[PATCH] powerpc/fault: fix wrong KUAP fault for IO_URING

[Zorro Lang]

On Wed, Jan 27, 2021 at 08:06:37PM -0700, Jens Axboe wrote:
> On 1/27/21 8:13 PM, Zorro Lang wrote:
> > On Thu, Jan 28, 2021 at 10:18:07AM +1000, Nicholas Piggin wrote:
> >> Excerpts from Jens Axboe's message of January 28, 2021 5:29 am:
> >>> On 1/27/21 9:38 AM, Christophe Leroy wrote:
> >>>>
> >>>>
> >>>> Le 27/01/2021 à 15:56, Zorro Lang a écrit :
> >>>>> On powerpc, io_uring test hit below KUAP fault on __do_page_fault.
> >>>>> The fail source line is:
> >>>>>
> >>>>>    if (unlikely(!is_user && bad_kernel_fault(regs, error_code, address, is_write)))
> >>>>>        return SIGSEGV;
> >>>>>
> >>>>> The is_user() is based on user_mod(regs) only. This's not suit for
> >>>>> io_uring, where the helper thread can assume the user app identity
> >>>>> and could perform this fault just fine. So turn to use mm to decide
> >>>>> if this is valid or not.
> >>>>
> >>>> I don't understand why testing is_user would be an issue. KUAP purpose
> >>>> it to block any unallowed access from kernel to user memory
> >>>> (Equivalent to SMAP on x86). So it really must be based on MSR_PR bit,
> >>>> that is what is_user provides.
> >>>>
> >>>> If the kernel access is legitimate, kernel should have opened
> >>>> userspace access then you shouldn't get this "Bug: Read fault blocked
> >>>> by KUAP!".
> >>>>
> >>>> As far as I understand, the fault occurs in
> >>>> iov_iter_fault_in_readable() which calls fault_in_pages_readable() And
> >>>> fault_in_pages_readable() uses __get_user() so it is a legitimate
> >>>> access and you really should get a KUAP fault.
> >>>>
> >>>> So the problem is somewhere else, I think you proposed patch just
> >>>> hides the problem, it doesn't fix it.
> >>>
> >>> If we do kthread_use_mm(), can we agree that the user access is valid?
> >>
> >> Yeah the io uring code is fine, provided it uses the uaccess primitives 
> >> like any other kernel code. It's looking more like a an arch/powerpc bug.
> >>
> >>> We should be able to copy to/from user space, and including faults, if
> >>> that's been done and the new mm assigned. Because it really should be.
> >>> If SMAP was a problem on x86, we would have seen it long ago.
> >>>
> >>> I'm assuming this may be breakage related to the recent uaccess changes
> >>> related to set_fs and friends? Or maybe recent changes on the powerpc
> >>> side?
> >>>
> >>> Zorro, did 5.10 work?
> >>
> >> Would be interesting to know.
> > 
> > Sure Nick and Jens, which 5.10 rc? version do you want to know ? Or any git
> > commit(be the HEAD) in 5.10 phase?
> 
> I forget which versions had what series of this, but 5.10 final - and if
> that fails, then 5.9 final. IIRC, 5.9 was pre any of these changes, and
> 5.10 definitely has them.

I justed built linux v5.10 with same .config file, and gave it same test.
v5.10 (HEAD=2c85ebc57b Linux 5.10) can't reproduce this bug:

# ./check generic/013 generic/051
FSTYP         -- xfs (non-debug)
PLATFORM      -- Linux/ppc64le ibm-p9z-xxx-xxxx 5.10.0 #3 SMP Thu Jan 28 04:12:14 EST 2021
MKFS_OPTIONS  -- -f -m crc=1,finobt=1,reflink=1,rmapbt=1,bigtime=1,inobtcount=1 /dev/sda3
MOUNT_OPTIONS -- -o context=system_u:object_r:root_t:s0 /dev/sda3 /mnt/xfstests/scratch

generic/013 138s ...  77s
generic/051 103s ...  143s
Ran: generic/013 generic/051
Passed all 2 tests

> 
> -- 
> Jens Axboe
>