Re: [3.18+] Can't boot with commit bd809af1 ("x86: Enable PAT to use cache mode translation tables")

[Peter Hurley <peter@hurleysoftware.com>]

VMware guys probably already know this but just in case

[ +cc Thomas Hellstrom ]

Jongman - you need to fix your mailer to use plaintext and not base64.

On 12/16/2014 01:46 AM, Jongman Heo wrote:
>> Sender : Juergen Gross<jgross@suse.com>
>> On 12/16/2014 07:29 AM, Jongman Heo wrote:
>>>> Sender : Juergen Gross
>>>> On 12/16/2014 05:40 AM, Jongman Heo wrote:
>>>>>> Sender : Juergen Gross
>>>>>> On 12/15/2014 08:52 AM, Jongman Heo wrote:
>>>>>>>> Sender : Juergen Gross
>>>>>>>> On 12/14/2014 06:07 AM, 허종만 wrote:
>>>>>>>>>
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> My Linux virtual machine on (Windows) VMWare workstation 10 can't boot with following commit.
>>>>>>>>>
>>>>>>>>> commit bd809af16e3ab1f8d55b3e2928c47c67e2a865d2
>>>>>>>>> Author: Juergen Gross
>>>>>>>>> Date:   Mon Nov 3 14:02:03 2014 +0100
>>>>>>>>>
>>>>>>>>>         x86: Enable PAT to use cache mode translation tables
>>>>>>>>>
>>>>>>>>> Unfortunately I can't see any console log.
>>>>>>>>
>>>>>>>> Hmm, weird. Could you provide some more information?
>>>>>>>>
>>>>>>>> Kernel config, hardware used, /proc/cpuinfo of working kernel?
>>>>>>>> Anything you see with earlyprintk enabled?
>>>>>>>>
>>>>>>>>
>>>>>>>> Juergen
>>>>>>>
>>>>>>> (Sorry for resending this email, previous one bounced from mailing list due to HTML format)
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> I'm using Fedora 21, with custom built kernel.
>>>>>>> Host PC is windows 7 64-bit, and running VMWare workstation 10 for guest Fedora Linux.
>>>>>>>
>>>>>>> With earlyprintk, just following message is printed.
>>>>>>>
>>>>>>>      early console in setup code
>>>>>>>
>>>>>>> and nothing more...
>>>>>>
>>>>>> Can you try attached diagnostic patch, please? I suspect a problem
>>>>>> regarding VMWares PAT emulation...
>>>>>>
>>>>>>
>>>>>> Juergen
>>>>>
>>>>> Hi,
>>>>>
>>>>> With the commit reverted, the patch doesn't apply.
>>>>
>>>> Sure.
>>>>
>>>>> Without revert, kernel (patch applied) doesn't boot and I can't see any message.
>>>>
>>>> What are your kernel parameters? There must be some message with the
>>>> diagnostic patch, as the first pr_info() is called before any other
>>>> part of the critical patch is becoming active. Could it be you have
>>>> instructed the kernel to be "quiet"? I'd recommend:
>>>>
>>>> earlyprintk=vga ignore_loglevel
>>>>
>>>> and no quiet. I don't know VMWare settings, so may be you can use
>>>> earlyprintk=ttyS0 instead of vga.
>>>>
>>>>>
>>>>> Let me show you my PAT values (the commit reverted)
>>>>>
>>>>> # dmesg | grep PAT
>>>>> [    0.000000] x86 PAT enabled: cpu 0, old 0x0, new 0x7010600070106
>>>>> [    0.314631] x86 PAT enabled: cpu 3, old 0x0, new 0x7010600070106
>>>>> [    0.314703] x86 PAT enabled: cpu 1, old 0x0, new 0x7010600070106
>>>>> [    0.314780] x86 PAT enabled: cpu 2, old 0x0, new 0x7010600070106
>>>>> [    0.314852] x86 PAT enabled: cpu 4, old 0x0, new 0x7010600070106
>>>>> [    0.314923] x86 PAT enabled: cpu 0, old 0x0, new 0x7010600070106
>>>>> [    0.314997] x86 PAT enabled: cpu 6, old 0x0, new 0x7010600070106
>>>>> [    0.315069] x86 PAT enabled: cpu 7, old 0x0, new 0x7010600070106
>>>>> [    0.315142] x86 PAT enabled: cpu 5, old 0x0, new 0x7010600070106
>>>>
>>>> These are the expected values. But these values are the ones which are
>>>> written, not the ones which have been read from the PAT MSR again.
>>>>
>>>> Without applying the critical patch you could add:
>>>>
>>>> rdmsrl(MSR_IA32_CR_PAT, pat);
>>>> printk(KERN_INFO "PAT read: cpu %d, 0x%Lx\n", smp_processor_id(), pat);
>>>>
>>>> at the end of pat_init() to verify VMWare is handling reads of the PAT
