Re: Remapping BOs and registers post BAR movements [was - Re: Question about supporting AMD eGPU hot plug case]

[Andrey Grodzovsky <andrey.grodzovsky@amd.com>]

On 2021-03-05 3:36 a.m., Christian König wrote:
> Hi Andrey,
> 
> Am 03.03.21 um 17:05 schrieb Andrey Grodzovsky:
>> Ping
>>
>> Andrey
>>
>> On 2021-03-02 12:48 p.m., Andrey Grodzovsky wrote:
>>> Adding PCI mailing list per Bjorn's advise (sorry, skipped the 
>>> address last time).
>>>
>>> On 2021-03-02 12:22 p.m., Christian König wrote:
>>>> Hi Andrey,
>>>>
>>>>> Can you elaborate more why you need such a lock on a global level 
>>>>> and not
>>>>> enough per driver (per device more specifically) local lock 
>>>>> acquired in
>>>>> rescan_preapare (before BARs shuffling starts) and released in 
>>>>> rescan_done
>>>>> (after BARs shuffling done) ? 
>>>>
>>>> because this adds inter device dependencies.
>>>>
>>>> E.g. when we have device A, B and C and each is taking it's own lock 
>>>> we end up with 3 locks.
>>>>
>>>> If those device have inter dependencies on their own, for example 
>>>> because of DMA-buf we will end up in quite a locking mess.
>>>>
>>>> Apart from that if we don't always keep the order A, B, C but 
>>>> instead end up with A, C, B lockdep will certainly start to complain.
>>>>
>>>> Regards,
>>>> Christian.
>>>
>>> But as far as I understand those locks are agnostic to each other - 
>>> each lock is taken only by it's respective device and not by others.
> 
> Well they should be agnostic to each other, but are they really? I mean 
> we could have inter device lock dependencies because of DMA-buf or good 
> knows what.
> 
>>> So I don't see deadlock danger here but, I do suspect lockdep might 
>>> throw a false warning...
> 
> Yeah and disabling or ignoring the lockdep warning could cause problems 
> because we don't see issues caused by inter device lock dependencies any 
> more.
> 
> I just have the feeling that this would cause a lot of problems.
> 
> Christian.

Sergei, can you please address Christian's suggestion bellow - quote
"
What we need instead is a upper level PCI layer rw-lock which we
can take on the read side in the driver
"

Andrey

> 
>>>
>>> Andrey
>>>
>>>>
>>>> Am 02.03.21 um 18:16 schrieb Andrey Grodzovsky:
>>>>> Per Bjorn's advise adding PCI mailing list.
>>>>>
>>>>> Christian, please reply on top of this mail and not on top the 
>>>>> earlier, original one.
>>>>>
>>>>> Andrey
>>>>>
>>>>> On 2021-03-02 11:49 a.m., Bjorn Helgaas wrote:
>>>>>> Please add linux-pci@vger.kernel.org to the cc list.  There's a 
>>>>>> lot of
>>>>>> useful context in this thread and it needs to be archived.
>>>>>> On Tue, Mar 02, 2021 at 11:37:25AM -0500, Andrey Grodzovsky wrote:
>>>>>>> Adding Sergei and Bjorn since you are suggesting some PCI related 
>>>>>>> changes.
>>>>>>>
>>>>>>> On 2021-03-02 6:28 a.m., Christian König wrote:
>>>>>>>> Yeah, I'm thinking about that as well for quite a while now.
>>>>>>>>
>>>>>>>> I'm not 100% sure what Sergei does in his patch set will work 
>>>>>>>> for us. He
>>>>>>>> basically tells the driver to stop any processing then shuffle 
>>>>>>>> around
>>>>>>>> the BARs and then tells him to start again.
>>>>>>>>
>>>>>>>> Because of the necessity to reprogram bridges there isn't much 
>>>>>>>> other way
>>>>>>>> how to do this, but we would need something like taking a lock,
>>>>>>>> returning and dropping the lock later on. If you do this for 
>>>>>>>> multiple
>>>>>>>> drivers at the same time this results in a locking nightmare.
>>>>>>>>
>>>>>>>> What we need instead is a upper level PCI layer rw-lock which we 
>>>>>>>> can
>>>>>>>> take on the read side in the driver, similar to our internal 
>>>>>>>> reset lock.
>>>>>>>
>>>>>>>
>>>>>>> Can you elaborate more why you need such a lock on a global level 
>>>>>>> and not
>>>>>>> enough per driver (per device more specifically) local lock 
>>>>>>> acquired in
>>>>>>> rescan_preapare (before BARs shuffling starts) and released in 
>>>>>>> rescan_done
>>>>>>> (after BARs shuffling done) ?
>>>>>>>
>>>>>>> Andrey
>>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>> Wiring up changing the mapping is the smaller problem in that 
>>>>>>>> picture.
>>>>>>>>
>>>>>>>> Christian.
>>>>>>>>
>>>>>>>> Am 01.03.21 um 22:14 schrieb Andrey Grodzovsky:
>>>>>>>>> Hi, I started looking into actual implementation for this, as
>>>>>>>>> reference point I am using Sege's NVME implementations (see 
>>>>>>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FYADRO-KNS%2Flinux%2Fcommit%2F58c375df086502538706ee23e8ef7c8bb5a4178f&data=04%7C01%7Candrey.grodzovsky%40amd.com%7C44bc68a68feb4b71836c08d8dd9b306a%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637503005898302093%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=nuTptAjGewG2T7tXOQr2h%2Bdr42lyg3m1%2Bfv0mJzd8tw%3D&reserved=0) 
>>>>>>>>>
>>>>>>>>> and our own amdgpu_device_resize_fb_bar. Our use case being more
>>>>>>>>> complicated then NVME's arises some questions or thoughts:
>>>>>>>>>
>>>>>>>>> Before the PCI core will move the BARs (VRAM and registers) we 
>>>>>>>>> have
>>>>>>>>> to stop the HW and SW and then unmap all MMIO mapped entitles 
>>>>>>>>> in the
>>>>>>>>> driver. This includes, registers of all types and VRAM placed BOs
>>>>>>>>> with CPU visibility. My concern is how to prevent accesses to all
>>>>>>>>> MMIO ranges between iounmap in rescan_preapare and ioremap in
>>>>>>>>> rescan_done. For register accesses we are again back to same issue
>>>>>>>>> we had with device unplug and GPU reset we are discussing with 
>>>>>>>>> Denis
>>>>>>>>> now. So here at least as first solution just to confirm the 
>>>>>>>>> feature
>>>>>>>>> works I can introduce low level (in register accessors) RW locking
>>>>>>>>> to avoid access until all register access driver pointers are
>>>>>>>>> remapped post BARs move. What more concerns me is how to - In
>>>>>>>>> rescan_preapare: collect all CPU accessible BOs placed in VRAM (i
>>>>>>>>> can use amdgpu specific list like I currently use for device
>>>>>>>>> unplug), lock them to prevent any further use of them, unpin them
>>>>>>>>> (and unmap). In rescan_preapare: remap them back and unlock 
>>>>>>>>> them for
>>>>>>>>> further use. Also, there is a difference between kernel BOs where
>>>>>>>>> indeed iounmap, ioremap should be enough and user space mapped BOs
>>>>>>>>> where it seems like in hot unplug, we need in rescan_preapare to
>>>>>>>>> only unamp them, in between drop new page faults with 
>>>>>>>>> VM_FAULT_RETRY
>>>>>>>>> and post rescan_done allow page faults to go through to refill all
>>>>>>>>> the user page tables with updated MMIO addresses.
>>>>>>>>>
>>>>>>>>> Let me know your thoughts.
>>>>>>>>>
>>>>>>>>> Andrey
>>>>>>>>>
>>>>>>>>> On 2021-02-26 4:03 p.m., Andrey Grodzovsky wrote:
>>>>>>>>>>
>>>>>>>>>> On 2021-02-26 1:44 a.m., Sergei Miroshnichenko wrote:
>>>>>>>>>>> On Thu, 2021-02-25 at 13:28 -0500, Andrey Grodzovsky wrote:
>>>>>>>>>>>> On 2021-02-25 2:00 a.m., Sergei Miroshnichenko wrote:
>>>>>>>>>>>>> On Wed, 2021-02-24 at 17:51 -0500, Andrey Grodzovsky wrote:
>>>>>>>>>>>>>> On 2021-02-24 1:23 p.m., Sergei Miroshnichenko wrote:
>>>>>>>>>>>>>>> ...
>>>>>>>>>>>>>> Are you saying that even without hot-plugging, while both 
>>>>>>>>>>>>>> nvme
>>>>>>>>>>>>>> and
>>>>>>>>>>>>>> AMD
>>>>>>>>>>>>>> card are present
>>>>>>>>>>>>>> right from boot, you still get BARs moving and MMIO ranges
>>>>>>>>>>>>>> reassigned
>>>>>>>>>>>>>> for NVME BARs
>>>>>>>>>>>>>> just because amdgpu driver will start resize of AMD card 
>>>>>>>>>>>>>> BARs and
>>>>>>>>>>>>>> this
>>>>>>>>>>>>>> will trigger NVMEs BARs move to
>>>>>>>>>>>>>> allow AMD card BARs to cover full range of VIDEO RAM ?
>>>>>>>>>>>>> Yes. Unconditionally, because it is unknown beforehand if 
>>>>>>>>>>>>> NVMe's
>>>>>>>>>>>>> BAR
>>>>>>>>>>>>> movement will help. In this particular case BAR movement is 
>>>>>>>>>>>>> not
>>>>>>>>>>>>> needed,
>>>>>>>>>>>>> but is done anyway.
>>>>>>>>>>>>>
>>>>>>>>>>>>> BARs are not moved one by one, but the kernel releases all the
>>>>>>>>>>>>> releasable ones, and then recalculates a new BAR layout to 
>>>>>>>>>>>>> fit them
>>>>>>>>>>>>> all. Kernel's algorithm is different from BIOS's, so NVME has
>>>>>>>>>>>>> appeared
>>>>>>>>>>>>> at a new place.
>>>>>>>>>>>>>
>>>>>>>>>>>>> This is triggered by following:
>>>>>>>>>>>>> - at boot, if BIOS had assigned not every BAR;
>>>>>>>>>>>>> - during pci_resize_resource();
>>>>>>>>>>>>> - during pci_rescan_bus() -- after a pciehp event or a 
>>>>>>>>>>>>> manual via
>>>>>>>>>>>>> sysfs.
>>>>>>>>>>>> By manual via sysfs you mean something like this - 'echo 1 >
>>>>>>>>>>>> /sys/bus/pci/drivers/amdgpu/0000\:0c\:00.0/remove && echo 1 >
>>>>>>>>>>>> /sys/bus/pci/rescan ' ? I am looking into how most reliably 
>>>>>>>>>>>> trigger
>>>>>>>>>>>> PCI
>>>>>>>>>>>> code to call my callbacks even without having external PCI 
>>>>>>>>>>>> cage for
>>>>>>>>>>>> GPU
>>>>>>>>>>>> (will take me some time to get it).
>>>>>>>>>>> Yeah, this is our way to go when a device can't be physically 
>>>>>>>>>>> removed
>>>>>>>>>>> or unpowered remotely. With just a bit shorter path:
>>>>>>>>>>>
>>>>>>>>>>>     sudo sh -c 'echo 1 > 
>>>>>>>>>>> /sys/bus/pci/devices/0000\:0c\:00.0/remove'
>>>>>>>>>>>     sudo sh -c 'echo 1 > /sys/bus/pci/rescan'
>>>>>>>>>>>
>>>>>>>>>>> Or, just a second command (rescan) is enough: a BAR movement 
>>>>>>>>>>> attempt
>>>>>>>>>>> will be triggered even if there were no changes in PCI topology.
>>>>>>>>>>>
>>>>>>>>>>> Serge
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Ok, able to trigger stub callbacks call after doing rescan from
>>>>>>>>>> sysfs. Also I see BARs movement
>>>>>>>>>>
>>>>>>>>>> [ 1860.968036] amdgpu 0000:09:00.0: BAR 0: assigned [mem
>>>>>>>>>> 0xc0000000-0xcfffffff 64bit pref]
>>>>>>>>>> [ 1860.968050] amdgpu 0000:09:00.0: BAR 0 updated: 0xe0000000 ->
>>>>>>>>>> 0xc0000000
>>>>>>>>>> [ 1860.968054] amdgpu 0000:09:00.0: BAR 2: assigned [mem
>>>>>>>>>> 0xd0000000-0xd01fffff 64bit pref]
>>>>>>>>>> [ 1860.968068] amdgpu 0000:09:00.0: BAR 2 updated: 0xf0000000 ->
>>>>>>>>>> 0xd0000000
>>>>>>>>>> [ 1860.968071] amdgpu 0000:09:00.0: BAR 5: assigned [mem
>>>>>>>>>> 0xf0100000-0xf013ffff]
>>>>>>>>>> [ 1860.968078] amdgpu 0000:09:00.0: BAR 5 updated: 0xfce00000 ->
>>>>>>>>>> 0xf0100000
>>>>>>>>>>
>>>>>>>>>> As expected, after the move the device becomes unresponsive as
>>>>>>>>>> unmap/ioremap wasn't done.
>>>>>>>>>>
>>>>>>>>>> Andrey
>>>>>>>>>>
>>>>>>>>
>>>>
>