[PATCH 1/3] arm/syscalls: Move address limit check in loop

[will.deacon@arm.com (Will Deacon)]

On Tue, Jul 25, 2017 at 01:01:17PM -0700, Thomas Garnier wrote:
> On Tue, Jul 25, 2017 at 3:38 AM, Russell King - ARM Linux
> <linux@armlinux.org.uk> wrote:
> > On Tue, Jul 25, 2017 at 01:28:01PM +0300, Leonard Crestez wrote:
> >> On Mon, 2017-07-24 at 10:07 -0700, Thomas Garnier wrote:
> >> > On Wed, Jul 19, 2017 at 10:58 AM, Thomas Garnier <thgarnie@google.com
> >> > > wrote:
> >> > >
> >> > > The work pending loop can call set_fs after addr_limit_user_check
> >> > > removed the _TIF_FSCHECK flag. To prevent the infinite loop, move
> >> > > the addr_limit_user_check call at the beginning of the loop.
> >> > >
> >> > > Fixes: 73ac5d6a2b6a ("arm/syscalls: Check address limit on user-
> >> > > mode return")
> >> > > Reported-by: Leonard Crestez <leonard.crestez@nxp.com>
> >> > > Signed-off-by: Thomas Garnier <thgarnie@google.com>
> >>
> >> > Any comments on this patch set?
> >>
> >> Tested-by: Leonard Crestez <leonard.crestez@nxp.com>
> >>
> >> This appears to fix the original issue of failing to boot from NFS when
> >> there are lots of alignment faults. But this is a very basic test
> >> relative to the reach of this change.
> >>
> >> However the original patch has been in linux-next for a while and
> >> apparently nobody else noticed system calls randomly hanging on arm.
> >>
> >> I assume maintainers need to give their opinion.
> >
> > I've already stated my opinion, which is different from what Linus has
> > requested of Thomas.  IMHO, the current approach is going to keep on
> > causing problems along the lines that I've already pointed out.
> 
> I understand. Do you think this problem apply to arm64 as well?

It's probably less of an issue for arm64 because we don't take alignment
faults from the kernel and I think the perf case would resolve itself by
throttling the event. However, I also don't see the advantage of doing
this in the work loop as opposed to leaving it until we're actually doing
the return to userspace.

I looked to see what you've done for x86, but it looks like you check/clear
the flag before the work pending loop (exit_to_usermode_loop), which
subsequently re-enables interrupts and exits when
EXIT_TO_USERMODE_LOOP_FLAGS are all clear. Since TIF_FSCHECK isn't included
in those flags, what stops it being set again by an irq and remaining set
for the return to userspace?

Will