omapdrm VRFB rotation
Hi,
Currently omapdrm supports TILER rotation only, which excludes omap3 and earlier SoCs. I have the hardware (N900/N950), time and will to implement VRFB rotation support in omapdrm driver, however, I'll appreciate some hints. Or, if there is already something ready, please point me to it so I can take it from where it is.
Besides partially reverting 517a8a9564c0dea98e6d4e2c7f0fe4cbb9b8c9ba and copying VRFB code from omapfb, is there anything else I shall take in consideration? Or, VRFB driver should not be a part of omapdrm, but a standalone one?
Regards,
Ivo
Sorry, mail was sent to the old Tomi's address.
Quoted 8 lines · Ivaylo Dimitrov
Hi,
Currently omapdrm supports TILER rotation only, which excludes omap3 and earlier SoCs. I have the hardware (N900/N950), time and will to implement VRFB rotation support in omapdrm driver, however, I'll appreciate some hints. Or, if there is already something ready, please point me to it so I can take it from where it is.
Besides partially reverting
517a8a9564c0dea98e6d4e2c7f0fe4cbb9b8c9baand copying VRFB code from omapfb, is there anything else I shall take in consideration? Or, VRFB driver should not be a part of omapdrm, but a standalone one?Regards,
Ivo
Hi,
Quoted 8 lines · Ivaylo Dimitrov
Sorry, mail was sent to the old Tomi's address.
Quoted 5 lines · Ivaylo Dimitrov
Hi,
Currently omapdrm supports TILER rotation only, which excludes omap3 and earlier SoCs. I have the hardware (N900/N950), time and will to implement VRFB rotation support in omapdrm driver, however, I'll appreciate some hints. Or, if there is already something ready, please point me to it so I can take it from where it is.
Besides partially reverting
517a8a9564c0dea98e6d4e2c7f0fe4cbb9b8c9baand copying VRFB code from omapfb, is there anything else I shall take in consideration? Or, VRFB driver should not be a part of omapdrm, but a standalone one?
We already have DMM driver in the omapdrm module, and I think VRFB fits along just fine. I don't think there ever has been any other users for VRFB than DSS, and as it's such an old IP, I don't think there ever will.
I don't have any particular hints in mind.
Do you have omap4/5 so you can test that DMM still works after your changes?
Tomi
Hi,
Quoted 5 lines · Tomi Valkeinen
Hi,
Quoted 2 lines · Ivaylo Dimitrov
Sorry, mail was sent to the old Tomi's address.
Quoted 9 lines
Quoted 6 lines · Ivaylo Dimitrov
Quoted 5 lines
Hi,
Currently omapdrm supports TILER rotation only, which excludes omap3 and earlier SoCs. I have the hardware (N900/N950), time and will to implement VRFB rotation support in omapdrm driver, however, I'll appreciate some hints. Or, if there is already something ready, please point me to it so I can take it from where it is.
Besides partially reverting
517a8a9564c0dea98e6d4e2c7f0fe4cbb9b8c9baand copying VRFB code from omapfb, is there anything else I shall take in consideration? Or, VRFB driver should not be a part of omapdrm, but a standalone one?We already have DMM driver in the omapdrm module, and I think VRFB fits along just fine. I don't think there ever has been any other users for VRFB than DSS, and as it's such an old IP, I don't think there ever will.
Ok, I guess if a need appears, it can always be moved out of omapdrm. The same applies to DMM/TILER code I guess.
Quoted 4 lines
I don't have any particular hints in mind.
Do you have omap4/5 so you can test that DMM still works after your changes?
Yes, I have motorola droid4 (4430/sgx540) to test with. Which brings another issue on the table - I was not able to find a way to allocate a TILER dma_buf. The only way seems to be by using omap_bo_xxx functions, which is a 'vendor' API. Do I miss something or omapdrm is lacks functionality? Also, is there any particular reason why TILER is not enabled by default for dma_buf BOs? Is it a limited resource (like, there is a finite number of BOs that can use TILER) or there is some other reason. Also, is it possible to 'migrate' non-TILER BO to a TILER one? The same issue will arise with VRFB (with its 12 contexts on omap3) as well.
So, in short:
- omap3: VRFB driver to be added to omapdrm.
