Re: [musl] [PATCH] remove extraneous syscall from fopen(3)

["Érico Nogueira" <ericonr@disroot.org>]

On Mon Aug 15, 2022 at 3:16 PM -03, Rich Felker wrote:
> On Mon, Aug 15, 2022 at 02:58:40PM -0300, Érico Nogueira wrote:
> > On Mon Aug 15, 2022 at 2:54 PM -03, Rich Felker wrote:
> > > On Mon, Aug 15, 2022 at 02:50:21PM -0300, Érico Nogueira wrote:
> > > > the __fdopen() call afterwards will set the close-on-exec flag with the
> > > > same syscall if "e" was specified in mode
> > > > ---
> > > >  src/stdio/fopen.c | 2 --
> > > >  1 file changed, 2 deletions(-)
> > > > 
> > > > diff --git a/src/stdio/fopen.c b/src/stdio/fopen.c
> > > > index e1b91e12..22b72edf 100644
> > > > --- a/src/stdio/fopen.c
> > > > +++ b/src/stdio/fopen.c
> > > > @@ -20,8 +20,6 @@ FILE *fopen(const char *restrict filename, const char *restrict mode)
> > > >  
> > > >  	fd = sys_open(filename, flags, 0666);
> > > >  	if (fd < 0) return 0;
> > > > -	if (flags & O_CLOEXEC)
> > > > -		__syscall(SYS_fcntl, fd, F_SETFD, FD_CLOEXEC);
> > > >  
> > > >  	f = __fdopen(fd, mode);
> > > >  	if (f) return f;
> > > > -- 
> > > > 2.37.2
> > >
> > > See commit 7765706c0584ed4a30e0b7a3ada742e490ef02b0
> > 
> > If the relevant part of that commit is that the flag is added
> > immediately after, would moving the SYS_fcntl call in __fdopen to the
> > top of the functon be acceptable?
>
> Oh, I missed that it also happens in __fdopen from the 'e' being
> present, and misunderstood your patch as just removing the fallback
> entirely.
>
> No, it's not acceptable to move the fcntl in __fdopen above the malloc
> because it would make fdopen modify the fd status on failure. I guess
> it's questionable whether we care "how soon" after the open it happens
> -- either way this is not a thread-safe fallback precluding fd leak on
> old/broken kernels. But since malloc may be application-provided,
> failure to set it before the malloc like we're doing now would be a
> "worse behavior" in some sense, exposing the incorrect fd state to a
> non-multithreaded application.

On some level, unless someone inherited a file descriptor or something
similar, I'd expect them to have used O_CLOEXEC if they are also using
"e" in mode. So hopefully this is not as much of a concern.

And I don't think fdopen setting the close-on-exec flag is behavior
users can rely on, seeing as glibc doesn't take "e" into account in
their fdopen implementation.

> So I'm not sure if it's a good idea to
> change this or not. Do you have reason to believe it's affecting
> performance in real-world usage?

From what testing I have done, a fcntl() call is essentially free, at
least when compared to the cost of open(). This commit was intended only
as cleanup.

>
> Rich