lkml.org 
[lkml]   [2019]   [Jan]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v3 0/6] Static calls
From
On January 14, 2019 3:27:55 PM PST, Andy Lutomirski <luto@kernel.org> wrote:
>On Mon, Jan 14, 2019 at 2:01 PM H. Peter Anvin <hpa@zytor.com> wrote:
>>
>> So I was already in the middle of composing this message when Andy
>posted:
>>
>> > I don't even think this is sufficient. I think we also need
>everyone
>> > who clears the bit to check if all bits are clear and, if so,
>remove
>> > the breakpoint. Otherwise we have a situation where, if you are in
>> > text_poke_bp() and you take an NMI (or interrupt or MCE or
>whatever)
>> > and that interrupt then hits the breakpoint, then you deadlock
>because
>> > no one removes the breakpoint.
>> >
>> > If we do this, and if we can guarantee that all CPUs make forward
>> > progress, then maybe the problem is solved. Can we guarantee
>something
>> > like all NMI handlers that might wait in a spinlock or for any
>other
>> > reason will periodically check if a sync is needed while they're
>> > spinning?
>>
>> So the really, really nasty case is when an asynchronous event on the
>> *patching* processor gets stuck spinning on a resource which is
>> unavailable due to another processor spinning on the #BP. We can
>disable
>> interrupts, but we can't stop NMIs from coming in (although we could
>> test in the NMI handler if we are in that condition and return
>> immediately; I'm not sure we want to do that, and we still have to
>deal
>> with #MC and what not.)
>>
>> The fundamental problem here is that we don't see the #BP on the
>> patching processor, in which case we could simply complete the
>patching
>> from the #BP handler on that processor.
>>
>> On 1/13/19 6:40 PM, H. Peter Anvin wrote:
>> > On 1/13/19 6:31 PM, H. Peter Anvin wrote:
>> >>
>> >> static cpumask_t text_poke_cpumask;
>> >>
>> >> static void text_poke_sync(void)
>> >> {
>> >> smp_wmb();
>> >> text_poke_cpumask = cpu_online_mask;
>> >> smp_wmb(); /* Should be optional on x86 */
>> >> cpumask_clear_cpu(&text_poke_cpumask, smp_processor_id());
>> >> on_each_cpu_mask(&text_poke_cpumask, text_poke_sync_cpu,
>NULL, false);
>> >> while (!cpumask_empty(&text_poke_cpumask)) {
>> >> cpu_relax();
>> >> smp_rmb();
>> >> }
>> >> }
>> >>
>> >> static void text_poke_sync_cpu(void *dummy)
>> >> {
>> >> (void)dummy;
>> >>
>> >> smp_rmb();
>> >> cpumask_clear_cpu(&poke_bitmask, smp_processor_id());
>> >> /*
>> >> * We are guaranteed to return with an IRET, either from the
>> >> * IPI or the #BP handler; this provides serialization.
>> >> */
>> >> }
>> >>
>> >
>> > The invariants here are:
>> >
>> > 1. The patching routine must set each bit in the cpumask after each
>event
>> > that requires synchronization is complete.
>> > 2. The bit can be (atomically) cleared on the target CPU only, and
>only in a
>> > place that guarantees a synchronizing event (e.g. IRET) before
>it may
>> > reaching the poked instruction.
>> > 3. At a minimum the IPI handler and #BP handler needs to clear the
>bit. It
>> > *is* also possible to clear it in other places, e.g. the NMI
>handler, if
>> > necessary as long as condition 2 is satisfied.
>> >
>>
>> OK, so with interrupts enabled *on the processor doing the patching*
>we
>> still have a problem if it takes an interrupt which in turn takes a
>#BP.
>> Disabling interrupts would not help, because but an NMI and #MC
>could
>> still cause problems unless we can guarantee that no path which may
>be
>> invoked by NMI/#MC can do text_poke, which seems to be a very
>aggressive
>> assumption.
>>
>> Note: I am assuming preemption is disabled.
>>
>> The easiest/sanest way to deal with this might be to switch the IDT
>(or
>> provide a hook in the generic exception entry code) on the patching
>> processor, such that if an asynchronous event comes in, we either
>roll
>> forward or revert. This is doable because the second sync we
>currently
>> do is not actually necessary per the hardware guys.
>
>This is IMO insanely complicated. I much prefer the kind of
>complexity that is more or less deterministic and easy to test to the
>kind of complexity (like this) that only happens in corner cases.
>
>I see two solutions here:
>
>1. Just suck it up and emulate the CALL. And find a way to write a
>test case so we know it works.
>
>2. Find a non-deadlocky way to make the breakpoint handler wait for
>the breakpoint to get removed, without any mucking at all with the
>entry code. And find a way to write a test case so we know it works.
>(E.g. stick an actual static_call call site *in text_poke_bp()* that
>fires once on boot so that the really awful recursive case gets
>exercised all the time.
>
>But if we're going to do any mucking with the entry code, let's just
>do the simple mucking to make emulating CALL work.
>
>--Andy

Ugh. So much for not really proofreading. Yes, I think the second solution is the right thing since I think I figured out how to do it without deadlock; see other mail.
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.

\
 
 \ /
  Last update: 2019-01-15 03:30    [W:0.290 / U:0.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site