Re: [PATCH] virtiofsd: Add `sigreturn` to the seccomp whitelist

[Christian Borntraeger]

Am 29.11.22 um 10:52 schrieb Christian Borntraeger:
> Am 29.11.22 um 10:42 schrieb Dr. David Alan Gilbert:
> > * Marc Hartmayer (mhartmay@linux.ibm.com) wrote:
> > > "Dr. David Alan Gilbert" <dgilbert@redhat.com> writes:
> > >
> > > > * Marc Hartmayer (mhartmay@linux.ibm.com) wrote:
> > > > > The virtiofsd currently crashes on s390x. This is because of a
> > > > > `sigreturn` system call. See audit log below:
> > > > >
> > > > > type=SECCOMP msg=audit(1669382477.611:459): auid=4294967295 uid=0 gid=0 ses=4294967295 
> > > > > subj=system_u:system_r:virtd_t:s0-s0:c0.c1023 pid=6649 comm="virtiofsd" exe="/usr/libexec/virtiofsd" sig=31 
> > > > > arch=80000016 syscall=119 compat=0 ip=0x3fff15f748a code=0x80000000AUID="unset" UID="root" 
> > > > > GID="root" ARCH=s390x SYSCALL=sigreturn
> > > >
> > > > I'm curious; doesn't that mean that some signal is being delivered and
> > > > you're returning?  Which one?
> > >
> > > code=0x80000000 means that the seccomp action SECCOMP_RET_KILL_PROCESS
> > > is taken => process is killed by a SIGSYS signal (31) [1].
> > >
> > > At least, that’s my understanding of this log message.
> > >
> > > [1] https://man7.org/linux/man-pages/man2/seccomp.2.html
> >
> > But isn't that the fallout rather than the cause ? i.e. seccomp
> > is sending a SIGSYS because the process used sigreturn, my question
> > is why did the process call sigreturn in the first place - it must
> > have received a signal to return from?
>
> Good question. virtiofsd seems to prepare itself for
>
> int fuse_set_signal_handlers(struct fuse_session *se)
> {
>      /*
>       * If we used SIG_IGN instead of the do_nothing function,
>       * then we would be unable to tell if we set SIG_IGN (and
>       * thus should reset to SIG_DFL in fuse_remove_signal_handlers)
>       * or if it was already set to SIG_IGN (and should be left
>       * untouched.
>       */
>      if (set_one_signal_handler(SIGHUP, exit_handler, 0) == -1 ||
>          set_one_signal_handler(SIGINT, exit_handler, 0) == -1 ||
>          set_one_signal_handler(SIGTERM, exit_handler, 0) == -1 ||
>          set_one_signal_handler(SIGPIPE, do_nothing, 0) == -1) {
>          return -1;
>      }
>
>
>
> Given that rt_sigreturn was already on the seccomp list it seems
> to be expected that those handlers are called.

For me, it seems to happen on shutdown:
                Stack trace of thread 1:
                #0  0x000003ffc06f348a __kernel_sigreturn (linux-vdso64.so.1 + 
0x48a)
                #1  0x000003ffc06f3488 __kernel_sigreturn (linux-vdso64.so.1 + 
0x488)
                #2  0x000003ff9af1be96 __GI___futex_abstimed_wait_cancelable64 
(libc.so.6 + 0x9be96)
                #3  0x000003ff9af211b4 __pthread_clockjoin_ex (libc.so.6 + 
0xa11b4)
                #4  0x000003ff9af2106e pthread_join@GLIBC_2.2 (libc.so.6 + 
0xa106e)
                #5  0x000002aa35d2fe36 fv_queue_cleanup_thread (virtiofsd + 
0x2fe36)
                #6  0x000002aa35d3152c stop_all_queues (virtiofsd + 0x3152c)
                #7  0x000002aa35d2869c main (virtiofsd + 0x2869c)
                #8  0x000003ff9aeb4872 __libc_start_call_main (libc.so.6 + 
0x34872)
                #9  0x000003ff9aeb4950 __libc_start_main@@GLIBC_2.34 (libc.so.6 
+ 0x34950)
                #10 0x000002aa35d290a0 .annobin_libvhost_user.c_end.startup 
(virtiofsd + 0x290a0)