[RFC] Strategy for 9533

[Paolo Carlini]

Hi,

thanking you all again for the feedback, I have a strategy in mind for 
fixing the bug which I'd like to ask your opinion on:

I- Removing all the O_NONBLOCK stuff consequence of 6746.
II- Removing also the underflow call in basic_filebuf::open.
III- Create a new, sort-of-underflow, funtion to be called _only_ upon 
file opening, which basically fills the buffer by way of a O_NOBLOCK 
read system call 
(http://www.opengroup.org/onlinepubs/007908799/xsh/read.html).

I and II above should avoid 9533, III should avoid 6746. All the normal 
operations would be O_NOBLOCK, as suggested by Nathan. Being III 
O_NOBLOCK regressions such as that I reported before today for pipes 
shouln't occur.

What do you think?
Paolo.