put (&memcg->css);return ret; } + +/** +* __memcg_kmem_uncharge_memcg: uncharge a kmem page +* @memcg: memcg to uncharge +* @nr_pages: number of pages to uncharge +*/ +void __memcg_kmem_uncharge_memcg (struct mem_cgroup *memcg, +unsigned int nr_pages) +{ +if (!cgroup_subsys_on_dfl (memory_cgrp_subsys)) +page_counter_uncharge (&memcg->kmem, nr_pages); + +page_counter_uncharge (&memcg->memory, nr_pages); +if (do_memsw_account ()) +page_counter_uncharge (&memcg->memsw, nr_pages); +}/** * __memcg_kmem_uncharge: uncharge a kmem page * @page: page to uncharge
@@ -2764,14 +2780,7 @@ void __memcg_kmem_uncharge(struct page *page, int order)return ;VM_BUG_ON_PAGE (mem_cgroup_is_root (memcg), page); - -if (!cgroup_subsys_on_dfl (memory_cgrp_subsys)) -page_counter_uncharge (&memcg->kmem, nr_pages); - -page_counter_uncharge (&memcg->memory, nr_pages); -if (do_memsw_account ()) -page_counter_uncharge (&memcg->memsw, nr_pages); - +__memcg_kmem_uncharge_memcg (memcg, nr_pages); page->mem_cgroup =NULL ;/* slab pages do not have PageKmemcg flag set */
Signature · 1 line
2.20.1
PATCH v5 4/7 mm: unify SLAB and SLUB page accounting
In reply to Roman GushchinCurrently the page accounting code is duplicated in SLAB and SLUB internals. Let's move it into new (un)charge_slab_page helpers in the slab_common.c file. These helpers will be responsible for statistics (global and memcg-aware) and memcg charging. So they are replacing direct memcg_(un)charge_slab() calls.
--- mm/slab.c | 19 +++---------------- mm/slab.h | 25 +++++++++++++++++++++++++ mm/slub.c | 14 ++------------ 3 files changed, 30 insertions(+), 28 deletions(-)
mm/slab.c +3 -16
@@ -1389,7 +1389,6 @@ static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) {struct page *page; -int nr_pages; flags |= cachep->allocflags;
@@ -1399,17 +1398,11 @@ static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,return NULL ; } -if (memcg_charge_slab (page, flags, cachep->gfporder, cachep)) { +if (charge_slab_page (page, flags, cachep->gfporder, cachep)) {__free_pages (page, cachep->gfporder);return NULL ; } - nr_pages = (1 << cachep->gfporder); -if (cachep->flags &SLAB_RECLAIM_ACCOUNT ) -mod_lruvec_page_state (page,NR_SLAB_RECLAIMABLE , nr_pages); -else -mod_lruvec_page_state (page,NR_SLAB_UNRECLAIMABLE , nr_pages); -__SetPageSlab (page);/* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ if (sk_memalloc_socks () &&page_is_pfmemalloc (page))
@@ -1424,12 +1417,6 @@ static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,static void kmem_freepages (struct kmem_cache *cachep,struct page *page) {int order = cachep->gfporder; -unsigned long nr_freed = (1 << order); - -if (cachep->flags &SLAB_RECLAIM_ACCOUNT ) -mod_lruvec_page_state (page,NR_SLAB_RECLAIMABLE , -nr_freed); -else -mod_lruvec_page_state (page,NR_SLAB_UNRECLAIMABLE , -nr_freed);BUG_ON (!PageSlab (page));__ClearPageSlabPfmemalloc (page);
@@ -1438,8 +1425,8 @@ static void kmem_freepages(struct kmem_cache *cachep, struct page *page) page->mapping =NULL ;if (current->reclaim_state) - current->reclaim_state->reclaimed_slab += nr_freed; -memcg_uncharge_slab (page, order, cachep); + current->reclaim_state->reclaimed_slab +=1 << order; +uncharge_slab_page (page, order, cachep);__free_pages (page, order); }
mm/slab.h +25 -0
@@ -205,6 +205,12 @@ ssize_t slabinfo_write(struct file *file, const char __user *buffer,void __kmem_cache_free_bulk (struct kmem_cache *,size_t ,void **);int __kmem_cache_alloc_bulk (struct kmem_cache *,gfp_t ,size_t ,void **); +static inline int cache_vmstat_idx (struct kmem_cache *s) +{ +return (s->flags &SLAB_RECLAIM_ACCOUNT ) ? +NR_SLAB_RECLAIMABLE :NR_SLAB_UNRECLAIMABLE ; +} +#ifdef CONFIG_MEMCG_KMEM /* List of all root caches. */
@@ -352,6 +358,25 @@ static inline void memcg_link_cache(struct kmem_cache *s,#endif /* CONFIG_MEMCG_KMEM */ +static __always_inline int charge_slab_page(struct page *page, +gfp_t gfp,int order, +struct kmem_cache *s) +{ +int ret =memcg_charge_slab (page, gfp, order, s); + +if (!ret) +mod_lruvec_page_state (page,cache_vmstat_idx (s),1 << order); + +return ret; +} + +static __always_inline void uncharge_slab_page(struct page *page, int order, +struct kmem_cache *s) +{ +mod_lruvec_page_state (page,cache_vmstat_idx (s), -(1 << order)); +memcg_uncharge_slab (page, order, s); +} +static inline struct kmem_cache *cache_from_obj (struct kmem_cache *s,void *x) {struct kmem_cache *cachep;
mm/slub.c +2 -12
@@ -1490,7 +1490,7 @@ static inline struct page *alloc_slab_page(struct kmem_cache *s,else page =__alloc_pages_node (node, flags, order); -if (page &&memcg_charge_slab (page, flags, order, s)) { +if (page &&charge_slab_page (page, flags, order, s)) {__free_pages (page, order); page =NULL ; }
@@ -1683,11 +1683,6 @@ static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)if (!page)return NULL ; -mod_lruvec_page_state (page, - (s->flags &SLAB_RECLAIM_ACCOUNT ) ? -NR_SLAB_RECLAIMABLE :NR_SLAB_UNRECLAIMABLE , -1 <<oo_order (oo)); -inc_slabs_node (s,page_to_nid (page), page->objects);return page;
@@ -1721,18 +1716,13 @@ static void __free_slab(struct kmem_cache *s, struct page *page)check_object (s, page, p,SLUB_RED_INACTIVE ); } -mod_lruvec_page_state (page, - (s->flags &SLAB_RECLAIM_ACCOUNT ) ? -NR_SLAB_RECLAIMABLE :NR_SLAB_UNRECLAIMABLE , - -pages); -__ClearPageSlabPfmemalloc (page);__ClearPageSlab (page); page->mapping =NULL ;if (current->reclaim_state) current->reclaim_state->reclaimed_slab += pages; -memcg_uncharge_slab (page, order, s); +uncharge_slab_page (page, order, s);__free_pages (page, order); }
Signature · 1 line
2.20.1
Re: [PATCH v5 4/7] mm: unify SLAB and SLUB page accounting
Quoted 9 lines · Roman Gushchin
Currently the page accounting code is duplicated in SLAB and SLUB internals. Let's move it into new (un)charge_slab_page helpers in the
slab_common.cfile. These helpers will be responsible for statistics (global and memcg-aware) and memcg charging. So they are replacing direct memcg_(un)charge_slab() calls.Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>
Makes sense even without the rest of the series,
Re: [PATCH v5 4/7] mm: unify SLAB and SLUB page accounting
In reply to Roman GushchinQuoted 9 lines · Roman Gushchin
Currently the page accounting code is duplicated in SLAB and SLUB internals. Let's move it into new (un)charge_slab_page helpers in the
slab_common.cfile. These helpers will be responsible for statistics (global and memcg-aware) and memcg charging. So they are replacing direct memcg_(un)charge_slab() calls.Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>
PATCH v5 5/7 mm: rework non-root kmem_cache lifecycle management
In reply to Roman GushchinThis commit makes several important changes in the lifecycle of a non-root kmem_cache, which also affect the lifecycle of a memory cgroup.
