drm/amdgpu: validate wait_for_idle before function call

Before making a function call to wait_for_idle,
validate the function pointer like we do in sw_init.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: validate resume before function call

Before making a function call to resume, validate
the function pointer like we do in sw_init.

Use the helper function amdgpu_ip_block_resume where
same checks and calls are repeated.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: validate suspend before function call

Before making a function call to suspend, validate
the function pointer like we do in sw_init.

Use the helper function amdgpu_ip_block_suspend where
same checks and calls are repeated.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: validate hw_fini before function call

Before making a function call to hw_fini, validate
the function pointer like we do in sw_init.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdkfd: fix the hang caused by the write reorder to fence_addr

make sure KFD_FENCE_INIT write to fence_addr before pm_send_query_status
called, to avoid qcm fence timeout caused by incorrect ordering.

Signed-off-by: Victor Zhao <[email protected]>
Reviewed-by: Philip Yang <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/gfx9: Add cleaner shader for GFX9.4.2

This commit adds the cleaner shader microcode for GFX9.4.2 GPUs. The
cleaner shader is a piece of GPU code that is used to clear or
initialize certain GPU resources, such as Local Data Share (LDS), Vector
General Purpose Registers (VGPRs), and Scalar General Purpose Registers
(SGPRs).

Clearing these resources is important for ensuring data isolation
between different workloads running on the GPU. Without the cleaner
shader, residual data from a previous workload could potentially be
accessed by a subsequent workload, leading to data leaks and incorrect
computation results.

The cleaner shader microcode is represented as an array of 32-bit words
(`gfx_9_4_2_cleaner_shader_hex`). This array is the binary
representation of the cleaner shader code, which is written in a
low-level GPU instruction set.

Also, this patch updates the `gfx_v9_0_sw_init` function to initialize
the cleaner shader if the MEC firmware version is 88 or higher. It sets
the `cleaner_shader_ptr` and `cleaner_shader_size` to the appropriate
values and attempts to initialize the cleaner shader.

When the cleaner shader feature is enabled, the AMDGPU driver loads this
array into a specific location in the GPU memory. The GPU then reads
this memory location to fetch and execute the cleaner shader
instructions.

The cleaner shader is executed automatically by the GPU at the end of
each workload, before the next workload starts. This ensures that all
GPU resources are in a clean state before the start of each workload.

This change ensures that the GPU memory is properly cleared between
different processes, preventing data leakage and enhancing security. It
also aligns with the serialization mechanism between KGD and KFD,
ensuring that the GPU state is consistent across different workloads.

Cc: Christian König <[email protected]>
Cc: Alex Deucher <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Suggested-by: Alex Deucher <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: fix typo for sdma6 constant fill packet

Fix typo for sdma6 constant fill packet

Signed-off-by: Frank Min <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: fix random data corruption for sdma 7

There is random data corruption caused by const fill, this is caused by
write compression mode not correctly configured.

So correct compression mode for const fill.

Signed-off-by: Frank Min <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: 3.2.306

This version brings along following fixes:
- Fix dcn401 idle optimization problem
- Fix cursor corruption on dcn35
- Fix DP LL compliance failures
- Fix SubVP Phantom VBlank End calculation

Acked-by: Tom Chung <[email protected]>
Signed-off-by: Aric Cyr <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: To change dcn301_init.h guard.

[why & How]
The original guard is wrongly to be set as for dcn30.
Changed it from 30 to 301.

Reviewed-by: Dillon Varone <[email protected]>
Signed-off-by: Bhuvanachandra Pinninti <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: update fullscreen status to SPL

[Why]
Current fullscreen check in SPL using dm_helpers is out-of-sync
with dc state. This causes an issue during minimal transition
where we pick an invalid intermediate state because the pre and
post fullscreen status are different.

[How]
Add sharpening_required flag to dc_stream_state. Use this flag to
indicate if we are in fullscreen or not. Propagate flag to SPL for
fullscreen status. Remove workaround in DML

Reviewed-by: Alvin Lee <[email protected]>
Signed-off-by: Samson Tam <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Add a Precise Delay Routine

Fix DP compliance failures 4.2.2.12, 4.3.1.21, 4.9.1.19
caused by imprecise delay on fsleep().

Reviewed-by: Aric Cyr <[email protected]>
Signed-off-by: Fangzhi Zuo <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Recalculate SubVP Phantom VBlank End in dml21

[WHY]
The phantom stream timing is copied from the main stream as most
parameters are identical, however some need to be recalculated.
Currently VBlank End is not recalculated and copied from the main
incorrectly.

[HOW]
Recalculate VBlank End for phantom stream timing.

Reviewed-by: Alvin Lee <[email protected]>
Signed-off-by: Dillon Varone <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: temp w/a for DP Link Layer compliance

[Why&How]
Disabling P-State support on full updates for DCN401 results in
introducing additional communication with SMU. A UCLK hard min message
to SMU takes 4 seconds to go through, which was due to DCN not allowing
pstate switch, which was caused by incorrect value for TTU watermark
before blanking the HUBP prior to DPG on for servicing the test request.

Fix the issue temporarily by disallowing pstate changes for compliance
test while test request handler is reworked for a proper fix.

Fixes: 67ea53a4bd9d ("drm/amd/display: Disable DCN401 UCLK P-State support on full updates")
Cc: Mario Limonciello <[email protected]>
Cc: Alex Deucher <[email protected]>
Reviewed-by: Dillon Varone <[email protected]>
Signed-off-by: Aurabindo Pillai <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Adding array index check to prevent memory corruption

[Why & How]
Array indices out of bound caused memory corruption. Adding checks to
ensure that array index stays in bound.

