Re: [PATCH v2 2/2] x86/xen: allow privcmd hypercalls to be preempted

[Andy Lutomirski]

On Wed, Dec 10, 2014 at 4:55 PM, Luis R. Rodriguez <mcgrof@xxxxxxxx> wrote:
> On Wed, Dec 10, 2014 at 03:51:48PM -0800, Andy Lutomirski wrote:
>> On Wed, Dec 10, 2014 at 3:34 PM, Luis R. Rodriguez
>> <mcgrof@xxxxxxxxxxxxxxxx> wrote:
>> > From: "Luis R. Rodriguez" <mcgrof@xxxxxxxx>
>> >
>> > Xen has support for splitting heavy work work into a series
>> > of hypercalls, called multicalls, and preempting them through
>> > what Xen calls continuation [0]. Despite this though without
>> > CONFIG_PREEMPT preemption won't happen and while enabling
>> > CONFIG_RT_GROUP_SCHED can at times help its not enough to
>> > make a system usable. Such is the case for example when
>> > creating a > 50 GiB HVM guest, we can get softlockups [1] with:.
>> >
>> > kernel: [  802.084335] BUG: soft lockup - CPU#1 stuck for 22s! [xend:31351]
>> >
>> > The softlock up triggers on the TASK_UNINTERRUPTIBLE hanger check
>> > (default 120 seconds), on the Xen side in this particular case
>> > this happens when the following Xen hypervisor code is used:
>> >
>> > xc_domain_set_pod_target() -->
>> >   do_memory_op() -->
>> >     arch_memory_op() -->
>> >       p2m_pod_set_mem_target()
>> >         -- long delay (real or emulated) --
>> >
>> > This happens on arch_memory_op() on the XENMEM_set_pod_target memory
>> > op even though arch_memory_op() can handle continuation via
>> > hypercall_create_continuation() for example.
>> >
>> > Machines over 50 GiB of memory are on high demand and hard to come
>> > by so to help replicate this sort of issue long delays on select
>> > hypercalls have been emulated in order to be able to test this on
>> > smaller machines [2].
>> >
>> > On one hand this issue can be considered as expected given that
>> > CONFIG_PREEMPT=n is used however we have forced voluntary preemption
>> > precedent practices in the kernel even for CONFIG_PREEMPT=n through
>> > the usage of cond_resched() sprinkled in many places. To address
>> > this issue with Xen hypercalls though we need to find a way to aid
>> > to the schedular in the middle of hypercalls. We are motivated to
>> > address this issue on CONFIG_PREEMPT=n as otherwise the system becomes
>> > rather unresponsive for long periods of time; in the worst case, at least
>> > only currently by emulating long delays on select io disk bound
>> > hypercalls, this can lead to filesystem corruption if the delay happens
>> > for example on SCHEDOP_remote_shutdown (when we call 'xl <domain> shutdown').
>> >
>> > We can address this problem by trying to check if we should schedule
>> > on the xen timer in the middle of a hypercall on the return from the
>> > timer interrupt. We want to be careful to not always force voluntary
>> > preemption though so to do this we only selectively enable preemption
>> > on very specific xen hypercalls.
>> >
>> > This enables hypercall preemption by selectively forcing checks for
>> > voluntary preempting only on ioctl initiated private hypercalls
>> > where we know some folks have run into reported issues [1].
>> >
>> > [0] http://xenbits.xen.org/gitweb/?p=xen.git;a=commitdiff;h=42217cbc5b3e84b8c145d8cfb62dd5de0134b9e8;hp=3a0b9c57d5c9e82c55dd967c84dd06cb43c49ee9
>> > [1] https://bugzilla.novell.com/show_bug.cgi?id=861093
>> > [2] http://ftp.suse.com/pub/people/mcgrof/xen/emulate-long-xen-hypercalls.patch
>> >
>> > Based on original work by: David Vrabel <david.vrabel@xxxxxxxxxx>
>> > Cc: Borislav Petkov <bp@xxxxxxx>
>> > Cc: David Vrabel <david.vrabel@xxxxxxxxxx>
>> > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
>> > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
>> > Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
>> > Cc: x86@xxxxxxxxxx
>> > Cc: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
>> > Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
>> > Cc: Masami Hiramatsu <masami.hiramatsu.pt@xxxxxxxxxxx>
>> > Cc: Jan Beulich <JBeulich@xxxxxxxx>
>> > Cc: linux-kernel@xxxxxxxxxxxxxxx
>> > Signed-off-by: Luis R. Rodriguez <mcgrof@xxxxxxxx>
>> > ---
>> >  arch/x86/kernel/entry_32.S | 21 +++++++++++++++++++++
>> >  arch/x86/kernel/entry_64.S | 17 +++++++++++++++++
>> >  drivers/xen/Makefile       |  2 +-
>> >  drivers/xen/preempt.c      | 17 +++++++++++++++++
>> >  drivers/xen/privcmd.c      |  2 ++
>> >  include/xen/xen-ops.h      | 26 ++++++++++++++++++++++++++
>> >  6 files changed, 84 insertions(+), 1 deletion(-)
>> >  create mode 100644 drivers/xen/preempt.c
>> >
>> > diff --git a/arch/x86/kernel/entry_32.S b/arch/x86/kernel/entry_32.S
>> > index 344b63f..40b5c0c 100644
>> > --- a/arch/x86/kernel/entry_32.S
>> > +++ b/arch/x86/kernel/entry_32.S
>> > @@ -982,7 +982,28 @@ ENTRY(xen_hypervisor_callback)
>> >  ENTRY(xen_do_upcall)
>> >  1:     mov %esp, %eax
>> >         call xen_evtchn_do_upcall
>> > +#ifdef CONFIG_PREEMPT
>> >         jmp  ret_from_intr
>> > +#else
>> > +       GET_THREAD_INFO(%ebp)
>> > +#ifdef CONFIG_VM86
>> > +       movl PT_EFLAGS(%esp), %eax      # mix EFLAGS and CS
>> > +       movb PT_CS(%esp), %al
>> > +       andl $(X86_EFLAGS_VM | SEGMENT_RPL_MASK), %eax
>> > +#else
>> > +       movl PT_CS(%esp), %eax
>> > +       andl $SEGMENT_RPL_MASK, %eax
>> > +#endif
>> > +       cmpl $USER_RPL, %eax
>> > +       jae resume_userspace            # returning to v8086 or userspace
>> > +       DISABLE_INTERRUPTS(CLBR_ANY)
>> > +       cmpb $0,PER_CPU_VAR(xen_in_preemptible_hcall)
>> > +       jz resume_kernel
>> > +       movb $0,PER_CPU_VAR(xen_in_preemptible_hcall)
>> > +       call cond_resched_irq
>> > +       movb $1,PER_CPU_VAR(xen_in_preemptible_hcall)
>> > +       jmp resume_kernel
>> > +#endif /* CONFIG_PREEMPT */
>> >         CFI_ENDPROC
>> >  ENDPROC(xen_hypervisor_callback)
>> >
>> > diff --git a/arch/x86/kernel/entry_64.S b/arch/x86/kernel/entry_64.S
>> > index c0226ab..0ccdd06 100644
>> > --- a/arch/x86/kernel/entry_64.S
>> > +++ b/arch/x86/kernel/entry_64.S
>> > @@ -1170,7 +1170,23 @@ ENTRY(xen_do_hypervisor_callback)   # do_hypervisor_callback(struct *pt_regs)
>> >         popq %rsp
>> >         CFI_DEF_CFA_REGISTER rsp
>> >         decl PER_CPU_VAR(irq_count)
>> > +#ifdef CONFIG_PREEMPT
>> >         jmp  error_exit
>> > +#else
>> > +       movl %ebx, %eax
>> > +       RESTORE_REST
>> > +       DISABLE_INTERRUPTS(CLBR_NONE)
>> > +       TRACE_IRQS_OFF
>> > +       GET_THREAD_INFO(%rcx)
>> > +       testl %eax, %eax
>> > +       je error_exit_user
>> > +       cmpb $0,PER_CPU_VAR(xen_in_preemptible_hcall)
>> > +       jz retint_kernel
>>
>> I think I understand this part.
>>
>> > +       movb $0,PER_CPU_VAR(xen_in_preemptible_hcall)
>>
>> Why?  Is the issue that, if preemptible hypercalls nest, you don't
>> want to preempt again?
>
>
> So this callback on the xen timer, without the CPU variable
> we would not be able to selectively preempt specific hypercalls
> and we'd a no-preempt kernel fully preemptive.
>
> I asked the same question, see:
>
> https://lkml.org/lkml/2014/12/3/756
>
>>
>> > +       call cond_resched_irq
>>
>> On !CONFIG_PREEMPT, there's no preempt_disable, right?  So how do you
>> guarantee that you don't preempt something you shouldn't?
>
