[PATCH 3/3] btrfs: Avoid live-lock in search_ioctl() on hardware with sub-page faults

Catalin Marinas catalin.marinas at arm.com
Mon Nov 29 07:36:41 PST 2021


On Mon, Nov 29, 2021 at 02:33:42PM +0100, Andreas Gruenbacher wrote:
> On Mon, Nov 29, 2021 at 1:22 PM Catalin Marinas <catalin.marinas at arm.com> wrote:
> > On Sat, Nov 27, 2021 at 07:05:39PM +0100, Andreas Gruenbacher wrote:
> > > We also still have fault_in_safe_writeable which is more difficult to
> > > fix, and fault_in_readable which we don't want to leave behind broken,
> > > either.
> >
> > fault_in_safe_writeable() can be done by using get_user() instead of
> > put_user() for arm64 MTE and probably SPARC ADI (an alternative is to
> > read the in-memory tags and compare them with the pointer).
> 
> So we'd keep the existing fault_in_safe_writeable() logic for the
> actual fault-in and use get_user() to check for sub-page faults? If
> so, then that should probably also be hidden in arch code.

That's what this series does when it probes the whole range in
fault_in_writeable(). The main reason was that it's more efficient to do
a read than a write on a large range (the latter dirtying the cache
lines).

> > For CHERI, that's different again since the fault_in_safe_writeable capability
> > encodes the read/write permissions independently.
> >
> > However, do we actually want to change the fault_in_safe_writeable() and
> > fault_in_readable() functions at this stage? I could not get any of them
> > to live-lock, though I only tried btrfs, ext4 and gfs2. As per the
> > earlier discussion, normal files accesses are guaranteed to make
> > progress. The only problematic one was O_DIRECT which seems to be
> > alright for the above filesystems (the fs either bails out after several
> > attempts or uses GUP to read which skips the uaccess altogether).
> 
> Only gfs2 uses fault_in_safe_writeable(). For buffered reads, progress
> is guaranteed because failures are at a byte granularity.
> 
> O_DIRECT reads and writes happen in device block size granularity, but
> the pages are grabbed with get_user_pages() before the copying
> happens. So by the time the copying happens, the pages are guaranteed
> to be resident, and we don't need to loop around fault_in_*().

For file reads, I couldn't triggered any mismatched tag faults with gfs2
and O_DIRECT, so it matches your description above. For file writes it
does trigger such faults, so I suspect it doesn't always use
get_user_pages() for writes. No live-lock though with the vanilla
kernel. My test uses a page with some mismatched tags in the middle.

ext4: no tag faults with O_DIRECT read/write irrespective of whether the
user buffer is page aligned or not.

btrfs: O_DIRECT file writes - no faults on page-aligned buffers, faults
on unaligned; file reads - tag faults on both aligned/unaligned buffers.
No live-lock.

So, some tag faults still happen even with O_DIRECT|O_SYNC but the
filesystems too care of continuous faulting.

> You've mentioned before that copying to/from struct page bypasses
> sub-page fault checking. If that is the case, then the checking
> probably needs to happen in iomap_dio_bio_iter and dio_refill_pages
> instead.

It's too expensive and not really worth it. With a buffered access, the
uaccess takes care of checking at the time of load/store (the hardware
does this for us). With a GUP, the access is done via the kernel mapping
with a match-all tag to avoid faults (kernel panic). We set the ABI
expectation some time ago that kernel accesses to user memory may not
always be tag-checked if the access is done via a GUP'ed page.

> > Happy to address them if there is a real concern, I just couldn't trigger it.
> 
> Hopefully it should now be clear why you couldn't. One way of
> reproducing with fault_in_safe_writeable() would be to use that in
> btrfs instead of fault_in_writeable(), of course.

Yes, that would trigger it again. I guess if we want to make this API
safer in general, we can add the checks to the other functions. Only
probing a few bytes at the start shouldn't cause a performance issue.

Thanks.

-- 
Catalin



More information about the linux-arm-kernel mailing list