Messages in this thread | | | Subject | Re: recent x86-64 nested NMI adjustments | From | Steven Rostedt <> | Date | Mon, 12 Mar 2012 09:16:41 -0400 |
| |
On Mon, 2012-03-12 at 12:10 +0000, Jan Beulich wrote: > Hi Steven, > > the explanation of 45d5a1683c04be28abdf5c04c27b1417e0374486 > seems bogus to me: When arriving from user mode, %rsp won't point > to the user stack anymore, as it gets switched away from during the > processing of the exception (the more that the IDT entry specifies a > separate stack anyway, which even guarantees this for kernel mode > entries).
No it is real, and I had a test program that exploited it. I'm not worried about the current %rsp, I'm worried about what %rsp is saved on the stack. Two things are used to check if the incoming NMI is nested or not.
1) if the on-stack "in-nmi" variable is set
2) if the saved %rsp is pointing to the NMI stack.
Note, #2 looks at the *saved* %rsp. Which is the %rsp at the time the NMI triggered. The second check is used to handle the case that a nested NMI came in after the previous NMI cleared the on-stack "in-nmi" variable, but before it calls the iret.
There are few cases that the stack can change in the NMI so the variable is also used.
There's a really good article on LWN about this :-)
https://lwn.net/Articles/484932/
(subscription required, but you should have one)
That said, I added a printk into the boot up to show me where the NMI stacks were located. Then I wrote a program that would pin itself to a CPU and change its stack pointer to point into the NMI stack of that CPU and then go into an infinite loop. I ran perf on this code and it became "invisible" to perf. That is, every time the NMI came in while this code was running, it incorrectly considered itself a nested NMI and returned, never recording the presence of this program.
After adding this patch, perf shows the task spending 99.9% of the time in this loop. Thus this is a real bug.
> > Further, a38449ef596b345e13a8f9b7d5cd9fedb8fcf921 makes the > (presumably superfluous) compare a 4-byte one, while the > documentation isn't really stating that selectors get pushed zero- > extended. Hence, if not reverting the first change altogether, I'd > minimally recommend converting the compare to a 2-byte one.
I'll let H. Peter answer this one, he's the Intel representative here.
-- Steve
| |