| Message ID | 20190429220825.156644-1-olvaffe@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/3] drm/virtio: set seqno for dma-fence | expand |
(Add missing CCs) On Mon, Apr 29, 2019 at 3:08 PM Chia-I Wu <olvaffe@gmail.com> wrote: > > This is motivated by having meaningful ftrace events, but it also > fixes use cases where dma_fence_is_later is called, such as in > sync_file_merge. > > In other drivers, fence creation and cmdbuf submission normally > happen atomically, > > mutex_lock(); > fence = dma_fence_create(..., ++timeline->seqno); > submit_cmdbuf(); > mutex_unlock(); > > and have no such issue. But in our driver, because most ioctls > queue commands into ctrlq, we do not want to grab a lock. Instead, > we set seqno to 0 when a fence is created, and update it when the > command is finally queued and the seqno is known. > > Signed-off-by: Chia-I Wu <olvaffe@gmail.com> > --- > drivers/gpu/drm/virtio/virtgpu_drv.h | 1 - > drivers/gpu/drm/virtio/virtgpu_fence.c | 17 ++++++++++------- > 2 files changed, 10 insertions(+), 8 deletions(-) > > diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.h b/drivers/gpu/drm/virtio/virtgpu_drv.h > index 491dec0712b3..90461feeafdb 100644 > --- a/drivers/gpu/drm/virtio/virtgpu_drv.h > +++ b/drivers/gpu/drm/virtio/virtgpu_drv.h > @@ -102,7 +102,6 @@ struct virtio_gpu_fence { > struct dma_fence f; > struct virtio_gpu_fence_driver *drv; > struct list_head node; > - uint64_t seq; > }; > #define to_virtio_fence(x) \ > container_of(x, struct virtio_gpu_fence, f) > diff --git a/drivers/gpu/drm/virtio/virtgpu_fence.c b/drivers/gpu/drm/virtio/virtgpu_fence.c > index 87d1966192f4..72b4f7561432 100644 > --- a/drivers/gpu/drm/virtio/virtgpu_fence.c > +++ b/drivers/gpu/drm/virtio/virtgpu_fence.c > @@ -40,16 +40,14 @@ bool virtio_fence_signaled(struct dma_fence *f) > { > struct virtio_gpu_fence *fence = to_virtio_fence(f); > > - if (atomic64_read(&fence->drv->last_seq) >= fence->seq) > + if (atomic64_read(&fence->drv->last_seq) >= fence->f.seqno) > return true; > return false; > } > > static void virtio_fence_value_str(struct dma_fence *f, char *str, int size) > { > - struct virtio_gpu_fence *fence = to_virtio_fence(f); > - > - snprintf(str, size, "%llu", fence->seq); > + snprintf(str, size, "%llu", f->seqno); > } > > static void virtio_timeline_value_str(struct dma_fence *f, char *str, int size) > @@ -76,6 +74,11 @@ struct virtio_gpu_fence *virtio_gpu_fence_alloc(struct virtio_gpu_device *vgdev) > return fence; > > fence->drv = drv; > + > + /* This only partially initializes the fence because the seqno is > + * unknown yet. The fence must not be used outside of the driver > + * until virtio_gpu_fence_emit is called. > + */ > dma_fence_init(&fence->f, &virtio_fence_ops, &drv->lock, drv->context, 0); > > return fence; > @@ -89,13 +92,13 @@ int virtio_gpu_fence_emit(struct virtio_gpu_device *vgdev, > unsigned long irq_flags; > > spin_lock_irqsave(&drv->lock, irq_flags); > - fence->seq = ++drv->sync_seq; > + fence->f.seqno = ++drv->sync_seq; > dma_fence_get(&fence->f); > list_add_tail(&fence->node, &drv->fences); > spin_unlock_irqrestore(&drv->lock, irq_flags); > > cmd_hdr->flags |= cpu_to_le32(VIRTIO_GPU_FLAG_FENCE); > - cmd_hdr->fence_id = cpu_to_le64(fence->seq); > + cmd_hdr->fence_id = cpu_to_le64(fence->f.seqno); > return 0; > } > > @@ -109,7 +112,7 @@ void virtio_gpu_fence_event_process(struct virtio_gpu_device *vgdev, > spin_lock_irqsave(&drv->lock, irq_flags); > atomic64_set(&vgdev->fence_drv.last_seq, last_seq); > list_for_each_entry_safe(fence, tmp, &drv->fences, node) { > - if (last_seq < fence->seq) > + if (last_seq < fence->f.seqno) > continue; > dma_fence_signal_locked(&fence->f); > list_del(&fence->node); > -- > 2.21.0.593.g511ec345e18-goog >
Hi Chia-I, On Mon, 29 Apr 2019 at 23:08, Chia-I Wu <olvaffe@gmail.com> wrote: > > This is motivated by having meaningful ftrace events, but it also > fixes use cases where dma_fence_is_later is called, such as in > sync_file_merge. > > In other drivers, fence creation and cmdbuf submission normally > happen atomically, > > mutex_lock(); > fence = dma_fence_create(..., ++timeline->seqno); > submit_cmdbuf(); > mutex_unlock(); > > and have no such issue. But in our driver, because most ioctls > queue commands into ctrlq, we do not want to grab a lock. Instead, > we set seqno to 0 when a fence is created, and update it when the > command is finally queued and the seqno is known. > > Signed-off-by: Chia-I Wu <olvaffe@gmail.com> The series looks great. For the lot: Reviewed-by: Emil Velikov <emil.velikov@collabora.com> -Emil
On Mon, Apr 29, 2019 at 03:08:23PM -0700, Chia-I Wu wrote: > This is motivated by having meaningful ftrace events, but it also > fixes use cases where dma_fence_is_later is called, such as in > sync_file_merge. > > In other drivers, fence creation and cmdbuf submission normally > happen atomically, > > mutex_lock(); > fence = dma_fence_create(..., ++timeline->seqno); > submit_cmdbuf(); > mutex_unlock(); > > and have no such issue. But in our driver, because most ioctls > queue commands into ctrlq, we do not want to grab a lock. Instead, > we set seqno to 0 when a fence is created, and update it when the > command is finally queued and the seqno is known. Series pushed to drm-misc-next. thanks, Gerd
diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.h b/drivers/gpu/drm/virtio/virtgpu_drv.h index 491dec0712b3..90461feeafdb 100644 --- a/drivers/gpu/drm/virtio/virtgpu_drv.h +++ b/drivers/gpu/drm/virtio/virtgpu_drv.h @@ -102,7 +102,6 @@ struct virtio_gpu_fence { struct dma_fence f; struct virtio_gpu_fence_driver *drv; struct list_head node; - uint64_t seq; }; #define to_virtio_fence(x) \ container_of(x, struct virtio_gpu_fence, f) diff --git a/drivers/gpu/drm/virtio/virtgpu_fence.c b/drivers/gpu/drm/virtio/virtgpu_fence.c index 87d1966192f4..72b4f7561432 100644 --- a/drivers/gpu/drm/virtio/virtgpu_fence.c +++ b/drivers/gpu/drm/virtio/virtgpu_fence.c @@ -40,16 +40,14 @@ bool virtio_fence_signaled(struct dma_fence *f) { struct virtio_gpu_fence *fence = to_virtio_fence(f); - if (atomic64_read(&fence->drv->last_seq) >= fence->seq) + if (atomic64_read(&fence->drv->last_seq) >= fence->f.seqno) return true; return false; } static void virtio_fence_value_str(struct dma_fence *f, char *str, int size) { - struct virtio_gpu_fence *fence = to_virtio_fence(f); - - snprintf(str, size, "%llu", fence->seq); + snprintf(str, size, "%llu", f->seqno); } static void virtio_timeline_value_str(struct dma_fence *f, char *str, int size) @@ -76,6 +74,11 @@ struct virtio_gpu_fence *virtio_gpu_fence_alloc(struct virtio_gpu_device *vgdev) return fence; fence->drv = drv; + + /* This only partially initializes the fence because the seqno is + * unknown yet. The fence must not be used outside of the driver + * until virtio_gpu_fence_emit is called. + */ dma_fence_init(&fence->f, &virtio_fence_ops, &drv->lock, drv->context, 0); return fence; @@ -89,13 +92,13 @@ int virtio_gpu_fence_emit(struct virtio_gpu_device *vgdev, unsigned long irq_flags; spin_lock_irqsave(&drv->lock, irq_flags); - fence->seq = ++drv->sync_seq; + fence->f.seqno = ++drv->sync_seq; dma_fence_get(&fence->f); list_add_tail(&fence->node, &drv->fences); spin_unlock_irqrestore(&drv->lock, irq_flags); cmd_hdr->flags |= cpu_to_le32(VIRTIO_GPU_FLAG_FENCE); - cmd_hdr->fence_id = cpu_to_le64(fence->seq); + cmd_hdr->fence_id = cpu_to_le64(fence->f.seqno); return 0; } @@ -109,7 +112,7 @@ void virtio_gpu_fence_event_process(struct virtio_gpu_device *vgdev, spin_lock_irqsave(&drv->lock, irq_flags); atomic64_set(&vgdev->fence_drv.last_seq, last_seq); list_for_each_entry_safe(fence, tmp, &drv->fences, node) { - if (last_seq < fence->seq) + if (last_seq < fence->f.seqno) continue; dma_fence_signal_locked(&fence->f); list_del(&fence->node);
This is motivated by having meaningful ftrace events, but it also fixes use cases where dma_fence_is_later is called, such as in sync_file_merge. In other drivers, fence creation and cmdbuf submission normally happen atomically, mutex_lock(); fence = dma_fence_create(..., ++timeline->seqno); submit_cmdbuf(); mutex_unlock(); and have no such issue. But in our driver, because most ioctls queue commands into ctrlq, we do not want to grab a lock. Instead, we set seqno to 0 when a fence is created, and update it when the command is finally queued and the seqno is known. Signed-off-by: Chia-I Wu <olvaffe@gmail.com> --- drivers/gpu/drm/virtio/virtgpu_drv.h | 1 - drivers/gpu/drm/virtio/virtgpu_fence.c | 17 ++++++++++------- 2 files changed, 10 insertions(+), 8 deletions(-)