On Thu, Nov 17, 2022 at 05:37:31AM +0900, Masahiro Yamada wrote:
> On Thu, Nov 17, 2022 at 4:01 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > On Fri, Oct 28, 2022 at 01:28:39AM +0900, Masahiro Yamada wrote:
> > > Jiri Slaby reported that building the kernel with AR=gcc-ar shows:
> > > /usr/bin/ar terminated with signal 13 [Broken pipe]
> > >
> > > Nathan Chancellor reported the latest AR=llvm-ar shows
> > > error: write on a pipe with no reader
> > >
> > > The latter occurs since LLVM commit 51b557adc131 ("Add an error message
> > > to the default SIGPIPE handler").
> > >
> > > The resulting vmlinux is correct, but it is better to silence it.
> > >
> > > 'head -n1' exits after reading the first line, so the pipe is closed.
> > >
> > > Use 'sed -n 1p' to eat the stream till the end.
> >
> > I think this is wrong because it needlessly consumes CPU time. SIGPIPE
> > is _needed_ to stop a process after we found what we needed, but it's up
> > to the caller (the shell here) to determine what to do about it.
> >
> > Similarly, that LLVM commit is wrong -- tools should _not_ catch their
> > own SIGPIPEs. They should be caught by their callers.
> >
> > For example, see:
> >
> > $ seq 10000 | head -n1
> > 1
> >
> > ^^^ no warnings from the shell (caller of "seq")
> > And you can see it _is_ being killed by SIGPIPE:
> >
> > $ strace seq 1000 | head -n1
> > ...
> > write(1, "1\n2\n3\n4\n5\n6\n7\n8\n9\n10\n11\n12\n13\n14"..., 8192) = 8192
> > 1
> > write(1, "\n1861\n1862\n1863\n1864\n1865\n1866\n1"..., 4096) = -1 EPIPE (Broken pipe)
> > --- SIGPIPE {si_signo=SIGPIPE, si_code=SI_USER, si_pid=3503448, si_uid=1000} ---
> > +++ killed by SIGPIPE +++
> >
> > If we use "sed -n 1p" seq will continue to run, consuming needless time
> > and CPU resources.
> >
> > So, I strongly think this is the wrong solution. SIGPIPE should be
> > ignored for ar, and LLVM should _not_ catch its own SIGPIPE.
> >
> > -Kees
>
>
> I thought of this - it is just wasting CPU time,
> but I did not come up with a better idea on the kbuild side.
>
> I do not want to use 2>/dev/null because it may hide
> non-SIGPIPE (i.e. real) errors.
Yes, I've opened an upstream LLVM bug for this:
https://github.com/llvm/llvm-project/issues/59037
--
Kees Cook
