Re: [BUG] arm64: an infinite loop in generic_perform_write()

[Al Viro]

On Thu, Jun 24, 2021 at 04:09:11PM +0100, Catalin Marinas wrote:
> On Thu, Jun 24, 2021 at 12:15:46PM +0100, Matthew Wilcox wrote:
> > On Thu, Jun 24, 2021 at 08:04:07AM +0100, Christoph Hellwig wrote:
> > > On Thu, Jun 24, 2021 at 04:24:46AM +0100, Matthew Wilcox wrote:
> > > > On Thu, Jun 24, 2021 at 11:10:41AM +0800, Chen Huang wrote:
> > > > > In userspace, I perform such operation:
> > > > > 
> > > > >  	fd = open("/tmp/test", O_RDWR | O_SYNC);
> > > > >         access_address = (char *)mmap(NULL, uio_size, PROT_READ, MAP_SHARED, uio_fd, 0);
> > > > >         ret = write(fd, access_address + 2, sizeof(long));
> > > > 
> > > > ... you know that accessing this at unaligned offsets isn't going to
> > > > work.  It's completely meaningless.  Why are you trying to do it?
> > > 
> > > We still should not cause an infinite loop in kernel space due to a
> > > a userspace programmer error.
> > 
> > They're running as root and they've mapped some device memory.  We can't
> > save them from themself.  Imagine if they'd done this to the NVMe BAR.

> We could change raw_copy_from_user() to fall back to 1-byte read in case
> of a fault or fix this corner case in the generic code. A quick hack,
> re-attempting the access with one byte:

No.  If nothing else, iov_iter_single_seg_count() is a bad kludge.
What's more, this "do a single-byte copy" fallback is punishing
a much more common case (memory pressure evicting the page) for the sake
of a corner case specific to one architecture that should've been dealt
with in its raw_copy_from_user().

NAKed-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx>

For some context, see include/linux/uaccess.h and description of requirements
for raw_copy_from_user() in there.