- 14 Aug, 2020 7 commits
-
-
Wenhui Sheng authored
After amdgpu driver loading successfully, we can use RAP debugfs interface <debugfs_dir>/dri/xxx/rap_test to trigger RAP test. Currently only L0 validate test is supported. v2: refine amdgpu_rap.h Signed-off-by:
Wenhui Sheng <Wenhui.Sheng@amd.com> Reviewed-by:
Guchun Chen <Guchun.Chen@amd.com> Reviewed-by:
Hawking Zhang <Hawking.Zhang@amd.com> Signed-off-by:
Alex Deucher <alexander.deucher@amd.com>
-
Wenhui Sheng authored
Enable the RAP TA loading path and add RAP test trigger interface. v2: fix potential mem leak issue Signed-off-by:
-
Wenhui Sheng authored
The RAP TA contains tests used to verify if RAP(Register Access Policy), or otherwise known as Security Policy is applied correctly by PSP BL&TOS. The RAP test is a measure to ensure that we reduce the avenue of complexity and mistakes when dealing with RAP in post-si execution, where debugging failures related to RAP is quite difficult and expensive. v2: add introduction for RAP TA Signed-off-by:
-
Tianci.Yin authored
On Navi1x, the SPM golden settings are lost after GFXOFF enter/exit, so reconfigure the golden settings after GFXOFF exit. Reviewed-by:
Tianci.Yin <tianci.yin@amd.com> Signed-off-by:
-
Tianci.Yin authored
On Navi1x, the SPM golden settings are lost after GFXOFF enter/exit, so reconfiguration is needed. Make the configuration code as an interface for future use. Reviewed-by:
Luben Tuikov <luben.tuikov@amd.com> Reviewed-by:
Feifei Xu <Feifei.Xu@amd.com> Signed-off-by:
-
Guchun Chen authored
Before ras recovery is issued, user could operate this debugfs node to enable/disable the harvest of all RAS IPs' ras error count registers, which will help keep hardware's registers' status instead of cleaning up them. Signed-off-by:
Dennis Li <Dennis.Li@amd.com> Signed-off-by:
-
Guchun Chen authored
Once ras recovery is issued by ras sync flood interrupt or ras controller interrupt, add this guard to bypass or execute ras error count register harvest of all IPs. Signed-off-by:
-
- 11 Aug, 2020 1 commit
-
-
Arunpravin authored
Access the exported P2P dmabuf over XGMI, if available. Otherwise, fall back to the existing PCIe method. Reviewed-by:
Christian König <christian.koenig@amd.com> Signed-off-by:
Arunpravin <apaneers@amd.com> Signed-off-by:
-
- 10 Aug, 2020 22 commits
-
-
Qinglang Miao authored
Convert cpu_to_le16(le16_to_cpu(E1) + E2) to use le16_add_cpu(). Signed-off-by:
Qinglang Miao <miaoqinglang@huawei.com> Signed-off-by:
-
Alex Deucher authored
This is not longer used as of the latest rework of this code so drop it to avoid a unused function warning. Acked-by:
Nirmoy Das <nirmoy.das@amd.com> Signed-off-by:
-
Daniel Kolesa authored
This adds ARM64 support into the DCN. This mainly enables support for Navi graphics cards. The dcn10 changes haven't been tested, since I don't have the relevant hardware available, but there is no way to conditionally disable them, so I've done them anyway. Signed-off-by:
Daniel Kolesa <daniel@octaforge.org> Signed-off-by:
-
Daniel Kolesa authored
GFP_KERNEL may and will sleep, and this is being executed in a non-preemptible context; this will mess things up since it's called inbetween DC_FP_START/END, and rescheduling will result in the DC_FP_END later being called in a different context (or just crashing if any floating point/vector registers/instructions are used after the call is resumed in a different context). Signed-off-by:
-
Jaehyun Chung authored
[Why] Stream disable sequence incorretly destroys HDCP session while stream is not blanked and while audio is not muted. This sequence causes a flash of corruption during mode change and an audio click. [How] Change sequence to blank stream before destroying HDCP session. Audio will also be muted by blanking the stream. Cc: stable@vger.kernel.org Signed-off-by:
Jaehyun Chung <jaehyun.chung@amd.com> Reviewed-by:
Alvin Lee <Alvin.Lee2@amd.com> Acked-by:
Qingqing Zhuo <qingqing.zhuo@amd.com> Signed-off-by:
-
Stylon Wang authored
