linux-kernel - Re: [PATCH 1/1] drm/amdkfd: Fix unaligned doorbell absolute offset for gfx8

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <1fb85296-6f2a-57d7-e443-7fa92699cb59@amd.com>
Date:   Wed, 27 Sep 2023 19:59:34 +0200
From:   Shashank Sharma <shashank.sharma@....com>
To:     Felix Kuehling <felix.kuehling@....com>,
        Arvind Yadav <Arvind.Yadav@....com>, Christian.Koenig@....com,
        alexander.deucher@....com, Felix.Kuehling@....co,
        Xinhui.Pan@....com, airlied@...il.com, daniel@...ll.ch,
        Mukul Joshi <mukul.joshi@....com>
Cc:     dri-devel@...ts.freedesktop.org, amd-gfx@...ts.freedesktop.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH 1/1] drm/amdkfd: Fix unaligned doorbell absolute offset
 for gfx8


On 27/09/2023 19:27, Felix Kuehling wrote:
>
> [+Mukul]
>
> On 2023-09-27 12:16, Arvind Yadav wrote:
>> This patch is to adjust the absolute doorbell offset
>> against the doorbell id considering the doorbell
>> size of 32/64 bit.
>>
>> Cc: Christian Koenig<christian.koenig@....com>
>> Cc: Alex Deucher<alexander.deucher@....com>
>> Signed-off-by: Shashank Sharma<shashank.sharma@....com>
>> Signed-off-by: Arvind Yadav<Arvind.Yadav@....com>
>> ---
>>   drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 11 ++++++++++-
>>   1 file changed, 10 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> index 0d3d538b64eb..c327f7f604d7 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
>> @@ -407,7 +407,16 @@ static int allocate_doorbell(struct qcm_process_device *qpd,
>>   
>>   	q->properties.doorbell_off = amdgpu_doorbell_index_on_bar(dev->adev,
>>   								  qpd->proc_doorbells,
>> -								  q->doorbell_id);
>> +								  0);
>
> Looks like we're now always calling amdgpu_doorbell_index_on_bar with 
> the third parameter = 0. So we could remove that parameter. It doesn't 
> do us any good and only causes bugs if we use any non-0 value.
>
Hey Felix,

Actually this was happening because in usermode KFD library the 
doorbell-size is for non-SOC15() hardware is hard coded to 4 bytes.

We added this fix just so that the library code doesn't need to be 
changed, but can still get aligned to kernel code.

GFX Usermode queue code uses this doorbell-index parameter, and it gives 
us the right offset.


Regards

Shashank

>
>> +
>> +	/* Adjust the absolute doorbell offset against the doorbell id considering
>> +	 * the doorbell size of 32/64 bit.
>> +	 */
>> +	if (!KFD_IS_SOC15(dev))
>> +		q->properties.doorbell_off += q->doorbell_id;
>> +	else
>> +		q->properties.doorbell_off += q->doorbell_id * 2;
>
> This could be simplified and generalized as
>
> 	q->properties.doorbell_off += q->doorbell_id * dev->kfd->device_info.doorbell_size / 4;

Agree, we can do this simplification.

- Shashank

> Regards,
>   Felix
>
>
>> +
>>   	return 0;
>>   }
>>