Reviewed-by: Charlene Liu <[email protected]>
Reviewed-by: Nicholas Kazlauskas <[email protected]>
Signed-off-by: Leo Chen <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Reuse subvp enable check for DCN401

Reuse subvp enable check from DCN32 for IGT testing of Sub-Viewport
feature on DCN401

Reviewed-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Aurabindo Pillai <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: w/a to program DISPCLK_R_GATE_DISABLE DCN35

[WHY & HOW]
Cursor corruption observed on USBC display with specific system setup with a
reboot. Cursor memory might still in the lightsleep state due to voltage
issue, we need program DISPCLK_R_GATE_DISABLE to avoid this issue only on
DCN35.

Reviewed-by: Nicholas Kazlauskas <[email protected]>
Signed-off-by: Yihan Zhu <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: temp w/a for dGPU to enter idle optimizations

[Why&How]
vblank immediate disable currently does not work for all asics. On
DCN401, the vblank interrupts never stop coming, and hence we never
get a chance to trigger idle optimizations.

Add a workaround to enable immediate disable only on APUs for now. This
adds a 2-frame delay for triggering idle optimization, which is a
negligible overhead.

Fixes: 58a261bfc967 ("drm/amd/display: use a more lax vblank enable policy for older ASICs")
Fixes: e45b6716de4b ("drm/amd/display: use a more lax vblank enable policy for DCN35+")
Cc: Mario Limonciello <[email protected]>
Cc: Alex Deucher <[email protected]>
Reviewed-by: Harry Wentland <[email protected]>
Reviewed-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Aurabindo Pillai <[email protected]>
Signed-off-by: Wayne Lin <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: clean the dummy sw_fini functions

Remove the dummy sw_fini functions for all
ip blocks.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Add hpd_source index check for dcn401 link encoder setup

This patch adds a boundary check for the hpd_source index during the
link encoder creation process for all dcn401 ip. The check ensures that the
index is within the valid range of the link_enc_hpd_regs array to
prevent out-of-bounds access.

Cc: Tom Chung <[email protected]>
Cc: Rodrigo Siqueira <[email protected]>
Cc: Roman Li <[email protected]>
Cc: Alex Hung <[email protected]>
Cc: Aurabindo Pillai <[email protected]>
Cc: Harry Wentland <[email protected]>
Cc: Hamza Mahfooz <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Roman Li <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Add hpd_source index check for dcn10 link encoder setup

This patch adds a boundary check for the hpd_source index during the
link encoder creation process for all dcn10 ip. The check ensures that the
index is within the valid range of the link_enc_hpd_regs array to
prevent out-of-bounds access.

Cc: Tom Chung <[email protected]>
Cc: Rodrigo Siqueira <[email protected]>
Cc: Roman Li <[email protected]>
Cc: Alex Hung <[email protected]>
Cc: Aurabindo Pillai <[email protected]>
Cc: Harry Wentland <[email protected]>
Cc: Hamza Mahfooz <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Roman Li <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Add hpd_source index check for DCE60/80/100/110/112/120 link encoders

This patch adds a boundary check for the hpd_source index during the
link encoder creation process for all DCExxx IP's. The check ensures
that the index is within the valid range of the link_enc_hpd_regs array
to prevent out-of-bounds access.

Cc: Tom Chung <[email protected]>
Cc: Rodrigo Siqueira <[email protected]>
Cc: Roman Li <[email protected]>
Cc: Alex Hung <[email protected]>
Cc: Aurabindo Pillai <[email protected]>
Cc: Harry Wentland <[email protected]>
Cc: Hamza Mahfooz <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Roman Li <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Use SPX as default in partition config

In certain cases - ex: when a reset is required on initialization - XCP
manager won't have a valid partition mode. In such cases, use SPX as the
default selected mode for which partition configuration details are
populated.

Fixes: 4ae86dc87850 ("drm/amdgpu: Add sysfs nodes to get xcp details")
Signed-off-by: Lijo Lazar <[email protected]>
Reported-by: Hao Zhou <[email protected]>
Reviewed-by: Asad Kamal <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: validate sw_fini before function call

Before making a function call to sw_fini, validate
the function pointer like we do in sw_init.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: clean the dummy sw_init functions

Remove the dummy sw_init functions for all
IP blocks.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: validate sw_init before function call

Before making a function call to sw_init, validate
the function pointer like we do in late_init.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdkfd: Not restore userptr buffer if kfd process has been removed

When kfd process has been terminated not restore userptr buffer after mmu
notifier invalidates a range.

Signed-off-by: Xiaogang Chen <[email protected]>
Reviewed-by: Felix Kuehling <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/pm: update deep sleep status on smu v14.0.2/3

disable deep sleep during the compute workload for the
potential performance loss on smu v14.0.2/3

Signed-off-by: Kenneth Feng <[email protected]>
Reviewed-by: Lijo Lazar <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/pm: update overdrive function on smu v14.0.2/3

update overdrive function on smu v14.0.2/3

Signed-off-by: Kenneth Feng <[email protected]>
Acked-by: Yang Wang <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Zero-initialize mqd backup memory

Zero-initialize mqd backup memory, otherwise the check for
'already-backed-up' could go wrong.

Signed-off-by: Lijo Lazar <[email protected]>
Reviewed-by: Yang Wang <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/pm: update the driver-fw interface file for smu v14.0.2/3

update the driver-fw interface file for smu v14.0.2/3

Signed-off-by: Kenneth Feng <[email protected]>
Reviewed-by: Yang Wang <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Ensure HPD source index is valid for dcn20/dcn201 link encoders

This patch adds a boundary check for the hpd_source index during the
link encoder creation process for dcn20/dcn201 IP's. The check ensures
that the index is within the valid range of the link_enc_hpd_regs array
to prevent out-of-bounds access.

Cc: Tom Chung <[email protected]>
Cc: Rodrigo Siqueira <[email protected]>
Cc: Roman Li <[email protected]>
Cc: Alex Hung <[email protected]>
Cc: Aurabindo Pillai <[email protected]>
Cc: Harry Wentland <[email protected]>
Cc: Hamza Mahfooz <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Roman Li <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Fix spelling mistake "tunndeling" -> "tunneling"

There is a spelling mistake in a dm_error message. Fix it.

