Skip to contentSkip to conversation

perf/x86: make perf callchain work without CONFIG_FRAME_POINTER

RFC PATCH Superseded Kairui Song 5 participants 14 messages Apr 8, 2019 – Apr 18, 2019

Subsystems (2)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

RFC PATCH v2 1/1 perf/x86: make perf callchain work without CONFIG_FRAME_POINTER

Currently perf callchain is not working properly with ORC unwinder, and sampling event from trace point. We'll get useless in kernel callchain like this:

perf 6429 [000] 22.498450: kmem:mm_page_alloc: page=0x176a17 pfn=1534487 order=0 migratetype=0 gfp_flags=GFP_KERNEL

    ffffffffbe23e32e __alloc_pages_nodemask+0x22e (/lib/modules/5.1.0-rc3+/build/vmlinux)
	7efdf7f7d3e8 __poll+0x18 (/usr/lib64/libc-2.28.so)
	5651468729c1 [unknown] (/usr/bin/perf)
	5651467ee82a main+0x69a (/usr/bin/perf)
	7efdf7eaf413 __libc_start_main+0xf3 (/usr/lib64/libc-2.28.so)
    5541f689495641d7 [unknown] ([unknown])

The root cause is within a trace point perf will try to dump the caller's register, but without CONFIG_FRAME_POINTER we can't get caller's BP as the frame pointer, so current frame pointer is returned instead. We get a register combination of caller IP and current BP, which confuse the unwinder and end the stacktrace early.

So in such case don't dump BP, and just let the unwinder start directly and skip until we reached the stack we wanted.

This make the callchain get the full kernel space stacktrace again:

perf 6503 [000] 1567.570191: kmem:mm_page_alloc: page=0x16c904 pfn=1493252 order=0 migratetype=0 gfp_flags=GFP_KERNEL

    ffffffffb523e2ae __alloc_pages_nodemask+0x22e (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52383bd __get_free_pages+0xd (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52fd28a __pollwait+0x8a (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb521426f perf_poll+0x2f (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52fe3e2 do_sys_poll+0x252 (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52ff027 __x64_sys_poll+0x37 (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb500418b do_syscall_64+0x5b (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb5a0008c entry_SYSCALL_64_after_hwframe+0x44 (/lib/modules/5.1.0-rc3+/build/vmlinux)
	7f71e92d03e8 __poll+0x18 (/usr/lib64/libc-2.28.so)
	55a22960d9c1 [unknown] (/usr/bin/perf)
	55a22958982a main+0x69a (/usr/bin/perf)
	7f71e9202413 __libc_start_main+0xf3 (/usr/lib64/libc-2.28.so)
    5541f689495641d7 [unknown] ([unknown])

---

Update from V1:
Get rid of a lot of unneccessary code and just don't dump a inaccurate BP, and use SP as the marker for target frame.

 arch/x86/events/core.c            | 18 +++++++++++++++---
 arch/x86/include/asm/stacktrace.h |  9 +++++++--
 2 files changed, 22 insertions(+), 5 deletions(-)
arch/x86/events/core.c +15 -3
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
 	cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+	return (regs->ip && regs->bp && regs->sp);
+}
+
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
 		return;
 	}
 
-	if (perf_callchain_store(entry, regs->ip))
+	if (valid_perf_registers(regs)) {
+		if (perf_callchain_store(entry, regs->ip))
+			return;
+		unwind_start(&state, current, regs, NULL);
+	} else if (regs->sp) {
+		unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+	} else {
 		return;
+	}
 
-	for (unwind_start(&state, current, regs, NULL); !unwind_done(&state);
-	     unwind_next_frame(&state)) {
+	for (; !unwind_done(&state); unwind_next_frame(&state)) {
 		addr = unwind_get_return_address(&state);
 		if (!addr || perf_callchain_store(entry, addr))
 			return;
arch/x86/include/asm/stacktrace.h +7 -2
index f335aad404a4..226077e20412 100644
@@ -98,18 +98,23 @@ struct stack_frame_ia32 {
     u32 return_address;
 };
 
+#ifdef CONFIG_FRAME_POINTER
 static inline unsigned long caller_frame_pointer(void)
 {
 	struct stack_frame *frame;
 
 	frame = __builtin_frame_address(0);
 
-#ifdef CONFIG_FRAME_POINTER
 	frame = frame->next_frame;
-#endif
 
 	return (unsigned long)frame;
 }
+#else
+static inline unsigned long caller_frame_pointer(void)
+{
+	return 0;
+}
+#endif
 
 void show_opcodes(struct pt_regs *regs, const char *loglvl);
 void show_ip(struct pt_regs *regs, const char *loglvl);
Signature · 1 line
2.20.1
Tested-by on v2 1/1
#1 lore raw
Quoted 38 lines · Kairui Song

Currently perf callchain is not working properly with ORC unwinder, and sampling event from trace point. We'll get useless in kernel callchain like this:

perf 6429 [000] 22.498450: kmem:mm_page_alloc: page=0x176a17 pfn=1534487 order=0 migratetype=0 gfp_flags=GFP_KERNEL

    ffffffffbe23e32e __alloc_pages_nodemask+0x22e (/lib/modules/5.1.0-rc3+/build/vmlinux)
	7efdf7f7d3e8 __poll+0x18 (/usr/lib64/libc-2.28.so)
	5651468729c1 [unknown] (/usr/bin/perf)
	5651467ee82a main+0x69a (/usr/bin/perf)
	7efdf7eaf413 __libc_start_main+0xf3 (/usr/lib64/libc-2.28.so)
    5541f689495641d7 [unknown] ([unknown])

The root cause is within a trace point perf will try to dump the caller's register, but without CONFIG_FRAME_POINTER we can't get caller's BP as the frame pointer, so current frame pointer is returned instead. We get a register combination of caller IP and current BP, which confuse the unwinder and end the stacktrace early.

So in such case don't dump BP, and just let the unwinder start directly and skip until we reached the stack we wanted.

This make the callchain get the full kernel space stacktrace again:

perf 6503 [000] 1567.570191: kmem:mm_page_alloc: page=0x16c904 pfn=1493252 order=0 migratetype=0 gfp_flags=GFP_KERNEL

    ffffffffb523e2ae __alloc_pages_nodemask+0x22e (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52383bd __get_free_pages+0xd (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52fd28a __pollwait+0x8a (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb521426f perf_poll+0x2f (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52fe3e2 do_sys_poll+0x252 (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb52ff027 __x64_sys_poll+0x37 (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb500418b do_syscall_64+0x5b (/lib/modules/5.1.0-rc3+/build/vmlinux)
    ffffffffb5a0008c entry_SYSCALL_64_after_hwframe+0x44 (/lib/modules/5.1.0-rc3+/build/vmlinux)
	7f71e92d03e8 __poll+0x18 (/usr/lib64/libc-2.28.so)
	55a22960d9c1 [unknown] (/usr/bin/perf)
	55a22958982a main+0x69a (/usr/bin/perf)
	7f71e9202413 __libc_start_main+0xf3 (/usr/lib64/libc-2.28.so)
    5541f689495641d7 [unknown] ([unknown])
Quoted 1 line

thanks,
jirka

Quoted 81 lines

---

Update from V1:
Get rid of a lot of unneccessary code and just don't dump a inaccurate BP, and use SP as the marker for target frame.

 arch/x86/events/core.c            | 18 +++++++++++++++---
 arch/x86/include/asm/stacktrace.h |  9 +++++++--
 2 files changed, 22 insertions(+), 5 deletions(-)
arch/x86/events/core.c +15 -3
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
 	cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+	return (regs->ip && regs->bp && regs->sp);
+}
+
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
 		return;
 	}
 
-	if (perf_callchain_store(entry, regs->ip))
+	if (valid_perf_registers(regs)) {
+		if (perf_callchain_store(entry, regs->ip))
+			return;
+		unwind_start(&state, current, regs, NULL);
+	} else if (regs->sp) {
+		unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+	} else {
 		return;
+	}
 
