Skip to content

L
linux
Switch branch/tag
  1. 21 Dec, 2011 3 commits
    • Steven Rostedt's avatar
      x86: Add workaround to NMI iret woes · 3f3c8b8c
      Steven Rostedt authored 
      In x86, when an NMI goes off, the CPU goes into an NMI context that
prevents other NMIs to trigger on that CPU. If an NMI is suppose to
trigger, it has to wait till the previous NMI leaves NMI context.
At that time, the next NMI can trigger (note, only one more NMI will
trigger, as only one can be latched at a time).

The way x86 gets out of NMI context is by calling iret. The problem
with this is that this causes problems if the NMI handle either
triggers an exception, or a breakpoint. Both the exception and the
breakpoint handlers will finish with an iret. If this happens while
in NMI context, the CPU will leave NMI context and a new NMI may come
in. As NMI handlers are not made to be re-entrant, this can cause
havoc with the system, not to mention, the nested NMI will write
all over the previous NMI's stack.

Linus Torvalds proposed the following workaround to this problem:

https://lkml.org/lkml/2010/7/14/264

"In fact, I wonder if we couldn't just do a software NMI disable
instead? Hav ea per-cpu variable (in the _core_ percpu areas that get
allocated statically) that points to the NMI stack frame, and just
make the NMI code itself do something like

 NMI entry:
 - load percpu NMI stack frame pointer
 - if non-zero we know we're nested, and should ignore this NMI:
    - we're returning to kernel mode, so return immediately by using
"popf/ret", which also keeps NMI's disabled in the hardware until the
"real" NMI iret happens.
    - before the popf/iret, use the NMI stack pointer to make the NMI
return stack be invalid and cause a fault
  - set the NMI stack pointer to the current stack pointer

 NMI exit (not the above "immediate exit because we nested"):
   clear the percpu NMI stack pointer
   Just do the iret.

Now, the thing is, now the "iret" is atomic. If we had a nested NMI,
we'll take a fault, and that re-does our "delayed" NMI - and NMI's
will stay masked.

And if we didn't have a nested NMI, that iret will now unmask NMI's,
and everything is happy."

I first tried to follow this advice but as I started implementing this
code, a few gotchas showed up.

One, is accessing per-cpu variables in the NMI handler.

The problem is that per-cpu variables use the %gs register to get the
variable for the given CPU. But as the NMI may happen in userspace,
we must first perform a SWAPGS to get to it. The NMI handler already
does this later in the code, but its too late as we have saved off
all the registers and we don't want to do that for a disabled NMI.

Peter Zijlstra suggested to keep all variables on the stack. This
simplifies things greatly and it has the added benefit of cache locality.

Two, faulting on the iret.

I really wanted to make this work, but it was becoming very hacky, and
I never got it to be stable. The iret already had a fault handler for
userspace faulting with bad segment registers, and getting NMI to trigger
a fault and detect it was very tricky. But for strange reasons, the system
would usually take a double fault and crash. I never figured out why
and decided to go with a simple "jmp" approach. The new approach I took
also simplified things.

Finally, the last problem with Linus's approach was to have the nested
NMI handler do a ret instead of an iret to give the first NMI NMI-context
again.

The problem is that ret is much more limited than an iret. I couldn't figure
out how to get the stack back where it belonged. I could have copied the
current stack, pushed the return onto it, but my fear here is that there
may be some place that writes data below the stack pointer. I know that
is not something code should depend on, but I don't want to chance it.
I may add this feature later, but for now, an NMI handler that loses NMI
context will not get it back.

Here's what is done:

When an NMI comes in, the HW pushes the interrupt stack frame onto the
per cpu NMI stack that is selected by the IST.

A special location on the NMI stack holds a variable that is set when
the first NMI handler runs. If this variable is set then we know that
this is a nested NMI and we process the nested NMI code.

There is still a race when this variable is cleared and an NMI comes
in just before the first NMI does the return. For this case, if the
variable is cleared, we also check if the interrupted stack is the
NMI stack. If it is, then we process the nested NMI code.

Why the two tests and not just test the interrupted stack?

If the first NMI hits a breakpoint and loses NMI context, and then it
hits another breakpoint and while processing that breakpoint we get a
nested NMI. When processing a breakpoint, the stack changes to the
breakpoint stack. If another NMI comes in here we can't rely on the
interrupted stack to be the NMI stack.

If the variable is not set and the interrupted task's stack is not the
NMI stack, then we know this is the first NMI and we can process things
normally. But in order to do so, we need to do a few things first.