Signed-off-by: Colin Ian King <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

Revert "drm/amdgpu/gfx9: put queue resets behind a debug option"

This reverts commit 7c1a2d8aba6cadde0cc542b2d805edc0be667e79.

Extended validation has completed successfully, so enable
these features by default.

Acked-by: Jiadong Zhu <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>
Cc: Jonathan Kim <[email protected]>
Cc: Jiadong Zhu <[email protected]>

drm/amdgpu: init saw registers for mmhub v1.0

This commits init registers in the Stand Along Walker
for mmhub v1.0, to support ISP use cases.

Signed-off-by: Zhu Lingshan <[email protected]>
Reported-and-tested-by: Du Bin <[email protected]>
Acked-by: Alex Deucher <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/discovery: add ISP discovery entries for old APUs

Raven1/2 and Picasso have ISP 2.0.0, however their ISP blocks
are not in the IP discovery table yet.

This commit fixes this issue by adding new ISP entries for
Raven and Picasso in the IP discovery table.

Signed-off-by: Alex Deucher <[email protected]>
Signed-off-by: Zhu Lingshan <[email protected]>
Acked-by: Alex Deucher <[email protected]>

drm/amd: Guard against bad data for ATIF ACPI method

If a BIOS provides bad data in response to an ATIF method call
this causes a NULL pointer dereference in the caller.

```
? show_regs (arch/x86/kernel/dumpstack.c:478 (discriminator 1))
? __die (arch/x86/kernel/dumpstack.c:423 arch/x86/kernel/dumpstack.c:434)
? page_fault_oops (arch/x86/mm/fault.c:544 (discriminator 2) arch/x86/mm/fault.c:705 (discriminator 2))
? do_user_addr_fault (arch/x86/mm/fault.c:440 (discriminator 1) arch/x86/mm/fault.c:1232 (discriminator 1))
? acpi_ut_update_object_reference (drivers/acpi/acpica/utdelete.c:642)
? exc_page_fault (arch/x86/mm/fault.c:1542)
? asm_exc_page_fault (./arch/x86/include/asm/idtentry.h:623)
? amdgpu_atif_query_backlight_caps.constprop.0 (drivers/gpu/drm/amd/amdgpu/amdgpu_acpi.c:387 (discriminator 2)) amdgpu
? amdgpu_atif_query_backlight_caps.constprop.0 (drivers/gpu/drm/amd/amdgpu/amdgpu_acpi.c:386 (discriminator 1)) amdgpu
```

It has been encountered on at least one system, so guard for it.

Fixes: d38ceaf99ed0 ("drm/amdgpu: add core driver (v4)")
Acked-by: Alex Deucher <[email protected]>
Signed-off-by: Mario Limonciello <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/swsmu: add automatic parameter to set_soft_freq_range

On chips that support it, you can specificy 0 and 0xffff for
min and max and the PMFW will use that to determine the optimal
min and max.  This enables optimal performance when the
user manually switches between performance levels using sysfs.
Previously we'd set soft min/max which could limit performance.

Reviewed-by: Kenneth Feng <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Fix off by one in current_memory_partition_show()

The >= ARRAY_SIZE() should be > ARRAY_SIZE() to prevent an out of
bounds read.

Fixes: 012be6f22c01 ("drm/amdgpu: Add sysfs interfaces for NPS mode")
Reviewed-by: Lijo Lazar <[email protected]>
Signed-off-by: Dan Carpenter <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/swsmu: default to fullscreen 3D profile for dGPUs

This uses more aggressive hueristics than the the bootup default
profile.  On windows the OS has a special fullscreen 3D mode
where this is used.  Since we don't have the equivalent on Linux
default to this profile for dGPUs.

Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/3618
Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/1500
Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3131
Fixes: c50fe289ed72 ("drm/amdgpu/swsmu: always force a state reprogram on init")
Reviewed-by: Kenneth Feng <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/swsmu: Only force workload setup on init

Needed to set the workload type at init time so that
we can apply the navi3x margin optimization.

Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/3618
Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3131
Fixes: c50fe289ed72 ("drm/amdgpu/swsmu: always force a state reprogram on init")
Reviewed-by: Kenneth Feng <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/radeon: Fix encoder->possible_clones

Include the encoder itself in its possible_clones bitmask.
In the past nothing validated that drivers were populating
possible_clones correctly, but that changed in commit
74d2aacbe840 ("drm: Validate encoder->possible_clones").
Looks like radeon never got the memo and is still not
following the rules 100% correctly.

This results in some warnings during driver initialization:
Bogus possible_clones: [ENCODER:46:TV-46] possible_clones=0x4 (full encoder mask=0x7)
WARNING: CPU: 0 PID: 170 at drivers/gpu/drm/drm_mode_config.c:615 drm_mode_config_validate+0x113/0x39c
...

Cc: Alex Deucher <[email protected]>
Cc: [email protected]
Fixes: 74d2aacbe840 ("drm: Validate encoder->possible_clones")
Reported-by: Erhard Furtner <[email protected]>
Closes: https://lore.kernel.org/dri-devel/20241009000321.418e4294@yea/
Tested-by: Erhard Furtner <[email protected]>
Signed-off-by: Ville Syrjälä <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/smu13: always apply the powersave optimization

It can avoid margin issues in some very demanding applications.

Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/3618
Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3131
Fixes: c50fe289ed72 ("drm/amdgpu/swsmu: always force a state reprogram on init")
Reviewed-by: Kenneth Feng <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: 3.2.305

- Add sharpening policy to plane state
- Clear pipe pointers on pipe reset
- Resolve correct MALL size for dcn401
- Read Sink emission rate capability
- IPX fixes
- Coverity fixes

Reviewed-by: Roman Li <[email protected]>
Signed-off-by: Aric Cyr <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: disable dynamic ODM when sharpness is enabled

[Why & How]
Disable dynamic ODM when sharpness is enabled

Reviewed-by: Alvin Lee <[email protected]>
Signed-off-by: Samson Tam <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Check returned status from core_link_read_dpcd

[WHAT]
The function core_link_read_dpcd returns status which is not used at
all, making them useless assignments.

