[PATCH] tracing: remove superfluous sub instructions

[PATCH] tracing: remove superfluous sub instructions

[Jiri Olsa]

hi,

I think there's no need for substracting MCOUNT_INSN_SIZE from the
IP parameter before calling the function trace (/graph) handler.

Maybe I overlooked something, but all the IP usage I saw ended
up in the kallsyms_lookup function, which does the lookup using the
functions' start/end boundaries to find the correct symbol for pointer.

Thus it seems to me there's no point in substracting the
MCOUNT_INSN_SIZE value from the IP parameter.

I tested for x86_64 and got proper results, I believe it's
the same case for x86_32.

wbr,
jirka


Signed-off-by: Jiri Olsa <jolsa@redhat.com>
---
 arch/x86/kernel/entry_32.S |    3 ---
 arch/x86/kernel/entry_64.S |    3 ---
 2 files changed, 0 insertions(+), 6 deletions(-)

diff --git a/arch/x86/kernel/entry_32.S b/arch/x86/kernel/entry_32.S
index 591e601..92cdb47 100644
--- a/arch/x86/kernel/entry_32.S
+++ b/arch/x86/kernel/entry_32.S
@@ -1122,7 +1122,6 @@ ENTRY(ftrace_caller)
 	pushl %edx
 	movl 0xc(%esp), %eax
 	movl 0x4(%ebp), %edx
-	subl $MCOUNT_INSN_SIZE, %eax
 
 .globl ftrace_call
 ftrace_call:
@@ -1168,7 +1167,6 @@ trace:
 	pushl %edx
 	movl 0xc(%esp), %eax
 	movl 0x4(%ebp), %edx
-	subl $MCOUNT_INSN_SIZE, %eax
 
 	call *ftrace_trace_function
 
@@ -1191,7 +1189,6 @@ ENTRY(ftrace_graph_caller)
 	movl 0xc(%esp), %edx
 	lea 0x4(%ebp), %eax
 	movl (%ebp), %ecx
-	subl $MCOUNT_INSN_SIZE, %edx
 	call prepare_ftrace_return
 	popl %edx
 	popl %ecx
diff --git a/arch/x86/kernel/entry_64.S b/arch/x86/kernel/entry_64.S
index d3b895f..662f8f6 100644
--- a/arch/x86/kernel/entry_64.S
+++ b/arch/x86/kernel/entry_64.S
@@ -75,7 +75,6 @@ ENTRY(ftrace_caller)
 
 	movq 0x38(%rsp), %rdi
 	movq 8(%rbp), %rsi
-	subq $MCOUNT_INSN_SIZE, %rdi
 
 GLOBAL(ftrace_call)
 	call ftrace_stub
@@ -115,7 +114,6 @@ trace:
 
 	movq 0x38(%rsp), %rdi
 	movq 8(%rbp), %rsi
-	subq $MCOUNT_INSN_SIZE, %rdi
 
 	call   *ftrace_trace_function
 
@@ -136,7 +134,6 @@ ENTRY(ftrace_graph_caller)
 	leaq 8(%rbp), %rdi
 	movq 0x38(%rsp), %rsi
 	movq (%rbp), %rdx
-	subq $MCOUNT_INSN_SIZE, %rsi
 
 	call	prepare_ftrace_return
 

Re: [PATCH] tracing: remove superfluous sub instructions

[Frederic Weisbecker]

On Tue, Jan 18, 2011 at 04:28:18PM +0100, Jiri Olsa wrote:
> hi,
> 
> I think there's no need for substracting MCOUNT_INSN_SIZE from the
> IP parameter before calling the function trace (/graph) handler.
> 
> Maybe I overlooked something, but all the IP usage I saw ended
> up in the kallsyms_lookup function, which does the lookup using the
> functions' start/end boundaries to find the correct symbol for pointer.
> 
> Thus it seems to me there's no point in substracting the
> MCOUNT_INSN_SIZE value from the IP parameter.
> 
> I tested for x86_64 and got proper results, I believe it's
> the same case for x86_32.
> 
> wbr,
> jirka
> 
> 
> Signed-off-by: Jiri Olsa <jolsa@redhat.com>

Well, that sounds right after all. If we are only interested in the
symbol, the instruction that follows "call mcount" is still relevant
as it must belong to the same function.

Steve?

Re: [PATCH] tracing: remove superfluous sub instructions

[Steven Rostedt]

On Tue, 2011-01-18 at 16:52 +0100, Frederic Weisbecker wrote:
> On Tue, Jan 18, 2011 at 04:28:18PM +0100, Jiri Olsa wrote:
> > hi,
> > 
> > I think there's no need for substracting MCOUNT_INSN_SIZE from the
> > IP parameter before calling the function trace (/graph) handler.
> > 
> > Maybe I overlooked something, but all the IP usage I saw ended
> > up in the kallsyms_lookup function, which does the lookup using the
> > functions' start/end boundaries to find the correct symbol for pointer.
> > 
> > Thus it seems to me there's no point in substracting the
> > MCOUNT_INSN_SIZE value from the IP parameter.
> > 
> > I tested for x86_64 and got proper results, I believe it's
> > the same case for x86_32.
> > 
> > wbr,
> > jirka
> > 
> > 
> > Signed-off-by: Jiri Olsa <jolsa@redhat.com>
> 
> Well, that sounds right after all. If we are only interested in the
> symbol, the instruction that follows "call mcount" is still relevant
> as it must belong to the same function.
> 
> Steve?

NAK, it will break function triggers/probes. The "func:traceon" and
"func:traceoff". They compare the ip to the call location of mcount.

-- Steve

Re: [PATCH] tracing: remove superfluous sub instructions

[Jiri Olsa]

On Tue, Jan 18, 2011 at 11:22:42AM -0500, Steven Rostedt wrote:
> On Tue, 2011-01-18 at 16:52 +0100, Frederic Weisbecker wrote:
> > On Tue, Jan 18, 2011 at 04:28:18PM +0100, Jiri Olsa wrote:
> > > hi,
> > > 
> > > I think there's no need for substracting MCOUNT_INSN_SIZE from the
> > > IP parameter before calling the function trace (/graph) handler.
> > > 
> > > Maybe I overlooked something, but all the IP usage I saw ended
> > > up in the kallsyms_lookup function, which does the lookup using the
> > > functions' start/end boundaries to find the correct symbol for pointer.
> > > 
> > > Thus it seems to me there's no point in substracting the
> > > MCOUNT_INSN_SIZE value from the IP parameter.
> > > 
> > > I tested for x86_64 and got proper results, I believe it's
> > > the same case for x86_32.
> > > 
> > > wbr,
> > > jirka
> > > 
> > > 
> > > Signed-off-by: Jiri Olsa <jolsa@redhat.com>
> > 
> > Well, that sounds right after all. If we are only interested in the
> > symbol, the instruction that follows "call mcount" is still relevant
> > as it must belong to the same function.
> > 
> > Steve?
> 
> NAK, it will break function triggers/probes. The "func:traceon" and
> "func:traceoff". They compare the ip to the call location of mcount.
> 
> -- Steve
> 
> 

ops, missed this one..

would it make sense to update the IP inside the function_trace_probe_call
function to save one instruction in the entry code used by all? or it's
not worth it..

jirka

Re: [PATCH] tracing: remove superfluous sub instructions

[Steven Rostedt]

On Tue, 2011-01-18 at 17:41 +0100, Jiri Olsa wrote:

> ops, missed this one..
> 
> would it make sense to update the IP inside the function_trace_probe_call
> function to save one instruction in the entry code used by all? or it's
> not worth it..

That would require function_trace_probe_call() to handle arch
dependencies. I would like to keep arch dependent code in the arch
subsystem when possible.

This change only saves a subtraction to a register. It only gets called
if we are tracing functions. This modification is far into the noise
that is caused by the tracer. I don't think it is worth it as it will
cause head aches with knowing how to handle the ip from the users of the
function tracer.

I do plan on making the function tracer a bit more generic for common
users. I would like to keep the ip at the call site of mcount.

-- Steve