Re: [PATCH 3/3] drm: omapdrm: Do no allocate non-scanout GEMs through DMM/TILER

From: Andrew Davis
Date: Mon Aug 29 2022 - 10:36:29 EST


On 8/29/22 8:24 AM, Ivaylo Dimitrov wrote:
Hi,


On 29.08.22 г. 5:51 ч., Yongqin Liu wrote:
Hi, Ivaylo

Sorry for the late response, and Thanks very much for the detailed explanations!

On Thu, 18 Aug 2022 at 18:23, Ivaylo Dimitrov
<ivo.g.dimitrov.75@xxxxxxxxx> wrote:

Hi,

On 17.08.22 г. 7:52 ч., Yongqin Liu wrote:
Hi, Ivaylo

On Mon, 15 Aug 2022 at 14:23, Ivaylo Dimitrov
<ivo.g.dimitrov.75@xxxxxxxxx> wrote:

Hi Liu,

On 14.08.22 г. 17:27 ч., Yongqin Liu wrote:
Hi, IvayIo

Thanks very much for the reply!

On Sat, 13 Aug 2022 at 14:58, Ivaylo Dimitrov
<ivo.g.dimitrov.75@xxxxxxxxx> wrote:

Hi Liu,

On 12.08.22 г. 7:35 ч., Yongqin Liu wrote:
Hi, Ivaylo, Tomi

We have one X15 Android AOSP master build, it could not have the home
screen displayed
on the hdmi monitor connected with this change, with the following
message printed on the serial console
        [  607.404205] omapdrm omapdrm.0: Failed to setup plane plane-0
        [  607.410522] omapdrm omapdrm.0: Failed to setup plane plane-1
        [  607.416381] omapdrm omapdrm.0: Failed to setup plane plane-2
        [  607.422088] omapdrm omapdrm.0: Failed to setup plane plane-3

       # for details, please check the link here: http://ix.io/47m1

It will work with home screen displayed on the hdmi monitor if this
change is reverted.

Is this the broken problem you talked about here?

And could you please give some suggestions on how to have the x15
Android build work with this change?


Make sure scanout (i.e. those to be displayed) buffers are actually
allocated as such - OMAP_BO_SCANOUT flag must be set when calling
omap_bo_new().

I am not familiar with this area, I am sorry if I asked quite silly questions:(
I googled omap_bo_new, and found it's a function of libdrm here[1], is
it what you meant here?


Yes, calling this function from userspace ends in kernel code the
$subject patch is part of.

If it's the omap_bo_new that we should pass OMAP_BO_SCANOUT when it is called,
then is it the correct way to update omap_bo_new to add the OMAP_BO_SCANOUT flag
before it calls omap_bo_new_impl?


omap_bo_new() is fine and does not need any updates/fixes, it is the
code that uses it (whoever it is, I am not familiar with the userspace
you are using) that shall pass correct flags (third parameter) when
calling it.

Sorry, I do not get the point here.
Like you said, the code that calls omap_bo_new needs to pass OMAP_BO_SCANOUT,
then IMO omap_bo_new should be the best place to add the OMAP_BO_SCANOUT flag,
(like via flags = flags | OMAP_BO_SCANOUT), that could help avoid
missing the flag by some code,
and also avoids hacks/changes on the possible blob binaries.

Do I misunderstand somewhere?
Or is there some case that OMAP_BO_SCANOUT shouldn't be passed when
omap_bo_new is called?


Exactly. You need to pass OMAP_BO_SCANOUT only when you want your
buffers to be 'scanout' buffers(i.e. buffers that can be displayed on
screen), which is not always the case - there is no need offscreen
buffers or pixmaps to be scanout capable, for example. There are more
cases like that.

The problem is that scanout buffer on OMAP4 allocate additional
resources in DMM/TILER (a piece of hardware) and those resources are
limited. Not only that, but DMM/TILER memory space eventually gets
fragmented over time (if you have lots of allocataoins/deallocations)
and you will start getting ENOMEM (or similar) errors.

Ofc, in your particular use case you may never hit such issues.

Thanks, I understand the cases now.


BTW you shall really find who and how uses OMAP BO API, in theory it
might use ioctls directly and not call omap_bo_xxx functions.

Do you mean the DRM_OMAP_GEM_NEW ioctl api?
There is no place in the AOSP tree to call that except the
omap_bo_new_impl function,
which is called by the omap_bo_new and omap_bo_new_tiled functions.
The omap_bo_new should not be called with the OMAP_BO_TILED flag,
while the omap_bo_new_tiled should be called with the OMAP_BO_TILED flag

Regarding to the omap_bo_new function, there are 2 places call it in
the AOSP tree:
#1 ./external/libkmsxx/kms++/src/omap/omapframebuffer.cpp
#2 ./device/ti/beagle_x15/gpu/gralloc.am57x.so

#1 seems not used in AOSP yet, and #2 is one blob binary we do not
have the source for.


I would bet on gralloc.am57x.so.
yeah, that's my guess as well.

strace
would be your friend there. or gdb, or whatever tools are used on
android. Or put some printfs() in omap_bo_new() that output the PID of
the calling process, etc.

Thanks a lot for these great suggestions! Will use them when possible.

And another question is that, since the userspace(libdrm) will be used
to work with different kernel versions,
like the old 4.14, 4.19, etc, do you think there will be problem to
pass  OMAP_BO_SCANOUT
from the userspace side with the old kernels(which does not have this change)?
does this change need to be backported to the old kernel versions?

There should not be any issue. The changes could be backported if one
hits the issues this $series is fixing, but there is no need.

Thanks for the confirmation!
I just boot-tested with adding OMAP_BO_SCANOUT in the omap_bo_new function,
and it worked with the current 4.14, 4.19, and the mainline kernels.
# via adding line "flags = flags | OMAP_BO_SCANOUT" in the omap_bo_new function.


sure, the point is that with this change *every* BO will be allocated as
scanout BO, potentially leading to the above explained issues.

get it.


And the last question is that, omap_bo_new might be called by some
property binaries what not everyone
could get the source to update, for such case what's your suggestions?


Hard to say without knowing what that library would be.

When I hit issues with closed blobs, sometimes I reverse-engineer them
to fix the issue, example:

https://github.com/maemo-leste/sgx-ddk-um/tree/master/dbm

This is REed libdbm from sgx-ddk-um 1.17.4948957, that is responsible
for allocating BOs (what omap_bo_new() does) but it uses DUMB buffers
API, instead of OMAP BO API.

I guess you are using some older version of sgx-ddk-um, so you may fix
in similar way. Or binary patch.

The blob binary that calls omap_bo_new is the gralloc.am57x.so here[2]:
any suggestions with it?
# sorry, I am not able to find out how you did the reverse-engineer
work# with the dbm repository shared here,
# not sure if you could give some tutorial steps for the similar
reverse-engineer# work with gralloc.am57x.so


Sorry, but it is like if you ask me to provide you with a tutorial on
how to do brain surgery :)

[2]: https://android.googlesource.com/device/ti/beagle-x15/+/refs/heads/master/gpu/gralloc.am57x.so


I investigated this a bit and it seems it calls omap_bo_new() in a
wrapper function like:

bo = omap_bo_new(dev, -page_size & (size + page_size - 1), ((param5 &
0x800000) != 0) | OMAP_BO_WC | OMAP_BO_MEM_CONTIG);

Didn't investigate further what param5 is, but it controls if
OMAP_BO_SCANOUT is passed to omap_bo_new or not.

However, this library was not made with upstream kernel in mind, as
AFAIK OMAP_BO_MEM_CONTIG never made it upstream:

https://yhbt.net/lore/all/2580272.MiZDHyRxZo@avalon/T/

@Tomi - any comment?