Currently each charged slab page has a page->mem_cgroup pointer to the memory cgroup and holds a reference to it. Kmem_caches are held by the memcg and are released with it. It means that none of kmem_caches are released unless at least one reference to the memcg exists, which is not optimal.
So the current scheme can be illustrated as:page->mem_cgroup->kmem_cache.
To implement the slab memory reparenting we need to invert the scheme
into: page->kmem_cache->mem_cgroup.
Let's make every page to hold a reference to the kmem_cache (we already have a stable pointer), and make kmem_caches to hold a single reference to the memory cgroup.
To make this possible we need to introduce a new percpu refcounter for non-root kmem_caches. The counter is initialized to the percpu mode, and is switched to atomic mode after deactivation, so we never shutdown an active cache. The counter is bumped for every charged page and also for every running allocation. So the kmem_cache can't be released unless all allocations complete.
To shutdown non-active empty kmem_caches, let's reuse the infrastructure of the RCU-delayed work queue, used previously for the deactivation. After the generalization, it's perfectly suited for our needs.
Since now we can release a kmem_cache at any moment after the deactivation, let's call sysfs_slab_remove() only from the shutdown path. It makes deactivation path simpler.
Because we don't set the page->mem_cgroup pointer, we need to change the way how memcg-level stats is working for slab pages. We can't use mod_lruvec_page_state() helpers anymore, so switch over to mod_lruvec_state().
* I used the following simple approach to test the performance (stolen from another patchset by T. Harding):
time find / -name fname-no-exist
echo 2 > /proc/sys/vm/drop_caches
repeat 10 times
Results (I've chosen best results in several runs):
orig patched
real 0m0.648s real 0m0.593s
user 0m0.148s user 0m0.162s
sys 0m0.295s sys 0m0.253s
real 0m0.581s real 0m0.649s
user 0m0.119s user 0m0.136s
sys 0m0.254s sys 0m0.250s
real 0m0.645s real 0m0.705s
user 0m0.138s user 0m0.138s
sys 0m0.263s sys 0m0.250s
real 0m0.691s real 0m0.718s
user 0m0.139s user 0m0.134s
sys 0m0.262s sys 0m0.253s
real 0m0.654s real 0m0.715s
user 0m0.146s user 0m0.128s
sys 0m0.247s sys 0m0.261s
real 0m0.675s real 0m0.717s
user 0m0.129s user 0m0.137s
sys 0m0.277s sys 0m0.248s
real 0m0.631s real 0m0.719s
user 0m0.137s user 0m0.134s
sys 0m0.255s sys 0m0.251s
real 0m0.622s real 0m0.715s
user 0m0.108s user 0m0.124s
sys 0m0.279s sys 0m0.264s
real 0m0.651s real 0m0.669s
user 0m0.139s user 0m0.139s
sys 0m0.252s sys 0m0.247s
real 0m0.671s real 0m0.632s
user 0m0.130s user 0m0.139s
sys 0m0.263s sys 0m0.245s
So it looks like the difference is not noticeable in this test.
--- include/linux/slab.h | 3 +- mm/memcontrol.c | 57 +++++++++++++++++++++--------- mm/slab.h | 82 +++++++++++++++++++++++++------------------- mm/slab_common.c | 74 +++++++++++++++++++++++---------------- mm/slub.c | 12 +------ 5 files changed, 135 insertions(+), 93 deletions(-)
include/linux/slab.h +2 -1
@@ -16,6 +16,7 @@#include <linux/overflow.h> #include <linux/types.h> #include <linux/workqueue.h> +#include <linux/percpu-refcount.h> /*
@@ -152,7 +153,6 @@ int kmem_cache_shrink(struct kmem_cache *);void memcg_create_kmem_cache (struct mem_cgroup *,struct kmem_cache *);void memcg_deactivate_kmem_caches (struct mem_cgroup *); -void memcg_destroy_kmem_caches (struct mem_cgroup *);/* * Please use this macro to create slab caches. Simply specify the
@@ -641,6 +641,7 @@ struct memcg_cache_params {struct mem_cgroup *memcg;struct list_head children_node;struct list_head kmem_caches_node; +struct percpu_ref refcnt;void (*work_fn)(struct kmem_cache *);union {
mm/memcontrol.c +41 -16
114 lines not on this page · raw message
mm/slab.h +47 -35
121 lines not on this page · raw message
mm/slab_common.c +44 -30
144 lines not on this page · raw message
mm/slub.c +1 -11
@@ -4018,18 +4018,8 @@ void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s) {/* * Called with all the locks held after a sched RCU grace period. -* Even if @s becomes empty after shrinking, we can't know that @s -* doesn't have allocations already in-flight and thus can't -* destroy @s until the associated memcg is released. -* -* However, let's remove the sysfs files for empty caches here. -* Each cache has a lot of interface files which aren't -* particularly useful for empty draining caches; otherwise, we can -* easily end up with millions of unnecessary sysfs files on -* systems which have a lot of memory and transient cgroups. */ -if (!__kmem_cache_shrink (s)) -sysfs_slab_remove (s); +__kmem_cache_shrink (s); }void __kmemcg_cache_deactivate (struct kmem_cache *s)
Signature · 1 line
2.20.1
Re: [PATCH v5 5/7] mm: rework non-root kmem_cache lifecycle management
Hello Roman,
Quoted 13 lines · Roman Gushchin
This commit makes several important changes in the lifecycle of a non-root
kmem_cache, which also affect the lifecycle of a memory cgroup.Currently each charged slab page has a
page->mem_cgrouppointer to the memory cgroup and holds a reference to it.Kmem_cachesare held by the memcg and are released with it. It means that none ofkmem_cachesare released unless at least one reference to the memcg exists, which is not optimal.So the current scheme can be illustrated as:
page->mem_cgroup->kmem_cache.