- omap3/omap4/omap5: dma_buf lacks TILER/VRFB support
Adding VRFB should be trivial, but I am more concerned about how to rotate dma_buf buffers. I imagine something like - on setting the plane "rotate" property, migrate physical address from 'normal' CMA memory to a TILER/VRFB one. Is that possible? What if we already have that bo mmap-ed? What if SGX MMU is already set up to access the memory? Or, is it possible to use TILER for read access, the same way VRFB can be set-up, so instead of writing through TILER/VRFB, DSS to be set-up to read FB memory through TILER/VRFB thus avoiding the need to migrate SG and/or page tables?
Thanks and regards,
Ivo
Quoted 11 lines · Ivaylo Dimitrov
Hi,
Quoted 8 lines · Tomi Valkeinen
Hi,
Quoted 5 lines · Ivaylo Dimitrov
Sorry, mail was sent to the old Tomi's address.
Quoted 2 lines · Ivaylo Dimitrov
Hi,
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Currently omapdrm supports TILER rotation only, which excludes omap3 and earlier SoCs. I have the hardware (N900/N950), time and will to implement VRFB rotation support in omapdrm driver, however, I'll appreciate some hints. Or, if there is already something ready, please point me to it so I can take it from where it is.
Besides partially reverting
517a8a9564c0dea98e6d4e2c7f0fe4cbb9b8c9baand copying VRFB code from omapfb, is there anything else I shall take in consideration? Or, VRFB driver should not be a part of omapdrm, but a standalone one?We already have DMM driver in the omapdrm module, and I think VRFB fits along just fine. I don't think there ever has been any other users for VRFB than DSS, and as it's such an old IP, I don't think there ever will.
Ok, I guess if a need appears, it can always be moved out of omapdrm. The same applies to DMM/TILER code I guess.
Yes.
Quoted 4 lines
Quoted 4 lines
I don't have any particular hints in mind.
Do you have omap4/5 so you can test that DMM still works after your changes?
So, first, a few clarifications about DMM/TILER. There are two somewhat separate things. I'm not sure what are the exactly correct words for these, but these are the terms I've used:
- DMM (sometimes called wrongly TILER 1D), which acts as a basic iommu. It doesn't do rotation, just mapping scattered pages to a contiguous memory view.
- TILER (or TILER 2D) which on top of the above, adds support for rotation.
Quoted 1 line
Yes, I have motorola droid4 (4430/sgx540) to test with. Which brings another issue on the table - I was not able to find a way to allocate a TILER
dma_buf. The only way seems to be by usingomap_bo_xxxfunctions, which is a 'vendor' API. Do I miss something or omapdrm is lacks
Yes, omap_bo_xxx is omap specific API. But I don't understand your question about omapdrm. omap_bo functions use omapdrm's API. The userspace can skip libdrm and just call the ioctls directly if they so want.
Quoted 1 line
functionality? Also, is there any particular reason why TILER is not enabled by default for
dma_bufBOs? Is it a limited resource (like, there is a finite number of BOs that can use TILER) or there is some
Yes, TILER is a limited resource, there's a maximum size for the currently mapped memory. Going through TILER is also slower, and while for 0 degree rotation the diff probably isn't much, it's still there.
Also TILER is not really quite supported, as I never got it working with all the YUV modes, and also because there's a HW issue, which may cause the device to lock up in some rare cases. And last, TILER creates such a memory layout that it's not possible to share it safely with dmabuf, as dmabuf lacks the features to properly describe the layout.
All that said, I think TILER should mostly work in upstream.
However, DMM is supported, and used by default. If you allocate a dumb buffer, it uses DMM by default.
Quoted 1 line
other reason. Also, is it possible to 'migrate' non-TILER BO to a TILER one? The same issue will arise with VRFB (with its 12 contexts on omap3) as well.
No, not possible to migrate.
Quoted 3 lines
So, in short:
- omap3: VRFB driver to be added to omapdrm.
- omap3/omap4/omap5:dma_buflacks TILER/VRFB support
Can you elaborate on what's missing?
Quoted 1 line
Adding VRFB should be trivial, but I am more concerned about how to rotate
dma_bufbuffers. I imagine something like - on setting the plane "rotate" property, migrate physical address from 'normal' CMA memory to a TILER/VRFB one. Is that possible? What if we already have that bo mmap-ed? What if SGX MMU is already set up to access the memory? Or, is it possible to use TILER for read access, the same way VRFB can be set-up, so instead of writing through TILER/VRFB, DSS to be set-up to read FB memory through TILER/VRFB thus avoiding the need to migrate SG and/or page tables?
I have to say I don't quite remember how VRFB works. But for TILER, we have to specifically allocate a TILER buffer. And then everyone else uses TILER 0 degree view, but DSS can use 0/90/180/270 view when showing it on the display.
