Message-ID: <f4de969e-69a9-34dd-5637-50685d32738a@arm.com>
Date:   Wed, 5 Oct 2022 17:15:49 +0100
From:   Robin Murphy <robin.murphy@....com>
To:     Guenter Roeck <linux@...ck-us.net>,
        Thorsten Leemhuis <regressions@...mhuis.info>
Cc:     Fenghua Yu <fenghua.yu@...el.com>, intel-gfx@...ts.freedesktop.org,
        Chris Wilson <chris@...is-wilson.co.uk>,
        Joerg Roedel <joro@...tes.org>, Will Deacon <will@...nel.org>,
        iommu@...ts.linux.dev, linux-kernel@...r.kernel.org,
        regressions@...ts.linux.dev,
        Janusz Krzysztofik <janusz.krzysztofik@...ux.intel.com>,
        Lucas De Marchi <lucas.demarchi@...el.com>,
        linux-hwmon@...r.kernel.org
Subject: Re: [core-for-CI][PATCH] iommu: Remove iova cpu hotplugging flushing

On 2022-10-05 16:25, Guenter Roeck wrote:
> On Wed, Oct 05, 2022 at 04:26:28PM +0200, Thorsten Leemhuis wrote:
>> [adding the coretemp maintainer (Fenghua Yu) and the appropriate mailing
>> list to the list of recipients, as there apparently is a coretemp bug
>> that results in a iommu change causing a regression]
>>
>> On 30.09.22 18:57, Janusz Krzysztofik wrote:
>>> I think this issue can hit any user with a platform that loads iommu and
>>> coretemp drivers.  Adding regressions@...ts.linux.dev to the loop.
>>
>> f598a497bc7d was merged for 5.13-rc1, which is quite a while ago, so at
>> least a quick revert is out of question as it might do more harm than
>> good. The authors of the commit are kinda responsible for fixing
>> situations like this; but well, did anybody ask the developers of the
>> coretemp driver kindly if they are aware of the problem and maybe even
>> willing to fix it? Doesn't look like it from here from search lore (hope
>> I didn't miss anything), so let's give it a try.
>>
>> Ciao, Thorsten
>>
>>> On Thursday, 22 September 2022 14:09:35 CEST Robin Murphy wrote:
>>>> On 22/09/2022 11:10 am, Janusz Krzysztofik wrote:
>>>>> From: Chris Wilson <chris@...is-wilson.co.uk>
>>>>>
>>>>> Manual revert of commit f598a497bc7d ("iova: Add CPU hotplug handler to
>>>>> flush rcaches").  It is trying to instantiate a cpuhp notifier from inside
>>>>> a cpuhp callback.  That code replaced intel_iommu implementation of
>>>>> flushing per-IOVA domain CPU rcaches which used a single instance of cpuhp
>>>>> held for the module lifetime.
>>>>
>>>> OK, *now* I see what's going on. It doesn't seem unreasonable to me for
>>>> bus notifiers to touch CPU hotplug - what seems more unexpected is the
>>>> coretemp driver creating and adding a platform device from inside a
>>>> hotplug callback.
> 
> It is only unexpected if it is documented that creating a platform driver
> from a hotplug callback is off limits.
> 
>>>>
>>>> Once we start trying to revert multiple unrelated bits of important
>>>> functionality from other subsystems because one driver is doing a weird
>>>> thing, maybe it's time to instead question whether that driver should be
>>>> doing a weird thing?
> 
> That isn't the point. This _used_ to work, after all. Maybe the functionality
> introduced with f598a497bc7d is important, but there is still a regression
> introduced by f598a497bc7d. Sure, maybe the coretemp driver is doing
> "a weird thing", but if some generic code is changed causing something to fail
> that previously worked, it is still a regression and the reponsibility of the
> person or team making the generic code change to fix the problems caused by
> that change.

