| Message ID | 20230927150832.335132-3-roman.gushchin@linux.dev (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm: improve performance of kernel memory accounting | expand |
On Wed, Sep 27, 2023 at 08:08:29AM -0700, Roman Gushchin wrote: > @@ -3001,6 +3001,47 @@ static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg) > return objcg; > } > > +static DEFINE_SPINLOCK(current_objcg_lock); > + > +static struct obj_cgroup *current_objcg_update(struct obj_cgroup *old) > +{ > + struct mem_cgroup *memcg; > + struct obj_cgroup *objcg; > + unsigned long flags; > + > + old = current_objcg_clear_update_flag(old); > + if (old) > + obj_cgroup_put(old); > + > + spin_lock_irqsave(¤t_objcg_lock, flags); > + rcu_read_lock(); > + memcg = mem_cgroup_from_task(current); > + for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) { > + objcg = rcu_dereference(memcg->objcg); > + if (objcg && obj_cgroup_tryget(objcg)) > + break; > + objcg = NULL; > + } > + rcu_read_unlock(); Can this tryget() actually fail when this is called on the current task during fork() and attach()? A cgroup cannot be offlined while there is a task in it. > @@ -6345,6 +6393,22 @@ static void mem_cgroup_move_task(void) > mem_cgroup_clear_mc(); > } > } > + > +#ifdef CONFIG_MEMCG_KMEM > +static void mem_cgroup_fork(struct task_struct *task) > +{ > + task->objcg = (struct obj_cgroup *)0x1; dup_task_struct() will copy this pointer from the old task. Would it be possible to bump the refcount here instead? That would save quite a bit of work during fork().
On Mon, Oct 02, 2023 at 04:12:54PM -0400, Johannes Weiner wrote: > On Wed, Sep 27, 2023 at 08:08:29AM -0700, Roman Gushchin wrote: > > @@ -3001,6 +3001,47 @@ static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg) > > return objcg; > > } > > > > +static DEFINE_SPINLOCK(current_objcg_lock); > > + > > +static struct obj_cgroup *current_objcg_update(struct obj_cgroup *old) > > +{ > > + struct mem_cgroup *memcg; > > + struct obj_cgroup *objcg; > > + unsigned long flags; > > + > > + old = current_objcg_clear_update_flag(old); > > + if (old) > > + obj_cgroup_put(old); > > + > > + spin_lock_irqsave(¤t_objcg_lock, flags); > > + rcu_read_lock(); > > + memcg = mem_cgroup_from_task(current); > > + for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) { > > + objcg = rcu_dereference(memcg->objcg); > > + if (objcg && obj_cgroup_tryget(objcg)) > > + break; > > + objcg = NULL; > > + } > > + rcu_read_unlock(); > > Can this tryget() actually fail when this is called on the current > task during fork() and attach()? A cgroup cannot be offlined while > there is a task in it. Highly theoretically it can if it races against a migration of the current task to another memcg and the previous memcg is getting offlined. I actually might make sense to apply the same approach for memcgs as well (saving a lazily-updating memcg pointer on task_struct). Then it will be possible to ditch this "for" loop. But I need some time to master the code and run benchmarks. Idk if it will make enough difference to justify the change. Btw, this is the rfc version, while there is a newer v1 version, which Andrew already picked for mm-unstable. Both of your comments still apply, just fyi. > > > @@ -6345,6 +6393,22 @@ static void mem_cgroup_move_task(void) > > mem_cgroup_clear_mc(); > > } > > } > > + > > +#ifdef CONFIG_MEMCG_KMEM > > +static void mem_cgroup_fork(struct task_struct *task) > > +{ > > + task->objcg = (struct obj_cgroup *)0x1; > > dup_task_struct() will copy this pointer from the old task. Would it > be possible to bump the refcount here instead? That would save quite a > bit of work during fork(). Yeah, it should be possible. It won't save a lot, but I agree it makes sense. I'll take a look and will prepare a separate patch for this. Thank you!