[Why] Resuming from suspend, CEA blocks from EDID are not parsed and no video modes can support YUV420. When this happens, output bpc cannot go over 8-bit with 4K modes on HDMI. [How] In amdgpu_dm_update_connector_after_detect(), drm_add_edid_modes() is called after drm_connector_update_edid_property() to fully parse EDID and update display info. Cc: stable@vger.kernel.org Signed-off-by:
Stylon Wang <stylon.wang@amd.com> Reviewed-by:
Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com> Acked-by:
-
Alvin Lee authored
[Why] When changing pixel formats for HDR (e.g. ARGB -> FP16) there are configurations that change from 2 pipes to 1 pipe. In these cases, it seems that disconnecting MPCC and doing a surface update at the same time(after unlocking) causes some registers to be updated slightly faster than others after unlocking (e.g. if the pixel format is updated to FP16 before the new surface address is programmed, we get corruption on the screen because the pixel formats aren't matching). We separate disconnecting MPCC from the rest of the pipe programming sequence to prevent this. [How] Move MPCC disconnect into separate operation than the rest of the pipe programming. Signed-off-by:
Aric Cyr <Aric.Cyr@amd.com> Acked-by:
-
Anthony Koo authored
[Why] Using FRAME_UPDATE will result in infopacket to be potentially updated one frame late. In commit stream scenarios for previously active stream, some stale infopacket data from previous config might be erroneously sent out on initial frame after stream is re-enabled. [How] Switch to using IMMEDIATE_UPDATE mode Signed-off-by:
Anthony Koo <Anthony.Koo@amd.com> Reviewed-by:
Ashley Thomas <Ashley.Thomas2@amd.com> Acked-by:
-
Anthony Koo authored
[Why] 1. There is a calculation that is using frame_time_in_us instead of last_render_time_in_us to calculate whether choosing an LFC multiplier would cause the inserted frame duration to be outside of range. 2. We do not handle unsigned integer subtraction correctly and it underflows to a really large value, which causes some logic errors. [How] 1. Fix logic to calculate 'within range' using last_render_time_in_us 2. Split out delta_from_mid_point_delta_in_us calculation to ensure we don't underflow and wrap around Signed-off-by:
-
Xiaodong Yan authored
[Why] The format in MPCC should be 444 [How] do not modify the mpcc black color according to pixel encoding format Signed-off-by:
Xiaodong Yan <Xiaodong.Yan@amd.com> Reviewed-by:
Eric Yang <eric.yang2@amd.com> Acked-by:
-
Alvin Lee authored
[Why] Caused pipe split regression Signed-off-by:
-
Aric Cyr authored
[Why] Typo in backlight refactor inctroduced wrong register offset. [How] Change DCE to DCN register map for PWRSEQ_REF_DIV Cc: stable@vger.kernel.org Signed-off-by:
-
Joshua Aberback authored
[Why] Register definitions are asic-specific, so functions that use registers of a particular asic should be static, to be exposed in asic-specific function pointer structures. [How] - make register-definition-using functions static - make some functions non-static, for future use - remove duplicate function definition Signed-off-by:
Joshua Aberback <joshua.aberback@amd.com> Reviewed-by:
-
Monk Liu authored
GFX10 KIQ will hang if we try below steps: modprobe amdgpu rmmod amdgpu modprobe amdgpu sched_hw_submission=4 Due to KIQ is always living there even after KMD unloaded thus when doing the realod KIQ will crash upon its register being programed by different values with the previous loading (the config like HQD addr, ring size, is easily changed if we alter the sched_hw_submission) the fix is we must inactive KIQ first before touching any of its registgers Signed-off-by:
Monk Liu <Monk.Liu@amd.com> Reviewed-by:
Emily Deng <Emily.Deng@amd.com> Signed-off-by:
-
shiwu.zhang authored
Update golden setting to improve performance on HPC and ML apps Signed-off-by:
shiwu.zhang <shiwu.zhang@amd.com> Tested-by:
gang.long <gang.long@amd.com> Reviewed-by:
-
Evan Quan authored
The UVD/VCE PG state is managed by UVD and VCE IP. It's error-prone to assume the bootup state in SMU based on the dpm status. Signed-off-by:
Evan Quan <evan.quan@amd.com> Acked-by:
-
Evan Quan authored
Correct the cached smu feature state on pp_features sysfs setting. Signed-off-by:
-
Liu ChengZhe authored
Some registers are not accessible to virtual function setup, so skip their initialization when in VF-SRIOV mode. v2: move SRIOV VF check into specify functions; modify commit description and comment. Signed-off-by:
Liu ChengZhe <ChengZhe.Liu@amd.com> Reviewed-by:
-
Jay Cornwall authored
s_barrier triggers a debug exception when issued with PRIV=1, DEBUG_EN=1. This causes spurious notifications to rocm-gdb. Clear MODE before issuing s_barrier and restore MODE afterwards in the context restore handler. Signed-off-by:
Jay Cornwall <jay.cornwall@amd.com> Tested-by:
Laurent Morichetti <laurent.morichetti@amd.com> Reviewed-by:
Felix Kuehling <Felix.Kuehling@amd.com> Signed-off-by:
-
Felix Kuehling authored
This reverts commit 0a5baee4. The change introduced a regression on some chips. Reverting until a proper solution can be found. Signed-off-by:
-
Felix Kuehling authored
This reverts commit ea368183. Needed due to conflicts when reverting "drm/amdkfd: Unify gfx9/gfx10 context save area layouts". Signed-off-by:
-
Christophe JAILLET authored
When '*sgt' is allocated, we must allocated 'sizeof(**sgt)' bytes instead of 'sizeof(*sg)'. The sizeof(*sg) is bigger than sizeof(**sgt) so this wastes memory but it won't lead to corruption. Fixes: f44ffd67 ("drm/amdgpu: add support for exporting VRAM using DMA-buf v3") Reviewed-by:
Christophe JAILLET <christophe.jaillet@wanadoo.fr> Signed-off-by:
-
- 07 Aug, 2020 5 commits
-
-
Sandeep Raghuraman authored
Reproducing bug report here: After hibernating and resuming, DPM is not enabled. This remains the case even if you test hibernate using the steps here: https://www.kernel.org/doc/html/latest/power/basic-pm-debugging.html I debugged the problem, and figured out that in the file hardwaremanager.c, in the function, phm_enable_dynamic_state_management(), the check 'if (!hwmgr->pp_one_vf && smum_is_dpm_running(hwmgr) && !amdgpu_passthrough(adev) && adev->in_suspend)' returns true for the hibernate case, and false for the suspend case. This means that for the hibernate case, the AMDGPU driver doesn't enable DPM (even though it should) and simply returns from that function. In the suspend case, it goes ahead and enables DPM, even though it doesn't need to. I debugged further, and found out that in the case of suspend, for the CIK/Hawaii GPUs, smum_is_dpm_running(hwmgr) returns false, while in the case of hibernate, smum_is_dpm_running(hwmgr) returns true. For CIK, the ci_is_dpm_running() function calls the ci_is_smc_ram_running() function, which is ultimately used to determine if DPM is currently enabled or not, and this seems to provide the wrong answer. I've changed the ci_is_dpm_running() function to instead use the same method that some other AMD GPU chips do (e.g Fiji), which seems to read the voltage controller. I've tested on my R9 390 and it seems to work correctly for both suspend and hibernate use cases, and has been stable so far. Bug: https://bugzilla.kernel.org/show_bug.cgi?id=208839Signed-off-by:
Sandeep Raghuraman <sandy.8925@gmail.com> Signed-off-by:
-
Dennis Li authored
Make sure to unlock the mutex when error happen v2: 1. correct syntax error in the commit comments 2. remove change-Id Acked-by:
-
Evan Quan authored
As VCN related dpm table setup needs VCN be in PG ungate state. Same logics applies to JPEG. V2: fix paste typo V3: code cosmetic Signed-off-by:
Matt Coffin <mcoffin13@gmail.com> Reviewed-by:
-
Evan Quan authored
Add lock protections and avoid unnecessary actions if the PG state is already the same as required. Signed-off-by:
-
Likun Gao authored
Swith default gpu reset method for sienna_cichlid to MODE1 reset. Signed-off-by:
Likun Gao <Likun.Gao@amd.com> Reviewed-by:
-
