2006-12-12 03:17:20

by Arnaldo Carvalho de Melo

[permalink] [raw]
Subject: [PATCH][SCSI]: Save some bytes in struct scsi_target

Before:

[acme@newtoy kpahole-2.6]$ pahole --cacheline 32 /tmp/scsi.o.before scsi_target
/* include/scsi/scsi_device.h:86 */
struct scsi_target {
struct scsi_device * starget_sdev_user; /* 0 4 */
struct list_head siblings; /* 4 8 */
struct list_head devices; /* 12 8 */
struct device dev; /* 20 300 */
/* --- cacheline 10 boundary (320 bytes) --- */
unsigned int reap_ref; /* 320 4 */
unsigned int channel; /* 324 4 */
unsigned int id; /* 328 4 */
unsigned int create:1; /* 332 4 */

/* XXX 31 bits hole, try to pack */

unsigned int pdt_1f_for_no_lun; /* 336 4 */
char scsi_level; /* 340 1 */

/* XXX 3 bytes hole, try to pack */

struct execute_work ew; /* 344 16 */
/* --- cacheline 11 boundary (352 bytes) was 8 bytes ago --- */
enum scsi_target_state state; /* 360 4 */
void * hostdata; /* 364 4 */
long unsigned int starget_data[0]; /* 368 0 */
}; /* size: 368, cachelines: 12 */
/* sum members: 365, holes: 1, sum holes: 3 */
/* bit holes: 1, sum bit holes: 31 bits */
/* last cacheline: 16 bytes */

After:

[acme@newtoy kpahole-2.6]$ pahole --cacheline 32 drivers/scsi/scsi.o scsi_target
/* include/scsi/scsi_device.h:86 */
struct scsi_target {
struct scsi_device * starget_sdev_user; /* 0 4 */
struct list_head siblings; /* 4 8 */
struct list_head devices; /* 12 8 */
struct device dev; /* 20 300 */
/* --- cacheline 10 boundary (320 bytes) --- */
unsigned int reap_ref; /* 320 4 */
unsigned int channel; /* 324 4 */
unsigned int id; /* 328 4 */
char scsi_level; /* 332 1 */
unsigned char create:1; /* 333 1 */

/* XXX 7 bits hole, try to pack */
/* XXX 2 bytes hole, try to pack */

unsigned int pdt_1f_for_no_lun; /* 336 4 */
struct execute_work ew; /* 340 16 */
/* --- cacheline 11 boundary (352 bytes) was 4 bytes ago --- */
enum scsi_target_state state; /* 356 4 */
void * hostdata; /* 360 4 */
long unsigned int starget_data[0]; /* 364 0 */
}; /* size: 364, cachelines: 12 */
/* sum members: 362, holes: 1, sum holes: 2 */
/* bit holes: 1, sum bit holes: 7 bits */
/* last cacheline: 12 bytes */

[acme@newtoy kpahole-2.6]$ codiff -V /tmp/scsi.o.before drivers/scsi/scsi.o
drivers/scsi/scsi.c:
struct scsi_target | -4
create:1;
from: unsigned int /* 332(31) 4(1) */
to: unsigned char /* 333(7) 1(1) */
scsi_level;
from: char /* 340(0) 1(0) */
to: char /* 332(0) 1(0) */
<SNIP offset changes>
1 struct changed

Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>

---

scsi_device.h | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/scsi/scsi_device.h b/include/scsi/scsi_device.h
index ebf31b1..ab245fc 100644
--- a/include/scsi/scsi_device.h
+++ b/include/scsi/scsi_device.h
@@ -175,11 +175,11 @@ struct scsi_target {
unsigned int channel;
unsigned int id; /* target id ... replace
* scsi_device.id eventually */
- unsigned int create:1; /* signal that it needs to be added */
+ char scsi_level;
+ unsigned char create:1; /* signal that it needs to be added */
unsigned int pdt_1f_for_no_lun; /* PDT = 0x1f */
/* means no lun present */

- char scsi_level;
struct execute_work ew;
enum scsi_target_state state;
void *hostdata; /* available to low-level driver */


