Hi all,
I've finally managed to clean up my pwrite/pread rework. Major changes since the first submission: - fixed a bunch of bugs, some of them discovered by the intel-gpu-tools gem test suite - extracted the clflush helper into common drm code as suggested by Chris Wilson. - extended the prefault helpers in pagemap.h instead of rolling our own. Also add a patch to not prefault for writes (which change the userspace memory) before we're committed to the write. Both address issues raised by Keith Packard.
Why this is cool stuff: - fixes the spurious -EFAULTS when handing in a pointer to a gem gtt mappings. See https://bugs.freedesktop.org/show_bug.cgi?id=38115 - Kills a bunch of code by dropping the practically useless ranged cpu read domain tracking and integrating the slow paths into the fast paths. - Fixes pwrite/pread to use snooped cpu access on llc machines, hereby decently speeding up microbenchmarks. - Paves the way for further clever tricks to reduce pressure on the mappable gtt area.
Reviews and comments highly welcome.
Yours, Daniel
Daniel Vetter (13): drm/i915: fall through pwrite_gtt_slow to the shmem slow path drm/i915: rewrite shmem_pwrite_slow to use copy_from_user drm/i915: rewrite shmem_pread_slow to use copy_to_user drm/i915: merge shmem_pwrite slow&fast-path drm/i915: merge shmem_pread slow&fast-path drm: add helper to clflush a virtual address range drm/i915: move clflushing into shmem_pread drm/i915: kill ranged cpu read domain support drm/i915: don't use gtt_pwrite on LLC cached objects drm/i915: don't call shmem_read_mapping unnecessarily mm: extend prefault helpers to fault in more than PAGE_SIZE drm/i915: drop gtt slowpath drm/i915: don't clobber userspace memory before commiting to the pread
drivers/gpu/drm/drm_cache.c | 23 ++ drivers/gpu/drm/i915/i915_drv.h | 7 - drivers/gpu/drm/i915/i915_gem.c | 824 +++++++++++---------------------------- include/drm/drmP.h | 1 + include/linux/pagemap.h | 28 +- 5 files changed, 273 insertions(+), 610 deletions(-)