Re: [PATCH] x86/cpu: Drop _init from *_cpu_cap functions

[Jan Beulich <jbeulich@xxxxxxxx>]

On 11.08.2022 12:34, Andrew Cooper wrote:
> On 11/08/2022 11:30, Ross Lagerwall wrote:
>>> From: Andrew Cooper <Andrew.Cooper3@xxxxxxxxxx>
>>> Sent: Thursday, August 11, 2022 11:21 AM
>>> To: Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>; 
>>> xen-devel@xxxxxxxxxxxxxxxxxxxx <xen-devel@xxxxxxxxxxxxxxxxxxxx>
>>> Cc: Jan Beulich <jbeulich@xxxxxxxx>; Roger Pau Monne 
>>> <roger.pau@xxxxxxxxxx>; Wei Liu <wl@xxxxxxx>
>>> Subject: Re: [PATCH] x86/cpu: Drop _init from *_cpu_cap functions 
>>>  
>>> On 11/08/2022 11:17, Ross Lagerwall wrote:
>>>> These functions may be called by init_amd() after the _init functions
>>>> have been purged during CPU hotplug or PV shim boot so drop the _init.
>>>>
>>>> Signed-off-by: Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>
>>> Hmm.  That's a bug in init_amd() I'd say.  These really shouldn't be
>>> used after __init.
>>>
>>> Which path exploded specifically?
>> The stack trace was:
>>
>> setup_force_cpu_cap
>> init_amd
>> identify_cpu
>> start_secondary
>>
>> In setup_force_cpu_cap() here:
>>
>>         /*
>>          * On pre-CLFLUSHOPT AMD CPUs, CLFLUSH is weakly ordered with
>>          * everything, including reads and writes to address, and
>>          * LFENCE/SFENCE instructions.
>>          */
>>         if (!cpu_has_clflushopt)
>>                 setup_force_cpu_cap(X86_BUG_CLFLUSH_MFENCE);
>>
>> which was recently introduced by:
>>
>> commit 062868a5a8b428b85db589fa9a6d6e43969ffeb9
>> Author: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
>> Date:   Thu Jun 9 14:23:07 2022 +0200
>>
>>     x86/amd: Work around CLFLUSH ordering on older parts
> 
> Bah, and that was also backported in a security fix, to everything back
> to 4.12 is broken.

4.13, but yes. Oh well.

It's actually odd that we use __set_bit() for X86_FEATURE_MFENCE_RDTSC (a
few lines up) but the more heavyweight function for X86_BUG_CLFLUSH_MFENCE.

Jan