Message-ID: <80dbe1de-c71c-4556-817d-3f06e67f38ba@amd.com>
Date:   Tue, 28 Nov 2023 00:09:41 -0600
From:   Mario Limonciello <mario.limonciello@....com>
To:     Takashi Sakamoto <o-takashi@...amocchi.jp>,
        Linux kernel regressions list <regressions@...ts.linux.dev>,
        a.mark.broadworth@...il.com, matthias.schrumpf@...enet.de,
        LKML <linux-kernel@...r.kernel.org>, aros@....com,
        bagasdotme@...il.com,
        "open list:PCI SUBSYSTEM" <linux-pci@...r.kernel.org>,
        Bjorn Helgaas <bhelgaas@...gle.com>,
        Borislav Petkov <bp@...en8.de>
Subject: Re: Regression from dcadfd7f7c74ef9ee415e072a19bdf6c085159eb

+ Boris

Maybe he has some ideas on this issue.

On 11/27/2023 23:24, Takashi Sakamoto wrote:
> Hi Mario
> 
> Following up on our last conversation, I purchase some hardware to
> attempt to retrieve outputs from serial port. Finally, I bought another
> mother board in used market which provides serial port from Super I/O
> chip (ASUS TUF Gaming X570-Plus). However, I have retrieved no helpful
> outputs yet when encountering the system reboot.

Did you up the loglevel to 8 to make sure you'll get all kernel output 
on the serial port, not just errors?

> 
> As you mentioned, I check whether PCIe AER is enabled or not in the
> running kernel (Ubuntu 23.04 linux-image-6.2.0-37-generic). It is
> certainly enabled, however I can see nothing in the output as I noted.
> 
> I experienced extra troubles relevant to AMD Ryzen machine and the issued
> PCIe device:
> 
> * ASRock X570 Phantom Gaming 4 with AMD Ryzen 5 3600X does not detect
>    the card. We can see no corresponding entry in lspci.
> * After associating the card to vfio-pci, lspci command can reboot the
>    system even if firewire-ohci driver is not loaded. I can regenerate it
>    in both Gigabyte AX370-Gaming 5/ASUS TUF Gaming X570-plus with AMD
>    Ryzen 2400G.

Rather than lspci, is it specifically config space access from sysfs? 
Does the output from the serial port change with IOMMU enabled vs disabled?

> 
> I'm plreased to see if you have extra ideas to get helpful output from
> the system. But I guess that I should start finding some workaround to
> avoid the issued access to register instead of investigating the reboot
> mechanism, sigh...
> 
> Anyway, thanks for your help. >

Can you check FCH::PM::S5_RESET_STATUS on next boot after failure has 
occurred?  It is available at MMIO FED80300 or through indirect IO 
access at 0xC0.

If MMIO doesn't work, double check FCH::PM_ISACONTROL bit 1 (described 
on page 296) to confirm if your system enables it.

The meanings of the different bits can be found in a recent PPR:
https://www.amd.com/content/dam/amd/en/documents/processor-tech-docs/programmer-references/55901_B1_pub_053.zip

Indirect IO is described on PDF page 294.

This will at least give us a hint what's going on in this case.