But I think the case is the same for VRFB too: if you want to use VRFB, all access has to go through VRFB, as the pixels are written to memory in a special way.
Tomi
Quoted 28 lines · Tomi Valkeinen
Quoted 11 lines · Ivaylo Dimitrov
Hi,
Quoted 8 lines · Tomi Valkeinen
Hi,
Quoted 5 lines · Ivaylo Dimitrov
Sorry, mail was sent to the old Tomi's address.
Quoted 2 lines · Ivaylo Dimitrov
Hi,
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Currently omapdrm supports TILER rotation only, which excludes omap3 and earlier SoCs. I have the hardware (N900/N950), time and will to implement VRFB rotation support in omapdrm driver, however, I'll appreciate some hints. Or, if there is already something ready, please point me to it so I can take it from where it is.
Besides partially reverting
517a8a9564c0dea98e6d4e2c7f0fe4cbb9b8c9baand copying VRFB code from omapfb, is there anything else I shall take in consideration? Or, VRFB driver should not be a part of omapdrm, but a standalone one?We already have DMM driver in the omapdrm module, and I think VRFB fits along just fine. I don't think there ever has been any other users for VRFB than DSS, and as it's such an old IP, I don't think there ever will.
Ok, I guess if a need appears, it can always be moved out of omapdrm. The same applies to DMM/TILER code I guess.
Yes.
Quoted 4 lines
Quoted 4 lines
I don't have any particular hints in mind.
Do you have omap4/5 so you can test that DMM still works after your changes?
Quoted 7 lines
So, first, a few clarifications about DMM/TILER. There are two somewhat separate things. I'm not sure what are the exactly correct words for these, but these are the terms I've used:
- DMM (sometimes called wrongly TILER 1D), which acts as a basic iommu. It doesn't do rotation, just mapping scattered pages to a contiguous memory view.
- TILER (or TILER 2D) which on top of the above, adds support for rotation.
Thanks for the clarification, I'll try to stick to those.
Quoted 4 lines
Quoted 1 line
Yes, I have motorola droid4 (4430/sgx540) to test with. Which brings another issue on the table - I was not able to find a way to allocate a TILER
dma_buf. The only way seems to be by usingomap_bo_xxxfunctions, which is a 'vendor' API. Do I miss something or omapdrm is lacksYes,
omap_bo_xxxis omap specific API. But I don't understand your question about omapdrm.omap_bofunctions use omapdrm's API. The userspace can skip libdrm and just call the ioctls directly if they so want.
omap_bo functions/ioctls do, but if we try to use GBM dma_buf API (gbm_bo_create() and friends), we cannot allocate anything but CMA memory which omapdrm refuses to rotate, IIUC. Well at least I was not able to make framebuffer rotate with GBM allocated BO.
The terms I use:dma_buf - BO that is allocated using GBM API and seems to end up here: https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/omapdrm/omap_gem.c#L1300. Seems it cannot go through the TILER.
omap_bo: BO that is allocated using omap_bo functions. IIUC this is basically a dumb buffer, but it can be allocated to be accesses through TILER by passing the appropriate flags to omap_bo_new().
Quoted 4 lines
Quoted 1 line
functionality? Also, is there any particular reason why TILER is not enabled by default for
dma_bufBOs? Is it a limited resource (like, there is a finite number of BOs that can use TILER) or there is someYes, TILER is a limited resource, there's a maximum size for the currently mapped memory. Going through TILER is also slower, and while for 0 degree rotation the diff probably isn't much, it's still there.
Yeah, so we should go through TILER only when needed. The same as VRFB.
Quoted 4 lines
Also TILER is not really quite supported, as I never got it working with all the YUV modes, and also because there's a HW issue, which may cause the device to lock up in some rare cases. And last, TILER creates such a memory layout that it's not possible to share it safely with dmabuf, as dmabuf lacks the features to properly describe the layout.
All that said, I think TILER should mostly work in upstream.
But, only from omap_bo BOs
Quoted 2 lines
However, DMM is supported, and used by default. If you allocate a dumb buffer, it uses DMM by default.
not for dma_buf BOs though. Not really an issue as on omap3 there is no DMM so we must use CMA either ways.
Quoted 10 lines
Quoted 1 line
other reason. Also, is it possible to 'migrate' non-TILER BO to a TILER one? The same issue will arise with VRFB (with its 12 contexts on omap3) as well.
No, not possible to migrate.
Quoted 3 lines
So, in short:
- omap3: VRFB driver to be added to omapdrm.
- omap3/omap4/omap5:dma_buflacks TILER/VRFB supportCan you elaborate on what's missing?
GBM BOs (which I call dma_buf BOs) lack TILER support
Quoted 3 lines
Quoted 1 line
Adding VRFB should be trivial, but I am more concerned about how to rotate
dma_bufbuffers. I imagine something like - on setting the plane "rotate" property, migrate physical address from 'normal' CMA memory to a TILER/VRFB one. Is that possible? What if we already have that bo mmap-ed? What if SGX MMU is already set up to access the memory? Or, is it possible to use TILER for read access, the same way VRFB can be set-up, so instead of writing through TILER/VRFB, DSS to be set-up to read FB memory through TILER/VRFB thus avoiding the need to migrate SG and/or page tables?I have to say I don't quite remember how VRFB works.
it has 12 'contexts' that can represent up to that number of 'images' in memory with rotated views
Quoted 2 lines
But for TILER, we have to specifically allocate a TILER buffer. And then everyone else uses TILER 0 degree view, but DSS can use 0/90/180/270 view when showing it on the display.
oh, so in theory omapdrm should be able to show dma_buf BOs rotated? We only lack a way to allocate TILER GBM BO, right?
I ported xf86-video-omap to use GBM API instead of omap_bo API and it works fine in native orientation. However, when it tries to set 'rotated' property to framebuffer, it receives an error. Should I trace to see why it fails in the kernel or this is not supported?
Quoted 2 lines
But I think the case is the same for VRFB too: if you want to use VRFB, all access has to go through VRFB, as the pixels are written to memory in a special way.
Yes, I think you are right.
Sorry for asking so much questions but I want to have a clear picture on what is supposed to work and what is missing.
Thanks,
Ivo
Quoted 40 lines · Ivaylo Dimitrov
Quoted 28 lines · Tomi Valkeinen
Quoted 11 lines · Ivaylo Dimitrov
Hi,
Quoted 8 lines · Tomi Valkeinen
Hi,
Quoted 5 lines · Ivaylo Dimitrov
Sorry, mail was sent to the old Tomi's address.
Quoted 2 lines · Ivaylo Dimitrov
Hi,
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Currently omapdrm supports TILER rotation only, which excludes omap3 and earlier SoCs. I have the hardware (N900/N950), time and will to implement VRFB rotation support in omapdrm driver, however, I'll appreciate some hints. Or, if there is already something ready, please point me to it so I can take it from where it is.
Besides partially reverting
517a8a9564c0dea98e6d4e2c7f0fe4cbb9b8c9baand copying VRFB code from omapfb, is there anything else I shall take in consideration? Or, VRFB driver should not be a part of omapdrm, but a standalone one?We already have DMM driver in the omapdrm module, and I think VRFB fits along just fine. I don't think there ever has been any other users for VRFB than DSS, and as it's such an old IP, I don't think there ever will.
Ok, I guess if a need appears, it can always be moved out of omapdrm. The same applies to DMM/TILER code I guess.
Yes.
Quoted 4 lines
Quoted 4 lines
I don't have any particular hints in mind.
Do you have omap4/5 so you can test that DMM still works after your changes?
Quoted 8 lines
So, first, a few clarifications about DMM/TILER. There are two somewhat separate things. I'm not sure what are the exactly correct words for these, but these are the terms I've used:
- DMM (sometimes called wrongly TILER 1D), which acts as a basic
iommu. It doesn't do rotation, just mapping scattered pages to a contiguous memory view.- TILER (or TILER 2D) which on top of the above, adds support for rotation.
Quoted 8 lines
Thanks for the clarification, I'll try to stick to those.
Quoted 4 lines
Quoted 1 line
Yes, I have motorola droid4 (4430/sgx540) to test with. Which brings another issue on the table - I was not able to find a way to allocate a TILER
dma_buf. The only way seems to be by usingomap_bo_xxxfunctions, which is a 'vendor' API. Do I miss something or omapdrm is lacksYes,
omap_bo_xxxis omap specific API. But I don't understand your question about omapdrm.omap_bofunctions use omapdrm's API. The userspace can skip libdrm and just call the ioctls directly if they so want.
omap_bofunctions/ioctls do, but if we try to use GBMdma_bufAPI (gbm_bo_create()and friends), we cannot allocate anything but CMA memory which omapdrm refuses to rotate, IIUC. Well at least I was not able to make framebuffer rotate with GBM allocated BO.