-	for (unwind_start(&state, current, regs, NULL); !unwind_done(&state);
-	     unwind_next_frame(&state)) {
+	for (; !unwind_done(&state); unwind_next_frame(&state)) {
 		addr = unwind_get_return_address(&state);
 		if (!addr || perf_callchain_store(entry, addr))
 			return;
arch/x86/include/asm/stacktrace.h +7 -2
index f335aad404a4..226077e20412 100644
@@ -98,18 +98,23 @@ struct stack_frame_ia32 {
     u32 return_address;
 };
 
+#ifdef CONFIG_FRAME_POINTER
 static inline unsigned long caller_frame_pointer(void)
 {
 	struct stack_frame *frame;
 
 	frame = __builtin_frame_address(0);
 
-#ifdef CONFIG_FRAME_POINTER
 	frame = frame->next_frame;
-#endif
 
 	return (unsigned long)frame;
 }
+#else
+static inline unsigned long caller_frame_pointer(void)
+{
+	return 0;
+}
+#endif
 
 void show_opcodes(struct pt_regs *regs, const char *loglvl);
 void show_ip(struct pt_regs *regs, const char *loglvl);
Signature · 1 line
2.20.1
maintainer
#2 lore raw
In reply to Kairui Song

I'll mostly defer to Josh on unwinding, but a few comments below.

arch/x86/events/core.c void arch_perf_update_userpage(struct perf_event *event, · line 2362

Quoted 13 lines · Kairui Song
arch/x86/events/core.c +5 -0
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
 	cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+	return (regs->ip && regs->bp && regs->sp);
+}

I'm unconvinced by this, with both guess and orc having !bp is perfectly valid.

Quoted 7 lines
arch/x86/events/core.c +0 -0
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
 		return;
 	}
 

arch/x86/events/core.c perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re · line 2383

Quoted 10 lines
arch/x86/events/core.c +8 -1
-	if (perf_callchain_store(entry, regs->ip))
+	if (valid_perf_registers(regs)) {
+		if (perf_callchain_store(entry, regs->ip))
+			return;
+		unwind_start(&state, current, regs, NULL);
+	} else if (regs->sp) {
+		unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+	} else {
 		return;
+	}

AFAICT if we, by pure accident, end up with !bp for ORC, then we initialize the unwind wrong.

Note that @regs is mostly trivially correct, except for that tracepoint case. So I don't think we should magic here.

Quoted 25 lines
arch/x86/events/core.c +1 -2
-	for (unwind_start(&state, current, regs, NULL); !unwind_done(&state);
-	     unwind_next_frame(&state)) {
+	for (; !unwind_done(&state); unwind_next_frame(&state)) {
 		addr = unwind_get_return_address(&state);
 		if (!addr || perf_callchain_store(entry, addr))
 			return;
 
arch/x86/include/asm/stacktrace.h +1 -1
index f335aad404a4..226077e20412 100644
@@ -98,18 +98,23 @@ struct stack_frame_ia32 {
     u32 return_address;
 };
 
+#ifdef CONFIG_FRAME_POINTER
 static inline unsigned long caller_frame_pointer(void)
 {
 	struct stack_frame *frame;
 
 	frame = __builtin_frame_address(0);
 
-#ifdef CONFIG_FRAME_POINTER
 	frame = frame->next_frame;

arch/x86/include/asm/stacktrace.h struct stack_frame_ia32 { · line 117

Quoted 10 lines
arch/x86/include/asm/stacktrace.h +6 -1
-#endif
 
 	return (unsigned long)frame;
 }
+#else
+static inline unsigned long caller_frame_pointer(void)
+{
+	return 0;
+}
+#endif

OK, that makes sense I guess.

#3 lore raw
Quoted 42 lines · Peter Zijlstra

I'll mostly defer to Josh on unwinding, but a few comments below.

arch/x86/events/core.c void arch_perf_update_userpage(struct perf_event *event, · line 2362

Quoted 13 lines · Kairui Song
arch/x86/events/core.c +5 -0
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
 	cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+	return (regs->ip && regs->bp && regs->sp);
+}

I'm unconvinced by this, with both guess and orc having !bp is perfectly valid.

