Tarantool development patches archive
 help / color / mirror / Atom feed
From: Cyrill Gorcunov <gorcunov@gmail.com>
To: Vladislav Shpilevoy <v.shpilevoy@tarantool.org>
Cc: tml <tarantool-patches@dev.tarantool.org>
Subject: Re: [Tarantool-patches] [PATCH v2 4/4] qsync: don't send negative timeouts into fiber_cond_wait_timeout
Date: Fri, 17 Jul 2020 10:25:50 +0300	[thread overview]
Message-ID: <20200717072550.GA2613@grain> (raw)
In-Reply-To: <c50046b5-59f2-2696-ce34-97aee98ad204@tarantool.org>

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@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?

  reply	other threads:[~2020-07-17  7:25 UTC|newest]

Thread overview: 12+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-07-14 14:44 [Tarantool-patches] [PATCH 0/4] qsync: continue cooking the code Cyrill Gorcunov
2020-07-14 14:44 ` [Tarantool-patches] [PATCH 1/4] qsync: add missing functions descriptions Cyrill Gorcunov
2020-07-14 14:44 ` [Tarantool-patches] [PATCH 2/4] qsync: drop txn_limbo_confirm_timeout Cyrill Gorcunov
2020-07-14 14:44 ` [Tarantool-patches] [PATCH 3/4] qsync: txn_limbo_wait_complete -- fix type conversion Cyrill Gorcunov
2020-07-14 14:44 ` [Tarantool-patches] [PATCH 4/4] qsync: don't send negative timeouts into fiber_cond_wait_timeout Cyrill Gorcunov
2020-07-14 14:53   ` [Tarantool-patches] [PATCH v2 " Cyrill Gorcunov
2020-07-16 22:45     ` Vladislav Shpilevoy
2020-07-17  7:25       ` Cyrill Gorcunov [this message]
2020-07-17 20:03         ` Vladislav Shpilevoy
2020-07-17 20:33           ` Cyrill Gorcunov
2020-07-17 20:43             ` Cyrill Gorcunov
2020-07-20 19:28 ` [Tarantool-patches] [PATCH 0/4] qsync: continue cooking the code Vladislav Shpilevoy

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20200717072550.GA2613@grain \
    --to=gorcunov@gmail.com \
    --cc=tarantool-patches@dev.tarantool.org \
    --cc=v.shpilevoy@tarantool.org \
    --subject='Re: [Tarantool-patches] [PATCH v2 4/4] qsync: don'\''t send negative timeouts into fiber_cond_wait_timeout' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox