ftrace for MIPS

From: Wu Zhangjin <[email protected]>

This patchset adds ftrace support for MIPS, including static ftracer, dynamic ftracer and function graph tracer, and also, a MIPS specific trace_clock_local() function is added to get high precision timestamp.

Differ from the v7, This v8 revision incoporates the feedbacks from David, Ralf and Maciej. thanks goes to them.

o Remove the needless -mlong-calls introduced in the old "tracing: add static function tracer support for MIPS"

static ftracer already support module tracing for there is an existing -mlong-calls for modules in arch/mips/Makefile:
MODFLAGS += -mlong-calls

o Make dynamic ftracer work without -mlong-calls.

Now, we need to consider the kernel and module respectively, the kernel part is without -mlong-calls, the module part is with -mlong-calls.

In scripts/recordmcount.pl, we need to get the right offset, and in arch/mips/kernel/ftrace.c, we need to replace the right instruction in ftrace_make_nop() & ftrace_make_call().

And what I must mention here:

I. we will not really get the offset of calling to _mcount in scripts/recordmcount.pl for the modules with -mlong-calls, we get the offset of "lui,v1,hi16_mcount".

II. and in the ftrace_make_nop(), we not really replace the instruction by a nop, but by a "b 1f" instead. so, ftrace_make_nop not mean its original meaning.

o Make function graph tracer work without -mlong-calls

'Cause -pg with and without -mlong-calls generate different number of instructions, we need to consider this and move the searching pointer to the right position in ftrace_get_parent_addr().

All of the updates have been tested with gcc-4.4/binutils-2.19.1 and gcc-4.5(including David's mcount-ra-address)/binutils-2.20 on a YeeLoong netbook. and the source code have been pushed into:

git://dev.lemote.com/rt4ls.git linux-mips/dev/ftrace-upstream

BTW: What about this revision? Hope it is the last revision from me, but I can not promise for the kernel developers are always "rigorous": they need the best solution ;) So, more feedbacks are welcome!

Thanks & Regards,
Wu Zhangjin

Wu Zhangjin (16):
tracing: convert trace_clock_local() as weak function
tracing: add mips_timecounter_read() for MIPS
tracing: add MIPS specific trace_clock_local()
tracing: add static function tracer support for MIPS
tracing: enable HAVE_FUNCTION_TRACE_MCOUNT_TEST for MIPS
tracing: add an endian argument to scripts/recordmcount.pl
tracing: add dynamic function tracer support for MIPS
tracing: add IRQENTRY_EXIT section for MIPS
tracing: define a new __time_notrace annotation flag
tracing: not trace the timecounter_read* in kernel/time/clocksource.c
tracing: not trace mips_timecounter_read() for MIPS
tracing: add function graph tracer support for MIPS
tracing: add dynamic function graph tracer for MIPS
tracing: make ftrace for MIPS work without -fno-omit-frame-pointer
tracing: reserve $12(t0) for mcount-ra-address of gcc 4.5
tracing: make function graph tracer work with -mmcount-ra-address

 arch/mips/Kconfig              |    5 +
 arch/mips/Makefile             |    9 ++
 arch/mips/include/asm/ftrace.h |   97 ++++++++++++++-
 arch/mips/include/asm/irq.h    |   29 +----
 arch/mips/include/asm/time.h   |   20 +++
 arch/mips/kernel/Makefile      |    9 ++
 arch/mips/kernel/csrc-r4k.c    |   42 ++++++
 arch/mips/kernel/ftrace.c      |  275 ++++++++++++++++++++++++++++++++++++++++
 arch/mips/kernel/irq.c         |   30 +++++
 arch/mips/kernel/mcount.S      |  189 +++++++++++++++++++++++++++
 arch/mips/kernel/mips_ksyms.c  |    5 +
 arch/mips/kernel/smp.c         |    3 +-
 arch/mips/kernel/smtc.c        |   21 ++-
 arch/mips/kernel/time.c        |    2 +
 arch/mips/kernel/trace_clock.c |   33 +++++
 arch/mips/kernel/vmlinux.lds.S |    1 +
 arch/mips/sgi-ip22/ip22-int.c  |    3 +-
 arch/mips/sgi-ip22/ip22-time.c |    3 +-
 include/linux/ftrace.h         |   12 ++
 kernel/time/clocksource.c      |    5 +-
 kernel/trace/trace_clock.c     |    2 +-
 scripts/Makefile.build         |    1 +
 scripts/recordmcount.pl        |   60 ++++++++-
 23 files changed, 812 insertions(+), 44 deletions(-)
 create mode 100644 arch/mips/kernel/ftrace.c
 create mode 100644 arch/mips/kernel/mcount.S
 create mode 100644 arch/mips/kernel/trace_clock.c

From: Wu Zhangjin <[email protected]>

trace_clock_local() is based on the arch-specific sched_clock(), in X86, it is tsc(64bit) based, which can give very high precision(about 1ns with 1GHz). but in MIPS, the sched_clock() is jiffies based, which can give only 10ms precison with 1000 HZ. which is not enough for tracing, especially for Real Time system.

so, we need to implement a MIPS specific sched_clock() to get higher precision. There is a tsc like clock counter register in MIPS, whose frequency is half of the processor, so, if the cpu frequency is 800MHz, the time precision reaches 2.5ns, which is very good for tracing, even for Real Time system.

but 'Cause it is only 32bit long, which will rollover quickly, so, such a sched_clock() will bring with extra load, which is not good for the whole system. so, we only need to implement a arch-specific trace_clock_local() for tracing. as a preparation, we convert it as a weak function.

The MIPS specific trace_clock_local() is coming in the next two patches.

---
 kernel/trace/trace_clock.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

index 20c5f92..06b8cd2 100644
@@ -26,7 +26,7 @@
  * Useful for tracing that does not cross to other CPUs nor
  * does it go through idle events.
  */
-u64 notrace trace_clock_local(void)
+u64 __weak notrace trace_clock_local(void)
 {
 	unsigned long flags;
 	u64 clock;

1.6.2.1

From: Wu Zhangjin <[email protected]>

This patch add a new function: mips_timecounter_read() to get high precision timestamp without extra lock.

It is based on the clock counter register and the timecounter/cyclecounter abstraction layer of kernel.

---
 arch/mips/include/asm/time.h |   20 ++++++++++++++++++++
 arch/mips/kernel/csrc-r4k.c  |   41 +++++++++++++++++++++++++++++++++++++++++
 arch/mips/kernel/time.c      |    2 ++
 3 files changed, 63 insertions(+), 0 deletions(-)

index df6a430..3fcc481 100644
@@ -15,6 +15,7 @@
 #define _ASM_TIME_H
 
 #include <linux/rtc.h>
+#include <linux/sched.h>
 #include <linux/spinlock.h>
 #include <linux/clockchips.h>
 #include <linux/clocksource.h>

@@ -73,8 +74,18 @@ static inline int mips_clockevent_init(void)
  */
 #ifdef CONFIG_CSRC_R4K_LIB
 extern int init_r4k_clocksource(void);
+extern int init_r4k_timecounter(void);
+extern u64 r4k_timecounter_read(void);
 #endif
 
+static inline u64 mips_timecounter_read(void)
+{
+#ifdef CONFIG_CSRC_R4K
+	return r4k_timecounter_read();
+#else
+	return sched_clock();
+#endif
+}
 static inline int init_mips_clocksource(void)
 {
 #ifdef CONFIG_CSRC_R4K

@@ -84,6 +95,15 @@ static inline int init_mips_clocksource(void)
 #endif
 }
 
+static inline int init_mips_timecounter(void)
+{
+#ifdef CONFIG_CSRC_R4K
+	return init_r4k_timecounter();
+#else
+	return 0;
+#endif
+}
+
 extern void clocksource_set_clock(struct clocksource *cs, unsigned int clock);
 extern void clockevent_set_clock(struct clock_event_device *cd,
 		unsigned int clock);

index 1f467d5..e38cca1 100644
@@ -156,4 +156,6 @@ void __init time_init(void)
 
 	if (!mips_clockevent_init() || !cpu_has_mfc0_count_bug())
 		init_mips_clocksource();
+	if (!cpu_has_mfc0_count_bug())
+		init_mips_timecounter();
 }

1.6.2.1

From: Wu Zhangjin <[email protected]>

This patch add the mips_timecounter_read() based high precision version of trace_clock_local().

