[PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[OGAWA Hirofumi]

[forgot to Cc: lkml, resend]

Hi,

Maybe, nobody using debug patch in atomic_dec_and_test()...  Well,
anyway, how about this?



stop_machine_first is just to see if it is first one or not.  So, there
is no reason to use atomic_dec_and_test(), and makes the value below 0.

I think it is not desirable, because this usage only triggers
atomic_dec_and_test() underflow debug patch. (the patch tests result
of atomic_dec_and_test() is < 0)

So, this uses test_and_clear_bit() instead.

Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
---

 arch/x86/kernel/alternative.c |    8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff -puN arch/x86/kernel/alternative.c~stop_machine-use-test_and_set_bit arch/x86/kernel/alternative.c
--- linux/arch/x86/kernel/alternative.c~stop_machine-use-test_and_set_bit	2012-05-23 02:48:01.000000000 +0900
+++ linux-hirofumi/arch/x86/kernel/alternative.c	2012-05-23 02:48:01.000000000 +0900
@@ -650,7 +650,7 @@ void *__kprobes text_poke(void *addr, co
  * Cross-modifying kernel text with stop_machine().
  * This code originally comes from immediate value.
  */
-static atomic_t stop_machine_first;
+static unsigned long stop_machine_first;
 static int wrote_text;
 
 struct text_poke_params {
@@ -664,7 +664,7 @@ static int __kprobes stop_machine_text_p
 	struct text_poke_param *p;
 	int i;
 
-	if (atomic_dec_and_test(&stop_machine_first)) {
+	if (test_and_clear_bit(0, &stop_machine_first)) {
 		for (i = 0; i < tpp->nparams; i++) {
 			p = &tpp->params[i];
 			text_poke(p->addr, p->opcode, p->len);
@@ -714,7 +714,7 @@ void *__kprobes text_poke_smp(void *addr
 	p.len = len;
 	tpp.params = &p;
 	tpp.nparams = 1;
-	atomic_set(&stop_machine_first, 1);
+	stop_machine_first = 1;
 	wrote_text = 0;
 	/* Use __stop_machine() because the caller already got online_cpus. */
 	__stop_machine(stop_machine_text_poke, (void *)&tpp, cpu_online_mask);
@@ -736,7 +736,7 @@ void __kprobes text_poke_smp_batch(struc
 {
 	struct text_poke_params tpp = {.params = params, .nparams = n};
 
-	atomic_set(&stop_machine_first, 1);
+	stop_machine_first = 1;
 	wrote_text = 0;
 	__stop_machine(stop_machine_text_poke, (void *)&tpp, cpu_online_mask);
 }
_

-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: [PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[Steven Rostedt]

On Wed, May 23, 2012 at 03:11:48AM +0900, OGAWA Hirofumi wrote:
> [forgot to Cc: lkml, resend]
> 
> Hi,
> 
> Maybe, nobody using debug patch in atomic_dec_and_test()...  Well,
> anyway, how about this?

What debug patch?

> 
> 
> 
> stop_machine_first is just to see if it is first one or not.  So, there
> is no reason to use atomic_dec_and_test(), and makes the value below 0.
> 
> I think it is not desirable, because this usage only triggers
> atomic_dec_and_test() underflow debug patch. (the patch tests result
> of atomic_dec_and_test() is < 0)

Well it should only underflow if you have a box with more than 2 billion
CPUs.