Quoted 6 lines
To implement the slab memory reparenting we need to invert the scheme
into:page->kmem_cache->mem_cgroup.Let's make every page to hold a reference to the
kmem_cache(we already have a stable pointer), and makekmem_cachesto hold a single reference to the memory cgroup.
Is there any reason why we can't reference both mem cgroup and kmem cache per each charged kmem page? I mean,
page->mem_cgroup references mem_cgroup page->kmem_cache references kmem_cache mem_cgroup references kmem_cache while it's online
TBO it seems to me that not taking a reference to mem cgroup per charged kmem page makes the code look less straightforward, e.g. as you mentioned in the commit log, we have to use mod_lruvec_state() for memcg pages and mod_lruvec_page_state() for root pages.
Quoted 8 lines
To make this possible we need to introduce a new percpu refcounter for non-root
kmem_caches. The counter is initialized to the percpu mode, and is switched to atomic mode after deactivation, so we never shutdown an active cache. The counter is bumped for every charged page and also for every running allocation. So thekmem_cachecan't be released unless all allocations complete.
Quoted 8 lines
To shutdown non-active empty
kmem_caches, let's reuse the infrastructure of the RCU-delayed work queue, used previously for the deactivation. After the generalization, it's perfectly suited for our needs.Since now we can release a
kmem_cacheat any moment after the deactivation, let's callsysfs_slab_remove()only from the shutdown path. It makes deactivation path simpler.
But a cache can be dangling for quite a while after cgroup was taken down, even after this patch, because there still can be pages charged to it. The reason why we call sysfs_slab_remove() is to delete associated files from sysfs ASAP. I'd try to preserve the current behavior if possible.
Quoted 15 lines
Because we don't set the
page->mem_cgrouppointer, we need to change the way how memcg-level stats is working for slab pages. We can't usemod_lruvec_page_state()helpers anymore, so switch over tomod_lruvec_state().mm/slab_common.c +1
-0@@ -727,9 +737,31 @@ static void kmemcg_schedule_work_after_rcu(struct rcu_head *head)queue_work (memcg_kmem_cache_wq, &s->memcg_params.work); } +static void kmemcg_cache_shutdown_after_rcu (struct kmem_cache *s)
mm/slab_common.c static void kmemcg_schedule_work_after_rcu(struct rcu_head *head) · line 750
Quoted 10 lines
mm/slab_common.c +10
-0+{ +WARN_ON (shutdown_cache (s)); +} + +static void kmemcg_queue_cache_shutdown (struct percpu_ref *percpu_ref) +{ + struct kmem_cache *s = container_of(percpu_ref, struct kmem_cache, + memcg_params.refcnt); + +spin_lock (&memcg_kmem_wq_lock);
This code may be called from irq context AFAIU so you should use irq-safe primitive.
mm/slab_common.c static void kmemcg_schedule_work_after_rcu(struct rcu_head *head) · line 756
Quoted 6 lines
mm/slab_common.c +6
-0+if (s->memcg_params.root_cache->memcg_params.dying) + goto unlock; + +WARN_ON (s->memcg_params.work_fn); + s->memcg_params.work_fn = kmemcg_cache_shutdown_after_rcu; +call_rcu (&s->memcg_params.rcu_head, kmemcg_schedule_work_after_rcu);
I may be totally wrong here, but I have a suspicion we don't really need rcu here.
As I see it, you add this code so as to prevent memcg_kmem_get_cache from dereferencing a destroyed kmem cache. Can't we continue using css_tryget_online for that? I mean, take rcu_read_lock() and try to get css reference. If you succeed, then the cgroup must be online, and css_offline won't be called until you unlock rcu, right? This means that the cache is guaranteed to be alive until then, because the cgroup holds a reference to all its kmem caches until it's taken offline.
Quoted 14 lines
mm/slab_common.c +5
-0+unlock: +spin_unlock (&memcg_kmem_wq_lock); +} +static void kmemcg_cache_deactivate_after_rcu (struct kmem_cache *s) {__kmemcg_cache_deactivate_after_rcu (s); +percpu_ref_kill (&s->memcg_params.refcnt); }static void kmemcg_cache_deactivate (struct kmem_cache *s)@@ -854,8 +861,15 @@ static int shutdown_memcg_caches(struct kmem_cache *s)static void flush_memcg_workqueue (struct kmem_cache *s)
mm/slab_common.c static int shutdown_memcg_caches(struct kmem_cache *s) · line 872
Quoted 10 lines
mm/slab_common.c +7
-0{ +/* +* memcg_params.dying is synchronized using slab_mutex AND +* memcg_kmem_wq_lock spinlock, because it's not always +* possible to grab slab_mutex. +*/ mutex_lock (&slab_mutex); +spin_lock (&memcg_kmem_wq_lock); s->memcg_params.dying = true; +spin_unlock (&memcg_kmem_wq_lock);
I would completely switch from the mutex to the new spin lock - acquiring them both looks weird.
Quoted 3 lines
mm/slab_common.c +0
-0mutex_unlock (&slab_mutex);/*
Re: [PATCH v5 5/7] mm: rework non-root kmem_cache lifecycle management
Quoted 5 lines · Vladimir Davydov
Quoted 5 lines
static void
flush_memcg_workqueue(struct kmem_cache *s){mm/slab_common.c +3
-0+/* +* memcg_params.dying is synchronized using slab_mutex AND +* memcg_kmem_wq_lock spinlock, because it's not always
Quoted 12 lines
mm/slab_common.c static int shutdown_memcg_caches(struct kmem_cache *s) · line 872
Quoted 6 lines
mm/slab_common.c +4
-0+* possible to grab slab_mutex. +*/ mutex_lock (&slab_mutex); +spin_lock (&memcg_kmem_wq_lock); s->memcg_params.dying = true; +spin_unlock (&memcg_kmem_wq_lock);I would completely switch from the mutex to the new spin lock - acquiring them both looks weird.
Quoted 3 lines
mm/slab_common.c +0
-0mutex_unlock (&slab_mutex);/*
There are places where the slab_mutex is held and sleeping functions like kvzalloc() are called. I understand that taking both mutex and spinlocks look ugly, but converting all the slab_mutex critical sections to spinlock critical sections will be a major undertaking by itself. So I would suggest leaving that for now.
Cheers,
Longman
Re: [PATCH v5 5/7] mm: rework non-root kmem_cache lifecycle management
Quoted 17 lines · Waiman Long
Quoted 16 lines · Vladimir Davydov
mm/slab_common.c static int shutdown_memcg_caches(struct kmem_cache *s) · line 872
Quoted 11 lines
static void
flush_memcg_workqueue(struct kmem_cache *s){mm/slab_common.c +7
-0+/* +* memcg_params.dying is synchronized using slab_mutex AND +* memcg_kmem_wq_lock spinlock, because it's not always +* possible to grab slab_mutex. +*/ mutex_lock (&slab_mutex); +spin_lock (&memcg_kmem_wq_lock); s->memcg_params.dying = true; +spin_unlock (&memcg_kmem_wq_lock);I would completely switch from the mutex to the new spin lock - acquiring them both looks weird.
Quoted 2 lines
mutex_unlock(&slab_mutex);
Quoted 7 lines
Quoted 1 line
Quoted 1 line
/*
There are places where the
slab_mutexis held and sleeping functions likekvzalloc()are called. I understand that taking both mutex and spinlocks look ugly, but converting all theslab_mutexcritical sections to spinlock critical sections will be a major undertaking by itself. So I would suggest leaving that for now.
I didn't mean that. I meant taking spin_lock wherever we need to access the 'dying' flag, even if slab_mutex is held. So that we don't need to take mutex_lock in flush_memcg_workqueue, where it's used solely for 'dying' synchronization.
Re: [PATCH v5 5/7] mm: rework non-root kmem_cache lifecycle management
Quoted 24 lines · Vladimir Davydov
Quoted 15 lines · Waiman Long
Quoted 14 lines · Vladimir Davydov
mm/slab_common.c static int shutdown_memcg_caches(struct kmem_cache *s) · line 872
Quoted 11 lines
static void
flush_memcg_workqueue(struct kmem_cache *s){mm/slab_common.c +7
-0+/* +* memcg_params.dying is synchronized using slab_mutex AND +* memcg_kmem_wq_lock spinlock, because it's not always +* possible to grab slab_mutex. +*/ mutex_lock (&slab_mutex); +spin_lock (&memcg_kmem_wq_lock); s->memcg_params.dying = true; +spin_unlock (&memcg_kmem_wq_lock);I would completely switch from the mutex to the new spin lock - acquiring them both looks weird.
Quoted 8 lines
Quoted 3 lines
mm/slab_common.c static int shutdown_memcg_caches(struct kmem_cache *s) · line 875
Quoted 3 lines
mm/slab_common.c +0
-0mutex_unlock (&slab_mutex);/* There are places where the
slab_mutexis held and sleeping functions likekvzalloc()are called. I understand that taking both mutex and spinlocks look ugly, but converting all theslab_mutexcritical sections to spinlock critical sections will be a major undertaking by itself. So I would suggest leaving that for now.
Quoted 4 lines
I didn't mean that. I meant taking
spin_lockwherever we need to access the 'dying' flag, even ifslab_mutexis held. So that we don't need to takemutex_lockinflush_memcg_workqueue, where it's used solely for 'dying' synchronization.
OK, that makes sense. Thanks for the clarification.
Cheers,
Longman
Re: [PATCH v5 5/7] mm: rework non-root kmem_cache lifecycle management
In reply to Vladimir DavydovQuoted 26 lines · Vladimir Davydov
Hello Roman,
Quoted 13 lines · Roman Gushchin
This commit makes several important changes in the lifecycle of a non-root
kmem_cache, which also affect the lifecycle of a memory cgroup.Currently each charged slab page has a
page->mem_cgrouppointer to the memory cgroup and holds a reference to it.Kmem_cachesare held by the memcg and are released with it. It means that none ofkmem_cachesare released unless at least one reference to the memcg exists, which is not optimal.So the current scheme can be illustrated as:
page->mem_cgroup->kmem_cache.Quoted 6 lines
To implement the slab memory reparenting we need to invert the scheme
into:page->kmem_cache->mem_cgroup.Let's make every page to hold a reference to the
kmem_cache(we already have a stable pointer), and makekmem_cachesto hold a single reference to the memory cgroup.Is there any reason why we can't reference both mem cgroup and kmem cache per each charged kmem page? I mean,
Quoted 8 lines
page->mem_cgroupreferencesmem_cgrouppage->kmem_cachereferenceskmem_cachemem_cgroupreferenceskmem_cachewhile it's onlineTBO it seems to me that not taking a reference to mem cgroup per charged kmem page makes the code look less straightforward, e.g. as you mentioned in the commit log, we have to use
mod_lruvec_state()for memcg pages andmod_lruvec_page_state()for root pages.
I think I completely missed the point here. In the following patch you move kmem caches from a child to the parent cgroup on offline (aka reparent them). That's why you can't maintain page->mem_cgroup. Sorry for misunderstanding.
Johannes Weiner
maintainerRe: [PATCH v5 5/7] mm: rework non-root kmem_cache lifecycle management
In reply to Roman Gushchinmm/memcontrol.c struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep) · line 2683
Quoted 8 lines · Roman Gushchin
mm/memcontrol.c +8
-0+ arr =rcu_dereference (cachep->memcg_params.memcg_caches); + +/* +* Make sure we will access the up-to-date value. The code updating +* memcg_caches issues a write barrier to match this (see +* memcg_create_kmem_cache()). +*/ + memcg_cachep =READ_ONCE (arr->entries[kmemcg_id]);
READ_ONCE() isn't an SMP barrier, it just prevents compiler muckery. This needs an explicit smp_rmb() to pair with the smp_wmb() on the other side.
I realize you're only moving this code, but it would be good to fix that up while you're there.
Re: [PATCH v5 5/7] mm: rework non-root kmem_cache lifecycle management
Quoted 13 lines · Johannes Weiner
mm/memcontrol.c struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep) · line 2683
Quoted 8 lines · Roman Gushchin
mm/memcontrol.c +8
-0+ arr =rcu_dereference (cachep->memcg_params.memcg_caches); + +/* +* Make sure we will access the up-to-date value. The code updating +* memcg_caches issues a write barrier to match this (see +* memcg_create_kmem_cache()). +*/ + memcg_cachep =READ_ONCE (arr->entries[kmemcg_id]);
READ_ONCE()isn't an SMP barrier, it just prevents compiler muckery. This needs an explicitsmp_rmb()to pair with thesmp_wmb()on the other side.
I believe rcu_dereference()/rcu_assign_pointer()/... are better replacements.
Quoted 3 lines
I realize you're only moving this code, but it would be good to fix that up while you're there.
Right. I'll try to fix it with new-ish rcu API in a separate patch preceding this one.
Thank you for looking into the series!
PATCH v5 6/7 mm: reparent slab memory on cgroup removal
In reply to Roman GushchinLet's reparent memcg slab memory on memcg offlining. This allows us to release the memory cgroup without waiting for the last outstanding kernel object (e.g. dentry used by another application).
So instead of reparenting all accounted slab pages, let's do reparent a relatively small amount of kmem_caches. Reparenting is performed as a part of the deactivation process.
Since the parent cgroup is already charged, everything we need to do is to splice the list of kmem_caches to the parent's kmem_caches list, swap the memcg pointer and drop the css refcounter for each kmem_cache and adjust the parent's css refcounter. Quite simple.
