Hi,
In the display core, we utilize floats and doubles units for calculating modesetting parameters. One side effect of our approach to use double-precision is the fact that we spread multiple FPU access across our driver, which means that we can accidentally clobber user space FPU state.
# Challenges
1. Keep in mind that this FPU code is ingrained in our display driver and performs several crucial tasks. Additionally, we already have multiple architectures available in the kernel and a large set of users; in other words, we prefer to avoid a radical approach that might break our user's system.
2. We share our display code with other OSs; thus, we need to maintain the interoperability between these two systems.
3. We need a mechanism for identifying which function uses FPU registers; fortunately, Peter Zijlstra wrote a series a couple of months ago where he introduced an FPU check for objtool. I used the following command for identifying the potential FPU usage:
./tools/objtool/objtool check -Ffa "drivers/gpu/drm/amd/display/dc/ANY_FILE.o"
4. Since our code heavily relies on FPU and the fact that we spread kernel_fpu_begin/end across multiple functions, we can have some complex scenarios that will require code refactoring. However, we want to avoid complicated changes since this is a formula to introduce regressions; we want something that allows us to fix it in small, safe, and reliable steps.
5. Unfortunately, for legacy reasons, we have some problems in how we program our FPU access, which in some weird scenarios can generate situations where we try to enter in the fpu mode multiple times or exit too early.
# Our approach
For trying to solve this problem, we came up with the following strategy:
1. Keep in mind that we are using kernel_fpu_begin/end spread in various areas and sometimes across multiple functions. If we try to move some of the functions to an isolated place, we can generate a situation where we can call the FPU protection more than once, causing multiple warnings. We can deal with this problem by adding a thin management layer around the kernel_fpu_begin/end used inside the display.
2. We will need a trace mechanism for this FPU management inside our display code.
3. After we get the thin layer that manages FPU, we can start to move each function that uses FPU to a centralized place. Our DQE runs multiple tests in different ASICs every week; we can take advantage of this to ensure that our FPU patches work does not introduce any regression. The idea is to work on a specific part of the code every week (e.g., week 1: DCN2, week 1: DCN2.1, etc.).
4. Finally, after we can isolate the FPU operations in a single place, we can altogether remove the FPU flags from other files and eliminate an unnecessary code introduced to deal with this problem. We can also remove the thin layer added in the step 3.
# This series
To maintain the interoperability between multiple OSes, we already have a define named DC_FP_START/END, which is a straightforward wrapper to kernel_fpu_begin/end in the Linux side. In this series, I decided to expand the scope of this DC_FP_* wrapper to trace FPU entrance and exit in the display code, but I also add a mechanism for managing the entrance and exit of kernel_fpu_begin/end. You can see the details on how I did that in the last two patches.
I also isolate a simple function that requires FPU access to demonstrate my strategy for isolating this FPU access in a single place. If this series gets accepted, the following steps consist of moving all FPU functions weekly until we isolate everything in the DML folder.
Cc: Harry Wentland harry.wentland@amd.com Cc: Anson Jacob Anson.Jacob@amd.com Cc: Christian König christian.koenig@amd.com Cc: Hersen Wu hersenxs.wu@amd.com Cc: Aric Cyr aric.cyr@amd.com Cc: Jun Lei jun.lei@amd.com Cc: Dmytro Laktyushkin dmytro.laktyushkin@amd.com Cc: Qingqing Zhuo qingqing.zhuo@amd.com
Change Since V3: - Use DML as the FPU folder instead create a new folder - Add a proper guard to x86 and PPC64
Change Since V2: - Make sure to compile FPU operation only when DCN is enabled (officially, we only enable it for x86). - Check cross-compile with ARM and x86_32. Everything looks fine. - Avoid call this_cpu_* operations between get/put_cpu_ptr. - Fix GCC warnings. - Update documentation. - Update our assert mechanism. - Remove unnecessary wrappers.
Changes since V1: - Use a better name for variables. - Update documentation. - Avoid preemption.
* See update details per commit message
Rodrigo Siqueira (4): drm/amd/display: Move specific DCN2x code that uses FPU to DML drm/amd/display: Add control mechanism for FPU drm/amd/display: Add control mechanism for FPU utilization drm/amd/display: Add DC_FP helper to check FPU state
.../gpu/drm/amd/display/amdgpu_dm/Makefile | 4 + .../amd/display/amdgpu_dm/amdgpu_dm_trace.h | 24 +++ .../gpu/drm/amd/display/amdgpu_dm/dc_fpu.c | 144 ++++++++++++++++++ .../gpu/drm/amd/display/amdgpu_dm/dc_fpu.h | 34 +++++ drivers/gpu/drm/amd/display/dc/dc_trace.h | 3 + .../drm/amd/display/dc/dcn20/dcn20_resource.c | 41 +---- .../drm/amd/display/dc/dcn20/dcn20_resource.h | 2 - .../drm/amd/display/dc/dcn21/dcn21_resource.c | 2 + drivers/gpu/drm/amd/display/dc/dml/Makefile | 4 + .../gpu/drm/amd/display/dc/dml/dcn2x/dcn2x.c | 102 +++++++++++++ .../gpu/drm/amd/display/dc/dml/dcn2x/dcn2x.h | 34 +++++ drivers/gpu/drm/amd/display/dc/os_types.h | 35 +---- 12 files changed, 358 insertions(+), 71 deletions(-) create mode 100644 drivers/gpu/drm/amd/display/amdgpu_dm/dc_fpu.c create mode 100644 drivers/gpu/drm/amd/display/amdgpu_dm/dc_fpu.h create mode 100644 drivers/gpu/drm/amd/display/dc/dml/dcn2x/dcn2x.c create mode 100644 drivers/gpu/drm/amd/display/dc/dml/dcn2x/dcn2x.h