One of the costs of superioctl has been the need to perform relocations
inside the kernel. The cost of mapping the buffers to the CPU and writing
data is fairly high, especially if those buffers have been mapped and read
by the GPU.
If we assume that buffers don't move around very often, we can have the
client compute the relocations itself using the previous GPU address. When
that object doesn't move, the kernel can skip computing and writing the
updated data.
Here's a patch which adds a new field to struct drm_bo_info_req called
'presumed_offset', and a new DRM_BO_HINT_PRESUMED_OFFSET that is set when
this field has been filled in by the client.
There are two separate optimizations performed when the presumed_offset is
correct:
1. i915_exec_reloc checks to see if all previous buffer offsets were guessed
correctly. If so, there's no need for it to look at *any* of the
relocations for a buffer. When this happens, it skips the whole
relocation process, simply returning success.
2. i915_apply_reloc checks to see if the target buffer offset was guessed
correctly. If so, it skips mapping the relocatee, computing the
relocation and writing the value. If no relocations are needed, the
relocatee should never be mapped to the CPU, and so the kernel shouldn't
need to wait for any fences to pass.
If drmMinor >= 6, the intel DDX driver will enable vblank events on both
pipes. If drmMinor >= 10 on pre-965 chipsets, the intel DDX driver will
swap the pipe<->plane mapping to allow for framebuffer compression on
laptop screens. This means the secondary vblank counter (corresponding
to pipe B) will be incremented when vblank interrupts occur.
Now Mesa waits for vblank events on whichever plane has a greater
portion of the displayed window. So it will happly ask to wait for the
primary counter even though that one won't increment.
So we can fix this in either the DDX driver, Mesa or the kernel (though
I thought we already had several times).
Since current (and previous) userspace assumes it's talking about a pipe
== plane situation and now uses planes when talking to the kernel, we
should probably just hide the mapping details there (indeed they already
are hidden there for vblank swaps), which this patch does.
So as far as userland is concerned, whether we call things planes or
pipes is irrelevant, as long as kernel developers understand that
userland hands them planes and they have to figure out which pipe that
corresponds to (which will typically be the same on 965+ hardware and
reversed on pre-965 mobile chips).
This fix is actually a bit of a cleanup too--it moves lock freeing to
drm_rmmap_locked and out of drm_lastclose. This makes it symmetrical with
addmap and also prevents the lock from being incorrectly freed from driver
mappings.
Jakob Bornecrantz
Dave Airlie
Dave Airlie
Jerome Glisse
Jakob Bornecrantz
Alan Hourihane
Keith Packard
José Fonseca
Jesse Barnes
Kristian Høgsberg
Jesse Barnes