Skip to contentSkip to conversation

Fix Navi cards crashing with FP exception on 5.6+

PATCH Applied Daniel Kolesa 2 participants 3 messages Apr 29, 2020 – Apr 30, 2020

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Fix Navi cards crashing with FP exception on 5.6+

This is not a real fix, but rather a workaround that should be applied to the 5.6 tree and newer. The idea here is to work around the compiler emitting FP instructions outside the DC_FP_START/DC_FP_END blocks, which it currently does, which causes crashes at least on ppc64le.

A proper solution would be to move all the floating point code into its own files, then compile those with hard-float and strictly wrap all calls into those with the DC_FP_START/DC_FP_END blocks, with the rest of the code being compiled without floating point, but that's too extensive of a change to do, and would not be possible in a stable tree.

The proper solution is already being discussed, it seems:

https://lore.kernel.org/lkml/CAG48ez2Sx4ELkM94aD_h_J7K7KBOeuGmvZLKRkg3n_f2WoZ_cg@mail.gmail.com/

Daniel Kolesa (1):
drm/amd/display: work around fp code being emitted outside of
DC_FP_START/END

 .../drm/amd/display/dc/dcn20/dcn20_resource.c | 31 ++++++++++++++-----
 1 file changed, 23 insertions(+), 8 deletions(-)
Signature · 6 lines
2.26.2

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#1 lore raw

PATCH 1/1 drm/amd/display: work around fp code being emitted outside of DC_FP_START/END

The dcn20_validate_bandwidth function would have code touching the incorrect registers emitted outside of the boundaries of the DC_FP_START/END macros, at least on ppc64le. Work around the problem by wrapping the whole function instead.

---
 .../drm/amd/display/dc/dcn20/dcn20_resource.c | 31 ++++++++++++++-----
 1 file changed, 23 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/display/dc/dcn20/dcn20_resource.c +23 -8
index e310d67..1b0bca9 100644
@@ -3034,25 +3034,32 @@ static bool dcn20_validate_bandwidth_internal(struct dc *dc, struct dc_state *co
 	return out;
 }
 
-
-bool dcn20_validate_bandwidth(struct dc *dc, struct dc_state *context,
-		bool fast_validate)
+/*
+ * This must be noinline to ensure anything that deals with FP registers
+ * is contained within this call; previously our compiling with hard-float
+ * would result in fp instructions being emitted outside of the boundaries
+ * of the DC_FP_START/END macros, which makes sense as the compiler has no
+ * idea about what is wrapped and what is not
+ *
+ * This is largely just a workaround to avoid breakage introduced with 5.6,
+ * ideally all fp-using code should be moved into its own file, only that
+ * should be compiled with hard-float, and all code exported from there
+ * should be strictly wrapped with DC_FP_START/END
+ */
+static noinline bool dcn20_validate_bandwidth_fp(struct dc *dc,
+		struct dc_state *context, bool fast_validate)
 {
 	bool voltage_supported = false;
 	bool full_pstate_supported = false;
 	bool dummy_pstate_supported = false;
 	double p_state_latency_us;
 
-	DC_FP_START();
 	p_state_latency_us = context->bw_ctx.dml.soc.dram_clock_change_latency_us;
 	context->bw_ctx.dml.soc.disable_dram_clock_change_vactive_support =
 		dc->debug.disable_dram_clock_change_vactive_support;
 
 	if (fast_validate) {
-		voltage_supported = dcn20_validate_bandwidth_internal(dc, context, true);
-
-		DC_FP_END();
-		return voltage_supported;
+		return dcn20_validate_bandwidth_internal(dc, context, true);
 	}
 
 	// Best case, we support full UCLK switch latency
@@ -3081,7 +3088,15 @@ bool dcn20_validate_bandwidth(struct dc *dc, struct dc_state *context,
 
 restore_dml_state:
 	context->bw_ctx.dml.soc.dram_clock_change_latency_us = p_state_latency_us;
+	return voltage_supported;
+}
 
+bool dcn20_validate_bandwidth(struct dc *dc, struct dc_state *context,
+		bool fast_validate)
+{
+	bool voltage_supported = false;
+	DC_FP_START();
+	voltage_supported = dcn20_validate_bandwidth_fp(dc, context, fast_validate);
 	DC_FP_END();
 	return voltage_supported;
 }
Signature · 6 lines
2.26.2

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Applied
#2 lore raw

Re: [PATCH 1/1] drm/amd/display: work around fp code being emitted outside of DC_FP_START/END

Quoted 7 lines · Daniel Kolesa

The dcn20_validate_bandwidth function would have code touching the incorrect registers emitted outside of the boundaries of the DC_FP_START/END macros, at least on ppc64le. Work around the problem by wrapping the whole function instead.

Applied. Thanks!

Alex

Quoted 68 lines
---
 .../drm/amd/display/dc/dcn20/dcn20_resource.c | 31 ++++++++++++++-----
 1 file changed, 23 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/display/dc/dcn20/dcn20_resource.c +23 -9
index e310d67..1b0bca9 100644
@@ -3034,25 +3034,32 @@ static bool dcn20_validate_bandwidth_internal(struct dc *dc, struct dc_state *co
        return out;
 }
 
-
-bool dcn20_validate_bandwidth(struct dc *dc, struct dc_state *context,
-               bool fast_validate)
+/*
+ * This must be noinline to ensure anything that deals with FP registers
+ * is contained within this call; previously our compiling with hard-float
+ * would result in fp instructions being emitted outside of the boundaries
+ * of the DC_FP_START/END macros, which makes sense as the compiler has no
+ * idea about what is wrapped and what is not
+ *
+ * This is largely just a workaround to avoid breakage introduced with 5.6,
+ * ideally all fp-using code should be moved into its own file, only that
+ * should be compiled with hard-float, and all code exported from there
+ * should be strictly wrapped with DC_FP_START/END
+ */
+static noinline bool dcn20_validate_bandwidth_fp(struct dc *dc,
+               struct dc_state *context, bool fast_validate)
 {
        bool voltage_supported = false;
        bool full_pstate_supported = false;
        bool dummy_pstate_supported = false;
        double p_state_latency_us;
 
-       DC_FP_START();
        p_state_latency_us = context->bw_ctx.dml.soc.dram_clock_change_latency_us;
        context->bw_ctx.dml.soc.disable_dram_clock_change_vactive_support =
                dc->debug.disable_dram_clock_change_vactive_support;
 
        if (fast_validate) {
-               voltage_supported = dcn20_validate_bandwidth_internal(dc, context, true);
-
-               DC_FP_END();
-               return voltage_supported;
+               return dcn20_validate_bandwidth_internal(dc, context, true);
        }
 
        // Best case, we support full UCLK switch latency
@@ -3081,7 +3088,15 @@ bool dcn20_validate_bandwidth(struct dc *dc, struct dc_state *context,
 
 restore_dml_state:
        context->bw_ctx.dml.soc.dram_clock_change_latency_us = p_state_latency_us;
+       return voltage_supported;
+}
 
+bool dcn20_validate_bandwidth(struct dc *dc, struct dc_state *context,
+               bool fast_validate)
+{
+       bool voltage_supported = false;
+       DC_FP_START();
+       voltage_supported = dcn20_validate_bandwidth_fp(dc, context, fast_validate);
        DC_FP_END();
        return voltage_supported;
 }
--

2.26.2

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx