Message-ID: <20220926235944.GE4196@paulmck-ThinkPad-P17-Gen-1>
Date:   Mon, 26 Sep 2022 16:59:44 -0700
From:   "Paul E. McKenney" <paulmck@...nel.org>
To:     Joel Fernandes <joel@...lfernandes.org>
Cc:     Uladzislau Rezki <urezki@...il.com>, rcu@...r.kernel.org,
        linux-kernel@...r.kernel.org, rushikesh.s.kadam@...el.com,
        neeraj.iitr10@...il.com, frederic@...nel.org, rostedt@...dmis.org
Subject: Re: [PATCH v6 1/4] rcu: Make call_rcu() lazy to save power

On Mon, Sep 26, 2022 at 07:47:50PM -0400, Joel Fernandes wrote:
> 
> 
> > On Sep 26, 2022, at 6:32 PM, Paul E. McKenney <paulmck@...nel.org> wrote:
> > 
> > On Mon, Sep 26, 2022 at 09:02:21PM +0000, Joel Fernandes wrote:
> >> On Mon, Sep 26, 2022 at 09:32:44PM +0200, Uladzislau Rezki wrote:
> >> [...]
> >>>>>> On my KVM machine the boot time is affected:
> >>>>>> 
> >>>>>> <snip>
> >>>>>> [    2.273406] e1000 0000:00:03.0 eth0: Intel(R) PRO/1000 Network Connection
> >>>>>> [   11.945283] e1000 0000:00:03.0 ens3: renamed from eth0
> >>>>>> [   22.165198] sr 1:0:0:0: [sr0] scsi3-mmc drive: 4x/4x cd/rw xa/form2 tray
> >>>>>> [   22.165206] cdrom: Uniform CD-ROM driver Revision: 3.20
> >>>>>> [   32.406981] sr 1:0:0:0: Attached scsi CD-ROM sr0
> >>>>>> [  104.115418] process '/usr/bin/fstype' started with executable stack
> >>>>>> [  104.170142] EXT4-fs (sda1): mounted filesystem with ordered data mode. Quota mode: none.
> >>>>>> [  104.340125] systemd[1]: systemd 241 running in system mode. (+PAM +AUDIT +SELINUX +IMA +APPARMOR +SMACK +SYSVINIT +UTMP +LIBCRYPTSETUP +GCRYPT +GNUTLS +ACL +XZ +LZ4 +SECCOMP +BLKID +ELFUTILS +KMOD -IDN2 +IDN -PCRE2 default-hierarchy=hybrid)
> >>>>>> [  104.340193] systemd[1]: Detected virtualization kvm.
> >>>>>> [  104.340196] systemd[1]: Detected architecture x86-64.
> >>>>>> [  104.359032] systemd[1]: Set hostname to <pc638>.
> >>>>>> [  105.740109] random: crng init done
> >>>>>> [  105.741267] systemd[1]: Reached target Remote File Systems.
> >>>>>> <snip>
> >>>>>> 
> >>>>>> 2 - 11 and second delay is between 32 - 104. So there are still users which must
> >>>>>> be waiting for "RCU" in a sync way.
> >>>>> 
> >>>>> I was wondering if you can compare boot logs and see which timestamp does the
> >>>>> slow down start from. That way, we can narrow down the callback. Also another
> >>>>> idea is, add "trace_event=rcu:rcu_callback,rcu:rcu_invoke_callback
> >>>>> ftrace_dump_on_oops" to the boot params, and then manually call
> >>>>> "tracing_off(); panic();" from the code at the first printk that seems off in
> >>>>> your comparison of good vs bad. For example, if "crng init done" timestamp is
> >>>>> off, put the "tracing_off(); panic();" there. Then grab the serial console
> >>>>> output to see what were the last callbacks that was queued/invoked.
> >>>> 
> >>>> We do seem to be in need of some way to quickly and easily locate the
> >>>> callback that needed to be _flush() due to a wakeup.
> >>>> 
> >>> <snip>
> >>> diff --git a/kernel/workqueue.c b/kernel/workqueue.c
> >>> index aeea9731ef80..fe1146d97f1a 100644
> >>> --- a/kernel/workqueue.c
> >>> +++ b/kernel/workqueue.c
> >>> @@ -1771,7 +1771,7 @@ bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
> >>> 
> >>>        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
> >>>                rwork->wq = wq;
> >>> -               call_rcu(&rwork->rcu, rcu_work_rcufn);
> >>> +               call_rcu_flush(&rwork->rcu, rcu_work_rcufn);
> >>>                return true;
> >>>        }
> >>> 
> >>> <snip>
> >>> 
> >>> ?
> >>> 
> >>> But it does not fully solve my boot-up issue. Will debug tomorrow further.
> >> 
> >> Ah, but at least its progress, thanks. Could you send me a patch to include
> >> in the next revision with details of this?
> >> 
> >>>> Might one more proactive approach be to use Coccinelle to locate such
> >>>> callback functions?  We might not want -all- callbacks that do wakeups
> >>>> to use call_rcu_flush(), but knowing which are which should speed up
> >>>> slow-boot debugging by quite a bit.
> >>>> 
> >>>> Or is there a better way to do this?
> >>>> 
> >>> I am not sure what Coccinelle is. If we had something automated that measures
> >>> a boot time and if needed does some profiling it would be good. Otherwise it
> >>> is a manual debugging mainly, IMHO.
> >> 
> >> Paul, What about using a default-off kernel CONFIG that splats on all lazy
> >> call_rcu() callbacks that do a wake up. We could use the trace hooks to do it
> >> in kernel I think. I can talk to Steve to get ideas on how to do that but I
> >> think it can be done purely from trace events (we might need a new
> >> trace_end_invoke_callback to fire after the callback is invoked). Thoughts?
> > 
> > Could you look for wakeups invoked between trace_rcu_batch_start() and
> > trace_rcu_batch_end() that are not from interrupt context?  This would
> > of course need to be associated with a task rather than a CPU.
> 
> Yes this sounds good, but we also need to know if the callbacks are lazy or not since wake-up is ok from a non lazy one. I think I’ll need a table to track that at queuing time.

Agreed.

> > Note that you would need to check for wakeups from interrupt handlers
> > even with the extra trace_end_invoke_callback().  The window where an
> > interrupt handler could do a wakeup would be reduced, but not eliminated.
> 
> True! Since this is a  debugging option, can we not just disable interrupts across callback invocation?

Not without terminally annoying lockdep, at least for any RCU callbacks
doing things like spin_lock_bh().

							Thanx, Paul

Powered by blists - more mailing lists