Re: [PATCH] /dev/mem: Bail out upon SIGKILL when reading memory.

From: Tetsuo Handa
Date: Thu Aug 22 2019 - 09:45:19 EST


Tetsuo Handa wrote:
> Greg Kroah-Hartman wrote:
> > Oh, nice! This shouldn't break anything that is assuming that the read
> > will complete before a signal is delivered, right?
> >
> > I know userspace handling of "short" reads is almost always not there...
>
> Since this check will give up upon SIGKILL, userspace won't be able to see
> the return value from read(). Thus, returning 0 upon SIGKILL will be safe. ;-)
> Maybe we also want to add cond_resched()...
>
> By the way, do we want similar check on write_mem() side?
> If aborting "write to /dev/mem" upon SIGKILL (results in partial write) is
> unexpected, we might want to ignore SIGKILL for write_mem() case.
> But copying data from killed threads (especially when killed by OOM killer
> and userspace memory is reclaimed by OOM reaper before write_mem() returns)
> would be after all unexpected. Then, it might be preferable to check SIGKILL
> on write_mem() side...
>

Ha, ha. syzbot reported the same problem using write_mem().
https://syzkaller.appspot.com/text?tag=CrashLog&x=1018055a600000
We want fatal_signal_pending() check on both sides.

By the way, write_mem() worries me whether there is possibility of replacing
kernel code/data with user-defined memory data supplied from userspace.
If write_mem() were by chance replaced with code that does

while (1);

we won't be able to return from write_mem() even if we added fatal_signal_pending() check.
Ditto for replacing local variables with unexpected values...