---
 arch/mips/kernel/Makefile      |    6 ++++++
 arch/mips/kernel/trace_clock.c |   33 +++++++++++++++++++++++++++++++++
 2 files changed, 39 insertions(+), 0 deletions(-)
 create mode 100644 arch/mips/kernel/trace_clock.c

index 8e26e9c..f228868 100644
@@ -8,6 +8,10 @@ obj-y		+= cpu-probe.o branch.o entry.o genex.o irq.o process.o \
 		   ptrace.o reset.o setup.o signal.o syscall.o \
 		   time.o topology.o traps.o unaligned.o watch.o
 
+ifdef CONFIG_FUNCTION_TRACER
+CFLAGS_REMOVE_trace_clock.o = -pg
+endif
+
 obj-$(CONFIG_CEVT_BCM1480)	+= cevt-bcm1480.o
 obj-$(CONFIG_CEVT_R4K_LIB)	+= cevt-r4k.o
 obj-$(CONFIG_MIPS_MT_SMTC)	+= cevt-smtc.o

@@ -24,6 +28,8 @@ obj-$(CONFIG_SYNC_R4K)		+= sync-r4k.o
 obj-$(CONFIG_STACKTRACE)	+= stacktrace.o
 obj-$(CONFIG_MODULES)		+= mips_ksyms.o module.o
 
+obj-$(CONFIG_FTRACE)		+= trace_clock.o
+
 obj-$(CONFIG_CPU_LOONGSON2)	+= r4k_fpu.o r4k_switch.o
 obj-$(CONFIG_CPU_MIPS32)	+= r4k_fpu.o r4k_switch.o
 obj-$(CONFIG_CPU_MIPS64)	+= r4k_fpu.o r4k_switch.o

new file mode 100644
index 0000000..0fe85ea
@@ -0,0 +1,33 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive for
+ * more details.
+ *
+ * Copyright (C) 2009 Lemote Inc. & DSLab, Lanzhou University, China
+ * Author: Wu Zhangjin <[email protected]>
+ */
+
+#include <linux/clocksource.h>
+#include <linux/sched.h>
+
+#include <asm/time.h>
+
+/*
+ * trace_clock_local(): the simplest and least coherent tracing clock.
+ *
+ * Useful for tracing that does not cross to other CPUs nor
+ * does it go through idle events.
+ */
+u64 trace_clock_local(void)
+{
+	unsigned long flags;
+	u64 clock;
+
+	raw_local_irq_save(flags);
+
+	clock = mips_timecounter_read();
+
+	raw_local_irq_restore(flags);
+
+	return clock;
+}

1.6.2.1

From: Wu Zhangjin <[email protected]>

If -pg of gcc is enabled with CONFIG_FUNCTION_TRACER=y. a calling to _mcount will be inserted into each kernel function. so, there is a possibility to trace the kernel functions in _mcount.

This patch add the MIPS specific _mcount support for static function tracing. by default, ftrace_trace_function is initialized as ftrace_stub(an empty function), so, the default _mcount will introduce very little overhead. after enabling ftrace in user-space, it will jump to a real tracing function and do static function tracing for us.

and -ffunction-sections is incompatible with -pg, so, disable it when ftracer is enabled.

---
 arch/mips/Kconfig              |    1 +
 arch/mips/Makefile             |    2 +
 arch/mips/include/asm/ftrace.h |   25 +++++++++++-
 arch/mips/kernel/Makefile      |    2 +
 arch/mips/kernel/mcount.S      |   83 ++++++++++++++++++++++++++++++++++++++++
 arch/mips/kernel/mips_ksyms.c  |    5 ++
 6 files changed, 117 insertions(+), 1 deletions(-)
 create mode 100644 arch/mips/kernel/mcount.S

index 03bd56a..6b33e88 100644
@@ -4,6 +4,7 @@ config MIPS
 	select HAVE_IDE
 	select HAVE_OPROFILE
 	select HAVE_ARCH_KGDB
+	select HAVE_FUNCTION_TRACER
 	# Horrible source of confusion.  Die, die, die ...
 	select EMBEDDED
 	select RTC_LIB if !LEMOTE_FULOONG2E

index 77f5021..0a2073e 100644
@@ -48,7 +48,9 @@ ifneq ($(SUBARCH),$(ARCH))
   endif
 endif
 
+ifndef CONFIG_FUNCTION_TRACER
 cflags-y := -ffunction-sections
+endif
 cflags-y += $(call cc-option, -mno-check-zero-division)
 
 ifdef CONFIG_32BIT

index 40a8c17..5f8ebcf 100644
@@ -1 +1,24 @@
-/* empty */
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive for
+ * more details.
+ *
+ * Copyright (C) 2009 DSLab, Lanzhou University, China
+ * Author: Wu Zhangjin <[email protected]>
+ */
+
+#ifndef _ASM_MIPS_FTRACE_H
+#define _ASM_MIPS_FTRACE_H
+
+#ifdef CONFIG_FUNCTION_TRACER
+
+#define MCOUNT_ADDR ((unsigned long)(_mcount))
+#define MCOUNT_INSN_SIZE 4		/* sizeof mcount call */
+
+#ifndef __ASSEMBLY__
+extern void _mcount(void);
+#define mcount _mcount
+
+#endif /* __ASSEMBLY__ */
+#endif /* CONFIG_FUNCTION_TRACER */
+#endif /* _ASM_MIPS_FTRACE_H */

index f228868..38e704a 100644
@@ -10,6 +10,7 @@ obj-y		+= cpu-probe.o branch.o entry.o genex.o irq.o process.o \
 
 ifdef CONFIG_FUNCTION_TRACER
 CFLAGS_REMOVE_trace_clock.o = -pg
+CFLAGS_REMOVE_early_printk.o = -pg
 endif
 
 obj-$(CONFIG_CEVT_BCM1480)	+= cevt-bcm1480.o

@@ -29,6 +30,7 @@ obj-$(CONFIG_STACKTRACE)	+= stacktrace.o
 obj-$(CONFIG_MODULES)		+= mips_ksyms.o module.o
 
 obj-$(CONFIG_FTRACE)		+= trace_clock.o
+obj-$(CONFIG_FUNCTION_TRACER)	+= mcount.o
 
 obj-$(CONFIG_CPU_LOONGSON2)	+= r4k_fpu.o r4k_switch.o
 obj-$(CONFIG_CPU_MIPS32)	+= r4k_fpu.o r4k_switch.o

index 225755d..1d04807 100644
@@ -13,6 +13,7 @@
 #include <asm/checksum.h>
 #include <asm/pgtable.h>
 #include <asm/uaccess.h>
+#include <asm/ftrace.h>
 
 extern void *__bzero(void *__s, size_t __count);
 extern long __strncpy_from_user_nocheck_asm(char *__to,

@@ -51,3 +52,7 @@ EXPORT_SYMBOL(csum_partial_copy_nocheck);
 EXPORT_SYMBOL(__csum_partial_copy_user);
 
 EXPORT_SYMBOL(invalid_pte_table);
+#ifdef CONFIG_FUNCTION_TRACER
+/* _mcount is defined in arch/mips/kernel/mcount.S */
+EXPORT_SYMBOL(_mcount);
+#endif

1.6.2.1

From: Wu Zhangjin <[email protected]>

There is an exisiting common ftrace_test_stop_func() in kernel/trace/ftrace.c, which is used to check the global variable ftrace_trace_stop to determine whether stop the function tracing.

This patch implepment the MIPS specific one to speedup the procedure.

Thanks goes to Zhang Le for Cleaning it up.

---
 arch/mips/Kconfig         |    1 +
 arch/mips/kernel/mcount.S |    3 +++
 2 files changed, 4 insertions(+), 0 deletions(-)

index 6b33e88..a9bd992 100644
@@ -5,6 +5,7 @@ config MIPS
 	select HAVE_OPROFILE
 	select HAVE_ARCH_KGDB
 	select HAVE_FUNCTION_TRACER
+	select HAVE_FUNCTION_TRACE_MCOUNT_TEST
 	# Horrible source of confusion.  Die, die, die ...
 	select EMBEDDED
 	select RTC_LIB if !LEMOTE_FULOONG2E

index cebcc3c..cbb45ed 100644
@@ -59,6 +59,9 @@
 	.endm
 
 NESTED(_mcount, PT_SIZE, ra)
+	lw	t0, function_trace_stop
+	bnez	t0, ftrace_stub
+	 nop
 	PTR_LA	t0, ftrace_stub
 	PTR_L	t1, ftrace_trace_function /* Prepare t1 for (1) */
 	bne	t0, t1, static_trace

1.6.2.1

From: Wu Zhangjin <[email protected]>

MIPS and some other architectures need this argument to handle big/little endian respectively.

---
 scripts/Makefile.build  |    1 +
 scripts/recordmcount.pl |    6 +++---
 2 files changed, 4 insertions(+), 3 deletions(-)

index 341b589..0b94d2f 100644
@@ -207,6 +207,7 @@ endif
 
 ifdef CONFIG_FTRACE_MCOUNT_RECORD
 cmd_record_mcount = set -e ; perl $(srctree)/scripts/recordmcount.pl "$(ARCH)" \
+	"$(if $(CONFIG_CPU_BIG_ENDIAN),big,little)" \
 	"$(if $(CONFIG_64BIT),64,32)" \
 	"$(OBJDUMP)" "$(OBJCOPY)" "$(CC)" "$(LD)" "$(NM)" "$(RM)" "$(MV)" \
 	"$(if $(part-of-module),1,0)" "$(@)";

index a4e2435..62a0001 100755
@@ -113,13 +113,13 @@ $P =~ s@.*/@@g;
 
 my $V = '0.1';
 
-if ($#ARGV != 10) {
-	print "usage: $P arch bits objdump objcopy cc ld nm rm mv is_module inputfile\n";
+if ($#ARGV != 11) {
+	print "usage: $P arch endian bits objdump objcopy cc ld nm rm mv is_module inputfile\n";
 	print "version: $V\n";
 	exit(1);
 }
 
-my ($arch, $bits, $objdump, $objcopy, $cc,
+my ($arch, $endian, $bits, $objdump, $objcopy, $cc,
     $ld, $nm, $rm, $mv, $is_module, $inputfile) = @ARGV;
 
 # This file refers to mcount and shouldn't be ftraced, so lets' ignore it

1.6.2.1

From: Wu Zhangjin <[email protected]>

With dynamic function tracer, by default, _mcount is defined as an "empty" function, it returns directly without any more action . When enabling it in user-space, it will jump to a real tracing function(ftrace_caller), and do the real job for us.

Differ from the static function tracer, dynamic function tracer provides two functions ftrace_make_call()/ftrace_make_nop() to enable/disable the tracing of some indicated kernel functions(set_ftrace_filter).

In the -v4 version, the implementation of this support is basically the same as X86 version does: _mcount is implemented as an empty function and ftrace_caller is implemented as a real tracing function respectively.

But in this version, to support module tracing with the help of -mlong-calls in arch/mips/Makefile:

MODFLAGS += -mlong-calls.

The stuff becomes a little more complex. We need to cope with two different type of calling to _mcount.

For the kernel part, the calling to _mcount(result of "objdump -hdr vmlinux"). is like this:

	108:   03e0082d        move    at,ra
	10c:   0c000000        jal     0 <fpcsr_pending>

10c: R_MIPS_26 _mcount
10c: R_MIPS_NONE *ABS*
10c: R_MIPS_NONE *ABS*
110: 00020021 nop

For the module with -mlong-calls, it looks like this:

	c:	3c030000 	lui	v1,0x0
			c: R_MIPS_HI16	_mcount
			c: R_MIPS_NONE	*ABS*
			c: R_MIPS_NONE	*ABS*
	10:	64630000 	daddiu	v1,v1,0
			10: R_MIPS_LO16	_mcount
			10: R_MIPS_NONE	*ABS*
			10: R_MIPS_NONE	*ABS*
	14:	03e0082d 	move	at,ra
	18:	0060f809 	jalr	v1

In the kernel version, there is only one "_mcount" string for every kernel function, so, we just need to match this one in mcount_regex of scripts/recordmcount.pl, but in the module version, we need to choose one of the two to match. Herein, I choose the first one with "R_MIPS_HI16 _mcount".

and In the kernel verion, without module tracing support, we just need to replace "jal _mcount" by "jal ftrace_caller" to do real tracing, and filter the tracing of some kernel functions via replacing it by a nop instruction.

but as we have described before, the instruction "jal ftrace_caller" only left 32bit length for the address of ftrace_caller, it will fail when calling from the module space. so, herein, we must replace something else.

the basic idea is loading the address of ftrace_caller to v1 via changing these two instructions:

	lui	v1,0x0
	addiu	v1,v1,0

If we want to enable the tracing, we need to replace the above instructions to:

	lui	v1, HI_16BIT_ftrace_caller
	addiu	v1, v1, LOW_16BIT_ftrace_caller

If we want to stop the tracing of the indicated kernel functions, we just need to replace the "jalr v1" to a nop instruction. but we need to replace two instructions and encode the above two instructions oursevles.

Is there a simpler solution? Yes! Here it is, in this version, we put _mcount and ftrace_caller together, which means the address of _mcount and ftrace_caller is the same:

_mcount:
ftrace_caller:

	j	ftrace_stub
	 nop

	...(do real tracing here)...

ftrace_stub:
jr ra
move ra, at

By default, the kernel functions call _mcount, and then jump to ftrace_stub and return. and when we want to do real tracing, we just need to remove that "j ftrace_stub", and it will run through the two "nop" instructions and then do the real tracing job.

what about filtering job? we just need to do this:

	 lui v1, hi_16bit_of_mcount        <--> b 1f (0x10000004)
	 addiu v1, v1, low_16bit_of_mcount
	 move at, ra
	 jalr v1
	 nop
	 				     1f: (rec->ip + 12)

In linux-mips64, there will be some local symbols, whose name are prefixed by $L, which need to be filtered. thanks goes to Steven for writing the mips64-specific function_regex.

In a conclusion, with RISC, things becomes easier with such a "stupid" trick, RISC is something like K.I.S.S, and also, there are lots of "simple" tricks in the whole ftrace support, thanks goes to Steven and the other folks for providing such a wonderful tracing framework!