Yes, I think that's true: libgbm doesn't support omap's TILER.
Quoted 2 lines
The terms I use:
dma_buf- BO that is allocated using GBM API and seems to end up here: https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/omapdrm/omap_gem.c#L1300. Seems it cannot go through the TILER.
Isn't that a gbm_bo? Or something =). dma_buf means a specific thing, which is not what you mean above.
Quoted 1 line
omap_bo: BO that is allocated usingomap_bofunctions. IIUC this is basically a dumb buffer, but it can be allocated to be accesses through TILER by passing the appropriate flags toomap_bo_new().
No, I don't think you can allocate a TILER buffer with omap_bo_new(), you need to use omap_bo_new_tiled().
Quoted 5 lines
Quoted 4 lines
Quoted 1 line
functionality? Also, is there any particular reason why TILER is not enabled by default for
dma_bufBOs? Is it a limited resource (like, there is a finite number of BOs that can use TILER) or there is someYes, TILER is a limited resource, there's a maximum size for the currently mapped memory. Going through TILER is also slower, and while for 0 degree rotation the diff probably isn't much, it's still there.
Quoted 8 lines
Yeah, so we should go through TILER only when needed. The same as VRFB.
Quoted 4 lines
Also TILER is not really quite supported, as I never got it working with all the YUV modes, and also because there's a HW issue, which may cause the device to lock up in some rare cases. And last, TILER creates such a memory layout that it's not possible to share it safely with dmabuf, as dmabuf lacks the features to properly describe the layout.
All that said, I think TILER should mostly work in upstream.
But, only from
omap_boBOs
Yes.
Quoted 4 lines
Quoted 2 lines
However, DMM is supported, and used by default. If you allocate a dumb buffer, it uses DMM by default.
not for
dma_bufBOs though. Not really an issue as on omap3 there is no DMM so we must use CMA either ways.
I think all allocations use DMM by default, if available. So gbm_bo should use DMM underneath on omap4/5.
Quoted 15 lines
Quoted 10 lines
Quoted 1 line
other reason. Also, is it possible to 'migrate' non-TILER BO to a TILER one? The same issue will arise with VRFB (with its 12 contexts on omap3) as well.
No, not possible to migrate.
Quoted 3 lines
So, in short:
- omap3: VRFB driver to be added to omapdrm.
- omap3/omap4/omap5:dma_buflacks TILER/VRFB supportCan you elaborate on what's missing?
GBM BOs (which I call
dma_bufBOs) lack TILER supportQuoted 2 lines
Quoted 1 line
Adding VRFB should be trivial, but I am more concerned about how to rotate
dma_bufbuffers. I imagine something like - on setting the plane "rotate" property, migrate physical address from 'normal' CMA memory to a TILER/VRFB one. Is that possible? What if we already have that bo mmap-ed? What if SGX MMU is already set up to access the memory? Or, is it possible to use TILER for read access, the same way VRFB can be set-up, so instead of writing through TILER/VRFB, DSS to be set-up to read FB memory through TILER/VRFB thus avoiding the need to migrate SG and/or page tables?
Quoted 8 lines
Quoted 1 line
I have to say I don't quite remember how VRFB works.
it has 12 'contexts' that can represent up to that number of 'images' in memory with rotated views
Quoted 2 lines
But for TILER, we have to specifically allocate a TILER buffer. And then everyone else uses TILER 0 degree view, but DSS can use 0/90/180/270 view when showing it on the display.
oh, so in theory omapdrm should be able to show
dma_bufBOs rotated? We only lack a way to allocate TILER GBM BO, right?
Yes.
Quoted 1 line
I ported xf86-video-omap to use GBM API instead of
omap_boAPI and it works fine in native orientation. However, when it tries to set 'rotated' property to framebuffer, it receives an error. Should I trace to see why it fails in the kernel or this is not supported?
Well, I have no idea what's going on there, but I just tested with the latest drm-misc-next, and at least a basic test with TILER works. I used kms++ and rottest.py.
Maybe you used omap_bo_new(), which does not give you a TILER buffer?
Quoted 6 lines
Quoted 2 lines
But I think the case is the same for VRFB too: if you want to use VRFB, all access has to go through VRFB, as the pixels are written to memory in a special way.
Yes, I think you are right.
Sorry for asking so much questions but I want to have a clear picture on what is supposed to work and what is missing.
No problem. I think it's very nice to get VRFB moved to omapdrm, as I believe that's the only feature missing from omapdrm which is supported in omapfb.
Tomi