On 18.02.22 17:05, Fujii Masao wrote:
>>> Or even backtrace should be logged by write_stderr() so that it's
>>> written to eventlog if necessary? I just wonder why
>>> backtrace_symbols_fd() is used only in ExceptionalCondition().
>>
>> Probably because it was simpler. It would also make sense to convert
>> the whole thing to use write_stderr() consistently.
> +1
> Attached is the updated version of the patch that uses write_stderr() to
> log the backtrace in assertion failure case.
>
> + if (nframes >= lengthof(buf))
> + appendStringInfo(&errtrace, "\n(backtrace limited to %zu
> frames)",
> + lengthof(buf));
I think this should actually print nframes, because that's the actual
number of frames returned.
> I found this doesn't work on FreeBSD, at least FreeBSD 13 that cfbot
> uses on Cirrus CI. When the backtrace is larger than 100, on FreeBSD,
> backtrace() seems to write the *99* (not 100) most recent function calls
> to the buffer. That is, the variable "nframes" is 99 while lengthof(buf)
> indicates 100. So the above message about backtrace limit will never be
> logged on FreeBSD. OTOH, on Linux and MacOS, backtrace() writes the 100
> most recent function calls. I'm not sure if such a behavior on FreeBSD
> is expected or a bug.
Well, the API is that you ask for up to N frames, and it gives you X
frames back. There is no requirement that X is as close to N as
possible. But that also means that there is no portable API for finding
out how many X it could give you if you had enough room (other than by
looping and doubling the buffer dynamically etc.).
Maybe we can catch this particular case if we made the condition
if (nframes >= lengthof(buf) - 1)
That might give slightly misleading information if you have exactly 99
or 100 frames, but overall it might still be better.