Please, note that kmem_cache->memcg_params.memcg isn't a stable pointer anymore. It's safe to read it under rcu_read_lock() or with slab_mutex held.
We can race with the slab allocation and deallocation paths. It's not a big problem: parent's charge and slab global stats are always correct, and we don't care anymore about the child usage and global stats. The child cgroup is already offline, so we don't use or show it anywhere.
Local slab stats (NR_SLAB_RECLAIMABLE and NR_SLAB_UNRECLAIMABLE) aren't used anywhere except count_shadow_nodes(). But even there it won't break anything: after reparenting "nodes" will be 0 on child level (because we're already reparenting shrinker lists), and on parent level page stats always were 0, and this patch won't change anything.
--- include/linux/slab.h | 4 ++-- mm/memcontrol.c | 14 ++++++++------ mm/slab.h | 21 ++++++++++++++++----- mm/slab_common.c | 21 ++++++++++++++++++--- 4 files changed, 44 insertions(+), 16 deletions(-)
include/linux/slab.h +2 -2
@@ -152,7 +152,7 @@ void kmem_cache_destroy(struct kmem_cache *);int kmem_cache_shrink (struct kmem_cache *);void memcg_create_kmem_cache (struct mem_cgroup *,struct kmem_cache *); -void memcg_deactivate_kmem_caches (struct mem_cgroup *); +void memcg_deactivate_kmem_caches (struct mem_cgroup *,struct mem_cgroup *);/* * Please use this macro to create slab caches. Simply specify the
@@ -638,7 +638,7 @@ struct memcg_cache_params {bool dying; };struct { -struct mem_cgroup *memcg; +struct mem_cgroup __rcu *memcg;struct list_head children_node;struct list_head kmem_caches_node;struct percpu_ref refcnt;
mm/memcontrol.c +8 -6
@@ -3224,15 +3224,15 @@ static void memcg_offline_kmem(struct mem_cgroup *memcg)*/ memcg->kmem_state =KMEM_ALLOCATED ; -memcg_deactivate_kmem_caches (memcg); - - kmemcg_id = memcg->kmemcg_id; -BUG_ON (kmemcg_id <0 ); - parent =parent_mem_cgroup (memcg);if (!parent) parent = root_mem_cgroup; +memcg_deactivate_kmem_caches (memcg, parent); + + kmemcg_id = memcg->kmemcg_id; +BUG_ON (kmemcg_id <0 ); +/* * Change kmemcg_id of this cgroup and all its descendants to the * parent's id, and then move all entries from this cgroup's list_lrus
@@ -3265,7 +3265,6 @@ static void memcg_free_kmem(struct mem_cgroup *memcg)if (memcg->kmem_state ==KMEM_ALLOCATED ) {WARN_ON (!list_empty (&memcg->kmem_caches));static_branch_dec (&memcg_kmem_enabled_key); -WARN_ON (page_counter_read (&memcg->kmem)); } }#else
@@ -4677,6 +4676,9 @@ mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)/* The following stuff does not apply to the root */ if (!parent) { +#ifdef CONFIG_MEMCG_KMEM +INIT_LIST_HEAD (&memcg->kmem_caches); +#endif root_mem_cgroup = memcg;return &memcg->css; }
mm/slab.h +16 -5
@@ -268,10 +268,18 @@ static __always_inline int memcg_charge_slab(struct page *page,struct lruvec *lruvec;int ret; - memcg = s->memcg_params.memcg; +rcu_read_lock (); + memcg =rcu_dereference (s->memcg_params.memcg); +while (memcg && !css_tryget_online (&memcg->css)) + memcg =parent_mem_cgroup (memcg); +rcu_read_unlock (); + +if (unlikely (!memcg)) +return true; + ret =memcg_kmem_charge_memcg (page, gfp, order, memcg);if (ret) -return ret; + goto out; lruvec =mem_cgroup_lruvec (page_pgdat (page), memcg);mod_lruvec_state (lruvec,cache_vmstat_idx (s),1 << order);
@@ -279,8 +287,9 @@ static __always_inline int memcg_charge_slab(struct page *page,/* transer try_charge() page references to kmem_cache */ percpu_ref_get_many (&s->memcg_params.refcnt,1 << order);css_put_many (&memcg->css,1 << order); - -return 0 ; +out: +css_put (&memcg->css); +return ret; }/*
@@ -293,10 +302,12 @@ static __always_inline void memcg_uncharge_slab(struct page *page, int order,struct mem_cgroup *memcg;struct lruvec *lruvec; - memcg = s->memcg_params.memcg; +rcu_read_lock (); + memcg =rcu_dereference (s->memcg_params.memcg); lruvec =mem_cgroup_lruvec (page_pgdat (page), memcg);mod_lruvec_state (lruvec,cache_vmstat_idx (s), -(1 << order));memcg_kmem_uncharge_memcg (page, order, memcg); +rcu_read_unlock ();percpu_ref_put_many (&s->memcg_params.refcnt,1 << order); }
mm/slab_common.c +18 -3
@@ -237,7 +237,7 @@ void memcg_link_cache(struct kmem_cache *s, struct mem_cgroup *memcg)list_add (&s->root_caches_node, &slab_root_caches); }else {css_get (&memcg->css); - s->memcg_params.memcg = memcg; +rcu_assign_pointer (s->memcg_params.memcg, memcg);list_add (&s->memcg_params.children_node, &s->memcg_params.root_cache->memcg_params.children); list_add(&s->memcg_params.kmem_caches_node,
@@ -252,7 +252,8 @@ static void memcg_unlink_cache(struct kmem_cache *s) } else {list_del (&s->memcg_params.children_node);list_del (&s->memcg_params.kmem_caches_node); -css_put (&s->memcg_params.memcg->css); +mem_cgroup_put (rcu_dereference_protected (s->memcg_params.memcg, +lockdep_is_held (&slab_mutex))); } }#else
@@ -776,11 +777,13 @@ static void kmemcg_cache_deactivate(struct kmem_cache *s)call_rcu (&s->memcg_params.rcu_head, kmemcg_schedule_work_after_rcu); } -void memcg_deactivate_kmem_caches (struct mem_cgroup *memcg) +void memcg_deactivate_kmem_caches (struct mem_cgroup *memcg, +struct mem_cgroup *parent) {int idx;struct memcg_cache_array *arr;struct kmem_cache *s, *c; +unsigned int nr_reparented; idx =memcg_cache_id (memcg);
@@ -798,6 +801,18 @@ void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)kmemcg_cache_deactivate (c); arr->entries[idx] =NULL ; } + nr_reparented =0 ; +list_for_each_entry (s, &memcg->kmem_caches, + memcg_params.kmem_caches_node) { +rcu_assign_pointer (s->memcg_params.memcg, parent); +css_put (&memcg->css); + nr_reparented++; + } +if (nr_reparented) { +list_splice_init (&memcg->kmem_caches, + &parent->kmem_caches); +css_get_many (&parent->css, nr_reparented); + }mutex_unlock (&slab_mutex);put_online_mems ();
Signature · 1 line
2.20.1
Quoted 30 lines · Roman Gushchin
Let's reparent memcg slab memory on memcg offlining. This allows us to release the memory cgroup without waiting for the last outstanding kernel object (e.g. dentry used by another application).