[HOW]
Print error messages if core_link_read_dpcd does not return DC_OK.

This fixes 2 UNUSED_VALUE issues reported by Coverity.

Reviewed-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Hung <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Clear pipe pointers on pipe reset

[Why]
We want to clean up unnecessary asserts, one of which is an assert in
resource_is_pipe_type that fires if a pipe has no stream and still has
pointers to other pipes ("dangling state"). This gets hit because pipes
are not properly cleaned up in reset_back_end_for_pipe. When resetting a
pipe, the existing MPCC / ODM combine pointers are no longer valid,
especially when we put ODM in bypass.

[How]
 - reset pipe pointers in reset_back_end_for_pipe
 - remove useless code to avoid confusion
     (a long time ago it had a reason to be there, not anymore)

Reviewed-by: Alvin Lee <[email protected]>
Signed-off-by: Joshua Aberback <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: add sharpening policy to plane state

[Why]
Pass in sharpening policy through plane state from control side

[How]
Add sharpener support through dc_caps.
Add sharpen policy to plane state and move to spl_input.
Pass sharpen policy from plane state to SPL.

Reviewed-by: Aric Cyr <[email protected]>
Signed-off-by: Samson Tam <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: resolve correct MALL size for dcn401

[WHY]
Code for dcn401 to calculate available MALL size for display was shared
with dcn32 and did not provide the correct result for all ASICs.

[HOW]
Add dcn401 specific function to properly calculate the available MALL
for display.

Reviewed-by: Chris Park <[email protected]>
Signed-off-by: Dillon Varone <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Read Sink emission rate capability

[WHY]
To get sink emission rate information for future
supported refresh rate calculation.

Reviewed-by: ChunTao Tso <[email protected]>
Signed-off-by: Robin Chen <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: print messages when programming shaper/3dlut fails

[WHAT & HOW]
Print error messages when programming shaper lut or 3dlut fails.

This fixes 5 UNUSED_VALUE issues reported by Coverity.

Reviewed-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Hung <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Full exit out of IPS2 when all allow signals have been cleared

[Why]
A race condition occurs between cursor movement and vertical interrupt control
thread from OS, with both threads trying to exit IPS2.
Vertical interrupt control thread clears the prev driver allow signal while not fully
finishing the IPS2 exit process.

[How]
We want to detect all the allow signals have been cleared before we perform the full exit.

Reviewed-by: Nicholas Kazlauskas <[email protected]>
Signed-off-by: Leo Chen <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Fix Coverity change for visual confirm

[Why && How]
Previous change for Coverity has caused regression on visual confirm
so fix it by reverting the part that affects visual confirm.

Reviewed-by: Chris Park <[email protected]>
Signed-off-by: Leo (Hanghong) Ma <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Add 3DLUT FL HW bug workaround

[Why]
There is a known HW bug that causes the internal 3DLUT fetch signal to
be lost at VREADY, regardless of whether the OTG lock is being held or
not. A workaround is necessary to make sure that this internal signal
stays up after OTG unlock.

[How]
Set the 3DLUT_ENABLE bit immediately before and after the unlock. Also
use VUPDATE_KEEPOUT to prevent lock transition in the region between
VSTARTUP and VREADY, which could cause issues with this WA sequence.

Also including misc. 3DLUT DMA-related sequence fixes to address a few
regressions causing corruption.

Reviewed-by: Dillon Varone <[email protected]>
Signed-off-by: Ilya Bakoulin <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Add check for headless for idle optimization

[Why]
Currently idle worker thread that checks for HPD while system is in IPS2
only supports headless and static screen use-cases.
In other display-off scenarios hotplug may not work.

[How]
For display-off only allow idle optimization when no display is connected.

Reviewed-by: Sun peng Li <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Increase idle worker HPD detection time

[Why]
Idle worker thread waits HPD_DETECTION_TIME for HPD processing complete.
Some displays require longer time for that.

[How]
Increase HPD_DETECTION_TIME to 100ms.

Reviewed-by: Sun peng Li <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Remove useless assignments

[WHAT & HOW]
"split_pipe" are assigned to test_pipe and then immediately are updated
to other values. The same also applies to "status" as well.

Similarly, "id", "dwb" and "unused_dpps" are assigned but the functions
immediately return, and thus they have no effects.

As a results, the assignments removed.

This fixes 5 UNUSED_VALUE issues reported by Coverity.

Reviewed-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Hung <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Check status from dpcd_get_tunneling_device_data

[WHAT & HOW]
dpcd_get_tunneling_device_data calls core_link_read_dpcd which can
fail. The status from core_link_read_dpcd should be checked and error
messages is printed in case of failures.

This fixes 1 UNUSED_VALUE issue reported by Coverity.

Reviewed-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Hung <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Check returns from drm_dp_dpcd_write

[WHAT & HOW]
drm_dp_dpcd_write() returns negative error on failure and thus returned
values need to be checked.

This fixes 3 UNUSED_VALUE issues reported by Coverity.

Reviewed-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Hung <[email protected]>
Signed-off-by: Roman Li <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Wait for reset on init completion

When reset on initialization is requested, wait for the reset to finish.
In cases where module is loaded after boot, this makes sure all
initialization work is done after a successful return of modprobe.

Signed-off-by: Lijo Lazar <[email protected]>
Reviewed-by: Ramesh Errabolu <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdkfd: Accounting pdd vram_usage for svm

Process device data pdd->vram_usage is read by rocm-smi via sysfs, this
is currently missing the svm_bo usage accounting, so "rocm-smi
--showpids" per process VRAM usage report is incorrect.

Add pdd->vram_usage accounting when svm_bo allocation and release,
change to atomic64_t type because it is updated outside process mutex
now.

Signed-off-by: Philip Yang <[email protected]>
Reviewed-by: Felix Kuehling <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/amdgpu: Fix double unlock in amdgpu_mes_add_ring

This patch addresses a double unlock issue in the amdgpu_mes_add_ring
function. The mutex was being unlocked twice under certain error
conditions, which could lead to undefined behavior.

