Re: [PATCH] Enable sub-second timeout in zsystem flock

[Daniel Shahaf <d.s@xxxxxxxxxxxxxxxxxx>]

Cedric Ware wrote on Sun, Mar 15, 2020 at 17:03:24 +0100:
> Daniel Shahaf (Sunday 2020-03-15):
> > Cedric Ware wrote on Sat, 14 Mar 2020 22:04 +0100:
> > > +++ zsh-5.8/Doc/Zsh/mod_system.yo	2020-03-08 18:39:32.672683779 +0100
> > > @@ -166,7 +166,7 @@
> > >  printed in the last case, but the parameter tt(ERRNO) will reflect
> > >  the error that occurred.
> > >  )
> > > -xitem(tt(zsystem flock) [ tt(-t) var(timeout) ] [ tt(-f) var(var) ] [tt(-er)] var(file))
> > > +xitem(tt(zsystem flock) [ tt(-t) var(timeout) ] [ tt(-i) var(interval) ] [ tt(-f) var(var) ] [tt(-er)] var(file))
> > >  item(tt(zsystem flock -u) var(fd_expr))(
> > 
> > Please s/var(var)/var(interval)/ in the body paragraphs of the item()().
> 
> I don't understand.  Option -f's parameter is actually called "var".
> The only occurrences of "var(var)" here are about -f, I didn't add them.

My bad; I'd misread the diff.  Sorry for the noise.

> > > +		timeout_tmp = timeout_param.u.d * 1e6;
> > > +		if ((timeout_tmp < 1) || (timeout_tmp > ZLONG_MAX / 2)) {
> > > +		    zwarnnam(nam, "flock: invalid timeout value");
> > 
> > I think the invalid value should be included in the error message,
> > either as a number, or as a string if needed: "flock: invalid timeout
> > value: '%s'".
> 
> Done, but it had to be as a string, because I didn't see a way to use
> a zlong in zwarnnam().

That doesn't seem to be supported.  It could be added, if needed.

> > dana, I see you advocated the opposite approach in 45283.  What's your
> > rationale?
> 
> I think that was about my earlier attempt to include the actual limit
> in the error message, not the input.  I reverted to a generic message,
> because I don't know how to make the test script check for an error
> message that depends on the system/compiler options.

How would the message depend on the system/compiler options?

In any case, you might be able to address this by specifying the
expectations as patterns:

    # Each line of a '>' and '?' chunk may be preceded by a '*', so the line
    # starts '*>' or '*?'.  This signifies that for any line with '*' in front
    # the actual output should be pattern matched against the corresponding
    # lines in the test output.  Each line following '>' or '?' must be a
    # valid pattern, so characters special to patterns such as parentheses
    # must be quoted with a backslash.  The EXTENDED_GLOB option is used for
    # all such patterns.

> > > +	zlong now;
> > > +	zlong end = time_clock_us() + timeout;
> > 
> > Could this sum overflow (for example, if the -t option is used)?
> 
> I don't think so: timeout is limited to ZLONG_MAX / 2, which
> corresponds to 140000 years or so if zlong is 64-bit.  For the sum
> to overflow, the current time as given by either gettimeofday() or
> clock_gettime(CLOCK_MONOTONIC) would have to also be greater than that.
> 

Sounds good.

> OTOH, if zlong is only 32-bit, there could be a problem, and in fact my
> function time_clock_us() might not work.  However, it was suggested in
> message 45250 that it shouldn't be my problem.  Should it be?  Should
> this whole patch be disabled somehow on systems that don't support any
> 64-bit type?

I think pws was saying there that failure to detect a 64-bit type _when
one is available_ isn't your problem; i.e., that you can trust configure
to detect a 64-bit type when one is available.

How likely is zsh 5.9 to be ported to a platform that doesn't have a
64-bit type?

If that's a concern, I'd recommend to issue a warning or disable the
feature when we detect that case.

I'm not sure whether that's a concern.

> > > +	    if (now + timeout_retry > end) {
> > 
> > Could this sum overflow (for example, if the -i option is used)?
> 
> It might if long and zlong are the same size, as timeout_retry
> is limited to LONG_MAX.  The program would then wait past the timeout
> by one interval.  But, again, this interval would have to be hundreds
> of thousands of years; if a user requests that, in my opinion, we might
> as well just wait forever.

Thanks, agreed.

> Still, would you like it better if I limited the interval to
> min(LONG_MAX, ZLONG_MAX / 2) instead of LONG_MAX?

Well, it sounds like that won't make any difference to bin_system_flock()'s
behaviour in practice (at least until someone has servers with uptimes
on the order of tens of kiloyears), so I don't have a strong preference.
I suppose I'd recommend whichever of these is more likely to remain
correct even if the code is copied elsewhere and adapted.

Thanks,

Daniel