[PATCH v2] drm/amdgpu: introduce new amdgpu_fence object to indicate the job embedded fence

Andrey Grodzovsky andrey.grodzovsky at amd.com
Tue Dec 14 17:03:51 UTC 2021


On 2021-12-14 6:15 a.m., Huang Rui wrote:
> The job embedded fence donesn't initialize the flags at
> dma_fence_init(). Then we will go a wrong way in
> amdgpu_fence_get_timeline_name callback and trigger a null pointer panic
> once we enabled the trace event here. So introduce new amdgpu_fence
> object to indicate the job embedded fence.
>
> [  156.131790] BUG: kernel NULL pointer dereference, address: 00000000000002a0
> [  156.131804] #PF: supervisor read access in kernel mode
> [  156.131811] #PF: error_code(0x0000) - not-present page
> [  156.131817] PGD 0 P4D 0
> [  156.131824] Oops: 0000 [#1] PREEMPT SMP PTI
> [  156.131832] CPU: 6 PID: 1404 Comm: sdma0 Tainted: G           OE     5.16.0-rc1-custom #1
> [  156.131842] Hardware name: Gigabyte Technology Co., Ltd. Z170XP-SLI/Z170XP-SLI-CF, BIOS F20 11/04/2016
> [  156.131848] RIP: 0010:strlen+0x0/0x20
> [  156.131859] Code: 89 c0 c3 0f 1f 80 00 00 00 00 48 01 fe eb 0f 0f b6 07 38 d0 74 10 48 83 c7 01 84 c0 74 05 48 39 f7 75 ec 31 c0 c3 48 89 f8 c3 <80> 3f 00 74 10 48 89 f8 48 83 c0 01 80 38 00 75 f7 48 29 f8 c3 31
> [  156.131872] RSP: 0018:ffff9bd0018dbcf8 EFLAGS: 00010206
> [  156.131880] RAX: 00000000000002a0 RBX: ffff8d0305ef01b0 RCX: 000000000000000b
> [  156.131888] RDX: ffff8d03772ab924 RSI: ffff8d0305ef01b0 RDI: 00000000000002a0
> [  156.131895] RBP: ffff9bd0018dbd60 R08: ffff8d03002094d0 R09: 0000000000000000
> [  156.131901] R10: 000000000000005e R11: 0000000000000065 R12: ffff8d03002094d0
> [  156.131907] R13: 000000000000001f R14: 0000000000070018 R15: 0000000000000007
> [  156.131914] FS:  0000000000000000(0000) GS:ffff8d062ed80000(0000) knlGS:0000000000000000
> [  156.131923] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [  156.131929] CR2: 000000amdgpu_job_alloc00000002a0 CR3: 000000001120a005 CR4: 00000000003706e0
> [  156.131937] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> [  156.131942] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> [  156.131949] Call Trace:
> [  156.131953]  <TASK>
> [  156.131957]  ? trace_event_raw_event_dma_fence+0xcc/0x200
> [  156.131973]  ? ring_buffer_unlock_commit+0x23/0x130
> [  156.131982]  dma_fence_init+0x92/0xb0
> [  156.131993]  amdgpu_fence_emit+0x10d/0x2b0 [amdgpu]
> [  156.132302]  amdgpu_ib_schedule+0x2f9/0x580 [amdgpu]
> [  156.132586]  amdgpu_job_run+0xed/0x220 [amdgpu]
>
> Signed-off-by: Huang Rui <ray.huang at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   1 +
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |   3 +-
>   drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c  | 117 ++++++++++++++-------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h   |   3 -