On Mon, Oct 02, 2023 at 03:03:48PM -0700, Roman Gushchin wrote: > On Mon, Oct 02, 2023 at 04:12:54PM -0400, Johannes Weiner wrote: > > On Wed, Sep 27, 2023 at 08:08:29AM -0700, Roman Gushchin wrote: > > > @@ -3001,6 +3001,47 @@ static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg) > > > return objcg; > > > } > > > > > > +static DEFINE_SPINLOCK(current_objcg_lock); > > > + > > > +static struct obj_cgroup *current_objcg_update(struct obj_cgroup *old) > > > +{ > > > + struct mem_cgroup *memcg; > > > + struct obj_cgroup *objcg; > > > + unsigned long flags; > > > + > > > + old = current_objcg_clear_update_flag(old); > > > + if (old) > > > + obj_cgroup_put(old); > > > + > > > + spin_lock_irqsave(¤t_objcg_lock, flags); > > > + rcu_read_lock(); > > > + memcg = mem_cgroup_from_task(current); > > > + for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) { > > > + objcg = rcu_dereference(memcg->objcg); > > > + if (objcg && obj_cgroup_tryget(objcg)) > > > + break; > > > + objcg = NULL; > > > + } > > > + rcu_read_unlock(); > > > > Can this tryget() actually fail when this is called on the current > > task during fork() and attach()? A cgroup cannot be offlined while > > there is a task in it. > > Highly theoretically it can if it races against a migration of the current > task to another memcg and the previous memcg is getting offlined. Ah right, if this runs between css_set_move_task() and ->attach(). The cache would be briefly updated to a parent in the old hierarchy, but then quickly reset from the ->attach(). Can you please add a comment along these lines? > I actually might make sense to apply the same approach for memcgs as well > (saving a lazily-updating memcg pointer on task_struct). Then it will be > possible to ditch this "for" loop. But I need some time to master the code > and run benchmarks. Idk if it will make enough difference to justify the change. Yeah the memcg pointer is slightly less attractive from an optimization POV because it already is a pretty direct pointer from task through the cset array. If you still want to look into it from a simplification POV that sounds reasonable, but IMO it would be fine with a comment. > > > @@ -6345,6 +6393,22 @@ static void mem_cgroup_move_task(void) > > > mem_cgroup_clear_mc(); > > > } > > > } > > > + > > > +#ifdef CONFIG_MEMCG_KMEM > > > +static void mem_cgroup_fork(struct task_struct *task) > > > +{ > > > + task->objcg = (struct obj_cgroup *)0x1; > > > > dup_task_struct() will copy this pointer from the old task. Would it > > be possible to bump the refcount here instead? That would save quite a > > bit of work during fork(). > > Yeah, it should be possible. It won't save a lot, but I agree it makes > sense. I'll take a look and will prepare a separate patch for this. I guess the hairiest part would be synchronizing against a migration because all these cgroup core callbacks are unlocked. Would it make sense to add ->fork_locked() and ->attach_locked() callbacks that are dispatched under the css_set_lock? Then this could be a simple if (p && !(p & 0x1)) obj_cgroup_get(), which would certainly be nice to workloads where fork() is hot, with little downside otherwise.