So instead of reparenting all accounted slab pages, let's do reparent a relatively small amount of
kmem_caches. Reparenting is performed as a part of the deactivation process.Since the parent cgroup is already charged, everything we need to do is to splice the list of
kmem_cachesto the parent'skmem_cacheslist, swap the memcg pointer and drop the css refcounter for eachkmem_cacheand adjust the parent's css refcounter. Quite simple.Please, note that
kmem_cache->memcg_params.memcgisn't a stable pointer anymore. It's safe to read it underrcu_read_lock()or withslab_mutexheld.We can race with the slab allocation and deallocation paths. It's not a big problem: parent's charge and slab global stats are always correct, and we don't care anymore about the child usage and global stats. The child cgroup is already offline, so we don't use or show it anywhere.
Local slab stats (
NR_SLAB_RECLAIMABLEandNR_SLAB_UNRECLAIMABLE) aren't used anywhere exceptcount_shadow_nodes(). But even there it won't break anything: after reparenting "nodes" will be 0 on child level (because we're already reparenting shrinker lists), and on parent level page stats always were 0, and this patch won't change anything.
Quoted 2 lines
Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>
This one looks good to me. I can't see why anything could possibly go wrong after this change.
Quoted 31 lines · Vladimir Davydov
Quoted 30 lines · Roman Gushchin
Let's reparent memcg slab memory on memcg offlining. This allows us to release the memory cgroup without waiting for the last outstanding kernel object (e.g. dentry used by another application).
So instead of reparenting all accounted slab pages, let's do reparent a relatively small amount of
kmem_caches. Reparenting is performed as a part of the deactivation process.Since the parent cgroup is already charged, everything we need to do is to splice the list of
kmem_cachesto the parent'skmem_cacheslist, swap the memcg pointer and drop the css refcounter for eachkmem_cacheand adjust the parent's css refcounter. Quite simple.Please, note that
kmem_cache->memcg_params.memcgisn't a stable pointer anymore. It's safe to read it underrcu_read_lock()or withslab_mutexheld.We can race with the slab allocation and deallocation paths. It's not a big problem: parent's charge and slab global stats are always correct, and we don't care anymore about the child usage and global stats. The child cgroup is already offline, so we don't use or show it anywhere.
Local slab stats (
NR_SLAB_RECLAIMABLEandNR_SLAB_UNRECLAIMABLE) aren't used anywhere exceptcount_shadow_nodes(). But even there it won't break anything: after reparenting "nodes" will be 0 on child level (because we're already reparenting shrinker lists), and on parent level page stats always were 0, and this patch won't change anything.
Quoted 5 lines
Quoted 2 lines
Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>This one looks good to me. I can't see why anything could possibly go wrong after this change.
Hi Vladimir!
Thank you for looking into the series. Really appreciate it!
It looks like outstanding questions are:
1) synchronization around the dying flag
2) removing CONFIG_SLOB in 2/7
3) early sysfs_slab_remove()
4) mem_cgroup_from_kmem in 7/7
Please, let me know if I missed anything.
I'm waiting now for Johanness's review, so I'll address these issues in background and post the next (and hopefully) final version.
Thanks!
Quoted 7 lines · Roman Gushchin
It looks like outstanding questions are:
1) synchronization around the dying flag
2) removingCONFIG_SLOBin 2/7
3) earlysysfs_slab_remove()
4)mem_cgroup_from_kmemin 7/7Please, let me know if I missed anything.
Also, I think that it might be possible to get rid of RCU call in kmem cache destructor, because the cgroup subsystem already handles it and we could probably piggyback - see my comment to 5/7. Not sure if it's really necessary, since we already have RCU in SLUB, but worth looking into, I guess, as it might simplify the code a bit.
Quoted 7 lines · Vladimir Davydov
Quoted 6 lines · Roman Gushchin
It looks like outstanding questions are:
1) synchronization around the dying flag
2) removingCONFIG_SLOBin 2/7
3) earlysysfs_slab_remove()
4)mem_cgroup_from_kmemin 7/7
Quoted 7 lines
Quoted 1 line
Please, let me know if I missed anything.
Also, I think that it might be possible to get rid of RCU call in kmem cache destructor, because the cgroup subsystem already handles it and we could probably piggyback - see my comment to 5/7. Not sure if it's really necessary, since we already have RCU in SLUB, but worth looking into, I guess, as it might simplify the code a bit.
Added to the list. Thank you!
Johannes Weiner
maintainerQuoted 48 lines · Roman Gushchin
Quoted 31 lines · Vladimir Davydov
Quoted 30 lines · Roman Gushchin
Let's reparent memcg slab memory on memcg offlining. This allows us to release the memory cgroup without waiting for the last outstanding kernel object (e.g. dentry used by another application).
So instead of reparenting all accounted slab pages, let's do reparent a relatively small amount of
kmem_caches. Reparenting is performed as a part of the deactivation process.Since the parent cgroup is already charged, everything we need to do is to splice the list of
kmem_cachesto the parent'skmem_cacheslist, swap the memcg pointer and drop the css refcounter for eachkmem_cacheand adjust the parent's css refcounter. Quite simple.Please, note that
kmem_cache->memcg_params.memcgisn't a stable pointer anymore. It's safe to read it underrcu_read_lock()or withslab_mutexheld.We can race with the slab allocation and deallocation paths. It's not a big problem: parent's charge and slab global stats are always correct, and we don't care anymore about the child usage and global stats. The child cgroup is already offline, so we don't use or show it anywhere.
Local slab stats (
NR_SLAB_RECLAIMABLEandNR_SLAB_UNRECLAIMABLE) aren't used anywhere exceptcount_shadow_nodes(). But even there it won't break anything: after reparenting "nodes" will be 0 on child level (because we're already reparenting shrinker lists), and on parent level page stats always were 0, and this patch won't change anything.Quoted 5 lines
Quoted 2 lines
Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>This one looks good to me. I can't see why anything could possibly go wrong after this change.
Hi Vladimir!
Thank you for looking into the series. Really appreciate it!
It looks like outstanding questions are:
1) synchronization around the dying flag
2) removingCONFIG_SLOBin 2/7
3) earlysysfs_slab_remove()
4)mem_cgroup_from_kmemin 7/7
Quoted 4 lines
Please, let me know if I missed anything.
I'm waiting now for Johanness's review, so I'll address these issues in background and post the next (and hopefully) final version.
The todo items here aside, the series looks good to me - although I'm glad that Vladimir gave it a much more informed review than I could.
PATCH v5 7/7 mm: fix /proc/kpagecgroup interface for slab pages
In reply to Roman GushchinSwitching to an indirect scheme of getting mem_cgroup pointer for !root slab pages broke /proc/kpagecgroup interface for them.
Let's fix it by learning page_cgroup_ino() how to get memcg pointer for slab pages.
--- mm/memcontrol.c | 5 ++++- mm/slab.h | 25 +++++++++++++++++++++++++ mm/slab_common.c | 1 + 3 files changed, 30 insertions(+), 1 deletion(-)
mm/memcontrol.c +4 -1
@@ -494,7 +494,10 @@ ino_t page_cgroup_ino(struct page *page)unsigned long ino =0 ;rcu_read_lock (); - memcg =READ_ONCE (page->mem_cgroup); +if (PageHead (page) &&PageSlab (page)) + memcg =memcg_from_slab_page (page); +else + memcg =READ_ONCE (page->mem_cgroup);while (memcg && !(memcg->css.flags &CSS_ONLINE )) memcg =parent_mem_cgroup (memcg);if (memcg)
mm/slab.h +25 -0
@@ -256,6 +256,26 @@ static inline struct kmem_cache *memcg_root_cache(struct kmem_cache *s)return s->memcg_params.root_cache; } +/* +* Expects a pointer to a slab page. Please note, that PageSlab() check +* isn't sufficient, as it returns true also for tail compound slab pages, +* which do not have slab_cache pointer set. +* So this function assumes that the page can pass PageHead() and PageSlab() +* checks. +*/ +static inline struct mem_cgroup *memcg_from_slab_page (struct page *page) +{ +struct kmem_cache *s; + +WARN_ON_ONCE (!rcu_read_lock_held ()); + + s =READ_ONCE (page->slab_cache); +if (s && !is_root_cache (s)) +return rcu_dereference (s->memcg_params.memcg); + +return NULL ; +} +/* * Charge the slab page belonging to the non-root kmem_cache. * Can be called for non-root kmem_caches only.
@@ -353,6 +373,11 @@ static inline struct kmem_cache *memcg_root_cache(struct kmem_cache *s)return s; } +static inline struct mem_cgroup *memcg_from_slab_page (struct page *page) +{ +return NULL ; +} +static inline int memcg_charge_slab (struct page *page,gfp_t gfp,int order,struct kmem_cache *s) {
mm/slab_common.c +1 -0
@@ -254,6 +254,7 @@ static void memcg_unlink_cache(struct kmem_cache *s)list_del (&s->memcg_params.kmem_caches_node);mem_cgroup_put (rcu_dereference_protected (s->memcg_params.memcg,lockdep_is_held (&slab_mutex))); + rcu_assign_pointer(s->memcg_params.memcg, NULL); } }#else
Signature · 1 line
2.20.1
Re: [PATCH v5 7/7] mm: fix /proc/kpagecgroup interface for slab pages
Quoted 6 lines · Roman Gushchin
Switching to an indirect scheme of getting
mem_cgrouppointer for !root slab pages broke/proc/kpagecgroupinterface for them.Let's fix it by learning
page_cgroup_ino()how to get memcg pointer for slab pages.
Quoted 8 lines
Reported-by Shakeel Butt <[email protected]>Signed-off-by Roman Gushchin <[email protected]>Reviewed-by Shakeel Butt <[email protected]>--- mm/memcontrol.c | 5 ++++- mm/slab.h | 25 +++++++++++++++++++++++++ mm/slab_common.c | 1 + 3 files changed, 30 insertions(+), 1 deletion(-)
What about mem_cgroup_from_kmem, see mm/list_lru.c? Shouldn't we fix it, too?
Hi, Andrew!
Is this patchset good to go? Or do you have any remaining concerns?
It has been carefully reviewed by Shakeel; and also Christoph and Waiman gave some attention to it.
Since commit 172b06c32b94 ("mm: slowly shrink slabs with a relatively") has been reverted, the memcg "leak" problem is open again, and I've heard from several independent people and companies that it's a real problem for them. So it will be nice to close it asap.
I suspect that the fix is too heavy for stable, unfortunately.
Please, let me know if you have any issues that preventing you from pulling it into the tree.
Thank you!
Quoted 149 lines · Roman Gushchin
# Why do we need this?
We've noticed that the number of dying cgroups is steadily growing on most of our hosts in production. The following investigation revealed an issue in userspace memory reclaim code [1], accounting of kernel stacks [2], and also the mainreason: slab objects.
The underlying problem is quite simple: any page charged
to a cgroup holds a reference to it, so the cgroup can't be reclaimed unless all charged pages are gone. If a slab object is actively used by other cgroups, it won't be reclaimed, and will prevent the origin cgroup from being reclaimed.Slab objects, and first of all vfs cache, is shared between cgroups, which are using the same underlying fs, and what's even more important, it's shared between multiple generations of the same workload. So if something is running periodically every time in a new cgroup (like how systemd works), we do accumulate multiple dying cgroups.
Strictly speaking pagecache isn't different here, but there is a key difference:
we disable protection and apply some extra pressure on LRUs of dying cgroups, and these LRUs contain all charged pages.
My experiments show that with the disabled kernel memory accounting the number of dying cgroups stabilizes at a relatively small number (~100, depends on memory pressure and cgroup creation rate), and with kernel memory accounting it grows pretty steadily up to several thousands.Memory cgroups are quite complex and big objects (mostly due to percpu stats), so it leads to noticeable memory losses. Memory occupied by dying cgroups is measured in hundreds of megabytes. I've even seen a host with more than 100Gb of memory wasted for dying cgroups. It leads to a degradation of performance with the uptime, and generally limits the usage of cgroups.
My previous attempt [3] to fix the problem by applying extra pressure on slab shrinker lists caused a regressions with xfs and ext4, and has been reverted [4]. The following attempts to find the right balance [5, 6] were not successful.
So instead of trying to find a maybe non-existing balance, let's do reparent the accounted slabs to the parent cgroup on cgroup removal.
# Implementation approach
There is however a significant problem with reparenting of slab memory:
there is no list of charged pages. Some of them are in shrinker lists, but not all. Introducing of a new list is really not an option.But fortunately there is a way forward: every slab page has a stable pointer to the corresponding
kmem_cache. So the idea is to reparentkmem_cachesinstead of slab pages.It's actually simpler and cheaper, but requires some underlying changes:
1) Makekmem_cachesto hold a single reference to the memory cgroup, instead of a separate reference per every slab page.
2) Stop settingpage->mem_cgrouppointer for memcg slab pages and usepage->kmem_cache->memcgindirection instead. It's used only on slab page release, so it shouldn't be a big issue.
