[PATCH 0/3] Add fault_major, fault_minor page fault trace events

[PATCH 0/3] Add fault_major, fault_minor page fault trace events

[Chris Redmon]

These changes add common trace events for major and minor page faults,
as well as adding these traces to the arm and arm64 architectures. These
traces offer useful information for determining the source of page faults
in realtime systems, as well as the time penalty for taking a major or
minor page fault.

I made an attempt to minimize the overhead when these tracepoints are not
enabled, but I'm willing to make more changes if desired.

Chris Redmon (3):
  tracing: Introduce traces for major and minor page faults
  arm: Utilize trace events for major and minor page faults
  arm64: Utilize trace events for major and minor page faults

 arch/arm/mm/fault.c          | 14 +++++++
 arch/arm64/mm/fault.c        | 14 +++++++
 include/trace/events/fault.h | 87 ++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 115 insertions(+)
 create mode 100644 include/trace/events/fault.h

-- 
2.12.2.599.gcf11a67

[PATCH 1/3] tracing: Introduce traces for major and minor page faults

[Chris Redmon]

Tracing for major page faults is helpful especially for diagnosing
realtime latency issues (such as failure to mlock() something on a
realtime code path)

Signed-off-by: Chris Redmon <credmonster@gmail.com>
---
 include/trace/events/fault.h | 87 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 87 insertions(+)
 create mode 100644 include/trace/events/fault.h

diff --git a/include/trace/events/fault.h b/include/trace/events/fault.h
new file mode 100644
index 000000000000..c4803f859750
--- /dev/null
+++ b/include/trace/events/fault.h
@@ -0,0 +1,87 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM fault
+
+#if !defined(_TRACE_FAULT_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_FAULT_H
+
+#include <asm/ptrace.h>
+#include <linux/tracepoint.h>
+#include <linux/types.h>
+
+/*
+ * Event class for major/minor page faults:
+ */
+DECLARE_EVENT_CLASS(fault_major_minor_class,
+
+	TP_PROTO(
+		unsigned long address, struct pt_regs *regs,
+		unsigned long status, ktime_t start_time,
+		unsigned int flags, unsigned int fault),
+
+	TP_ARGS(address, regs, status, start_time, flags, fault),
+
+	TP_STRUCT__entry(
+		__field(unsigned long, address)
+		__field(unsigned long, instruction_pointer)
+		__field(unsigned long, status)
+		__field(unsigned int, duration)
+		__field(unsigned int, flags)
+		__field(unsigned int, fault)
+	),
+
+	TP_fast_assign(
+		__entry->address = address;
+		__entry->instruction_pointer = (regs ? instruction_pointer(regs) : 0);
+		__entry->status = (unsigned long) status;
+		__entry->duration = ktime_to_us(ktime_sub(ktime_get(), start_time));
+		__entry->flags = flags;
+		__entry->fault = fault;
+	),
+
+	TP_printk("address=%pf ip=%pf status=0x%lx duration=%uus flags=%s fault=%s",
+		(void *)__entry->address, (void *)__entry->instruction_pointer,
+		__entry->status,
+		__entry->duration,
+		__print_flags(__entry->flags, "|",
+			      {FAULT_FLAG_WRITE, "WRITE"},
+			      {FAULT_FLAG_MKWRITE, "MKWRITE"},
+			      {FAULT_FLAG_ALLOW_RETRY, "ALLOW_RETRY"},
+			      {FAULT_FLAG_RETRY_NOWAIT, "RETRY_NOWAIT"},
+			      {FAULT_FLAG_KILLABLE, "KILLABLE"},
+			      {FAULT_FLAG_TRIED, "TRIED"},
+			      {FAULT_FLAG_USER, "USER"}),
+		__print_flags(__entry->fault, "|",
+			      {VM_FAULT_OOM, "OOM"},
+			      {VM_FAULT_SIGBUS, "SIGBUS"},
+			      {VM_FAULT_MAJOR, "MAJOR"},
+			      {VM_FAULT_WRITE, "WRITE"},
+			      {VM_FAULT_HWPOISON, "HWPOISON"},
+			      {VM_FAULT_HWPOISON_LARGE, "HWPOISON_LARGE"},
+			      {VM_FAULT_NOPAGE, "NOPAGE"},
+			      {VM_FAULT_LOCKED, "LOCKED"},
+			      {VM_FAULT_RETRY, "RETRY"},
+			      {VM_FAULT_FALLBACK, "FALLBACK"})
+		)
+);
+
+#define DEFINE_PAGE_FAULT_MAJOR_MINOR_EVENT(name)			\
+DEFINE_EVENT(fault_major_minor_class, name,				\
+	TP_PROTO(unsigned long address,	struct pt_regs *regs,		\
+		  unsigned long status, ktime_t start_time,		\
+		  unsigned int flags, unsigned int fault),		\
+	TP_ARGS(address, regs, status, start_time, flags, fault));
+
+/*
+ * Tracepoint for major page faults:
+ */
+DEFINE_PAGE_FAULT_MAJOR_MINOR_EVENT(fault_major);
+
+/*
+ * Tracepoint for minor page faults:
+ */
+DEFINE_PAGE_FAULT_MAJOR_MINOR_EVENT(fault_minor);
+
+#endif /* _TRACE_FAULT_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>
-- 
2.12.2.599.gcf11a67

