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

>=20
>=20>=20
>=20> On Sat, Aug 01, 2026 at 05:19:25AM +0000, Zqiang wrote:
> >=20=20
>=20>=20=20
>=20>  Hello Zqiang,
> >=20=20
>=20>  On Thu, Jul 09, 2026 at 06:06:02PM +0800, Zqiang wrote:
> >=20=20
>=20>  >=20
>=20>  > 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 de=
lay
> >  > timer before flush_work(&sdp->work).
> >  >=20
>=20>  > However, if the timer_delete_sync() returns 1 means that it succ=
essfully
> >  > deleted an pending timer before it had a chance to fire, also mean=
s that
> >  > the sdp->work cannot be queued, the subsequent flush_work(&sdp->wo=
rk)
> >  > will returns immediately without waiting for anything, this causes=
 SRCU
> >  > callbacks to not be processed.
> >  >=20
>=20>  > 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.
> >  >=20
>=20>  > Signed-off-by: Zqiang <[email protected]>
> >  > ---
> >  > kernel/rcu/srcutree.c | 6 +++++-
> >  > 1 file changed, 5 insertions(+), 1 deletion(-)
> >  >=20
>=20>  > 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 =3D per_cpu_ptr(ssp->sda, cpu);
> >  >=20
>=20>  > - 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);
> >  >=20
>=20>  I started seeing this on my tests, it is not trivial to decode thi=
s one,
> >  but, I can try harder if _really_ needed.
> >=20=20
>=20>=20=20
>=20>  The scenario I can think of is that we missed the call to srcu_bar=
rier()
> >  before cleanup_srcu_struct():
> >=20=20
>=20>  loop_add()
> >  ->blk_mq_alloc_tag_set
> >  init_srcu_struct(&set->tags_srcu)
> >=20=20
>=20>  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_callb=
ack);
> >  } return error=20
>=20>=20=20
>=20>  goto out_free_mq_map:
> >  ....
> >  cleanup_srcu_struct(&set->tags_srcu)
> >  -> trigger WARN_ON(timer_delete_sync(&sdp->delay_work)
> >=20=20
>=20>  Can you try the following patch?
> >=20=20
>=20>  I tried it, and it does not help -- the WARN still fires at the sa=
me rate. I think the analysis points at the wrong call site.
> >=20=20
>=20>  Setup: linux-next-20260731 (arm64), 32 vCPU VM, HZ=3D1000, PROVE_L=
OCKING
> >  and DEBUG_OBJECTS_TIMERS enabled, reproducer stress-ng --loop 32
> >  --timeout 60s.
> >=20
>=20Thanks for provide testing methods, I will also testing it.
>=20
>=20>=20
>=20> baseline 3 x WARN srcutree.c:706 in 60s
> >  + your blk-mq patch 4 x WARN srcutree.c:706 in 60s
> >=20=20
>=20>  (3 vs 4 is just jitter on a one-jiffy race, not a regression.)
> >=20=20
>=20>  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 pat=
h
> >  your patch touches. All four splats in the patched run have the same
> >  trace:
> >=20=20
>=20>  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
> >=20
>=20This may trigger a new srcu grace period again during the window peri=
od between
> srcu-barrier() and cleanup_srcu_struct().
>=20
>=20>=20
>=20> I also put a pr_warn() at out_cleanup_tags_srcu: to be sure -- it f=
ired
> >  zero times over the whole run, so loop_add() never takes that error =
path
> >  in this workload.
> >=20=20
>=20>  Why do you wangt to have this
> >  WARN_ON(timer_delete_sync(&sdp->delay_work)) ?
> >=20=20
>=20>  timer_delete_sync() !=3D 0 "a timer was armed", not "callbacks are
> >  pending".
> >=20
>=20There are only two types of return values for timer_delete_sync(),
> return 0 or 1, the timer_delete_sync() !=3D 0 means that this timer
> was pending and has been deactivated, right?

How about this?

diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index b9fe57ff9100..62da31ec2bd0 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -751,7 +751,8 @@ void cleanup_srcu_struct(struct srcu_struct *ssp)

                // Call srcu_barrier() before this cleanup_srcu_struct()
                // to avoid triggering this WARN_ON().
-               if (WARN_ON(timer_delete_sync(&sdp->delay_work)) &&
+               if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist) &&
+                               timer_delete_sync(&sdp->delay_work)) &&
                                        rcu_cpu_beenfullyonline(sdp->cpu)=
)
                        queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
                flush_work(&sdp->work);

Thanks
Zqiang

>=20
>=20Thanks
> Zqiang
>=20
>=20>=20
>=20>=20
>=20>
>