[patch 0/6] Notify page fault call chain

[patch 0/6] Notify page fault call chain

[Anil S Keshavamurthy]

Hi Andrew,

Currently in the do_page_fault() code path, we call notify_die(DIE_PAGE_FAULT, ...)
to notify the page fault. The only interested components for this page fault
notifications are Kprobes and/or kdb. Since notify_die() is highly overloaded,
this page fault notification is currently being sent to other components
registered with register_die_notification() which uses the same die_chain to loop
for all the registered components.

In order to optimize the do_page_fault() code path, this critical page fault notification
is now moved to different call chain and the test results showed great improvements.

Patches for i386, x86_64, ia64, powerpc and sparc64 follows this mail.

Please apply.

Thanks,
Anil Keshavamurthy

--

[patch 1/6] Notify page fault call chain for i386

[Anil S Keshavamurthy]

[-- Attachment #1: notify_page_fault_i386.patch --]
[-- Type: text/plain, Size: 3185 bytes --]

---
 arch/i386/kernel/traps.c  |   12 ++++++++++++
 arch/i386/mm/fault.c      |    4 ++--
 include/asm-i386/kdebug.h |   15 +++++++++++++++
 3 files changed, 29 insertions(+), 2 deletions(-)

Index: linux-2.6.17-rc1-mm3/arch/i386/kernel/traps.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/i386/kernel/traps.c
+++ linux-2.6.17-rc1-mm3/arch/i386/kernel/traps.c
@@ -93,6 +93,7 @@ asmlinkage void machine_check(void);
 
 static int kstack_depth_to_print = 24;
 ATOMIC_NOTIFIER_HEAD(i386die_chain);
+ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
 
 int register_die_notifier(struct notifier_block *nb)
 {
@@ -107,6 +108,17 @@ int unregister_die_notifier(struct notif
 }
 EXPORT_SYMBOL(unregister_die_notifier);
 
+int register_page_fault_notifier(struct notifier_block *nb)
+{
+	vmalloc_sync_all();
+	return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
+}
+
+int unregister_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
+}
+
 static inline int valid_stack_ptr(struct thread_info *tinfo, void *p)
 {
 	return	p > (void *)tinfo &&
Index: linux-2.6.17-rc1-mm3/arch/i386/mm/fault.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/i386/mm/fault.c
+++ linux-2.6.17-rc1-mm3/arch/i386/mm/fault.c
@@ -321,7 +321,7 @@ fastcall void __kprobes do_page_fault(st
 	if (unlikely(address >= TASK_SIZE)) {
 		if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
 			return;
-		if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
+		if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 						SIGSEGV) == NOTIFY_STOP)
 			return;
 		/*
@@ -331,7 +331,7 @@ fastcall void __kprobes do_page_fault(st
 		goto bad_area_nosemaphore;
 	}
 
-	if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
+	if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 					SIGSEGV) == NOTIFY_STOP)
 		return;
 
Index: linux-2.6.17-rc1-mm3/include/asm-i386/kdebug.h
===================================================================
--- linux-2.6.17-rc1-mm3.orig/include/asm-i386/kdebug.h
+++ linux-2.6.17-rc1-mm3/include/asm-i386/kdebug.h
@@ -19,7 +19,10 @@ struct die_args {
 
 extern int register_die_notifier(struct notifier_block *);
 extern int unregister_die_notifier(struct notifier_block *);
+extern int register_page_fault_notifier(struct notifier_block *);
+extern int unregister_page_fault_notifier(struct notifier_block *);
 extern struct atomic_notifier_head i386die_chain;
+extern struct atomic_notifier_head notify_page_fault_chain;
 
 
 /* Grossly misnamed. */
@@ -53,4 +56,16 @@ static inline int notify_die(enum die_va
 	return atomic_notifier_call_chain(&i386die_chain, val, &args);
 }
 
+static inline int notify_page_fault(enum die_val val, const 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 atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
+}
 #endif