>>>> MSR properly.
>>>>
>>>> Juergen
>>>>
>>>
>>> Hi,
>>>
>>> With earlyprintk=vga, I can see the log.
>>> But due to call trace, I can't see what the pat value is.
>>>
>>> Call chain is as follows.
>>>
>>>    i386_start_kernel -> start_kernel -> setup_arch ->
>>>    mtrr_bp_init -> get_mtrr_state -> pat_init ->
>>>    pat_init_cache_mode_entry -> update_cache_mode_entry ->
>>>    early_idt_handler -> dump_stack
>>>
>>> So, I blocked update_cache_mode_entry() call like below...
>>>
>>> --- a/arch/x86/mm/pat.c
>>> +++ b/arch/x86/mm/pat.c
>>> @@ -182,11 +182,12 @@ void pat_init_cache_modes(void)
>>>          u64 pat;
>>>   
>>>          rdmsrl(MSR_IA32_CR_PAT, pat);
>>> +       pr_info("read pat %0llx\n", pat);
>>>          pat_msg[32] = 0;
>>>          for (i = 7; i >= 0; i--) {
>>>                  cache = pat_get_cache_mode((pat >> (i * 8)) & 7,
>>>                                             pat_msg + 4 * i);
>>> -               update_cache_mode_entry(i, cache);
>>> +               //update_cache_mode_entry(i, cache);
>>>          }
>>>          pr_info("PAT configuration [0-7]: %s\n", pat_msg);
>>>   }
>>> @@ -238,9 +239,13 @@ void pat_init(void)
>>>                  rdmsrl(MSR_IA32_CR_PAT, boot_pat_state);
>>>   
>>>          wrmsrl(MSR_IA32_CR_PAT, pat);
>>> +       pr_info("about to write pat %0llx\n", pat);
>>>   
>>>          if (boot_cpu)
>>>                  pat_init_cache_modes();
>>> +
>>> +       rdmsrl(MSR_IA32_CR_PAT, pat);
>>> +       printk(KERN_INFO "PAT read: cpu %d, 0x%Lx\n", smp_processor_id(), pat);
>>>   }
>>>   
>>>
>>> Then boot is fine, and PAT values are as follows.
>>>
>>>
>>> # dmesg|grep -i "pat "
>>> [    0.000000] about to write pat 7010600070106
>>> [    0.000000] read pat 0
>>> [    0.000000] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.000000] PAT read: cpu 0, 0x0
>>> [    0.320559] about to write pat 7010600070106
>>> [    0.320876] read pat 0
>>> [    0.321090] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.321260] PAT read: cpu 5, 0x0
>>> [    0.321403] about to write pat 7010600070106
>>> [    0.321818] read pat 0
>>> [    0.322033] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.322205] PAT read: cpu 6, 0x0
>>> [    0.322334] about to write pat 7010600070106
>>> [    0.322417] read pat 0
>>> [    0.322479] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.322573] PAT read: cpu 0, 0x0
>>> [    0.322703] about to write pat 7010600070106
>>> [    0.323012] read pat 0
>>> [    0.323228] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.323400] PAT read: cpu 1, 0x0
>>> [    0.323537] about to write pat 7010600070106
>>> [    0.323833] read pat 0
>>> [    0.324055] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.324224] PAT read: cpu 7, 0x0
>>> [    0.324362] about to write pat 7010600070106
>>> [    0.324662] read pat 0
>>> [    0.324877] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.325048] PAT read: cpu 2, 0x0
>>> [    0.325185] about to write pat 7010600070106
>>> [    0.325483] read pat 0
>>> [    0.325695] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.325863] PAT read: cpu 4, 0x0
>>> [    0.325997] about to write pat 7010600070106
>>> [    0.326288] read pat 0
>>> [    0.326507] PAT configuration [0-7]: UC  UC  UC  UC  UC  UC  UC  UC
>>> [    0.326677] PAT read: cpu 3, 0x0
>>
>> Okay, so VMWare doesn't seem to return the correct PAT MSR value.
>>
>> I suggest you try "nopat" as kernel option. This should disable all the
>> PAT handling and VMWare can't wreck the kernel this way.
>>
>> I'll write a patch which detects this VMWare bug by checking the PAT
>> value after writing it.
>>
>> Thanks for reporting that case,
>>
>>
>> Juergen
>>
>>
> 
> OK, my VMWare works with "nopat" option.
> 
> Thanks~.N