[PATCH] drm/amdgpu:fix race condition

[PATCH] drm/amdgpu:fix race condition

[Monk Liu]

sequence is protected by spinlock so don't access sequence
in paramter seq when invoking this function.

~0 means to get the latest sequence number and 0 means none to
get.

Change-Id: Ib7a03f3cf5594deeb4ad333cc59b47a6bddfd1ad
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
index 6d86eae..71d5f25 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
@@ -277,6 +277,13 @@ struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
 
 	spin_lock(&ctx->ring_lock);
 
+	if (seq == ~0) {
+		seq = ctx->rings[ring->idx].sequence - 1;
+	} else if(seq == 0) {
+		spin_unlock(&ctx->ring_lock);
+		return NULL;
+	}
+
 	if (seq >= cring->sequence) {
 		spin_unlock(&ctx->ring_lock);
 		return ERR_PTR(-EINVAL);
-- 
2.7.4

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [PATCH] drm/amdgpu:fix race condition

[Christian König]

Am 10.04.2017 um 09:18 schrieb Monk Liu:
> sequence is protected by spinlock so don't access sequence
> in paramter seq when invoking this function.
>
> ~0 means to get the latest sequence number and 0 means none to
> get.
>
> Change-Id: Ib7a03f3cf5594deeb4ad333cc59b47a6bddfd1ad
> Signed-off-by: Monk Liu <Monk.Liu@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 7 +++++++
>   1 file changed, 7 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> index 6d86eae..71d5f25 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> @@ -277,6 +277,13 @@ struct fence *amdgpu_ctx_get_fence(struct amdgpu_ctx *ctx,
>   
>   	spin_lock(&ctx->ring_lock);
>   
> +	if (seq == ~0) {

That must be ~0ull, otherwise you would be checking a 32bit value here.

Additional to that please define something in amdgpu_drm.h for that value.

> +		seq = ctx->rings[ring->idx].sequence - 1;
> +	} else if(seq == 0) {
> +		spin_unlock(&ctx->ring_lock);
> +		return NULL;
> +	}
> +

That is unnecessary the "if (seq + amdgpu_sched_jobs < cring->sequence) 
{" will catch this case.

Regards,
Christian.

>   	if (seq >= cring->sequence) {
>   		spin_unlock(&ctx->ring_lock);
>   		return ERR_PTR(-EINVAL);


_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

答复: [PATCH] drm/amdgpu:fix race condition

[Liu, Monk]

Oh yeah~~ thanks 

-----邮件原件-----
发件人: Christian König [mailto:deathsimple@vodafone.de] 
发送时间: Monday, April 10, 2017 3:22 PM
收件人: Liu, Monk <Monk.Liu@amd.com>; amd-gfx@lists.freedesktop.org
主题: Re: [PATCH] drm/amdgpu:fix race condition

Am 10.04.2017 um 09:18 schrieb Monk Liu:
> sequence is protected by spinlock so don't access sequence in paramter 
> seq when invoking this function.
>
> ~0 means to get the latest sequence number and 0 means none to get.
>
> Change-Id: Ib7a03f3cf5594deeb4ad333cc59b47a6bddfd1ad
> Signed-off-by: Monk Liu <Monk.Liu@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 7 +++++++
>   1 file changed, 7 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> index 6d86eae..71d5f25 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> @@ -277,6 +277,13 @@ struct fence *amdgpu_ctx_get_fence(struct 
> amdgpu_ctx *ctx,
>   
>   	spin_lock(&ctx->ring_lock);
>   
> +	if (seq == ~0) {

That must be ~0ull, otherwise you would be checking a 32bit value here.

Additional to that please define something in amdgpu_drm.h for that value.

> +		seq = ctx->rings[ring->idx].sequence - 1;
> +	} else if(seq == 0) {
> +		spin_unlock(&ctx->ring_lock);
> +		return NULL;
> +	}
> +

That is unnecessary the "if (seq + amdgpu_sched_jobs < cring->sequence) {" will catch this case.

Regards,
Christian.

>   	if (seq >= cring->sequence) {
>   		spin_unlock(&ctx->ring_lock);
>   		return ERR_PTR(-EINVAL);


_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

答复: [PATCH] drm/amdgpu:fix race condition

[Liu, Monk]

>That is unnecessary the "if (seq + amdgpu_sched_jobs < cring->sequence) {" will catch this case.

