Re: Commit 282d8998e997 (srcu: Prevent expedited GPs and blocking readers from consuming CPU) cause qemu boot slow

[zhangfei.gao@xxxxxxxxxxx]

On 2022/6/14 下午10:17, Paul E. McKenney wrote:
> On Tue, Jun 14, 2022 at 10:03:35PM +0800, zhangfei.gao@xxxxxxxxxxx wrote:
> > On 2022/6/14 下午8:19, Neeraj Upadhyay wrote:
> > > > 5.18-rc4 based               ~8sec
> > > >
> > > > 5.19-rc1                     ~2m43sec
> > > >
> > > > 5.19-rc1+fix1                 ~19sec
> > > >
> > > > 5.19-rc1-fix2                 ~19sec
> > > If you try below diff on top of either 5.19-rc1+fix1 or 5.19-rc1-fix2 ;
> > > does it show any difference in boot time?
> > >
> > > --- a/kernel/rcu/srcutree.c
> > > +++ b/kernel/rcu/srcutree.c
> > > @@ -706,7 +706,7 @@ static void srcu_schedule_cbs_snp(struct srcu_struct
> > > *ssp, struct srcu_node *snp
> > >    */
> > >   static void srcu_gp_end(struct srcu_struct *ssp)
> > >   {
> > > -       unsigned long cbdelay;
> > > +       unsigned long cbdelay = 1;
> > >          bool cbs;
> > >          bool last_lvl;
> > >          int cpu;
> > > @@ -726,7 +726,9 @@ static void srcu_gp_end(struct srcu_struct *ssp)
> > >          spin_lock_irq_rcu_node(ssp);
> > >          idx = rcu_seq_state(ssp->srcu_gp_seq);
> > >          WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
> > > -       cbdelay = !!srcu_get_delay(ssp);
> > > +       if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq),
> > > READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
> > > +               cbdelay = 0;
> > > +
> > >          WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
> Thank you both for the testing and the proposed fix!
>
> > Test here:
> > qemu: https://github.com/qemu/qemu/tree/stable-6.1
> > kernel:
> > https://github.com/Linaro/linux-kernel-uadk/tree/uacce-devel-5.19-srcu-test
> > (in case test patch not clear, push in git tree)
> >
> > Hardware: aarch64
> >
> > 1. 5.18-rc6
> > real    0m8.402s
> > user    0m3.015s
> > sys     0m1.102s
> >
> > 2. 5.19-rc1
> > real    2m41.433s
> > user    0m3.097s
> > sys     0m1.177s
> >
> > 3. 5.19-rc1 + fix1 from Paul
> > real    2m43.404s
> > user    0m2.880s
> > sys     0m1.214s
> >
> > 4. 5.19-rc1 + fix2: fix1 + Remove "if (!jbase)" block
> > real    0m15.262s
> > user    0m3.003s
> > sys     0m1.033s
> >
> > When build kernel in the meantime, load time become longer.
> >
> > 5. 5.19-rc1 + fix3: fix1 + SRCU_MAX_NODELAY_PHASE 1000000
> > real    0m15.215s
> > user    0m2.942s
> > sys    0m1.172s
> >
> > 6. 5.19-rc1 + fix4: fix1 + Neeraj's change of srcu_gp_end　
> > real    1m23.936s
> > user    0m2.969s
> > sys    0m1.181s
> And thank you for the testing!
>
> Could you please try fix3 + Neeraj's change of srcu_gp_end?
>
> That is, fix1 + SRCU_MAX_NODELAY_PHASE 1000000 + Neeraj's change of
> srcu_gp_end.
>
> Also, at what value of SRCU_MAX_NODELAY_PHASE do the boot
> times start rising?  This is probably best done by starting with
> SRCU_MAX_NODELAY_PHASE=100000 and dividing by (say) ten on each run
> until boot time becomes slow, followed by a binary search between the
> last two values.  (The idea is to bias the search so that fast boot
> times are the common case.)

SRCU_MAX_NODELAY_PHASE 100 becomes slower.


8. 5.19-rc1 + fix6: fix4 + SRCU_MAX_NODELAY_PHASE 1000000

real 0m11.154s ~12s

user 0m2.919s

sys 0m1.064s



9. 5.19-rc1 + fix7: fix4 + SRCU_MAX_NODELAY_PHASE 10000

real 0m11.258s

user 0m3.113s

sys 0m1.073s



10. 5.19-rc1 + fix8: fix4 + SRCU_MAX_NODELAY_PHASE 100

real 0m30.053s ~ 32s

user 0m2.827s

sys 0m1.161s



By the way, if build kernel on the board in the meantime (using memory), 
time become much longer.

real 1m2.763s



11. 5.19-rc1 + fix9: fix4 + SRCU_MAX_NODELAY_PHASE 1000

real 0m11.443s

user 0m3.022s

sys 0m1.052s


Thanks

> > More test details: https://docs.qq.com/doc/DRXdKalFPTVlUbFN5
> And thank you for these details.
>
> 							Thanx, Paul