On Wed, Feb 26, 2014 at 11:14:41AM +0100, Vegard Nossum wrote:
> On 26 February 2014 09:43, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Wed, Feb 19, 2014 at 02:24:41PM -0800, David Rientjes wrote:
> >> On Wed, 19 Feb 2014, Xishi Qiu wrote:
> >>
> >> > Here is a warning, I don't whether it is relative to my hardware.
> >> > If set "kmemcheck=1 nowatchdog", it can boot.
> >> >
> >> > code:
> >> > ...
> >> > pte = kmemcheck_pte_lookup(address);
> >> > if (!pte)
> >> > return false;
> >> >
> >> > WARN_ON_ONCE(in_nmi());
> >> >
> >> > if (error_code & 2)
> >> > ...
> >
> > That code seems to assume NMI context cannot fault; this is false since
> > a while back (v3.9 or thereabouts).
> >
> >> > [ 10.920757] [<ffffffff810452c1>] kmemcheck_fault+0xb1/0xc0
> >> > [ 10.920760] [<ffffffff814d262b>] __do_page_fault+0x39b/0x4c0
> >> > [ 10.920763] [<ffffffff814d2829>] do_page_fault+0x9/0x10
> >> > [ 10.920765] [<ffffffff814cf222>] page_fault+0x22/0x30
> >> > [ 10.920774] [<ffffffff8101eb02>] intel_pmu_handle_irq+0x142/0x3a0
> >> > [ 10.920777] [<ffffffff814d0655>] perf_event_nmi_handler+0x35/0x60
> >> > [ 10.920779] [<ffffffff814cfe83>] nmi_handle+0x63/0x150
> >> > [ 10.920782] [<ffffffff814cffd3>] default_do_nmi+0x63/0x290
> >> > [ 10.920784] [<ffffffff814d02a8>] do_nmi+0xa8/0xe0
> >> > [ 10.920786] [<ffffffff814cf527>] end_repeat_nmi+0x1e/0x2e
> >
> > And this does indeed show a fault from NMI context; which is totally
> > expected.
> >
> > kmemcheck needs to be fixed; but I've no clue how any of that works.
>
> IIRC the reason we don't support page faults in NMI context is that we
> may already be handling an existing fault (or trap) when the NMI hits.
> So that would mess up kmemcheck's working state. I don't really see
> that anything has changed in this respect lately, so it could always
> have been broken.
>
> I think the way we dealt with this before was just to make sure than
> NMI handlers don't access any kmemcheck-tracked memory (i.e. to make
> sure that all memory touched by NMI handlers has been marked NOTRACK).
> And the purpose of this warning is just to tell us that something
> inside an NMI triggered a page fault (in this specific case, it seems
> to be intel_pmu_handle_irq).
>
> I guess there are two ways forward:
>
> - create a stack of things that kmemcheck is working on, so that we
> handle recursive page faults
That's what perf and ftrace do. We keep a 4 layer stack using things
like:
static inline int get_recursion_context(int *recursion)
{
int rctx;
if (in_nmi())
rctx = 3;
else if (in_irq())
rctx = 2;
else if (in_softirq())
rctx = 1;
else
rctx = 0;
if (recursion[rctx])
return -1;
recursion[rctx]++;
barrier();
return rctx;
}
> - try to figure out why intel_pmu_handle_irq() faults and add a
> (kmemcheck-specific) workaround for it
Well, that's easy, we access user memory, which might or might not be
there.
We do this for a number of reasons; one is to read the code and decode
the current basic block to find the previous instruction; see
intel_pmu_pebs_fixup_ip() another is to try and walk the userspace
framepointers, see perf_callchain_user().
In all cases we use 'atomic' accesses which return short copies in case
of failure; we take the fault handler exception path, and we abort the
operation.
> Incidentally, do you remember what exactly changed wrt page faults in
> NMI context?
Sure; commit 3f3c8b8c4b2a34776c3470142a7c8baafcda6eb0 and a fair number
of 'fixes', in particular: 7fbb98c5cb07563d3ee08714073a8e5452a96be2.
These patches made it possible to take faults from NMI context.
Previously this was not possible because we return from the fault using
IRET and IRET unconditionally re-enables NMIs, which is a bit of a
problem when you're still running the NMI handler.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>