Re: [Bug-tar] opening files with O_NONBLOCK causes problems

[Kevin Fox]

On Fri, 2012-01-06 at 06:08 -0800, Eric Blake wrote:
> On 01/06/2012 06:52 AM, Ron Kerry wrote:
> >>> Why should gtar open FIFO files?
> >>
> >> The question is not why an archiver opens a FIFO file, but what it does
> >> after opening a file O_NONBLOCK (the TOCTTOU race is eliminated by
> >> switching stat()/open() to open()/fstat() filtering, and once we have
> >> ascertained that an open fd is not a FIFO, if we can then use fcntl() to
> >> remove the O_NONBLOCK, hopefully that will resolve the situation with
> >> DMF).
> >>
> > 
> > I am a bit lost in this discussion.  As far as I know, O_NONBLOCK has no
> > effect whatsoever on an open() system call,
> 
> Wrong.  Per POSIX, O_NONBLOCK on a FIFO controls whether the
> open(O_RDONLY) blocks until a writer is present, or succeeds right away;

On another HSM, HPSS, O_NONBLOCK on an open has a similar affect.
Opening a file that isn't on disk will cause the open to block until it
is retrieved off of tape. Setting O_NONBLOCK on open will let the open
succeed right away.

Kevin

> likewise, open(O_WRONLY) uses O_NONBLOCK to fail to open if there is no
> reader.  Which means that if you open a FIFO with no writer, you've
> blocked the program unless you used O_NONBLOCK.  (The same thing goes
> for block and character devices that support non-blocking operation, as
> well as any implementation extension file types that support
> non-blocking; apparently DMF is such an implementation where even
> regular files support non-blocking).
> 
> POSIX states that the use of O_NONBLOCK on non-FIFO files (more
> correctly, files that don't support non-blocking operation) has
> unspecified effect on whether fcntl() can observe the bit set, and if
> the bit is set, unspecified effects on what it does to reads and writes
> on that fd: http://austingroupbugs.net/view.php?id=141
> 
> So the question at hand is avoiding the block (using O_NONBLOCK during
> open()) while avoiding the unspecified behavior (clearing O_NONBLOCK on
> regular files before reading them).
> 
> > it only has an effect when
> > you do a read or a write.  So you open the file without O_NONBLOCK, you
> > fstat the file, and if it's a pipe you close it. I do not see the need
> > for using O_NONBLOCK.
> 
> Use of O_NONBLOCK is necessary to avoid hanging while waiting for a
> writer, before you ever get a chance to fstat() the fd to see if it was
> a FIFO to close it.
>