The fix ensures that the mutex is unlocked only once before jumping to
the clean_up_memory label. The unlock operation is moved to just before
the goto statement within the conditional block that checks the return
value of amdgpu_ring_init. This prevents the second unlock attempt after
the clean_up_memory label, which is no longer necessary as the mutex is
already unlocked by this point in the code flow.

This change resolves the potential double unlock and maintains the
correct mutex handling throughout the function.

Fixes below:
Commit d0c423b64765 ("drm/amdgpu/mes: use ring for kernel queue
submission"), leads to the following Smatch static checker warning:

drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c:1240 amdgpu_mes_add_ring()
warn: double unlock '&adev->mes.mutex_hidden' (orig line 1213)

drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
    1143 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
    1144                         int queue_type, int idx,
    1145                         struct amdgpu_mes_ctx_data *ctx_data,
    1146                         struct amdgpu_ring **out)
    1147 {
    1148         struct amdgpu_ring *ring;
    1149         struct amdgpu_mes_gang *gang;
    1150         struct amdgpu_mes_queue_properties qprops = {0};
    1151         int r, queue_id, pasid;
    1152
    1153         /*
    1154          * Avoid taking any other locks under MES lock to avoid circular
    1155          * lock dependencies.
    1156          */
    1157         amdgpu_mes_lock(&adev->mes);
    1158         gang = idr_find(&adev->mes.gang_id_idr, gang_id);
    1159         if (!gang) {
    1160                 DRM_ERROR("gang id %d doesn't exist\n", gang_id);
    1161                 amdgpu_mes_unlock(&adev->mes);
    1162                 return -EINVAL;
    1163         }
    1164         pasid = gang->process->pasid;
    1165
    1166         ring = kzalloc(sizeof(struct amdgpu_ring), GFP_KERNEL);
    1167         if (!ring) {
    1168                 amdgpu_mes_unlock(&adev->mes);
    1169                 return -ENOMEM;
    1170         }
    1171
    1172         ring->ring_obj = NULL;
    1173         ring->use_doorbell = true;
    1174         ring->is_mes_queue = true;
    1175         ring->mes_ctx = ctx_data;
    1176         ring->idx = idx;
    1177         ring->no_scheduler = true;
    1178
    1179         if (queue_type == AMDGPU_RING_TYPE_COMPUTE) {
    1180                 int offset = offsetof(struct amdgpu_mes_ctx_meta_data,
    1181                                       compute[ring->idx].mec_hpd);
    1182                 ring->eop_gpu_addr =
    1183                         amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
    1184         }
    1185
    1186         switch (queue_type) {
    1187         case AMDGPU_RING_TYPE_GFX:
    1188                 ring->funcs = adev->gfx.gfx_ring[0].funcs;
    1189                 ring->me = adev->gfx.gfx_ring[0].me;
    1190                 ring->pipe = adev->gfx.gfx_ring[0].pipe;
    1191                 break;
    1192         case AMDGPU_RING_TYPE_COMPUTE:
    1193                 ring->funcs = adev->gfx.compute_ring[0].funcs;
    1194                 ring->me = adev->gfx.compute_ring[0].me;
    1195                 ring->pipe = adev->gfx.compute_ring[0].pipe;
    1196                 break;
    1197         case AMDGPU_RING_TYPE_SDMA:
    1198                 ring->funcs = adev->sdma.instance[0].ring.funcs;
    1199                 break;
    1200         default:
    1201                 BUG();
    1202         }
    1203
    1204         r = amdgpu_ring_init(adev, ring, 1024, NULL, 0,
    1205                              AMDGPU_RING_PRIO_DEFAULT, NULL);
    1206         if (r)
    1207                 goto clean_up_memory;
    1208
    1209         amdgpu_mes_ring_to_queue_props(adev, ring, &qprops);
    1210
    1211         dma_fence_wait(gang->process->vm->last_update, false);
    1212         dma_fence_wait(ctx_data->meta_data_va->last_pt_update, false);
    1213         amdgpu_mes_unlock(&adev->mes);
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

    1214
    1215         r = amdgpu_mes_add_hw_queue(adev, gang_id, &qprops, &queue_id);
    1216         if (r)
    1217                 goto clean_up_ring;
                         ^^^^^^^^^^^^^^^^^^

    1218
    1219         ring->hw_queue_id = queue_id;
    1220         ring->doorbell_index = qprops.doorbell_off;
    1221
    1222         if (queue_type == AMDGPU_RING_TYPE_GFX)
    1223                 sprintf(ring->name, "gfx_%d.%d.%d", pasid, gang_id, queue_id);
    1224         else if (queue_type == AMDGPU_RING_TYPE_COMPUTE)
    1225                 sprintf(ring->name, "compute_%d.%d.%d", pasid, gang_id,
    1226                         queue_id);
    1227         else if (queue_type == AMDGPU_RING_TYPE_SDMA)
    1228                 sprintf(ring->name, "sdma_%d.%d.%d", pasid, gang_id,
    1229                         queue_id);
    1230         else
    1231                 BUG();
    1232
    1233         *out = ring;
    1234         return 0;
    1235
    1236 clean_up_ring:
    1237         amdgpu_ring_fini(ring);
    1238 clean_up_memory:
    1239         kfree(ring);
--> 1240         amdgpu_mes_unlock(&adev->mes);
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

    1241         return r;
    1242 }

Fixes: d0c423b64765 ("drm/amdgpu/mes: use ring for kernel queue submission")
Cc: Christian König <[email protected]>
Cc: Alex Deucher <[email protected]>
Cc: Hawking Zhang <[email protected]>
Suggested-by: Jack Xiao <[email protected]>
Reported by: Dan Carpenter <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Jack Xiao <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Add hpd_source index out-of-bounds check for dcn3x link encoder creation

This patch adds a boundary check for the hpd_source index during the
link encoder creation process for dcn3x IP's. The check ensures that the
index is within the valid range of the link_enc_hpd_regs array to
prevent out-of-bounds access.

