Re: Clearing O_NONBLOCK from a pipe may lose data

[Corinna Vinschen <corinna-cygwin at cygwin dot com>]

On Feb 20 08:56, Thomas Wolff wrote:
> 
> Am 20.02.2015 um 00:47 schrieb Andrey Repin:
> >Greetings, Thomas Wolff!
> >
> >>Am 19.02.2015 um 10:51 schrieb Corinna Vinschen:
> >>>On Feb 18 22:08, Lasse Collin wrote:
> >>>>(Please Cc me when replying, I'm not subscribed to the list.)
> >>>>
> >>>>Hi!
> >>>>
> >>>>I suspect that there is a bug in Cygwin:
> >>>>
> >>>>1. Create a pipe with both ends in blocking mode (O_NONBLOCK
> >>>>     is not set).
> >>>>2. The writer sets its end to non-blocking mode.
> >>>>3. The writer writes to the pipe.
> >>>>4. The writer restores its end of the pipe to blocking mode
> >>>>     before the reader has read anything from the pipe.
> >>>>5. The writer closes its end of the pipe.
> >>>>6. The reader reads from the pipe in blocking mode. The last
> >>>>     bytes written by the writer never appear at the reader,
> >>>>     thus data is silently lost.
> >>>>
> >>>>Omitting the step 4 above makes the problem go away.
> >>>I can imagine.  A few years back, when changing the pipe code to
> >>>using overlapped IO, we stumbled over a problem in Windows.  When
> >>>closing an overlapped pipe while I/O is still ongoing, Windows
> >>>simply destroys the pipe buffers without flushing the data to the
> >>>reader.  This is not much of a problem for blocking IO, but it
> >>>obviously is for non-blocking.
> >>>
> >>>The workaround for this behaviour is this:  If the pipe is closed, and
> >>>this is the writing side of a nonblocking pipe, a background thread gets
> >>>started which keeps the overlapped structure open and continues to wait
> >>>for IO completion (i.e. the data has been sent to the reader).
> >>>
> >>>However, if you switch back to blocking before closing the pipe, the
> >>>aforementioned mechanism does not kick in.
> >>Could not "switching back to blocking" simply be handled like closing as
> >>far as the waiting is concerned,
> >>thus effectively flushing the pipe buffer?
> >You can't "just flush" it, if the receiving end isn't reading from it.
> By flushing I meant actually waiting until it's been consumed at the
> other end in this case, if that's technically feasible.

You mean the actual act of changing the descriptor from non-blocking
to blocking, as in fcntl(fd, F_SETFL), shall perform the same action
of waiting as the close call on non-blocking descriptors does?

> I see no strict requirement that the fcntl call removing O_NONBLOCK from
> a file descriptor should itself still be handled as nonblocking (it can
> well be argued that the flag is changed first and then the call is
> allowed to block) - and even if this were not proper it is certainly
> more acceptable than losing data.

I'm not sure that works as desired, but it's probably worth a try.  An
fcntl method for pipes has to be added (there is none yet, it's all done
in fhandler_base::fcntl), then the F_SETFL command would have to be
augmented to create a thread calling FlushFileBuffers (which is
*supposed* to work on pipe handles but I never tried it myself), and the
fcntl call would have to wait for thread completion, allowing
interruption by signals (calling cygwait, that is).

The question with stuff like this is usually, how long are you willing
to wait?  You never know what the reader side of a pipe is doing.  It
might just be busy and intends to read from the pipe in a second, a
minute, or an hour.


Corinna

-- 
Corinna Vinschen                  Please, send mails regarding Cygwin to
Cygwin Maintainer                 cygwin AT cygwin DOT com
Red Hat

Attachment: pgprfChCAFbRT.pgp
Description: PGP signature