> 
> So, this uses test_and_clear_bit() instead.
> 
> Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
> ---
> 
>  arch/x86/kernel/alternative.c |    8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff -puN arch/x86/kernel/alternative.c~stop_machine-use-test_and_set_bit arch/x86/kernel/alternative.c
> --- linux/arch/x86/kernel/alternative.c~stop_machine-use-test_and_set_bit	2012-05-23 02:48:01.000000000 +0900
> +++ linux-hirofumi/arch/x86/kernel/alternative.c	2012-05-23 02:48:01.000000000 +0900
> @@ -650,7 +650,7 @@ void *__kprobes text_poke(void *addr, co
>   * Cross-modifying kernel text with stop_machine().
>   * This code originally comes from immediate value.
>   */
> -static atomic_t stop_machine_first;
> +static unsigned long stop_machine_first;

The down side to this is that it adds 4 more bytes on a 64bit
machine. (sizeof(unsigned log) == 8 and sizeof(atomic_t) == 4)

You could probably also set it to -1, and do a atomic_inc_and_test(),
would that also cause the debug to trigger too?

-- Steve

>  static int wrote_text;
>  
>  struct text_poke_params {
> @@ -664,7 +664,7 @@ static int __kprobes stop_machine_text_p
>  	struct text_poke_param *p;
>  	int i;
>  
> -	if (atomic_dec_and_test(&stop_machine_first)) {
> +	if (test_and_clear_bit(0, &stop_machine_first)) {
>  		for (i = 0; i < tpp->nparams; i++) {
>  			p = &tpp->params[i];
>  			text_poke(p->addr, p->opcode, p->len);
> @@ -714,7 +714,7 @@ void *__kprobes text_poke_smp(void *addr
>  	p.len = len;
>  	tpp.params = &p;
>  	tpp.nparams = 1;
> -	atomic_set(&stop_machine_first, 1);
> +	stop_machine_first = 1;
>  	wrote_text = 0;
>  	/* Use __stop_machine() because the caller already got online_cpus. */
>  	__stop_machine(stop_machine_text_poke, (void *)&tpp, cpu_online_mask);
> @@ -736,7 +736,7 @@ void __kprobes text_poke_smp_batch(struc
>  {
>  	struct text_poke_params tpp = {.params = params, .nparams = n};
>  
> -	atomic_set(&stop_machine_first, 1);
> +	stop_machine_first = 1;
>  	wrote_text = 0;
>  	__stop_machine(stop_machine_text_poke, (void *)&tpp, cpu_online_mask);
>  }

Re: [PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[OGAWA Hirofumi]

Steven Rostedt <rostedt@goodmis.org> writes:

> On Wed, May 23, 2012 at 03:11:48AM +0900, OGAWA Hirofumi wrote:
>> [forgot to Cc: lkml, resend]
>> 
>> Hi,
>> 
>> Maybe, nobody using debug patch in atomic_dec_and_test()...  Well,
>> anyway, how about this?
>
> What debug patch?

It is this patch. I got this from -mm (akpm series), I don't know
whether -mm is still using though.

The patch below will detect atomic counter underflows.  This has been
test-driven in the -RT patchset for some time.  qdisc_destroy() triggered
it sometimes (in a seemingly nonfatal way, during device shutdown) - with
DaveM suggesting that it is most likely a bug in the networking code.  So
it would be nice to have this in -mm for some time to validate all atomic
counters on a broader base.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
---

Change it to atomic check.

Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
---

 arch/x86/include/asm/atomic.h      |   13 +++++++++++++
 arch/x86/include/asm/atomic64_32.h |    5 ++++-
 arch/x86/include/asm/atomic64_64.h |   12 ++++++++++++
 3 files changed, 29 insertions(+), 1 deletion(-)

diff -puN arch/x86/include/asm/atomic.h~debug_atomic_t-underflows-atomic arch/x86/include/asm/atomic.h
--- linux/arch/x86/include/asm/atomic.h~debug_atomic_t-underflows-atomic	2012-04-03 17:29:38.000000000 +0900
+++ linux-hirofumi/arch/x86/include/asm/atomic.h	2012-04-03 17:29:38.000000000 +0900
@@ -6,6 +6,18 @@
 #include <asm/processor.h>
 #include <asm/alternative.h>
 #include <asm/cmpxchg.h>
+#include <asm/bug.h>
+
+#define ATOMIC_UNDERFLOW_CHECK(v)	do {				\
+	unsigned char __sf;						\
+	/* if (atomic_read(v) < 0) */					\
+	__asm__ __volatile__("sets %0"					\
+			     : "=qm" (__sf)				\
+			     : /* no input */				\
+			     : "memory");				\
+	WARN(__sf, KERN_ERR "atomic counter underflow: %d\n",		\
+	     atomic_read(v));						\
+} while(0)
 
 /*
  * Atomic operations that C can't guarantee us.  Useful for
@@ -123,6 +135,7 @@ static inline int atomic_dec_and_test(at
 	asm volatile(LOCK_PREFIX "decl %0; sete %1"
 		     : "+m" (v->counter), "=qm" (c)
 		     : : "memory");
+	ATOMIC_UNDERFLOW_CHECK(v);
 	return c != 0;
 }
 
diff -puN arch/x86/include/asm/atomic64_32.h~debug_atomic_t-underflows-atomic arch/x86/include/asm/atomic64_32.h
--- linux/arch/x86/include/asm/atomic64_32.h~debug_atomic_t-underflows-atomic	2012-04-03 17:29:38.000000000 +0900
+++ linux-hirofumi/arch/x86/include/asm/atomic64_32.h	2012-04-03 17:29:38.000000000 +0900
@@ -244,7 +244,10 @@ static inline void atomic64_dec(atomic64
  */
 static inline int atomic64_dec_and_test(atomic64_t *v)
 {
-	return atomic64_dec_return(v) == 0;
+	long long ret = atomic64_dec_return(v);
+	WARN(ret < 0, KERN_ERR "atomic counter underflow: %lld\n",
+	     atomic64_read(v));
+	return ret == 0;
 }
 
 /**
diff -puN arch/x86/include/asm/atomic64_64.h~debug_atomic_t-underflows-atomic arch/x86/include/asm/atomic64_64.h
--- linux/arch/x86/include/asm/atomic64_64.h~debug_atomic_t-underflows-atomic	2012-04-03 17:29:38.000000000 +0900
+++ linux-hirofumi/arch/x86/include/asm/atomic64_64.h	2012-04-03 17:29:38.000000000 +0900
@@ -5,6 +5,17 @@
 #include <asm/alternative.h>
 #include <asm/cmpxchg.h>
 
+#define ATOMIC64_UNDERFLOW_CHECK(v)	do {				\
+	unsigned char __sf;						\
+	/* if (atomic64_read(v) < 0) */					\
+	__asm__ __volatile__("sets %0"					\
+			     : "=qm" (__sf)				\
+			     : /* no input */				\
+			     : "memory");				\
+	WARN(__sf, KERN_ERR "atomic counter underflow: %ld\n",		\
+	     atomic64_read(v));						\
+} while(0)
+
 /* The 64-bit atomic type */
 
 #define ATOMIC64_INIT(i)	{ (i) }
@@ -121,6 +132,7 @@ static inline int atomic64_dec_and_test(
 	asm volatile(LOCK_PREFIX "decq %0; sete %1"
 		     : "=m" (v->counter), "=qm" (c)
 		     : "m" (v->counter) : "memory");
+	ATOMIC64_UNDERFLOW_CHECK(v);
 	return c != 0;
 }
 
_

>> stop_machine_first is just to see if it is first one or not.  So, there
>> is no reason to use atomic_dec_and_test(), and makes the value below 0.
>> 
>> I think it is not desirable, because this usage only triggers
>> atomic_dec_and_test() underflow debug patch. (the patch tests result
>> of atomic_dec_and_test() is < 0)
>
> Well it should only underflow if you have a box with more than 2 billion
> CPUs.

It meant < 0, not underflow INT_MIN.

>> -static atomic_t stop_machine_first;
>> +static unsigned long stop_machine_first;
>
> The down side to this is that it adds 4 more bytes on a 64bit
> machine. (sizeof(unsigned log) == 8 and sizeof(atomic_t) == 4)

Oh, sure. If nobody has interest, unfortunately I will use this as my
local patch...