Quoted 7 lines
arch/x86/events/core.c +0 -0
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
 		return;
 	}
 

arch/x86/events/core.c perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re · line 2383

Quoted 10 lines
arch/x86/events/core.c +8 -1
-	if (perf_callchain_store(entry, regs->ip))
+	if (valid_perf_registers(regs)) {
+		if (perf_callchain_store(entry, regs->ip))
+			return;
+		unwind_start(&state, current, regs, NULL);
+	} else if (regs->sp) {
+		unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+	} else {
 		return;
+	}

AFAICT if we, by pure accident, end up with !bp for ORC, then we initialize the unwind wrong.

Quoted 2 lines

Note that @regs is mostly trivially correct, except for that tracepoint case. So I don't think we should magic here.

Ah, I didn't quite understand this code before, and I still don't really, but I guess the issue is that @regs can be either real or fake.

In the real @regs case, we just want to always unwind starting from regs->sp.

But in the fake @regs case, we should instead unwind from the current frame, skipping all frames until we hit the fake regs->sp. Because starting from fake/incomplete regs is most likely going to cause problems with ORC (or DWARF for other arches).

The idea of a fake regs is fragile and confusing. Is it possible to just pass in the "skip" stack pointer directly instead? That should work for both FP and non-FP. And I _think_ there's no need to ever capture regs->bp anyway -- the stack pointer should be sufficient.

In other words, either regs should be "real", and skip_sp is NULL; or regs should be NULL and skip_sp should have a value.

Signature · 1 line
Josh
#4 lore raw
Quoted 58 lines · Josh Poimboeuf
Quoted 42 lines · Peter Zijlstra

I'll mostly defer to Josh on unwinding, but a few comments below.

arch/x86/events/core.c void arch_perf_update_userpage(struct perf_event *event, · line 2362

Quoted 13 lines · Kairui Song
arch/x86/events/core.c +5 -0
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
    cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+   return (regs->ip && regs->bp && regs->sp);
+}

I'm unconvinced by this, with both guess and orc having !bp is perfectly valid.

Quoted 7 lines
arch/x86/events/core.c +0 -0
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
            return;
    }
 

arch/x86/events/core.c perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re · line 2383

Quoted 10 lines
arch/x86/events/core.c +8 -1
-   if (perf_callchain_store(entry, regs->ip))
+   if (valid_perf_registers(regs)) {
+           if (perf_callchain_store(entry, regs->ip))
+                   return;
+           unwind_start(&state, current, regs, NULL);
+   } else if (regs->sp) {
+           unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+   } else {
            return;
+   }

AFAICT if we, by pure accident, end up with !bp for ORC, then we initialize the unwind wrong.

Quoted 2 lines

Note that @regs is mostly trivially correct, except for that tracepoint case. So I don't think we should magic here.

Ah, I didn't quite understand this code before, and I still don't really, but I guess the issue is that @regs can be either real or fake.

In the real @regs case, we just want to always unwind starting from regs->sp.

But in the fake @regs case, we should instead unwind from the current frame, skipping all frames until we hit the fake regs->sp. Because starting from fake/incomplete regs is most likely going to cause problems with ORC (or DWARF for other arches).

Quoted 4 lines

The idea of a fake regs is fragile and confusing. Is it possible to just pass in the "skip" stack pointer directly instead? That should work for both FP and non-FP. And I _think_ there's no need to ever capture regs->bp anyway -- the stack pointer should be sufficient.

Hi, that will break some other usage, if perf_callchain_kernel is called but it won't unwind to the callsite (could be produced by attach an ebpf call to kprobe), things will also go wrong. It should start with given registers when the register is valid.
And it's true with omit frame pointer BP value could be anything, so 0 is also valid, I think I need to find a better way to tell if we could start with the registers value or direct start unwinding and skip until got the stack.

Quoted 6 lines

In other words, either regs should be "real", and skip_sp is NULL; or regs should be NULL and skip_sp should have a value.

--
Josh

--
Best Regards,
Kairui Song

#5 lore raw
Quoted 64 lines · Kairui Song
Quoted 58 lines · Josh Poimboeuf
Quoted 42 lines · Peter Zijlstra

I'll mostly defer to Josh on unwinding, but a few comments below.

arch/x86/events/core.c void arch_perf_update_userpage(struct perf_event *event, · line 2362

Quoted 13 lines · Kairui Song
arch/x86/events/core.c +5 -0
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
    cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+   return (regs->ip && regs->bp && regs->sp);
+}

I'm unconvinced by this, with both guess and orc having !bp is perfectly valid.

Quoted 7 lines
arch/x86/events/core.c +0 -0
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
            return;
    }
 

arch/x86/events/core.c perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re · line 2383

Quoted 10 lines
arch/x86/events/core.c +8 -1
-   if (perf_callchain_store(entry, regs->ip))
+   if (valid_perf_registers(regs)) {
+           if (perf_callchain_store(entry, regs->ip))
+                   return;
+           unwind_start(&state, current, regs, NULL);
+   } else if (regs->sp) {
+           unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+   } else {
            return;
+   }

AFAICT if we, by pure accident, end up with !bp for ORC, then we initialize the unwind wrong.

Quoted 2 lines

Note that @regs is mostly trivially correct, except for that tracepoint case. So I don't think we should magic here.

Ah, I didn't quite understand this code before, and I still don't really, but I guess the issue is that @regs can be either real or fake.

In the real @regs case, we just want to always unwind starting from regs->sp.

But in the fake @regs case, we should instead unwind from the current frame, skipping all frames until we hit the fake regs->sp. Because starting from fake/incomplete regs is most likely going to cause problems with ORC (or DWARF for other arches).

Quoted 4 lines

The idea of a fake regs is fragile and confusing. Is it possible to just pass in the "skip" stack pointer directly instead? That should work for both FP and non-FP. And I _think_ there's no need to ever capture regs->bp anyway -- the stack pointer should be sufficient.

Quoted 8 lines

Hi, that will break some other usage, if perf_callchain_kernel is called but it won't unwind to the callsite (could be produced by attach an ebpf call to kprobe), things will also go wrong. It should start with given registers when the register is valid.
And it's true with omit frame pointer BP value could be anything, so 0 is also valid, I think I need to find a better way to tell if we could start with the registers value or direct start unwinding and skip until got the stack.

Hey, what is the status of the fix?
We're hitting it from bpf as well.
kernel stack traces are not working in tracepoints.

#6 lore raw
In reply to Kairui Song
Quoted 65 lines · Kairui Song
Quoted 58 lines · Josh Poimboeuf
Quoted 42 lines · Peter Zijlstra

I'll mostly defer to Josh on unwinding, but a few comments below.

arch/x86/events/core.c void arch_perf_update_userpage(struct perf_event *event, · line 2362

Quoted 13 lines · Kairui Song
arch/x86/events/core.c +5 -0
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
    cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+   return (regs->ip && regs->bp && regs->sp);
+}

I'm unconvinced by this, with both guess and orc having !bp is perfectly valid.

Quoted 7 lines
arch/x86/events/core.c +0 -0
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
            return;
    }
 

arch/x86/events/core.c perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re · line 2383

Quoted 10 lines
arch/x86/events/core.c +8 -1
-   if (perf_callchain_store(entry, regs->ip))
+   if (valid_perf_registers(regs)) {
+           if (perf_callchain_store(entry, regs->ip))
+                   return;
+           unwind_start(&state, current, regs, NULL);
+   } else if (regs->sp) {
+           unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+   } else {
            return;
+   }

AFAICT if we, by pure accident, end up with !bp for ORC, then we initialize the unwind wrong.

Quoted 2 lines

Note that @regs is mostly trivially correct, except for that tracepoint case. So I don't think we should magic here.

Ah, I didn't quite understand this code before, and I still don't really, but I guess the issue is that @regs can be either real or fake.

In the real @regs case, we just want to always unwind starting from regs->sp.

But in the fake @regs case, we should instead unwind from the current frame, skipping all frames until we hit the fake regs->sp. Because starting from fake/incomplete regs is most likely going to cause problems with ORC (or DWARF for other arches).

Quoted 4 lines

The idea of a fake regs is fragile and confusing. Is it possible to just pass in the "skip" stack pointer directly instead? That should work for both FP and non-FP. And I _think_ there's no need to ever capture regs->bp anyway -- the stack pointer should be sufficient.

