Re: 5e6669387e ("of/platform: Pause/resume sync state during init .."): [ 3.192726] WARNING: CPU: 1 PID: 1 at drivers/base/core.c:688 device_links_supplier_sync_state_resume
From: Frank Rowand
Date: Tue Dec 03 2019 - 23:17:44 EST
On 12/3/19 4:50 PM, Saravana Kannan wrote:
> On Tue, Dec 3, 2019 at 1:10 PM Rob Herring <robh+dt@xxxxxxxxxx> wrote:
>>
>> On Tue, Dec 3, 2019 at 2:05 PM Saravana Kannan <saravanak@xxxxxxxxxx> wrote:
>>>
>>> On Tue, Dec 3, 2019 at 1:01 AM Frank Rowand <frowand.list@xxxxxxxxx> wrote:
>>>>
>>>> On 12/2/19 3:19 PM, Saravana Kannan wrote:
>>>>> On Sun, Dec 1, 2019 at 7:00 AM kernel test robot <lkp@xxxxxxxxx> wrote:
>>>>>>
>>>>>> Greetings,
>>>>>>
>>>>>> 0day kernel testing robot got the below dmesg and the first bad commit is
>>>>>>
>>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>>>>>>
>>>>>> commit 5e6669387e2287f25f09fd0abd279dae104cfa7e
>>>>>> Author: Saravana Kannan <saravanak@xxxxxxxxxx>
>>>>>> AuthorDate: Wed Sep 4 14:11:24 2019 -0700
>>>>>> Commit: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
>>>>>> CommitDate: Fri Oct 4 17:30:19 2019 +0200
>>>>>>
>>>>>> of/platform: Pause/resume sync state during init and of_platform_populate()
>>>>>>
>>>>>> When all the top level devices are populated from DT during kernel
>>>>>> init, the supplier devices could be added and probed before the
>>>>>> consumer devices are added and linked to the suppliers. To avoid the
>>>>>> sync_state() callback from being called prematurely, pause the
>>>>>> sync_state() callbacks before populating the devices and resume them
>>>>>> at late_initcall_sync().
>>>>>>
>>>>>> Similarly, when children devices are populated from a module using
>>>>>> of_platform_populate(), there could be supplier-consumer dependencies
>>>>>> between the children devices that are populated. To avoid the same
>>>>>> problem with sync_state() being called prematurely, pause and resume
>>>>>> sync_state() callbacks across of_platform_populate().
>>>>>>
>>>>>> Signed-off-by: Saravana Kannan <saravanak@xxxxxxxxxx>
>>>>>> Link: https://lore.kernel.org/r/20190904211126.47518-6-saravanak@xxxxxxxxxx
>>>>>> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
>>>>>>
>>>>>> fc5a251d0f driver core: Add sync_state driver/bus callback
>>>>>> 5e6669387e of/platform: Pause/resume sync state during init and of_platform_populate()
>>>>>> 81b6b96475 Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux; tag 'dma-mapping-5.5' of git://git.infradead.org/users/hch/dma-mapping
>>>>>> +-------------------------------------------------------------------------+------------+------------+------------+
>>>>>> | | fc5a251d0f | 5e6669387e | 81b6b96475 |
>>>>>> +-------------------------------------------------------------------------+------------+------------+------------+
>>>>>> | boot_successes | 30 | 0 | 0 |
>>>>>> | boot_failures | 1 | 11 | 22 |
>>>>>> | Oops:#[##] | 1 | | |
>>>>>> | EIP:unmap_vmas | 1 | | |
>>>>>> | PANIC:double_fault | 1 | | |
>>>>>> | Kernel_panic-not_syncing:Fatal_exception | 1 | | |
>>>>>> | WARNING:at_drivers/base/core.c:#device_links_supplier_sync_state_resume | 0 | 11 | 22 |
>>>>>> | EIP:device_links_supplier_sync_state_resume | 0 | 11 | 22 |
>>>>>> +-------------------------------------------------------------------------+------------+------------+------------+
>>>>>>
>>>>>> If you fix the issue, kindly add following tag
>>>>>> Reported-by: kernel test robot <lkp@xxxxxxxxx>
>>>>>>
>>>>>> [ 3.186107] OF: /testcase-data/phandle-tests/consumer-b: #phandle-cells = 2 found -1
>>>>>> [ 3.188595] platform testcase-data:testcase-device2: IRQ index 0 not found
>>>>>> [ 3.191047] ### dt-test ### end of unittest - 199 passed, 0 failed
>>>>>> [ 3.191932] ------------[ cut here ]------------
>>>>>> [ 3.192571] Unmatched sync_state pause/resume!
