Re: [PATCH] abort: Only flush file-based stdio streams before termination

[Florian Weimer <fweimer at redhat dot com>]

On 08/21/2017 10:58 AM, Andreas Schwab wrote:
> On Aug 21 2017, Florian Weimer <fweimer@redhat.com> wrote:
> 
>>> _IO_cleanup must do locking as well, of course.
>>
>> That means that a process cannot terminate if flockfile on a stream has
>> been called without a matching funlockfile.  I don't think this is
>> permitted by POSIX, and wouldn't be a desirable implementation, either.
> 
> Is it?  That would simply be a programming error.  Since POSIX requires
> locking on stdio I don't see how it can require exit to use no locking.

Is exit a function which references a FILE * object?  What about fflush
(NULL)?

POSIX says this:

“
All functions that reference (FILE *) objects, except those with names
ending in _unlocked, shall behave as if they use flockfile() and
funlockfile() internally to obtain ownership of these (FILE *) objects.
”

I have no idea whether this expresses an intent that only explicit FILE
* references cause locking and, potentially blocking, or if this wording
is the result of a quick specification hack to add thread safety to stdio.

>> In a hypothetical, from-scratch stdio implementation, it should be
>> possible to implement flush-once without locking, but it requires
>> careful ordering of buffer pointer updates (or two locks instead of one).
> 
> I don't think making stdio lock-free is desirable.

But neither is blocking on exit because a thread happens to have called
flockfile on a stdio stream.

Thanks,
Florian