> Not sure I follow, but in essence this is no different then the use
> of cond_resched() on !CONFIG_PREEMPT, the only thing here is we are
> in interrupt context. If this is about abuse of making !CONFIG_PREEMPT
> voluntarily preemptive at select points then I had similar concerns
> and David pointed out to me the wide use of cond_resched() on the
> kernel.
>
>> Is the idea
>> that these events will only fire nested *directly* inside a
>> preemptible hypercall?
>
> Yeah its the timer interrupt that would trigger the above.
>
>> Also, should you check that IRQs were on when
>> the event fired?  (Are they on in pt_regs?)
>
> Right before this xen_evtchn_do_upcall() is issued, which
> saves pt_regs and then restores them.
>
>> > +       movb $1,PER_CPU_VAR(xen_in_preemptible_hcall)
>> > +       jmp retint_kernel
>> > +#endif /* CONFIG_PREEMPT */
>> >         CFI_ENDPROC
>>
>> All that being said, this is IMO a bit gross.
>
> That was my first reaction hence my original attempt to try to get away from
> this. I've learned David also tried to not go down this route too before,
> more on this below.
>
>>  You've added a bunch of
>> asm that's kind of like a parallel error_exit,
>
> yeah.. I first tried to macro'tize this but it looked hairy, if we
> wanted to do that... (although the name probably ain't best)
>
> 32-bit:
> .macro test_from_kernel kernel_ret:req
>         GET_THREAD_INFO(%ebp)
> #ifdef CONFIG_VM86
>         movl PT_EFLAGS(%esp), %eax      # mix EFLAGS and CS
>         movb PT_CS(%esp), %al
>         andl $(X86_EFLAGS_VM | SEGMENT_RPL_MASK), %eax
> #else
>         /*
>          * We can be coming here from child spawned by kernel_thread().
>          */
>         movl PT_CS(%esp), %eax
>         andl $SEGMENT_RPL_MASK, %eax
> #endif
>         cmpl $USER_RPL, %eax
>         jb \kernel_ret
> .endm
>
> 64-bit:
>         .macro test_from_kernel kernel_ret:req
>         movl %ebx,%eax
>         RESTORE_REST
>         DISABLE_INTERRUPTS(CLBR_NONE)
>         TRACE_IRQS_OFF
>         GET_THREAD_INFO(%rcx)
>         testl %eax,%eax
>         jne \kernel_ret
>         .endm
>
>> and the error entry and
>> exit code is hairy enough that this scares me.
>
> yeah...
>
>>  Can you do this mostly in C instead?  This would look a nicer if it could be:
>>
>>     call xen_evtchn_do_upcall
>>     popq %rsp
>>     CFI_DEF_CFA_REGISTER rsp
>>     decl PER_CPU_VAR(irq_count)
>> +  call xen_end_upcall
>>     jmp error_exit
>
> It seems David tried this originally eons ago:
>
> http://lists.xen.org/archives/html/xen-devel/2014-02/msg01101.html
>
> and the strategy was shifted based on Jan's feedback that we could
> not sched as we're on the IRQ stack. David evolved the strategy
> to asm and to use preempt_schedule_irq(), this new iteratin just
> revisits the same old but tries to generalize scheduling on IRQ
> context on very special circumstances.

Indeed.  But look more closely at my proposed one-line asm patch:
we're not on the irq stack.

Also, to make this more obviously safe wrt preempting at the wrong
time, would it be possible to check regs->ip and make sure it's
pointing exactly where you expect before scheduling?  (You'd need one
more line of asm to get pt_regs in xen_end_upcall.)

--Andy, the stack switching maestro extraordinaire :)

>
>> Where xen_end_upcall would be witten in C, nokprobes and notrace (if
>> needed) and would check pt_regs and whatever else and just call
>> schedule if needed?
>
> If there's a way to do it it'd be great. I am not sure if we can though.
>
>   Luis



-- 
Andy Lutomirski
AMA Capital Management, LLC
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/