CP_RB_DOORBELL_RANGE_LOWER/UPPER and CP_MEC_DOORBELL_RANGE_LOWER/UPPER
are used for waking up an idle scheduler and for power gating support.
Usually the first few doorbells in pci doorbell bar are used for RB
and all leftover for MEC. This patch fixes the incorrect settings.
Theoretically, gfx ring doorbells should come before all MEC doorbells
to be consistent with the design. However, since the doorbell
allocations are agreed by all and we are not free to change them, also
considering the kernel MEC ring doorbells which are before gfx ring
doorbells are not used often, we compromise by leaving the doorbell
allocations unchanged.
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>
Temporarily removing eviction fences to avoid triggering them by
accident is no longer necessary due to the fence_owner logic in
amdgpu_sync_resv.
As a result the ef_list usage of amdgpu_amdkfd_remove_eviction_fence
and amdgpu_amdkfd_add_eviction_fence are no longer needed.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Use FENCE_OWNER_KFD to synchronize PT/PD initialization and clearing
of page table entries. This avoids triggering KFD eviction fences on
the PD reservation objects of compute VMs.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
sriov's gpu_recover inside xgpu_ai_mailbox_flr_work would cause duplicate recover in TDR.
TDR's gpu_recover would be triggered by amdgpu_job_timedout,
that could avoid vk-cts failure by unexpected recover.
Signed-off-by: Wentao Lou <Wentao.Lou@amd.com>
Acked-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Bisected guest kernel changes crashing qemu. Landed at
"6c1cd97bda drm/virtio: fix resource id handling". Looked again, and
noticed we where not only leaking *some* ids, but *all* ids. The old
code never ever called virtio_gpu_resource_id_put().
So, commit 6c1cd97bda effectively makes the linux kernel starting
re-using IDs after releasing them, and apparently virglrenderer can't
deal with that. Oops.
This patch puts a temporary stopgap into place for the 5.0 release.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208140409.15280-1-kraxel@redhat.com
drm/tegra: Changes for v5.1-rc1
This set of changes starts of with some refactoring of the CEC support
to make it reusable on Tegra210 and later. Following are a couple of
fixes for HDMI audio support (via HDA).
The bulk here is a set of preparatory patches working towards enabling
Tegra186 support for host1x and VIC. Additional patches will be needed
to fully enable this, but they're not quite ready yet.
To round things off, this also adds support for configuring the SOR
crossbar using device tree, and fixes a couple of job-related issues in
the host1x code.
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Thierry Reding <thierry.reding@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208144721.25830-1-thierry.reding@gmail.com
On the D3 and E3 SoCs the LVDS PLL clock output provides the dot clock
to the DU channels, even when the LVDS outputs are not in use. Enable
and disable the LVDS clock output when enabling or disabling a CRTC
connected to the DPAD0 output.
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
On the D3 and E3 platforms, the LVDS internal PLL supplies the pixel
clock to the DU. This works automatically for LVDS outputs as the LVDS
encoder is enabled through the bridge API, enabling the internal PLL and
clock output. However, when using the DU DPAD output with the LVDS
outputs turned off, the LVDS PLL needs to be controlled manually. Add an
API to do so, to be called by the DU driver.
The drivers/gpu/drm/rcar-du/ directory has to be treated as obj-y
unconditionally, as the LVDS driver could be built-in while the DU
driver is compiled as a module.
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
On the D3 and E3 SoCs the LVDS encoder has an extended internal PLL and
supplies a clock to the DU. That clock is used not only for the LVDS
outputs but also for the DPAD output. The LVDS encoder thus needs to be
available to the DU even when its output is disabled. Don't fail probe
in that case on D3 and E3.
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Before the driver fully moved to drm_bridge and drm_panel, it was
necessary to parse DT and locate encoder and connector nodes. The
connector node is now unused and can be removed as a parameter to
rcar_du_encoder_init(). As a consequence rcar_du_encoders_init_one() can
be greatly simplified, removing most of the DT parsing.
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Add an of_node_put when the result of of_graph_get_remote_port_parent is
not available.
Add a second of_node_put if no encoder is selected (encoder remains NULL).
The semantic match that finds the first problem is as follows
(http://coccinelle.lip6.fr):
// <smpl>
@r exists@
local idexpression e;
expression x;
@@
e = of_graph_get_remote_port_parent(...);
... when != x = e
when != true e == NULL
when != of_node_put(e)
when != of_fwnode_handle(e)
(
return e;
|
*return ...;
)
// </smpl>
Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Kieran Bingham <kieran.bingham+renesas@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Split the color management hooks along the single vs. double
buffered registers line. Of the currently programmed registers
GAMMA_MODE and the ilk+ pipe CSC are double buffered, the
LUTS and CHV CGM block are single buffered.
The double buffered register will be programmed during the
normal pipe update with evasion, and also during pipe enable
so that the settings will already be correct when the pipe
starts up before the planes are enabled.
The single buffered registers are currently programmed before
the vblank evade. Which is totally wrong, but we'll correct
that later.
v2: Add some docs to explain the two vfuncs (Matt,Uma)
Rebase
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-6-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
On g4x+ the pipe gamma enable bit for the primary plane affects
the pipe bottom color as well. The same for the pipe csc enable
bit on ilk+. Thus we must configure those bits correctly even
when the primary plane is disabled.
To make the feasible let's split those settings from the
plane_ctl() function into a seprate funciton that we can
call from the ->disable_plane() hook as well.
For consistency we'll do that on all the plane types. While
that has no real benefits at this time, it'll become useful
when we start to control the pipe gamma/csc enable bits
dynamically when we overhaul the color management code.
On pre-g4x there doesn't appear to be any way to gamma
correct the pipe bottom color, but sticking to the same
pattern doesn't hurt. And it'll still help us to do
crtc state readout correctly for the pipe gamma enable
bit for the color management overhaul.
An alternative apporach would be to still precompute these
bits into plane_state->ctl, but that would require that we
run through the plane check even when the plane isn't logically
enabled on any crtc. Currently that condition causes us to
short circuit the entire thing and not call ->check_plane().
There would also be some chicken and egg problems with
->check_plane() vs. crtc color state check that would
requite splitting certain things into multiple steps.
So all in all this seems like the easier route.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-2-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Currently gathers of a hung job are getting NOP'ed and a restarted CDMA
executes the NOP'ed gathers. There shouldn't be a reason to not restart
CDMA execution starting with a next job, avoiding the unnecessary churning
with gathers NOP'ing.
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
There is a chance that the last job has been completed at the time of
CDMA timeout handler invocation. In this case there is no need to complete
the completed job.
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Host1x doesn't have information about jobs inter-dependency, that is
something that will become available once host1x will get a proper
jobs scheduler implementation. Currently a hang job causes other unrelated
jobs to be canceled, that is a relic from downstream driver which is
irrelevant to upstream. Let's cancel only the hanging job and not to touch
other jobs in queue.
Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
The crossbar configuration is usually the same across all designs for a
given SoC generation. But sometimes there are designs that require some
other configuration.
Implement support for parsing the crossbar configuration from a device
tree. If the crossbar configuration is not present in the device tree,
fall back to the default crossbar configuration.
Signed-off-by: Thierry Reding <treding@nvidia.com>
The SOR has a crossbar that can map each lane of the SOR to each of the
SOR pads. The mapping is usually the same across designs for a specific
SoC generation, but every now and then there's a design that doesn't.
Allow the crossbar configuration to be specified in device tree to make
it possible to support these designs.
Signed-off-by: Thierry Reding <treding@nvidia.com>
The version of VIC found in Tegra186 and later incorporates improvements
with regards to context isolation. As part of those improvements, stream
ID registers were added that allow to specify separate stream IDs for
the Falcon microcontroller and the VIC memory interface.
While it is possible to also set the stream ID dynamically at runtime to
allow userspace contexts to be completely separated, this commit doesn't
implement that yet. Instead, the static VIC stream ID is programmed when
the Falcon is booted. This ensures that memory accesses by the Falcon or
the VIC are properly translated via the SMMU.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Upon driver failure, the driver core will take care of clearing the
driver data, so there's no need to do so explicitly in the driver.
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
On Tegra186 and later, the ARM SMMU provides an input address space that
is 48 bits wide. However, memory clients can only address up to 40 bits.
If the geometry is used as-is, allocations of IOVA space can end up in a
region that cannot be addressed by the memory clients.
To fix this, restrict the IOVA space to the DMA mask of the host1x
device. Note that, technically, the IOVA space needs to be restricted to
the intersection of the DMA masks for all clients that are attached to
the IOMMU domain. In practice using the DMA mask of the host1x device is
sufficient because all host1x clients share the same DMA mask.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Move initialization of the shared IOMMU domain after the host1x device
has been initialized. At this point all the Tegra DRM clients have been
attached to the shared IOMMU domain.
This is important because Tegra186 and later use an ARM SMMU, for which
the driver defers setting up the geometry for a domain until a device is
attached to it. This is to ensure that the domain is properly set up for
a specific ARM SMMU instance, which is unknown at allocation time.
