On Thu, 8 Aug 2019 at 23:29, Rob Herring robh@kernel.org wrote:
Up until now, a single shared GPU address space was used. This is not ideal as there's no protection between processes and doesn't work for supporting the same GPU/CPU VA feature. Most importantly, this will hopefully mitigate Alyssa's fear of WebGL, whatever that is.
Most of the changes here are moving struct drm_mm and struct panfrost_mmu objects from the per device struct to the per FD struct. The critical function is panfrost_mmu_as_get() which handles allocating and switching the h/w address spaces.
Cc: Tomeu Vizoso tomeu.vizoso@collabora.com Cc: David Airlie airlied@linux.ie Cc: Daniel Vetter daniel@ffwll.ch Cc: Robin Murphy robin.murphy@arm.com Cc: Steven Price steven.price@arm.com Cc: Alyssa Rosenzweig alyssa.rosenzweig@collabora.com Signed-off-by: Rob Herring robh@kernel.org
This depends on madvise support (now in drm-misc) and the heap/no-exec series (just the rework). Seeems to be working pretty well for me, but more testing would be helpful. I've run multiple 'glmark2-es2-drm --off-screen' instances and Gnome Shell. Running more than 8 clients (at least for T860) will hit the address space switch code paths.
Rob
drivers/gpu/drm/panfrost/TODO | 4 - drivers/gpu/drm/panfrost/panfrost_device.c | 2 + drivers/gpu/drm/panfrost/panfrost_device.h | 24 ++- drivers/gpu/drm/panfrost/panfrost_drv.c | 31 ++- drivers/gpu/drm/panfrost/panfrost_gem.c | 15 +- drivers/gpu/drm/panfrost/panfrost_gem.h | 3 + drivers/gpu/drm/panfrost/panfrost_job.c | 12 +- drivers/gpu/drm/panfrost/panfrost_mmu.c | 220 +++++++++++++++------ drivers/gpu/drm/panfrost/panfrost_mmu.h | 8 + 9 files changed, 239 insertions(+), 80 deletions(-)
[snip]
@@ -413,7 +414,28 @@ panfrost_open(struct drm_device *dev, struct drm_file *file) panfrost_priv->pfdev = pfdev; file->driver_priv = panfrost_priv;
return panfrost_job_open(panfrost_priv);
spin_lock_init(&panfrost_priv->mm_lock);/* 4G enough for now. can be 48-bit */panfrost_priv->mm.color_adjust = panfrost_drm_mm_color_adjust;drm_mm_init(&panfrost_priv->mm, SZ_32M >> PAGE_SHIFT, (SZ_4G - SZ_32M) >> PAGE_SHIFT);
Drive-by-comment: drm_mm_init will trample your color_adjust. Intentional?