From: Hui Zhu drain_all_stock() sets FLUSHING_CACHED_CHARGE before calling schedule_drain_work() to queue per-CPU drain work. When the target CPU is isolated (cpu_is_isolated() == true), the work is silently not queued, but FLUSHING_CACHED_CHARGE stays set. Every subsequent drain_all_stock() then sees the bit and skips this stock entirely, so the entry is effectively pinned until something else on that CPU runs drain_local_*_stock() and clears the bit -- which on a long- isolated CPU may never happen. The original idea was to actually perform the drain from the calling CPU on behalf of the isolated one, by adding a lock around the per-CPU stock so that a remote drainer could safely touch it. In practice this turned out to be intrusive: the stock data structures and their fast paths (consume_stock(), refill_stock(), the obj_stock helpers) are deliberately designed around current-CPU-only access, and retrofitting cross-CPU serialisation onto them adds non-trivial locking and PREEMPT_RT concerns for very little gain. Looking at the actual amount of charge that can accumulate in a single per-CPU stock, it is bounded and small, so leaving an isolated CPU's stock undrained for a while is not a real problem. The only real bug is that the stuck FLUSHING_CACHED_CHARGE bit prevents future drain_all_stock() callers from re-attempting once the CPU is no longer isolated. Fix this minimally by clearing FLUSHING_CACHED_CHARGE when the work could not be queued because the target CPU is isolated. The cached charge itself is left in place; it will be released the next time the CPU runs drain_local_*_stock() (e.g. after leaving isolation, or if the isolated CPU itself calls drain_all_stock() -- in that case cpu == curcpu causes drain_local_memcg_stock() to be invoked directly), and the next drain_all_stock() call is free to retry instead of skipping the stock forever. Fixes: 6a792697a53a ("memcg: do not drain charge pcp caches on remote isolated cpus") Signed-off-by: Hui Zhu --- Changelog: v3: According to the comments of Shakeel Butt, keep the whole schedule_drain_work() body under a single guard(rcu)() instead of splitting it with scoped_guard(rcu){}. v2: According to the comments of Waiman Long, updated fixes. mm/memcontrol.c | 26 +++++++++++++++++++++----- 1 file changed, 21 insertions(+), 5 deletions(-) diff --git a/mm/memcontrol.c b/mm/memcontrol.c index 6dc4888a90f3..42a0d0d6284a 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -2256,7 +2256,8 @@ static bool is_memcg_drain_needed(struct memcg_stock_pcp *stock, return flush; } -static void schedule_drain_work(int cpu, struct work_struct *work) +static void +schedule_drain_work(int cpu, struct work_struct *work, unsigned long *flags) { /* * Protect housekeeping cpumask read and work enqueue together @@ -2265,8 +2266,21 @@ static void schedule_drain_work(int cpu, struct work_struct *work) * pending work on newly isolated CPUs. */ guard(rcu)(); - if (!cpu_is_isolated(cpu)) - queue_work_on(cpu, memcg_wq, work); + if (cpu_is_isolated(cpu)) { + /* + * The target CPU is isolated: the drain work was not + * queued. Clear FLUSHING_CACHED_CHARGE so a future + * drain_all_stock() call can retry instead of skipping + * this stock forever. + * + * The cached charge itself is left untouched here; it + * will be drained the next time drain_local_stock() runs + * on this CPU (e.g. after it leaves isolation). + */ + clear_bit(FLUSHING_CACHED_CHARGE, flags); + return; + } + queue_work_on(cpu, memcg_wq, work); } /* @@ -2299,7 +2313,8 @@ void drain_all_stock(struct mem_cgroup *root_memcg) if (cpu == curcpu) drain_local_memcg_stock(&memcg_st->work); else - schedule_drain_work(cpu, &memcg_st->work); + schedule_drain_work(cpu, &memcg_st->work, + &memcg_st->flags); } if (!test_bit(FLUSHING_CACHED_CHARGE, &obj_st->flags) && @@ -2309,7 +2324,8 @@ void drain_all_stock(struct mem_cgroup *root_memcg) if (cpu == curcpu) drain_local_obj_stock(&obj_st->work); else - schedule_drain_work(cpu, &obj_st->work); + schedule_drain_work(cpu, &obj_st->work, + &obj_st->flags); } } migrate_enable(); -- 2.43.0