On Tue, Nov 17, 2015 at 11:16:09PM +0100, Daniel Vetter wrote: > On Thu, Nov 5, 2015 at 12:56 PM, Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> wrote: > > The tricky part is reviewing the i915_gem_release_mmap() callers to > > ensure that they have the right barrier. If you make > > i915_gem_release_mmap() assert it holds an rpm ref, and then make the > > i915_gem_release_all_mmaps() unlink the node directly that should do the > > trick. > > I think the easier solution would be to add a mutex for the > fault_list. We call release_mmap from a lot of places, We don't though. The only times we do are when we touching hw registers (or gsm): set_tiling_ioctl - which also may unbind and so needs rpm fence_write - which needs rpm to write the reigsters vma_unbind - which needs rpm to write through the gsm set_caching - which needs rpm to write through the gsm and currently by rpm itself. I think it is certainly reasonable to use the rpm barriers for the faulted list. The only one where we have to actually ensure we hold rpm is the set_tiling_ioctl. -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/intel-gfx
