wait_until fails with "Invalid argument" when called too soon after reboot

[maflcko]

wait_until accepts a time point that may be in the past. This generally works fine unless the time point is before the machine booted up.

The boost implementation will first calculate the difference to the current time: https://github.com/boostorg/thread/blob/boost-1.70.0/include/boost/thread/pthread/condition_variable_fwd.hpp#L270

Right afterward, the difference is added to the internal clock (detail::internal_chrono_clock::now() + d). However, the internal clock only starts when the machine was booted up (and not when the unix epoch started). Thus, the resulting timeout that is passed to do_wait_until may be negative (in both values, seconds and nanoseconds). Then pthread_cond_timedwait throws "Invalid argument".

[maflcko]

Boost 1.66 works fine, and Boost 1.67 is the first version that is broken.

#167 landed in 1.67, so it might be related.

[maflcko]

Specifically this changed line: https://github.com/boostorg/thread/pull/167/files#diff-54d08155c062cef6acafd6452effd61cL344

[maflcko]

I wrote a test in bitcoin/bitcoin#18232 to illustrate the issue

[fanquake]

Boost 1.66 works fine, and Boost 1.76 is the first version that is broken.

Do you mean Boost 1.72.0 here (latest release)? So the broken range is 1.66.0 to 1.72.0?

[maflcko]

I see two solutions, where at least one is needed to fix the problem:

• Restore the if ( d <= CD::zero() ) return cv_status::timeout;
• Pass a valid timestruct to pthread_cond_timedwait (http://man7.org/linux/man-pages/man3/pthread_cond_timedwait.3p.html says the nsecs value can not be negative: "The abstime argument specified a nanosecond value less than zero or greater than or equal to 1000 million." )

[luke-jr]

I'm not sure the clock is guaranteed to be positive, so I think the latter fix is probably correct?