---
 arch/mips/Kconfig              |    2 +
 arch/mips/include/asm/ftrace.h |    9 +++
 arch/mips/kernel/Makefile      |    3 +-
 arch/mips/kernel/ftrace.c      |  112 ++++++++++++++++++++++++++++++++++++++++
 arch/mips/kernel/mcount.S      |   29 ++++++++++
 scripts/recordmcount.pl        |   54 +++++++++++++++++++
 6 files changed, 208 insertions(+), 1 deletions(-)
 create mode 100644 arch/mips/kernel/ftrace.c

index a9bd992..147fbbc 100644
@@ -6,6 +6,8 @@ config MIPS
 	select HAVE_ARCH_KGDB
 	select HAVE_FUNCTION_TRACER
 	select HAVE_FUNCTION_TRACE_MCOUNT_TEST
+	select HAVE_DYNAMIC_FTRACE
+	select HAVE_FTRACE_MCOUNT_RECORD
 	# Horrible source of confusion.  Die, die, die ...
 	select EMBEDDED
 	select RTC_LIB if !LEMOTE_FULOONG2E

index 5f8ebcf..7094a40 100644
@@ -19,6 +19,15 @@
 extern void _mcount(void);
 #define mcount _mcount
 
+#ifdef CONFIG_DYNAMIC_FTRACE
+static inline unsigned long ftrace_call_adjust(unsigned long addr)
+{
+	return addr;
+}
+
+struct dyn_arch_ftrace {
+};
+#endif /*  CONFIG_DYNAMIC_FTRACE */
 #endif /* __ASSEMBLY__ */
 #endif /* CONFIG_FUNCTION_TRACER */
 #endif /* _ASM_MIPS_FTRACE_H */

index 38e704a..3cda3ab 100644
@@ -10,6 +10,7 @@ obj-y		+= cpu-probe.o branch.o entry.o genex.o irq.o process.o \
 
 ifdef CONFIG_FUNCTION_TRACER
 CFLAGS_REMOVE_trace_clock.o = -pg
+CFLAGS_REMOVE_ftrace.o = -pg
 CFLAGS_REMOVE_early_printk.o = -pg
 endif
 

@@ -30,7 +31,7 @@ obj-$(CONFIG_STACKTRACE)	+= stacktrace.o
 obj-$(CONFIG_MODULES)		+= mips_ksyms.o module.o
 
 obj-$(CONFIG_FTRACE)		+= trace_clock.o
-obj-$(CONFIG_FUNCTION_TRACER)	+= mcount.o
+obj-$(CONFIG_FUNCTION_TRACER)	+= mcount.o ftrace.o
 
 obj-$(CONFIG_CPU_LOONGSON2)	+= r4k_fpu.o r4k_switch.o
 obj-$(CONFIG_CPU_MIPS32)	+= r4k_fpu.o r4k_switch.o

index cbb45ed..ffc4259 100644
@@ -58,6 +58,33 @@
 	 move ra, AT
 	.endm
 
+#ifdef CONFIG_DYNAMIC_FTRACE
+
+NESTED(ftrace_caller, PT_SIZE, ra)
+	.globl _mcount
+_mcount:
+	b	ftrace_stub
+	 nop
+	lw	t0, function_trace_stop
+	bnez	t0, ftrace_stub
+	 nop
+
+	MCOUNT_SAVE_REGS
+
+	move	a0, ra		/* arg1: next ip, selfaddr */
+	.globl ftrace_call
+ftrace_call:
+	nop	/* a placeholder for the call to a real tracing function */
+	 move	a1, AT		/* arg2: the caller's next ip, parent */
+
+	MCOUNT_RESTORE_REGS
+	.globl ftrace_stub
+ftrace_stub:
+	RETURN_BACK
+	END(ftrace_caller)
+
+#else	/* ! CONFIG_DYNAMIC_FTRACE */
+
 NESTED(_mcount, PT_SIZE, ra)
 	lw	t0, function_trace_stop
 	bnez	t0, ftrace_stub

@@ -82,5 +109,7 @@ ftrace_stub:
 	RETURN_BACK
 	END(_mcount)
 
+#endif	/* ! CONFIG_DYNAMIC_FTRACE */
+
 	.set at
 	.set reorder

1.6.2.1

From: Wu Zhangjin <[email protected]>

This patch add a new section for MIPS to record the block of the hardirq handling for function graph tracer(print_graph_irq) via adding the __irq_entry annotation to the the entrypoints of the hardirqs(the block with irq_enter()...irq_exit()).

Thanks goes to Steven & Frederic Weisbecker for their feedbacks.

---
 arch/mips/include/asm/irq.h    |   29 ++---------------------------
 arch/mips/kernel/irq.c         |   30 ++++++++++++++++++++++++++++++
 arch/mips/kernel/smp.c         |    3 ++-
 arch/mips/kernel/smtc.c        |   21 ++++++++++++++-------
 arch/mips/kernel/vmlinux.lds.S |    1 +
 arch/mips/sgi-ip22/ip22-int.c  |    3 ++-
 arch/mips/sgi-ip22/ip22-time.c |    3 ++-
 7 files changed, 53 insertions(+), 37 deletions(-)

index 09b08d0..0696036 100644
@@ -113,36 +113,11 @@ do {									\
 
 #endif
 
-/*
- * do_IRQ handles all normal device IRQ's (the special
- * SMP cross-CPU interrupts have their own specific
- * handlers).
- *
- * Ideally there should be away to get this into kernel/irq/handle.c to
- * avoid the overhead of a call for just a tiny function ...
- */
-#define do_IRQ(irq)							\
-do {									\
-	irq_enter();							\
-	__DO_IRQ_SMTC_HOOK(irq);					\
-	generic_handle_irq(irq);					\
-	irq_exit();							\
-} while (0)
+extern void do_IRQ(unsigned int irq);
 
 #ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
-/*
- * To avoid inefficient and in some cases pathological re-checking of
- * IRQ affinity, we have this variant that skips the affinity check.
- */
-
 
-#define do_IRQ_no_affinity(irq)						\
-do {									\
-	irq_enter();							\
-	__NO_AFFINITY_IRQ_SMTC_HOOK(irq);				\
-	generic_handle_irq(irq);					\
-	irq_exit();							\
-} while (0)
+extern void do_IRQ_no_affinity(unsigned int irq);
 
 #endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
 

index 7b845ba..dee8d5f 100644
@@ -22,6 +22,7 @@
 #include <linux/seq_file.h>
 #include <linux/kallsyms.h>
 #include <linux/kgdb.h>
+#include <linux/ftrace.h>
 
 #include <asm/atomic.h>
 #include <asm/system.h>

@@ -150,3 +151,32 @@ void __init init_IRQ(void)
 		kgdb_early_setup = 1;
 #endif
 }
+
+/*
+ * do_IRQ handles all normal device IRQ's (the special
+ * SMP cross-CPU interrupts have their own specific
+ * handlers).
+ */
+void __irq_entry do_IRQ(unsigned int irq)
+{
+	irq_enter();
+	__DO_IRQ_SMTC_HOOK(irq);
+	generic_handle_irq(irq);
+	irq_exit();
+}
+
+#ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
+/*
+ * To avoid inefficient and in some cases pathological re-checking of
+ * IRQ affinity, we have this variant that skips the affinity check.
+ */
+
+void __irq_entry do_IRQ_no_affinity(unsigned int irq)
+{
+	irq_enter();
+	__NO_AFFINITY_IRQ_SMTC_HOOK(irq);
+	generic_handle_irq(irq);
+	irq_exit();
+}
+
+#endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */

index e72e684..6cdca19 100644
@@ -32,6 +32,7 @@
 #include <linux/cpumask.h>
 #include <linux/cpu.h>
 #include <linux/err.h>
+#include <linux/ftrace.h>
 
 #include <asm/atomic.h>
 #include <asm/cpu.h>

@@ -130,7 +131,7 @@ asmlinkage __cpuinit void start_secondary(void)
 /*
  * Call into both interrupt handlers, as we share the IPI for them
  */
-void smp_call_function_interrupt(void)
+void __irq_entry smp_call_function_interrupt(void)
 {
 	irq_enter();
 	generic_smp_call_function_single_interrupt();

index 24630fd..75034a8 100644
@@ -25,6 +25,7 @@
 #include <linux/interrupt.h>
 #include <linux/kernel_stat.h>
 #include <linux/module.h>
+#include <linux/ftrace.h>
 
 #include <asm/cpu.h>
 #include <asm/processor.h>

@@ -939,23 +940,29 @@ static void ipi_call_interrupt(void)
 
 DECLARE_PER_CPU(struct clock_event_device, mips_clockevent_device);
 
