[Intel-gfx] [PATCH 07/12] lib/igt_kms: Add support for the OUT_FENCE_PTR property

Tue Nov 22 13:50:53 UTC 2016

On Tue, Nov 22, 2016 at 02:12:59PM +0100, Daniel Vetter wrote:
>On Tue, Nov 22, 2016 at 12:37:47PM +0000, Brian Starkey wrote:
>> On Tue, Nov 22, 2016 at 12:10:52PM +0000, Chris Wilson wrote:
>> > On Tue, Nov 22, 2016 at 11:54:57AM +0000, Brian Starkey wrote:
>> > > On Tue, Nov 22, 2016 at 11:06:00AM +0000, Chris Wilson wrote:
>> > > >On Tue, Nov 22, 2016 at 10:53:51AM +0000, Brian Starkey wrote:
>> > > >>Hi Gustavo,
>> > > >>
>> > > >>A little late to the party here, but I was blocked by our internal
>> > > >>contributions process...
>> > > >>
>> > > >>I didn't see these end up in my checkout yet though, so I guess they
>> > > >>aren't picked up yet.
>> > > >>
>> > > >>On Mon, Nov 14, 2016 at 06:59:21PM +0900, Gustavo Padovan wrote:
>> > > >>>From: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
>> > > >>>
>> > > >>>Add support for the OUT_FENCE_PTR property to enable setting out fences for
>> > > >>>atomic commits.
>> > > >>>
>> > > >>>Signed-off-by: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
>> > > >>>---
>> > > >>>lib/igt_kms.c | 20 +++++++++++++++++++-
>> > > >>>lib/igt_kms.h |  3 +++
>> > > >>>2 files changed, 22 insertions(+), 1 deletion(-)
>> > > >>>
>> > > >>>diff --git a/lib/igt_kms.c b/lib/igt_kms.c
>> > > >>>index 4748c0a..f25e1eb 100644
>> > > >>>--- a/lib/igt_kms.c
>> > > >>>+++ b/lib/igt_kms.c
>> > > >>>@@ -175,7 +175,8 @@ const char *igt_crtc_prop_names[IGT_NUM_CRTC_PROPS] = {
>> > > >>>	"DEGAMMA_LUT",
>> > > >>>	"GAMMA_LUT",
>> > > >>>	"MODE_ID",
>> > > >>>-	"ACTIVE"
>> > > >>>+	"ACTIVE",
>> > > >>>+	"OUT_FENCE_PTR"
>> > > >>>};
>> > > >>>
>> > > >>>const char *igt_connector_prop_names[IGT_NUM_CONNECTOR_PROPS] = {
>> > > >>>@@ -2103,6 +2104,9 @@ static void igt_atomic_prepare_crtc_commit(igt_pipe_t *pipe_obj, drmModeAtomicRe
>> > > >>>		igt_atomic_populate_crtc_req(req, pipe_obj, IGT_CRTC_ACTIVE, !!output);
>> > > >>>	}
>> > > >>>
>> > > >>>+	if (pipe_obj->out_fence_ptr)
>> > > >>>+		igt_atomic_populate_crtc_req(req, pipe_obj, IGT_CRTC_OUT_FENCE_PTR, pipe_obj->out_fence_ptr);
>> > > >>>+
>> > > >>>	/*
>> > > >>>	 *	TODO: Add all crtc level properties here
>> > > >>>	 */
>> > > >>>@@ -2683,6 +2687,20 @@ igt_pipe_set_gamma_lut(igt_pipe_t *pipe, void *ptr, size_t length)
>> > > >>>}
>> > > >>>
>> > > >>>/**
>> > > >>>+ * igt_pipe_set_out_fence_ptr:
>> > > >>>+ * @pipe: pipe pointer to which background color to be set
>> > > >>>+ * @fence_ptr: out fence pointer
>> > > >>
>> > > >>I don't think fence_ptr can be int *. It needs to be a pointer to a
>> > > >>64-bit type.
>> > > >>
>> > > >>>+ *
>> > > >>>+ * Sets the out fence pointer that will be passed to the kernel in
>> > > >>>+ * the atomic ioctl. When the kernel returns the out fence pointer
>> > > >>>+ * will contain the fd number of the out fence created by KMS.
>> > > >>>+ */
>> > > >>>+void igt_pipe_set_out_fence_ptr(igt_pipe_t *pipe, int *fence_ptr)
>> > > >>>+{
>> > > >>>+	pipe->out_fence_ptr = (uint64_t) fence_ptr;
>> > > >>>+}
>> > > >>>+
>> > > >>>+/**
>> > > >>> * igt_crtc_set_background:
>> > > >>> * @pipe: pipe pointer to which background color to be set
>> > > >>> * @background: background color value in BGR 16bpc
>> > > >>>diff --git a/lib/igt_kms.h b/lib/igt_kms.h
>> > > >>>index 344f931..02d7bd1 100644
