[Cluster-devel] [PATCH v7 05/19] iov_iter: Introduce fault_in_iov_iter_writeable

[Thomas Gleixner]

On Fri, Aug 27 2021 at 16:22, Tony Luck wrote:
> On Fri, Aug 27, 2021 at 09:57:10PM +0000, Al Viro wrote:
>> On Fri, Aug 27, 2021 at 09:48:55PM +0000, Al Viro wrote:
>> 
>> > 	[btrfs]search_ioctl()
>> > Broken with memory poisoning, for either variant of semantics.  Same for
>> > arm64 sub-page permission differences, I think.
>> 
>> 
>> > So we have 3 callers where we want all-or-nothing semantics - two in
>> > arch/x86/kernel/fpu/signal.c and one in btrfs.  HWPOISON will be a problem
>> > for all 3, AFAICS...
>> > 
>> > IOW, it looks like we have two different things mixed here - one that wants
>> > to try and fault stuff in, with callers caring only about having _something_
>> > faulted in (most of the users) and one that wants to make sure we *can* do
>> > stores or loads on each byte in the affected area.
>> > 
>> > Just accessing a byte in each page really won't suffice for the second kind.
>> > Neither will g-u-p use, unless we teach it about HWPOISON and other fun
>> > beasts...  Looks like we want that thing to be a separate primitive; for
>> > btrfs I'd probably replace fault_in_pages_writeable() with clear_user()
>> > as a quick fix for now...
>> > 
>> > Comments?
>> 
>> Wait a sec...  Wasn't HWPOISON a per-page thing?  arm64 definitely does have
>> smaller-than-page areas with different permissions, so btrfs search_ioctl()
>> has a problem there, but arch/x86/kernel/fpu/signal.c doesn't have to deal
>> with that...
>> 
>> Sigh...  I really need more coffee...
>
> On Intel poison is tracked at the cache line granularity. Linux
> inflates that to per-page (because it can only take a whole page away).
> For faults triggered in ring3 this is pretty much the same thing because
> mm/memory_failure.c unmaps the page ... so while you see a #MC on first
> access, you get #PF when you retry. The x86 fault handler sees a magic
> signature in the page table and sends a SIGBUS.
>
> But it's all different if the #MC is triggerd from ring0. The machine
> check handler can't unmap the page. It just schedules task_work to do
> the unmap when next returning to the user.
>
> But if your kernel code loops and tries again without a return to user,
> then your get another #MC.

But that's not the case for restore_fpregs_from_user() when it hits #MC.

restore_fpregs_from_user()
  ...
  ret = __restore_fpregs_from_user(buf, xrestore, fx_only)
  
  /* Try to handle #PF, but anything else is fatal. */
  if (ret != -EFAULT)
     return -EINVAL;

Now let's look at __restore_fpregs_from_user()

__restore_fpregs_from_user()
   return $FPUVARIANT_rstor_from_user_sigframe()

which all end up in user_insn(). user_insn() returns 0 or the negated
trap number, which results in -EFAULT for #PF, but for #MC the negated
trap number is -18 i.e. != -EFAULT. IOW, there is no endless loop.

This used to be a problem before commit:

  aee8c67a4faa ("x86/fpu: Return proper error codes from user access functions")

and as the changelog says the initial reason for this was #GP going into
the fault path, but I'm pretty sure that I also discussed the #MC angle with
Borislav back then. Should have added some more comments there
obviously.

Thanks,

        tglx