>   4 files changed, 80 insertions(+), 44 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 9f017663ac50..fcaf6e9703f9 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -444,6 +444,7 @@ struct amdgpu_sa_bo {
>   
>   int amdgpu_fence_slab_init(void);
>   void amdgpu_fence_slab_fini(void);
> +bool is_job_embedded_fence(struct dma_fence *f);
>   
>   /*
>    * IRQS.
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 5625f7736e37..444a19eb2248 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -4483,9 +4483,8 @@ int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
>   
>   			ptr = &ring->fence_drv.fences[j];
>   			old = rcu_dereference_protected(*ptr, 1);
> -			if (old && test_bit(AMDGPU_FENCE_FLAG_EMBED_IN_JOB_BIT, &old->flags)) {
> +			if (old && is_job_embedded_fence(old))
>   				RCU_INIT_POINTER(*ptr, NULL);
> -			}
>   		}
>   		/* after all hw jobs are reset, hw fence is meaningless, so force_completion */
>   		amdgpu_fence_driver_force_completion(ring);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> index 3b7e86ea7167..3a81249b5660 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> @@ -77,16 +77,28 @@ void amdgpu_fence_slab_fini(void)
>    * Cast helper
>    */
>   static const struct dma_fence_ops amdgpu_fence_ops;
> +static const struct dma_fence_ops amdgpu_job_fence_ops;
>   static inline struct amdgpu_fence *to_amdgpu_fence(struct dma_fence *f)
>   {
>   	struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
>   
> -	if (__f->base.ops == &amdgpu_fence_ops)
> +	if (__f->base.ops == &amdgpu_fence_ops ||
> +	    __f->base.ops == &amdgpu_job_fence_ops)
>   		return __f;
>   
>   	return NULL;
>   }
>   
> +bool is_job_embedded_fence(struct dma_fence *f)
> +{
> +	struct amdgpu_fence *__f = container_of(f, struct amdgpu_fence, base);
> +
> +	if (__f->base.ops == &amdgpu_job_fence_ops)
> +		return true;
> +
> +	return false;
> +}
> +
>   /**
>    * amdgpu_fence_write - write a fence value
>    *
> @@ -158,19 +170,18 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd
>   	}
>   
>   	seq = ++ring->fence_drv.sync_seq;
> -	if (job != NULL && job->job_run_counter) {
> +	if (job && job->job_run_counter) {
>   		/* reinit seq for resubmitted jobs */
>   		fence->seqno = seq;
>   	} else {
> -		dma_fence_init(fence, &amdgpu_fence_ops,
> -				&ring->fence_drv.lock,
> -				adev->fence_context + ring->idx,
> -				seq);
> -	}
> -
> -	if (job != NULL) {
> -		/* mark this fence has a parent job */
> -		set_bit(AMDGPU_FENCE_FLAG_EMBED_IN_JOB_BIT, &fence->flags);
> +		if (job)
> +			dma_fence_init(fence, &amdgpu_job_fence_ops,
> +				       &ring->fence_drv.lock,
> +				       adev->fence_context + ring->idx, seq);
> +		else
> +			dma_fence_init(fence, &amdgpu_fence_ops,
> +				       &ring->fence_drv.lock,
> +				       adev->fence_context + ring->idx, seq);
>   	}


It's probably me missing something but why can't we just move setting of 
AMDGPU_FENCE_FLAG_EMBED_IN_JOB_BIT
to before dma_fence_init here or even into amdgpu_job_alloc instead of 
all the refactoring ?

Andrey


>   
>   	amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
> @@ -643,16 +654,14 @@ static const char *amdgpu_fence_get_driver_name(struct dma_fence *fence)
>   
>   static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
>   {
> -	struct amdgpu_ring *ring;
> +	return (const char *)to_amdgpu_fence(f)->ring->name;
> +}
>   
> -	if (test_bit(AMDGPU_FENCE_FLAG_EMBED_IN_JOB_BIT, &f->flags)) {
> -		struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
> +static const char *amdgpu_job_fence_get_timeline_name(struct dma_fence *f)
> +{
> +	struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
>   
> -		ring = to_amdgpu_ring(job->base.sched);
> -	} else {
> -		ring = to_amdgpu_fence(f)->ring;
> -	}
> -	return (const char *)ring->name;
> +	return (const char *)to_amdgpu_ring(job->base.sched)->name;
>   }
>   
>   /**
> @@ -665,18 +674,25 @@ static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
>    */
>   static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
>   {
> -	struct amdgpu_ring *ring;
> +	if (!timer_pending(&to_amdgpu_fence(f)->ring->fence_drv.fallback_timer))
> +		amdgpu_fence_schedule_fallback(to_amdgpu_fence(f)->ring);
>   
> -	if (test_bit(AMDGPU_FENCE_FLAG_EMBED_IN_JOB_BIT, &f->flags)) {
> -		struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
> +	return true;
> +}
>   
> -		ring = to_amdgpu_ring(job->base.sched);
> -	} else {
> -		ring = to_amdgpu_fence(f)->ring;
> -	}
> +/**
> + * amdgpu_job_fence_enable_signaling - enable signalling on job fence
> + * @f: fence
> + *
> + * This is the simliar function with amdgpu_fence_enable_signaling above, it
> + * only handles the job embedded fence.
> + */
> +static bool amdgpu_job_fence_enable_signaling(struct dma_fence *f)
> +{
> +	struct amdgpu_job *job = container_of(f, struct amdgpu_job, hw_fence);
>   
> -	if (!timer_pending(&ring->fence_drv.fallback_timer))
> -		amdgpu_fence_schedule_fallback(ring);
> +	if (!timer_pending(&to_amdgpu_ring(job->base.sched)->fence_drv.fallback_timer))
> +		amdgpu_fence_schedule_fallback(to_amdgpu_ring(job->base.sched));
>   
>   	return true;
>   }
> @@ -692,19 +708,23 @@ static void amdgpu_fence_free(struct rcu_head *rcu)
>   {
>   	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
>   
> -	if (test_bit(AMDGPU_FENCE_FLAG_EMBED_IN_JOB_BIT, &f->flags)) {
> -	/* free job if fence has a parent job */
> -		struct amdgpu_job *job;
> -
> -		job = container_of(f, struct amdgpu_job, hw_fence);
> -		kfree(job);
> -	} else {
>   	/* free fence_slab if it's separated fence*/
> -		struct amdgpu_fence *fence;
> +	kmem_cache_free(amdgpu_fence_slab, to_amdgpu_fence(f));
> +}
>   
> -		fence = to_amdgpu_fence(f);
> -		kmem_cache_free(amdgpu_fence_slab, fence);
> -	}
> +/**
> + * amdgpu_job_fence_free - free up the job with embedded fence
> + *
> + * @rcu: RCU callback head
> + *
> + * Free up the job with embedded fence after the RCU grace period.
> + */
> +static void amdgpu_job_fence_free(struct rcu_head *rcu)
> +{
> +	struct dma_fence *f = container_of(rcu, struct dma_fence, rcu);
> +
> +	/* free job if fence has a parent job */
> +	kfree(container_of(f, struct amdgpu_job, hw_fence));
>   }
>   
>   /**
> @@ -720,6 +740,19 @@ static void amdgpu_fence_release(struct dma_fence *f)
>   	call_rcu(&f->rcu, amdgpu_fence_free);
>   }
>   
> +/**
> + * amdgpu_job_fence_release - callback that job embedded fence can be freed
> + *
> + * @f: fence
> + *
> + * This is the simliar function with amdgpu_fence_release above, it
> + * only handles the job embedded fence.
> + */
> +static void amdgpu_job_fence_release(struct dma_fence *f)
> +{
> +	call_rcu(&f->rcu, amdgpu_job_fence_free);
> +}
> +
>   static const struct dma_fence_ops amdgpu_fence_ops = {
>   	.get_driver_name = amdgpu_fence_get_driver_name,
>   	.get_timeline_name = amdgpu_fence_get_timeline_name,
> @@ -727,6 +760,12 @@ static const struct dma_fence_ops amdgpu_fence_ops = {
>   	.release = amdgpu_fence_release,
>   };
>   
> +static const struct dma_fence_ops amdgpu_job_fence_ops = {
> +	.get_driver_name = amdgpu_fence_get_driver_name,
> +	.get_timeline_name = amdgpu_job_fence_get_timeline_name,
> +	.enable_signaling = amdgpu_job_fence_enable_signaling,
> +	.release = amdgpu_job_fence_release,
> +};
>   
>   /*
>    * Fence debugfs
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> index 4d380e79752c..c29554cf6e63 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> @@ -53,9 +53,6 @@ enum amdgpu_ring_priority_level {
>   #define AMDGPU_FENCE_FLAG_INT           (1 << 1)
>   #define AMDGPU_FENCE_FLAG_TC_WB_ONLY    (1 << 2)
>   
> -/* fence flag bit to indicate the face is embedded in job*/
> -#define AMDGPU_FENCE_FLAG_EMBED_IN_JOB_BIT		(DMA_FENCE_FLAG_USER_BITS + 1)
> -
>   #define to_amdgpu_ring(s) container_of((s), struct amdgpu_ring, sched)
>   
>   #define AMDGPU_IB_POOL_SIZE	(1024 * 1024)


More information about the amd-gfx mailing list