On Thu, Sep 12, 2024 at 10:07 AM Jesper Dangaard Brouer <hawk@xxxxxxxxxx> wrote: > > This patch reintroduces and generalizes the "stats_flush_ongoing" concept > to avoid redundant flushes if there is an ongoing flush at cgroup root > level, addressing production lock contention issues on the global cgroup > rstat lock. > > In this revision userspace readers will wait for the ongoing flusher to > complete before returning, to avoid reading out-dated stats just before > they get updated. Generally in-kernel users will attempt to skip the > flush in-order to get out of the lock contention state. Some in-kernel > users of the cgroup_rstat_flush() API depend on waiting for the flush to > complete before continuing. This patch introduce the call > cgroup_rstat_flush_relaxed() with a wait_for_flush option to satisfy both > use-cases. > > At Cloudflare, we observed significant performance degradation due to > lock contention on the rstat lock, primarily caused by kswapd. The > specific mem_cgroup_flush_stats() call inlined in shrink_node, which > takes the rstat lock, is particularly problematic. > > On our 12 NUMA node machines, each with a kswapd kthread per NUMA node, we > noted severe lock contention on the rstat lock, causing 12 CPUs to waste > cycles spinning every time kswapd runs. Fleet-wide stats (/proc/N/schedstat) > for kthreads revealed that we are burning an average of 20,000 CPU cores > fleet-wide on kswapd, primarily due to spinning on the rstat lock. > > Here's a brief overview of the issue: > - __alloc_pages_slowpath calls wake_all_kswapds, causing all kswapdN threads > to wake up simultaneously. > - The kswapd thread invokes shrink_node (via balance_pgdat), triggering the > cgroup rstat flush operation as part of its work. > - balance_pgdat() has a NULL value in target_mem_cgroup, causing > mem_cgroup_flush_stats() to flush with root_mem_cgroup. > > The kernel previously addressed this with a "stats_flush_ongoing" concept, > which was removed in commit 7d7ef0a4686a ("mm: memcg: restore subtree stats > flushing"). This patch reintroduces and generalizes the concept to apply to > all users of cgroup rstat, not just memcg. > > It have been a general theme to replace mem_cgroup_flush_stats() with > mem_cgroup_flush_stats_ratelimited every time we see a new case of this > issue. This will hide the contention issue until something starves the > kthread that does the periodic 2 second flush (for 2 periods). In > production we are seeing kthreads getting starved longer than 20 seconds. > This often happens in connection with OOM killer. This recreates the > kswapd lock contention situation at a very unfortunate point in time. > Thus, it makes sense to have this ongoing flusher lock contention > protection in place. > > In this patch only a root cgroup can become the ongoing flusher, as this solves > the production issue. Letting other levels becoming ongoing flusher cause root > cgroup to contend on the lock again. > > This change significantly reduces lock contention, especially in > environments with multiple NUMA nodes, thereby improving overall system > performance. > > Fixes: 7d7ef0a4686a ("mm: memcg: restore subtree stats flushing"). > Signed-off-by: Jesper Dangaard Brouer <hawk@xxxxxxxxxx> > --- To reiterate my response on v10, I prefer that this problem is handled at the reclaim flushing site. This started as a nice simple and generic approach, but ended up being complex and tailored to handle the reclaim flushing case. I do have some comments on the current implementation nonetheless. > v11: > - Address Yosry request to wait-for-flush for userspace readers > > V10: https://lore.kernel.org/all/172547884995.206112.808619042206173396.stgit@firesoul/ > > block/blk-cgroup.c | 2 - > include/linux/cgroup.h | 1 > include/linux/memcontrol.h | 1 > kernel/cgroup/rstat.c | 133 ++++++++++++++++++++++++++++++++++++++++++-- > mm/memcontrol.c | 40 +++++++++---- > mm/vmscan.c | 2 - > mm/zswap.c | 2 - > 7 files changed, 157 insertions(+), 24 deletions(-) > > diff --git a/block/blk-cgroup.c b/block/blk-cgroup.c > index 37e6cc91d576..058393e7665a 100644 > --- a/block/blk-cgroup.c > +++ b/block/blk-cgroup.c > @@ -1200,7 +1200,7 @@ static int blkcg_print_stat(struct seq_file *sf, void *v) > if (!seq_css(sf)->parent) > blkcg_fill_root_iostats(); > else > - cgroup_rstat_flush(blkcg->css.cgroup); > + cgroup_rstat_flush_relaxed(blkcg->css.cgroup, true); > > rcu_read_lock(); > hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { > diff --git a/include/linux/cgroup.h b/include/linux/cgroup.h > index 2150ca60394b..ff65bc100ca5 100644 > --- a/include/linux/cgroup.h > +++ b/include/linux/cgroup.h > @@ -691,6 +691,7 @@ void cgroup_rstat_updated(struct cgroup *cgrp, int cpu); > void cgroup_rstat_flush(struct cgroup *cgrp); > void cgroup_rstat_flush_hold(struct cgroup *cgrp); > void cgroup_rstat_flush_release(struct cgroup *cgrp); > +int cgroup_rstat_flush_relaxed(struct cgroup *cgrp, bool wait_for_flush); We now have 4 different flavors of rstat flushing: 1. cgroup_rstat_flush() -> normal flush: lock, flush, unlock 2. cgroup_rstat_flush_hold() -> same, but keep the lock held 3. cgroup_rstat_flush_relaxed(wait_for_flush=true) -> if someone is already flushing us, wait, otherwise normal 4. cgroup_rstat_flush_relaxed(wait_for_flush=false) -> if someone is already flushing us, wait, otherwise normal Why do we need the third one? Can't we just keep using (1) for now for userspace readers? > > /* > * Basic resource stats. > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h > index 030d34e9d117..7e24c5e1327f 100644 > --- a/include/linux/memcontrol.h > +++ b/include/linux/memcontrol.h > @@ -1026,6 +1026,7 @@ unsigned long lruvec_page_state_local(struct lruvec *lruvec, > enum node_stat_item idx); > > void mem_cgroup_flush_stats(struct mem_cgroup *memcg); > +void mem_cgroup_flush_stats_relaxed(struct mem_cgroup *memcg); > void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg); > > void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val); > diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c > index a06b45272411..80a4b949138f 100644 > --- a/kernel/cgroup/rstat.c > +++ b/kernel/cgroup/rstat.c > @@ -11,6 +11,9 @@ > > static DEFINE_SPINLOCK(cgroup_rstat_lock); > static DEFINE_PER_CPU(raw_spinlock_t, cgroup_rstat_cpu_lock); > +static struct cgroup *cgrp_rstat_ongoing_flusher = NULL; > +static struct task_struct *cgrp_rstat_ongoing_flusher_ID = NULL; rstat_ongoing_flush_cgrp and rstat_ongoing_flush_task are probably clearer names. > +static DEFINE_MUTEX(cgrp_rstat_ongoing_flusher_serialize); > > static void cgroup_base_stat_flush(struct cgroup *cgrp, int cpu); > > @@ -299,6 +302,68 @@ static inline void __cgroup_rstat_unlock(struct cgroup *cgrp, int cpu_in_loop) > spin_unlock_irq(&cgroup_rstat_lock); > } > > +static inline bool cgroup_is_root(struct cgroup *cgrp) > +{ > + return cgroup_parent(cgrp) == NULL; > +} > + > +/** > + * cgroup_rstat_trylock_flusher - Trylock that checks for on ongoing flusher > + * @cgrp: target cgroup > + * @strict: always lock and ignore/skip ongoing flusher checks > + * > + * Function return value follow trylock semantics. Returning true when lock is > + * obtained. Returning false when not locked and it detected flushing can be > + * skipped as another ongoing flusher is taking care of the flush. > + * > + * For callers that depend on flush completing before returning