> 
> Takashi Sakamoto
> 
> On Wed, Nov 08, 2023 at 02:16:44PM +0900, Takashi Sakamoto wrote:
>> Hi Mario,
>>
>> On Tue, Nov 07, 2023 at 03:27:08PM -0600, Mario Limonciello wrote:
>>> +linux-pci / Bjorn
>>> On 11/7/2023 06:17, Takashi Sakamoto wrote:
>>>> Hi Mario,
>>>>
>>>> Thanks for the report.
>>>>
>>>> I apologize for the inconvenience you and your reporter facing, however
>>>> I can not avoid to say that the problem appears to be specific to the AMD
>>>> Ryzen machines.
>>>
>>> Unfortunately I don't have this 1394 hardware myself.  I was just looking at
>>> another completely unrelated issue on Bugzilla and noticed the report come
>>> up in my search and wanted to ensure it's on your radar already as the
>>> author as it's lingered a while.
>>
>> It is your misfortune to face this kind of machine trouble.
>>
>> In the report[1], Matthias Schrumpf and Mark Broadworth noted to use AMD
>> Ryzen 7 5800X on B550/X570 chipsets, and insert VT6307 in their PCIe bus.
>> I guess that the device attends PCI bridge (ASM1083) since VT6307 has PCI
>> interface only.
>>
>> We can see MCE error in another report[2]. Unfortunately, the reporter,
>> Ian Donnelly, have less suspiction about machine architecture, and never
>> provides hardware information. But I believe that it comes from AMD Ryzen
>> machine. I transcribe the error here:
>>
>> ```
>> [    0.860834] mce: [Hardware Error]: Machine check events logged
>> [    0.860834] microcode: CPU20: patch_level=0x0a201025
>> [    0.860835] microcode: CPU21: patch_level=0x0a201025
>> [    0.860836] microcode: CPU23: patch_level=0x0a201025
>> [    0.860836] microcode: CPU22: patch_level=0x0a201025
>> [    0.860837] mce: [Hardware Error]: CPU 17: Machine Check: 0 Bank 0: bc00080001010135
>> [    0.860845] fbcon: Taking over console
>> [    0.860847] mce: [Hardware Error]: TSC 0 ADDR fca000f0 MISC d012000000000000 IPID 1000b000000000
>> [    0.860854] mce: [Hardware Error]: PROCESSOR 2:a20f10 TIME 1696955537 SOCKET 0 APIC b microcode a201025
>> [    0.860860] microcode: CPU0: patch_level=0x0a201025
>> [    0.861676] microcode: Microcode Update Driver: v2.2.
>> ```
>>
>> Additionally, as I note in the PR[3], I observed cache-coherence failure
>> over memory dedicated for DMA transmission. The mapping is created by
>> `dmam_alloc_coherent()` and no need to have extra care such as streaming
>> API. However, the combination of ASM1083 and VT6307 provides me bogus
>> values from the memory in AMD Ryzen machine, and I can see no issue in
>> Intel machines.
>>
>> Essentially, the host system reboots when firewire-ohci module in guest
>> system probes the PCI device for 1394 OHCI hardware provided by PCI
>> pass-though[4].
>>
>>>> I've already received the similar report[1], and have been
>>>> investigating it in the last few weeks, then got the insight. Please take
>>>> a look at my short report about it in PR to Linus for 6.7-rc1:
>>>> https://lore.kernel.org/lkml/20231105144852.GA165906@workstation.local/
>>>>
>>>> I can confirm that I have been abe to reproduce the problem on AMD Ryzen
>>>> machine. However, it's important to note that I have not observed the
>>>> problem on the following systems:
>>>
>>> Any chance you (or anyone with the issue) has a serial output available?
>>> I think it would be really good to look at the circumstances surrounding the
>>> reboot.
>>>
>>>>
>>>> * Intel machine (Sandy Bridge and Skylake generations)
>>>> * AMD machines predating Ryzen (Sempron 145)
>>>> * Machines using different 1394 OHCI hardware from other vendors such as
>>>>     TI
>>>> * VIA VT6307 connected directly to PCI slot (i.e. without the issued
>>>>     PCIe/PCI bridge)
>>>>
>>>> Currently, I have not been able to obtain any useful debug output from
>>>> the Linux system or any hardware error reports when the system reboots.
>>>> It seems that the system reboots spontaneously. My assumption at this
>>>> point is that AMD Ryzen machines detect a specific hardware error
>>>> triggered by Ryzen machine quirk related to the combination of the Asmedia
>>>> ASM1083/1085 and VIA VT6306/6307/6308, leading to power reset.
>>>>
>>>
>>> Recent kernels have enabled PCI AER.  Could that be factoring in perhaps?
>>
>> I ordered equipments for the workflow, and waiting for shipping, since
>> my motherboard has no interface for serial output.
>>
>> (However, I predict that we can no helpful output via the interface.)
>>
>>>> I genuinely appreciate your assistance in debugging this elusive
>>>> hardware issue. If any workaround specific to AMD Ryzen machine quirk is
>>>> required in PCI driver for 1394 OHCI hardware, I'm willing to apply it.
>>>> However, it is preferable to figure out the reboot mechanism at first,
>>>> I think.
>>>
>>> Does the BIOS on these machines enable a watchdog timer?  If so, I'd suggest
>>> disabling that for a starting point.
>>   
>> For consumer use, the machine has no such function, I think. For
>> your information, this is the machine information I used:
>>
>> * Ryzen 5 2400G
>> * Gigabyte Technology Co., Ltd. AX370-Gaming 5/AX370-Gaming 5
>>      * BIOS F51h 02/09/2023
>>
>>> How about if you compile as a module and then modprobe.blacklist the module
>>> on kernel command line and load it later.  Can you trigger the fault/reboot
>>> this way?  If so, it at least rules out some conditions that happen during a
>>> race at boot.
>>
>> Nowadays FireWire software stack is optional in the most of
>> distributions. I can encounter the same issue at deferred probing enough
>> after booting up, even if the load of system is very low.
>>
>>> Looking more closely at the change, I would guess the fault is specifically
>>> in get_cycle_time().  I can see that the VIA devices do set
>>> QUIRK_CYCLE_TIMER which will cause additional reads.
>>
>> I've already tested with the driver compiled without these codes, but the
>> system reboots again.
>>
>>> Another guesses worth looking at is to see if iommu=pt or amd_iommu=off
>>> help.
>>>
>>> If either of those help it could point at being a problem with
>>> get_cycle_time() and IOMMU.  The older systems you mentioned working
>>> probably didn't enable IOMMU by default but most AMD Ryzen systems do.
>>
>> I already suspect platform IOMMU and kernel implementation, however it
>> is helpless to disable AMD SVM and IOMMU in BIOS settings. Of course, it
>> is helpless as well to provide any options to iommu in kernel command line.
>>
>> If I had any opportunity to access to AMD machines for enterprise-grade
>> usage somehow, I would have done it. However, I am a private-time
>> contributor and what I can access to is the ones for consumer use
>> without any hardware support for RAS reporting.
>>
>>
>> [1] https://bugzilla.kernel.org/show_bug.cgi?id=217993
>> [2] https://bugzilla.kernel.org/show_bug.cgi?id=217994
>> [3] https://lore.kernel.org/lkml/20231105144852.GA165906@workstation.local/
>> [4] https://lore.kernel.org/lkml/20231016155657.GA7904@workstation.local/
>>
>> Thanks
>>
>> Takashi Sakamoto

Powered by blists - more mailing lists