Note that AFAICS I don't think anything's actually broken, and this is 
merely a lockdep false-positive. The coretemp device itself will not be 
associated with the IOMMU, so the IOMMU notifier will never get as far 
as taking any further locks in that particular instance.

Of course I *can* try writing the patch to fix things properly if I have 
to, but fair warning; I'm not familiar with this driver or the relevant 
hardware or the subsystem, and from a brief look it will involve some 
significant redesign that I have every chance of getting wrong. Plus I'm 
not sure I can test the hotplug stuff at all since the x86 box I have to 
hand only seems to have a single coretemp device.

The fact is, the wacky thing it's doing with platform_device_add() 
doesn't actually work *all* that well anyway:

$ sudo rmmod coretemp
$ echo 0 | sudo tee /sys/bus/platform/drivers_autoprobe
0
$ sudo modprobe coretemp

[7169271.187103] BUG: kernel NULL pointer dereference, address: 
0000000000000418
[7169271.187127] #PF: supervisor write access in kernel mode
[7169271.187131] #PF: error_code(0x0002) - not-present page
[7169271.187134] PGD 0 P4D 0
[7169271.187139] Oops: 0002 [#1] SMP PTI
[7169271.187144] CPU: 0 PID: 16 Comm: cpuhp/0 Not tainted 
5.13.0-52-generic #59~20.04.1-Ubuntu
[7169271.187150] Hardware name: LENOVO 30B6S08J03/1030, BIOS S01KT29A 
06/20/2016
[7169271.187152] RIP: 0010:create_core_data+0x3cb/0x510 [coretemp]
[7169271.187163] Code: 44 89 e7 e8 67 99 7f c8 85 c0 75 17 0f b6 45 b9 
41 83 46 24 01 69 c0 18 fc ff ff 41 03 46 08 41 89 46 04 48 8b 45 b0 4c 
63 fb <4e> 89 b4 f8 10 04 00 00 48 8b 00 41 8b 56 24 48 89 45 a0 85 d2 7e
[7169271.187167] RSP: 0018:ffffa5ddc015fd98 EFLAGS: 00010203
[7169271.187172] RAX: 0000000000000000 RBX: 0000000000000001 RCX: 
0000000000000002
[7169271.187175] RDX: 0000000000000000 RSI: ffffffff89207b30 RDI: 
ffffa5ddc015fd40
[7169271.187178] RBP: ffffa5ddc015fe00 R08: 0000000000000000 R09: 
ffff8e049c04c800
[7169271.187181] R10: 0000000000019460 R11: 0000000000000000 R12: 
0000000000000000
[7169271.187184] R13: 000000000000005f R14: ffff8e049c04c800 R15: 
0000000000000001
[7169271.187187] FS:  0000000000000000(0000) GS:ffff8e0b5f600000(0000) 
knlGS:0000000000000000
[7169271.187191] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[7169271.187194] CR2: 0000000000000418 CR3: 0000000190672002 CR4: 
00000000003706f0
[7169271.187198] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 
0000000000000000
[7169271.187200] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 
0000000000000400
[7169271.187203] Call Trace:
[7169271.187206]  <TASK>
[7169271.187212]  coretemp_cpu_online+0x14f/0x180 [coretemp]
[7169271.187220]  ? create_core_data+0x510/0x510 [coretemp]
[7169271.187226]  cpuhp_invoke_callback+0x10b/0x430
[7169271.187237]  cpuhp_thread_fun+0x92/0x150
[7169271.187244]  smpboot_thread_fn+0xd0/0x170
[7169271.187253]  ? sort_range+0x30/0x30
[7169271.187260]  kthread+0x12b/0x150
[7169271.187264]  ? set_kthread_struct+0x40/0x40
[7169271.187269]  ret_from_fork+0x22/0x30
[7169271.187280]  </TASK>

Consider that a bug report, unless of course it's documented somewhere 
that users aren't allowed to turn off autoprobe ;)

Thanks,
Robin.

Powered by blists - more mailing lists