Quoted 9 lines

Hi, that will break some other usage, if perf_callchain_kernel is called but it won't unwind to the callsite (could be produced by attach an ebpf call to kprobe), things will also go wrong. It should start with given registers when the register is valid.
And it's true with omit frame pointer BP value could be anything, so 0 is also valid, I think I need to find a better way to tell if we could start with the registers value or direct start unwinding and skip until got the stack.

Hi, sorry I might have some misunderstanding. Adding an extra argument (eg. skip_sp) to indicate if it should just unwind from the current frame, and use SP as the "skip mark", should work well.

And I also think the "fake"/"real" reg is fragile, could we abuse another eflag (just like PERF_EFLAGS_EXACT) to indicate the regs are partially dumped fake registers? So perf_callchain_kernel just check if it's a "partial registers", and in such case it can start unwinding and skip until it get to SP. This make it easier to tell if the registers are "fake".

Signature · 2 lines
Best Regards,
Kairui Song
maintainer
#7 lore raw
Quoted 3 lines · Kairui Song

And I also think the "fake"/"real" reg is fragile, could we abuse another eflag (just like PERF_EFLAGS_EXACT) to indicate the regs are partially dumped fake registers?

Sure, the SDM seems to suggest bits 1,3,5,15 are 'available'. We've already used 3 and 5, and I think we can use !X86_EFLAGS_FIXED to indicate a fake regs set. Any real regs set will always have that set.

#8 lore raw
Quoted 9 lines · Peter Zijlstra
Quoted 3 lines · Kairui Song

And I also think the "fake"/"real" reg is fragile, could we abuse another eflag (just like PERF_EFLAGS_EXACT) to indicate the regs are partially dumped fake registers?

Sure, the SDM seems to suggest bits 1,3,5,15 are 'available'. We've already used 3 and 5, and I think we can use !X86_EFLAGS_FIXED to indicate a fake regs set. Any real regs set will always have that set.

Thanks! This is a good idea. Will update accordingly in V3 later.

--
Best Regards,
Kairui Song

#9 lore raw
In reply to Kairui Song
Quoted 80 lines · Kairui Song
Quoted 65 lines · Kairui Song
Quoted 58 lines · Josh Poimboeuf
Quoted 42 lines · Peter Zijlstra

I'll mostly defer to Josh on unwinding, but a few comments below.

arch/x86/events/core.c void arch_perf_update_userpage(struct perf_event *event, · line 2362

Quoted 13 lines · Kairui Song
arch/x86/events/core.c +5 -0
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
    cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+   return (regs->ip && regs->bp && regs->sp);
+}

I'm unconvinced by this, with both guess and orc having !bp is perfectly valid.

Quoted 7 lines
arch/x86/events/core.c +0 -0
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
            return;
    }
 

arch/x86/events/core.c perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re · line 2383

Quoted 10 lines
arch/x86/events/core.c +8 -1
-   if (perf_callchain_store(entry, regs->ip))
+   if (valid_perf_registers(regs)) {
+           if (perf_callchain_store(entry, regs->ip))
+                   return;
+           unwind_start(&state, current, regs, NULL);
+   } else if (regs->sp) {
+           unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+   } else {
            return;
+   }

AFAICT if we, by pure accident, end up with !bp for ORC, then we initialize the unwind wrong.

Quoted 2 lines

Note that @regs is mostly trivially correct, except for that tracepoint case. So I don't think we should magic here.

Ah, I didn't quite understand this code before, and I still don't really, but I guess the issue is that @regs can be either real or fake.

In the real @regs case, we just want to always unwind starting from regs->sp.

But in the fake @regs case, we should instead unwind from the current frame, skipping all frames until we hit the fake regs->sp. Because starting from fake/incomplete regs is most likely going to cause problems with ORC (or DWARF for other arches).

Quoted 4 lines

The idea of a fake regs is fragile and confusing. Is it possible to just pass in the "skip" stack pointer directly instead? That should work for both FP and non-FP. And I _think_ there's no need to ever capture regs->bp anyway -- the stack pointer should be sufficient.

