libstdc++/9533: Regression: Can't read from tty with ifstream

[Paolo Carlini]

Pétur Runólfsson wrote:

>>    Patching: indeed, reverting this hunk of 6746 commit:
>>    
>>    +
>>    +	    // Set input to nonblocking for fifos.
>>    +	    if (__mode & ios_base::in)
>>    +	      fcntl(this->fd(), F_SETFL, O_NONBLOCK);
>>    +
>>    
>>    Fixes the problem without regressing on 6746.
>>    
>>
>Interesting. I would have thought that removing this would
>cause the call to underflow() in basic_filebuf::open() to
>block, causing the test case to wait for input *before*
>printing out the prompt (this seems to be the reason that
>non-blocking input is specified).
>
Yes, you are right. My bad.
I didn't notice that since was using a further reduced :( testcase.
Also, we got a regression elsewhere by simply reverting that hunk (not 
on 6746): see
the libstdc++ list.

>So far, this is what I think is going on:
>* non-blocking input is specified so underflow() doesn't
>  block (which is wrong, underflow() should block).
>* underflow() must not block because it is called from
>  open() which must not block.
>* open() calls underflow() so that in_avail() will return
>  non-zero after open() (libstdc++/6746).
>
Agreed.

>However, I don't see how calling underflow() from open()
>fixes libstdc++/6746.
>
I concur, but in fact it does.

> Unless I am missing something,
>readsome will return up to BUFSIZ characters, and then
>return 0 until the end of time.
>  
>
Paolo.