Cc: Tom Chung <[email protected]>
Cc: Rodrigo Siqueira <[email protected]>
Cc: Roman Li <[email protected]>
Cc: Alex Hung <[email protected]>
Cc: Aurabindo Pillai <[email protected]>
Cc: Harry Wentland <[email protected]>
Cc: Hamza Mahfooz <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Roman Li <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/mes: fix issue of writing to the same log buffer from 2 MES pipes

With Unified MES enabled in gfx12, need separate event log buffer for the
2 MES pipes to avoid data overwrite.

Signed-off-by: Michael Chen <[email protected]>
Reviewed-by: Jack Xiao <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Show current compute partition on VF

Enable sysfs node for current compute partition mode on VFs also.

Signed-off-by: Lijo Lazar <[email protected]>
Reviewed-by: Vignesh Chander <[email protected]>
Tested-by: Vignesh Chander <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Fetch NPS mode for GCv9.4.3 VFs

Use the memory ranges published in discovery table to deduce NPS mode
of GC v9.4.3 VFs.

Signed-off-by: Lijo Lazar <[email protected]>
Reviewed-by: Vignesh Chander <[email protected]>
Tested-by: Vignesh Chander <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: prevent BO_HANDLES error from being overwritten

Before this patch, if multiple BO_HANDLES chunks were submitted,
the error -EINVAL would be correctly set but could be overwritten
by the return value from amdgpu_cs_p1_bo_handles(). This patch
ensures that if there are multiple BO_HANDLES, we stop.

Fixes: fec5f8e8c6bc ("drm/amdgpu: disallow multiple BO_HANDLES chunks in one submit")
Signed-off-by: Mohammed Anees <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Pierre-Eric Pelloux-Prayer <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/pm: Fill pcie recov cntr to metrics 1.6

Fill pcie other end recovery counter to metrics 1.6

v2: Add separate function to check recovery counter support

Signed-off-by: Asad Kamal <[email protected]>
Reviewed-by: Lijo Lazar <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/pm: Update SMUv13.0.6 PMFW headers

Update pmfw headers for smuv13.0.6 to version 0xE

Signed-off-by: Asad Kamal <[email protected]>
Reviewed-by: Lijo Lazar <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: enable enforce_isolation sysfs node on VFs

It should be enabled on both bare metal and VFs.

Fixes: e189be9b2e38 ("drm/amdgpu: Add enforce_isolation sysfs attribute")
Signed-off-by: Alex Deucher <[email protected]>
Cc: Srinivasan Shanmugam <[email protected]>
Cc: Amber Lin <[email protected]>
Reviewed-by: Srinivasan Shanmugam <[email protected]>

drm/amdgpu: Add NPS switch support for GC 9.4.3

Add dynamic NPS switch support for GC 9.4.3 variants. Only GC v9.4.3 and
GC v9.4.4 currently support this. NPS switch is only supported if an SOC
supports multiple NPS modes.

Signed-off-by: Lijo Lazar <[email protected]>
Signed-off-by: Rajneesh Bhardwaj <[email protected]>
Reviewed-by: Feifei Xu <[email protected]>
Reviewed-by: Hawking Zhang <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/gfx12: Apply Isolation Enforcement to GFX & Compute rings

This commit applies isolation enforcement to the GFX and Compute rings
in the gfx_v12_0 module.

The commit sets `amdgpu_gfx_enforce_isolation_ring_begin_use` and
`amdgpu_gfx_enforce_isolation_ring_end_use` as the functions to be
called when a ring begins and ends its use, respectively.

`amdgpu_gfx_enforce_isolation_ring_begin_use` is called when a ring
begins its use. This function cancels any scheduled
`enforce_isolation_work` and, if necessary, signals the Kernel Fusion
Driver (KFD) to stop the runqueue.

`amdgpu_gfx_enforce_isolation_ring_end_use` is called when a ring ends
its use. This function schedules `enforce_isolation_work` to be run
after a delay.

These functions are part of the Enforce Isolation Handler, which
enforces shader isolation on AMD GPUs to prevent data leakage between
different processes.

Cc: Christian König <[email protected]>
Cc: Alex Deucher <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: optimize fn gfx_v12_ring_insert_nop

Optimize gfx_v12_ring_insert_nop() to call
optimized version of amdgpu_ring_insert_nop
instead of calling amdgpu_ring_write for number
of nop times.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: optimize fn gfx_v11_ring_insert_nop

Optimize gfx_v11_ring_insert_nop() to call
optimized version of amdgpu_ring_insert_nop
instead of calling amdgpu_ring_write for number
of nop times.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: optimize fn gfx_v10_ring_insert_nop

Optimize gfx_v10_ring_insert_nop() to call
optimized version of amdgpu_ring_insert_nop
instead of calling amdgpu_ring_write for number
of nop times.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: optimize fn gfx_v9_ring_insert_nop

Optimize gfx_v9_ring_insert_nop() to call
optimized version of amdgpu_ring_insert_nop
instead of calling amdgpu_ring_write for number
of nop times.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: optimize fn gfx_v9_4_3_ring_insert_nop

Optimize gfx_v9_4_3_ring_insert_nop() to call
optimized version of amdgpu_ring_insert_nop
instead of calling amdgpu_ring_write for number
of nop times.

Signed-off-by: Sunil Khatri <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: optimize insert_nop using multi dwords

Optimize the ring_insert_nop fn for n dwords in one
step rather then call to amdgpu_ring_write for each
nop packet. This avoid function call for each nop
packet and also wptr is updated once only.

Signed-off-by: Sunil Khatri <[email protected]>
Suggested-by: Christian König <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Check gmc requirement for reset on init

Add a callback to check if there is any condition detected by GMC block
for reset on init. One case is if a pending NPS change request is
detected. If reset is done because of NPS switch, refresh NPS info from
discovery table.