But what if user land pass in seq 0 on purpose , but the latest cring->sequence value is only 3 ?
Without this "if (seq == 0)" the logic seems goes to a wrong direction

BR Monk

-----邮件原件-----
发件人: Christian König [mailto:deathsimple@vodafone.de] 
发送时间: Monday, April 10, 2017 3:22 PM
收件人: Liu, Monk <Monk.Liu@amd.com>; amd-gfx@lists.freedesktop.org
主题: Re: [PATCH] drm/amdgpu:fix race condition

Am 10.04.2017 um 09:18 schrieb Monk Liu:
> sequence is protected by spinlock so don't access sequence in paramter 
> seq when invoking this function.
>
> ~0 means to get the latest sequence number and 0 means none to get.
>
> Change-Id: Ib7a03f3cf5594deeb4ad333cc59b47a6bddfd1ad
> Signed-off-by: Monk Liu <Monk.Liu@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 7 +++++++
>   1 file changed, 7 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> index 6d86eae..71d5f25 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> @@ -277,6 +277,13 @@ struct fence *amdgpu_ctx_get_fence(struct 
> amdgpu_ctx *ctx,
>   
>   	spin_lock(&ctx->ring_lock);
>   
> +	if (seq == ~0) {

That must be ~0ull, otherwise you would be checking a 32bit value here.

Additional to that please define something in amdgpu_drm.h for that value.

> +		seq = ctx->rings[ring->idx].sequence - 1;
> +	} else if(seq == 0) {
> +		spin_unlock(&ctx->ring_lock);
> +		return NULL;
> +	}
> +

That is unnecessary the "if (seq + amdgpu_sched_jobs < cring->sequence) {" will catch this case.

Regards,
Christian.

>   	if (seq >= cring->sequence) {
>   		spin_unlock(&ctx->ring_lock);
>   		return ERR_PTR(-EINVAL);


_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: 答复: [PATCH] drm/amdgpu:fix race condition

[Christian König]

Am 10.04.2017 um 09:26 schrieb Liu, Monk:
>> That is unnecessary the "if (seq + amdgpu_sched_jobs < cring->sequence) {" will catch this case.
> But what if user land pass in seq 0 on purpose , but the latest cring->sequence value is only 3 ?
> Without this "if (seq == 0)" the logic seems goes to a wrong direction

If cring->sequence is only 3 cring->fences[0] will be NULL, so we will 
just return NULL anyway (We might want to add a comment explaining that).

Christian.

>
> BR Monk
>
> -----邮件原件-----
> 发件人: Christian König [mailto:deathsimple@vodafone.de]
> 发送时间: Monday, April 10, 2017 3:22 PM
> 收件人: Liu, Monk <Monk.Liu@amd.com>; amd-gfx@lists.freedesktop.org
> 主题: Re: [PATCH] drm/amdgpu:fix race condition
>
> Am 10.04.2017 um 09:18 schrieb Monk Liu:
>> sequence is protected by spinlock so don't access sequence in paramter
>> seq when invoking this function.
>>
>> ~0 means to get the latest sequence number and 0 means none to get.
>>
>> Change-Id: Ib7a03f3cf5594deeb4ad333cc59b47a6bddfd1ad
>> Signed-off-by: Monk Liu <Monk.Liu@amd.com>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 7 +++++++
>>    1 file changed, 7 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> index 6d86eae..71d5f25 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> @@ -277,6 +277,13 @@ struct fence *amdgpu_ctx_get_fence(struct
>> amdgpu_ctx *ctx,
>>    
>>    	spin_lock(&ctx->ring_lock);
>>    
>> +	if (seq == ~0) {
> That must be ~0ull, otherwise you would be checking a 32bit value here.
>
> Additional to that please define something in amdgpu_drm.h for that value.
>
>> +		seq = ctx->rings[ring->idx].sequence - 1;
>> +	} else if(seq == 0) {
>> +		spin_unlock(&ctx->ring_lock);
>> +		return NULL;
>> +	}
>> +
> That is unnecessary the "if (seq + amdgpu_sched_jobs < cring->sequence) {" will catch this case.
>
> Regards,
> Christian.
>
>>    	if (seq >= cring->sequence) {
>>    		spin_unlock(&ctx->ring_lock);
>>    		return ERR_PTR(-EINVAL);
>

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: 答复: [PATCH] drm/amdgpu:fix race condition

[Christian König]

Am 10.04.2017 um 12:16 schrieb Liu, Monk:
> Is that so ?
>
> Seq is already type of "uint64_t", GCC still compiles ~0 as 0xffffffff, but not 0xffffffffffffffff ?

