Re: [PATCH v4 3/4] drm/amd/display: Add control mechanism for FPU utilization

Christian König <christian.koenig@xxxxxxx> · Tue, 27 Jul 2021 09:00:10 +0200

Am 27.07.21 um 02:52 schrieb Rodrigo Siqueira:
DC invokes DC_FPU_START/END in multiple parts of the code; this can
create a situation where we invoke this FPU operation in a nested way or
exit too early. For avoiding this situation, this commit adds a
mechanism where dc_fpu_begin/end manages the access to
kernel_fpu_begin/end.

Change since V3:
- Rebase

Change since V2:
- Christian: Do not use this_cpu_* between get/put_cpu_ptr().

Change since V1:
- Use a better variable names
- Use get_cpu_ptr and put_cpu_ptr to better balance preemption enable
and disable

Cc: Harry Wentland <harry.wentland@xxxxxxx>
Cc: Anson Jacob <Anson.Jacob@xxxxxxx>
Cc: Christian König <christian.koenig@xxxxxxx>
Cc: Hersen Wu <hersenxs.wu@xxxxxxx>
Cc: Aric Cyr <aric.cyr@xxxxxxx>
Cc: Jun Lei <jun.lei@xxxxxxx>
Cc: Dmytro Laktyushkin <dmytro.laktyushkin@xxxxxxx>
Cc: Qingqing Zhuo <qingqing.zhuo@xxxxxxx>
Reported-by: kernel test robot <lkp@xxxxxxxxx>
Signed-off-by: Rodrigo Siqueira <Rodrigo.Siqueira@xxxxxxx>

Reviewed-by: Christian König <christian.koenig@xxxxxxx>

---
  .../amd/display/amdgpu_dm/amdgpu_dm_trace.h   | 13 ++--
  .../gpu/drm/amd/display/amdgpu_dm/dc_fpu.c    | 77 +++++++++++++------
  drivers/gpu/drm/amd/display/dc/dc_trace.h     |  4 +-
  3 files changed, 63 insertions(+), 31 deletions(-)

diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_trace.h b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_trace.h
index 230bb12c405e..fdcaea22b456 100644
--- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_trace.h
+++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm_trace.h
@@ -638,23 +638,26 @@ TRACE_EVENT(amdgpu_refresh_rate_track,
  );
  
  TRACE_EVENT(dcn_fpu,
-	    TP_PROTO(bool begin, const char *function, const int line),
-	    TP_ARGS(begin, function, line),
+	    TP_PROTO(bool begin, const char *function, const int line, const int recursion_depth),
+	    TP_ARGS(begin, function, line, recursion_depth),
  
  	    TP_STRUCT__entry(
  			     __field(bool, begin)
  			     __field(const char *, function)
  			     __field(int, line)
+			     __field(int, recursion_depth)
  	    ),
  	    TP_fast_assign(
  			   __entry->begin = begin;
  			   __entry->function = function;
  			   __entry->line = line;
+			   __entry->recursion_depth = recursion_depth;
  	    ),
-	    TP_printk("%s()+%d: %s",
+	    TP_printk("%s: recursion_depth: %d: %s()+%d:",
+		      __entry->begin ? "begin" : "end",
+		      __entry->recursion_depth,
  		      __entry->function,
-		      __entry->line,
-		      __entry->begin ? "begin" : "end"
+		      __entry->line
  	    )
  );
  
diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/dc_fpu.c b/drivers/gpu/drm/amd/display/amdgpu_dm/dc_fpu.c
index d4ab27bf8d04..33807d746e76 100644
--- a/drivers/gpu/drm/amd/display/amdgpu_dm/dc_fpu.c
+++ b/drivers/gpu/drm/amd/display/amdgpu_dm/dc_fpu.c
@@ -33,6 +33,19 @@
  #include <asm/cputable.h>
  #endif
  
