[Intel-xe] [PATCH v2 1/2] drm/xe/uapi: add exec_queue_id member to drm_xe_wait_user_fence structure
Zeng, Oak
oak.zeng at intel.com
Wed Dec 6 03:26:05 UTC 2023
See one question below. With that addressed, patch is Reviewed-by: Oak Zeng <oak.zeng at intel.com>
> -----Original Message-----
> From: Intel-xe <intel-xe-bounces at lists.freedesktop.org> On Behalf Of Bommu
> Krishnaiah
> Sent: Tuesday, December 5, 2023 9:45 PM
> To: intel-xe at lists.freedesktop.org
> Cc: Dugast, Francois <francois.dugast at intel.com>; Vivi, Rodrigo
> <rodrigo.vivi at intel.com>
> Subject: [Intel-xe] [PATCH v2 1/2] drm/xe/uapi: add exec_queue_id member to
> drm_xe_wait_user_fence structure
>
> remove the num_engines/instances members from drm_xe_wait_user_fence
> structure and add a exec_queue_id member
>
> Right now this is only checking if the engine list is sane and nothing
> else. In the end every operation with this IOCTL is a soft check.
> So, let's formalize that and only use this IOCTL to wait on the fence.
>
> exec_queue_id member will help to user space to get proper error code
> from kernel while in exec_queue reset
>
> Signed-off-by: Bommu Krishnaiah <krishnaiah.bommu at intel.com>
> Signed-off-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
> Signed-off-by: Francois Dugast <francois.dugast at intel.com>
> Acked-by: Matthew Brost <matthew.brost at intel.com>
> ---
> drivers/gpu/drm/xe/xe_wait_user_fence.c | 56 +------------------------
> include/uapi/drm/xe_drm.h | 16 ++-----
> 2 files changed, 4 insertions(+), 68 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_wait_user_fence.c
> b/drivers/gpu/drm/xe/xe_wait_user_fence.c
> index 4d5c2555ce41..447e4cfd204c 100644
> --- a/drivers/gpu/drm/xe/xe_wait_user_fence.c
> +++ b/drivers/gpu/drm/xe/xe_wait_user_fence.c
> @@ -58,29 +58,7 @@ static const enum xe_engine_class
> user_to_xe_engine_class[] = {
> [DRM_XE_ENGINE_CLASS_COMPUTE] = XE_ENGINE_CLASS_COMPUTE,
> };
>
> -static int check_hw_engines(struct xe_device *xe,
> - struct drm_xe_engine_class_instance *eci,
> - int num_engines)
> -{
> - int i;
> -
> - for (i = 0; i < num_engines; ++i) {
> - enum xe_engine_class user_class =
> - user_to_xe_engine_class[eci[i].engine_class];
> -
> - if (eci[i].gt_id >= xe->info.tile_count)
> - return -EINVAL;
> -
> - if (!xe_gt_hw_engine(xe_device_get_gt(xe, eci[i].gt_id),
> - user_class, eci[i].engine_instance, true))
> - return -EINVAL;
> - }
> -
> - return 0;
> -}
> -
> -#define VALID_FLAGS (DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP | \
> - DRM_XE_UFENCE_WAIT_FLAG_ABSTIME)
> +#define VALID_FLAGS DRM_XE_UFENCE_WAIT_FLAG_ABSTIME
> #define MAX_OP DRM_XE_UFENCE_WAIT_OP_LTE
>
> static long to_jiffies_timeout(struct xe_device *xe,
> @@ -132,12 +110,8 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
> void *data,
> struct xe_device *xe = to_xe_device(dev);
> DEFINE_WAIT_FUNC(w_wait, woken_wake_function);
> struct drm_xe_wait_user_fence *args = data;
> - struct drm_xe_engine_class_instance
> eci[XE_HW_ENGINE_MAX_INSTANCE];
> - struct drm_xe_engine_class_instance __user *user_eci =
> - u64_to_user_ptr(args->instances);
> u64 addr = args->addr;
> int err;
> - bool no_engines = args->flags &
> DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP;
> long timeout;
> ktime_t start;
>
> @@ -151,41 +125,13 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
> void *data,
> if (XE_IOCTL_DBG(xe, args->op > MAX_OP))
> return -EINVAL;
>
> - if (XE_IOCTL_DBG(xe, no_engines &&
> - (args->num_engines || args->instances)))
> - return -EINVAL;
> -
> - if (XE_IOCTL_DBG(xe, !no_engines && !args->num_engines))
> - return -EINVAL;
> -
> if (XE_IOCTL_DBG(xe, addr & 0x7))
> return -EINVAL;
>
> - if (XE_IOCTL_DBG(xe, args->num_engines >
> XE_HW_ENGINE_MAX_INSTANCE))
> - return -EINVAL;
> -
> - if (!no_engines) {
> - err = copy_from_user(eci, user_eci,
> - sizeof(struct drm_xe_engine_class_instance) *
> - args->num_engines);
> - if (XE_IOCTL_DBG(xe, err))
> - return -EFAULT;
> -
> - if (XE_IOCTL_DBG(xe, check_hw_engines(xe, eci,
> - args->num_engines)))
> - return -EINVAL;
> - }
> -
> timeout = to_jiffies_timeout(xe, args);
>
> start = ktime_get();
>
> - /*
> - * FIXME: Very simple implementation at the moment, single wait queue
> - * for everything. Could be optimized to have a wait queue for every
> - * hardware engine. Open coding as 'do_compare' can sleep which
> doesn't
> - * work with the wait_event_* macros.
> - */
Above comment suggested a optimization to use per engine based wait queue.
You deleted this comment, but it is still per device based wait queue.
Personally I think it is okay to use per device base wait queue. So I think it
Is fine to delete this comment though...
Oak
> add_wait_queue(&xe->ufence_wq, &w_wait);
> for (;;) {
> err = do_compare(addr, args->value, args->mask, args->op);
> diff --git a/include/uapi/drm/xe_drm.h b/include/uapi/drm/xe_drm.h
> index eb03a49c17a1..d922641e11e8 100644
> --- a/include/uapi/drm/xe_drm.h
> +++ b/include/uapi/drm/xe_drm.h
> @@ -1030,8 +1030,7 @@ struct drm_xe_wait_user_fence {
> /** @op: wait operation (type of comparison) */
> __u16 op;
>
> -#define DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP (1 << 0) /* e.g. Wait on
> VM bind */
> -#define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME (1 << 1)
> +#define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME (1 << 0)
> /** @flags: wait flags */
> __u16 flags;
>
> @@ -1064,17 +1063,8 @@ struct drm_xe_wait_user_fence {
> */
> __s64 timeout;
>
> - /**
> - * @num_engines: number of engine instances to wait on, must be zero
> - * when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP set
> - */
> - __u64 num_engines;
> -
> - /**
> - * @instances: user pointer to array of drm_xe_engine_class_instance to
> - * wait on, must be NULL when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP
> set
> - */
> - __u64 instances;
> + /** @exec_queue_id: exec_queue_id returned from
> xe_exec_queue_create_ioctl */
> + __u32 exec_queue_id;
>
> /** @reserved: Reserved */
> __u64 reserved[2];
> --
> 2.25.1
More information about the Intel-xe
mailing list