On Tue, Oct 03, 2023 at 10:22:55AM -0400, Johannes Weiner wrote: > On Mon, Oct 02, 2023 at 03:03:48PM -0700, Roman Gushchin wrote: > > On Mon, Oct 02, 2023 at 04:12:54PM -0400, Johannes Weiner wrote: > > > On Wed, Sep 27, 2023 at 08:08:29AM -0700, Roman Gushchin wrote: > > > > @@ -3001,6 +3001,47 @@ static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg) > > > > return objcg; > > > > } > > > > > > > > +static DEFINE_SPINLOCK(current_objcg_lock); > > > > + > > > > +static struct obj_cgroup *current_objcg_update(struct obj_cgroup *old) > > > > +{ > > > > + struct mem_cgroup *memcg; > > > > + struct obj_cgroup *objcg; > > > > + unsigned long flags; > > > > + > > > > + old = current_objcg_clear_update_flag(old); > > > > + if (old) > > > > + obj_cgroup_put(old); > > > > + > > > > + spin_lock_irqsave(¤t_objcg_lock, flags); > > > > + rcu_read_lock(); > > > > + memcg = mem_cgroup_from_task(current); > > > > + for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) { > > > > + objcg = rcu_dereference(memcg->objcg); > > > > + if (objcg && obj_cgroup_tryget(objcg)) > > > > + break; > > > > + objcg = NULL; > > > > + } > > > > + rcu_read_unlock(); > > > > > > Can this tryget() actually fail when this is called on the current > > > task during fork() and attach()? A cgroup cannot be offlined while > > > there is a task in it. > > > > Highly theoretically it can if it races against a migration of the current > > task to another memcg and the previous memcg is getting offlined. > > Ah right, if this runs between css_set_move_task() and ->attach(). The > cache would be briefly updated to a parent in the old hierarchy, but > then quickly reset from the ->attach(). Even simpler: rcu_read_lock(); memcg = mem_cgroup_from_task(current); --------- Here the task can be moved to another memcg and the previous one can be offlined, making objcg fully detached. --------- for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) { objcg = rcu_dereference(memcg->objcg); if (objcg && obj_cgroup_tryget(objcg)) --------- Objcg can be NULL here or it can be not NULL, but loose the last reference between the objcg check and obj_cgroup_tryget(). --------- break; objcg = NULL; } rcu_read_unlock(); > > Can you please add a comment along these lines? Sure, will do. > > > I actually might make sense to apply the same approach for memcgs as well > > (saving a lazily-updating memcg pointer on task_struct). Then it will be > > possible to ditch this "for" loop. But I need some time to master the code > > and run benchmarks. Idk if it will make enough difference to justify the change. > > Yeah the memcg pointer is slightly less attractive from an > optimization POV because it already is a pretty direct pointer from > task through the cset array. > > If you still want to look into it from a simplification POV that > sounds reasonable, but IMO it would be fine with a comment. I'll come back with some numbers, hard to speculate without it. In this case the majority of savings came from not bumping and decreasing a percpu objcg refcounter on the slab allocation path - that was quite surprising to me. > > > > > @@ -6345,6 +6393,22 @@ static void mem_cgroup_move_task(void) > > > > mem_cgroup_clear_mc(); > > > > } > > > > } > > > > + > > > > +#ifdef CONFIG_MEMCG_KMEM > > > > +static void mem_cgroup_fork(struct task_struct *task) > > > > +{ > > > > + task->objcg = (struct obj_cgroup *)0x1; > > > > > > dup_task_struct() will copy this pointer from the old task. Would it > > > be possible to bump the refcount here instead? That would save quite a > > > bit of work during fork(). > > > > Yeah, it should be possible. It won't save a lot, but I agree it makes > > sense. I'll take a look and will prepare a separate patch for this. > > I guess the hairiest part would be synchronizing against a migration > because all these cgroup core callbacks are unlocked. Yep. > > Would it make sense to add ->fork_locked() and ->attach_locked() > callbacks that are dispatched under the css_set_lock? Then this could > be a simple if (p && !(p & 0x1)) obj_cgroup_get(), which would > certainly be nice to workloads where fork() is hot, with little > downside otherwise. Maybe, but then the question is if it really worth it. In the final version the update path doesn't need a spinlock, so it's quite cheap and happens once on the first allocation, so Idk if it's worth it at all, but I'll take a look. I think the bigger question I have here (and probably worth a lsfmmbpf/plumbers discussion) - what if we introduce a cgroup mount (or even Kconfig) option to prohibit moving tasks between cgroups and rely solely on fork to enter the right cgroup (a-la namespaces). I start thinking that this is the right path long-term, things will be not only more reliable, but we also can ditch a lot of synchronization and get better performance. Obviously not a small project. Thanks!
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h index ab94ad4597d0..84425bfe4124 100644 --- a/include/linux/memcontrol.h +++ b/include/linux/memcontrol.h @@ -553,6 +553,16 @@ static inline bool folio_memcg_kmem(struct folio *folio) return folio->memcg_data & MEMCG_DATA_KMEM; } +static inline bool current_objcg_needs_update(struct obj_cgroup *objcg) +{ + return (struct obj_cgroup *)((unsigned long)objcg & 0x1); +} + +static inline struct obj_cgroup * +current_objcg_clear_update_flag(struct obj_cgroup *objcg) +{ + return (struct obj_cgroup *)((unsigned long)objcg & ~0x1); +} #else static inline bool folio_memcg_kmem(struct folio *folio) diff --git a/include/linux/sched.h b/include/linux/sched.h index 77f01ac385f7..60de42715b56 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1443,6 +1443,10 @@ struct task_struct { struct mem_cgroup *active_memcg; #endif +#ifdef CONFIG_MEMCG_KMEM + struct obj_cgroup *objcg; +#endif + #ifdef CONFIG_BLK_CGROUP struct gendisk *throttle_disk; #endif diff --git a/mm/memcontrol.c b/mm/memcontrol.c index 16ac2a5838fb..7f33a503d600 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -3001,6 +3001,47 @@ static struct obj_cgroup *__get_obj_cgroup_from_memcg(struct mem_cgroup *memcg) return objcg; } +static DEFINE_SPINLOCK(current_objcg_lock); + +static struct obj_cgroup *current_objcg_update(struct obj_cgroup *old) +{ + struct mem_cgroup *memcg; + struct obj_cgroup *objcg; + unsigned long flags; + + old = current_objcg_clear_update_flag(old); + if (old) + obj_cgroup_put(old); + + spin_lock_irqsave(¤t_objcg_lock, flags); + rcu_read_lock(); + memcg = mem_cgroup_from_task(current); + for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) { + objcg = rcu_dereference(memcg->objcg); + if (objcg && obj_cgroup_tryget(objcg)) + break; + objcg = NULL; + } + rcu_read_unlock(); + + WRITE_ONCE(current->objcg, objcg); + spin_unlock_irqrestore(¤t_objcg_lock, flags); + + return objcg; +} + +static inline void current_objcg_set_needs_update(struct task_struct *task) +{ + struct obj_cgroup *objcg; + unsigned long flags; + + spin_lock_irqsave(¤t_objcg_lock, flags); + objcg = READ_ONCE(task->objcg); + objcg = (struct obj_cgroup *)((unsigned long)objcg | 0x1); + WRITE_ONCE(task->objcg, objcg); + spin_unlock_irqrestore(¤t_objcg_lock, flags); +} + __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void) { struct mem_cgroup *memcg; @@ -3008,19 +3049,26 @@ __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void) if (in_task()) { memcg = current->active_memcg; + if (unlikely(memcg)) + goto from_memcg; - /* Memcg to charge can't be determined. */ - if (likely(!memcg) && (!current->mm || (current->flags & PF_KTHREAD))) - return NULL; + objcg = READ_ONCE(current->objcg); + if (unlikely(current_objcg_needs_update(objcg))) + objcg = current_objcg_update(objcg); + + if (objcg) { + obj_cgroup_get(objcg); + return objcg; + } } else { memcg = this_cpu_read(int_active_memcg); - if (likely(!memcg)) - return NULL; + if (unlikely(memcg)) + goto from_memcg; } + return NULL; +from_memcg: rcu_read_lock(); - if (!memcg) - memcg = mem_cgroup_from_task(current); objcg = __get_obj_cgroup_from_memcg(memcg); rcu_read_unlock(); return objcg; @@ -6345,6 +6393,22 @@ static void mem_cgroup_move_task(void) mem_cgroup_clear_mc(); } } + +#ifdef CONFIG_MEMCG_KMEM +static void mem_cgroup_fork(struct task_struct *task) +{ + task->objcg = (struct obj_cgroup *)0x1; +} + +static void mem_cgroup_exit(struct task_struct *task) +{ + struct obj_cgroup *objcg = current_objcg_clear_update_flag(task->objcg); + + if (objcg) + obj_cgroup_put(objcg); +} +#endif + #else /* !CONFIG_MMU */ static int mem_cgroup_can_attach(struct cgroup_taskset *tset) { @@ -6359,7 +6423,7 @@ static void mem_cgroup_move_task(void) #endif #ifdef CONFIG_LRU_GEN -static void mem_cgroup_attach(struct cgroup_taskset *tset) +static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset) { struct task_struct *task; struct cgroup_subsys_state *css; @@ -6377,10 +6441,29 @@ static void mem_cgroup_attach(struct cgroup_taskset *tset) task_unlock(task); } #else +static void mem_cgroup_lru_gen_attach(struct cgroup_taskset *tset) {} +#endif /* CONFIG_LRU_GEN */ + +#ifdef CONFIG_MEMCG_KMEM +static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset) +{ + struct task_struct *task; + struct cgroup_subsys_state *css; + + cgroup_taskset_for_each(task, css, tset) + current_objcg_set_needs_update(task); +} +#else +static void mem_cgroup_kmem_attach(struct cgroup_taskset *tset) {} +#endif /* CONFIG_MEMCG_KMEM */ + +#if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM) static void mem_cgroup_attach(struct cgroup_taskset *tset) { + mem_cgroup_lru_gen_attach(tset); + mem_cgroup_kmem_attach(tset); } -#endif /* CONFIG_LRU_GEN */ +#endif static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value) { @@ -6824,9 +6907,15 @@ struct cgroup_subsys memory_cgrp_subsys = { .css_reset = mem_cgroup_css_reset, .css_rstat_flush = mem_cgroup_css_rstat_flush, .can_attach = mem_cgroup_can_attach, +#if defined(CONFIG_LRU_GEN) || defined(CONFIG_MEMCG_KMEM) .attach = mem_cgroup_attach, +#endif .cancel_attach = mem_cgroup_cancel_attach, .post_attach = mem_cgroup_move_task, +#ifdef CONFIG_MEMCG_KMEM + .fork = mem_cgroup_fork, + .exit = mem_cgroup_exit, +#endif .dfl_cftypes = memory_files, .legacy_cftypes = mem_cgroup_legacy_files, .early_init = 0,
To charge a freshly allocated kernel object to a memory cgroup, the kernel needs to obtain an objcg pointer. Currently it does it indirectly by obtaining the memcg pointer first and then calling to __get_obj_cgroup_from_memcg(). Usually tasks spend their entire life belonging to the same object cgroup. So it makes sense to save the objcg pointer on task_struct directly, so it can be obtained faster. It requires some work on fork, exit and cgroup migrate paths, but these paths are way colder. To avoid any costly synchronization the following rules are applied: 1) A task sets it's objcg pointer itself. 2) If a task is being migrated to another cgroup, the least significant bit of the objcg pointer is set. 3) On the allocation path the objcg pointer is obtained locklessly using the READ_ONCE() macro and the least significant bit is checked. If it set, the task updates it's objcg before proceeding with an allocation. 4) Operations 1) and 4) are synchronized via a new spinlock, so that if a task is moved twice, the update bit can't be lost. This allows to keep the hot path fully lockless. Because the task is keeping a reference to the objcg, it can't go away while the task is alive. This commit doesn't change the way the remote memcg charging works. Signed-off-by: Roman Gushchin (Cruise) <roman.gushchin@linux.dev> --- include/linux/memcontrol.h | 10 ++++ include/linux/sched.h | 4 ++ mm/memcontrol.c | 107 +++++++++++++++++++++++++++++++++---- 3 files changed, 112 insertions(+), 9 deletions(-)