Re: [PATCH] srcu: Queue sdp->work when the delay timer is successfully deleted

["Zqiang" <qiang.zhang@linux.dev>]

> 
> On Sat, Aug 01, 2026 at 05:19:25AM +0000, Zqiang wrote:
> 
> > 
> > Hello Zqiang,
> >  
> >  On Thu, Jul 09, 2026 at 06:06:02PM +0800, Zqiang wrote:
> >  
> >  > 
> >  > In the cleanup_srcu_struct(), when iterating over per-cpu's srcu_data,
> >  > the timer_delete_sync(&sdp->delay_work) is called to cancel the delay
> >  > timer before flush_work(&sdp->work).
> >  > 
> >  > However, if the timer_delete_sync() returns 1 means that it successfully
> >  > deleted an pending timer before it had a chance to fire, also means that
> >  > the sdp->work cannot be queued, the subsequent flush_work(&sdp->work)
> >  > will returns immediately without waiting for anything, this causes SRCU
> >  > callbacks to not be processed.
> >  > 
> >  > Fix this by checking the return value of timer_delete_sync(), if it
> >  > returns 1, explicitly queue sdp->work so that the following flush_work()
> >  > can correctly wait for the work to complete.
> >  > 
> >  > Signed-off-by: Zqiang <qiang.zhang@linux.dev>
> >  > ---
> >  > kernel/rcu/srcutree.c | 6 +++++-
> >  > 1 file changed, 5 insertions(+), 1 deletion(-)
> >  > 
> >  > diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
> >  > index 7c2f7cc131f7..02c322b7c6f1 100644
> >  > --- a/kernel/rcu/srcutree.c
> >  > +++ b/kernel/rcu/srcutree.c
> >  > @@ -725,7 +725,11 @@ void cleanup_srcu_struct(struct srcu_struct *ssp)
> >  > for_each_possible_cpu(cpu) {
> >  > struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
> >  > 
> >  > - timer_delete_sync(&sdp->delay_work);
> >  > + //In most scenarios, calling srcu_barrier before cleanup
> >  > + //will not trigger WARN_ON().
> >  > + if (WARN_ON(timer_delete_sync(&sdp->delay_work)) &&
> >  > + rcu_cpu_beenfullyonline(sdp->cpu))
> >  > + queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
> >  > 
> >  I started seeing this on my tests, it is not trivial to decode this one,
> >  but, I can try harder if _really_ needed.
> > 
> >  
> >  The scenario I can think of is that we missed the call to srcu_barrier()
> >  before cleanup_srcu_struct():
> >  
> >  loop_add()
> >  ->blk_mq_alloc_tag_set
> >  init_srcu_struct(&set->tags_srcu)
> >  
> >  blk_mq_alloc_set_map_and_rqs() {
> >  ->__blk_mq_alloc_rq_maps()
> >  ->__blk_mq_alloc_map_and_rqs() return error
> >  goto out_unwind: __blk_mq_free_map_and_rqs()
> >  ->blk_mq_free_rq_map()
> >  ->blk_mq_free_tags()
> >  ->call_srcu(&set->tags_srcu, &tags->rcu_head, blk_mq_free_tags_callback);
> >  } return error 
> >  
> >  goto out_free_mq_map:
> >  ....
> >  cleanup_srcu_struct(&set->tags_srcu)
> >  -> trigger WARN_ON(timer_delete_sync(&sdp->delay_work)
> >  
> >  Can you try the following patch?
> > 
> I tried it, and it does not help -- the WARN still fires at the same rate. I think the analysis points at the wrong call site.
> 
> Setup: linux-next-20260731 (arm64), 32 vCPU VM, HZ=1000, PROVE_LOCKING
> and DEBUG_OBJECTS_TIMERS enabled, reproducer stress-ng --loop 32
> --timeout 60s.
> 

Thanks for provide testing methods, I will also testing it.

>  baseline 3 x WARN srcutree.c:706 in 60s
>  + your blk-mq patch 4 x WARN srcutree.c:706 in 60s
> 
>  (3 vs 4 is just jitter on a one-jiffy race, not a regression.)
> 
> The reason it cannot help is that the splat comes from
> blk_mq_free_tag_set(), not from the blk_mq_alloc_tag_set() error path
> your patch touches. All four splats in the patched run have the same
> trace:
> 
>  cleanup_srcu_struct+0x274/0x450 (P)
>  blk_mq_free_tag_set+0x1a4/0x1e0
>  loop_remove+0x2c/0x78
>  loop_control_ioctl+0x248/0x2a0
>  __arm64_sys_ioctl+0x9c0/0xb00

This may trigger a new srcu grace period again during the window period between
srcu-barrier() and cleanup_srcu_struct().

> 
> I also put a pr_warn() at out_cleanup_tags_srcu: to be sure -- it fired
> zero times over the whole run, so loop_add() never takes that error path
> in this workload.
> 
> Why do you wangt to have this
> WARN_ON(timer_delete_sync(&sdp->delay_work)) ?
> 
> timer_delete_sync() != 0 "a timer was armed", not "callbacks are
> pending".

There are only two types of return values for timer_delete_sync(),
return 0 or 1,  the timer_delete_sync() != 0 means that this timer
was pending and has been deactivated, right?

Thanks
Zqiang



>