Skip to contentSkip to conversation

perf kvm/riscv: Fix event key collision between interrupts and exceptions

PATCH New [email protected] 2 participants 2 messages Jul 4

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf kvm/riscv: Fix event key collision between interrupts and exceptions

From: Fangyu Yu <[email protected]>

RISC-V scause encodes interrupts and exceptions in the same cause field, with bit 63 on RV64 and bit 31 on RV32 used to distinguish interrupts from exceptions. The lower cause bits overlap between the two namespaces, so masking off the interrupt bit makes distinct IRQ and exception events map to the same key.

This affects perf kvm stat exit-reason decoding. For example, IRQ_S_TIMER and EXC_LOAD_ACCESS both use cause code 5, and can be misreported as the wrong exit reason.

Fix this by using the full scause value as the lookup key in event_get_key(). Also split TRAP() into TRAP_EXC() and TRAP_IRQ() so interrupt entries are encoded with the architecture-specific IRQ bit set, matching the value reported by hardware.

Fixes ceea279f9376 ("perf kvm stat: Remove use of the arch directory")
---
 .../perf/util/kvm-stat-arch/kvm-stat-riscv.c  |  4 +--
 .../util/kvm-stat-arch/riscv_trap_types.h     | 27 ++++++++++---------
 2 files changed, 15 insertions(+), 16 deletions(-)
tools/perf/util/kvm-stat-arch/kvm-stat-riscv.c +1 -3
index 967bba261a47..829096a41a3b 100644
@@ -23,10 +23,8 @@ static const char * const __kvm_events_tp[] = {
 static void event_get_key(struct perf_sample *sample,
 			  struct event_key *key)
 {
-	int xlen = 64; // TODO: 32-bit support.
-
 	key->info = 0;
-	key->key = perf_sample__intval(sample, kvm_exit_reason(EM_RISCV)) & ~CAUSE_IRQ_FLAG(xlen);
+	key->key = perf_sample__intval(sample, kvm_exit_reason(EM_RISCV));
 	key->exit_reasons = riscv_exit_reasons;
 }
 
tools/perf/util/kvm-stat-arch/riscv_trap_types.h +14 -13
index aa5d24fab4ee..da1f1ce55326 100644
@@ -38,20 +38,21 @@
 #define EXC_VIRTUAL_INST_FAULT 22
 #define EXC_STORE_GUEST_PAGE_FAULT 23
 
-#define TRAP(x) { x, #x }
+#define TRAP_EXC(x) { x, #x }
+#define TRAP_IRQ(x) { (x) | CAUSE_IRQ_FLAG(__riscv_xlen), #x }
 
 #define kvm_riscv_trap_class \
-	TRAP(IRQ_S_SOFT), TRAP(IRQ_VS_SOFT), TRAP(IRQ_M_SOFT), \
-	TRAP(IRQ_S_TIMER), TRAP(IRQ_VS_TIMER), TRAP(IRQ_M_TIMER), \
-	TRAP(IRQ_S_EXT), TRAP(IRQ_VS_EXT), TRAP(IRQ_M_EXT), \
-	TRAP(IRQ_S_GEXT), TRAP(IRQ_PMU_OVF), \
-	TRAP(EXC_INST_MISALIGNED), TRAP(EXC_INST_ACCESS), TRAP(EXC_INST_ILLEGAL), \
-	TRAP(EXC_BREAKPOINT), TRAP(EXC_LOAD_MISALIGNED), TRAP(EXC_LOAD_ACCESS), \
-	TRAP(EXC_STORE_MISALIGNED), TRAP(EXC_STORE_ACCESS), TRAP(EXC_SYSCALL), \
-	TRAP(EXC_HYPERVISOR_SYSCALL), TRAP(EXC_SUPERVISOR_SYSCALL), \
-	TRAP(EXC_INST_PAGE_FAULT), TRAP(EXC_LOAD_PAGE_FAULT), \
-	TRAP(EXC_STORE_PAGE_FAULT), TRAP(EXC_INST_GUEST_PAGE_FAULT), \
-	TRAP(EXC_LOAD_GUEST_PAGE_FAULT), TRAP(EXC_VIRTUAL_INST_FAULT), \
-	TRAP(EXC_STORE_GUEST_PAGE_FAULT)
+	TRAP_IRQ(IRQ_S_SOFT), TRAP_IRQ(IRQ_VS_SOFT), TRAP_IRQ(IRQ_M_SOFT), \
+	TRAP_IRQ(IRQ_S_TIMER), TRAP_IRQ(IRQ_VS_TIMER), TRAP_IRQ(IRQ_M_TIMER), \
+	TRAP_IRQ(IRQ_S_EXT), TRAP_IRQ(IRQ_VS_EXT), TRAP_IRQ(IRQ_M_EXT), \
+	TRAP_IRQ(IRQ_S_GEXT), TRAP_IRQ(IRQ_PMU_OVF), \
+	TRAP_EXC(EXC_INST_MISALIGNED), TRAP_EXC(EXC_INST_ACCESS), TRAP_EXC(EXC_INST_ILLEGAL), \
+	TRAP_EXC(EXC_BREAKPOINT), TRAP_EXC(EXC_LOAD_MISALIGNED), TRAP_EXC(EXC_LOAD_ACCESS), \
+	TRAP_EXC(EXC_STORE_MISALIGNED), TRAP_EXC(EXC_STORE_ACCESS), TRAP_EXC(EXC_SYSCALL), \
+	TRAP_EXC(EXC_HYPERVISOR_SYSCALL), TRAP_EXC(EXC_SUPERVISOR_SYSCALL), \
+	TRAP_EXC(EXC_INST_PAGE_FAULT), TRAP_EXC(EXC_LOAD_PAGE_FAULT), \
+	TRAP_EXC(EXC_STORE_PAGE_FAULT), TRAP_EXC(EXC_INST_GUEST_PAGE_FAULT), \
+	TRAP_EXC(EXC_LOAD_GUEST_PAGE_FAULT), TRAP_EXC(EXC_VIRTUAL_INST_FAULT), \
+	TRAP_EXC(EXC_STORE_GUEST_PAGE_FAULT)
 
