Skip to contentSkip to conversation

Add page fault trace event definitions

PATCH In review Francis Deslauriers 5 participants 6 messages May 9, 2013 – May 13, 2013

Subsystems (5)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 Add page fault trace event definitions

Add page_fault_entry and page_fault_exit event definitions. It will allow each architecture to instrument their page faults.

---
 include/trace/events/fault.h |   51 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 51 insertions(+)
 create mode 100644 include/trace/events/fault.h
include/trace/events/fault.h new file +51 -0
new file mode 100644
index 0000000..522ddee
@@ -0,0 +1,51 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM fault
+
+#if !defined(_TRACE_FAULT_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_FAULT_H
+
+#include <linux/tracepoint.h>
+
+TRACE_EVENT(page_fault_entry,
+
+	TP_PROTO(struct pt_regs *regs, unsigned long address,
+					int write_access),
+
+	TP_ARGS(regs, address, write_access),
+
+	TP_STRUCT__entry(
+		__field(	unsigned long,	ip	)
+		__field(	unsigned long,	addr	)
+		__field(	uint8_t,	write	)
+	),
+
+	TP_fast_assign(
+		__entry->ip	= regs ? instruction_pointer(regs) : 0UL;
+		__entry->addr	= address;
+		__entry->write	= !!write_access;
+	),
+
+	TP_printk("ip=%lu addr=%lu write_access=%d",
+		__entry->ip, __entry->addr, __entry->write)
+);
+
+TRACE_EVENT(page_fault_exit,
+
+	TP_PROTO(int result),
+
+	TP_ARGS(result),
+
+	TP_STRUCT__entry(
+		__field(	int,	res	)
+	),
+
+	TP_fast_assign(
+		__entry->res	= result;
+	),
+
+	TP_printk("result=%d", __entry->res)
+);
+
+#endif /* _TRACE_FAULT_H */
+/* This part must be outside protection */
+#include <trace/define_trace.h>
Signature · 7 lines
1.7.10.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#1 lore raw

PATCH 2/2 x86:Instruments page fault trace event

Reviewed-by RaphaA<<l Beamonte <[email protected]>
---
 arch/x86/mm/fault.c |   11 +++++++++++
 mm/memory.c         |    5 +++++
 2 files changed, 16 insertions(+)
arch/x86/mm/fault.c +11 -0
index 654be4a..e227828 100644
@@ -20,6 +20,9 @@
 #include <asm/kmemcheck.h>		/* kmemcheck_*(), ...		*/
 #include <asm/fixmap.h>			/* VSYSCALL_START		*/
 
+#define CREATE_TRACE_POINTS
+#include <trace/events/fault.h>		/* trace_page_fault_*(), ...	*/
+
 /*
  * Page fault error code bits:
  *
@@ -756,12 +759,18 @@ __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 
 		if (likely(show_unhandled_signals))
 			show_signal_msg(regs, error_code, address, tsk);
+		trace_page_fault_entry(regs, address, error_code & PF_WRITE);
 
 		tsk->thread.cr2		= address;
 		tsk->thread.error_code	= error_code;
 		tsk->thread.trap_nr	= X86_TRAP_PF;
 
 		force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
+		/*
+		 * Using -1 here, since there is no VM_FAULT flag to identify
+		 * user accesses triggering SIGSEGV.
+		 */
+		trace_page_fault_exit(-1);
 
 		return;
 	}
@@ -1185,7 +1194,9 @@ good_area:
 	 * make sure we exit gracefully rather than endlessly redo
 	 * the fault:
 	 */
+	trace_page_fault_entry(regs, address, write);
 	fault = handle_mm_fault(mm, vma, address, flags);
+	trace_page_fault_exit(fault);
 
 	if (unlikely(fault & (VM_FAULT_RETRY|VM_FAULT_ERROR))) {
 		if (mm_fault_error(regs, error_code, address, fault))
mm/memory.c +5 -0
index 6dc1882..0bd86f8 100644
@@ -67,6 +67,8 @@
 #include <asm/tlbflush.h>
 #include <asm/pgtable.h>
 
+#include <trace/events/fault.h>
+
 #include "internal.h"
 
 #ifdef LAST_NID_NOT_IN_PAGE_FLAGS
@@ -1829,8 +1831,11 @@ long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 				if (foll_flags & FOLL_NOWAIT)
 					fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
 
+				trace_page_fault_entry(0, start,
+						foll_flags & FOLL_WRITE);
 				ret = handle_mm_fault(mm, vma, start,
 							fault_flags);
+				trace_page_fault_exit(ret);
 
 				if (ret & VM_FAULT_ERROR) {
 					if (ret & VM_FAULT_OOM)
Signature · 7 lines
1.7.10.4

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
#2 lore raw
In reply to Francis Deslauriers
Quoted 2 lines · Francis Deslauriers

Add page_fault_entry and page_fault_exit event definitions. It will allow each architecture to instrument their page faults.

