Jane Jian [Wed, 18 Dec 2019 10:53:46 +0000 (18:53 +0800)]
drm/amdgpu: disable VCN2.5 ib test for Arcturus sriov
currently using TMR loading VCN fw MMSCH would fail
to init after FLR, just disable ib test for temporarily
daily testing, continuing debug with mm team.
Jonathan Kim [Thu, 12 Dec 2019 16:46:05 +0000 (11:46 -0500)]
drm/amdgpu: add perfmons accessible during df c-states
During DF C-State, Perfmon counters outside of range 1D700-1D7FF will
encounter SLVERR affecting xGMI performance monitoring. PerfmonCtr[7:4]
is being added to avoid SLVERR during read since it falls within this
range. PerfmonCtl[7:4] is being added in order to arm PerfmonCtr[7:4].
Since PerfmonCtl[7:4] exists outside of range 1D700-1D7FF, DF routines
will be enabled to opportunistically re-arm PerfmonCtl[7:4] on retry
after SLVERR.
Huang Rui [Mon, 16 Dec 2019 07:02:50 +0000 (15:02 +0800)]
drm/amdkfd: expose num_sdma_queues_per_engine data field to topology node (v2)
Thunk driver would like to know the num_sdma_queues_per_engine data, however
this data relied on different asic specific. So it's better to get it from kfd
driver.
Jane Jian [Mon, 16 Dec 2019 09:04:01 +0000 (17:04 +0800)]
drm/amdgpu: enable VCN0 and VCN1 sriov instances support for Arcturus
v1: compared to bare-metal: sriov support psp loading VCN firmware; only one
encoding ring would be used in each instance.
v2: keep unchange for bare-metal VCN2.5 hw_init, just add a flag with sriov
and also remove multiple lines.
v3: squash in warning fix
Jane Jian [Mon, 16 Dec 2019 08:24:13 +0000 (16:24 +0800)]
drm/amdgpu: skip VCN2.5 power gating and clock gating for sriov Arcturus
v1: skip gating in serveral called functions by power gating and clock gating
v2: from suggestion, skip setting gate in both set function, which is where
it being called.
Alex Deucher [Tue, 10 Dec 2019 21:21:44 +0000 (16:21 -0500)]
drm/amdgpu: wait for all rings to drain before runtime suspending
Add a safety check to runtime suspend to make sure all outstanding
fences have signaled before we suspend. Doesn't fix any known issue.
We already do this via the fence driver suspend function, but we
just force completion rather than bailing. This bails on runtime
suspend so we can try again later once the fences are signaled to
avoid missing any outstanding work.
drm/amdgpu: Switch from system_highpri_wq to system_unbound_wq
This is to avoid queueing jobs to same CPU during XGMI hive reset
because there is a strict timeline for when the reset commands
must reach all the GPUs in the hive.
It is used to synchronize N threads at a rendevouz point before execution
of critical code that has to be started by all the threads at approximatly
the same time.
v2: Remove mention of reset use case, improve doc.
Monk Liu [Tue, 17 Dec 2019 10:16:44 +0000 (18:16 +0800)]
drm/amdgpu: fix KIQ ring test fail in TDR of SRIOV
issues:
MEC is ruined by the amdkfd_pre_reset after VF FLR done
fix:
amdkfd_pre_reset() would ruin MEC after hypervisor finished the VF FLR,
the correct sequence is do amdkfd_pre_reset before VF FLR but there is
a limitation to block this sequence:
if we do pre_reset() before VF FLR, it would go KIQ way to do register
access and stuck there, because KIQ probably won't work by that time
(e.g. you already made GFX hang)
Monk Liu [Tue, 17 Dec 2019 10:18:31 +0000 (18:18 +0800)]
drm/amdgpu: fix double gpu_recovery for NV of SRIOV
issues:
gpu_recover() is re-entered by the mailbox interrupt
handler mxgpu_nv.c
fix:
we need to bypass the gpu_recover() invoke in mailbox
interrupt as long as the timeout is not infinite (thus the TDR
will be triggered automatically after time out, no need to invoke
gpu_recover() through mailbox interrupt.
Yintian Tao [Tue, 17 Dec 2019 03:43:40 +0000 (11:43 +0800)]
drm/amd/powerplay: skip soc clk setting under pp one vf
Under sriov pp one vf mode, there is no need to set
soc clk under pp one vf because smu firmware will depend
on the mclk to set the appropriate soc clk for it.