--

[patch 2/6] Notify page fault call chain for x86_64

[Anil S Keshavamurthy]

[-- Attachment #1: notify_page_fault_x86_64.patch --]
[-- Type: text/plain, Size: 3303 bytes --]

---
 arch/x86_64/kernel/traps.c  |   12 ++++++++++++
 arch/x86_64/mm/fault.c      |    4 ++--
 include/asm-x86_64/kdebug.h |   16 ++++++++++++++++
 3 files changed, 30 insertions(+), 2 deletions(-)

Index: linux-2.6.17-rc1-mm3/arch/x86_64/kernel/traps.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/x86_64/kernel/traps.c
+++ linux-2.6.17-rc1-mm3/arch/x86_64/kernel/traps.c
@@ -71,6 +71,7 @@ asmlinkage void machine_check(void);
 asmlinkage void spurious_interrupt_bug(void);
 
 ATOMIC_NOTIFIER_HEAD(die_chain);
+ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
 
 int register_die_notifier(struct notifier_block *nb)
 {
@@ -85,6 +86,17 @@ int unregister_die_notifier(struct notif
 }
 EXPORT_SYMBOL(unregister_die_notifier);
 
+int register_page_fault_notifier(struct notifier_block *nb)
+{
+	vmalloc_sync_all();
+	return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
+}
+
+int unregister_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
+}
+
 static inline void conditional_sti(struct pt_regs *regs)
 {
 	if (regs->eflags & X86_EFLAGS_IF)
Index: linux-2.6.17-rc1-mm3/arch/x86_64/mm/fault.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/x86_64/mm/fault.c
+++ linux-2.6.17-rc1-mm3/arch/x86_64/mm/fault.c
@@ -348,7 +348,7 @@ asmlinkage void __kprobes do_page_fault(
 			if (vmalloc_fault(address) >= 0)
 				return;
 		}
-		if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
+		if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 						SIGSEGV) == NOTIFY_STOP)
 			return;
 		/*
@@ -358,7 +358,7 @@ asmlinkage void __kprobes do_page_fault(
 		goto bad_area_nosemaphore;
 	}
 
-	if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
+	if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 					SIGSEGV) == NOTIFY_STOP)
 		return;
 
Index: linux-2.6.17-rc1-mm3/include/asm-x86_64/kdebug.h
===================================================================
--- linux-2.6.17-rc1-mm3.orig/include/asm-x86_64/kdebug.h
+++ linux-2.6.17-rc1-mm3/include/asm-x86_64/kdebug.h
@@ -15,7 +15,10 @@ struct die_args {
 
 extern int register_die_notifier(struct notifier_block *);
 extern int unregister_die_notifier(struct notifier_block *);
+extern int register_page_fault_notifier(struct notifier_block *);
+extern int unregister_page_fault_notifier(struct notifier_block *);
 extern struct atomic_notifier_head die_chain;
+extern struct atomic_notifier_head notify_page_fault_chain;
 
 /* Grossly misnamed. */
 enum die_val {
@@ -47,6 +50,19 @@ static inline int notify_die(enum die_va
 	return atomic_notifier_call_chain(&die_chain, val, &args);
 } 
 
+static inline int notify_page_fault(enum die_val val, const 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 atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
+}
+
 extern int printk_address(unsigned long address);
 extern void die(const char *,struct pt_regs *,long);
 extern void __die(const char *,struct pt_regs *,long);

--

[patch 3/6] Notify page fault call chain for ia64

[Anil S Keshavamurthy]

[-- Attachment #1: notify_page_fault_ia64.patch --]
[-- Type: text/plain, Size: 2768 bytes --]

---
 arch/ia64/kernel/traps.c  |   11 +++++++++++
 arch/ia64/mm/fault.c      |    2 +-
 include/asm-ia64/kdebug.h |   16 ++++++++++++++++
 3 files changed, 28 insertions(+), 1 deletion(-)