Signed-off-by: Lijo Lazar <[email protected]>
Reviewed-by: Feifei Xu <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Place NPS mode request on unload

If a user has requested NPS mode switch, place the request through PSP
during unload of the driver. For devices which are part of a hive, all
requests are placed together. If one of them fails, revert back to the
current NPS mode.

Signed-off-by: Lijo Lazar <[email protected]>
Signed-off-by: Rajneesh Bhardwaj <[email protected]>
Reviewed-by: Feifei Xu <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/radeon: always set GEM function pointer

Make sure to always set the GEM function pointer even for in kernel
allocations. This fixes a NULL pointer deref caused by switching to GEM
references.

Signed-off-by: Christian König <[email protected]>
Fixes: fd69ef05029f ("drm/radeon: use GEM references instead of TTMs")
Acked-by: Alex Deucher <[email protected]>
Tested-by: Hans de Goede <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: fix dm_suspend/resume arguments to ip_block

"build failure after merge of the amdgpu tree"
dm_suspend/dm_resume functions argument mismatch
not caught in validation as it was under config
CONFIG_DEBUG_KERNEL_DC which wasnt enabled by
default.

Change argument from adev to ip_block.

Fixes: 982d7f9bfe4a ("drm/amdgpu: update the handle ptr in suspend")
Fixes: 7feb4f3ad8be ("drm/amdgpu: update the handle ptr in resume")
Reported-by: Stephen Rothwell <[email protected]>
Signed-off-by: Sunil Khatri <[email protected]>
Acked-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: no need to log error in multi ring write

No need to log error in multi ring write as its taken
care during ring commit.

This is inline with change done in amdgpu_ring_write.

Signed-off-by: Sunil Khatri <[email protected]>
Suggested-by: Christian König <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: move error log from ring write to commit

Move the error message from ring write as an optimization
to avoid printing that message on every write instead
print once during commit if it exceeds write the allocated
size i.e ring->count_dw.

Also we do not want to log the error message in between a
ring write and complete the write as its mostly not harmful
as it will overwrite stale data only as GPU read from ring
is faster than CPU write to ring.

This reduces the size of amdgpu.ko module by around
600 Kb as write is very often used function and hence
the print.

Signed-off-by: Sunil Khatri <[email protected]>
Suggested-by: Christian König <[email protected]>
Reviewed-by: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: fix typos

Fix typos in comments: "wether -> whether".

Signed-off-by: Andrew Kreimer <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Remove the while loop from amdgpu_job_prepare_job

While loop makes it sound like amdgpu_vmid_grab() potentially needs to be
called multiple times to produce a fence, while in reality all code paths
either return an error, assign a valid job->vmid or assign a vmid which
will be valid once the returned fence signals.

Therefore we can remove the loop to make it clear the call does not need
to be repeated.

Reviewed-by: Christian König <[email protected]>
Signed-off-by: Tvrtko Ursulin <[email protected]>
Cc: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Drop impossible condition from amdgpu_job_prepare_job

Fence has been initialised to NULL so no need to test it.

Reviewed-by: Christian König <[email protected]>
Signed-off-by: Tvrtko Ursulin <[email protected]>
Cc: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: disable SG displays on cyan skillfish

These parts were mainly for compute workloads, but they have
a display that was available for the console.  These chips
should support SG display, but I don't know that the support
was ever validated on Linux so disable it by default. It can
still be enabled by setting sg_display=1 for those that
want to play with it.  These systems also generally had large
carve outs so SG display was less of a factor.

Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/3356
Reviewed-by: Harry Wentland <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Use drm_print_memory_stats helper from fdinfo

Convert fdinfo memory stats to use the common drm_print_memory_stats
helper.

This achieves alignment with the common keys as documented in
drm-usage-stats.rst, adding specifically drm-total- key the driver was
missing until now.

Additionally I made the code stop skipping total size for objects which
currently do not have a backing store, and I added resident, active and
purgeable reporting.

Legacy keys have been preserved, with the outlook of only potentially
removing only the drm-memory- when the time gets right.

The example output now looks like this:

 pos: 0
 flags: 02100002
 mnt_id: 24
 ino: 1239
 drm-driver: amdgpu
 drm-client-id: 4
 drm-pdev: 0000:04:00.0
 pasid: 32771
 drm-total-cpu: 0
 drm-shared-cpu: 0
 drm-active-cpu: 0
 drm-resident-cpu: 0
 drm-purgeable-cpu: 0
 drm-total-gtt: 2392 KiB
 drm-shared-gtt: 0
 drm-active-gtt: 0
 drm-resident-gtt: 2392 KiB
 drm-purgeable-gtt: 0
 drm-total-vram: 44564 KiB
 drm-shared-vram: 31952 KiB
 drm-active-vram: 0
 drm-resident-vram: 44564 KiB
 drm-purgeable-vram: 0
 drm-memory-vram: 44564 KiB
 drm-memory-gtt:  2392 KiB
 drm-memory-cpu:  0 KiB
 amd-memory-visible-vram: 44564 KiB
 amd-evicted-vram: 0 KiB
 amd-evicted-visible-vram: 0 KiB
 amd-requested-vram: 44564 KiB
 amd-requested-visible-vram: 11952 KiB
 amd-requested-gtt: 2392 KiB
 drm-engine-compute: 46464671 ns

v2:
 * Track purgeable via AMDGPU_GEM_CREATE_DISCARDABLE.

Acked-by: Daniel Vetter <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Signed-off-by: Tvrtko Ursulin <[email protected]>
Cc: Alex Deucher <[email protected]>
Cc: Christian König <[email protected]>
Cc: Daniel Vetter <[email protected]>
Cc: Rob Clark <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Drop unused fence argument from amdgpu_vmid_grab_used

Fence argument is unused so lets drop it.