>>>>>> [ 3.192726] WARNING: CPU: 1 PID: 1 at drivers/base/core.c:688 device_links_supplier_sync_state_resume+0x27/0xc0
>>>>>> [ 3.195084] Modules linked in:
>>>>>> [ 3.195494] CPU: 1 PID: 1 Comm: swapper/0 Tainted: G T 5.4.0-rc1-00005-g5e6669387e228 #1
>>>>>> [ 3.196674] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
>>>>>> [ 3.197693] EIP: device_links_supplier_sync_state_resume+0x27/0xc0
>>>>>> [ 3.198680] Code: 00 00 00 3e 8d 74 26 00 57 56 31 d2 53 b8 a0 d0 d9 c1 e8 6c b6 38 00 a1 e4 d0 d9 c1 85 c0 75 13 68 84 ba c4 c1 e8 29 30 b1 ff <0f> 0b 58 eb 7f 8d 74 26 00 83 e8 01 85 c0 a3 e4 d0 d9 c1 75 6f 8b
>>>>>> [ 3.201560] EAX: 00000022 EBX: 00000000 ECX: 00000000 EDX: 00000000
>>>>>> [ 3.202466] ESI: 000001ab EDI: c02c7f80 EBP: c1e87d27 ESP: c02c7f20
>>>>>> [ 3.203301] DS: 007b ES: 007b FS: 00d8 GS: 0000 SS: 0068 EFLAGS: 00010282
>>>>>> [ 3.204258] CR0: 80050033 CR2: bfa1bf98 CR3: 01f28000 CR4: 00140690
>>>>>> [ 3.205022] DR0: 00000000 DR1: 00000000 DR2: 00000000 DR3: 00000000
>>>>>> [ 3.205919] DR6: fffe0ff0 DR7: 00000400
>>>>>> [ 3.206529] Call Trace:
>>>>>> [ 3.207011] ? of_platform_sync_state_init+0x13/0x16
>>>>>> [ 3.207719] ? do_one_initcall+0xda/0x260
>>>>>> [ 3.208247] ? kernel_init_freeable+0x110/0x197
>>>>>> [ 3.208906] ? rest_init+0x120/0x120
>>>>>> [ 3.209369] ? kernel_init+0xa/0x100
>>>>>> [ 3.209775] ? ret_from_fork+0x19/0x24
>>>>>> [ 3.210283] ---[ end trace 81d0f2d2ee65199b ]---
>>>>>> [ 3.210955] ALSA device list:
>>>>>
>>>>> Rob/Frank,
>>>>>
>>>>> This seems to be an issue with the unit test code not properly
>>>>> cleaning up the state after it's done.
>>>>>
>>>>> Specifically, unittest_data_add() setting up of_root on systems where
>>>>> there's no device tree (of_root == NULL). It doesn't clean up of_root
>>>>> after the tests are done. This affects the of_have_populated_dt() API
>>>>> that in turn affects calls to
>>>>> device_links_supplier_sync_state_pause/resume(). I think unittests
>>>>> shouldn't affect the of_have_populated_dt() API.
>>>> There are at least a couple of reasons why the unittest devicetree data
>>>> needs to remain after the point where devicetree unittests currently
>>>> complete. So cleaning up (removing the data) is not an option.
>>>>
>>>> I depend on the unittest devicetree entries still existing after the system
>>>> boots and I can log into a shell for some validation of the final result of
>>>> the devicetree data.
>>>
>>> IMHO unittests shouldn't have a residual impact on the system after
>>> they are done. So, I'll agree to disagree on this one.
>>
>> They shouldn't be enabled in a production system either. Why would you
>> want the extra boot time?
>
> Should we ask the kernel test robot folks to not enable OF unittest
No. If unittests are breaking other code I want to know that.
> then? It broke my patch, but I wouldn't be surprised if it's silently
> breaking other stuff too. I think we need to do option 4 below.
>
>>
>>>> There is also a desire for the devicetree unittests to be able to be loaded
>>>> as a module. That work is not yet scheduled, but I do not want to preclude
>>>> the possibility. If unittests are loaded from a module then they will
>>>> need some devicetree data to exist that is created in early boot. That
>>>> data will be in the devicetree when of_platform_sync_state_init() is
>>>> invoked.
>>>
>>> On a normal system, FDT is parsed and of_root is set (or not set) very
>>> early on during setup_arch() before any of the initcall levels are
>>> run. The return value of of_have_populated_dt() isn't expected to
>>> change across initcall levels. But because of the way the unittest is
>>> written (the of_root is changed at late_initcall() level) the return
>>> value of of_have_populated_dt() changes across initcall levels. I
>>> think that's a real problem with the unittest -- it's breaking API
>>> semantics.
>>
>> I think what's really desired here is a 'Am I booting using DT' call.
>
> I think the community has decided to use of_have_populated_dt() as
> that call. So, we shouldn't break it.
Have you analyzed each and every use of of_have_populated_dt() to verify
that? I have not yet looked at each of them.
The function was created with one user for a specific purpose and the use
of it has grown over the years. I was not going to modify it to have
the specific meaning of "Am I booting using DT" (thus being able to ignore
the existence of the unittest data in the devicetree) without first examining
each of the users of the of_have_populated_dt(). [[ This possible change was
one of the solutions I considered before I examined what the actual problem
leading to the WARNing was. ]]
>
>>
>>> of_have_populated_dt() is being used to check if DT is present in the
>>> system and different things are done based on that. We can't have that
>>> value change across initcall levels.
>>>
>>> Couple of thoughts:
>>> 1. Don't run unit test if there is no live DT in the system?
>>
>> That's pretty much the only case I do run. I use UML to run the tests.
>
> Ah, makes sense.
>
>>> 2. If you don't want to do (1), then at least set up the unit test
>>> data during setup_arch() instead of doing it at some initcall level?
>>
>> That further breaks making it a module. The plan is also to move to
>> kunit which probably will preclude some hacky hook into setup_arch().
>> Side effects may need to be fixed for kunit though.
>
> Yup.
>
>>> 3. Can you use overlays for the unit tests if they are loaded as a module?
>>
>> That was the idea, yes.
>>
>>
>> 4. Make running the unittests a command line option instead of running
>> if enabled. Still has side effects, but you have to explicitly run it.
I am assuming "command line option" means the kernel boot command line, not
a command line interface.
I would prefer not. It is a debug option. There is no need to add the extra
complexity of an additional switch to control it. Configure it in or configure
it out.
>
> Hmm... this is another good option. I think this should be done. Do we
> have a consensus on this?
Why would you even ask if there was consensus on something that has not
even been discussed?
>> A module would still be my preference. If only there was someone
>> interested in making everything a module... ;)>
> :)
>
>>>>> I was looking into writing a unittest patch to fix this, but I don't
>>>>> know enough about the FDT parsing code to make sure I don't leak any
>>>>> memory or free stuff that's in use. I'm not sure I can simply set
>>>>> of_root = NULL if it was NULL before the unittest started. Let me know
>>>>> how I should proceed or if you plan to write up a patch for this.
>>>>
>>>> Based on the above, "clean up" of the unittest data is not the solution.
>>>>
>>>> I haven't looked at the mechanism in device_links_supplier_sync_state_resume()
>>>> that leads to the WARN yet. But is does not seem reasonable for that code
>>>> to be so sensitive to what valid data is in the devicetree that a WARN results.
>>>
>>> Sure, I could easily fix it to work around this. But this seems to be
>>> a genuine problem with the unittest setup IMO.
>
> I'll go ahead and do this (basically always doing it instead of
> checking on of_have_populated_dt()) but I don't want us to forget this
> unittest issue.
Thank you for planning to do this fix.
The unittest issue will not be forgotten. The possible impacts of unittest on
other users of devicetree is something I am very sensitive to and have
thought about quite a bit.
-Frank
>
> -Saravana
>