Thanks.
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: [PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[OGAWA Hirofumi]

OGAWA Hirofumi <hirofumi@mail.parknet.co.jp> writes:

>> The down side to this is that it adds 4 more bytes on a 64bit
>> machine. (sizeof(unsigned log) == 8 and sizeof(atomic_t) == 4)

Another patch without additional 4bytes. This simply change
atomic_dec_and_test() to atomic_xchg().

Thanks.
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>


stop_machine_first is just to see if it is first one or not.  In this
usage, atomic_dec_and_test() makes value less than 0.

I think it is not desirable, because it only triggers
atomic_dec_and_test() less than 0 debug patch. (the patch tests result
of atomic_dec_and_test() is < 0)

So, this uses atomic_xchg() instead.

Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
---

 arch/x86/kernel/alternative.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff -puN arch/x86/kernel/alternative.c~stop_machine-use-atomic_xchg arch/x86/kernel/alternative.c
--- linux/arch/x86/kernel/alternative.c~stop_machine-use-atomic_xchg	2012-05-23 13:10:03.000000000 +0900
+++ linux-hirofumi/arch/x86/kernel/alternative.c	2012-05-23 13:10:03.000000000 +0900
@@ -664,7 +664,7 @@ static int __kprobes stop_machine_text_p
 	struct text_poke_param *p;
 	int i;
 
-	if (atomic_dec_and_test(&stop_machine_first)) {
+	if (atomic_xchg(&stop_machine_first, 0)) {
 		for (i = 0; i < tpp->nparams; i++) {
 			p = &tpp->params[i];
 			text_poke(p->addr, p->opcode, p->len);
_

Re: [PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[Steven Rostedt]

On Wed, 2012-05-23 at 13:19 +0900, OGAWA Hirofumi wrote:
> OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>
> 
> 
> stop_machine_first is just to see if it is first one or not.  In this
> usage, atomic_dec_and_test() makes value less than 0.
> 
> I think it is not desirable, because it only triggers
> atomic_dec_and_test() less than 0 debug patch. (the patch tests result
> of atomic_dec_and_test() is < 0)
> 
> So, this uses atomic_xchg() instead.

Acked-by: Steven Rostedt <rostedt@goodmis.org>

-- Steve

> 
> Signed-off-by: OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: [PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[Steven Rostedt]

On Wed, 2012-05-23 at 13:19 +0900, OGAWA Hirofumi wrote:
> OGAWA Hirofumi <hirofumi@mail.parknet.co.jp> writes:
> 
> >> The down side to this is that it adds 4 more bytes on a 64bit
> >> machine. (sizeof(unsigned log) == 8 and sizeof(atomic_t) == 4)
> 
> Another patch without additional 4bytes. This simply change
> atomic_dec_and_test() to atomic_xchg().
> 

If nobody picked this up, you might want to resend it with my acked-by.
(The second patch, not the first). As patches added to replies are
usually ignored.

-- Steve

Re: [PATCH] Use test_and_clear_bit() instead atomic_dec_and_test() for stop_machine

[OGAWA Hirofumi]

Steven Rostedt <rostedt@goodmis.org> writes:

> On Wed, 2012-05-23 at 13:19 +0900, OGAWA Hirofumi wrote:
>> OGAWA Hirofumi <hirofumi@mail.parknet.co.jp> writes:
>> 
>> >> The down side to this is that it adds 4 more bytes on a 64bit
>> >> machine. (sizeof(unsigned log) == 8 and sizeof(atomic_t) == 4)
>> 
>> Another patch without additional 4bytes. This simply change
>> atomic_dec_and_test() to atomic_xchg().
>> 
>
> If nobody picked this up, you might want to resend it with my acked-by.
> (The second patch, not the first). As patches added to replies are
> usually ignored.

Thanks. I was going to add that to my personal patchset, I will try it again.
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>