2006-12-12 03:52:28

by Matthew Wilcox

[permalink] [raw]
Subject: Re: [PATCH][SCSI]: Save some bytes in struct scsi_target

On Tue, Dec 12, 2006 at 01:17:18AM -0200, Arnaldo Carvalho de Melo wrote:
> }; /* size: 368, cachelines: 12 */
> }; /* size: 364, cachelines: 12 */

Saving space is always good ;-)

> - unsigned int create:1; /* signal that it needs to be added */
> + char scsi_level;
> + unsigned char create:1; /* signal that it needs to be added */
> unsigned int pdt_1f_for_no_lun; /* PDT = 0x1f */
> /* means no lun present */
>
> - char scsi_level;

However, pdt_1f_for_no_lun is really only one bit, saving another 4 bytes.

> struct execute_work ew;
> enum scsi_target_state state;

enums are a bit of a pain. Even though scsi_target_state uses only two
values, it's represented as an int. Unless you're on arm-eabi, when
it'll use less. And even then, it won't use less than a byte, as it has
to be addressable. I wonder if we can turn scsi_target_state into a
bit. That'll save another 8 bytes total.

2006-12-12 13:26:10

by Arnaldo Carvalho de Melo

[permalink] [raw]
Subject: Re: [PATCH][SCSI]: Save some bytes in struct scsi_target

On Mon, Dec 11, 2006 at 08:52:22PM -0700, Matthew Wilcox wrote:
> On Tue, Dec 12, 2006 at 01:17:18AM -0200, Arnaldo Carvalho de Melo wrote:
> > }; /* size: 368, cachelines: 12 */
> > }; /* size: 364, cachelines: 12 */
>
> Saving space is always good ;-)
>
> > - unsigned int create:1; /* signal that it needs to be added */
> > + char scsi_level;
> > + unsigned char create:1; /* signal that it needs to be added */
> > unsigned int pdt_1f_for_no_lun; /* PDT = 0x1f */
> > /* means no lun present */
> >
> > - char scsi_level;
>
> However, pdt_1f_for_no_lun is really only one bit, saving another 4 bytes.
>
> > struct execute_work ew;
> > enum scsi_target_state state;
>
> enums are a bit of a pain. Even though scsi_target_state uses only two
> values, it's represented as an int. Unless you're on arm-eabi, when
> it'll use less. And even then, it won't use less than a byte, as it has
> to be addressable. I wonder if we can turn scsi_target_state into a
> bit. That'll save another 8 bytes total.

I guess we could use:

enum scsi_target_state state:1;

And make the enum entries start with 0 and not 1 as is today, no? With
that we get down to:

}; /* size: 356, cachelines: 12 */
/* last cacheline: 4 bytes */

Anything else to save these 4 bytes and get down to 11 cachelines per
scsi_target instance? Following patch is on top of the previous one.

Signed-off-by: Arnaldo Carvalho de Melo <[email protected]>

diff --git a/include/scsi/scsi_device.h b/include/scsi/scsi_device.h
index ab245fc..772f834 100644
--- a/include/scsi/scsi_device.h
+++ b/include/scsi/scsi_device.h
@@ -157,7 +157,7 @@ #define scmd_printk(prefix, scmd, fmt, a
dev_printk(prefix, &(scmd)->device->sdev_gendev, fmt, ##a)

enum scsi_target_state {
- STARGET_RUNNING = 1,
+ STARGET_RUNNING = 0,
STARGET_DEL,
};

@@ -176,12 +176,12 @@ struct scsi_target {
unsigned int id; /* target id ... replace
* scsi_device.id eventually */
char scsi_level;
+ enum scsi_target_state state:1;
unsigned char create:1; /* signal that it needs to be added */
- unsigned int pdt_1f_for_no_lun; /* PDT = 0x1f */
+ unsigned char pdt_1f_for_no_lun:1; /* PDT = 0x1f */
/* means no lun present */

struct execute_work ew;
- enum scsi_target_state state;
void *hostdata; /* available to low-level driver */
unsigned long starget_data[0]; /* for the transport */
/* starget_data must be the last element!!!! */