>> > > >>>--- a/lib/igt_kms.h
>> > > >>>+++ b/lib/igt_kms.h
>> > > >>>@@ -110,6 +110,7 @@ enum igt_atomic_crtc_properties {
>> > > >>>       IGT_CRTC_GAMMA_LUT,
>> > > >>>       IGT_CRTC_MODE_ID,
>> > > >>>       IGT_CRTC_ACTIVE,
>> > > >>>+       IGT_CRTC_OUT_FENCE_PTR,
>> > > >>>       IGT_NUM_CRTC_PROPS
>> > > >>>};
>> > > >>>
>> > > >>>@@ -298,6 +299,7 @@ struct igt_pipe {
>> > > >>>
>> > > >>>	uint64_t mode_blob;
>> > > >>>	bool mode_changed;
>> > > >>>+	uint64_t out_fence_ptr;
>> > > >>
>> > > >>IMO this should be:
>> > > >>
>> > > >>	int64_t *out_fence_ptr;
>> > > >
>> > > >In userspace, fences are *fd*, a plain int. It is only the uabi that we
>> > > >pass pointers as u64 to the kernel, and indeed that should be limited to
>> > > >the uabi wrapper.
>> > > >-Chris
>> > >
>> > > Where's the uabi wrapper in this case?
>> > >
>> > > Wherever it is, afaik someone needs to have 64-bit type for the kernel
>> > > to stash its fd in - on the kernel side out_fence_ptr is
>> > > (s64 __user *), so if there's not a 64-bit variable on the other end
>> > > of it then someone's going to have a bad day.
>> >
>> > We do not have pointers in the uabi because they are different sizes on
>> > different platforms. The uabi must be a u64 representation of a user
>> > address to store the result - that is what we pass to the crtc set
>> > property ioctl.
>>
>> Sure, but igt_pipe is not a uabi structure. By storing a uint64_t here
>> we're making it needlessly opaque what the value is actually meant to
>> be - which is the address of a 64-bit signed integer.
>>
>> Regardless, tests cannot set out_fence_ptr to the address of an int, I
>> hope we can agree on that. Where that detail gets taken care of I
>> don't much mind - but this code as-is is incorrect.
>>
>> By making igt_pipe.out_fence_ptr an (int64_t *) I thought we'd be
>> letting the compiler warn anyone else away from incorrect code.
>>
>> > That it has been futher managled not to pass around fd
>> > is an interesting twist, but ideally that sillyness should not make
>> > itself into our API.
>>
>> Allowing the kernel and userspace to have different ideas about how
>> big an int is doesn't sound so silly to me. It may not be a
>> theoretical problem forever.
>
>What Chris means is that you want to have an int out_fence in igt_pipe,
>and just pass the address of that into the OUT_FENCE_PTR property.

Storing the fence itself in igt_pipe instead of a pointer to it is a
different matter (and it isn't what's implemented in this patch).

It still doesn't change the fact that you can't do as you suggest -
you cannot just pass the address of an int in the OUT_FENCE_PTR
property.

In the kernel, put_user(fd, fence_state[i].out_fence_ptr); is going to
write 8 bytes. If out_fence_ptr is the address of a 4-byte variable,
then obviously that's not going to work out so well.

>In
>userspace we want to directly handle the fd, not a pointer to an fd. Like
>Chris explained, the pointer-to-fd-cast-to-u64 is just to be able to reuse
>the atomic ioctl as transport, it's not a reasonable interface within
>userspace.

I don't really follow this bit. At some point, something in userspace
is going to need to take care of the fact that the kernel needs to
have an 8-byte container to write into.

-Brian

>-Daniel
>-- 
>Daniel Vetter
>Software Engineer, Intel Corporation
>http://blog.ffwll.ch