On December 5, 2022 8:24:41 PM PST, Masahiro Yamada <masahiroy@xxxxxxxxxx> wrote:
>On Thu, Nov 17, 2022 at 7:07 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>>
>> 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
>
>
>
>BTW, Python does something similar by default.
>(noisy back-trace for SIGPIPE)
>
>
>
>
>
>masahiro@zoe:/tmp$ cat test.py
>#!/usr/bin/python3
>for i in range(4000):
> print(i)
>
>masahiro@zoe:/tmp$ ./test.py | head -n1
>0
>Traceback (most recent call last):
> File "/tmp/./test.py", line 3, in <module>
> print(i)
>BrokenPipeError: [Errno 32] Broken pipe
Eww. Well, same problem, IMO. For any Python scripts that are going to have potentially truncated output, they need to do:
from signal import signal, SIGPIPE, SIG_DFL
signal(SIGPIPE,SIG_DFL)
>This page
>https://www.geeksforgeeks.org/broken-pipe-error-in-python/
>
>suggests some workarounds.
(As suggested in this page.)
>What would you suggest for python scripts?
They need to be fixed. A command line tool internally catching SIGPIPE is wrong. :)
-Kees
--
Kees Cook