- 06 Aug, 2020 5 commits
-
-
Nicholas Kazlauskas authored
[Why] This was added in the past to solve the issue of not knowing when to stall for medium and full updates in DM. Since DC is ultimately decides what requires bandwidth changes we wanted to make use of it directly to determine this. The problem is that we can't actually pass any of the stream or surface updates into DC global validation, so we don't actually check if the new configuration is valid - we just validate the old existing config instead and stall for outstanding commits to finish. There's also the problem of grabbing the DRM private object for pageflips which can lead to page faults in the case where commits execute out of order and free a DRM private object state that was still required for commit tail. [How] Now that we reset the plane in DM with the same conditions DC checks we can have planes go through DC validation and we know when we need to check and stall based on whether the stream or planes changed. We mark lock_and_validation_needed whenever we've done this, so just go back to using that instead of dm_determine_update_type_for_commit. Since we'll skip resetting the plane for a pageflip we will no longer grab the DRM private object for pageflips as well, avoiding the page fault issued caused by pageflipping under load with commits executing out of order. Signed-off-by:
Rodrigo Siqueira <Rodrigo.Siqueira@amd.com> Signed-off-by:
-
Nicholas Kazlauskas authored
[Why] MEDIUM or FULL updates can require global validation or affect bandwidth. By treating these all simply as surface updates we aren't actually passing this through DC global validation. [How] There's currently no way to pass surface updates through DC global validation, nor do I think it's a good idea to change the interface to accept these. DC global validation itself is currently stateless, and we can move our update type checking to be stateless as well by duplicating DC surface checks in DM based on DRM properties. We wanted to rely on DC automatically determining this since DC knows best, but DM is ultimately what fills in everything into DC plane state so it does need to know as well. There are basically only three paths that we exercise in DM today: 1) Cursor (async update) 2) Pageflip (fast update) 3) Full pipe programming (medium/full updates) Which means that anything that's more than a pageflip really needs to go down path #3. So this change duplicates all the surface update checks based on DRM state instead inside of should_reset_plane(). Next step is dropping dm_determine_update_type_for_commit and we no longer require the old DC state at all for global validation. Optimization can come later so we don't reset DC planes at all for MEDIUM udpates and avoid validation, but we might require some extra checks in DM to achieve this. Signed-off-by:
Hersen Wu <hersenxs.wu@amd.com> Reviewed-by:
-
Nicholas Kazlauskas authored
[Why] So we're not racing with userspace or deadlocking DM. [How] These flags are now stored on dm_plane_state itself and acquried and validated during commit_check, so just use those instead. Signed-off-by:
-
Nicholas Kazlauskas authored
[Why] We're racing with userspace as the flags could potentially change from when we acquired and validated them in commit_check. [How] We unfortunately can't drop this function in its entirety from prepare_planes since we don't know the afb->address at commit_check time yet. So instead of querying new tiling_flags and tmz_surface use the ones from the plane_state directly. While we're at it, also update the force_disable_dcc option based on the state from atomic check. Signed-off-by:
-
Nicholas Kazlauskas authored
[Why] Enabling or disable DCC or switching between tiled and linear formats can require bandwidth updates. They're currently skipping all DC validation by being treated as purely surface updates. [How] Treat tiling_flag changes (which encode DCC state) as a condition for resetting the plane. Signed-off-by:
-
