| Message ID | 20211210025228.158196-1-longman@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [PATCH-next,v2] mm/memcg: Properly handle memcg_stock access for PREEMPT_RT | expand |
On 2021-12-09 21:52:28 [-0500], Waiman Long wrote: … > --- a/mm/memcontrol.c > +++ b/mm/memcontrol.c … > @@ -2210,7 +2211,7 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) > struct memcg_stock_pcp *stock; > unsigned long flags; > > - local_irq_save(flags); > + local_lock_irqsave(&memcg_stock.lock, flags); Why is this one using the lock? It isn't accessing irq_obj, right? > stock = this_cpu_ptr(&memcg_stock); > if (stock->cached != memcg) { /* reset if necessary */ > @@ -2779,29 +2780,28 @@ static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg) > * which is cheap in non-preempt kernel. The interrupt context object stock > * can only be accessed after disabling interrupt. User context code can > * access interrupt object stock, but not vice versa. > + * > + * This task and interrupt context optimization is disabled for PREEMPT_RT > + * as there is no performance gain in this case. > */ > static inline struct obj_stock *get_obj_stock(unsigned long *pflags) > { > - struct memcg_stock_pcp *stock; > - > - if (likely(in_task())) { > + if (likely(in_task()) && !IS_ENABLED(CONFIG_PREEMPT_RT)) { > *pflags = 0UL; > preempt_disable(); > - stock = this_cpu_ptr(&memcg_stock); > - return &stock->task_obj; > + return this_cpu_ptr(&memcg_stock.task_obj); > } We usually add the local_lock_t to the object it protects, struct obj_stock it this case. That would give you two different locks (instead of one) so you wouldn't have to use preempt_disable() to avoid lockdep's complains. Also it would warn you if you happen to use that obj_stock in !in_task() which is isn't possible now. The only downside would be that drain_local_stock() needs to acquire two locks. > > - local_irq_save(*pflags); > - stock = this_cpu_ptr(&memcg_stock); > - return &stock->irq_obj; > + local_lock_irqsave(&memcg_stock.lock, *pflags); > + return this_cpu_ptr(&memcg_stock.irq_obj); > } > > static inline void put_obj_stock(unsigned long flags) > { > - if (likely(in_task())) > + if (likely(in_task()) && !IS_ENABLED(CONFIG_PREEMPT_RT)) > preempt_enable(); > else > - local_irq_restore(flags); > + local_unlock_irqrestore(&memcg_stock.lock, flags); > } > > /* Sebastian
On 12/10/21 08:01, Sebastian Andrzej Siewior wrote: > On 2021-12-09 21:52:28 [-0500], Waiman Long wrote: > … >> --- a/mm/memcontrol.c >> +++ b/mm/memcontrol.c > … >> @@ -2210,7 +2211,7 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) >> struct memcg_stock_pcp *stock; >> unsigned long flags; >> >> - local_irq_save(flags); >> + local_lock_irqsave(&memcg_stock.lock, flags); > Why is this one using the lock? It isn't accessing irq_obj, right? Well, the lock isn't just for irq_obj. It protects the whole memcg_stock structure which include irq_obj. Sometimes, data in irq_obj (or task_obj) will get transfer to nr_pages and vice versa. So it is easier to use one single lock for the whole thing. > >> stock = this_cpu_ptr(&memcg_stock); >> if (stock->cached != memcg) { /* reset if necessary */ >> @@ -2779,29 +2780,28 @@ static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg) >> * which is cheap in non-preempt kernel. The interrupt context object stock >> * can only be accessed after disabling interrupt. User context code can >> * access interrupt object stock, but not vice versa. >> + * >> + * This task and interrupt context optimization is disabled for PREEMPT_RT >> + * as there is no performance gain in this case. >> */ >> static inline struct obj_stock *get_obj_stock(unsigned long *pflags) >> { >> - struct memcg_stock_pcp *stock; >> - >> - if (likely(in_task())) { >> + if (likely(in_task()) && !IS_ENABLED(CONFIG_PREEMPT_RT)) { >> *pflags = 0UL; >> preempt_disable(); >> - stock = this_cpu_ptr(&memcg_stock); >> - return &stock->task_obj; >> + return this_cpu_ptr(&memcg_stock.task_obj); >> } > We usually add the local_lock_t to the object it protects, struct > obj_stock it this case. > That would give you two different locks (instead of one) so you wouldn't > have to use preempt_disable() to avoid lockdep's complains. Also it > would warn you if you happen to use that obj_stock in !in_task() which > is isn't possible now. > The only downside would be that drain_local_stock() needs to acquire two > locks. > As said above, having separate locks will complicate the interaction between irq_obj and the broader memcg_stock fields. Besides throughput is a less important matrix for PREEMPT_RT, so I am not trying to optimize throughput performance for PREEMPT_RT here. Cheers, Longman
On 2021-12-10 11:29:31 [-0500], Waiman Long wrote: > > On 12/10/21 08:01, Sebastian Andrzej Siewior wrote: > > On 2021-12-09 21:52:28 [-0500], Waiman Long wrote: > > … > > > --- a/mm/memcontrol.c > > > +++ b/mm/memcontrol.c > > … > > > @@ -2210,7 +2211,7 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) > > > struct memcg_stock_pcp *stock; > > > unsigned long flags; > > > - local_irq_save(flags); > > > + local_lock_irqsave(&memcg_stock.lock, flags); > > Why is this one using the lock? It isn't accessing irq_obj, right? > Well, the lock isn't just for irq_obj. It protects the whole memcg_stock > structure which include irq_obj. Sometimes, data in irq_obj (or task_obj) > will get transfer to nr_pages and vice versa. So it is easier to use one > single lock for the whole thing. This needs way better documentation what protects what any why. I don't like the quick slapping for RT only usage without any kind of explanation. Once you think you know it is irq_obj only you end up here where you have the lock again for no obvious reason. > > Cheers, > Longman Sebastian
On 12/10/21 11:34, Sebastian Andrzej Siewior wrote: > On 2021-12-10 11:29:31 [-0500], Waiman Long wrote: >> On 12/10/21 08:01, Sebastian Andrzej Siewior wrote: >>> On 2021-12-09 21:52:28 [-0500], Waiman Long wrote: >>> … >>>> --- a/mm/memcontrol.c >>>> +++ b/mm/memcontrol.c >>> … >>>> @@ -2210,7 +2211,7 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) >>>> struct memcg_stock_pcp *stock; >>>> unsigned long flags; >>>> - local_irq_save(flags); >>>> + local_lock_irqsave(&memcg_stock.lock, flags); >>> Why is this one using the lock? It isn't accessing irq_obj, right? >> Well, the lock isn't just for irq_obj. It protects the whole memcg_stock >> structure which include irq_obj. Sometimes, data in irq_obj (or task_obj) >> will get transfer to nr_pages and vice versa. So it is easier to use one >> single lock for the whole thing. > This needs way better documentation what protects what any why. > I don't like the quick slapping for RT only usage without any kind of > explanation. Once you think you know it is irq_obj only you end up here > where you have the lock again for no obvious reason. Sure, I will update the patch description and add comments to document that. Let's see what other feedback I have before I make the update. Cheers, Longman
diff --git a/mm/memcontrol.c b/mm/memcontrol.c index a09a7d2e0b1b..8bed8e2993e4 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -2097,6 +2097,7 @@ struct obj_stock { }; struct memcg_stock_pcp { + local_lock_t lock; struct mem_cgroup *cached; /* this never be root cgroup */ unsigned int nr_pages; struct obj_stock task_obj; @@ -2145,7 +2146,7 @@ static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages) if (nr_pages > MEMCG_CHARGE_BATCH) return ret; - local_irq_save(flags); + local_lock_irqsave(&memcg_stock.lock, flags); stock = this_cpu_ptr(&memcg_stock); if (memcg == stock->cached && stock->nr_pages >= nr_pages) { @@ -2153,7 +2154,7 @@ static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages) ret = true; } - local_irq_restore(flags); + local_unlock_irqrestore(&memcg_stock.lock, flags); return ret; } @@ -2189,7 +2190,7 @@ static void drain_local_stock(struct work_struct *dummy) * drain_stock races is that we always operate on local CPU stock * here with IRQ disabled */ - local_irq_save(flags); + local_lock_irqsave(&memcg_stock.lock, flags); stock = this_cpu_ptr(&memcg_stock); drain_obj_stock(&stock->irq_obj); @@ -2198,7 +2199,7 @@ static void drain_local_stock(struct work_struct *dummy) drain_stock(stock); clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags); - local_irq_restore(flags); + local_unlock_irqrestore(&memcg_stock.lock, flags); } /* @@ -2210,7 +2211,7 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) struct memcg_stock_pcp *stock; unsigned long flags; - local_irq_save(flags); + local_lock_irqsave(&memcg_stock.lock, flags); stock = this_cpu_ptr(&memcg_stock); if (stock->cached != memcg) { /* reset if necessary */ @@ -2223,7 +2224,7 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages) if (stock->nr_pages > MEMCG_CHARGE_BATCH) drain_stock(stock); - local_irq_restore(flags); + local_unlock_irqrestore(&memcg_stock.lock, flags); } /* @@ -2779,29 +2780,28 @@ static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg) * which is cheap in non-preempt kernel. The interrupt context object stock * can only be accessed after disabling interrupt. User context code can * access interrupt object stock, but not vice versa. + * + * This task and interrupt context optimization is disabled for PREEMPT_RT + * as there is no performance gain in this case. */ static inline struct obj_stock *get_obj_stock(unsigned long *pflags) { - struct memcg_stock_pcp *stock; - - if (likely(in_task())) { + if (likely(in_task()) && !IS_ENABLED(CONFIG_PREEMPT_RT)) { *pflags = 0UL; preempt_disable(); - stock = this_cpu_ptr(&memcg_stock); - return &stock->task_obj; + return this_cpu_ptr(&memcg_stock.task_obj); } - local_irq_save(*pflags); - stock = this_cpu_ptr(&memcg_stock); - return &stock->irq_obj; + local_lock_irqsave(&memcg_stock.lock, *pflags); + return this_cpu_ptr(&memcg_stock.irq_obj); } static inline void put_obj_stock(unsigned long flags) { - if (likely(in_task())) + if (likely(in_task()) && !IS_ENABLED(CONFIG_PREEMPT_RT)) preempt_enable(); else - local_irq_restore(flags); + local_unlock_irqrestore(&memcg_stock.lock, flags); } /* @@ -7088,9 +7088,12 @@ static int __init mem_cgroup_init(void) cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL, memcg_hotplug_cpu_dead); - for_each_possible_cpu(cpu) - INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work, - drain_local_stock); + for_each_possible_cpu(cpu) { + struct memcg_stock_pcp *stock = per_cpu_ptr(&memcg_stock, cpu); + + INIT_WORK(&stock->work, drain_local_stock); + local_lock_init(&stock->lock); + } for_each_node(node) { struct mem_cgroup_tree_per_node *rtpn;
Direct calls to local_irq_{save/restore}() and preempt_{enable/disable}() are not appropriate for PREEMPT_RT. To provide better PREEMPT_RT support, change local_irq_{save/restore}() to local_lock_irq{save/restore}() and add a local_lock_t to struct memcg_stock_pcp. Also disable the task and interrupt context optimization for obj_stock as there will be no performance gain in the case of PREEMPT_RT. In this case, task obj_stock will be there but remain unused. Note that preempt_enable() and preempt_disable() in get_obj_stock() and put_obj_stock() are not replaced by local_lock() and local_unlock() as it is possible that a task accessing task_obj may get interrupted and then access irq_obj concurrently. So using local_lock for task_obj access may cause lockdep splat. Signed-off-by: Waiman Long <longman@redhat.com> --- mm/memcontrol.c | 41 ++++++++++++++++++++++------------------- 1 file changed, 22 insertions(+), 19 deletions(-)