+/**
+ * DOC: DC FPU manipulation overview
+ *
+ * DC core uses FPU operations in multiple parts of the code, which requires a
+ * more specialized way to manage these areas' entrance. To fulfill this
+ * requirement, we created some wrapper functions that encapsulate
+ * kernel_fpu_begin/end to better fit our need in the display component. In
+ * summary, in this file, you can find functions related to FPU operation
+ * management.
+ */
+
+static DEFINE_PER_CPU(int, fpu_recursion_depth);
+
  /**
   * dc_fpu_begin - Enables FPU protection
   * @function_name: A string containing the function name for debug purposes
@@ -48,22 +61,30 @@
   */
  void dc_fpu_begin(const char *function_name, const int line)
  {
-	TRACE_DCN_FPU(true, function_name, line);
+	int *pcpu;
+
+	pcpu = get_cpu_ptr(&fpu_recursion_depth);
+	*pcpu += 1;
  
+	if (*pcpu == 1) {
  #if defined(CONFIG_X86)
-	kernel_fpu_begin();
+		kernel_fpu_begin();
  #elif defined(CONFIG_PPC64)
-	if (cpu_has_feature(CPU_FTR_VSX_COMP)) {
-		preempt_disable();
-		enable_kernel_vsx();
-	} else if (cpu_has_feature(CPU_FTR_ALTIVEC_COMP)) {
-		preempt_disable();
-		enable_kernel_altivec();
-	} else if (!cpu_has_feature(CPU_FTR_FPU_UNAVAILABLE)) {
-		preempt_disable();
-		enable_kernel_fp();
-	}
+		if (cpu_has_feature(CPU_FTR_VSX_COMP)) {
+			preempt_disable();
+			enable_kernel_vsx();
+		} else if (cpu_has_feature(CPU_FTR_ALTIVEC_COMP)) {
+			preempt_disable();
+			enable_kernel_altivec();
+		} else if (!cpu_has_feature(CPU_FTR_FPU_UNAVAILABLE)) {
+			preempt_disable();
+			enable_kernel_fp();
+		}
  #endif
+	}
+
+	TRACE_DCN_FPU(true, function_name, line, *pcpu);
+	put_cpu_ptr(&fpu_recursion_depth);
  }
  
  /**
@@ -78,19 +99,27 @@ void dc_fpu_begin(const char *function_name, const int line)
   */
  void dc_fpu_end(const char *function_name, const int line)
  {
-	TRACE_DCN_FPU(false, function_name, line);
+	int *pcpu;
+
+	pcpu = get_cpu_ptr(&fpu_recursion_depth);
+	*pcpu -= 1;
+	if (*pcpu <= 0) {
  #if defined(CONFIG_X86)
-	kernel_fpu_end();
+		kernel_fpu_end();
  #elif defined(CONFIG_PPC64)
-	if (cpu_has_feature(CPU_FTR_VSX_COMP)) {
-		disable_kernel_vsx();
-		preempt_enable();
-	} else if (cpu_has_feature(CPU_FTR_ALTIVEC_COMP)) {
-		disable_kernel_altivec();
-		preempt_enable();
-	} else if (!cpu_has_feature(CPU_FTR_FPU_UNAVAILABLE)) {
-		disable_kernel_fp();
-		preempt_enable();
-	}
+		if (cpu_has_feature(CPU_FTR_VSX_COMP)) {
+			disable_kernel_vsx();
+			preempt_enable();
+		} else if (cpu_has_feature(CPU_FTR_ALTIVEC_COMP)) {
+			disable_kernel_altivec();
+			preempt_enable();
+		} else if (!cpu_has_feature(CPU_FTR_FPU_UNAVAILABLE)) {
+			disable_kernel_fp();
+			preempt_enable();
+		}
  #endif
+	}
+
+	TRACE_DCN_FPU(false, function_name, line, *pcpu);
+	put_cpu_ptr(&fpu_recursion_depth);
  }
diff --git a/drivers/gpu/drm/amd/display/dc/dc_trace.h b/drivers/gpu/drm/amd/display/dc/dc_trace.h
index d598ba697e45..c711797e5c9e 100644
--- a/drivers/gpu/drm/amd/display/dc/dc_trace.h
+++ b/drivers/gpu/drm/amd/display/dc/dc_trace.h
@@ -38,5 +38,5 @@
  #define TRACE_DCN_CLOCK_STATE(dcn_clocks) \
  	trace_amdgpu_dm_dc_clocks_state(dcn_clocks)
  
-#define TRACE_DCN_FPU(begin, function, line) \
-	trace_dcn_fpu(begin, function, line)
+#define TRACE_DCN_FPU(begin, function, line, ref_count) \
+	trace_dcn_fpu(begin, function, line, ref_count)







