On 2017å¹´09æ??12æ?¥ 19:00, Christian König wrote: > NAK for the CS part (but I think that was just as prove of concept > anyway). > > Dave's sync object implementation is already done for this, the > problem is that isn't merged yet into amd-staging-4.12 (but should be > in 4-13). Yeah, I already realized it before, if only vm part, I agree syncobj can satisfy the synchronization between SDMA and other engine. But from your old dependency perspective, we also should cover wait_ioctl and wait_any/all_ioctl, which should also can wait fence from syncfile, so it makes sense that cs_ioctl and va_ioctl return syncfile fd to UMD. Compared with only retuning fence seq number, UMD can be convenience to pass fd to any where it needs to sync and as dependency or to semaphore, and they don't need to construct ip_type/ip_instance/ctx_id/ring any more, even the caching fences in ctx becomes no meanless if don't consider compatibility. Regards, David Zhou > > The VM part looks similar to what I had in mind, but we should base > this on Dave's sync_obj stuff as well. > > Regards, > Christian. > > Am 12.09.2017 um 12:23 schrieb Chunming Zhou: >> *** BLURB HERE *** >> >> Chunming Zhou (5): >> drm/amdgpu: introduce sync file for CS returning fence >> drm/amdgpu: add syncfile chunk support >> drm/amdgpu: add syncfile support to wait ioctl >> drm/amdgpu: add syncfile fence support to wait_any/all ioctl >> drm/amdgpu: expand mapping ioctl to return fence to UMD by using >> syncfile >> >> drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 90 >> ++++++++++++++++++++++++--------- >> drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 24 +++++++-- >> include/uapi/drm/amdgpu_drm.h | 13 ++++- >> 3 files changed, 96 insertions(+), 31 deletions(-) >> >
