drm/msm: Defer VMA unmap for fb unpins
authorRob Clark <robin.clark@oss.qualcomm.com>
Sun, 29 Jun 2025 20:13:24 +0000 (13:13 -0700)
committerRob Clark <robin.clark@oss.qualcomm.com>
Sat, 5 Jul 2025 00:48:38 +0000 (17:48 -0700)
commit3bebfd53af0f7c8ea55094ba7b8b8b907024bb7b
treee2c1ce0d462cc4bcddabdcfd1a71cef764e5fcd9
parent8d4c21718cebf316daa36d06c7206d72e738fa16
drm/msm: Defer VMA unmap for fb unpins

With the conversion to drm_gpuvm, we lost the lazy VMA cleanup, which
means that fb cleanup/unpin when pageflipping to new scanout buffers
immediately unmaps the scanout buffer.  This is costly (with tlbinv,
it can be 4-6ms for a 1080p scanout buffer, and more for higher
resolutions)!

To avoid this, introduce a vma_ref, which is incremented whenever
userspace has a GEM handle or dma-buf fd.  When unpinning if the
vm is the kms->vm we defer tearing down the VMA until the vma_ref
drops to zero.  If the buffer is still part of a flip-chain then
userspace will be holding some sort of reference to the BO, either
via a GEM handle and/or dma-buf fd.  So this avoids unmapping the VMA
when there is a strong possibility that it will be needed again.

Signed-off-by: Rob Clark <robin.clark@oss.qualcomm.com>
Tested-by: Antonino Maniscalco <antomani103@gmail.com>
Reviewed-by: Antonino Maniscalco <antomani103@gmail.com>
Patchwork: https://patchwork.freedesktop.org/patch/661538/
drivers/gpu/drm/msm/msm_drv.c
drivers/gpu/drm/msm/msm_drv.h
drivers/gpu/drm/msm/msm_fb.c
drivers/gpu/drm/msm/msm_gem.c
drivers/gpu/drm/msm/msm_gem.h
drivers/gpu/drm/msm/msm_gem_prime.c