Nirmoy Das [Fri, 6 Dec 2019 15:55:49 +0000 (16:55 +0100)]
drm/amdgpu: replace vm_pte's run-queue list with drm gpu scheds list
drm_sched_entity_init() takes drm gpu scheduler list instead of
drm_sched_rq list. This makes conversion of drm_sched_rq list
to drm gpu scheduler list unnecessary
Nirmoy Das [Thu, 5 Dec 2019 10:38:00 +0000 (11:38 +0100)]
drm/scheduler: rework entity creation
Entity currently keeps a copy of run_queue list and modify it in
drm_sched_entity_set_priority(). Entities shouldn't modify run_queue
list. Use drm_gpu_scheduler list instead of drm_sched_rq list
in drm_sched_entity struct. In this way we can select a runqueue based
on entity/ctx's priority for a drm scheduler.
Alex Deucher [Thu, 12 Dec 2019 22:43:36 +0000 (17:43 -0500)]
drm/amdgpu/pm_runtime: update usage count in fence handling
Increment the usage count in emit fence, and decrement in
process fence to make sure the GPU is always considered in
use while there are fences outstanding. We always wait for
the engines to drain in runtime suspend, but in practice
that only covers short lived jobs for gfx. This should
cover us for longer lived fences.
Zhan Liu [Mon, 16 Dec 2019 19:56:50 +0000 (14:56 -0500)]
drm/amd/powerplay: Add SMU WMTABLE Validity Check for Renoir
[Why]
SMU watermark table (WMTABLE) validity check is missing on Renoir.
This validity check is very useful for checking whether
WMTABLE is updated successfully.
Nikola Cornij [Tue, 3 Dec 2019 22:01:12 +0000 (17:01 -0500)]
drm/amd/display: Add debug option to override DSC target bpp increment
[why]
It's required for debug purposes.
[how]
Add a dsc_bpp_increment_div debug option that overrides DPCD
BITS_PER_PIXEL_INCREMENT value. The value dsc_bpp_increment_div should
be set to is the one after parsing, i.e. it could be 1, 2, 4, 8 or 16
(meaning 1pix, 1/2pix, ..., 1/16pix).
Samson Tam [Wed, 4 Dec 2019 23:35:15 +0000 (18:35 -0500)]
drm/amd/display: fix missing cursor on some rotated SLS displays
[Why]
Cursor disappears for some SLS displays that are rotated 180
and 270 degrees. This occurs when there is no pipe split being
done ( ex. 3 or more displays ). The cursor calculations assume
pipe splitting is done so when it calculates the new cursor
position in hwss.set_cursor_position(), it is out-of-bounds so
it disables the cursor in hubp.set_cursor_position().
[How]
In non pipe split cases, calculate cursor using viewport size
( width or height ) instead of viewport size * 2 ( the two
because pipe splitting divides the rectangle into two ).
Sung Lee [Wed, 4 Dec 2019 23:36:07 +0000 (18:36 -0500)]
drm/amd/display: Lower DPP DTO only when safe
[Why]
A corner case currently exists where DPP DTO is lowered before
pipes are updated to a higher viewport. This causes underflow
as the DPPCLK is too low for the current viewport.
[How]
Only lower DPP DTO when it is safe to lower, or if
the newer clocks are higher than the current ones.
Jun Lei [Mon, 25 Nov 2019 15:58:44 +0000 (10:58 -0500)]
drm/amd/display: support virtual DCN
[why]
DAL3 should support SRIOV
[how]
Add support for the virtual dal flag. This flag should skip
most/all of DC construction since the HW isn't accessible, but
still construct WindowsDM (almost) normally but with only SW display
targets
Sung Lee [Mon, 2 Dec 2019 21:45:16 +0000 (16:45 -0500)]
drm/amd/display: Fix update_bw_bounding_box Calcs
[Why]
Previously update_bw_bounding_box for RN was commented out
due to incorrect values causing BSOD on Hybrid Graphics.
However, commenting out this function also may cause issues
such as underflow in certain cases such as 2x4K displays.
[How]
Fix dram_speed_mts calculations.