1) Set the stack variable that tells us that we are in an NMI handler

2) Make two copies of the interrupt stack frame.
   One copy is used to return on iret
   The other is used to restore the first one if we have a nested NMI.

This is what the stack will look like:

	  +-------------------------+
	  | original SS             |
	  | original Return RSP     |
	  | original RFLAGS         |
	  | original CS             |
	  | original RIP            |
	  +-------------------------+
	  | temp storage for rdx    |
	  +-------------------------+
	  | NMI executing variable  |
	  +-------------------------+
	  | Saved SS                |
	  | Saved Return RSP        |
	  | Saved RFLAGS            |
	  | Saved CS                |
	  | Saved RIP               |
	  +-------------------------+
	  | copied SS               |
	  | copied Return RSP       |
	  | copied RFLAGS           |
	  | copied CS               |
	  | copied RIP              |
	  +-------------------------+
	  | pt_regs                 |
	  +-------------------------+

The original stack frame contains what the HW put in when we entered
the NMI.

We store %rdx as a temp variable to use. Both the original HW stack
frame and this %rdx storage will be clobbered by nested NMIs so we
can not rely on them later in the first NMI handler.

The next item is the special stack variable that is set when we execute
the rest of the NMI handler.

Then we have two copies of the interrupt stack. The second copy is
modified by any nested NMIs to let the first NMI know that we triggered
a second NMI (latched) and that we should repeat the NMI handler.

If the first NMI hits an exception or breakpoint that takes it out of
NMI context, if a second NMI comes in before the first one finishes,
it will update the copied interrupt stack to point to a fix up location
to trigger another NMI.

When the first NMI calls iret, it will instead jump to the fix up
location. This fix up location will copy the saved interrupt stack back
to the copy and execute the nmi handler again.

Note, the nested NMI knows enough to check if it preempted a previous
NMI handler while it is in the fixup location. If it has, it will not
modify the copied interrupt stack and will just leave as if nothing
happened. As the NMI handle is about to execute again, there's no reason
to latch now.

To test all this, I forced the NMI handler to call iret and take itself
out of NMI context. I also added assemble code to write to the serial to
make sure that it hits the nested path as well as the fix up path.
Everything seems to be working fine.

Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: H. Peter Anvin <hpa@linux.intel.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Paul Turner <pjt@google.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Signed-off-by: default avatarSteven Rostedt <rostedt@goodmis.org>
      3f3c8b8c
    • Steven Rostedt's avatar
      x86: Document the NMI handler about not using paranoid_exit · 1fd466ef
      Steven Rostedt authored 
      Linus cleaned up the NMI handler but it still needs some comments to
explain why it uses save_paranoid but not paranoid_exit. Just to keep
others from adding that in the future, document why it's not used.

Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andi Kleen <andi@firstfloor.org>
Signed-off-by: default avatarSteven Rostedt <rostedt@goodmis.org>
      1fd466ef
    • Linus Torvalds's avatar
      x86: Do not schedule while still in NMI context · 549c89b9
      Linus Torvalds authored 
      The NMI handler uses the paranoid_exit routine that checks the
NEED_RESCHED flag, and if it is set and the return is for userspace,
then interrupts are enabled, the stack is swapped to the thread's stack,
and schedule is called. The problem with this is that we are still in an
NMI context until an iret is executed. This means that any new NMIs are
now starved until an interrupt or exception occurs and does the iret.

As NMIs can not be masked and can interrupt any location, they are
treated as a special case. NEED_RESCHED should not be set in an NMI
handler. The interruption by the NMI should not disturb the work flow
for scheduling. Any IPI sent to a processor after sending the
NEED_RESCHED would have to wait for the NMI anyway, and after the IPI
finishes the schedule would be called as required.

There is no reason to do anything special leaving an NMI. Remove the
call to paranoid_exit and do a simple return. This not only fixes the
bug of starved NMIs, but it also cleans up the code.

Link: http://lkml.kernel.org/r/CA+55aFzgM55hXTs4griX5e9=v_O+=ue+7Rj0PTD=M7hFYpyULQ@mail.gmail.comAcked-by: default avatarAndi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: "H. Peter Anvin" <hpa@linux.intel.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Paul Turner <pjt@google.com>
Signed-off-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: default avatarSteven Rostedt <rostedt@goodmis.org>
      549c89b9
  3. 09 Dec, 2011 34 commits
  5. 08 Dec, 2011 3 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7