[PATCH 2/3] arm: Utilize trace events for major and minor page faults

[Chris Redmon]

Signed-off-by: Chris Redmon <credmonster@gmail.com>
---
 arch/arm/mm/fault.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/arch/arm/mm/fault.c b/arch/arm/mm/fault.c
index ff8b0aa2dfde..e59514f85e01 100644
--- a/arch/arm/mm/fault.c
+++ b/arch/arm/mm/fault.c
@@ -29,6 +29,9 @@
 
 #include "fault.h"
 
+#define CREATE_TRACE_POINTS
+#include <trace/events/fault.h>
+
 #ifdef CONFIG_MMU
 
 #ifdef CONFIG_KPROBES
@@ -262,6 +265,8 @@ do_page_fault(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
 	struct mm_struct *mm;
 	int fault, sig, code;
 	unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
+	ktime_t start_time;
+	bool trace_fault_enabled = false;
 
 	if (notify_page_fault(regs, fsr))
 		return 0;
@@ -285,6 +290,11 @@ do_page_fault(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
 	if (fsr & FSR_WRITE)
 		flags |= FAULT_FLAG_WRITE;
 
+	if (trace_fault_major_enabled() || trace_fault_minor_enabled()) {
+		start_time = ktime_get();
+		trace_fault_enabled = true;
+	}
+
 	/*
 	 * As per x86, we may deadlock here.  However, since the kernel only
 	 * validly references user space from well defined areas of the code,
@@ -330,10 +340,14 @@ do_page_fault(unsigned long addr, unsigned int fsr, struct pt_regs *regs)
 			tsk->maj_flt++;
 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 					regs, addr);
+			if (trace_fault_enabled)
+				trace_fault_major(addr, regs, fsr, start_time, flags, fault);
 		} else {
 			tsk->min_flt++;
 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 					regs, addr);
+			if (trace_fault_enabled)
+				trace_fault_minor(addr, regs, fsr, start_time, flags, fault);
 		}
 		if (fault & VM_FAULT_RETRY) {
 			/* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
-- 
2.12.2.599.gcf11a67

[PATCH 3/3] arm64: Utilize trace events for major and minor page faults

[Chris Redmon]

Signed-off-by: Chris Redmon <credmonster@gmail.com>
---
 arch/arm64/mm/fault.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/arch/arm64/mm/fault.c b/arch/arm64/mm/fault.c
index 1b35b8bddbfb..d3097ff6fb8e 100644
--- a/arch/arm64/mm/fault.c
+++ b/arch/arm64/mm/fault.c
@@ -42,6 +42,9 @@
 #include <asm/pgtable.h>
 #include <asm/tlbflush.h>
 
+#define CREATE_TRACE_POINTS
+#include <trace/events/fault.h>
+
 struct fault_info {
 	int	(*fn)(unsigned long addr, unsigned int esr,
 		      struct pt_regs *regs);
@@ -315,6 +318,8 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr,
 	int fault, sig, code;
 	unsigned long vm_flags = VM_READ | VM_WRITE;
 	unsigned int mm_flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
+	ktime_t start_time;
+	bool trace_fault_enabled = false;
 
 	if (notify_page_fault(regs, esr))
 		return 0;
@@ -351,6 +356,11 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr,
 			die("Accessing user space memory outside uaccess.h routines", regs, esr);
 	}
 
+	if (trace_fault_major_enabled() || trace_fault_minor_enabled()) {
+		start_time = ktime_get();
+		trace_fault_enabled = true;
+	}
+
 	/*
 	 * As per x86, we may deadlock here. However, since the kernel only
 	 * validly references user space from well defined areas of the code,
@@ -395,10 +405,14 @@ static int __kprobes do_page_fault(unsigned long addr, unsigned int esr,
 			tsk->maj_flt++;
 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs,
 				      addr);
+			if (trace_fault_enabled)
+				trace_fault_major(addr, regs, esr, start_time, mm_flags, fault);
 		} else {
 			tsk->min_flt++;
 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs,
 				      addr);
+			if (trace_fault_enabled)
+				trace_fault_minor(addr, regs, esr, start_time, mm_flags, fault);
 		}
 		if (fault & VM_FAULT_RETRY) {
 			/*
-- 
2.12.2.599.gcf11a67

[PATCH 0/3] Add fault_major, fault_minor page fault trace events

[Will Deacon]

On Wed, Apr 12, 2017 at 10:20:37PM -0400, Chris Redmon wrote:
> These changes add common trace events for major and minor page faults,
> as well as adding these traces to the arm and arm64 architectures. These
> traces offer useful information for determining the source of page faults
> in realtime systems, as well as the time penalty for taking a major or
> minor page fault.
> 
> I made an attempt to minimize the overhead when these tracepoints are not
> enabled, but I'm willing to make more changes if desired.

Why is this necessary, given that we already report these types of fault
as perf sw events?

Will