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Loading the firmware requires an allocation of IOVA space to make sure
that the VIC's Falcon microcontroller can read the firmware if address
translation via the SMMU is enabled.
However, the allocation currently happens at a time where the geometry
of an IOMMU domain may not have been initialized yet. This happens for
example on Tegra186 and later where an ARM SMMU is used. Domains which
are created by the ARM SMMU driver postpone the geometry setup until a
device is attached to the domain. This is because IOMMU domains aren't
attached to a specific IOMMU instance at allocation time and hence the
input address space, which defines the geometry, is not known yet.
Work around this by postponing the firmware load until it is needed at
the time where a channel is opened to the VIC. At this time the shared
IOMMU domain's geometry has been properly initialized.
As a byproduct this allows the Tegra DRM to be created in the absence
of VIC firmware, since the VIC initialization no longer fails if the
firmware can't be found.
Based on an earlier patch by Dmitry Osipenko <digetx@gmail.com>.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Tegra DRM clients need access to their parent, so store a pointer to it
upon registration. It's technically possible to get at this by going via
the host1x client's parent and getting the driver data, but that's quite
complicated and not very transparent. It's much more straightforward and
natural to let the children know about their parent.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
The host1x CDMA push buffer is terminated by a special opcode (RESTART)
that tells the CDMA to wrap around to the beginning of the push buffer.
To accomodate the RESTART opcode, an extra 4 bytes are allocated on top
of the 512 * 8 = 4096 bytes needed for the 512 slots (1 slot = 2 words)
that are used for other commands passed to CDMA. This requires that two
memory pages are allocated, but most of the second page (4092 bytes) is
never used.
Decrease the number of slots to 511 so that the RESTART opcode fits
within the page. Adjust the push buffer wraparound code to take into
account push buffer sizes that are not a power of two.
Signed-off-by: Thierry Reding <treding@nvidia.com>
The HOST1X_CHANNEL_DMAEND is an offset relative to the value written to
the HOST1X_CHANNEL_DMASTART register, but it is currently treated as an
absolute address. This can cause SMMU faults if the CDMA fetches past a
pushbuffer's IOMMU mapping.
Properly setting the DMAEND prevents the CDMA from fetching beyond that
address and avoid such issues. This is currently not observed because a
whole (almost) page of essentially scratch space absorbs any excessive
prefetching by CDMA. However, changing the number of slots in the push
buffer can trigger these SMMU faults.
Signed-off-by: Thierry Reding <treding@nvidia.com>
On Tegra186 and later, the ARM SMMU provides an input address space that
is 48 bits wide. However, memory clients can only address up to 40 bits.
If the geometry is used as-is, allocations of IOVA space can end up in a
region that is not addressable by the memory clients.
To fix this, restrict the IOVA space to the DMA mask of the host1x
device.
Signed-off-by: Thierry Reding <treding@nvidia.com>
Tegra186 and later support 40 bits of address space. Additional
registers need to be programmed to store the full 40 bits of push
buffer addresses.
Since command stream gathers can also reside in buffers in a 40-bit
address space, a new variant of the GATHER opcode is also introduced.
It takes two parameters: the first parameter contains the lower 32
bits of the address and the second parameter contains bits 32 to 39.
Signed-off-by: Thierry Reding <treding@nvidia.com>
The CDMA push buffer can currently only handle opcodes that take a
single word parameter. However, the host1x implementation on Tegra186
and later supports opcodes that require multiple words as parameters.
Unfortunately the way the push buffer is structured, these wide opcodes
cannot simply be composed of two regular opcodes because that could
result in the wide opcode being split across the end of the push buffer
and the final RESTART opcode required to wrap the push buffer around
would break the wide opcode.
One way to fix this would be to remove the concept of slots to simplify
push buffer operations. However, that's not entirely trivial and should
be done in a separate patch. For now, simply use a different function
to push four-word opcodes into the push buffer. Technically only three
words are pushed, with the fourth word used as padding to preserve the
2-word alignment required by the slots abstraction. The fourth word is
always a NOP opcode.
Additional care must be taken when the end of the push buffer is
reached. If a four-word opcode doesn't fit into the push buffer without
being split by the boundary, NOP opcodes will be introduced and the new
wide opcode placed at the beginning of the push buffer.
Signed-off-by: Thierry Reding <treding@nvidia.com>
When processing command streams, make sure the host1x's stream ID is
programmed for the channel so that addresses are properly translated
through the SMMU.
Signed-off-by: Thierry Reding <treding@nvidia.com>
