Hi,
On Fri, Aug 03, 2018 at 03:39:24PM +0200, Oleg Nesterov wrote:
> On 08/02, Dmitry Safonov wrote:
> > 2018-07-31 1:56 GMT+01:00 Ivan Delalande <colona@xxxxxxxxxx>:
> > > We were seeing unexplained segfaults in coreutils processes and other
> > > basic utilities that we tracked down to binfmt_elf failing to load
> > > segments for ld.so. Digging further, the actual problem seems to occur
> > > when a process gets sigkilled while it is still being loaded by the
> > > kernel. In our case when _do_page_fault goes for a retry it will return
> > > early as it first checks for fatal_signal_pending(), so load_elf_interp
> > > also returns with error and as a result search_binary_handler will
> > > force_sigsegv() which is pretty confusing as nothing actually failed
> > > here.
> > >
> > > Fixes: 19d860a140be ("handle suicide on late failure exits in execve() in search_binary_handler()")
> > > Reference: https://lkml.org/lkml/2013/2/14/5
> > > Signed-off-by: Ivan Delalande <colona@xxxxxxxxxx>
> >
> > +Cc: Oleg Nesterov <oleg@xxxxxxxxxx>
> > +Cc: Andy Lutomirski <luto@xxxxxxxxxx>
>
> Thanks...
>
> and sorry, I fail to understand the problem and what/how this patch tries to fix.
>
> Hmm. After I read the next email from Dmitry it seems to me that the whole purpose
> of this patch is to avoid print_fatal_signal()? If yes, the changelog should clearly
> explain this.
Sorry about that, yes this is purely to avoid printing the segfault
messages for these processes when they were in fact killed.
I'll definitely send a v2 to clarify that, and probably add the helpful
message Dimitry suggested as well.
> > > --- a/fs/exec.c
> > > +++ b/fs/exec.c
> > > @@ -1656,7 +1656,8 @@ int search_binary_handler(struct linux_binprm *bprm)
> > > if (retval < 0 && !bprm->mm) {
> > > /* we got to flush_old_exec() and failed after it */
> > > read_unlock(&binfmt_lock);
> > > - force_sigsegv(SIGSEGV, current);
> > > + if (!fatal_signal_pending(current))
> > > + force_sigsegv(SIGSEGV, current);
>
> I won't argue, but may be force_sigsegv() should check fatal_signal_pending()
> itself. setup_rt_frame() can too fail if fatal_signal_pending() by the same
> reason.
I'm not sure, I think it would feel out of place in force_sigsegv() as
other callers might not expect this check in different contexts. I could
add a similar call to fatal_signal_pending() in signal_setup_done()
though, if you think we can hit the same problem from setup_rt_frame().
Thanks,
--
Ivan Delalande
Arista Networks