Update from proper index of clock_limits[]
Samson Tam [Thu, 28 Nov 2019 20:55:01 +0000 (15:55 -0500)]
drm/amd/display: fix 270 degree rotation for mixed-SLS mode
[Why]
When we rotate 270 in mixed SLS mode, the recouts occupy the
right side of the display. So all the recout_skip_v values
are relative to the left side of the display. This causes
adjust_vp_and_init_for_seamless_clip() to incorrectly increase
the data->viewport.height for that recout. The rotation looks
like the bottom half is duplicated twice.
[How]
recout.x values are being adjusted based on
stream->timing.h_border_left. Instead of using h_border_left,
use dst.x to represent the border. Shift dst.x by the amount of
stream->timing.h_border_left and set
stream->timing.h_border_left to 0. Do all the calculations
and then revert stream->timing.h_border_left and
stream->dst.x back to their original values.
When calculating pipe_ctx->plane_res.scl_data.h_active,
make sure to use the original stream->timing.h_border_left
value.
Noah Abradjian [Fri, 29 Nov 2019 18:48:36 +0000 (13:48 -0500)]
drm/amd/display: Remove reliance on pipe indexing
[Why]
In certain instances, there was a reliance on pipe indexing being accurate. However, this
assumption fails with harvesting of pipes 1 or 2, which can occur in production B6 parts.
HW hang would occur as a result.
[How]
Use hubp index for mpcc, and do mpc_init for all theoretical pipes (including disabled ones).
Paul Hsieh [Thu, 28 Nov 2019 02:44:39 +0000 (10:44 +0800)]
drm/amd/display: check link status before disable stream
[Why]
1. Set second screen only then unplug external monitor
2. Enter to S4 then plug in external monitor
3. Resume from S4, eDP will not turn off when OS set
second screen only
Sometimes OS will not set eDP power up cause eDP dpms_off
keep true then driver skipp disable stream
[How]
When drvier try to disable stream, add link status condition
Nikola Cornij [Tue, 26 Nov 2019 20:18:31 +0000 (15:18 -0500)]
drm/amd/display: Map ODM memory correctly when doing ODM combine
[why]
Up to 4 ODM memory pieces are required per ODM combine and cannot
overlap, i.e. each ODM "session" has to use its own memory pieces.
The ODM-memory mapping is currently broken for generic case.
The maximum number of memory pieces is ASIC-dependent, but it's always
big enough to satisfy maximum number of ODM combines. Memory pieces
are mapped as a bit-map, i.e. one memory piece corresponds to one bit.
The OPTC doing ODM needs to select memory pieces by setting the
corresponding bits, making sure there's no overlap with other OPTC
instances that might be doing ODM.
The current mapping works only for OPTC instance indexes smaller than
3. For instance indexes 3 and up it practically maps no ODM memory,
causing black, gray or white screen in display configs that include
ODM on OPTC instance 3 or up.
[how]
Statically map two unique ODM memory pieces for each OPTC instance
and piece them together when programming ODM combine mode.
Charlene Liu [Thu, 21 Nov 2019 02:23:47 +0000 (21:23 -0500)]
drm/amd/display: Add warmup escape call support
Add warmup escape support, for diags, in a way that is possible to
choose a new or an existing sequence. For achieving this goal, this
commit adds separated MCIF buffer as VCN request.
Josip Pavic [Tue, 26 Nov 2019 17:26:14 +0000 (12:26 -0500)]
drm/amd/display: fix regamma build optimization
[Why]
When the global variable pow_buffer_ptr is set to -1, by definition
optimizations should not be used to build the regamma. Since
translate_from_linear_space unconditionally increments this global, it
inadvertently enables the optimization.
[How]
Increment pow_buffer_ptr only if it is not -1.
Aric Cyr [Sat, 23 Nov 2019 22:15:51 +0000 (17:15 -0500)]
drm/amd/display: Remove integer scaling code from DC and fix cursor
[Why]
Scaling better handled by upper layers before pipe splitting.
[How]
Remove DC code for integer scaling and force cursor update if
viewport or scaling changes occur to prevent underflow from
invalid cursor position.
Martin Leung [Fri, 22 Nov 2019 00:13:54 +0000 (19:13 -0500)]
drm/amd/display: Enable Seamless Boot Transition for Multiple Streams
[why]
dc previously had bugs that interfered with the ability to inherit a
timing from a device with multiple streams (without flash/blanking).
After this fix there is still a dependency on UEFI support.
[how]
fixed 3 bugs: loaded MPC state, changed bw_optimize flag to a counter
instead of a boolean, and reading dpp/disp clk from HW to ensure we
don't raise the clock's when we're not supposed to.