Index: linux-2.6.17-rc1-mm3/arch/ia64/kernel/traps.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/ia64/kernel/traps.c
+++ linux-2.6.17-rc1-mm3/arch/ia64/kernel/traps.c
@@ -31,6 +31,7 @@ fpswa_interface_t *fpswa_interface;
 EXPORT_SYMBOL(fpswa_interface);
 
 ATOMIC_NOTIFIER_HEAD(ia64die_chain);
+ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
 
 int
 register_die_notifier(struct notifier_block *nb)
@@ -46,6 +47,16 @@ unregister_die_notifier(struct notifier_
 }
 EXPORT_SYMBOL_GPL(unregister_die_notifier);
 
+int register_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
+}
+
+int unregister_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
+}
+
 void __init
 trap_init (void)
 {
Index: linux-2.6.17-rc1-mm3/arch/ia64/mm/fault.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/ia64/mm/fault.c
+++ linux-2.6.17-rc1-mm3/arch/ia64/mm/fault.c
@@ -84,7 +84,7 @@ ia64_do_page_fault (unsigned long addres
 	/*
 	 * This is to handle the kprobes on user space access instructions
 	 */
-	if (notify_die(DIE_PAGE_FAULT, "page fault", regs, code, TRAP_BRKPT,
+	if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, code, TRAP_BRKPT,
 					SIGSEGV) == NOTIFY_STOP)
 		return;
 
Index: linux-2.6.17-rc1-mm3/include/asm-ia64/kdebug.h
===================================================================
--- linux-2.6.17-rc1-mm3.orig/include/asm-ia64/kdebug.h
+++ linux-2.6.17-rc1-mm3/include/asm-ia64/kdebug.h
@@ -40,7 +40,10 @@ struct die_args {
 
 extern int register_die_notifier(struct notifier_block *);
 extern int unregister_die_notifier(struct notifier_block *);
+extern int register_page_fault_notifier(struct notifier_block *);
+extern int unregister_page_fault_notifier(struct notifier_block *);
 extern struct atomic_notifier_head ia64die_chain;
+extern struct atomic_notifier_head notify_page_fault_chain;
 
 enum die_val {
 	DIE_BREAK = 1,
@@ -86,4 +89,17 @@ static inline int notify_die(enum die_va
 	return atomic_notifier_call_chain(&ia64die_chain, val, &args);
 }
 
+static inline int notify_page_fault(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 atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
+}
 #endif

--

[patch 4/6] Notify page fault call chain for powerpc

[Anil S Keshavamurthy]

[-- Attachment #1: notify_page_fault_powerpc.patch --]
[-- Type: text/plain, Size: 2795 bytes --]

---
 arch/powerpc/kernel/traps.c  |   11 +++++++++++
 arch/powerpc/mm/fault.c      |    2 +-
 include/asm-powerpc/kdebug.h |    9 +++++++++
 3 files changed, 21 insertions(+), 1 deletion(-)

Index: linux-2.6.17-rc1-mm3/arch/powerpc/kernel/traps.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/powerpc/kernel/traps.c
+++ linux-2.6.17-rc1-mm3/arch/powerpc/kernel/traps.c
@@ -75,6 +75,7 @@ EXPORT_SYMBOL(__debugger_fault_handler);
 #endif
 
 ATOMIC_NOTIFIER_HEAD(powerpc_die_chain);
+ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
 
 int register_die_notifier(struct notifier_block *nb)
 {
@@ -88,6 +89,16 @@ int unregister_die_notifier(struct notif
 }
 EXPORT_SYMBOL(unregister_die_notifier);
 
+int register_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
+}
+
+int unregister_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
+}
+
 /*
  * Trap & Exception support
  */
Index: linux-2.6.17-rc1-mm3/arch/powerpc/mm/fault.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/powerpc/mm/fault.c
+++ linux-2.6.17-rc1-mm3/arch/powerpc/mm/fault.c
@@ -142,7 +142,7 @@ int __kprobes do_page_fault(struct pt_re
 	is_write = error_code & ESR_DST;
 #endif /* CONFIG_4xx || CONFIG_BOOKE */
 