So, you have couple of options:

1. Ask TI for upstream-compatible library.
check is in progress, but it would take quite a long time I guess
2. Try to push OMAP_BO_MEM_CONTIG patch upstream.
hmm, sounds like one impossible thing...
3. Modify omap_bo_new() to something like:
.
#define OMAP_BO_MEM_CONTIG      0x00000008      /* only use contiguous dma mem */
.
if (flags & OMAP_BO_MEM_CONTIG)
    flags |= OMAP_BO_SCANOUT;
.
This will not achieve exactly what OMAP_BO_MEM_CONTIG is supposed to do,
but should make it work, at least.

This looks like the only doable thing at the moment, maybe one change
needs to be submitted to the mesa/drm repository.
I can submit a request on your #3 change to the mesa/drm repository
for discussion after some check if you do not mind.


I doubt mesa/drm will accept such hack, I think you will need to support your drm clone (with the above fix) until TI fixes the closed library.



Hi all,

Just got around to reading this thread. I work with the TI gralloc lib
and can generate new versions as needed (I was probably the one who compiled
the version you have now). I've wanted to have our gralloc layer open source'd
as there is nothing really propriety in it (and I re-wrote a lot of it already)
and to avoid issues like this. But it interacts with the GPU code in some places,
so it's up to Imagination :(. The actual code in question if it helps is:

if(ui32Flags & PVRSRV_MEM_CACHED)
flags &= ~OMAP_BO_CACHE_MASK;
else
flags |= OMAP_BO_WC;

if (ui32Flags & PVRSRV_HAP_CONTIG)
flags |= OMAP_BO_SCANOUT;

flags &= ~OMAP_BO_TILED_MASK;
flags |= 0x00000008;
flags |= OMAP_BO_WC;

bo = omap_bo_new(dev, size, flags);

As you can see we use 0x00000008 (OMAP_BO_MEM_CONTIG) unconditionally.
This was a hack added since even non-scanout buffers sometimes need
to be contiguous (video decoder surfaces), but we had no way back
then to communicate this to the gralloc layer. I think your best
bet would be to modify the gralloc lib to not do that, or put it
under the CONTIG check.

If you tell me what the code should look like, I can rebuild the
lib and post a copy.

Long term, I'd like to start using DMA-BUF Heaps for CMA memory
allocations in gralloc and elsewhere, then drop out the DMM/TILER
support from OMAPDRM, since it never really belonged there in
the first place (being a IOMMU unrelated to the display/GPU).

Thanks,
Andrew


Regards,
Ivo

Thanks,
Yongqin Liu

On Thu, 17 Feb 2022 at 23:29, Ivaylo Dimitrov
<ivo.g.dimitrov.75@xxxxxxxxx> wrote:



On 17.02.22 г. 14:46 ч., Tomi Valkeinen wrote:
Hi,

On 19/01/2022 12:23, Ivaylo Dimitrov wrote:
On devices with DMM, all allocations are done through either DMM or
TILER.
DMM/TILER being a limited resource means that such allocations will start
to fail before actual free memory is exhausted. What is even worse is
that
with time DMM/TILER space gets fragmented to the point that even if we
have
enough free DMM/TILER space and free memory, allocation fails because
there
is no big enough free block in DMM/TILER space.

Such failures can be easily observed with OMAP xorg DDX, for example -
starting few GUI applications (so buffers for their windows are
allocated)
and then rotating landscape<->portrait while closing and opening new
windows soon results in allocation failures.

Fix that by mapping buffers through DMM/TILER only when really needed,
like, for scanout buffers.

Doesn't this break users that get a buffer from omapdrm and expect it to
be contiguous?


If you mean dumb buffer, then no, this does not break users as dumb
buffers are allocated as scanout:

https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/omapdrm/omap_gem.c#L603

If you mean omap_bo allocated buffers, then if users want
linear(scanout) buffer, then they request it explicitly by passing
OMAP_BO_SCANOUT.

Ivo