[PATCH] add missing mutex lock to amdgpu_get_xgmi_hive() (v2)

[StDenis, Tom]

On 2019-01-07 12:00 p.m., Grodzovsky, Andrey wrote:
> 
> 
> On 01/07/2019 11:53 AM, StDenis, Tom wrote:
>> On 2019-01-07 11:51 a.m., Grodzovsky, Andrey wrote:
>>>
>>> On 01/07/2019 11:36 AM, StDenis, Tom wrote:
>>>> On 2019-01-07 11:33 a.m., Grodzovsky, Andrey wrote:
>>>>> On 01/07/2019 11:16 AM, Liu, Shaoyun wrote:
>>>>>> I think it's reasonable to use the hive  specific lock for hive  specific functions.
>>>>>> The changes is acked-by  Shaoyun.liu < Shaoyun.liu at amd.com>
>>>>>>
>>>>>> -----Original Message-----
>>>>>> From: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> On Behalf Of StDenis, Tom
>>>>>> Sent: Monday, January 7, 2019 10:16 AM
>>>>>> To: amd-gfx at lists.freedesktop.org
>>>>>> Cc: StDenis, Tom <Tom.StDenis at amd.com>
>>>>>> Subject: [PATCH] add missing mutex lock to amdgpu_get_xgmi_hive() (v2)
>>>>>>
>>>>>> v2: Move locks around in other functions so that this function can stand on its own.  Also only hold the hive specific lock for add/remove device instead of the driver global lock so you can't add/remove devices in parallel from one hive.
>>>>>>
>>>>>> Signed-off-by: Tom St Denis <tom.stdenis at amd.com>
>>>>>> ---
>>>>>>        drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |  2 +-
>>>>>>        drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c   | 36 ++++++++++++++--------
>>>>>>        drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.h   |  2 +-
>>>>>>        3 files changed, 25 insertions(+), 15 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>> index 39d5d058b2c7..13d8e2ad2f7a 100644
>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>> @@ -3525,7 +3525,7 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
>>>>>>        	 * by different nodes. No point also since the one node already executing
>>>>>>        	 * reset will also reset all the other nodes in the hive.
>>>>>>        	 */
>>>>>> -	hive = amdgpu_get_xgmi_hive(adev);
>>>>>> +	hive = amdgpu_get_xgmi_hive(adev, 0);
>>>>>>        	if (hive && adev->gmc.xgmi.num_physical_nodes > 1 &&
>>>>>>        	    !mutex_trylock(&hive->hive_lock))
>>>>>>        		return 0;
>>>>> Let's say i have device 0 in hive A and it just got a gpu reset and at
>>>>> the same time device 1 is being added to same have though
>>>>> amdgpu_xgmi_add_device, hive->hive_lock is acquired by this new device
>>>>> being added and so gpu reset for device 0 bails out on
>>>>> '!mutex_trylock(&hive->hive_lock))' without completing the reset.
>>>>> Also in general i feel a bit uncomfortable about the confusing
>>>>> acquisition scheme in the function and  the fact that you take the
>>>>> hive->hive_lock inside amdgpu_get_xgmi_hive but release is still outside
>>>>> of the function.
>>>> Is adding a device while resetting a device even a valid operation
>>>> anyways?
>>> In theory it's valid if you have hot pluggable devices
>>>> I think this means more so that the reset logic is broken.  Instead
>>>> there should be a per-hive reset lock that is taken and that is tested
>>>> instead.
>>>>
>>>> Tom
>>> The hive->hive_lock was added exactly for this purpose and used only for
>>> that purpose. Maybe the naming i gave it wasn't reflective of it's
>>> purpose :)
>>
>> But the add/remove should use per-hive locks not the global lock... :-)
>>
>> (I'm honestly not trying to bike shed I just thought the get_hive
>> function looked wrong :-)).
>>
>> Tom
> 
> Totally agree with you, if Shayun (who origianlly added the global
> xgmi_mutex) is fine with switching to per hive mutex then me too, I just
> point out the problem with gpu reset and as you said we then need to
> rename the existing hive_lock into reset_lock and then and another per
> hive lock to do what you propose. Also - is there a way to not take the
> hive lock inside amdgpu_get_xgmi_hive but release it outside ? AFAIK
> it's an opening for problems where people use it but forget to call
> release.

I wanted to take the per-hive lock inside get_hive because it also takes 
the global lock so that add/remove couldn't happen in parallel.

For instance, deleting the last node while adding a new node means the 
per-hive mutex could be in limbo (because remove will delete the lock).

Adding a per-hive reset lock would fix the remaining issues no?

Tom