mbox series

[0/9] drm: Support GEM SHMEM fbdev without shadow FB

Message ID 20220303205839.28484-1-tzimmermann@suse.de
Headers show
Series drm: Support GEM SHMEM fbdev without shadow FB | expand

Message

Thomas Zimmermann March 3, 2022, 8:58 p.m. UTC
Rework the fbdev deferred-I/O to not interfere with fields of struct
page. Make the code more flexible and implement GEM SHMEM mmap on top
of it.

This patchset removes the need for a special shadow framebuffer for
fbdev mmap when using GEM SHMEM. SHMEM pages are now mmap'ed from
/dev/fb directly.

Patches 2 and 3 rework the fbdev deferred I/O code. It now allows
drivers to have better control of the mmap operations. All references
to fields in struct page are gone. The rsp state is help in a 
separate pageref structure.

Patches 4 to 7 provide callbacks an helpers to implement deferred I/O
with DRM drivers. Specifically, patch 6 introduces a callback to create
a dumb buffer for fbdev. This will be useful for many drivers that
currently cannot use generic fbdev emulation because of special placement
needs of the BO, such as amdgpu or radeon. The drivers can handle the
differences to regular dumb buffers in their new callback implementation.

Patch 8 extends the GEM SHMEM memory manager with a new helper for fbdev
dumb-buffer creation. The returned BO has it's mmap set up to implement
deferred I/O with SHMEM pages. No additional shadow buffer is requires
any longer. Many drivers can immediatelly benefit from this change.

Patch 9 extends virtgpu to support fbdev dumb buffers. It's build on
top of GEM SHMEM, but has some modifications that need to be implemented
for fbdev as well.

There's no immediate fbdev performance improvement from this patchset.
Most of all, it removes unnecessary shadow framebuffers and rsp memcpys.
A shadow fb for a FullHD display is ~8 MiB, which we now save. The patches
do reduce latency between drawing to the fbdev buffer to displaying
on the screen. Watching a video on the fbdev console felt smoother and
had less flickering.

Thomas Zimmermann (9):
  drm/simpledrm: Use fbdev defaults for shadow buffering
  fbdev: Put mmap for deferred I/O into drivers
  fbdev: Track deferred-I/O pages in pageref struct
  fbdev: Export helper for implementing page_mkwrite
  drm/fb-helper: Separate deferred I/O from shadow buffers
  drm/fb-helper: Provide callback to create fbdev dumb buffers
  drm/fb-helper: Provide fbdev deferred-I/O helpers
  drm/gem-shmem: Implement fbdev dumb buffer and mmap helpers
  drm/virtio: Implement dumb_create_fbdev with GEM SHMEM helpers

 drivers/gpu/drm/drm_client.c            |  14 +-
 drivers/gpu/drm/drm_crtc_internal.h     |   3 +
 drivers/gpu/drm/drm_dumb_buffers.c      |  36 +++-
 drivers/gpu/drm/drm_fb_helper.c         |  79 +++++++--
 drivers/gpu/drm/drm_gem_shmem_helper.c  | 204 ++++++++++++++++++++---
 drivers/gpu/drm/gud/gud_drv.c           |   2 +-
 drivers/gpu/drm/hyperv/hyperv_drm_drv.c |   2 +-
 drivers/gpu/drm/mgag200/mgag200_drv.c   |   2 +-
 drivers/gpu/drm/tiny/cirrus.c           |   2 +-
 drivers/gpu/drm/tiny/gm12u320.c         |   2 +-
 drivers/gpu/drm/tiny/simpledrm.c        |   3 +-
 drivers/gpu/drm/udl/udl_drv.c           |   2 +-
 drivers/gpu/drm/virtio/virtgpu_drv.c    |   2 +
 drivers/gpu/drm/virtio/virtgpu_drv.h    |   2 +
 drivers/gpu/drm/virtio/virtgpu_object.c |  54 +++++-
 drivers/gpu/drm/vkms/vkms_drv.c         |   2 +-
 drivers/gpu/drm/vmwgfx/vmwgfx_fb.c      |   6 +-
 drivers/hid/hid-picolcd_fb.c            |   1 +
 drivers/staging/fbtft/fbtft-core.c      |   6 +-
 drivers/video/fbdev/broadsheetfb.c      |   6 +-
 drivers/video/fbdev/core/fb_defio.c     | 211 +++++++++++++++++-------
 drivers/video/fbdev/core/fbmem.c        |  19 +--
 drivers/video/fbdev/hecubafb.c          |   1 +
 drivers/video/fbdev/hyperv_fb.c         |   6 +-
 drivers/video/fbdev/metronomefb.c       |   6 +-
 drivers/video/fbdev/sh_mobile_lcdcfb.c  |  12 +-
 drivers/video/fbdev/smscufx.c           |   8 +-
 drivers/video/fbdev/ssd1307fb.c         |   1 +
 drivers/video/fbdev/udlfb.c             |   8 +-
 drivers/video/fbdev/xen-fbfront.c       |   6 +-
 include/drm/drm_client.h                |   3 +-
 include/drm/drm_drv.h                   |  27 +++
 include/drm/drm_fb_helper.h             |   9 +
 include/drm/drm_gem_shmem_helper.h      |  63 ++++++-
 include/linux/fb.h                      |  12 +-
 35 files changed, 668 insertions(+), 154 deletions(-)


base-commit: 710a019ad85e96e66f7d75ee7f4733cdd8a2b0d0
prerequisite-patch-id: c2b2f08f0eccc9f5df0c0da49fa1d36267deb11d
prerequisite-patch-id: c67e5d886a47b7d0266d81100837557fda34cb24
prerequisite-patch-id: 6e1032c6302461624f33194c8b8f37103a3fa6ef

