Skip to contentSkip to conversation

cfq: reorder cfq_queue removing padding on 64bit & allowing more objects/slab in it's kmem_cache

PATCH Applied Richard Kennedy 4 participants 6 messages Feb 15, 2010 – Feb 22, 2010 Block layer

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 cfq: reorder cfq_queue removing padding on 64bit & allowing more objects/slab in it's kmem_cache

This removes 8 bytes of padding from struct cfq_queue on 64 bit builds, shrinking it's size to 256 bytes, so fitting into 1 fewer cachelines and allowing 1 more object/slab in it's kmem_cache.

----
patch against 2.6.33-rc8
tested on x86_64 AMDX2

regards
Richard

block/cfq-iosched.c +3 -3
index 023f4e6..db4c127 100644
@@ -116,10 +116,10 @@ struct cfq_queue {
 	unsigned long dispatch_start;
 	unsigned int allocated_slice;
 	/* time when first request from queue completed and slice started. */
+	unsigned int slice_dispatch;
 	unsigned long slice_start;
 	unsigned long slice_end;
 	long slice_resid;
-	unsigned int slice_dispatch;
 
 	/* pending metadata requests */
 	int meta_pending;
@@ -130,13 +130,13 @@ struct cfq_queue {
 	unsigned short ioprio, org_ioprio;
 	unsigned short ioprio_class, org_ioprio_class;
 
+	pid_t pid;
+
 	unsigned int seek_samples;
 	u64 seek_total;
 	sector_t seek_mean;
 	sector_t last_request_pos;
 
-	pid_t pid;
-
 	struct cfq_rb_root *service_tree;
 	struct cfq_queue *new_cfqq;
 	struct cfq_group *cfqg;
#1 lore raw
Quoted 19 lines · Richard Kennedy

This removes 8 bytes of padding from struct cfq_queue on 64 bit builds, shrinking it's size to 256 bytes, so fitting into 1 fewer cachelines and allowing 1 more object/slab in it's kmem_cache.

----
patch against 2.6.33-rc8
tested on x86_64 AMDX2

regards
Richard

block/cfq-iosched.c +0 -0
index 023f4e6..db4c127 100644

block/cfq-iosched.c struct cfq_queue { · line 119

Quoted 4 lines
block/cfq-iosched.c +1 -0
@@ -116,10 +116,10 @@ struct cfq_queue {
 	unsigned long dispatch_start;
 	unsigned int allocated_slice;
 	/* time when first request from queue completed and slice started. */
+	unsigned int slice_dispatch;

Can you move it above the comment. The comment describes "slice_start" and now comment and variable have been separated.

Vivek

Quoted 24 lines
block/cfq-iosched.c +2 -3
 	unsigned long slice_start;
 	unsigned long slice_end;
 	long slice_resid;
-	unsigned int slice_dispatch;
 
 	/* pending metadata requests */
 	int meta_pending;
@@ -130,13 +130,13 @@ struct cfq_queue {
 	unsigned short ioprio, org_ioprio;
 	unsigned short ioprio_class, org_ioprio_class;
 
+	pid_t pid;
+
 	unsigned int seek_samples;
 	u64 seek_total;
 	sector_t seek_mean;
 	sector_t last_request_pos;
 
-	pid_t pid;
-
 	struct cfq_rb_root *service_tree;
 	struct cfq_queue *new_cfqq;
 	struct cfq_group *cfqg;
#2 lore raw
Quoted 28 lines · Vivek Goyal
Quoted 19 lines · Richard Kennedy

This removes 8 bytes of padding from struct cfq_queue on 64 bit builds, shrinking it's size to 256 bytes, so fitting into 1 fewer cachelines and allowing 1 more object/slab in it's kmem_cache.

----
patch against 2.6.33-rc8
tested on x86_64 AMDX2

regards
Richard

block/cfq-iosched.c +0 -0
index 023f4e6..db4c127 100644

block/cfq-iosched.c struct cfq_queue { · line 119

Quoted 4 lines
block/cfq-iosched.c +1 -0
@@ -116,10 +116,10 @@ struct cfq_queue {
 	unsigned long dispatch_start;
 	unsigned int allocated_slice;
 	/* time when first request from queue completed and slice started. */
+	unsigned int slice_dispatch;

Can you move it above the comment. The comment describes "slice_start" and now comment and variable have been separated.

Quoted 1 line

Vivek

Sure, no problem. I'll send out a new version.
Richard

#3 lore raw
In reply to Vivek Goyal

This removes 8 bytes of padding from struct cfq_queue on 64 bit builds, shrinking it's size to 256 bytes, so fitting into 1 fewer cachelines and allowing 1 more object/slab in it's kmem_cache.

----
patch against 2.6.33-rc8
tested on x86_64 AMDX2

block/cfq-iosched.c +3 -3
index 023f4e6..e3dedfd 100644
@@ -115,11 +115,11 @@ struct cfq_queue {
 	/* time when queue got scheduled in to dispatch first request. */
 	unsigned long dispatch_start;
 	unsigned int allocated_slice;
+	unsigned int slice_dispatch;
 	/* time when first request from queue completed and slice started. */
 	unsigned long slice_start;
 	unsigned long slice_end;
 	long slice_resid;
-	unsigned int slice_dispatch;
 
 	/* pending metadata requests */
 	int meta_pending;
@@ -130,13 +130,13 @@ struct cfq_queue {
 	unsigned short ioprio, org_ioprio;
 	unsigned short ioprio_class, org_ioprio_class;
 
+	pid_t pid;
+
 	unsigned int seek_samples;
 	u64 seek_total;
 	sector_t seek_mean;
 	sector_t last_request_pos;
 
-	pid_t pid;
-
 	struct cfq_rb_root *service_tree;
 	struct cfq_queue *new_cfqq;
 	struct cfq_group *cfqg;
Reviewed-by
#4 lore raw
Quoted 3 lines · Richard Kennedy

This removes 8 bytes of padding from struct cfq_queue on 64 bit builds, shrinking it's size to 256 bytes, so fitting into 1 fewer cachelines and allowing 1 more object/slab in it's kmem_cache.

OK, I ran pahole to verify your findings:

$ pahole -C cfq_queue build/master/block/cfq-iosched.o 
struct cfq_queue {
...
         unsigned int               allocated_slice;      /*   136     4  */

/* XXX 4 bytes hole, try to pack */

long unsigned int slice_start; /* 144 8 */

...
pid_t pid; /* 216 4 */

/* XXX 4 bytes hole, try to pack */

struct cfq_rb_root * service_tree; /* 224 8 */

...

        /* size: 264, cachelines: 5, members: 34 */
        /* sum members: 256, holes: 2, sum holes: 8 */
        /* last cacheline: 8 bytes */

After applying the patch, it indeed does save a cacheline.

Applied
#5 lore raw
In reply to Richard Kennedy
Quoted 3 lines · Richard Kennedy

This removes 8 bytes of padding from struct cfq_queue on 64 bit builds, shrinking it's size to 256 bytes, so fitting into 1 fewer cachelines and allowing 1 more object/slab in it's kmem_cache.

Thanks, applied!

Signature · 1 line
Jens Axboe