Reviewed-by: Christian König <[email protected]>
Signed-off-by: Tvrtko Ursulin <[email protected]>
Cc: Christian König <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: partially revert powerplay `__counted_by` changes

Partially revert
commit 0ca9f757a0e2 ("drm/amd/pm: powerplay: Add `__counted_by` attribute for flexible arrays")

The count attribute for these arrays does not get set until
after the arrays are allocated and populated leading to false
UBSAN warnings.

Fixes: 0ca9f757a0e2 ("drm/amd/pm: powerplay: Add `__counted_by` attribute for flexible arrays")
Reviewed-by: Mario Limonciello <[email protected]>
Reviewed-by: Lijo Lazar <[email protected]>
Closes: https://gitlab.freedesktop.org/drm/amd/-/issues/3662
Signed-off-by: Alex Deucher <[email protected]>

Documentation/gpu: Document the situation with unqualified drm-memory-

Currently it is not well defined what is drm-memory- compared to other
categories.

In practice the only driver which emits these keys is amdgpu and in them
exposes the current resident buffer object memory (including shared).

To prevent any confusion, document that drm-memory- is deprecated and an
alias for drm-resident-memory-.

While at it also clarify that the reserved sub-string 'memory' refers to
the memory region component, and also clarify the intended semantics of
other memory categories.

v2:
 * Also mark drm-memory- as deprecated.
 * Add some more text describing memory categories. (Alex)

v3:
 * Semantics of the amdgpu drm-memory is actually as drm-resident.

Reviewed-by: Rob Clark <[email protected]>
Acked-by: Daniel Vetter <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Signed-off-by: Tvrtko Ursulin <[email protected]>
Cc: Alex Deucher <[email protected]>
Cc: Christian König <[email protected]>
Cc: Rob Clark <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdkfd: SMI report dropped event count

Add new SMI event to report the dropped event count.

When the event kfifo is full, drop count is not zero, or no enough space
left to store the event message, increase drop count.

After reading event out from kfifo, if event was dropped, drop_count is
not zero, generate a dropped event record and reset drop count to zero.

Signed-off-by: Philip Yang <[email protected]>
Reviewed-by: James Zhu <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Add sysfs interfaces for NPS mode

Add a sysfs interface to see available NPS modes to switch to -

cat /sys/bus/pci/devices/../available_memory_paritition

Make the current_memory_partition sysfs node read/write for requesting a
new NPS mode. The request is only cached and at a later point a driver
unload/reload is required to switch to the new NPS mode.

Ex:
echo NPS1 > /sys/bus/pci/devices/../current_memory_paritition
echo NPS4 > /sys/bus/pci/devices/../current_memory_paritition

The above interfaces will be available only if the SOC supports more than
one NPS mode.

Also modify the current memory partition sysfs logic to be more
generic.

Signed-off-by: Lijo Lazar <[email protected]>
Reviewed-by: Rajneesh Bhardwaj <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu: Add gmc interface to request NPS mode

Add a common interface in GMC to request NPS mode through PSP. Also add
a variable in hive and gmc control to track the last requested mode.

Signed-off-by: Rajneesh Bhardwaj <[email protected]>
Signed-off-by: Lijo Lazar <[email protected]>
Reviewed-by: Feifei Xu <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amdgpu/gfx10: Apply Isolation Enforcement to GFX & Compute rings

This commit applies isolation enforcement to the GFX and Compute rings
in the gfx_v10_0 module.

The commit sets `amdgpu_gfx_enforce_isolation_ring_begin_use` and
`amdgpu_gfx_enforce_isolation_ring_end_use` as the functions to be
called when a ring begins and ends its use, respectively.

`amdgpu_gfx_enforce_isolation_ring_begin_use` is called when a ring
begins its use. This function cancels any scheduled
`enforce_isolation_work` and, if necessary, signals the Kernel Fusion
Driver (KFD) to stop the runqueue.

`amdgpu_gfx_enforce_isolation_ring_end_use` is called when a ring ends
its use. This function schedules `enforce_isolation_work` to be run
after a delay.

These functions are part of the Enforce Isolation Handler, which
enforces shader isolation on AMD GPUs to prevent data leakage between
different processes.

Cc: Christian König <[email protected]>
Cc: Alex Deucher <[email protected]>
Signed-off-by: Srinivasan Shanmugam <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: fix hibernate entry for DCN35+

Since, two suspend-resume cycles are required to enter hibernate and,
since we only need to enable idle optimizations in the first cycle
(which is pretty much equivalent to s2idle). We can check in_s0ix, to
prevent the system from entering idle optimizations before it actually
enters hibernate (from display's perspective). Also, call
dc_set_power_state() before dc_allow_idle_optimizations(), since it's
safer to do so because dc_set_power_state() writes to DMUB.

Acked-by: Alex Deucher <[email protected]>
Signed-off-by: Hamza Mahfooz <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: Fetch the EDID from _DDC if available for eDP

Some manufacturers have intentionally put an EDID that differs from
the EDID on the internal panel on laptops.

Attempt to fetch this EDID if it exists and prefer it over the EDID
that is provided by the panel. If a user prefers to use the EDID from
the panel, offer a DC debugging parameter that would disable this.

Signed-off-by: Mario Limonciello <[email protected]>
Reviewed-by: Alex Hung <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: remove redundant freesync parser for DP

When updating connector under drm_edid infrastructure, many calculations
and validations are already done and become redundant inside AMD driver.
Remove those driver-specific code in favor of the DRM common code.

Signed-off-by: Melissa Wen <[email protected]>
Co-developed-by: Mario Limonciello <[email protected]>
Signed-off-by: Mario Limonciello <[email protected]>
Reviewed-by: Alex Hung <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>

drm/amd/display: always call connector_update when parsing freesync_caps

Update connector caps with drm_edid data before parsing info for
freesync.

Signed-off-by: Melissa Wen <[email protected]>
Signed-off-by: Mario Limonciello <[email protected]>
Reviewed-by: Alex Hung <[email protected]>
Tested-by: Daniel Wheeler <[email protected]>
Signed-off-by: Rodrigo Siqueira <[email protected]>
Signed-off-by: Alex Deucher <[email protected]>