I'm wondering if this tracepoint could handle other page faults, like faults in kernel memory(vmalloc, kmmio, etc...)

And if we decide to support those faults, add a type annotate in TP_printk would be much helpful for user, to let user know what type of page faults happened.

Thanks.

Quoted 57 lines
---
 include/trace/events/fault.h |   51 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 51 insertions(+)
 create mode 100644 include/trace/events/fault.h
include/trace/events/fault.h new file +41 -0
new file mode 100644
index 0000000..522ddee
@@ -0,0 +1,51 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM fault
+
+#if !defined(_TRACE_FAULT_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_FAULT_H
+
+#include <linux/tracepoint.h>
+
+TRACE_EVENT(page_fault_entry,
+
+	TP_PROTO(struct pt_regs *regs, unsigned long address,
+					int write_access),
+
+	TP_ARGS(regs, address, write_access),
+
+	TP_STRUCT__entry(
+		__field(	unsigned long,	ip	)
+		__field(	unsigned long,	addr	)
+		__field(	uint8_t,	write	)
+	),
+
+	TP_fast_assign(
+		__entry->ip	= regs ? instruction_pointer(regs) : 0UL;
+		__entry->addr	= address;
+		__entry->write	= !!write_access;
+	),
+
+	TP_printk("ip=%lu addr=%lu write_access=%d",
+		__entry->ip, __entry->addr, __entry->write)
+);
+
+TRACE_EVENT(page_fault_exit,
+
+	TP_PROTO(int result),
+
+	TP_ARGS(result),
+
+	TP_STRUCT__entry(
+		__field(	int,	res	)
+	),
+

include/trace/events/fault.h line 51

Quoted 10 lines
include/trace/events/fault.h new file +10 -0
+	TP_fast_assign(
+		__entry->res	= result;
+	),
+
+	TP_printk("result=%d", __entry->res)
+);
+
+#endif /* _TRACE_FAULT_H */
+/* This part must be outside protection */
+#include <trace/define_trace.h>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#3 lore raw
Quoted 10 lines · zhangwei(Jovi)
Quoted 2 lines · Francis Deslauriers

Add page_fault_entry and page_fault_exit event definitions. It will allow each architecture to instrument their page faults.

I'm wondering if this tracepoint could handle other page faults, like faults in kernel memory(vmalloc, kmmio, etc...)

And if we decide to support those faults, add a type annotate in TP_printk would be much helpful for user, to let user know what type of page faults happened.

The plan for x86 was to switch the IDT so that any exception could get a trace event without any overhead in normal operation. This has been in the process for quite some time but looks like it was getting very close.

-hpa

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

#4 lore raw
Quoted 8 lines · * H. Peter Anvin ([email protected]) wrote:
Quoted 7 lines · zhangwei(Jovi)
Quoted 2 lines · Francis Deslauriers

Add page_fault_entry and page_fault_exit event definitions. It will allow each architecture to instrument their page faults.

I'm wondering if this tracepoint could handle other page faults, like faults in kernel memory(vmalloc, kmmio, etc...)

Quoted 7 lines
Quoted 3 lines

And if we decide to support those faults, add a type annotate in TP_printk would be much helpful for user, to let user know what type of page faults happened.

The plan for x86 was to switch the IDT so that any exception could get a trace event without any overhead in normal operation. This has been in the process for quite some time but looks like it was getting very close.

Hi Peter,

Who is leading this IDT instrumentation effort ?

Since we have tracepoints in interrupt handlers nowadays, I wonder what makes traps so much more special than interrupts to require the arch-specific complexity of the IDT switcharoo trick ? If I had to guess, the reason for this would be the page fault handler, which is called way too frequently for its own good. The number of page faults triggered by COW on process fork has been impressively high for the past couple of years.

IMHO, this should be one extra reason for quickly allowing people to trace those page faults, so they can get an idea of their tremendous performance impact. This could speed up the efforts on transparent huge pages, which seems to be a viable long-term solution to this page-size scalability issue.

By default, my 3.5 Linux kernel (Debian) has:

$ cat /sys/kernel/mm/transparent_hugepage/enabled
always [madvise] never

I think transparent huge pages will become generally useful when enabled by default, and when they will handle the page cache in addition to anonymous pages.[1]

Thanks,

Mathieu

[1] Documentation/vm/transhuge.txt

Signature · 9 lines
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
maintainer
#5 lore raw
Quoted 4 lines · Mathieu Desnoyers

* H. Peter Anvin ([email protected]) wrote:

Who is leading this IDT instrumentation effort ?

Seiji has been doing most of the work. I've just been busy doing other things but I need to start getting this tidied up, and hopefully this can get into 3.11.

https://lkml.org/lkml/2013/4/5/401

-- Steve

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>