Re: Query regarding srcu_funnel_exp_start()

From: Paul E. McKenney
Date: Sun Oct 29 2017 - 15:24:28 EST


On Sat, Oct 28, 2017 at 09:19:52AM +0530, Neeraj Upadhyay wrote:
> On 10/28/2017 03:50 AM, Paul E. McKenney wrote:
> >On Fri, Oct 27, 2017 at 10:15:04PM +0530, Neeraj Upadhyay wrote:
> >>On 10/27/2017 05:56 PM, Paul E. McKenney wrote:
> >>>On Fri, Oct 27, 2017 at 02:23:07PM +0530, Neeraj Upadhyay wrote:
> >>>>Hi,
> >>>>
> >>>>One query regarding srcu_funnel_exp_start() function in
> >>>>kernel/rcu/srcutree.c.
> >>>>
> >>>>static void srcu_funnel_exp_start(struct srcu_struct *sp, struct
> >>>>srcu_node *snp,
> >>>> unsigned long s)
> >>>>{
> >>>> <snip>
> >>>> if (!ULONG_CMP_LT(sp->srcu_gp_seq_needed_exp, s))
> >>>> sp->srcu_gp_seq_needed_exp = s;
> >>>> <snip>
> >>>>}
> >>>>
> >>>>Why is sp->srcu_gp_seq_needed_exp set to 's' if srcu_gp_seq_needed_exp is >=
> >>>>'s'. Shouldn't srcu_gp_seq_needed_exp be equal to the greater of both?
> >>>
> >>>Let's suppose that it is incorrect as currently written. Can you
> >>>construct a test case demonstrating a failure of some sort, then provide
> >>>a fix?
> >>
> >>Will check this. Might take some time to build a test case.
> >
> >Fair enough!
> >
> >I suggest checking to see if kernel/rcu/rcuperf.c can do what you need for
> >this test. (Might not with a single test, but perhaps a before-and-after
> >comparison. Or maybe you really do need to add some test code somewhere.)
>
> Thanks for the suggestion, will try that out.
>
> >>>To start with, if it is currently incorrect, what would be the nature
> >>>of the failure?
> >>>
> >>> Thanx, Paul
> >>>
> >>
> >>Hi Paul,
> >>
> >>I see below scenario, where new gp won't be expedited. Please correct
> >>me if I am missing something here.
> >>
> >>1. CPU0 calls synchronize_srcu_expedited()
> >>
> >>synchronize_srcu_expedited()
> >> __synchronize_srcu()
> >> __call_srcu()
> >> s = rcu_seq_snap(&sp->srcu_gp_seq); // lets say
> >>srcu_gp_seq = 0;
> >> // s = 0x100
> >
> >Looks like you have one hex digit and then two binary digits, but why not?
> >(RCU_SEQ_STATE_MASK is 3 rather than 0xff >
>
> Yeah, sorry I confused myself while representing the values. 0x100
> need to be replaced with b'100' and 0x200 with b'1000'.

Sounds like something I would do! ;-)

> >> sdp->srcu_gp_seq_needed = s // 0x100
> >> needgp = true
> >> sdp->srcu_gp_seq_needed_exp = s // 0x100
> >> srcu_funnel_gp_start()
> >> sp->srcu_gp_seq_needed_exp = s;
> >> srcu_gp_start(sp);
> >> rcu_seq_start(&sp->srcu_gp_seq);
> >>
> >>2. CPU1 calls normal synchronize_srcu()
> >>
> >>synchronize_srcu()
> >> __synchronize_srcu(sp, true)
> >> __call_srcu()
> >> s = rcu_seq_snap(&sp->srcu_gp_seq); // srcu_gp_seq = 1
> >> // s= 0x200
> >> sdp->srcu_gp_seq_needed = s; // 0x200
> >> srcu_funnel_gp_start()
> >> smp_store_release(&sp->srcu_gp_seq_needed, s); // 0x200
> >>
> >>3. CPU3 calls synchronize_srcu_expedited()
> >>
> >>synchronize_srcu_expedited()
> >> __synchronize_srcu()
> >> __call_srcu()
> >> s = rcu_seq_snap(&sp->srcu_gp_seq); // srcu_gp_seq = 1
> >> // s = 0x200
> >> sdp->srcu_gp_seq_needed_exp = s // 0x200
> >> srcu_funnel_exp_start(sp, sdp->mynode, s);
> >> // sp->srcu_gp_seq_needed_exp = 0x100
> >> // s = 0x200 ; sp->srcu_gp_seq_needed_exp is not updated
> >> if (!ULONG_CMP_LT(sp->srcu_gp_seq_needed_exp, s))
> >> sp->srcu_gp_seq_needed_exp = s;
> >
> >Seems plausible, but you should be able to show the difference in
> >grace-period duration with a test.
> >
>
> Ok sure, will attempt that.
>
> >While you are in srcu_funnel_exp_start(), should it be rechecking
> >rcu_seq_done(&sp->srcu_gp_seq, s) as well as the current
> >ULONG_CMP_GE(snp->srcu_gp_seq_needed_exp, s) under the lock?
> >Why or why not?
> >
> > Thanx, Paul
>
> Hi Paul,
>
> I don't see how it will impact. I have put markers in code snippet
> below to explain my points. My understanding is
>
> * rcu_seq_done check @a is a fastpath return, and avoid contention
> for snp lock, if the gp has already elapsed.
>
> * Checking it @b, inside srcu_node lock might not make any
> difference, as sp->srcu_gp_seq counter portion is updated
> under srcu_struct lock. Also, we cannot lock srcu_struct at this
> point, as it will cause lock contention among multiple CPUs.
>
> * Checking rcu_seq_done @c also does not impact, as we have already
> done all the work of traversing the entire parent chain and if
> rcu_seq_done() is true srcu_gp_seq_needed_exp will be greater
> than or equal to 's'.
>
> srcu_gp_end()
> raw_spin_lock_irq_rcu_node(sp);
> rcu_seq_end(&sp->srcu_gp_seq);
> gpseq = rcu_seq_current(&sp->srcu_gp_seq);
> if (ULONG_CMP_LT(sp->srcu_gp_seq_needed_exp, gpseq))
> sp->srcu_gp_seq_needed_exp = gpseq;
> raw_spin_unlock_irq_rcu_node(sp);
>
> static void srcu_funnel_exp_start(...)
> {
> <snip>
>
> for (; snp != NULL; snp = snp->srcu_parent) {
> if (rcu_seq_done(&sp->srcu_gp_seq, s) || /* a */
> ULONG_CMP_GE(READ_ONCE(snp->srcu_gp_seq_needed_exp), s))
> return;
> raw_spin_lock_irqsave_rcu_node(snp, flags);
> /* b */
> if (ULONG_CMP_GE(snp->srcu_gp_seq_needed_exp, s)) {
> raw_spin_unlock_irqrestore_rcu_node(snp, flags);
> return;
> }
> <snip>
> raw_spin_unlock_irqrestore_rcu_node(snp, flags);
> }
> raw_spin_lock_irqsave_rcu_node(sp, flags);
> /* c */
> if (!ULONG_CMP_LT(sp->srcu_gp_seq_needed_exp, s))
> sp->srcu_gp_seq_needed_exp = s;
> raw_spin_unlock_irqrestore_rcu_node(sp, flags);
> }

That does match my understanding, thank you for taking the time
to go through it! Especially given that my understanding has
proven to be wrong from time to time. ;-)

Thanx, Paul