a strict option > + * is provided. > + */ > +static bool cgroup_rstat_trylock_flusher(struct cgroup *cgrp, bool strict) > +{ > + struct cgroup *ongoing; > + > + if (strict) > + goto lock; Might as well just have a cgroup_rstat_lock_flusher() function instead of the boolean parameter, which is also consistent with the lock/trylock semantics you are following. > + > + /* > + * Check if ongoing flusher is already taking care of this. Descendant > + * check is necessary due to cgroup v1 supporting multiple root's. > + */ > + ongoing = READ_ONCE(cgrp_rstat_ongoing_flusher); > + if (ongoing && cgroup_is_descendant(cgrp, ongoing)) The ongoing flusher may be going away, so cgroup_is_descendant() may be a UAF as I pointed out a few times. I actually think even taking a ref here may not work as it may be a flush from cgroup_rstat_exit(). One thing we can do is get the root of cgrp (probably cgrp->root->cgrp?) and compare it to cgrp_rstat_ongoing_flusher without ever dereferencing it. If the ongoing flusher is in fact the root of cgrp, this implies a ref on it anyway so we can dereference it if needed. That would obviously need a comment to explain it. > + return false; > + > + /* Grab right to be ongoing flusher */ > + if (!ongoing && cgroup_is_root(cgrp)) { > + struct cgroup *old; > + > + old = cmpxchg(&cgrp_rstat_ongoing_flusher, NULL, cgrp); > + if (old) { > + /* Lost race for being ongoing flusher */ > + if (cgroup_is_descendant(cgrp, old)) > + return false; > + } > + /* Due to lock yield combined with strict mode record ID */ This needs a more detailed comment. > + WRITE_ONCE(cgrp_rstat_ongoing_flusher_ID, current); This will overwrite the ID if we lost the race but are not a descendant of the flusher that one the race, right? > + } > +lock: > + __cgroup_rstat_lock(cgrp, -1); > + > + return true; > +} > + > +static void cgroup_rstat_unlock_flusher(struct cgroup *cgrp) > +{ > + if (cgrp == READ_ONCE(cgrp_rstat_ongoing_flusher) && > + READ_ONCE(cgrp_rstat_ongoing_flusher_ID) == current) { > + WRITE_ONCE(cgrp_rstat_ongoing_flusher_ID, NULL); > + WRITE_ONCE(cgrp_rstat_ongoing_flusher, NULL); > + } > + > + __cgroup_rstat_unlock(cgrp, -1); > +} > + > /* see cgroup_rstat_flush() */ > static void cgroup_rstat_flush_locked(struct cgroup *cgrp) > __releases(&cgroup_rstat_lock) __acquires(&cgroup_rstat_lock) > @@ -333,6 +398,19 @@ static void cgroup_rstat_flush_locked(struct cgroup *cgrp) > } > } > > +static int __cgroup_rstat_flush(struct cgroup *cgrp, bool strict) > +{ > + might_sleep(); > + > + if (!cgroup_rstat_trylock_flusher(cgrp, strict)) > + return false; > + > + cgroup_rstat_flush_locked(cgrp); > + cgroup_rstat_unlock_flusher(cgrp); > + > + return true; > +} > + > /** > * cgroup_rstat_flush - flush stats in @cgrp's subtree > * @cgrp: target cgroup > @@ -348,11 +426,49 @@ static void cgroup_rstat_flush_locked(struct cgroup *cgrp) > */ > __bpf_kfunc void cgroup_rstat_flush(struct cgroup *cgrp) > { > - might_sleep(); > + __cgroup_rstat_flush(cgrp, true); > +} > > - __cgroup_rstat_lock(cgrp, -1); > - cgroup_rstat_flush_locked(cgrp); > - __cgroup_rstat_unlock(cgrp, -1); > +int cgroup_rstat_flush_relaxed(struct cgroup *cgrp, bool wait_for_flush) > +{ > + bool flushed = __cgroup_rstat_flush(cgrp, false); > + > + if (!flushed && wait_for_flush) { Isn't the code below essentially open-coding completions in a less efficient way? Anyway as I mentioned above I don't believe we really need wait_for_flush for now. I was hoping we can make this work for both in-kernel and userspace flushers, but I don't think we want to add it just for userspace flushers. > + /* > + * Reaching here we know an ongoing flusher is running, that > + * will take care of flushing for