Quoted 9 lines

Hi, that will break some other usage, if perf_callchain_kernel is called but it won't unwind to the callsite (could be produced by attach an ebpf call to kprobe), things will also go wrong. It should start with given registers when the register is valid.
And it's true with omit frame pointer BP value could be anything, so 0 is also valid, I think I need to find a better way to tell if we could start with the registers value or direct start unwinding and skip until got the stack.

Hi, sorry I might have some misunderstanding. Adding an extra argument (eg. skip_sp) to indicate if it should just unwind from the current frame, and use SP as the "skip mark", should work well.

Quoted 6 lines

And I also think the "fake"/"real" reg is fragile, could we abuse another eflag (just like PERF_EFLAGS_EXACT) to indicate the regs are partially dumped fake registers? So perf_callchain_kernel just check if it's a "partial registers", and in such case it can start unwinding and skip until it get to SP. This make it easier to tell if the registers are "fake".

If you do the regs->eflags thing to mark the regs as fake in (perf_arch_fetch_caller_regs()), then I don't think skip_sp would be needed, because regs->sp could probably mark the skip point.

Instead I was actually hoping we could get rid of fake regs and perf_arch_fetch_caller_regs() altogether, because it's a nasty hack. But I don't know what else those fake regs are used for.

Signature · 1 line
Josh
maintainer
#10 lore raw
Quoted 7 lines · Josh Poimboeuf

If you do the regs->eflags thing to mark the regs as fake in (perf_arch_fetch_caller_regs()), then I don't think skip_sp would be needed, because regs->sp could probably mark the skip point.

Instead I was actually hoping we could get rid of fake regs and perf_arch_fetch_caller_regs() altogether, because it's a nasty hack. But I don't know what else those fake regs are used for.

This is the generic perf generate a sample path. It doesn't know the context. Normally it is an interrupt or exception of sorts and we simply pass the pt_regs from that down the chain and all is good.

It is just for a few software events, such as SW_CONTEXT_SWITCH, SW_MIGRATIONS and all the TP muck that we do not have regs to pass down.

These regs are used by:

- SAMPLE_IP,
- SAMPLE_CALLCHAIN (this here issue),
- SAMPLE_REGS_INTR,
- a few misc bits

There's actually comment on top of perf_fetch_caller_regs() that tries to explain what is needed -- although I suppose that could do with an update.

#11 lore raw
Quoted 18 lines · Peter Zijlstra
Quoted 7 lines · Josh Poimboeuf

If you do the regs->eflags thing to mark the regs as fake in (perf_arch_fetch_caller_regs()), then I don't think skip_sp would be needed, because regs->sp could probably mark the skip point.

Instead I was actually hoping we could get rid of fake regs and perf_arch_fetch_caller_regs() altogether, because it's a nasty hack. But I don't know what else those fake regs are used for.

This is the generic perf generate a sample path. It doesn't know the context. Normally it is an interrupt or exception of sorts and we simply pass the pt_regs from that down the chain and all is good.

It is just for a few software events, such as SW_CONTEXT_SWITCH, SW_MIGRATIONS and all the TP muck that we do not have regs to pass down.

These regs are used by:

Quoted 8 lines

- SAMPLE_IP,
- SAMPLE_CALLCHAIN (this here issue),
- SAMPLE_REGS_INTR,
- a few misc bits

There's actually comment on top of perf_fetch_caller_regs() that tries to explain what is needed -- although I suppose that could do with an update.

Ok, I guess it makes a little more sense now to my perf-ignorant brain. An improved comment would definitely help.

Sounds like using a reserved regs->flags bit to indicate fake regs is the way to go.

Signature · 1 line
Josh
#12 lore raw
In reply to Josh Poimboeuf
Quoted 89 lines · Josh Poimboeuf
Quoted 80 lines · Kairui Song
Quoted 65 lines · Kairui Song
Quoted 58 lines · Josh Poimboeuf
Quoted 42 lines · Peter Zijlstra

I'll mostly defer to Josh on unwinding, but a few comments below.

arch/x86/events/core.c void arch_perf_update_userpage(struct perf_event *event, · line 2362

Quoted 13 lines · Kairui Song
arch/x86/events/core.c +5 -0
index e2b1447192a8..6075a4f94376 100644
@@ -2355,6 +2355,12 @@ void arch_perf_update_userpage(struct perf_event *event,
    cyc2ns_read_end();
 }
 
+static inline int
+valid_perf_registers(struct pt_regs *regs)
+{
+   return (regs->ip && regs->bp && regs->sp);
+}

I'm unconvinced by this, with both guess and orc having !bp is perfectly valid.