 #endif /* ARCH_PERF_RISCV_TRAP_TYPES_H */
Signature · 7 lines
2.50.1


_______________________________________________
linux-riscv mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-riscv
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [Medium] Using the compiler predefined macro `__riscv_xlen` in a static array initializer causes build…

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] Using the compiler predefined macro __riscv_xlen in a static array initializer causes build failures on non-RISC-V architectures and breaks cross-architecture analysis. --

commit e5247b904831a2084b1e8757d98d388a7fc008c0
Author: Fangyu Yu <[email protected]>

perf kvm/riscv: Fix event key collision between interrupts and exceptions

This commit updates the RISC-V KVM stat exit-reason decoding to use the full scause value, avoiding key collisions between interrupts and exceptions. It splits the TRAP macro into TRAP_EXC and TRAP_IRQ to encode interrupts with the architecture-specific IRQ bit matching the hardware value.

tools/perf/util/kvm-stat-arch/kvm-stat-riscv.c static const char * const __kvm_events_tp[] = { · line 27

Quoted 13 lines
tools/perf/util/kvm-stat-arch/kvm-stat-riscv.c +1 -3
index 967bba261a475..829096a41a3b2 100644
@@ -23,10 +23,8 @@ static const char * const __kvm_events_tp[] = {
 static void event_get_key(struct perf_sample *sample,
 			  struct event_key *key)
 {
-	int xlen = 64; // TODO: 32-bit support.
-
 	key->info = 0;
-	key->key = perf_sample__intval(sample, kvm_exit_reason(EM_RISCV)) & ~CAUSE_IRQ_FLAG(xlen);
+	key->key = perf_sample__intval(sample, kvm_exit_reason(EM_RISCV));

[Severity: Medium]
Does this code introduce a regression by preventing dynamic cross-architecture bitness decoding when analyzing traces? With the dynamic mask removed here, the lookup now depends entirely on the statically sized table.

[ ... ]

tools/perf/util/kvm-stat-arch/riscv_trap_types.h line 42

Quoted 11 lines
tools/perf/util/kvm-stat-arch/riscv_trap_types.h +2 -1
index aa5d24fab4ee3..da1f1ce55326c 100644
@@ -38,20 +38,21 @@
 #define EXC_VIRTUAL_INST_FAULT 22
 #define EXC_STORE_GUEST_PAGE_FAULT 23
 
-#define TRAP(x) { x, #x }
+#define TRAP_EXC(x) { x, #x }
+#define TRAP_IRQ(x) { (x) | CAUSE_IRQ_FLAG(__riscv_xlen), #x }

[Severity: Medium]
Does using the compiler predefined macro __riscv_xlen in this static array initializer cause a regression by breaking the build on non-RISC-V architectures?

When building perf cross-platform, this macro is not defined on hosts like x86_64 or ARM64, and the static initialization in riscv_exit_reasons expands this macro.

Also, does baking the host's compile-time word size into TRAP_IRQ introduce a regression for cross-platform analysis? For example, if a 64-bit perf binary processes a 32-bit KVM trace, it will expect bit 63 to be set for interrupts, which would fail to decode 32-bit KVM events using bit 31.

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1