us, but for caller to read > + * accurate stats we want to wait for this ongoing flusher. > + * > + * TODO: When lock becomes mutex and no-yielding this code can > + * be simplifed as we can just sleep on the mutex lock. > + */ > + struct task_struct *id, *cur_id; > + u64 timeout; > + > + id = READ_ONCE(cgrp_rstat_ongoing_flusher_ID); > + timeout = jiffies_64 + msecs_to_jiffies(50); > + > + if (!id) > + return false; > + > + cond_resched(); > + /* We might get lucky and flush already completed */ > + cur_id = READ_ONCE(cgrp_rstat_ongoing_flusher_ID); > + > + /* Due to lock yield, make sure "id" flusher completes */ > + while (cur_id == id && time_before64(jiffies_64, timeout)) { > + cond_resched(); > + /* Use mutex to reduce stress on global lock */ > + mutex_lock(&cgrp_rstat_ongoing_flusher_serialize); > + __cgroup_rstat_lock(cgrp, -1); > + /* Get lock with ongoing can happen due to yielding */ > + cur_id = READ_ONCE(cgrp_rstat_ongoing_flusher_ID); > + __cgroup_rstat_unlock(cgrp, -1); > + mutex_unlock(&cgrp_rstat_ongoing_flusher_serialize); > + } > + } > + > + return flushed; > } > > /** > @@ -368,8 +484,11 @@ void cgroup_rstat_flush_hold(struct cgroup *cgrp) > __acquires(&cgroup_rstat_lock) > { > might_sleep(); > - __cgroup_rstat_lock(cgrp, -1); > - cgroup_rstat_flush_locked(cgrp); > + > + if (cgroup_rstat_trylock_flusher(cgrp, false)) > + cgroup_rstat_flush_locked(cgrp); > + else > + __cgroup_rstat_lock(cgrp, -1); > } > > /** > @@ -379,7 +498,7 @@ void cgroup_rstat_flush_hold(struct cgroup *cgrp) > void cgroup_rstat_flush_release(struct cgroup *cgrp) > __releases(&cgroup_rstat_lock) > { > - __cgroup_rstat_unlock(cgrp, -1); > + cgroup_rstat_unlock_flusher(cgrp); > } > > int cgroup_rstat_init(struct cgroup *cgrp) > diff --git a/mm/memcontrol.c b/mm/memcontrol.c > index 71fe2a95b8bd..6694f7a859b5 100644 > --- a/mm/memcontrol.c > +++ b/mm/memcontrol.c > @@ -871,12 +871,26 @@ static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val) > } > } > > -static void do_flush_stats(struct mem_cgroup *memcg) > +static void do_flush_stats(struct mem_cgroup *memcg, bool wait_for_flush) > { > - if (mem_cgroup_is_root(memcg)) > - WRITE_ONCE(flush_last_time, jiffies_64); > + bool flushed = cgroup_rstat_flush_relaxed(memcg->css.cgroup, > + wait_for_flush); > > - cgroup_rstat_flush(memcg->css.cgroup); > + if (mem_cgroup_is_root(memcg) && flushed) > + WRITE_ONCE(flush_last_time, jiffies_64); > +} > + > +static void __mem_cgroup_flush_stats(struct mem_cgroup *memcg, > + bool wait_for_flush) > +{ > + if (mem_cgroup_disabled()) > + return; > + > + if (!memcg) > + memcg = root_mem_cgroup; > + > + if (memcg_vmstats_needs_flush(memcg->vmstats)) > + do_flush_stats(memcg, wait_for_flush); > } > > /* > @@ -890,21 +904,19 @@ static void do_flush_stats(struct mem_cgroup *memcg) > */ > void mem_cgroup_flush_stats(struct mem_cgroup *memcg) > { > - if (mem_cgroup_disabled()) > - return; > - > - if (!memcg) > - memcg = root_mem_cgroup; > + __mem_cgroup_flush_stats(memcg, true); > +} > > - if (memcg_vmstats_needs_flush(memcg->vmstats)) > - do_flush_stats(memcg); > +void mem_cgroup_flush_stats_relaxed(struct mem_cgroup *memcg) > +{ > + __mem_cgroup_flush_stats(memcg, false); > } > > void mem_cgroup_flush_stats_ratelimited(struct mem_cgroup *memcg) > { > /* Only flush if the periodic flusher is one full cycle late */ > if (time_after64(jiffies_64, READ_ONCE(flush_last_time) + 2*FLUSH_TIME)) > - mem_cgroup_flush_stats(memcg); > + mem_cgroup_flush_stats_relaxed(memcg); mem_cgroup_flush_stats_ratelimited() now basically says: only flush if the periodic flusher is late but no one else is flushing, if we end up pursuing this we probably want to start documenting these variants.