Date:   Thu, 24 Nov 2022 08:47:47 +0100
From:   Juergen Gross <jgross@...e.com>
To:     Josh Poimboeuf <jpoimboe@...nel.org>,
        Andrew Cooper <Andrew.Cooper3@...rix.com>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        "Paul E. McKenney" <paulmck@...nel.org>,
        "sfr@...b.auug.org.au" <sfr@...b.auug.org.au>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "sstabellini@...nel.org" <sstabellini@...nel.org>,
        "boris.ostrovsky@...cle.com" <boris.ostrovsky@...cle.com>,
        "xen-devel@...ts.xenproject.org" <xen-devel@...ts.xenproject.org>,
        Thomas Gleixner <tglx@...utronix.de>
Subject: Re: objtool warning for next-20221118

On 24.11.22 06:28, Juergen Gross wrote:
> On 24.11.22 03:39, Josh Poimboeuf wrote:
>> On Wed, Nov 23, 2022 at 09:03:40AM -0800, Josh Poimboeuf wrote:
>>> On Wed, Nov 23, 2022 at 10:52:09AM +0000, Andrew Cooper wrote:
>>>>> Well, if you return from arch_cpu_idle_dead() you're back in the idle
>>>>> loop -- exactly where you would be if you were to bootstrap the whole
>>>>> CPU -- provided you have it remember the whole state (easier with a
>>>>> vCPU).
>>>
>>> play_dead() really needs sane semantics.  Not only does it introduce a
>>> surprise to the offlining code in do_idle(), it also skips the entire
>>> hotplug state machine.  Not sure if that introduces any bugs, but at the
>>> very least it's subtle and surprising.
>>>
>>>>> But maybe I'm missing something, lets add Xen folks on.
>>>>
>>>> Calling VCPUOP_down on oneself always succeeds, but all it does is
>>>> deschedule the vCPU.
>>>>
>>>> It can be undone at a later point by a different vcpu issuing VCPUOP_up
>>>> against the previously-downed CPU, at which point the vCPU gets rescheduled.
>>>>
>>>> This is why the VCPUOP_down hypercall returns normally.  All state
>>>> really is intact.
>>>>
>>>> As for what Linux does, this is how xen_pv_cpu_up() currently behaves.
>>>> If you want to make Xen behave more everything else, then bug a BUG()
>>>> after VCPUOP_down, and adjust xen_pv_cpu_up() to skip its initialised
>>>> check and always use VCPUOP_initialise to bring the vCPU back online.
>>>
>>> Or we could do what sev_es_play_dead() does and just call start_cpu0()
>>> after the hypercall returns?
>>
>> Something like so (untested).  This is only the x86 bits.
>>
>> I think I convinced myself that start_cpu0() isn't buggy.  I'm looking
>> at other cleanups, e.g. converging cpu_bringup_and_idle() with
>> start_secondary().
>>
>> I can pick it up again next week, post-turkey.
>>
>> diff --git a/arch/x86/include/asm/smp.h b/arch/x86/include/asm/smp.h
>> index b4dbb20dab1a..e6d1d2810e38 100644
>> --- a/arch/x86/include/asm/smp.h
>> +++ b/arch/x86/include/asm/smp.h
>> @@ -93,9 +93,10 @@ static inline void __cpu_die(unsigned int cpu)
>>       smp_ops.cpu_die(cpu);
>>   }
>> -static inline void play_dead(void)
>> +static inline void __noreturn play_dead(void)
>>   {
>>       smp_ops.play_dead();
>> +    BUG();
>>   }
>>   static inline void smp_send_reschedule(int cpu)
>> diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
>> index 26e8f57c75ad..8e2841deb1eb 100644
>> --- a/arch/x86/kernel/process.c
>> +++ b/arch/x86/kernel/process.c
>> @@ -700,7 +700,7 @@ EXPORT_SYMBOL(boot_option_idle_override);
>>   static void (*x86_idle)(void);
>>   #ifndef CONFIG_SMP
>> -static inline void play_dead(void)
>> +static inline void __noreturn play_dead(void)
>>   {
>>       BUG();
>>   }
>> diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
>> index 55cad72715d9..d8b12ac1a7c5 100644
>> --- a/arch/x86/kernel/smpboot.c
>> +++ b/arch/x86/kernel/smpboot.c
>> @@ -1833,9 +1833,12 @@ void native_play_dead(void)
>>       play_dead_common();
>>       tboot_shutdown(TB_SHUTDOWN_WFS);
>> -    mwait_play_dead();    /* Only returns on failure */
>> +    mwait_play_dead();    /* Only returns if mwait is not supported */
>> +
>>       if (cpuidle_play_dead())
>>           hlt_play_dead();
>> +
>> +    BUG();
>>   }
>>   #else /* ... !CONFIG_HOTPLUG_CPU */
>> diff --git a/arch/x86/xen/smp_pv.c b/arch/x86/xen/smp_pv.c
>> index 480be82e9b7b..30dc904ca990 100644
>> --- a/arch/x86/xen/smp_pv.c
>> +++ b/arch/x86/xen/smp_pv.c
>> @@ -385,17 +385,9 @@ static void xen_pv_play_dead(void) /* used only with 
>> HOTPLUG_CPU */
>>   {
>>       play_dead_common();
>>       HYPERVISOR_vcpu_op(VCPUOP_down, xen_vcpu_nr(smp_processor_id()), NULL);
>> -    cpu_bringup();
>> -    /*
>> -     * commit 4b0c0f294 (tick: Cleanup NOHZ per cpu data on cpu down)
>> -     * clears certain data that the cpu_idle loop (which called us
>> -     * and that we return from) expects. The only way to get that
>> -     * data back is to call:
>> -     */
>> -    tick_nohz_idle_enter();
>> -    tick_nohz_idle_stop_tick_protected();
>> -    cpuhp_online_idle(CPUHP_AP_ONLINE_IDLE);
>> +    /* FIXME: converge cpu_bringup_and_idle() and start_secondary() */
>> +    cpu_bringup_and_idle();
> 
> I think this will leak stack memory. Multiple cpu offline/online cycles of
> the same cpu will finally exhaust the idle stack.

The attached patch seems to work fine.

The __noreturn annotation seems to trigger an objtool warning, though, in
spite of the added BUG() at the end of xen_pv_play_dead():

arch/x86/xen/smp_pv.o: warning: objtool: xen_pv_play_dead() falls through to 
next function xen_pv_cpu_die()


Juergen


View attachment "0001-x86-xen-don-t-let-xen_pv_play_dead-return.patch" of type "text/x-patch" (3215 bytes)

Download attachment "OpenPGP_0xB0DE9DD628BF132F.asc" of type "application/pgp-keys" (3099 bytes)

Download attachment "OpenPGP_signature" of type "application/pgp-signature" (496 bytes)

Powered by blists - more mailing lists