[PATCH] drm/amdgpu: clean up and unify hw fence handling
Christian König
christian.koenig at amd.com
Thu Aug 28 07:03:42 UTC 2025
On 27.08.25 18:03, Alex Deucher wrote:
> Decouple the amdgpu fence from the amdgpu_job structure.
> This lets us clean up the separate fence ops for the embedded
> fence and other fences. This also allows us to allocate the
> vm fence up front when we allocate the job.
>
> Cc: David.Wu3 at amd.com
> Cc: christian.koenig at amd.com
> Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 2 +-
> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
> drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 111 +++-----------------
> drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 12 ++-
> drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 36 ++++---
> drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 3 +-
> drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 3 +
> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 +-
> 8 files changed, 58 insertions(+), 117 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
> index f81608330a3d0..7ea3cb6491b1b 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
> @@ -1902,7 +1902,7 @@ static void amdgpu_ib_preempt_mark_partial_job(struct amdgpu_ring *ring)
> continue;
> }
> job = to_amdgpu_job(s_job);
> - if (preempted && (&job->hw_fence.base) == fence)
> + if (preempted && (&job->hw_fence->base) == fence)
> /* mark the job as preempted */
> job->preemption_status |= AMDGPU_IB_PREEMPTED;
> }
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index e117494e80547..efd86f095323a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -6494,7 +6494,7 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
> *
> * job->base holds a reference to parent fence
> */
> - if (job && dma_fence_is_signaled(&job->hw_fence.base)) {
> + if (job && dma_fence_is_signaled(&job->hw_fence->base)) {
> job_signaled = true;
> dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
> goto skip_hw_reset;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> index 2d58aefbd68a7..0afe2427c25cc 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> @@ -45,16 +45,11 @@
> * 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 ||
> - __f->base.ops == &amdgpu_job_fence_ops)
> - return __f;
> -
> - return NULL;
> + return __f;
> }
>
> /**
> @@ -110,35 +105,18 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
> {
While at it I would also drop the "struct dma_fence **f" parameter.
> struct amdgpu_device *adev = ring->adev;
> struct dma_fence *fence;
> - struct amdgpu_fence *am_fence;
> struct dma_fence __rcu **ptr;
> uint32_t seq;
> int r;
>
> - if (!af) {
> - /* create a separate hw fence */
> - am_fence = kzalloc(sizeof(*am_fence), GFP_KERNEL);
> - if (!am_fence)
> - return -ENOMEM;
> - } else {
> - am_fence = af;
> - }
> - fence = &am_fence->base;
> - am_fence->ring = ring;
> + fence = &af->base;
> + af->ring = ring;
>
> seq = ++ring->fence_drv.sync_seq;
> - am_fence->seq = seq;
> - if (af) {
> - dma_fence_init(fence, &amdgpu_job_fence_ops,
> - &ring->fence_drv.lock,
> - adev->fence_context + ring->idx, seq);
> - /* Against remove in amdgpu_job_{free, free_cb} */
> - dma_fence_get(fence);
> - } else {
> - dma_fence_init(fence, &amdgpu_fence_ops,
> - &ring->fence_drv.lock,
> - adev->fence_context + ring->idx, seq);
> - }
> + af->seq = seq;
> + dma_fence_init(fence, &amdgpu_fence_ops,
> + &ring->fence_drv.lock,
> + adev->fence_context + ring->idx, seq);
>
> amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
> seq, flags | AMDGPU_FENCE_FLAG_INT);
> @@ -683,18 +661,22 @@ void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
Please nuke that complete function.
As far as I can see it is unecessary now and the whole concept was always a completely broken idea since it signals fences in an incorrect order.
> for (i = 0; i <= ring->fence_drv.num_fences_mask; i++) {
> ptr = &ring->fence_drv.fences[i];
> old = rcu_dereference_protected(*ptr, 1);
> - if (old && old->ops == &amdgpu_job_fence_ops) {
> + if (old) {
> + struct amdgpu_fence *af;
> struct amdgpu_job *job;
>
> /* For non-scheduler bad job, i.e. failed ib test, we need to signal
> * it right here or we won't be able to track them in fence_drv
> * and they will remain unsignaled during sa_bo free.
> */
> - job = container_of(old, struct amdgpu_job, hw_fence.base);
> - if (!job->base.s_fence && !dma_fence_is_signaled(old))
> - dma_fence_signal(old);
> - RCU_INIT_POINTER(*ptr, NULL);
> - dma_fence_put(old);
> + af = container_of(old, struct amdgpu_fence, base);
> + job = af->job;
> + if (job) {
> + if (!job->base.s_fence && !dma_fence_is_signaled(old))
> + dma_fence_signal(old);
> + RCU_INIT_POINTER(*ptr, NULL);
> + dma_fence_put(old);
> + }
> }
> }
> }
> @@ -830,13 +812,6 @@ static const char *amdgpu_fence_get_timeline_name(struct dma_fence *f)
> return (const char *)to_amdgpu_fence(f)->ring->name;
> }
>
> -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.base);
> -
> - return (const char *)to_amdgpu_ring(job->base.sched)->name;
> -}
> -
> /**
> * amdgpu_fence_enable_signaling - enable signalling on fence
> * @f: fence
> @@ -853,23 +828,6 @@ static bool amdgpu_fence_enable_signaling(struct dma_fence *f)
> return true;
> }
>
> -/**
> - * 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.base);
> -
> - 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;
> -}
> -
> /**
> * amdgpu_fence_free - free up the fence memory
> *
> @@ -885,21 +843,6 @@ static void amdgpu_fence_free(struct rcu_head *rcu)
> kfree(to_amdgpu_fence(f));
> }
>
> -/**
> - * 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.base));
> -}
> -
> /**
> * amdgpu_fence_release - callback that fence can be freed
> *
> @@ -913,19 +856,6 @@ 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,
> @@ -933,13 +863,6 @@ 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_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> index 7d9bcb72e8dd3..192ed1bd2ec63 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> @@ -128,7 +128,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
> struct amdgpu_device *adev = ring->adev;
> struct amdgpu_ib *ib = &ibs[0];
> struct dma_fence *tmp = NULL;
> - struct amdgpu_fence *af;
> + struct amdgpu_fence *af, *vm_af;
> bool need_ctx_switch;
> struct amdgpu_vm *vm;
> uint64_t fence_ctx;
> @@ -154,12 +154,15 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
> csa_va = job->csa_va;
> gds_va = job->gds_va;
> init_shadow = job->init_shadow;
> - af = &job->hw_fence;
> + af = job->hw_fence;
> /* Save the context of the job for reset handling.
> * The driver needs this so it can skip the ring
> * contents for guilty contexts.
> */
> af->context = job->base.s_fence ? job->base.s_fence->finished.context : 0;
> + vm_af = job->hw_vm_fence;
> + /* the vm fence is also part of the job's context */
> + vm_af->context = job->base.s_fence ? job->base.s_fence->finished.context : 0;
We already have the fence_ctx around as local variable, but at the moment it uses job->base.s_fence->scheduled.context.
Should be trivial to switch that to job->base.s_fence->finished.context so that we don't need to duplicate the "job->base.s_fence ? " part multiple times.
> } else {
> vm = NULL;
> fence_ctx = 0;
> @@ -167,7 +170,10 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
> csa_va = 0;
> gds_va = 0;
> init_shadow = false;
> - af = NULL;
> + af = kzalloc(sizeof(*af), GFP_NOWAIT);
> + if (!af)
> + return -ENOMEM;
> + vm_af = NULL;
> }
>
> if (!ring->sched.ready) {
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> index 311e97c96c4e0..4189a2b8cb674 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
> @@ -138,7 +138,7 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
> ring->funcs->reset) {
> dev_err(adev->dev, "Starting %s ring reset\n",
> s_job->sched->name);
> - r = amdgpu_ring_reset(ring, job->vmid, &job->hw_fence);
> + r = amdgpu_ring_reset(ring, job->vmid, job->hw_fence);
> if (!r) {
> atomic_inc(&ring->adev->gpu_reset_counter);
> dev_err(adev->dev, "Ring %s reset succeeded\n",
> @@ -185,6 +185,8 @@ int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm,
> struct drm_sched_entity *entity, void *owner,
> unsigned int num_ibs, struct amdgpu_job **job)
> {
> + struct amdgpu_fence *af;
> +
> if (num_ibs == 0)
> return -EINVAL;
>
> @@ -192,6 +194,23 @@ int amdgpu_job_alloc(struct amdgpu_device *adev, struct amdgpu_vm *vm,
> if (!*job)
> return -ENOMEM;
>
> + af = kzalloc(sizeof(struct amdgpu_fence), GFP_KERNEL);
> + if (!af) {
> + kfree(*job);
> + return -ENOMEM;
> + }
> + /* assign the job so we can clean up in amdgpu_fence_driver_clear_job_fences() */
> + af->job = *job;
That becomes unecessary when we nuke amdgpu_fence_driver_clear_job_fences().
> + (*job)->hw_fence = af;
> +
> + af = kzalloc(sizeof(struct amdgpu_fence), GFP_KERNEL);
> + if (!af) {
> + kfree((*job)->hw_fence);
> + kfree(*job);
> + return -ENOMEM;
> + }
Not a must have, but goto cleanup would probably look nicer.
> + (*job)->hw_vm_fence = af;
> +
> (*job)->vm = vm;
>
> amdgpu_sync_create(&(*job)->explicit_sync);
> @@ -251,8 +270,8 @@ void amdgpu_job_free_resources(struct amdgpu_job *job)
> /* Check if any fences where initialized */
> if (job->base.s_fence && job->base.s_fence->finished.ops)
> f = &job->base.s_fence->finished;
> - else if (job->hw_fence.base.ops)
> - f = &job->hw_fence.base;
> + else if (job->hw_fence)
> + f = &job->hw_fence->base;
> else
> f = NULL;
>
> @@ -268,11 +287,7 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
>
> amdgpu_sync_free(&job->explicit_sync);
>
> - /* only put the hw fence if has embedded fence */
> - if (!job->hw_fence.base.ops)
> - kfree(job);
> - else
> - dma_fence_put(&job->hw_fence.base);
> + kfree(job);
> }
>
> void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
> @@ -301,10 +316,7 @@ void amdgpu_job_free(struct amdgpu_job *job)
> if (job->gang_submit != &job->base.s_fence->scheduled)
> dma_fence_put(job->gang_submit);
>
> - if (!job->hw_fence.base.ops)
> - kfree(job);
> - else
> - dma_fence_put(&job->hw_fence.base);
> + kfree(job);
> }
>
> struct dma_fence *amdgpu_job_submit(struct amdgpu_job *job)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
> index 931fed8892cc1..077b2414a24b7 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
> @@ -48,7 +48,8 @@ struct amdgpu_job {
> struct drm_sched_job base;
> struct amdgpu_vm *vm;
> struct amdgpu_sync explicit_sync;
> - struct amdgpu_fence hw_fence;
> + struct amdgpu_fence *hw_fence;
> + struct amdgpu_fence *hw_vm_fence;
> struct dma_fence *gang_submit;
> uint32_t preamble_status;
> uint32_t preemption_status;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> index 7670f5d82b9e4..ceadd7e4bdb58 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> @@ -143,6 +143,9 @@ struct amdgpu_fence {
> struct amdgpu_ring *ring;
> ktime_t start_timestamp;
>
> + /* store the job for cleanup */
> + struct amdgpu_job *job;
> +
That can be dropped as well.
Regards,
Christian.
> /* wptr for the fence for resets */
> u64 wptr;
> /* fence context for resets */
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index bf42246a3db2f..a287718506aa6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -772,7 +772,6 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
> bool cleaner_shader_needed = false;
> bool pasid_mapping_needed = false;
> struct dma_fence *fence = NULL;
> - struct amdgpu_fence *af;
> unsigned int patch;
> int r;
>
> @@ -835,12 +834,9 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job,
> }
>
> if (vm_flush_needed || pasid_mapping_needed || cleaner_shader_needed) {
> - r = amdgpu_fence_emit(ring, &fence, NULL, 0);
> + r = amdgpu_fence_emit(ring, &fence, job->hw_vm_fence, 0);
> if (r)
> return r;
> - /* this is part of the job's context */
> - af = container_of(fence, struct amdgpu_fence, base);
> - af->context = job->base.s_fence ? job->base.s_fence->finished.context : 0;
> }
>
> if (vm_flush_needed) {
More information about the amd-gfx
mailing list