[Tarantool-patches] [PATCH v2 4/4] qsync: don't send negative timeouts into fiber_cond_wait_timeout

[Cyrill Gorcunov]

On Fri, Jul 17, 2020 at 12:45:16AM +0200, Vladislav Shpilevoy wrote:
> Thanks for the patch!
> 
> On 14.07.2020 16:53, Cyrill Gorcunov wrote:
> > Basically our timeout is calculated via (a - b), where
> > @a is a constant positive value fetched once, in turn
> > the @b is rather a dynamic value thus the result may
> > be negative. libev uses assert() call to catch such
> > values when passed to timers setup. Thus lets intercept
> > potential assert() trigger and exit early if timeout
> > is already expired.
> > 
> > Signed-off-by: Cyrill Gorcunov <gorcunov at gmail.com>
> > ---
> > There were a typo, so I force-updated the branch
> > 
> >  src/box/txn_limbo.c | 8 ++++++--
> >  1 file changed, 6 insertions(+), 2 deletions(-)
> > 
> > diff --git a/src/box/txn_limbo.c b/src/box/txn_limbo.c
> > index d5b887d36..0924952b7 100644
> > --- a/src/box/txn_limbo.c
> > +++ b/src/box/txn_limbo.c
> > @@ -174,8 +174,10 @@ txn_limbo_wait_complete(struct txn_limbo *limbo, struct txn_limbo_entry *entry)
> >  	double start_time = fiber_clock();
> >  	while (true) {
> >  		double deadline = start_time + replication_synchro_timeout;
> > -		bool cancellable = fiber_set_cancellable(false);
> >  		double timeout = deadline - fiber_clock();
> > +		if (timeout < 0)
> > +			goto do_rollback;
> > +		bool cancellable = fiber_set_cancellable(false);
> 
> I added timeout = -1; here and tried to commit a sync transaction. I
> got timed out error, no assertions. Please, tell me, how to reproduce
> the assertion you mention in the commit message. Otherwise I don't see
> why would we need the < 0 check. If it is done somewhere inside
> fiber_cond_wait_timeout anyway.

Look, here is libev code

---
noinline
void
ev_timer_start (EV_P_ ev_timer *w) EV_THROW
{
  if (expect_false (ev_is_active (w)))
    return;

  ev_at (w) += mn_now;

  assert (("libev: ev_timer_start called with negative timer repeat value", w->repeat >= 0.));
---

To trigger this assert we have to enter idle cycle, then manually change
replication_synchro_timeout via cfg (make it less than it was initially)
which should lead to negative timeout.

To be fair I don't know how to force it without error injection. Lets
assume we have initial fiber clock 1 (the clocks are increasing everytime
libev does a new polling cycle). Thus

	// replication_synchro_timeout = 2

	double start_time = fiber_clock();	// 1
	while (true) {
		double deadline = start_time + replication_synchro_timeout;
		// => 3
		double timeout = deadline - fiber_clock();
		// => 2
		bool cancellable = fiber_set_cancellable(false);
		int rc = fiber_cond_wait_timeout(&limbo->wait_cond, timeout);
		//
		// enter into resched cycle
		// manually change replication_synchro_timeout to 1
		// next cycle starts

		double deadline = start_time + replication_synchro_timeout;
		// start_time didn't changed
		// => 1 + 1 = 2
		double timeout = deadline - fiber_clock();
		// assume several resched cycles passed, thus
		// fiber_clock returns 3, thus
		// => 2 - 3 => -1
	...

Am I missing something obvious here?