-	if (notify_die(DIE_PAGE_FAULT, "page_fault", regs, error_code,
+	if (notify_page_fault(DIE_PAGE_FAULT, "page_fault", regs, error_code,
 				11, SIGSEGV) == NOTIFY_STOP)
 		return 0;
 
Index: linux-2.6.17-rc1-mm3/include/asm-powerpc/kdebug.h
===================================================================
--- linux-2.6.17-rc1-mm3.orig/include/asm-powerpc/kdebug.h
+++ linux-2.6.17-rc1-mm3/include/asm-powerpc/kdebug.h
@@ -18,7 +18,10 @@ struct die_args {
 
 extern int register_die_notifier(struct notifier_block *);
 extern int unregister_die_notifier(struct notifier_block *);
+extern int register_page_fault_notifier(struct notifier_block *);
+extern int unregister_page_fault_notifier(struct notifier_block *);
 extern struct atomic_notifier_head powerpc_die_chain;
+extern struct atomic_notifier_head notify_page_fault_chain;
 
 /* Grossly misnamed. */
 enum die_val {
@@ -36,5 +39,11 @@ static inline int notify_die(enum die_va
 	return atomic_notifier_call_chain(&powerpc_die_chain, val, &args);
 }
 
+static inline int notify_page_fault(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 atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
+}
+
 #endif /* __KERNEL__ */
 #endif /* _ASM_POWERPC_KDEBUG_H */

--

[patch 5/6] Notify page fault call chain for sparc64

[Anil S Keshavamurthy]

[-- Attachment #1: notify_page_fault_sparc64.patch --]
[-- Type: text/plain, Size: 2886 bytes --]

---
 arch/sparc64/kernel/traps.c  |   11 +++++++++++
 arch/sparc64/mm/fault.c      |    2 +-
 include/asm-sparc64/kdebug.h |   16 ++++++++++++++++
 3 files changed, 28 insertions(+), 1 deletion(-)

Index: linux-2.6.17-rc1-mm3/arch/sparc64/kernel/traps.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/sparc64/kernel/traps.c
+++ linux-2.6.17-rc1-mm3/arch/sparc64/kernel/traps.c
@@ -44,6 +44,7 @@
 #endif
 
 ATOMIC_NOTIFIER_HEAD(sparc64die_chain);
+ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
 
 int register_die_notifier(struct notifier_block *nb)
 {
@@ -57,6 +58,16 @@ int unregister_die_notifier(struct notif
 }
 EXPORT_SYMBOL(unregister_die_notifier);
 
+int register_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
+}
+
+int unregister_page_fault_notifier(struct notifier_block *nb)
+{
+	return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
+}
+
 /* When an irrecoverable trap occurs at tl > 0, the trap entry
  * code logs the trap state registers at every level in the trap
  * stack.  It is found at (pt_regs + sizeof(pt_regs)) and the layout
Index: linux-2.6.17-rc1-mm3/arch/sparc64/mm/fault.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/arch/sparc64/mm/fault.c
+++ linux-2.6.17-rc1-mm3/arch/sparc64/mm/fault.c
@@ -263,7 +263,7 @@ asmlinkage void __kprobes do_sparc64_fau
 
 	fault_code = get_thread_fault_code();
 
-	if (notify_die(DIE_PAGE_FAULT, "page_fault", regs,
+	if (notify_page_fault(DIE_PAGE_FAULT, "page_fault", regs,
 		       fault_code, 0, SIGSEGV) == NOTIFY_STOP)
 		return;
 
Index: linux-2.6.17-rc1-mm3/include/asm-sparc64/kdebug.h
===================================================================
--- linux-2.6.17-rc1-mm3.orig/include/asm-sparc64/kdebug.h
+++ linux-2.6.17-rc1-mm3/include/asm-sparc64/kdebug.h
@@ -17,7 +17,10 @@ struct die_args {
 
 extern int register_die_notifier(struct notifier_block *);
 extern int unregister_die_notifier(struct notifier_block *);
+extern int register_page_fault_notifier(struct notifier_block *);
+extern int unregister_page_fault_notifier(struct notifier_block *);
 extern struct atomic_notifier_head sparc64die_chain;
+extern struct atomic_notifier_head notify_page_fault_chain;
 
 extern void bad_trap(struct pt_regs *, long);
 
@@ -46,4 +49,17 @@ static inline int notify_die(enum die_va
 	return atomic_notifier_call_chain(&sparc64die_chain, val, &args);
 }
 
+static inline int notify_page_fault(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 atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
+}
+
+
 #endif

--

[patch 6/6] Kprobes registers for notify page fault

[Anil S Keshavamurthy]

[-- Attachment #1: notify_page_fault_kprobes.patch --]
[-- Type: text/plain, Size: 955 bytes --]

Kprobes now registers for the page fault notifications.

---
 kernel/kprobes.c |    8 ++++++++
 1 file changed, 8 insertions(+)

Index: linux-2.6.17-rc1-mm3/kernel/kprobes.c
===================================================================
--- linux-2.6.17-rc1-mm3.orig/kernel/kprobes.c
+++ linux-2.6.17-rc1-mm3/kernel/kprobes.c
@@ -544,6 +544,11 @@ static struct notifier_block kprobe_exce
 	.priority = 0x7fffffff /* we need to notified first */
 };
 
+static struct notifier_block kprobe_page_fault_nb = {
+	.notifier_call = kprobe_exceptions_notify,
+	.priority = 0x7fffffff /* we need to notified first */
+};
+
 int __kprobes register_jprobe(struct jprobe *jp)
 {
 	/* Todo: Verify probepoint is a function entry point */
@@ -654,6 +659,9 @@ static int __init init_kprobes(void)
 	if (!err)
 		err = register_die_notifier(&kprobe_exceptions_nb);
 
+	if (!err)
+		err = register_page_fault_notifier(&kprobe_page_fault_nb);
+
 	return err;
 }
 

--

Re: [patch 2/6] Notify page fault call chain for x86_64

[Andi Kleen]

On Wednesday 19 April 2006 21:01, Anil S Keshavamurthy wrote:

You seem to be missing a description/rationale here.



-Andi

> ---
> arch/x86_64/kernel/traps.c  |   12 ++++++++++++
> arch/x86_64/mm/fault.c      |    4 ++--
> include/asm-x86_64/kdebug.h |   16 ++++++++++++++++
> 3 files changed, 30 insertions(+), 2 deletions(-)

Re: [patch 2/6] Notify page fault call chain for x86_64

[Keshavamurthy Anil S]

On Wed, Apr 19, 2006 at 11:07:03PM +0200, Andi Kleen wrote:
> On Wednesday 19 April 2006 21:01, Anil S Keshavamurthy wrote:
> 
> You seem to be missing a description/rationale here.
Sorry about that, I had the description in [patch 0/6] and looks like 
I missed you cc'ing in that patch. Any way I am copying it again here.

   "Currently    in    the    do_page_fault()    code    path,   we   call
   notify_die(DIE_PAGE_FAULT, ...) to notify the page fault. The only 
   interested components for this page fault notifications  are  Kprobes  
   and/or  kdb. Since notify_die() is highly overloaded, this  page  fault  
   notification  is  currently  being  sent  to  other components
   registered   with  register_die_notification()  which  uses  the  same
   die_chain to loop for all the registered components.

   In order to optimize the do_page_fault() code path, this critical page
   fault notification is now moved to different call chain and the 
   test results showed great improvements.

   Patches for i386, x86_64, ia64, powerpc and sparc64 follows this mail."

This patch introduces the exclusive {register/unregister}_page_fault_notifier()
for use by Kprobes and kdb.


-Anil