-void ipi_decode(struct smtc_ipi *pipi)
+static void __irq_entry smtc_clock_tick_interrupt(void)
 {
 	unsigned int cpu = smp_processor_id();
 	struct clock_event_device *cd;
+	int irq = MIPS_CPU_IRQ_BASE + 1;
+
+	irq_enter();
+	kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
+	cd = &per_cpu(mips_clockevent_device, cpu);
+	cd->event_handler(cd);
+	irq_exit();
+}
+
+void ipi_decode(struct smtc_ipi *pipi)
+{
 	void *arg_copy = pipi->arg;
 	int type_copy = pipi->type;
-	int irq = MIPS_CPU_IRQ_BASE + 1;
 
 	smtc_ipi_nq(&freeIPIq, pipi);
 
 	switch (type_copy) {
 	case SMTC_CLOCK_TICK:
-		irq_enter();
-		kstat_incr_irqs_this_cpu(irq, irq_to_desc(irq));
-		cd = &per_cpu(mips_clockevent_device, cpu);
-		cd->event_handler(cd);
-		irq_exit();
+		smtc_clock_tick_interrupt();
 		break;
 
 	case LINUX_SMP_IPI:

index 162b299..f25df73 100644
@@ -46,6 +46,7 @@ SECTIONS
 		SCHED_TEXT
 		LOCK_TEXT
 		KPROBES_TEXT
+		IRQENTRY_TEXT
 		*(.text.*)
 		*(.fixup)
 		*(.gnu.warning)

index 0ecd5fe..383f11d 100644
@@ -13,6 +13,7 @@
 #include <linux/init.h>
 #include <linux/kernel_stat.h>
 #include <linux/interrupt.h>
+#include <linux/ftrace.h>
 
 #include <asm/irq_cpu.h>
 #include <asm/sgi/hpc3.h>

@@ -150,7 +151,7 @@ static void indy_local1_irqdispatch(void)
 
 extern void ip22_be_interrupt(int irq);
 
-static void indy_buserror_irq(void)
+static void __irq_entry indy_buserror_irq(void)
 {
 	int irq = SGI_BUSERR_IRQ;
 

index c8f7d23..603fc91 100644
@@ -16,6 +16,7 @@
 #include <linux/interrupt.h>
 #include <linux/kernel_stat.h>
 #include <linux/time.h>
+#include <linux/ftrace.h>
 
 #include <asm/cpu.h>
 #include <asm/mipsregs.h>

@@ -115,7 +116,7 @@ __init void plat_time_init(void)
 }
 
 /* Generic SGI handler for (spurious) 8254 interrupts */
-void indy_8254timer_irq(void)
+void __irq_entry indy_8254timer_irq(void)
 {
 	int irq = SGI_8254_0_IRQ;
 	ULONG cnt;

1.6.2.1

From: Wu Zhangjin <[email protected]>

Ftrace gets the timestamp via time relative functions, when enabling function graph tracer, if we also trace these functions, the system will go into recusion and then hang there. So, we must not trace them. and there are two existing methods help this job.

The first one is annotating them with notrace, another method is removing the -pg for the whole file which included them:

CFLAGS_REMOVE_ftrace.o = -pg

But there is a particular situation, that is some archs(i.e MIPS, Microblaze) need to notrace the common time relative source code, we can not simply annotate them with the above methods, but need to consider the archs respectively. this patch does it!

This patch introduce a new annotation flag: __time_notrace, if your platform need it, define it in your arch specific ftrace.h:

	#define __time_notrace notrace

if only function graph trace need it, wrap it:

	#ifdef CONFIG_FUNCTION_GRAPH_TRACER
	 #define __time_notrace notrace
	#endif

otherwise, ignore it!

---
 include/linux/ftrace.h |   12 ++++++++++++
 1 files changed, 12 insertions(+), 0 deletions(-)

index 0b4f97d..c3f2f0f 100644
@@ -511,4 +511,16 @@ static inline void trace_hw_branch_oops(void) {}
 
 #endif /* CONFIG_HW_BRANCH_TRACER */
 
+/* arch specific notrace for time relative functions, If your arch need it,
+ * define it in the arch specific ftrace.h
+ *
+ *	#define __time_notrace notrace
+ *
+ * otherwise, ignore it!
+ */
+
+#ifndef __time_notrace
+ #define __time_notrace
+#endif
+
 #endif /* _LINUX_FTRACE_H */

1.6.2.1

From: Wu Zhangjin <[email protected]>

Some platforms(i.e. MIPS) need these two functions to get the precise timestamp, we use __time_notrace(added in the last patch) to annotate it. By default, __time_notrace is empty, so, this patch have no influence to the original functions, but if you really not need to trace them, just add the following line into the arch specific ftrace.h:

	#define __time_notrace notrace

If you only need it for function graph tracer, wrap it:

	#ifdef CONFIG_FUNCTION_GRAPH_TRACER
	 #define __time_notrace notrace
	#endif

please get more information from include/linux/ftrace.h

---
 kernel/time/clocksource.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)

index 5e18c6a..b00476a 100644
@@ -30,6 +30,7 @@
 #include <linux/sched.h> /* for spin_unlock_irq() using preempt_count() m68k */
 #include <linux/tick.h>
 #include <linux/kthread.h>
+#include <linux/ftrace.h>
 
 void timecounter_init(struct timecounter *tc,
 		      const struct cyclecounter *cc,

@@ -52,7 +53,7 @@ EXPORT_SYMBOL(timecounter_init);
  * The first call to this function for a new time counter initializes
  * the time tracking and returns an undefined result.
  */
-static u64 timecounter_read_delta(struct timecounter *tc)
+static u64 __time_notrace timecounter_read_delta(struct timecounter *tc)
 {
 	cycle_t cycle_now, cycle_delta;
 	u64 ns_offset;

@@ -72,7 +73,7 @@ static u64 timecounter_read_delta(struct timecounter *tc)
 	return ns_offset;
 }
 
-u64 timecounter_read(struct timecounter *tc)
+u64 __time_notrace timecounter_read(struct timecounter *tc)
 {
 	u64 nsec;
 

1.6.2.1

From: Wu Zhangjin <[email protected]>

We use mips_timecounter_read() to get the timestamp in MIPS, so, it's better to not trace it and it's subroutines, otherwise, it will goto recursion(hang) when using function graph tracer. we use the __notrace_funcgraph annotation to indicate them not to be traced.

And there are two common functions called by mips_timecounter_read(), we define the __time_notrace macro to ensure they are not to be traced either.

---
 arch/mips/include/asm/ftrace.h |    6 ++++++
 arch/mips/kernel/csrc-r4k.c    |    5 +++--
 2 files changed, 9 insertions(+), 2 deletions(-)

index 7094a40..aa7c80b 100644
@@ -28,6 +28,12 @@ static inline unsigned long ftrace_call_adjust(unsigned long addr)
 struct dyn_arch_ftrace {
 };
 #endif /*  CONFIG_DYNAMIC_FTRACE */
+
+/* not trace the timecounter_read* in kernel/time/clocksource.c */
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+ #define __time_notrace notrace
+#endif
+
 #endif /* __ASSEMBLY__ */
 #endif /* CONFIG_FUNCTION_TRACER */
 #endif /* _ASM_MIPS_FTRACE_H */

index 4e7705f..0c1bf80 100644
@@ -8,6 +8,7 @@
 #include <linux/clocksource.h>
 #include <linux/init.h>
 #include <linux/sched.h>
+#include <linux/ftrace.h>
 
 #include <asm/time.h>
 

@@ -42,7 +43,7 @@ static struct timecounter r4k_tc = {
 	.cc = NULL,
 };
 
-static cycle_t r4k_cc_read(const struct cyclecounter *cc)
+static cycle_t __notrace_funcgraph r4k_cc_read(const struct cyclecounter *cc)
 {
 	return read_c0_count();
 }

@@ -66,7 +67,7 @@ int __init init_r4k_timecounter(void)
 	return 0;
 }
 
-u64 r4k_timecounter_read(void)
+u64 __notrace_funcgraph r4k_timecounter_read(void)
 {
 	u64 clock;
 

1.6.2.1

From: Wu Zhangjin <[email protected]>

The implementation of function graph tracer for MIPS is a little different from X86.

in MIPS, gcc(with -pg) only transfer the caller's return address(at) and the _mcount's return address(ra) to us.

For the kernel part without -mlong-calls:

move at, ra
jal _mcount

For the module part with -mlong-calls:

lui v1, hi16bit_of_mcount
addiu v1, v1, low16bit_of_mcount
move at, ra
jal _mcount

Without -mlong-calls,

if the function is a leaf, it will not save the return address(ra):

ffffffff80101298 <au1k_wait>:

ffffffff80101298:       67bdfff0        daddiu  sp,sp,-16
ffffffff8010129c:       ffbe0008        sd      s8,8(sp)
ffffffff801012a0:       03a0f02d        move    s8,sp
ffffffff801012a4:       03e0082d        move    at,ra
ffffffff801012a8:       0c042930        jal     ffffffff8010a4c0 <_mcount>
ffffffff801012ac:       00020021        nop

so, we can hijack it directly in _mcount, but if the function is non-leaf, the return address is saved in the stack.

ffffffff80133030 <copy_process>:

ffffffff80133030:       67bdff50        daddiu  sp,sp,-176
ffffffff80133034:       ffbe00a0        sd      s8,160(sp)
ffffffff80133038:       03a0f02d        move    s8,sp
ffffffff8013303c:       ffbf00a8        sd      ra,168(sp)
ffffffff80133040:       ffb70098        sd      s7,152(sp)
ffffffff80133044:       ffb60090        sd      s6,144(sp)
ffffffff80133048:       ffb50088        sd      s5,136(sp)
ffffffff8013304c:       ffb40080        sd      s4,128(sp)
ffffffff80133050:       ffb30078        sd      s3,120(sp)
ffffffff80133054:       ffb20070        sd      s2,112(sp)
ffffffff80133058:       ffb10068        sd      s1,104(sp)
ffffffff8013305c:       ffb00060        sd      s0,96(sp)
ffffffff80133060:       03e0082d        move    at,ra
ffffffff80133064:       0c042930        jal     ffffffff8010a4c0 <_mcount>
ffffffff80133068:       00020021        nop

but we can not get the exact stack address(which saved ra) directly in _mcount, we need to search the content of at register in the stack space or search the "s{d,w} ra, offset(sp)" instruction in the text. 'Cause we can not prove there is only a match in the stack space, so, we search the text instead.

as we can see, if the first instruction above "move at, ra" is not a store instruction, there should be a leaf function, so we hijack the at register directly via putting &return_to_handler into it, otherwise, we search the "s{d,w} ra, offset(sp)" instruction to get the stack offset, and then the stack address. we use the above copy_process() as an example, we at last find "ffbf00a8", 0xa8 is the stack offset, we plus it with s8(fp), that is the stack address, we hijack the content via writing the &return_to_handler in.

If with -mlong-calls, since there are two more instructions above "move at, ra", so, we can move the pointer to the position above "lui v1, hi16bit_of_mcount".

---
 arch/mips/Kconfig         |    1 +
 arch/mips/kernel/ftrace.c |  106 +++++++++++++++++++++++++++++++++++++++++++++
 arch/mips/kernel/mcount.S |   42 ++++++++++++++++++
 3 files changed, 149 insertions(+), 0 deletions(-)

index 147fbbc..de690fd 100644
@@ -8,6 +8,7 @@ config MIPS
 	select HAVE_FUNCTION_TRACE_MCOUNT_TEST
 	select HAVE_DYNAMIC_FTRACE
 	select HAVE_FTRACE_MCOUNT_RECORD
+	select HAVE_FUNCTION_GRAPH_TRACER
 	# Horrible source of confusion.  Die, die, die ...
 	select EMBEDDED
 	select RTC_LIB if !LEMOTE_FULOONG2E

1.6.2.1

From: Wu Zhangjin <[email protected]>

This patch make function graph tracer work with dynamic function tracer.

To share the source code of dynamic function tracer(MCOUNT_SAVE_REGS), and avoid restoring the whole saved registers, we need to restore the ra register from the stack.

(NOTE: This not work with 32bit! need to ensure why!)

---
 arch/mips/kernel/ftrace.c |   21 +++++++++++++++++++++
 arch/mips/kernel/mcount.S |   14 ++++++++++++--
 2 files changed, 33 insertions(+), 2 deletions(-)

index 65a3f8a..e981a49 100644
@@ -115,6 +115,27 @@ int __init ftrace_dyn_arch_init(void *data)
 
 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 
+#ifdef CONFIG_DYNAMIC_FTRACE
+
+extern void ftrace_graph_call(void);
+#define JMP	0x08000000	/* jump to target directly */
+#define CALL_FTRACE_GRAPH_CALLER \
+	jump_insn_encode(JMP, (unsigned long)(&ftrace_graph_caller))
+#define FTRACE_GRAPH_CALL_IP	((unsigned long)(&ftrace_graph_call))
+
+int ftrace_enable_ftrace_graph_caller(void)
+{
+	return ftrace_modify_code(FTRACE_GRAPH_CALL_IP,
+				  CALL_FTRACE_GRAPH_CALLER);
+}
+
+int ftrace_disable_ftrace_graph_caller(void)
+{
+	return ftrace_modify_code(FTRACE_GRAPH_CALL_IP, ftrace_nop);
+}
+
+#endif				/* !CONFIG_DYNAMIC_FTRACE */
+
 #define S_RA_SP	(0xafbf << 16)	/* s{d,w} ra, offset(sp) */
 #define S_R_SP	(0xafb0 << 16)  /* s{d,w} R, offset(sp) */
 #define OFFSET_MASK	0xffff	/* stack offset range: 0 ~ PT_SIZE */

index b50e38d..98d4690 100644
@@ -77,6 +77,13 @@ ftrace_call:
 	nop	/* a placeholder for the call to a real tracing function */
 	 move	a1, AT		/* arg2: the caller's next ip, parent */
 
+#ifdef CONFIG_FUNCTION_GRAPH_TRACER
+	.globl ftrace_graph_call
+ftrace_graph_call:
+	nop
+	 nop
+#endif
+
 	MCOUNT_RESTORE_REGS
 	.globl ftrace_stub
 ftrace_stub:

@@ -124,10 +131,13 @@ ftrace_stub:
 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 
 NESTED(ftrace_graph_caller, PT_SIZE, ra)
+#ifdef CONFIG_DYNAMIC_FTRACE
+	PTR_L	a1, PT_R31(sp)	/* load the original ra from the stack */
+#else
 	MCOUNT_SAVE_REGS
-
-	PTR_LA	a0, PT_R1(sp)	/* arg1: &AT -> a0 */
 	move	a1, ra		/* arg2: next ip, selfaddr */
+#endif
+	PTR_LA	a0, PT_R1(sp)	/* arg1: &AT -> a0 */
 	jal	prepare_ftrace_return
 	 move	a2, fp		/* arg3: frame pointer */
 

1.6.2.1

From: Wu Zhangjin <[email protected]>

When remove the -fno-omit-frame-pointer, gcc will not save the frame pointer for us, we need to save one ourselves.

---
 arch/mips/include/asm/ftrace.h |   57 ++++++++++++++++++++++++++++++++++++++++
 arch/mips/kernel/ftrace.c      |   56 ++++++++++++++++++++++++++++----------
 arch/mips/kernel/mcount.S      |    8 +++++
 3 files changed, 106 insertions(+), 15 deletions(-)

index aa7c80b..f78d763 100644
@@ -19,6 +19,62 @@
 extern void _mcount(void);
 #define mcount _mcount
 
+#define safe_load(load, src, dst, error)		\
+do {							\
+	asm volatile (					\
+		"1: " load " %[" STR(dst) "], 0(%[" STR(src) "])\n"\
+		"   li %[" STR(error) "], 0\n"		\
+		"2:\n"					\
+							\
+		".section .fixup, \"ax\"\n"		\
+		"3: li %[" STR(error) "], 1\n"		\
+		"   j 2b\n"				\
+		".previous\n"				\
+							\
+		".section\t__ex_table,\"a\"\n\t"	\
+		STR(PTR) "\t1b, 3b\n\t"			\
+		".previous\n"				\
+							\
+		: [dst] "=&r" (dst), [error] "=r" (error)\
+		: [src] "r" (src)			\
+		: "memory"				\
+	);						\
+} while (0)
+
+#define safe_store(store, src, dst, error)	\
+do {						\
+	asm volatile (				\
+		"1: " store " %[" STR(src) "], 0(%[" STR(dst) "])\n"\
+		"   li %[" STR(error) "], 0\n"	\
+		"2:\n"				\
+						\
+		".section .fixup, \"ax\"\n"	\
+		"3: li %[" STR(error) "], 1\n"	\
+		"   j 2b\n"			\
+		".previous\n"			\
+						\
+		".section\t__ex_table,\"a\"\n\t"\
+		STR(PTR) "\t1b, 3b\n\t"		\
+		".previous\n"			\
+						\
+		: [error] "=r" (error)		\
+		: [dst] "r" (dst), [src] "r" (src)\
+		: "memory"			\
+	);					\
+} while (0)
+
+#define safe_load_code(dst, src, error) \
+	safe_load(STR(lw), src, dst, error)
+#define safe_store_code(src, dst, error) \
+	safe_store(STR(sw), src, dst, error)
+
+#define safe_load_stack(dst, src, error) \
+	safe_load(STR(PTR_L), src, dst, error)
+
+#define safe_store_stack(src, dst, error) \
+	safe_store(STR(PTR_S), src, dst, error)
+
+
 #ifdef CONFIG_DYNAMIC_FTRACE
 static inline unsigned long ftrace_call_adjust(unsigned long addr)
 {

@@ -27,6 +83,7 @@ static inline unsigned long ftrace_call_adjust(unsigned long addr)
 
 struct dyn_arch_ftrace {
 };
+
 #endif /*  CONFIG_DYNAMIC_FTRACE */
 
 /* not trace the timecounter_read* in kernel/time/clocksource.c */

index 98d4690..bdfef2c 100644
@@ -139,7 +139,15 @@ NESTED(ftrace_graph_caller, PT_SIZE, ra)
 #endif
 	PTR_LA	a0, PT_R1(sp)	/* arg1: &AT -> a0 */
 	jal	prepare_ftrace_return
+#ifdef CONFIG_FRAME_POINTER
 	 move	a2, fp		/* arg3: frame pointer */
+#else
+#ifdef CONFIG_64BIT
+	 PTR_LA	a2, PT_SIZE(sp)
+#else
+	 PTR_LA	a2, (PT_SIZE+8)(sp)
+#endif
+#endif
 
 	MCOUNT_RESTORE_REGS
 	RETURN_BACK

1.6.2.1

From: Wu Zhangjin <[email protected]>

A new option -mmcount-ra-address for gcc 4.5 have been sent by David Daney <[email protected]> in the thread "MIPS: Add option to pass return address location to _mcount", which help to record the location of the return address(ra) for the function graph tracer of MIPS to hijack the return address easier and safer. that option used the $12(t0) register by default, so, we reserve it for it, and use t1,t2,t3 instead of t0,t1,t2.

---
 arch/mips/kernel/mcount.S |   26 +++++++++++++-------------
 1 files changed, 13 insertions(+), 13 deletions(-)

index bdfef2c..522e91c 100644
@@ -65,8 +65,8 @@ NESTED(ftrace_caller, PT_SIZE, ra)
 _mcount:
 	b	ftrace_stub
 	 nop
-	lw	t0, function_trace_stop
-	bnez	t0, ftrace_stub
+	lw	t1, function_trace_stop
+	bnez	t1, ftrace_stub
 	 nop
 
 	MCOUNT_SAVE_REGS

@@ -93,21 +93,21 @@ ftrace_stub:
 #else	/* ! CONFIG_DYNAMIC_FTRACE */
 
 NESTED(_mcount, PT_SIZE, ra)
-	lw	t0, function_trace_stop
-	bnez	t0, ftrace_stub
+	lw	t1, function_trace_stop
+	bnez	t1, ftrace_stub
 	 nop
-	PTR_LA	t0, ftrace_stub
-	PTR_L	t1, ftrace_trace_function /* Prepare t1 for (1) */
-	bne	t0, t1, static_trace
+	PTR_LA	t1, ftrace_stub
+	PTR_L	t2, ftrace_trace_function /* Prepare t2 for (1) */
+	bne	t1, t2, static_trace
 	 nop
 
 #ifdef	CONFIG_FUNCTION_GRAPH_TRACER
-	PTR_L	t2, ftrace_graph_return
-	bne	t0, t2, ftrace_graph_caller
+	PTR_L	t3, ftrace_graph_return
+	bne	t1, t3, ftrace_graph_caller
 	 nop
-	PTR_LA	t0, ftrace_graph_entry_stub
-	PTR_L	t2, ftrace_graph_entry
-	bne	t0, t2, ftrace_graph_caller
+	PTR_LA	t1, ftrace_graph_entry_stub
+	PTR_L	t3, ftrace_graph_entry
+	bne	t1, t3, ftrace_graph_caller
 	 nop
 #endif
 	b	ftrace_stub

@@ -117,7 +117,7 @@ static_trace:
 	MCOUNT_SAVE_REGS
 
 	move	a0, ra		/* arg1: next ip, selfaddr */
-	jalr	t1	/* (1) call *ftrace_trace_function */
+	jalr	t2		/* (1) call *ftrace_trace_function */
 	 move	a1, AT		/* arg2: the caller's next ip, parent */
 
 	MCOUNT_RESTORE_REGS

1.6.2.1

From: Wu Zhangjin <[email protected]>

That thread "MIPS: Add option to pass return address location to _mcount" from "David Daney <[email protected]>" have added a new option -mmcount-ra-address to gcc(4.5) for MIPS to transfer the location of the return address to _mcount.

Benefit from this new feature, function graph tracer on MIPS will be easier and safer to hijack the return address of the kernel function, which will save some overhead and make the whole thing more reliable.

In this patch, at first, try to enable the option -mmcount-ra-address in arch/mips/Makefile with cc-option, if gcc support it, it will be enabled, otherwise, no side effect.

and then, we need to support this new option of gcc 4.5 and also support the old gcc versions.

with _mcount in the old gcc versions, it's not easy to get the location of return address(tracing: add function graph tracer support for MIPS),
so, we do it in a C function: ftrace_get_parent_addr(ftrace.c), but with -mmcount-ra-address, only several instructions need to get what we want, so, I put into asm(mcount.S). and also, as the $12(t0) is used by -mmcount-ra-address for transferring the localtion of return address to _mcount, we need to save it into the stack and restore it when enabled dynamic function tracer, 'Cause we have called "ftrace_call" before "ftrace_graph_caller", which may destroy $12(t0).

(Thanks to David for providing that -mcount-ra-address and giving the
idea of KBUILD_MCOUNT_RA_ADDRESS, both of them have made the whole thing more beautiful!)

---
 arch/mips/Makefile        |    7 +++++++
 arch/mips/kernel/ftrace.c |   24 +++++++++++++++++-------
 arch/mips/kernel/mcount.S |   14 ++++++++++++++
 3 files changed, 38 insertions(+), 7 deletions(-)

index 0a2073e..50039e8 100644
@@ -51,6 +51,13 @@ endif
 ifndef CONFIG_FUNCTION_TRACER
 cflags-y := -ffunction-sections
 endif
+ifdef CONFIG_FUNCTION_GRAPH_TRACER
+  ifndef KBUILD_MCOUNT_RA_ADDRESS
+    ifeq ($(call cc-option-yn,-mmcount-ra-address), y)
+      cflags-y += -mmcount-ra-address -DKBUILD_MCOUNT_RA_ADDRESS
+    endif
+  endif
+endif
 cflags-y += $(call cc-option, -mno-check-zero-division)
 
 ifdef CONFIG_32BIT

index e363fc6..68b0670 100644
@@ -148,6 +148,7 @@ int ftrace_disable_ftrace_graph_caller(void)
 
 #endif				/* !CONFIG_DYNAMIC_FTRACE */
 
+#ifndef KBUILD_MCOUNT_RA_ADDRESS
 #define S_RA_SP	(0xafbf << 16)	/* s{d,w} ra, offset(sp) */
 #define S_R_SP	(0xafb0 << 16)  /* s{d,w} R, offset(sp) */
 #define OFFSET_MASK	0xffff	/* stack offset range: 0 ~ PT_SIZE */

@@ -201,6 +202,8 @@ unsigned long ftrace_get_parent_addr(unsigned long self_addr,
 	return 0;
 }
 
+#endif
+
 /*
  * Hook the return address and push it in the stack of return addrs
  * in current thread info.

@@ -218,19 +221,26 @@ void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
 		return;
 
 	/* "parent" is the stack address saved the return address of the caller
-	 * of _mcount, for a leaf function not save the return address in the
-	 * stack address, so, we "emulate" one in _mcount's stack space, and
-	 * hijack it directly, but for a non-leaf function, it will save the
-	 * return address to the its stack space, so, we can not hijack the
-	 * "parent" directly, but need to find the real stack address,
+	 * of _mcount.
+	 *
+	 * if the gcc < 4.5, a leaf function does not save the return address
+	 * in the stack address, so, we "emulate" one in _mcount's stack space,
+	 * and hijack it directly, but for a non-leaf function, it save the
+	 * return address to the its own stack space, we can not hijack it
+	 * directly, but need to find the real stack address,
 	 * ftrace_get_parent_addr() does it!
+	 *
+	 * if gcc>= 4.5, with the new -mmcount-ra-address option, for a
+	 * non-leaf function, the location of the return address will be saved
+	 * to $12 for us, and for a leaf function, only put a zero into $12. we
+	 * do it in ftrace_graph_caller of mcount.S.
 	 */
 
 	/* old = *parent; */
 	safe_load_stack(old, parent, faulted);
 	if (unlikely(faulted))
 		goto out;
-
+#ifndef KBUILD_MCOUNT_RA_ADDRESS
 	parent = (unsigned long *)ftrace_get_parent_addr(self_addr, old,
 							 (unsigned long)parent,
 							 fp);

@@ -238,7 +248,7 @@ void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
 	 * ra, stop function graph tracer and return */
 	if (parent == 0)
 		goto out;
-
+#endif
 	/* *parent = return_hooker; */
 	safe_store_stack(return_hooker, parent, faulted);
 	if (unlikely(faulted))

index 522e91c..0a9cfdb 100644
@@ -70,6 +70,9 @@ _mcount:
 	 nop
 
 	MCOUNT_SAVE_REGS
+#ifdef KBUILD_MCOUNT_RA_ADDRESS
+	PTR_S	t0, PT_R12(sp)	/* t0 saved the location of the return address(at) by -mmcount-ra-address */
+#endif
 
 	move	a0, ra		/* arg1: next ip, selfaddr */
 	.globl ftrace_call

@@ -133,11 +136,22 @@ ftrace_stub:
 NESTED(ftrace_graph_caller, PT_SIZE, ra)
 #ifdef CONFIG_DYNAMIC_FTRACE
 	PTR_L	a1, PT_R31(sp)	/* load the original ra from the stack */
+#ifdef KBUILD_MCOUNT_RA_ADDRESS
+	PTR_L	t0, PT_R12(sp)	/* load the original t0 from the stack */
+#endif
 #else
 	MCOUNT_SAVE_REGS
 	move	a1, ra		/* arg2: next ip, selfaddr */
 #endif
+
+#ifdef KBUILD_MCOUNT_RA_ADDRESS
+	bnez	t0, 1f		/* non-leaf func: t0 saved the location of the return address */
+	 nop
+	PTR_LA	t0, PT_R1(sp)	/* leaf func: get the location of at(old ra) from our own stack */
+1:	move	a0, t0		/* arg1: the location of the return address */
+#else
 	PTR_LA	a0, PT_R1(sp)	/* arg1: &AT -> a0 */
+#endif
 	jal	prepare_ftrace_return
 #ifdef CONFIG_FRAME_POINTER
 	 move	a2, fp		/* arg3: frame pointer */

1.6.2.1

Hmm, the patch adds the endian argument to the command line, but does nothing else with it. Is there something missing from the patch or is it just a left over from earlier iterations ?

Thanks,

tglx

It is used in:

[PATCH v8 07/16] tracing: add dynamic function tracer support for MIPS

Steven told me to split it out ;)

Thanks & Regards,
Wu Zhangjin

Heh, IIRC I think I meant to split out the recordmcount.pl code first. But this is fine too ;-)

-- Steve

Fair enough. Missed that.

Thanks,

tglx

-- Steve

Hmm, I'm not convinced that this is really a huge overhead.

First of all the rollover happens once every 10 seconds on a 800MHz machine.

Secondly we have a lockless implementation of extending 32bit counters to 63 bit which is used at least by ARM to provide a high resolution sched_clock implementation. See include/linux/cnt32_63.h and the users in arch/

But that's a problem which can be discussed seperately and does not affect the rest of the tracing infrastructure. I really would recommend that you implement a sched_clock for the r4k machines based on cnt32_63 and measure the overhead. Having a fine granular sched_clock in general is probably not a bad thing.

Thanks,

tglx

Thanks very much for your prompt,

Will have a look at cnt32_to_63() in include/linux/cnt32_to_63.h later.

Thanks & Regards,
Wu Zhangjin

please cc me on this discuss too. I have working ftrace implementation in my tree and I need improve timing too. I have similar patch as MIPS use. But I am not able to use it via timecounters. Something is weird there that's why I am open to find out any sensible and accepted solution.

Thanks,
Michal

Michal Simek, Ing. (M.Eng)
w: www.monstr.eu p: +42-0-721842854
Maintainer of Linux kernel 2.6 Microblaze Linux - http://www.monstr.eu/fdt/
Microblaze U-BOOT custodian

On Sat, Nov 14, 2009 at 02:30:31PM +0800, Wu Zhangjin wrote:

Time to talk about how to get the whole thing merged.

I'm reasonably happy with the arch part of the patchset. There are two remaining issues in the arch part - cache flushing will not work as is on a SMP system and the whole code patching is going to fail if the kernel text is replicated with CONFIG_REPLICATE_KTEXT which is an IP27-only option.

I suggest we should go ahready and just disallow dynamic ftrace on on these problematic configurations for now and deal with them later, so something like

	select HAVE_DYNAMIC_FTRACE if !SMP && !REPLICATE_KTEXT

This leaves the non-arch bits to merge; patches which touch code outside of arch/mips are:

[PATCH v8 01/16] tracing: convert trace_clock_local() as weak function [PATCH v8 06/16] tracing: add an endian argument to
[PATCH v8 07/16] tracing: add dynamic function tracer support for MIPS [PATCH v8 09/16] tracing: define a new __time_notrace annotation flag [PATCH v8 10/16] tracing: not trace the timecounter_read* in

Ralf

Makes sense.

I'm going to comment on the patches.

Thanks,

tglx