Comments

Javier Martinez Canillas March 8, 2022, 9:13 a.m. UTC | #1
Hello Thomas,

On 3/3/22 21:58, Thomas Zimmermann wrote:
> Rework the fbdev deferred-I/O to not interfere with fields of struct
> page. Make the code more flexible and implement GEM SHMEM mmap on top
> of it.
> 
> This patchset removes the need for a special shadow framebuffer for
> fbdev mmap when using GEM SHMEM. SHMEM pages are now mmap'ed from
> /dev/fb directly.
>

Interesting. I wonder if you have any performance improvements after
dropping the shadow buffer.

> Patches 2 and 3 rework the fbdev deferred I/O code. It now allows
> drivers to have better control of the mmap operations. All references
> to fields in struct page are gone. The rsp state is help in a 
> separate pageref structure.
>

That's a very nice cleanup. This really was a huge layering violation.
 
> Patches 4 to 7 provide callbacks an helpers to implement deferred I/O
> with DRM drivers. Specifically, patch 6 introduces a callback to create
> a dumb buffer for fbdev. This will be useful for many drivers that
> currently cannot use generic fbdev emulation because of special placement
> needs of the BO, such as amdgpu or radeon. The drivers can handle the
> differences to regular dumb buffers in their new callback implementation.
> 
> Patch 8 extends the GEM SHMEM memory manager with a new helper for fbdev
> dumb-buffer creation. The returned BO has it's mmap set up to implement
> deferred I/O with SHMEM pages. No additional shadow buffer is requires
> any longer. Many drivers can immediatelly benefit from this change.
> 
> Patch 9 extends virtgpu to support fbdev dumb buffers. It's build on
> top of GEM SHMEM, but has some modifications that need to be implemented
> for fbdev as well.
> 
> There's no immediate fbdev performance improvement from this patchset.
> Most of all, it removes unnecessary shadow framebuffers and rsp memcpys.
> A shadow fb for a FullHD display is ~8 MiB, which we now save. The patches
> do reduce latency between drawing to the fbdev buffer to displaying
> on the screen. Watching a video on the fbdev console felt smoother and
> had less flickering.
>

Awesome. And you also answered here the question I had above.
Thomas Zimmermann March 8, 2022, 9:44 a.m. UTC | #2
Hi Javier

Am 08.03.22 um 10:13 schrieb Javier Martinez Canillas:
> Hello Thomas,
> 
> On 3/3/22 21:58, Thomas Zimmermann wrote:
>> Rework the fbdev deferred-I/O to not interfere with fields of struct
>> page. Make the code more flexible and implement GEM SHMEM mmap on top
>> of it.
>>
>> This patchset removes the need for a special shadow framebuffer for
>> fbdev mmap when using GEM SHMEM. SHMEM pages are now mmap'ed from
>> /dev/fb directly.
>>
> 
> Interesting. I wonder if you have any performance improvements after
> dropping the shadow buffer.
> 
>> Patches 2 and 3 rework the fbdev deferred I/O code. It now allows
>> drivers to have better control of the mmap operations. All references
>> to fields in struct page are gone. The rsp state is help in a
>> separate pageref structure.
>>
> 
> That's a very nice cleanup. This really was a huge layering violation.
>   
>> Patches 4 to 7 provide callbacks an helpers to implement deferred I/O
>> with DRM drivers. Specifically, patch 6 introduces a callback to create
>> a dumb buffer for fbdev. This will be useful for many drivers that
>> currently cannot use generic fbdev emulation because of special placement
>> needs of the BO, such as amdgpu or radeon. The drivers can handle the
>> differences to regular dumb buffers in their new callback implementation.
>>
>> Patch 8 extends the GEM SHMEM memory manager with a new helper for fbdev
>> dumb-buffer creation. The returned BO has it's mmap set up to implement
>> deferred I/O with SHMEM pages. No additional shadow buffer is requires
>> any longer. Many drivers can immediatelly benefit from this change.
>>
>> Patch 9 extends virtgpu to support fbdev dumb buffers. It's build on
>> top of GEM SHMEM, but has some modifications that need to be implemented
>> for fbdev as well.
>>
>> There's no immediate fbdev performance improvement from this patchset.
>> Most of all, it removes unnecessary shadow framebuffers and rsp memcpys.
>> A shadow fb for a FullHD display is ~8 MiB, which we now save. The patches
>> do reduce latency between drawing to the fbdev buffer to displaying
>> on the screen. Watching a video on the fbdev console felt smoother and
>> had less flickering.
>>
> 
> Awesome. And you also answered here the question I had above.

There's no significant improvement in performance.

But when I play Bug Buck Bunny in mplayer's benchmark mode, the 
displayed animation looks smoother than before. The test finishes in 73 
seconds instead of 74. I guess that one-second improvement comes from 
the memcpy that we now save. I use a somewhat recent Core i7 for 
testing. On old hardware, we might get more performance out of it. I can 
really only guess here.

Best regards
Thomas

>