You mean that it sign extend the value? Good point, that could be as well.

> I'll change to use ~0ull anyway,

Yeah, that should make it a bit more clear what happens here.

Christian.

>
> BR Monk
>
>
> -----邮件原件-----
> 发件人: Christian König [mailto:deathsimple@vodafone.de]
> 发送时间: Monday, April 10, 2017 4:10 PM
> 收件人: Liu, Monk <Monk.Liu@amd.com>; amd-gfx@lists.freedesktop.org
> 主题: Re: [PATCH] drm/amdgpu:fix race condition
>
> Am 10.04.2017 um 09:39 schrieb Monk Liu:
>> sequence is protected by spinlock so don't access sequence in paramter
>> seq when invoking this function.
>>
>> ~0 means to get the latest sequence number and 0 means none to get.
> Well you are still not using ~0ull as I said. Using ~0 here will cause 0xffffffff to be used instead of 0xffffffffffffffff.
>
> Additional to that please add a define into amdgpu_drm.h to document this interface change.
>
> Christian.
>
>> Change-Id: Ib7a03f3cf5594deeb4ad333cc59b47a6bddfd1ad
>> Signed-off-by: Monk Liu <Monk.Liu@amd.com>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 3 +++
>>    1 file changed, 3 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> index 6d86eae..c24bb90 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
>> @@ -277,6 +277,9 @@ struct fence *amdgpu_ctx_get_fence(struct
>> amdgpu_ctx *ctx,
>>    
>>    	spin_lock(&ctx->ring_lock);
>>    
>> +	if (seq == ~0)
>> +		seq = ctx->rings[ring->idx].sequence - 1;
>> +
>>    	if (seq >= cring->sequence) {
>>    		spin_unlock(&ctx->ring_lock);
>>    		return ERR_PTR(-EINVAL);
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx


_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

答复: [PATCH] drm/amdgpu:fix race condition

[Liu, Monk]

Is that so ?

Seq is already type of "uint64_t", GCC still compiles ~0 as 0xffffffff, but not 0xffffffffffffffff ?
I'll change to use ~0ull anyway, 

BR Monk


-----邮件原件-----
发件人: Christian König [mailto:deathsimple@vodafone.de] 
发送时间: Monday, April 10, 2017 4:10 PM
收件人: Liu, Monk <Monk.Liu@amd.com>; amd-gfx@lists.freedesktop.org
主题: Re: [PATCH] drm/amdgpu:fix race condition

Am 10.04.2017 um 09:39 schrieb Monk Liu:
> sequence is protected by spinlock so don't access sequence in paramter 
> seq when invoking this function.
>
> ~0 means to get the latest sequence number and 0 means none to get.

Well you are still not using ~0ull as I said. Using ~0 here will cause 0xffffffff to be used instead of 0xffffffffffffffff.

Additional to that please add a define into amdgpu_drm.h to document this interface change.

Christian.

>
> Change-Id: Ib7a03f3cf5594deeb4ad333cc59b47a6bddfd1ad
> Signed-off-by: Monk Liu <Monk.Liu@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 3 +++
>   1 file changed, 3 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> index 6d86eae..c24bb90 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c
> @@ -277,6 +277,9 @@ struct fence *amdgpu_ctx_get_fence(struct 
> amdgpu_ctx *ctx,
>   
>   	spin_lock(&ctx->ring_lock);
>   
> +	if (seq == ~0)
> +		seq = ctx->rings[ring->idx].sequence - 1;
> +
>   	if (seq >= cring->sequence) {
>   		spin_unlock(&ctx->ring_lock);
>   		return ERR_PTR(-EINVAL);


_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx