Skip to contentSkip to conversation

workqueues: microoptimize set_wq_data()

PATCH In review Oleg Nesterov 3 participants 5 messages Feb 24, 2010 – Feb 25, 2010 Workqueue

lore
Share & tools
lore
#0 lore raw

PATCH 2/3 workqueues: microoptimize set_wq_data()

The comment correctly states that the _PENDING bit must be set and we even have the BUG_ON() check. But this means there is no need to set WORK_STRUCT_PENDING explicitely and load work_data_bits() twice, we can rely on WORK_STRUCT_FLAG_MASK which contains _PENDING.

Shaves 32 bytes from workqueue.o.

---

 kernel/workqueue.c |    7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
kernel/workqueue.c renamed from kernel/workqueue.c~2_CLEANUP_SET_DATA +2 -5
@@ -220,12 +220,9 @@ struct cpu_workqueue_struct *wq_per_cpu(
 static inline void set_wq_data(struct work_struct *work,
 				struct cpu_workqueue_struct *cwq)
 {
-	unsigned long new;
-
-	BUG_ON(!work_pending(work));
-
-	new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
+	unsigned long new = (unsigned long)cwq;
 	new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
+	BUG_ON(!(new & (1UL << WORK_STRUCT_PENDING)));
 	atomic_long_set(&work->data, new);
 }
 
maintainer
#1 lore raw
Quoted 18 lines · Oleg Nesterov

The comment correctly states that the _PENDING bit must be set and we even have the BUG_ON() check. But this means there is no need to set WORK_STRUCT_PENDING explicitely and load work_data_bits() twice, we can rely on WORK_STRUCT_FLAG_MASK which contains _PENDING.

Shaves 32 bytes from workqueue.o.

---

 kernel/workqueue.c |    7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)
kernel/workqueue.c renamed from kernel/workqueue.c~2_CLEANUP_SET_DATA +0 -0
@@ -220,12 +220,9 @@ struct cpu_workqueue_struct *wq_per_cpu(
 static inline void set_wq_data(struct work_struct *work,
 				struct cpu_workqueue_struct *cwq)

kernel/workqueue.c struct cpu_workqueue_struct *wq_per_cpu( · line 226

Quoted 10 lines
kernel/workqueue.c renamed from kernel/workqueue.c~2_CLEANUP_SET_DATA +2 -5
 {
-	unsigned long new;
-
-	BUG_ON(!work_pending(work));
-
-	new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
+	unsigned long new = (unsigned long)cwq;
 	new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
+	BUG_ON(!(new & (1UL << WORK_STRUCT_PENDING)));
 	atomic_long_set(&work->data, new);

Will apply under cmwq patches for the next merge window.

Thanks.

Signature · 1 line
tejun
#2 lore raw

Hello, again.

Quoted 7 lines · Tejun Heo
Quoted 7 lines
kernel/workqueue.c renamed from kernel/workqueue.c~2_CLEANUP_SET_DATA +0 -3
@@ -220,12 +220,9 @@ struct cpu_workqueue_struct *wq_per_cpu(
 static inline void set_wq_data(struct work_struct *work,
 				struct cpu_workqueue_struct *cwq)
 {
-	unsigned long new;
-
-	BUG_ON(!work_pending(work));
Quoted 8 lines

kernel/workqueue.c struct cpu_workqueue_struct *wq_per_cpu( · line 226

Quoted 6 lines
kernel/workqueue.c renamed from kernel/workqueue.c~2_CLEANUP_SET_DATA +2 -2
-
-	new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
+	unsigned long new = (unsigned long)cwq;
 	new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
+	BUG_ON(!(new & (1UL << WORK_STRUCT_PENDING)));
 	atomic_long_set(&work->data, new);

Will apply under cmwq patches for the next merge window.

Turns out I already have a patch which kills the second work_data_bits() dereferencing in the series. The first one is now in the cmwq series which is about to be posted again.

Thanks.

Signature · 1 line
tejun
#3 lore raw

Hi,

Quoted 20 lines · Tejun Heo

Hello, again.

Quoted 7 lines · Tejun Heo
Quoted 7 lines
kernel/workqueue.c renamed from kernel/workqueue.c~2_CLEANUP_SET_DATA +0 -3
@@ -220,12 +220,9 @@ struct cpu_workqueue_struct *wq_per_cpu(
 static inline void set_wq_data(struct work_struct *work,
 				struct cpu_workqueue_struct *cwq)
 {
-	unsigned long new;
-
-	BUG_ON(!work_pending(work));
Quoted 8 lines

kernel/workqueue.c struct cpu_workqueue_struct *wq_per_cpu( · line 226

Quoted 6 lines
kernel/workqueue.c renamed from kernel/workqueue.c~2_CLEANUP_SET_DATA +2 -2
-
-	new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
+	unsigned long new = (unsigned long)cwq;
 	new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
+	BUG_ON(!(new & (1UL << WORK_STRUCT_PENDING)));
 	atomic_long_set(&work->data, new);

Will apply under cmwq patches for the next merge window.

Quoted 3 lines

Turns out I already have a patch which kills the second work_data_bits() dereferencing in the series. The first one is now in the cmwq series which is about to be posted again.

OK, good.

Tejun, where can I find your cmwq patches? I'd like to take a look.

Thanks,

Oleg.

#4 lore raw

Hello,

Quoted 7 lines · Oleg Nesterov
Quoted 3 lines

Turns out I already have a patch which kills the second work_data_bits() dereferencing in the series. The first one is now in the cmwq series which is about to be posted again.

OK, good.

Tejun, where can I find your cmwq patches? I'd like to take a look.

The third take was about a month ago.

http://thread.gmane.org/gmane.linux.kernel/939353

The fourth take is just around the corner.

Thanks.

Signature · 1 line
tejun