Quoted 7 lines
arch/x86/events/core.c +0 -0
 void
 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 {
@@ -2366,11 +2372,17 @@ perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re
            return;
    }
 

arch/x86/events/core.c perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *re · line 2383

Quoted 10 lines
arch/x86/events/core.c +8 -1
-   if (perf_callchain_store(entry, regs->ip))
+   if (valid_perf_registers(regs)) {
+           if (perf_callchain_store(entry, regs->ip))
+                   return;
+           unwind_start(&state, current, regs, NULL);
+   } else if (regs->sp) {
+           unwind_start(&state, current, NULL, (unsigned long *)regs->sp);
+   } else {
            return;
+   }

AFAICT if we, by pure accident, end up with !bp for ORC, then we initialize the unwind wrong.

Quoted 2 lines

Note that @regs is mostly trivially correct, except for that tracepoint case. So I don't think we should magic here.

Ah, I didn't quite understand this code before, and I still don't really, but I guess the issue is that @regs can be either real or fake.

In the real @regs case, we just want to always unwind starting from regs->sp.

But in the fake @regs case, we should instead unwind from the current frame, skipping all frames until we hit the fake regs->sp. Because starting from fake/incomplete regs is most likely going to cause problems with ORC (or DWARF for other arches).

Quoted 4 lines

The idea of a fake regs is fragile and confusing. Is it possible to just pass in the "skip" stack pointer directly instead? That should work for both FP and non-FP. And I _think_ there's no need to ever capture regs->bp anyway -- the stack pointer should be sufficient.

Quoted 9 lines

Hi, that will break some other usage, if perf_callchain_kernel is called but it won't unwind to the callsite (could be produced by attach an ebpf call to kprobe), things will also go wrong. It should start with given registers when the register is valid.
And it's true with omit frame pointer BP value could be anything, so 0 is also valid, I think I need to find a better way to tell if we could start with the registers value or direct start unwinding and skip until got the stack.

Hi, sorry I might have some misunderstanding. Adding an extra argument (eg. skip_sp) to indicate if it should just unwind from the current frame, and use SP as the "skip mark", should work well.

Quoted 6 lines

And I also think the "fake"/"real" reg is fragile, could we abuse another eflag (just like PERF_EFLAGS_EXACT) to indicate the regs are partially dumped fake registers? So perf_callchain_kernel just check if it's a "partial registers", and in such case it can start unwinding and skip until it get to SP. This make it easier to tell if the registers are "fake".

Quoted 8 lines

If you do the regs->eflags thing to mark the regs as fake in (perf_arch_fetch_caller_regs()), then I don't think skip_sp would be needed, because regs->sp could probably mark the skip point.

Instead I was actually hoping we could get rid of fake regs and perf_arch_fetch_caller_regs() altogether, because it's a nasty hack. But I don't know what else those fake regs are used for.

Despite it's hacky, it seems not necessary to dump every register. And is there a straight way to get caller's regs in the trace point? It seems more trouble some. Or if we just use the regs inside the tracepoint, but it would need even more workaround (eg. unwind one frame before do anything).

Signature · 2 lines
Best Regards,
Kairui Song
maintainer
#13 lore raw
Quoted 5 lines · Kairui Song

Despite it's hacky, it seems not necessary to dump every register. And is there a straight way to get caller's regs in the trace point? It seems more trouble some. Or if we just use the regs inside the tracepoint, but it would need even more workaround (eg. unwind one frame before do anything).

Sure, and the current code doesn't fill out every reg either, but given SAMPLE_REGS_INTR I think we should at least clear the entire regs before passing it on, or otherwise make sure SAMPLE_REGS_INTR isn't allowed on these events.