* [PATCH] drm/rockchip: vop2: Give each RK3588 video port its own cluster
@ 2026-11-21 7:55 Arun Iyengar via B4 Relay
2026-11-21 8:15 ` sashiko-bot
0 siblings, 1 reply; 2+ messages in thread
From: Arun Iyengar via B4 Relay @ 2026-11-21 7:55 UTC (permalink / raw)
To: Sandy Huang, Heiko Stübner, Andy Yan, Maarten Lankhorst,
Maxime Ripard, Thomas Zimmermann, David Airlie, Simona Vetter,
Sascha Hauer
Cc: dri-devel, linux-rockchip, linux-arm-kernel, linux-kernel,
Chaoyi Chen, Damon Ding, Cristian Ciocaltea, Arun Iyengar
From: Arun Iyengar <aiyengar1990@gmail.com>
Damon Ding asked for POST_BUF_EMPTY to be fixed in the layer setup, and
Chaoyi Chen pointed out that a video port's layers must stay on one AXI
port. On RK3588 Cluster0/1 fetch through AXI0 and Cluster2/3 through
AXI1. Each video port takes the first free primary window, so VP0 and
VP2 get Cluster0 and Cluster1, both on AXI0. A 4K120 and a 4K60 XR24
primary need about 6 GB/s there, one port delivers about 4.5 GB/s, and
VP0 underflows in every frame. aclk_vop at 750 or 850 MHz is the same.
On RK3588 give VP N Cluster N as primary plane. Boards without VP1 now
get Cluster2 instead of Cluster1 as VP2's primary. RK3568 and RK3576 are
unchanged.
Fixes: 5a028e8f062f ("drm/rockchip: vop2: Add support for rk3588")
Signed-off-by: Arun Iyengar <aiyengar1990@gmail.com>
---
Tested on Orange Pi 5, HDMI 4K120 + DP 4K60, v7.3 plus local HDMI FRL and
750 MHz aclk_vop (not in mainline): 7207 bad frames in 60 s before, 0 after.
Applies with Chaoyi Chen's 8K series [1]: VP0 keeps Cluster0 and Cluster1.
[1] https://lore.kernel.org/all/20261116084555.1061-4-kernel@airkyi.com/
Developed with Claude (Anthropic), which wrote the code and the commit
message. I reviewed every change.
---
drivers/gpu/drm/rockchip/rockchip_drm_vop2.c | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c b/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
index 76d151b8bc52..303d9b987a7c 100644
--- a/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
+++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
@@ -2583,6 +2583,12 @@ static bool vop2_is_mirror_win(struct vop2_win *win)
}
}
+/* RK3588 primary per video port; Cluster0/1 on AXI0, Cluster2/3 on AXI1. */
+static const u8 rk3588_vp_primary[] = {
+ ROCKCHIP_VOP2_CLUSTER0, ROCKCHIP_VOP2_CLUSTER1,
+ ROCKCHIP_VOP2_CLUSTER2, ROCKCHIP_VOP2_CLUSTER3,
+};
+
static int vop2_create_crtcs(struct vop2 *vop2)
{
const struct vop2_data *vop2_data = vop2->data;
@@ -2655,6 +2661,10 @@ static int vop2_create_crtcs(struct vop2 *vop2)
if (!(win->data->possible_vp_mask & BIT(vp->id)))
continue;
+ if (vop2->version == VOP_VERSION_RK3588 &&
+ win->data->phys_id != rk3588_vp_primary[vp->id])
+ continue;
+
if (vop2_is_mirror_win(win))
continue;
---
base-commit: c408466a4dfaee4b12587e926c5b85e68c8b3c40
change-id: 20261121-vop2-axi-primary-971fcba0d98e
Best regards,
--
Arun Iyengar <aiyengar1990@gmail.com>
^ permalink raw reply related [flat|nested] 2+ messages in thread
* Re: [PATCH] drm/rockchip: vop2: Give each RK3588 video port its own cluster
2026-11-21 7:55 [PATCH] drm/rockchip: vop2: Give each RK3588 video port its own cluster Arun Iyengar via B4 Relay
@ 2026-11-21 8:15 ` sashiko-bot
0 siblings, 0 replies; 2+ messages in thread
From: sashiko-bot @ 2026-11-21 8:15 UTC (permalink / raw)
To: Arun Iyengar; +Cc: dri-devel
Thank you for your contribution! Sashiko AI review found 2 potential
issues to consider:
- [Medium] drm/rockchip: vop2: default zpos inversion for overlays in
vop2_create_crtcs()
- [Medium] drm/rockchip: vop2: overlay planes and VP0+VP1 still violate
AXI separation
It also noticed 1 related pre-existing issue this patch does not
introduce:
- [High] drm/rockchip: vop2: port node leak and UAF on error in
vop2_create_crtcs() [1]
--
commit 7cc91c6d385e03fd7bff780807113fe9561a15db
Author: Arun Iyengar <aiyengar1990@gmail.com>
drm/rockchip: vop2: Give each RK3588 video port its own cluster
This commit assigns Cluster N as the primary plane for video port N on
RK3588 so that VP0 and VP2 do not share AXI0 for their primary planes.
