drm/i915: Move GEM activity tracking into a common struct reservation_object
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 12:58:44 +0000 (13:58 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 19:53:50 +0000 (20:53 +0100)
commitd07f0e59b2c762584478920cd2d11fba2980a94a
treea9fc0cade87570bdec356493e6744ee5223bdaaf
parentf0cd518206e1a47e57bc251e1faba9d38eadcc59
drm/i915: Move GEM activity tracking into a common struct reservation_object

In preparation to support many distinct timelines, we need to expand the
activity tracking on the GEM object to handle more than just a request
per engine. We already use the struct reservation_object on the dma-buf
to handle many fence contexts, so integrating that into the GEM object
itself is the preferred solution. (For example, we can now share the same
reservation_object between every consumer/producer using this buffer and
skip the manual import/export via dma-buf.)

v2: Reimplement busy-ioctl (by walking the reservation object), postpone
the ABI change for another day. Similarly use the reservation object to
find the last_write request (if active and from i915) for choosing
display CS flips.

Caveats:

 * busy-ioctl: busy-ioctl only reports on the native fences, it will not
warn of stalls (in set-domain-ioctl, pread/pwrite etc) if the object is
being rendered to by external fences. It also will not report the same
busy state as wait-ioctl (or polling on the dma-buf) in the same
circumstances. On the plus side, it does retain reporting of which
*i915* engines are engaged with this object.

 * non-blocking atomic modesets take a step backwards as the wait for
render completion blocks the ioctl. This is fixed in a subsequent
patch to use a fence instead for awaiting on the rendering, see
"drm/i915: Restore nonblocking awaits for modesetting"

 * dynamic array manipulation for shared-fences in reservation is slower
than the previous lockless static assignment (e.g. gem_exec_lut_handle
runtime on ivb goes from 42s to 66s), mainly due to atomic operations
(maintaining the fence refcounts).

 * loss of object-level retirement callbacks, emulated by VMA retirement
tracking.

 * minor loss of object-level last activity information from debugfs,
could be replaced with per-vma information if desired

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-21-chris@chris-wilson.co.uk
15 files changed:
drivers/gpu/drm/i915/i915_debugfs.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_batch_pool.c
drivers/gpu/drm/i915/i915_gem_dmabuf.c
drivers/gpu/drm/i915/i915_gem_dmabuf.h [deleted file]
drivers/gpu/drm/i915/i915_gem_execbuffer.c
drivers/gpu/drm/i915/i915_gem_gtt.c
drivers/gpu/drm/i915/i915_gem_gtt.h
drivers/gpu/drm/i915/i915_gem_request.c
drivers/gpu/drm/i915/i915_gem_request.h
drivers/gpu/drm/i915/i915_gpu_error.c
drivers/gpu/drm/i915/intel_atomic_plane.c
drivers/gpu/drm/i915/intel_display.c
drivers/gpu/drm/i915/intel_drv.h