3) Introduce a refcounter for non-root slab caches. It's required to be able to destroykmem_cacheswhen they become empty and release the associated memory cgroup.There is a bonus: currently we do release empty
kmem_cacheson cgroup removal, however all other are waiting for the releasing of the memory cgroup. These refactorings allowkmem_cachesto be released as soon as they become inactive and free.Some additional implementation details are provided in corresponding commit messages.
# Results
Below is the average number of dying cgroups on two groups of our production hosts. They do run some sort of web frontend workload, the memory pressure is moderate. As we can see, with the kernel memory reparenting the number stabilizes in 60s range; however with the original version it grows almost linearly and doesn't show any signs of plateauing. The difference in slab and percpu usage between patched and unpatched versions also grows linearly. In 7 days it exceeded 200Mb.
day 0 1 2 3 4 5 6 7
original 56 362 628 752 1070 1250 1490 1560
patched 23 46 51 55 60 57 67 69
mem diff(Mb) 22 74 123 152 164 182 214 241# History
v5:
1) fixed a compilation warning around missingkmemcg_queue_cache_shutdown()
2) s/rcu_read_lock()/rcu_read_unlock() inmemcg_kmem_get_cache()v4:
1) removed excessive memcg != parent check inmemcg_deactivate_kmem_caches()
2) fixedrcu_read_lock()usage inmemcg_charge_slab()
3) fixed synchronization around dying flag inkmemcg_queue_cache_shutdown()
4) refreshed test results data
5) reworkedPageTail()checks inmemcg_from_slab_page()
6) added some comments in multiple placesv3:
1) reworked memcgkmem_cachesearch on allocation path
2) fixed/proc/kpagecgroupinterfacev2:
1) switched to percpukmem_cacherefcounter
2) a reference tokmem_cacheis held during the allocation
3) slabs stats are fixed for !MEMCG case (and the refactoring is separated into a standalone patch)
4)kmem_cachereparenting is performed from deactivatation contextv1:
https://lkml.org/lkml/2019/4/17/1095# Links
[1]: commit
68600f623d69("mm: don't miss the last page because of round-off error")
[2]: commit9b6f7e163cd0("mm: rework memcg kernel stack accounting") [3]: commit172b06c32b94("mm: slowly shrink slabs with a relatively small number of objects")
[4]: commita9a238e83fbb("Revert "mm: slowly shrink slabs with a relatively small number of objects")
[5]: https://lkml.org/lkml/2019/1/28/1865
[6]: https://marc.info/?l=linux-mm&m=155064763626437&w=2Roman Gushchin (7):
mm: postponekmem_cachememcg pointer initialization to
memcg_link_cache()
mm: generalize postponed non-rootkmem_cachedeactivation
mm: introduce__memcg_kmem_uncharge_memcg()
mm: unify SLAB and SLUB page accounting
mm: rework non-rootkmem_cachelifecycle management
mm: reparent slab memory on cgroup removal
mm: fix/proc/kpagecgroupinterface for slab pagesinclude/linux/memcontrol.h | 10 +++ include/linux/slab.h | 13 +-- mm/memcontrol.c | 101 ++++++++++++++++------- mm/slab.c | 25 ++---- mm/slab.h | 137 ++++++++++++++++++++++++------- mm/slab_common.c | 162 +++++++++++++++++++++---------------- mm/slub.c | 36 ++------- 7 files changed, 299 insertions(+), 185 deletions(-)Signature · 1 line
2.20.1
Andrew Morton
maintainerQuoted 10 lines · On Wed, 22 May 2019 21:43:54 +0000 Roman Gushchin <[email protected]> wrote:
Is this patchset good to go? Or do you have any remaining concerns?
It has been carefully reviewed by Shakeel; and also Christoph and Waiman gave some attention to it.
Since commit
172b06c32b94("mm: slowly shrink slabs with a relatively") has been reverted, the memcg "leak" problem is open again, and I've heard from several independent people and companies that it's a real problem for them. So it will be nice to close it asap.
Quoted 4 lines
I suspect that the fix is too heavy for stable, unfortunately.
Please, let me know if you have any issues that preventing you from pulling it into the tree.
I looked, and put it on ice for a while, hoping to hear from mhocko/hannes. Did they look at the earlier versions?
Quoted 12 lines · Andrew Morton
Quoted 10 lines · On Wed, 22 May 2019 21:43:54 +0000 Roman Gushchin <[email protected]> wrote:
Is this patchset good to go? Or do you have any remaining concerns?
It has been carefully reviewed by Shakeel; and also Christoph and Waiman gave some attention to it.
Since commit
172b06c32b94("mm: slowly shrink slabs with a relatively") has been reverted, the memcg "leak" problem is open again, and I've heard from several independent people and companies that it's a real problem for them. So it will be nice to close it asap.
Quoted 7 lines
Quoted 4 lines
I suspect that the fix is too heavy for stable, unfortunately.
Please, let me know if you have any issues that preventing you from pulling it into the tree.
I looked, and put it on ice for a while, hoping to hear from mhocko/hannes. Did they look at the earlier versions?
Johannes has definitely looked at one of early versions of the patchset, and one of the outcomes was his own patchset about pushing memcg stats up by the tree, which eliminated the need to deal with memcg stats on kmem_cache reparenting.
The problem and the proposed solution have been discussed on latest LSFMM, and I didn't hear any opposition. So I assume that Michal is at least not against the idea in general. A careful code review is always welcome, of course.
Thanks!
Michal Hocko
maintainerQuoted 26 lines · Roman Gushchin
Quoted 12 lines · Andrew Morton
Quoted 10 lines · On Wed, 22 May 2019 21:43:54 +0000 Roman Gushchin <[email protected]> wrote:
Is this patchset good to go? Or do you have any remaining concerns?
It has been carefully reviewed by Shakeel; and also Christoph and Waiman gave some attention to it.
Since commit
172b06c32b94("mm: slowly shrink slabs with a relatively") has been reverted, the memcg "leak" problem is open again, and I've heard from several independent people and companies that it's a real problem for them. So it will be nice to close it asap.Quoted 7 lines
Quoted 4 lines
I suspect that the fix is too heavy for stable, unfortunately.
Please, let me know if you have any issues that preventing you from pulling it into the tree.
I looked, and put it on ice for a while, hoping to hear from mhocko/hannes. Did they look at the earlier versions?
Johannes has definitely looked at one of early versions of the patchset, and one of the outcomes was his own patchset about pushing memcg stats up by the tree, which eliminated the need to deal with memcg stats on
kmem_cachereparenting.
Quoted 4 lines
The problem and the proposed solution have been discussed on latest LSFMM, and I didn't hear any opposition. So I assume that Michal is at least not against the idea in general. A careful code review is always welcome, of course.
I didn't get to review this properly (ETOOBUSY). This is a tricky area so a careful review is definitely due. I would really appreciate if Vladimir could have a look. I understand he is busy with other stuff but a highlevel review from him would be really helpful.
Signature · 2 lines
Michal Hocko SUSE Labs