[PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES

[PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES

[Christoph Lameter]

ia64_do_page_fault is a path critical for system performance. The code to call
notify_die() should not be compiled into that critical path if the system
is not configured to use KPROBES.

Signed-off-by: Christoph Lameter <clameter@sgi.com>

Index: linux-2.6.13-rc7/arch/ia64/mm/fault.c
=================================--- linux-2.6.13-rc7.orig/arch/ia64/mm/fault.c	2005-08-23 20:39:14.000000000 -0700
+++ linux-2.6.13-rc7/arch/ia64/mm/fault.c	2005-08-25 13:04:57.000000000 -0700
@@ -103,12 +103,16 @@ ia64_do_page_fault (unsigned long addres
 		goto bad_area_no_up;
 #endif
 
+#ifdef CONFIG_KPROBES
 	/*
-	 * This is to handle the kprobes on user space access instructions
+	 * This is to handle the kprobes on user space access instructions.
+	 * This is a path criticial for system performance. So only
+	 * process this notifier if we are compiled with kprobes support.
 	 */
 	if (notify_die(DIE_PAGE_FAULT, "page fault", regs, code, TRAP_BRKPT,
 					SIGSEGV) = NOTIFY_STOP)
 		return;
+#endif
 
 	down_read(&mm->mmap_sem);
 

RE: [PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES is configured.

[Luck, Tony]

>Please do not generate any code if the feature cannot ever be 
>used (CONFIG_KPROBES off). With this patch we still have lots of 
>unnecessary code being executed on each page fault.

I can (eventually) wrap this call inside the #ifdef CONFIG_KPROBES.

But I'd like to keep following leads on making the overhead as
low as possible for those people that do have KPROBES configured
(which may be most people if OS distributors ship kernels with
this enabled).

-Tony

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES is configured.

[Andi Kleen]

On Wednesday 31 August 2005 01:05, Luck, Tony wrote:
> >Please do not generate any code if the feature cannot ever be
> >used (CONFIG_KPROBES off). With this patch we still have lots of
> >unnecessary code being executed on each page fault.
>
> I can (eventually) wrap this call inside the #ifdef CONFIG_KPROBES.

At least the original die notifiers were designed as a generic debugger
interface, not a kprobes specific thing. So I don't think it's a good idea.
Given most debuggers don't need the early page fault hook and it's
mostly needed for a special case in kprobes, but it doesn't seem nice to only 
offer a subset of the hooks with specific config options.

Also with the inline the test should be essentially a single test of 
a global variable and jump. Hardly a big performance issue, no? 

-Andi

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if

[Christoph Lameter]

On Wed, 31 Aug 2005, Andi Kleen wrote:

> Also with the inline the test should be essentially a single test of 
> a global variable and jump. Hardly a big performance issue, no? 

There are multiple effects of this code.

- Additional cacheline in use in the page fault handler 
  increasing the cache foot print.

- There are registers in use for checking the global variable.

- The compilers will reserve registers for the code that is never 
  executed which may affect other elements of performance. From the 
  register perspective a function call may be better on ia64.

Certainly not a big effect (if we make sure the compiler knows that 
this test mostly fails and insure that the variable is in 
__mostly_read) but this is a frequently executed code path and the code
is there without purpose if CONFIG_KPROBES is off.

It wont get too bad unless lots of other people have similar ideas about 
fixing their race conditions using similar methods. But we will be setting 
a bad precedent if we allow this.

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES is configured.

[Andi Kleen]

On Wednesday 31 August 2005 01:54, Christoph Lameter wrote:

> Certainly not a big effect (if we make sure the compiler knows that
> this test mostly fails and insure that the variable is in
> __mostly_read) 

Currently neither, but that could be easily fixed.

> but this is a frequently executed code path and the code 
> is there without purpose if CONFIG_KPROBES is off.

Well if you really worry about it then it might be better to do some dynamic
code patching to make generic and distribution kernels too.

> It wont get too bad unless lots of other people have similar ideas about
> fixing their race conditions using similar methods. But we will be setting
> a bad precedent if we allow this.

Agreed on the general direction, but I didn't see an alternative
for kprobes for this. Well actually the hook could be maybe
right now moved into the part before kernel oops which is much less
frequently executed, but then it'll likely move up again once
kprobes support user space probes.

-Andi

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if

[David S. Miller]

From: Andi Kleen <ak@suse.de>
Subject: Re: [PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES is configured.
Date: Wed, 31 Aug 2005 01:38:08 +0200

> On Wednesday 31 August 2005 01:05, Luck, Tony wrote:
> > >Please do not generate any code if the feature cannot ever be
> > >used (CONFIG_KPROBES off). With this patch we still have lots of
> > >unnecessary code being executed on each page fault.
> >
> > I can (eventually) wrap this call inside the #ifdef CONFIG_KPROBES.
> 
> At least the original die notifiers were designed as a generic debugger
> interface, not a kprobes specific thing. So I don't think it's a good idea.

Me neither, I think a way too big deal is being made about
about this by the ia64 folks.  Just put the dang hook in
there unconditionally already :-)

Re:[PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES

[Christoph Lameter]

On Fri, 26 Aug 2005, Rusty Lynch wrote:

> Just to be sure everyone understands the overhead involved, kprobes only 
> registers a single notifier.  If kprobes is disabled (CONFIG_KPROBES is
> off) then the overhead on a page fault is the overhead to execute an empty
> notifier chain.

Its the overhead of using registers to pass parameters, performing a 
function call that does nothing etc. A waste of computing resources. All 
of that unconditionally in a performance critical execution path that 
is executed a gazillion times for an optional feature that I frankly 
find not useful at all and that is disabled by default.

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES is configured.

[Andi Kleen]

On Saturday 27 August 2005 01:05, Christoph Lameter wrote:
> On Fri, 26 Aug 2005, Rusty Lynch wrote:
> > Just to be sure everyone understands the overhead involved, kprobes only
> > registers a single notifier.  If kprobes is disabled (CONFIG_KPROBES is
> > off) then the overhead on a page fault is the overhead to execute an
> > empty notifier chain.
>
> Its the overhead of using registers to pass parameters, performing a
> function call that does nothing etc. A waste of computing resources. All
> of that unconditionally in a performance critical execution path that
> is executed a gazillion times for an optional feature that I frankly
> find not useful at all and that is disabled by default.

In the old days notifier_call_chain used to be inline. Then someone looking
at code size out of lined it. Perhaps it should be inlined again or notifier.h
could supply a special faster inline version for time critical code.

Then it would be simple if (global_var != NULL) { ... } in the fast path.
In addition the call chain could be declared __read_mostly.

I suspect with these changes Christoph's concerns would go away, right?

-Andi

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES is configured.

[Rusty Lynch]

On Sat, Aug 27, 2005 at 02:24:25AM +0200, Andi Kleen wrote:
> On Saturday 27 August 2005 01:05, Christoph Lameter wrote:
> > On Fri, 26 Aug 2005, Rusty Lynch wrote:
> > > Just to be sure everyone understands the overhead involved, kprobes only
> > > registers a single notifier.  If kprobes is disabled (CONFIG_KPROBES is
> > > off) then the overhead on a page fault is the overhead to execute an
> > > empty notifier chain.
> >
> > Its the overhead of using registers to pass parameters, performing a
> > function call that does nothing etc. A waste of computing resources. All
> > of that unconditionally in a performance critical execution path that
> > is executed a gazillion times for an optional feature that I frankly
> > find not useful at all and that is disabled by default.
> 
> In the old days notifier_call_chain used to be inline. Then someone looking
> at code size out of lined it. Perhaps it should be inlined again or notifier.h
> could supply a special faster inline version for time critical code.
> 
> Then it would be simple if (global_var != NULL) { ... } in the fast path.
> In addition the call chain could be declared __read_mostly.
> 
> I suspect with these changes Christoph's concerns would go away, right?
> 
> -Andi

So, assuming inlining the notifier_call_chain would address Christoph's
conserns, is the following patch something like what you are sugesting?  
This would make all the kdebug.h::notify_die() calls use the inline version. 

(WARNING:  The following has only been tested on ia64.)

 include/asm-i386/kdebug.h    |    2 +-
 include/asm-ia64/kdebug.h    |    2 +-
 include/asm-ppc64/kdebug.h   |    2 +-
 include/asm-sparc64/kdebug.h |    2 +-
 include/asm-x86_64/kdebug.h  |    2 +-
 include/linux/notifier.h     |   18 ++++++++++++++++++
 kernel/sys.c                 |   14 +-------------
 7 files changed, 24 insertions(+), 18 deletions(-)

Index: linux-2.6.13/include/linux/notifier.h
=================================--- linux-2.6.13.orig/include/linux/notifier.h
+++ linux-2.6.13/include/linux/notifier.h
@@ -72,5 +72,23 @@ extern int notifier_call_chain(struct no
 #define CPU_DOWN_FAILED		0x0006 /* CPU (unsigned)v NOT going down */
 #define CPU_DEAD		0x0007 /* CPU (unsigned)v dead */
 
+static inline int fast_notifier_call_chain(struct notifier_block **n,
+					   unsigned long val, void *v)
+{
+	int ret=NOTIFY_DONE;
+	struct notifier_block *nb = *n;
+
+	while(nb)
+	{
+		ret=nb->notifier_call(nb,val,v);
+		if(ret&NOTIFY_STOP_MASK)
+		{
+			return ret;
+		}
+		nb=nb->next;
+	}
+	return ret;
+}
+
 #endif /* __KERNEL__ */
 #endif /* _LINUX_NOTIFIER_H */
Index: linux-2.6.13/kernel/sys.c
=================================--- linux-2.6.13.orig/kernel/sys.c
+++ linux-2.6.13/kernel/sys.c
@@ -169,19 +169,7 @@ EXPORT_SYMBOL(notifier_chain_unregister)
  
 int notifier_call_chain(struct notifier_block **n, unsigned long val, void *v)
 {
-	int ret=NOTIFY_DONE;
-	struct notifier_block *nb = *n;
-
-	while(nb)
-	{
-		ret=nb->notifier_call(nb,val,v);
-		if(ret&NOTIFY_STOP_MASK)
-		{
-			return ret;
-		}
-		nb=nb->next;
-	}
-	return ret;
+	return fast_notifier_call_chain(n, val, v);
 }
 
 EXPORT_SYMBOL(notifier_call_chain);
Index: linux-2.6.13/include/asm-ia64/kdebug.h
=================================--- linux-2.6.13.orig/include/asm-ia64/kdebug.h
+++ linux-2.6.13/include/asm-ia64/kdebug.h
@@ -55,7 +55,7 @@ static inline int notify_die(enum die_va
 		.signr  = sig
 	};
 
-	return notifier_call_chain(&ia64die_chain, val, &args);
+	return fast_notifier_call_chain(&ia64die_chain, val, &args);
 }
 
 #endif
Index: linux-2.6.13/include/asm-i386/kdebug.h
=================================--- linux-2.6.13.orig/include/asm-i386/kdebug.h
+++ linux-2.6.13/include/asm-i386/kdebug.h
@@ -44,7 +44,7 @@ enum die_val {
 static inline int notify_die(enum die_val val,char *str,struct pt_regs *regs,long err,int trap, int sig)
 {
 	struct die_args args = { .regs=regs, .str=str, .err=err, .trapnr=trap,.signr=sig };
-	return notifier_call_chain(&i386die_chain, val, &args);
+	return fast_notifier_call_chain(&i386die_chain, val, &args);
 }
 
 #endif
Index: linux-2.6.13/include/asm-ppc64/kdebug.h
=================================--- linux-2.6.13.orig/include/asm-ppc64/kdebug.h
+++ linux-2.6.13/include/asm-ppc64/kdebug.h
@@ -37,7 +37,7 @@ enum die_val {
 static inline int notify_die(enum die_val val,char *str,struct pt_regs *regs,long err,int trap, int sig)
 {
 	struct die_args args = { .regs=regs, .str=str, .err=err, .trapnr=trap,.signr=sig };
-	return notifier_call_chain(&ppc64_die_chain, val, &args);
+	return fast_notifier_call_chain(&ppc64_die_chain, val, &args);
 }
 
 #endif
Index: linux-2.6.13/include/asm-sparc64/kdebug.h
=================================--- linux-2.6.13.orig/include/asm-sparc64/kdebug.h
+++ linux-2.6.13/include/asm-sparc64/kdebug.h
@@ -46,7 +46,7 @@ static inline int notify_die(enum die_va
 				 .trapnr	= trap,
 				 .signr		= sig };
 
-	return notifier_call_chain(&sparc64die_chain, val, &args);
+	return fast_notifier_call_chain(&sparc64die_chain, val, &args);
 }
 
 #endif
Index: linux-2.6.13/include/asm-x86_64/kdebug.h
=================================--- linux-2.6.13.orig/include/asm-x86_64/kdebug.h
+++ linux-2.6.13/include/asm-x86_64/kdebug.h
@@ -38,7 +38,7 @@ enum die_val { 
 static inline int notify_die(enum die_val val,char *str,struct pt_regs *regs,long err,int trap, int sig)
 { 
 	struct die_args args = { .regs=regs, .str=str, .err=err, .trapnr=trap,.signr=sig }; 
-	return notifier_call_chain(&die_chain, val, &args); 
+	return fast_notifier_call_chain(&die_chain, val, &args);
 } 
 
 extern int printk_address(unsigned long address);

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if

[Christoph Lameter]

On Mon, 29 Aug 2005, Rusty Lynch wrote:

> So, assuming inlining the notifier_call_chain would address Christoph's
> conserns, is the following patch something like what you are sugesting?  
> This would make all the kdebug.h::notify_die() calls use the inline version. 

Please do not generate any code if the feature cannot ever be 
used (CONFIG_KPROBES off). With this patch we still have lots of 
unnecessary code being executed on each page fault.

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if KPROBES is configured.

[Matthew Wilcox]

On Mon, Aug 29, 2005 at 05:19:05PM -0700, Rusty Lynch wrote:
> So, assuming inlining the notifier_call_chain would address Christoph's
> conserns, is the following patch something like what you are sugesting?  
> This would make all the kdebug.h::notify_die() calls use the inline version. 

I think we need something more like this ...

include/linux/notifier.h:
+static inline int notifier_call_chain(struct notifier_block **n,
+					unsigned long val, void *v)
+{
+	if (n)
+		return __notifier_call_chain(n, val, v);
+	return NOTIFY_DONE;
+}
kernel/sys.c:
-int notifier_call_chain(struct notifier_block **n, unsigned long val, void *v)
+int __notifier_call_chain(struct notifier_block **n, unsigned long val, void *v)
-EXPORT_SYMBOL(notifier_call_chain);
+EXPORT_SYMBOL(__notifier_call_chain);

That way everyone gets both the quick test and the global size reduction.

-- 
"Next the statesmen will invent cheap lies, putting the blame upon 
the nation that is attacked, and every man will be glad of those
conscience-soothing falsities, and will diligently study them, and refuse
to examine any refutations of them; and thus he will by and by convince 
himself that the war is just, and will thank God for the better sleep 
he enjoys after this process of grotesque self-deception." -- Mark Twain

Re: [PATCH] Only process_die notifier in ia64_do_page_fault if

[Christoph Lameter]

On Tue, 30 Aug 2005, Matthew Wilcox wrote:

> On Mon, Aug 29, 2005 at 05:19:05PM -0700, Rusty Lynch wrote:
> > So, assuming inlining the notifier_call_chain would address Christoph's
> > conserns, is the following patch something like what you are sugesting?  
> > This would make all the kdebug.h::notify_die() calls use the inline version. 
> 
> I think we need something more like this ...
> 
> include/linux/notifier.h:
> +static inline int notifier_call_chain(struct notifier_block **n,
> +					unsigned long val, void *v)
> +{
> +	if (n)
> +		return __notifier_call_chain(n, val, v);
> +	return NOTIFY_DONE;
> +}
> kernel/sys.c:
> -int notifier_call_chain(struct notifier_block **n, unsigned long val, void *v)
> +int __notifier_call_chain(struct notifier_block **n, unsigned long val, void *v)
> -EXPORT_SYMBOL(notifier_call_chain);
> +EXPORT_SYMBOL(__notifier_call_chain);
> 
> That way everyone gets both the quick test and the global size reduction.

And then do

#ifndef CONFIG_KPROBES

#define ia64die_chain 0

#endif

in include/asm-ia64/kdebug.h?

Otherwise we still check a notifier chain that cannot ever be 
activated.

But then the patch is essentially the same as the last one I proposed.