On 30/04/2025 19:37, Devarsh Thakkar wrote:
Hi Tomi
Thanks for the review.
<snip>
@@ -2025,7 +2101,7 @@ int dispc_plane_check(struct dispc_device
*dispc, u32 hw_plane,
const struct drm_plane_state *state,
u32 hw_videoport)
{
- bool lite = dispc->feat->vid_lite[hw_plane];
+ bool lite = dispc->feat->vid_info[hw_plane].is_lite;
I don't think this is correct. You can't access the vid_info[] with the
hw-id.
I don't think hw_id is getting passed to hw_plane here. The
dispc_plane_check is called from tidss_plane_atomic_check which passes
hw_plane as tplane->hw_plane_id and this index starts from actually
instantiated planes i.e. from 0 and are contiguous as these are
Well, if tplane->hw_plane_id is not the HW plane id (i.e. it's misnamed
now), and tidss_plane.c calls dispc_plane_enable() with
tplane->hw_plane_id as the hw_plane parameter, which is used as a HW
plane ID... Then... One of these is wrong, no?
populated from vid_order array (hw_plane_id =
feat->vid_order[tidss->num_planes];) and not the hw_id index.
So for e.g. for AM62L even though hw_id is 1 for VIDL hw_plane is
getting passed as 0 and that's how it is able to access the first and
only member of vid_info struct and read the properties correctly and
function properly as seen in test logs [1].
If for AM62L the tplane->hw_plane_id is 0, the the dispc_plane_enable()
call would enable the wrong plane, wouldn't it?
But even if it all works, I think this highlights how confusing it is...
u32 fourcc = state->fb->format->format;
bool need_scaling = state->src_w >> 16 != state->crtc_w ||
state->src_h >> 16 != state->crtc_h;
@@ -2096,7 +2172,7 @@ void dispc_plane_setup(struct dispc_device
*dispc, u32 hw_plane,
const struct drm_plane_state *state,
u32 hw_videoport)
{
- bool lite = dispc->feat->vid_lite[hw_plane];
+ bool lite = dispc->feat->vid_info[hw_plane].is_lite;
Here too.
Here also hw_plane is getting passed as 0 and not the hw_id which is 1
for AM62L.
u32 fourcc = state->fb->format->format;
u16 cpp = state->fb->format->cpp[0];
u32 fb_width = state->fb->pitches[0] / cpp;
@@ -2210,7 +2286,7 @@ static void dispc_k2g_plane_init(struct
dispc_device *dispc)
/* MFLAG_START = MFLAGNORMALSTARTMODE */
REG_FLD_MOD(dispc, DISPC_GLOBAL_MFLAG_ATTRIBUTE, 0, 6, 6);
- for (hw_plane = 0; hw_plane < dispc->feat->num_planes;
hw_plane++) {
+ for (hw_plane = 0; hw_plane < dispc->feat->num_vids; hw_plane++) {
u32 size = dispc_vid_get_fifo_size(dispc, hw_plane);
u32 thr_low, thr_high;
u32 mflag_low, mflag_high;
@@ -2226,7 +2302,7 @@ static void dispc_k2g_plane_init(struct
dispc_device *dispc)
dev_dbg(dispc->dev,
"%s: bufsize %u, buf_threshold %u/%u, mflag threshold
%u/%u preload %u\n",
- dispc->feat->vid_name[hw_plane],
+ dispc->feat->vid_info[hw_plane].name,
Here hw_plane is not actually the hw-id (anymore), but elsewhere in this
function it is used as a hw-id, which is no longer correct.
For accessing vid_info hw_plane needs to be used which is the index of
actually instantiated planes and I see it as correctly being passed for
AM62L too. hw_id is only for dispc_k3_vid* functions where we need to
skip the not-instantiated vid regions by adding the offset per the hw_id
index.
Hmm, sorry, I don't follow. If we use the same variable, hw_plane, to
access the vid_info[], and as a parameter to functions that take
hw_plane, e.g., dispc_vid_set_buf_threshold(), isn't one of those uses
wrong?
Oh, wait... I think I see it now. For some functions using the hw_id as
the hw_plane parameter is fine, as they access the VID's registers by
just using, e.g. dispc_vid_write(), which gets the address correctly
from dispc->base_vid[hw_plane], as that one is indexed from 0 to num_vids.
But some functions use registers that have bits based on the hw_id (like
dispc_k3_vid_write_irqstatus), and then we use the hw_id for the
hw_plane parameter. If that function were to also write a vid register,
using the passed hw_plane, it wouldn't work, but I guess we don't do that.
It feels broken... We can't have 'hw_plane' that's sometimes the HW id
(i.e. 1 for AM62L), and sometimes the driver's index (i.e. 0 for AM62L).
size,
thr_high, thr_low,
mflag_high, mflag_low,
@@ -2265,7 +2341,7 @@ static void dispc_k3_plane_init(struct
dispc_device *dispc)
/* MFLAG_START = MFLAGNORMALSTARTMODE */
REG_FLD_MOD(dispc, DISPC_GLOBAL_MFLAG_ATTRIBUTE, 0, 6, 6);
- for (hw_plane = 0; hw_plane < dispc->feat->num_planes;
hw_plane++) {
+ for (hw_plane = 0; hw_plane < dispc->feat->num_vids; hw_plane++) {
u32 size = dispc_vid_get_fifo_size(dispc, hw_plane);
u32 thr_low, thr_high;
u32 mflag_low, mflag_high;
@@ -2281,7 +2357,7 @@ static void dispc_k3_plane_init(struct
dispc_device *dispc)
dev_dbg(dispc->dev,
"%s: bufsize %u, buf_threshold %u/%u, mflag threshold
%u/%u preload %u\n",
- dispc->feat->vid_name[hw_plane],
+ dispc->feat->vid_info[hw_plane].name,
And here.
All these issues make me wonder whether we have the right model. It's
just too easy to get the usage wrong.
I'm not sure which way to go here.
Fix the current issues? It's a bit cumbersome to go from hw-id to the
index (needs a search), just to get some hw properties.
Or go back to the earlier one, with a vid array containing unused slots?
That makes the for loops a bit harder.
I need to think about it...
Hmm, I don't think so, it seems to look fine to me and work fine too. I
have tested thoroughly for AM62L (which has uninstantiated vid region)
along with AM62x and AM62A with all planes displayed simultaneously. If
you want I can put on some test logs, create some dummy holes for VID
regions in AM62 and AM62A to put this on to some further negative tests.
>
Also if naming convention is confusing (hw_id vs hw_plane) then maybe we
can use something else like vid_idx ??
It is confusing. But I think it's also broken, in the sense that e.g.
dispc_k3_vid_write_irqstatus() has hw_plane parameter. But it's actually
hw_id.
I'm not sure if naming them differently helps here. It's super
confusing. What indices do we have?
- The lowest level HW IDs, e.g. for DISPC_VID_IRQSTATUS()
- The index for the dispc->vid_info[]
- The index to tidss->planes[]
- drm_plane->index
Originally I kept the drm_plane and the HW index separate, so that the
dispc.c doesn't really deal with the drm_plane at all. But I wonder if
we need to change that, as drm_plane pointer can't really be
"understood" wrong, whereas an two indices are easy to mix.
Tomi
