[PATCH] JFS: reduce number of synchronous transactions

Use an inline pxd list rather than an xad list in the xadlock.
When the number of extents being modified can fit with the xadlock,
a transaction can be committed asynchronously.  Using a list of
pxd's instead of xad's allows us to fit 4 extents, rather than 2.

Signed-off-by: Dave Kleikamp <shaggy@austin.ibm.com>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This commit is contained in:
Dave Kleikamp 2005-05-02 12:24:46 -06:00 committed by Linus Torvalds
parent ac09f698f1
commit 66f3131f54
1 changed files with 40 additions and 23 deletions

View File

@ -1712,7 +1712,7 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
struct maplock *maplock;
struct xdlistlock *xadlock;
struct pxd_lock *pxdlock;
pxd_t *pxd;
pxd_t *page_pxd;
int next, lwm, hwm;
ip = tlck->ip;
@ -1722,7 +1722,7 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
lrd->log.redopage.type = cpu_to_le16(LOG_XTREE);
lrd->log.redopage.l2linesize = cpu_to_le16(L2XTSLOTSIZE);
pxd = &lrd->log.redopage.pxd;
page_pxd = &lrd->log.redopage.pxd;
if (tlck->type & tlckBTROOT) {
lrd->log.redopage.type |= cpu_to_le16(LOG_BTROOT);
@ -1752,9 +1752,9 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
* applying the after-image to the meta-data page.
*/
lrd->type = cpu_to_le16(LOG_REDOPAGE);
// *pxd = mp->cm_pxd;
PXDaddress(pxd, mp->index);
PXDlength(pxd,
// *page_pxd = mp->cm_pxd;
PXDaddress(page_pxd, mp->index);
PXDlength(page_pxd,
mp->logical_size >> tblk->sb->s_blocksize_bits);
lrd->backchain = cpu_to_le32(lmLog(log, tblk, lrd, tlck));
@ -1776,25 +1776,31 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
tlck->flag |= tlckUPDATEMAP;
xadlock->flag = mlckALLOCXADLIST;
xadlock->count = next - lwm;
if ((xadlock->count <= 2) && (tblk->xflag & COMMIT_LAZY)) {
if ((xadlock->count <= 4) && (tblk->xflag & COMMIT_LAZY)) {
int i;
pxd_t *pxd;
/*
* Lazy commit may allow xtree to be modified before
* txUpdateMap runs. Copy xad into linelock to
* preserve correct data.
*
* We can fit twice as may pxd's as xads in the lock
*/
xadlock->xdlist = &xtlck->pxdlock;
memcpy(xadlock->xdlist, &p->xad[lwm],
sizeof(xad_t) * xadlock->count);
for (i = 0; i < xadlock->count; i++)
xadlock->flag = mlckALLOCPXDLIST;
pxd = xadlock->xdlist = &xtlck->pxdlock;
for (i = 0; i < xadlock->count; i++) {
PXDaddress(pxd, addressXAD(&p->xad[lwm + i]));
PXDlength(pxd, lengthXAD(&p->xad[lwm + i]));
p->xad[lwm + i].flag &=
~(XAD_NEW | XAD_EXTENDED);
pxd++;
}
} else {
/*
* xdlist will point to into inode's xtree, ensure
* that transaction is not committed lazily.
*/
xadlock->flag = mlckALLOCXADLIST;
xadlock->xdlist = &p->xad[lwm];
tblk->xflag &= ~COMMIT_LAZY;
}
@ -1836,8 +1842,8 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
if (tblk->xflag & COMMIT_TRUNCATE) {
/* write NOREDOPAGE for the page */
lrd->type = cpu_to_le16(LOG_NOREDOPAGE);
PXDaddress(pxd, mp->index);
PXDlength(pxd,
PXDaddress(page_pxd, mp->index);
PXDlength(page_pxd,
mp->logical_size >> tblk->sb->
s_blocksize_bits);
lrd->backchain =
@ -1872,22 +1878,32 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
* deleted page itself;
*/
tlck->flag |= tlckUPDATEMAP;
xadlock->flag = mlckFREEXADLIST;
xadlock->count = hwm - XTENTRYSTART + 1;
if ((xadlock->count <= 2) && (tblk->xflag & COMMIT_LAZY)) {
if ((xadlock->count <= 4) && (tblk->xflag & COMMIT_LAZY)) {
int i;
pxd_t *pxd;
/*
* Lazy commit may allow xtree to be modified before
* txUpdateMap runs. Copy xad into linelock to
* preserve correct data.
*
* We can fit twice as may pxd's as xads in the lock
*/
xadlock->xdlist = &xtlck->pxdlock;
memcpy(xadlock->xdlist, &p->xad[XTENTRYSTART],
sizeof(xad_t) * xadlock->count);
xadlock->flag = mlckFREEPXDLIST;
pxd = xadlock->xdlist = &xtlck->pxdlock;
for (i = 0; i < xadlock->count; i++) {
PXDaddress(pxd,
addressXAD(&p->xad[XTENTRYSTART + i]));
PXDlength(pxd,
lengthXAD(&p->xad[XTENTRYSTART + i]));
pxd++;
}
} else {
/*
* xdlist will point to into inode's xtree, ensure
* that transaction is not committed lazily.
*/
xadlock->flag = mlckFREEXADLIST;
xadlock->xdlist = &p->xad[XTENTRYSTART];
tblk->xflag &= ~COMMIT_LAZY;
}
@ -1918,7 +1934,7 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
* header ?
*/
if (tlck->type & tlckTRUNCATE) {
pxd_t tpxd; /* truncated extent of xad */
pxd_t pxd; /* truncated extent of xad */
int twm;
/*
@ -1947,8 +1963,9 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
* applying the after-image to the meta-data page.
*/
lrd->type = cpu_to_le16(LOG_REDOPAGE);
PXDaddress(pxd, mp->index);
PXDlength(pxd, mp->logical_size >> tblk->sb->s_blocksize_bits);
PXDaddress(page_pxd, mp->index);
PXDlength(page_pxd,
mp->logical_size >> tblk->sb->s_blocksize_bits);
lrd->backchain = cpu_to_le32(lmLog(log, tblk, lrd, tlck));
/*
@ -1966,7 +1983,7 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
lrd->log.updatemap.type = cpu_to_le16(LOG_FREEPXD);
lrd->log.updatemap.nxd = cpu_to_le16(1);
lrd->log.updatemap.pxd = pxdlock->pxd;
tpxd = pxdlock->pxd; /* save to format maplock */
pxd = pxdlock->pxd; /* save to format maplock */
lrd->backchain =
cpu_to_le32(lmLog(log, tblk, lrd, NULL));
}
@ -2035,7 +2052,7 @@ static void xtLog(struct jfs_log * log, struct tblock * tblk, struct lrd * lrd,
pxdlock = (struct pxd_lock *) xadlock;
pxdlock->flag = mlckFREEPXD;
pxdlock->count = 1;
pxdlock->pxd = tpxd;
pxdlock->pxd = pxd;
jfs_info("xtLog: truncate ip:0x%p mp:0x%p count:%d "
"hwm:%d", ip, mp, pxdlock->count, hwm);