Nouveau - Nov 2023 - [PATCH] driver: gpu: Fixing warning directly dereferencing a rcu pointer

Hey,

Den 2023-11-13 kl. 09:10, skrev Abhinav Singh:
> This patch fixes a sparse warning with this message
> "warning:dereference of noderef expression". In this context it
means we
> are dereferencing a __rcu tagged pointer directly.
>
> We should not be directly dereferencing a rcu pointer, rather we should
> be using rcu helper function rcu_dereferece() inside rcu read critical
> section to get a normal pointer which can be dereferenced.
>
> I tested with qemu with this command
> qemu-system-x86_64 \
> 	-m 2G \
> 	-smp 2 \
> 	-kernel bzImage \
> 	-append "console=ttyS0 root=/dev/sda earlyprintk=serial
net.ifnames=0" \
> 	-drive file=bullseye.img,format=raw \
> 	-net user,host=10.0.2.10,hostfwd=tcp:127.0.0.1:10021-:22 \
> 	-net nic,model=e1000 \
> 	-enable-kvm \
> 	-nographic \
> 	-pidfile vm.pid \
> 	2>&1 | tee vm.log
> with lockdep enabled.
>
> Signed-off-by: Abhinav Singh <singhabhinav9051571833 at gmail.com>
> ---
>   drivers/gpu/drm/nouveau/nv04_fence.c | 4 +++-
>   1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/nouveau/nv04_fence.c
b/drivers/gpu/drm/nouveau/nv04_fence.c
> index 5b71a5a5cd85..e62bad1ac720 100644
> --- a/drivers/gpu/drm/nouveau/nv04_fence.c
> +++ b/drivers/gpu/drm/nouveau/nv04_fence.c
> @@ -39,7 +39,9 @@ struct nv04_fence_priv {
>   static int
>   nv04_fence_emit(struct nouveau_fence *fence)
>   {
> -	struct nvif_push *push = fence->channel->chan.push;
> +	rcu_read_lock();
> +	struct nvif_push *push =
rcu_dereference(fence->channel)->chan.push;
> +	rcu_read_unlock();
>   	int ret = PUSH_WAIT(push, 2);
>   	if (ret == 0) {
>   		PUSH_NVSQ(push, NV_SW, 0x0150, fence->base.seqno);
I'm not an expert in nouveau fence channel lifetime, but I'm pretty sure
this should probably be a rcu_dereference_protected, since a fence 
likely can't lose its channel before its command to signal is submitted.

But in case it's not, I would at least advise to check for 
fence->channel lifetime before submitting a patch like this. At least 
the original code warned about it not being 100% correct.

Cheers,

~Maarten

On 11/13/23 09:24, Maarten Lankhorst wrote:
> Hey,
> 
> Den 2023-11-13 kl. 09:10, skrev Abhinav Singh:
>> This patch fixes a sparse warning with this message
>> "warning:dereference of noderef expression". In this context
it means we
>> are dereferencing a __rcu tagged pointer directly.
>>
>> We should not be directly dereferencing a rcu pointer, rather we should
>> be using rcu helper function rcu_dereferece() inside rcu read critical
>> section to get a normal pointer which can be dereferenced.
>>
>> I tested with qemu with this command
>> qemu-system-x86_64 \
>> ????-m 2G \
>> ????-smp 2 \
>> ????-kernel bzImage \
>> ????-append "console=ttyS0 root=/dev/sda earlyprintk=serial
net.ifnames=0" \
>> ????-drive file=bullseye.img,format=raw \
>> ????-net user,host=10.0.2.10,hostfwd=tcp:127.0.0.1:10021-:22 \
>> ????-net nic,model=e1000 \
>> ????-enable-kvm \
>> ????-nographic \
>> ????-pidfile vm.pid \
>> ????2>&1 | tee vm.log
>> with lockdep enabled.
>>
>> Signed-off-by: Abhinav Singh <singhabhinav9051571833 at
gmail.com>
>> ---
>> ? drivers/gpu/drm/nouveau/nv04_fence.c | 4 +++-
>> ? 1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/nouveau/nv04_fence.c
b/drivers/gpu/drm/nouveau/nv04_fence.c
>> index 5b71a5a5cd85..e62bad1ac720 100644
>> --- a/drivers/gpu/drm/nouveau/nv04_fence.c
>> +++ b/drivers/gpu/drm/nouveau/nv04_fence.c
>> @@ -39,7 +39,9 @@ struct nv04_fence_priv {
>> ? static int
>> ? nv04_fence_emit(struct nouveau_fence *fence)
>> ? {
>> -??? struct nvif_push *push = fence->channel->chan.push;
>> +??? rcu_read_lock();
>> +??? struct nvif_push *push =
rcu_dereference(fence->channel)->chan.push;
>> +??? rcu_read_unlock();
>> ????? int ret = PUSH_WAIT(push, 2);
>> ????? if (ret == 0) {
>> ????????? PUSH_NVSQ(push, NV_SW, 0x0150, fence->base.seqno);
> 
> I'm not an expert in nouveau fence channel lifetime, but I'm pretty
sure this should probably be a rcu_dereference_protected, since a fence likely
can't lose its channel before its command to signal is submitted.
Yes, before nouveau_fence_emit() did not add this fence to the fence
context's
pending list ->channel doesn't need any protection. We can probably just
use
unrcu_pointer(), as in [1].

[1]
https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/nouveau/nouveau_fence.c#L210
> 
> But in case it's not, I would at least advise to check for
fence->channel lifetime before submitting a patch like this. At least the
original code warned about it not being 100% correct.
> 
> Cheers,
> 
> ~Maarten
>