2024-03-18 04:13:00

by wangxiaoming321

Subject: [PATCH] drm/xe: Failed getting VMA cause display stuck

The failure of binding VMA is duing to interrupt,
So it needs to retry while return fail.

Signed-off-by: wangxiaoming321 <[email protected]>
---
drivers/gpu/drm/xe/xe_vm.c | 3 +++
1 file changed, 3 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
index 99aa5ffb0ef1..d33476b631e1 100644
--- a/drivers/gpu/drm/xe/xe_vm.c
+++ b/drivers/gpu/drm/xe/xe_vm.c
@@ -2621,6 +2621,9 @@ static int __xe_vma_op_execute(struct xe_vm *vm, struct xe_vma *vma,
}
drm_exec_fini(&exec);

+ if (err == -ERESTARTSYS)
+ goto retry_userptr;
+
if (err == -EAGAIN) {
lockdep_assert_held_write(&vm->lock);

--
2.25.1

2024-03-18 04:46:42

by Matthew Brost

Subject: Re: [PATCH] drm/xe: Failed getting VMA cause display stuck

On Mon, Mar 18, 2024 at 12:12:44PM +0800, wangxiaoming321 wrote:
> The failure of binding VMA is duing to interrupt,
> So it needs to retry while return fail.
>
> Signed-off-by: wangxiaoming321 <[email protected]>
> ---
> drivers/gpu/drm/xe/xe_vm.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index 99aa5ffb0ef1..d33476b631e1 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -2621,6 +2621,9 @@ static int __xe_vma_op_execute(struct xe_vm *vm, struct xe_vma *vma,
> }
> drm_exec_fini(&exec);
>
> + if (err == -ERESTARTSYS)
> + goto retry_userptr;

This doesn't look right. If the user presses ctrl-c we'd restart?

Matt

> +
> if (err == -EAGAIN) {
> lockdep_assert_held_write(&vm->lock);
>
> --
> 2.25.1
>

2024-03-18 05:25:17

by Matthew Brost

Subject: Re: [PATCH] drm/xe: Failed getting VMA cause display stuck

On Mon, Mar 18, 2024 at 04:43:12AM +0000, Matthew Brost wrote:
> On Mon, Mar 18, 2024 at 12:12:44PM +0800, wangxiaoming321 wrote:
> > The failure of binding VMA is duing to interrupt,
> > So it needs to retry while return fail.
> >
> > Signed-off-by: wangxiaoming321 <[email protected]>
> > ---
> > drivers/gpu/drm/xe/xe_vm.c | 3 +++
> > 1 file changed, 3 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> > index 99aa5ffb0ef1..d33476b631e1 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.c
> > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > @@ -2621,6 +2621,9 @@ static int __xe_vma_op_execute(struct xe_vm *vm, struct xe_vma *vma,
> > }
> > drm_exec_fini(&exec);
> >
> > + if (err == -ERESTARTSYS)
> > + goto retry_userptr;
>
> This doesn't look right. If the user presses ctrl-c we'd restart?

Typo above. s/restart/retry.

To be clear this patch seems incorrect. This is traced to
dma_fence_wait_timeout, if it is interupted, that function return
-ERESTARTSYS which would should returned to user space via the IOCTL
failing. That is what the exisiting code should be doing.

Matt

>
> Matt
>
> > +
> > if (err == -EAGAIN) {
> > lockdep_assert_held_write(&vm->lock);
> >
> > --
> > 2.25.1
> >