According to Dave, once you've started an L2T flush, all L2T accesses
will be blocked until the flush completes. This fixes a consistent
3-4ms stall between the ioctl and running the job, and 3DMMES Taiji
goes from 27fps to 110fps.
v2: Leave a note about why we don't need to wait for completion.
Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: 57692c94dc ("drm/v3d: Introduce a new DRM driver for Broadcom V3D V3.x+")
Reviewed-by: Dave Emett <david.emett@broadcom.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20181203222438.25417-4-eric@anholt.net
This is the write combiner for TMU writes. You're supposed to flush
that at job end if you had dirtied any cachelines. Flushing it at job
start then doesn't make any sense.
Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: 57692c94dc ("drm/v3d: Introduce a new DRM driver for Broadcom V3D V3.x+")
Reviewed-by: Dave Emett <david.emett@broadcom.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20181203222438.25417-3-eric@anholt.net
I missed one case during the recent revert of the replace_fence
interface change.
Fixes: 0b258ed1a2 drm: revert "expand replace_fence to support timeline point v2"
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Link: https://patchwork.freedesktop.org/patch/266134/
This reverts commit 9a09a42369.
The whole interface isn't thought through. Since this function can't
fail we actually can't allocate an object to store the sync point.
Sorry, I should have taken the lead on this from the very beginning and
reviewed it more thoughtfully. Going to propose a new interface as a
follow up change.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Link: https://patchwork.freedesktop.org/patch/265580/
The core scheduler tells us when the job is pushed to the scheduler's
queue, and I had the job_run functions saying when they actually queue
the job to the hardware. By adding tracepoints for the very top of
the ioctls and the IRQs signaling job completion, "perf record -a -e
v3d:.\* -e gpu_scheduler:.\* <job>; perf script" gets you a pretty
decent timeline.
Signed-off-by: Eric Anholt <eric@anholt.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20181201005759.28093-5-eric@anholt.net
Reviewed-by: Dave Emett <david.emett@broadcom.com>
The TFU can copy from raster, UIF, and SAND input images to UIF output
images, with optional mipmap generation. This will certainly be
useful for media EGL image input, but is also useful immediately for
mipmap generation without bogging the V3D core down.
For now we only run the queue 1 job deep, and don't have any hang
recovery (though I don't think we should need it, with TFU). Queuing
multiple jobs in the HW will require synchronizing the YUV coefficient
regs updates since they don't get FIFOed with the job.
v2: Change the ioctl to IOW instead of IOWR, always set COEF0, explain
why TFU is AUTH, clarify the syncing docs, drop the unused TFU
interrupt regs (you're expected to use the hub's), don't take
&bo->base for NULL bos.
v3: Fix a little whitespace alignment (noticed by checkpatch), rebase
on drm_sched_job_cleanup() changes.
Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Dave Emett <david.emett@broadcom.com> (v2)
Link: https://patchwork.freedesktop.org/patch/264607/
The extra to_v3d_bo() calls came from copying this from the vc4
driver, which stored the cma gem object in the structs.
v2: Fix an unused var warning
Signed-off-by: Eric Anholt <eric@anholt.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20181108161654.19888-4-eric@anholt.net
Reviewed-by: Boris Brezillon <boris.brezillon@bootlin.com> (v1)
Let's support simultaneous submissions to multiple engines.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Huang Rui <ray.huang@amd.com>
Link: https://patchwork.kernel.org/patch/10626149/
flags can be used by driver to decide whether need to block wait submission.
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
SIgned-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.kernel.org/patch/10641339/
Once we push the job, the scheduler could run it and free it. So, if
we want to reference their fences, we need to grab them before then.
I haven't seen this happen in many days of conformance test runtime,
but let's still close the race.
Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: 57692c94dc ("drm/v3d: Introduce a new DRM driver for Broadcom V3D V3.x+")
Link: https://patchwork.freedesktop.org/patch/254119/
Reviewed-by: Boris Brezillon <boris.brezillon@bootlin.com>
we can place a fence to a timeline point after expanded.
v2: change func parameter order
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/246543/
we can fetch timeline point fence after expanded.
v2: The parameter fence is the result of the function and should come last.
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/246541/
entity has a scheduler field and we don't need the sched argument
in any of the functions where entity is provided.
Signed-off-by: Nayan Deshmukh <nayan26deshmukh@gmail.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Since our seqno value comes from a counter associated with the GPU
ring, not the entity (aka client), they'll be completed out of order.
There's actually no need for this code at all, since we don't have
enable_signaling() and thus DMA_FENCE_SIGNALED_BIT will be set before
we could be called.
Signed-off-by: Eric Anholt <eric@anholt.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20180605190302.18279-2-eric@anholt.net
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>
Between creation and queueing of a job, you need to prevent any other
job from being created and queued. Otherwise the scheduler's fences
may be signaled out of seqno order.
v2: move mutex unlock to the error label.
Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: 57692c94dc ("drm/v3d: Introduce a new DRM driver for Broadcom V3D V3.x+")
Link: https://patchwork.freedesktop.org/patch/msgid/20180606174851.12433-1-eric@anholt.net
Reviewed-by: Lucas Stach <l.stach@pengutronix.de>
This driver will be used to support Mesa on the Broadcom 7268 and 7278
platforms.
V3D 3.3 introduces an MMU, which means we no longer need CMA or vc4's
complicated CL/shader validation scheme. This massively changes the
GEM behavior, so I've forked off to a new driver.
v2: Mark SUBMIT_CL as needing DRM_AUTH. coccinelle fixes from kbuild
test robot. Drop personal git link from MAINTAINERS. Don't
double-map dma-buf imported BOs. Add kerneldoc about needing MMU
eviction. Drop prime vmap/unmap stubs. Delay mmap offset setup
to mmap time. Use drm_dev_init instead of _alloc. Use
ktime_get() for wait_bo timeouts. Drop drm_can_sleep() usage,
since we don't modeset. Switch page tables back to WC (debug
change to coherent had slipped in). Switch
drm_gem_object_unreference_unlocked() to
drm_gem_object_put_unlocked(). Simplify overflow mem handling by
not sharing overflow mem between jobs.
v3: no changes
v4: align submit_cl to 64 bits (review by airlied), check zero flags in
other ioctls.
Signed-off-by: Eric Anholt <eric@anholt.net>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch> (v4)
Acked-by: Dave Airlie <airlied@linux.ie> (v3, requested submit_cl change)
Link: https://patchwork.freedesktop.org/patch/msgid/20180430181058.30181-3-eric@anholt.net