Noah Abradjian [Fri, 22 Nov 2019 21:07:24 +0000 (16:07 -0500)]
drm/amd/display: Collapse resource arrays when pipe is disabled
[Why]
Currently, pipe resources are assigned to an index that matches the pipe position.
However, if pipe 1 or 2 is disabled, there will be a gap in the arrays which causes a crash when iterating based on pipe_count.
[How]
Fix resource construct to assign resources to minimum available array index.
Noah Abradjian [Fri, 22 Nov 2019 16:47:52 +0000 (11:47 -0500)]
drm/amd/display: Use pipe_count for num of opps
[Why]
There is one opp per pipe. For certain RN parts, the fourth pipe is disabled, so there is no opp for it.
res_cap->num_opp is hardcoded to 4, so if we use that to iterate over opps we will crash.
[How]
Use the pipe_count value instead, which is not hardcoded and so will have the correct number.
Amanda Liu [Thu, 21 Nov 2019 21:06:57 +0000 (16:06 -0500)]
drm/amd/display: Reinstate LFC optimization
[why]
We want to streamline the calculations made when entering LFC.
Previously, the optimizations led to screen tearing and were backed out
to unblock development.
[how]
Integrate other calculations parameters, as well as screen tearing,
fixes with the original LFC calculation optimizations.
Noah Abradjian [Mon, 18 Nov 2019 18:59:57 +0000 (13:59 -0500)]
drm/amd/display: Add wait for flip not pending on pipe unlock
[Why]
Lack of proper timing caused intermittent underflow on unplug external DP.
A previous fix was invalid and caused S0i3 regression, so had to be reverted.
[How]
When unlocking pipe, wait for no pipes to have flip pending before unlocking.
Leo Liu [Thu, 12 Dec 2019 15:28:02 +0000 (10:28 -0500)]
drm/amdgpu/vcn1.0: use its own idle handler and begin use funcs
Because VCN1.0 power management and DPG mode are managed together with
JPEG1.0 under both HW and FW, so separated them from general VCN code.
Also the multiple instances case got removed, since VCN1.0 HW just have
a single instance.
changzhu [Thu, 12 Dec 2019 05:46:06 +0000 (13:46 +0800)]
drm/amdgpu: enable gfxoff for raven1 refresh
When smu version is larger than 0x41e2b, it will load
raven_kicker_rlc.bin.To enable gfxoff for raven_kicker_rlc.bin,it
needs to avoid adev->pm.pp_feature &= ~PP_GFXOFF_MASK when it loads
raven_kicker_rlc.bin.
Roman Li [Wed, 27 Nov 2019 19:47:32 +0000 (14:47 -0500)]
drm/amd/display: add missing dcn link encoder regs
[Why]
The earlier change: "check phy dpalt lane count config"
uses link encoder registers not defined properly.
That caused regression with mst-enabled display not
lighting up.
Colin Ian King [Thu, 12 Dec 2019 18:16:57 +0000 (18:16 +0000)]
drm/amd/powerplay: fix various dereferences of a pointer before it is null checked
There are several occurrances of the pointer hwmgr being dereferenced
before it is null checked. Fix these by performing the dereference
of hwmgr after it has been null checked.
Addresses-Coverity: ("Dereference before null check") Fixes: c9ffa427db34e6 ("drm/amd/powerplay: enable pp one vf mode for vega10") Signed-off-by: Colin Ian King <[email protected]> Signed-off-by: Alex Deucher <[email protected]> Signed-off-by: Alex Deucher <[email protected]>
Philip Yang [Wed, 30 Jan 2019 20:29:34 +0000 (15:29 -0500)]
drm/amdkfd: queue kfd interrupt work to different CPU
Because queue_work schedule the work on the same CPU the interrupt
handler is running, if there are many interrupts pending, it takes
longer time for work queue to start, or even worse system will hang.
v2: queue work to same NUMA node for better cache locality
v3: handle cpumask_next wraparound case
Timothy Pearson [Sat, 7 Dec 2019 22:48:09 +0000 (16:48 -0600)]
amdgpu: Wrap FPU dependent functions in dc20
dc20 containes several FPU-dependent functions without proper FPU
kernel mode enable/disable wrappers. Add the required wrappers
for both x86 and POWER.
projects / linux.git / log