On Tue, May 16, 2017 at 09:29:37AM +0100, Matthew Auld wrote: > In i915 we are aiming to support huge GTT pages for the GPU, and to > complement this we also want to enable THP for our shmem backed objects. > Even though THP is supported in shmemfs it can only be enabled through > the huge= mount option, but for users of the kernel mounted shm_mnt like > i915, we are a little stuck. There is the sysfs knob shmem_enabled to > either forcefully enable/disable the feature, but that seems to only be > useful for testing purposes. What we propose is to expose a driver > overridable huge option as part of shmem_inode_info to control the use > of THP for a given mapping. I don't like this. It's kinda hacky. Is there a reason why i915 cannot mount a new tmpfs for own use? Or other option would be to change default to SHMEM_HUGE_ADVISE and wire up fadvise handle to control per-file allocation policy. -- Kirill A. Shutemov _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
