Re: [PATCH 0/6] use memcpy_mcsafe() for copy_to_iter()

[Andy Lutomirski]

On Tue, May 1, 2018 at 8:34 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx>
wrote:

> On Tue, May 1, 2018 at 8:22 PM Dan Williams <dan.j.williams@xxxxxxxxx>
> wrote:

> > All that to say that having a typical RAM page covering poisoned pmem
> > would complicate the 'clear badblocks' implementation.

> Ugh, ok.

> I guess the good news is that your patches aren't so big, and don't really
> affect anything else.


I pondered this a bit.  Doing better might be a big pain in the arse.  The
interesting case is where ordinary kernel code (memcpy, plain old memory
operands, etc) access faulty pmem.  This means that there's no extable
entry around.  If we actually try to recover, we have a few problems:

  - We can't sanely skip the instruction without causing random errors.

  - If the access was through the kernel direct map, then we could plausibly
remap a different page in place of the faulty page.  The problem is that,
if the page is *writable* and we share it between more than one faulty
page, then we're enabling a giant information leak.  But we still need to
figure out how we're supposed to invalidate the old mapping from a random,
potentially atomic context.

  - If the access is through kmap or similar, then we're talking about
modifying a PTE out from under kernel code that really isn't expecting us
to modify it.

  - How are we supposed to signal the process or fail a syscall?  The fault
could have come from interrupt context, softirq context, kernel thread
context, etc, and figuring out who's to blame seems quite awkward and
fragile.

All that being said, I suspect that we still have issues even with accesses
to user VAs that are protected by extable entries.  The whole #MC mechanism
is a supremely shitty interface for recoverable errors (especially on
Intel), and I'm a bit scared of what happens if the offending access is,
say, inside a perf NMI.

Dan, is there any chance you could put some pressure on the architecture
folks to invent an entirely new, less shitty way to tell the OS about
recoverable memory errors?  And to make it testable by normal people?
Needing big metal EINJ hardware to test the house of cards that is #MC is
just awful and means that there are few enough kernel developers that are
actually able to test that I can probably count them on one hand.  And I'm
not one of them...