> diff --git a/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c b/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
> index 76d151b8bc52b..303d9b987a7c5 100644
> --- a/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
> +++ b/drivers/gpu/drm/rockchip/rockchip_drm_vop2.c
> @@ -2583,6 +2583,12 @@ static bool vop2_is_mirror_win(struct vop2_win *win)
> }
> }
>
> +/* RK3588 primary per video port; Cluster0/1 on AXI0, Cluster2/3 on AXI1. */
> +static const u8 rk3588_vp_primary[] = {
> + ROCKCHIP_VOP2_CLUSTER0, ROCKCHIP_VOP2_CLUSTER1,
> + ROCKCHIP_VOP2_CLUSTER2, ROCKCHIP_VOP2_CLUSTER3,
> +};
> +
> static int vop2_create_crtcs(struct vop2 *vop2)
> {
> const struct vop2_data *vop2_data = vop2->data;
> @@ -2655,6 +2661,10 @@ static int vop2_create_crtcs(struct vop2 *vop2)
> if (!(win->data->possible_vp_mask & BIT(vp->id)))
> continue;
>
> + if (vop2->version == VOP_VERSION_RK3588 &&
> + win->data->phys_id != rk3588_vp_primary[vp->id])
> + continue;
> +
> if (vop2_is_mirror_win(win))
> continue;
[Severity: Medium]
When an earlier video port is unused (such as VP0 + VP2 without VP1), does
skipping Cluster1-win0 here give the first overlay plane a lower default
zpos than VP2's primary plane?
vop2_win_init() assigns win->win_id sequentially (Cluster0..3 get 0..3),
and vop2_plane_init() sets each plane's initial zpos to win->win_id:
drivers/gpu/drm/rockchip/rockchip_drm_vop2.c:vop2_plane_init() {
...
drm_plane_create_zpos_property(&win->base, win->win_id, 0,
vop2->registered_num_wins - 1);
...
}
When VP1 is disabled, VP2 claims Cluster2-win0 (default zpos 2) as its
primary plane, while the unclaimed Cluster1-win0 (default zpos 1) is
registered as the first DRM_PLANE_TYPE_OVERLAY plane in the loop below. If
userspace enables that overlay plane on VP2 or VP3 without explicitly
setting zpos, drm_atomic_normalize_zpos() places the overlay underneath the
primary plane.
[Severity: Medium]
Does filtering only the primary plane loop actually keep a video port's
layers on one AXI bus, or avoid AXI0 sharing on VP0 + VP1 boards?
In the overlay loop right below, possible_crtcs is still computed solely
from win->data->possible_vp_mask (BIT(0) | BIT(1) | BIT(2) | BIT(3) for all
eight windows in rk3588_vop_win_data[]), and neither possible_crtcs nor
atomic_check restricts a video port's planes to one axi_bus_id. On a board
with VP0 and VP2 active and VP1 disabled, the skipped Cluster1-win0
(axi_bus_id 0) becomes the first overlay plane and can be attached directly
to VP2 alongside its Cluster2-win0 (axi_bus_id 1) primary plane, as can
Esmart0/1.
In addition, because rk3588_vp_primary[] maps VP0 to Cluster0 and VP1 to
Cluster1, dual-display RK3588 boards wired to VP0 and VP1 (such as Rock 5B,
Orange Pi 5 Plus/Max, and NanoPC-T6) still put both 4K primary planes on
AXI0.
[Severity: High]
This problem wasn't introduced by this patch, but if vop2_create_crtcs()
fails after acquiring port nodes or initializing planes or CRTCs [1],
does the early return leak vp->crtc.port references and leave freed planes
and CRTCs linked in drm->mode_config?
In the first loop, of_graph_get_port_by_id() stores a device_node reference
in vp->crtc.port, and the following loops link embedded win->base and
vp->crtc objects into drm->mode_config.plane_list and crtc_list. Returning
an error without calling vop2_destroy_crtcs() leaks the port nodes and
lets component_bind() free the devm-allocated struct vop2 while its planes
and CRTCs remain on drm_device, causing a use-after-free when
rockchip_drm_bind() unwinds through drm_mode_config_cleanup().
[1] https://sashiko.dev/#/bug/linux-81b4737a-b947-4652-ad38-8f28fe971257
--
Sashiko AI review · https://sashiko.dev/#/patchset/20261121-vop2-axi-primary-v1-1-8d5c74905207@gmail.com?part=1
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-11-21 8:15 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-21 7:55 [PATCH] drm/rockchip: vop2: Give each RK3588 video port its own cluster Arun Iyengar via B4 Relay
2026-11-21 8:15 ` sashiko-bot
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox