Messages in this thread | | | Date | Wed, 14 Jul 2010 11:10:20 -0700 | Subject | Re: [patch 1/2] x86_64 page fault NMI-safe | From | Linus Torvalds <> |
| |
On Wed, Jul 14, 2010 at 10:06 AM, Mathieu Desnoyers <mathieu.desnoyers@efficios.com> wrote: >> >> This patch (1/2) doesn't look horrible per se. I have no problems with >> it. I just want to understand why it is needed.
[ And patch 2/2 is much more intrusive, and touches a critical path too.. If it was just the 1/2 series, I don't think I would care. For the 2/2, I think I'd want to explore all the alternative options ]
> The problem originally addressed by this patch is the case where a NMI handler > try to access vmalloc'd per-cpu data, which goes as follow: > > - One CPU does a fork(), which copies the basic kernel mappings. > - Perf allocates percpu memory for buffer control data structures. > This mapping does not get copied. > - Tracing is activated. > - switch_to() to the newly forked process which missed the new percpu > allocation. > - We take a NMI, which touches the vmalloc'd percpu memory in the Perf tracing > handler, therefore leading to a page fault in NMI context. Here, we might be > in the middle of switch_to(), where ->current might not be in sync with the > current cr3 register.
Ok. I was wondering why anybody would allocate core percpu variables so late that this would ever be an issue, but I guess perf is a reasonable such case. And reasonable to do from NMI.
That said - grr. I really wish there was some other alternative than adding yet more complexity to the exception return path. That "iret re-enables NMI's unconditionally" thing annoys me.
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.
Doesn't the above sound like a good solution? In other words, we solve the whole problem by simply _fixing_ the crazy Intel "iret-vs-NMI" semantics. And we don't need to change the hotpath, and we'll just _allow_ nested faults within NMI's.
Am I missing something? Maybe I'm not as clever as I think I am... But I _feel_ clever.
Linus -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |