In order to support out-of-line error capture, we need to remove the
active request from HW and put it to one side while a worker compresses
and stores all the details associated with that request. (As that
compression may take an arbitrary user-controlled amount of time, we
want to let the engine continue running on other workloads while the
hanging request is dumped.) Not only do we need to remove the active
request, but we also have to remove its context and all requests that
were dependent on it (both in flight, queued and future submission).
Finally once the capture is complete, we need to be able to resubmit the
request and its dependents and allow them to execute.
v2: Replace stack recursion with a simple list.
v3: Check all the parents, not just the first, when searching for a
stuck ancestor!
References: https://gitlab.freedesktop.org/drm/intel/issues/738
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200116184754.2860848-2-chris@chris-wilson.co.uk
(cherry picked from commit 32ff621fd7)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
As we use a mutex to serialise the first acquire (as it may be a lengthy
operation), but only an atomic decrement for the release, we have to
be careful in case a second thread races and completes both
acquire/release as the first finishes its acquire.
Thread A Thread B
i915_active_acquire i915_active_acquire
atomic_read() == 0 atomic_read() == 0
mutex_lock() mutex_lock()
atomic_read() == 0
ref->active();
atomic_inc()
mutex_unlock()
atomic_read() == 1
i915_active_release
atomic_dec_and_test() -> 0
ref->retire()
atomic_inc() -> 1
mutex_unlock()
So thread A has acquired the ref->active_count but since the ref was
still active at the time, it did not initialise it. By switching the
check inside the mutex to an atomic increment only if already active, we
close the race.
Fixes: c9ad602fea ("drm/i915: Split i915_active.mutex into an irq-safe spinlock for the rbtree")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200126102346.1877661-3-chris@chris-wilson.co.uk
(cherry picked from commit ac0e331a62)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Currently we create a new mmap_offset for every call to
mmap_offset_ioctl. This exposes ourselves to an abusive client that may
simply create new mmap_offsets ad infinitum, which will exhaust physical
memory and the virtual address space. In addition to the exhaustion, a
very long linear list of mmap_offsets causes other clients using the
object to incur long list walks -- these long lists can also be
generated by simply having many clients generate their own mmap_offset.
However, we can simply use the drm_vma_node itself to manage the file
association (allow/revoke) dropping our need to keep an mmo per-file.
Then if we keep a small rbtree of per-type mmap_offsets, we can lookup
duplicate requests quickly.
Fixes: cc662126b4 ("drm/i915: Introduce DRM_I915_GEM_MMAP_OFFSET")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
Reviewed-by: Abdiel Janulgue <abdiel.janulgue@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200120104924.4000706-3-chris@chris-wilson.co.uk
(cherry picked from commit 7865559872)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
intel_prepare_plane_fb() will always pin plane_state->hw.fb whenever
it is present. We copy that from the master plane to the slave plane,
but we fail to copy the corresponding ggtt view. Thus when it comes time
to pin the slave plane's fb we use some stale ggtt view left over from
the last time the plane was used as a non-slave plane. If that previous
use involved 90/270 degree rotation or remapping we'll try to shuffle
the pages of the new fb around accordingingly. However the new
fb may be backed by a bo with less pages than what the ggtt view
rotation/remapped info requires, and so we we trip a GEM_BUG().
Steps to reproduce on icl:
1. plane 1: whatever
plane 6: largish !NV12 fb + 90 degree rotation
2. plane 1: smallish NV12 fb
plane 6: make invisible so it gets slaved to plane 1
3. GEM_BUG()
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Closes: https://gitlab.freedesktop.org/drm/intel/issues/951
Fixes: 1f594b209f ("drm/i915: Remove special case slave handling during hw programming, v3.")
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200110183228.8199-1-ville.syrjala@linux.intel.com
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
(cherry picked from commit 103605e0d1)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
The post-fastset "does anyone still need a full modeset?" for
port sync looks busted. The outer loop bails out of a full modeset
is still needed by the current crtc, and then we skip forcing
a full modeset on the related crtcs. That's totally the opposite
of what we want.
The MST path has the logic mostly the other way around so it
looks correct. To fix the port sync case let's follow the MST
logic for both. So, if the current crtc already needs a modeset
we do nothing. otherwise we check if any of the related crtcs
needs a modeset, and if so we force a full modeset for the
current crtc.
And while at let's change the else if to a plain if to so
we don't have needless coupling between the MST and port sync
checks.
Cc: José Roberto de Souza <jose.souza@intel.com>
Cc: Manasi Navare <manasi.d.navare@intel.com>
Fixes: 05a8e45136 ("drm/i915/display: Use external dependency loop for port sync")
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200115190813.17971-1-ville.syrjala@linux.intel.com
Reviewed-by: José Roberto de Souza <jose.souza@intel.com>
(cherry picked from commit d0eed1545f)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
amd-drm-next-5.6-2020-02-05:
amdgpu:
- EDC fixes for Arcturus
- GDDR6 memory training fixe
- Fix for reading gfx clockgating registers while in GFXOFF state
- i2c freq fixes
- Misc display fixes
- TLB invalidation fix when using semaphores
- VCN 2.5 instancing fixes
- Switch raven1 gfxoff to a blacklist
- Coreboot workaround for KV/KB
- Root cause dongle fixes for display and revert workaround
- Enable GPU reset for renoir and navi
- Navi overclocking fixes
- Fix up confusing warnings in display clock validation on raven
amdkfd:
- SDMA fix
radeon:
- Misc LUT fixes
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Alex Deucher <alexdeucher@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200206035458.3894-1-alexander.deucher@amd.com
The DMA direction is only used by the DMA API, so there is no use in
setting it when a buffer object isn't mapped with the DMA API.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Tested-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
This partially reverts the DMA API support that was recently merged
because it was causing performance regressions on older Tegra devices.
Unfortunately, the cache maintenance performed by dma_map_sg() and
dma_unmap_sg() causes performance to drop by a factor of 10.
The right solution for this would be to cache mappings for buffers per
consumer device, but that's a bit involved. Instead, we simply revert to
the old behaviour of sharing IOVA mappings when we know that devices can
do so (i.e. they share the same IOMMU domain).
Cc: <stable@vger.kernel.org> # v5.5
Reported-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Tested-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Older Tegra devices only allow addressing 32 bits of memory, so whether
or not the host1x is attached to an IOMMU doesn't matter. host1x IOMMU
attachment is only needed on devices that can address memory beyond the
32-bit boundary and where the host1x doesn't support the wide GATHER
opcode that allows it to access buffers at higher addresses.
Cc: <stable@vger.kernel.org> # v5.5
Signed-off-by: Thierry Reding <treding@nvidia.com>
Tested-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Previously, the syfs functionality for restoring the default powerplay
table was sourcing it's information from the currently-staged powerplay
table.
This patch adds a step to cache the first overdrive table that we see on
boot, so that it can be used later to "restore" the powerplay table
v2: sqaush my original with Matt's fix
Bug: https://gitlab.freedesktop.org/drm/amd/issues/1020
Signed-off-by: Matt Coffin <mcoffin13@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org # 5.5.x
[Why]
When we disable a connector we don't explicitly remove it from the module so the
display is still cached(SW) in the hdcp_module.
SST: no issues because we can only have 1 display per link
MST: We have x displays per link, now if we disable 1 we don't remove it from the
module so the module has x display cached(SW).
If we try to enable HDCP, psp verification will fail because we are reporting x
displays while the HW only has x-1 display enabled
[How]
Check the callback for when we disable stream and call remove display.
Signed-off-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Reviewed-by: Rodrigo Siqueira <Rodrigo.Siqueira@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[why]
Remove a backslash symbol accidentally left in increase bpp function
when computing mst dsc configuration.
Signed-off-by: Mikita Lipski <mikita.lipski@amd.com>
Reviewed-by: Zhan Liu <zhan.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The sdma_queue_count increment should be done before
execute_queues_cpsch(), which calls pm_calc_rlib_size() where
sdma_queue_count is used to calculate whether over_subscription is
triggered.
With the previous code, when a SDMA queue is created,
compute_queue_count in pm_calc_rlib_size() is one more than the
actual compute queue number, because the queue_count has been
incremented while sdma_queue_count has not. This patch fixes that.
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
In current code we're essentially drawing the cursor on every pipe
that contains it. This only works when the planes have the same
scaling for src to dest rect, otherwise we'll get "double cursor" where
one cursor is incorrectly filtered and offset from the real position.
[How]
Without dedicated cursor planes on DCN we require at least one pipe
that matches the scaling of the current timing.
This is an optimization and workaround for the most common case where
the top-most plane is not scaled but the bottom-most plane is scaled.
Whenever a pipe has a parent pipe in the blending tree whose recout
fully contains the current pipe we can disable the pipe.
This only applies when the pipe is actually visible of course.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The OR routing logic in NVKM does not expect to receive supervisor
interrupts until the DD has provided consistent information on the
ORs it's using and the EVO/NVD assembly state to match.
The combination of changing window ownership + core channel update
during display init triggered a situation where we'd disconnect an
OR from the pad it was meant to still be driving on some systems.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
For various complicated reasons, we need to avoid sending a core update
method during display init. Something, which we've been required to do
on GV100 and up because we've been assigning windows to heads there and
the HW is rather picky about when that's allowed.
This moves window assignment into the modesetting path at a point where
it's much safer to send our first update methods to NVDisplay.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Commit fd67e9c6ed ("drm/tegra: Do not implement runtime PM") replaced
the generic runtime PM usage by a host1x bus-specific implementation in
order to work around some assumptions baked into runtime PM that are in
conflict with the requirements in the Tegra DRM driver.
Unfortunately the new runtime PM callbacks are not setup yet at the time
when the SOR driver first needs to resume the device to register the SOR
pad clock, and accesses to register will cause the system to hang.
Note that this only happens on Tegra124 and Tegra210 because those are
the only SoCs where the SOR pad clock is registered from the SOR driver.
Later generations use a SOR pad clock provided by the BPMP.
Fix this by moving the registration of the SOR pad clock after the
host1x client has been registered. That's somewhat suboptimal because
this could potentially, though it's very unlikely, cause the Tegra DRM
to be probed if the SOR happens to be the last subdevice to register,
only to be immediately removed again if the SOR pad output clock fails
to register. That's just a minor annoyance, though, and doesn't justify
implementing a workaround.
Fixes: fd67e9c6ed ("drm/tegra: Do not implement runtime PM")
Signed-off-by: Thierry Reding <treding@nvidia.com>
If the driver fails to probe, make sure to disable runtime PM again.
While at it, make the cleanup code in ->remove() symmetric.
Signed-off-by: Thierry Reding <treding@nvidia.com>
