Add a EXT2FS-specific implementation for lseek(SEEK_DATA).

The lseek(SEEK_DATA) optimization logic could be simply borrowed from ufs side.
See, https://reviews.freebsd.org/D19599.

Reviewed by:    pfg
MFC after:      1 week

Differential Revision:    https://reviews.freebsd.org/D23605
This commit is contained in:
Fedor Uporov 2020-02-18 16:39:57 +00:00
parent 56cd269eac
commit 3767ed5b11
3 changed files with 160 additions and 29 deletions

View File

@ -139,6 +139,47 @@ ext4_bmapext(struct vnode *vp, int32_t bn, int64_t *bnp, int *runp, int *runb)
return (error);
}
static int
readindir(struct vnode *vp, e2fs_lbn_t lbn, e2fs_daddr_t daddr, struct buf **bpp)
{
struct buf *bp;
struct mount *mp;
struct ext2mount *ump;
int error;
mp = vp->v_mount;
ump = VFSTOEXT2(mp);
bp = getblk(vp, lbn, mp->mnt_stat.f_iosize, 0, 0, 0);
if ((bp->b_flags & B_CACHE) == 0) {
KASSERT(daddr != 0,
("readindir: indirect block not in cache"));
bp->b_blkno = blkptrtodb(ump, daddr);
bp->b_iocmd = BIO_READ;
bp->b_flags &= ~B_INVAL;
bp->b_ioflags &= ~BIO_ERROR;
vfs_busy_pages(bp, 0);
bp->b_iooffset = dbtob(bp->b_blkno);
bstrategy(bp);
#ifdef RACCT
if (racct_enable) {
PROC_LOCK(curproc);
racct_add_buf(curproc, bp, 0);
PROC_UNLOCK(curproc);
}
#endif
curthread->td_ru.ru_inblock++;
error = bufwait(bp);
if (error != 0) {
brelse(bp);
return (error);
}
}
*bpp = bp;
return (0);
}
/*
* Indirect blocks are now on the vnode for the file. They are given negative
* logical block numbers. Indirect blocks are addressed by the negative
@ -228,34 +269,9 @@ ext2_bmaparray(struct vnode *vp, daddr_t bn, daddr_t *bnp, int *runp, int *runb)
*/
if (bp)
bqrelse(bp);
bp = getblk(vp, metalbn, bsize, 0, 0, 0);
if ((bp->b_flags & B_CACHE) == 0) {
#ifdef INVARIANTS
if (!daddr)
panic("ext2_bmaparray: indirect block not in cache");
#endif
bp->b_blkno = blkptrtodb(ump, daddr);
bp->b_iocmd = BIO_READ;
bp->b_flags &= ~B_INVAL;
bp->b_ioflags &= ~BIO_ERROR;
vfs_busy_pages(bp, 0);
bp->b_iooffset = dbtob(bp->b_blkno);
bstrategy(bp);
#ifdef RACCT
if (racct_enable) {
PROC_LOCK(curproc);
racct_add_buf(curproc, bp, 0);
PROC_UNLOCK(curproc);
}
#endif
curthread->td_ru.ru_inblock++;
error = bufwait(bp);
if (error) {
brelse(bp);
return (error);
}
}
error = readindir(vp, metalbn, daddr, &bp);
if (error != 0)
return (error);
daddr = ((e2fs_daddr_t *)bp->b_data)[ap->in_off];
if (num == 1 && daddr && runp) {
@ -296,6 +312,107 @@ ext2_bmaparray(struct vnode *vp, daddr_t bn, daddr_t *bnp, int *runp, int *runb)
return (0);
}
static e2fs_lbn_t
lbn_count(struct ext2mount *ump, int level)
{
e2fs_lbn_t blockcnt;
for (blockcnt = 1; level > 0; level--)
blockcnt *= MNINDIR(ump);
return (blockcnt);
}
int
ext2_bmap_seekdata(struct vnode *vp, off_t *offp)
{
struct buf *bp;
struct indir a[EXT2_NIADDR + 1], *ap;
struct inode *ip;
struct mount *mp;
struct ext2mount *ump;
e2fs_daddr_t bn, daddr, nextbn;
uint64_t bsize;
off_t numblks;
int error, num, num1, off;
bp = NULL;
error = 0;
ip = VTOI(vp);
mp = vp->v_mount;
ump = VFSTOEXT2(mp);
if (vp->v_type != VREG || (ip->i_flags & SF_SNAPSHOT) != 0)
return (EINVAL);
if (*offp < 0 || *offp >= ip->i_size)
return (ENXIO);
bsize = mp->mnt_stat.f_iosize;
for (bn = *offp / bsize, numblks = howmany(ip->i_size, bsize);
bn < numblks; bn = nextbn) {
if (bn < EXT2_NDADDR) {
daddr = ip->i_db[bn];
if (daddr != 0)
break;
nextbn = bn + 1;
continue;
}
ap = a;
error = ext2_getlbns(vp, bn, ap, &num);
if (error != 0)
break;
MPASS(num >= 2);
daddr = ip->i_ib[ap->in_off];
ap++, num--;
for (nextbn = EXT2_NDADDR, num1 = num - 1; num1 > 0; num1--)
nextbn += lbn_count(ump, num1);
if (daddr == 0) {
nextbn += lbn_count(ump, num);
continue;
}
for (; daddr != 0 && num > 0; ap++, num--) {
if (bp != NULL)
bqrelse(bp);
error = readindir(vp, ap->in_lbn, daddr, &bp);
if (error != 0)
return (error);
/*
* Scan the indirect block until we find a non-zero
* pointer.
*/
off = ap->in_off;
do {
daddr = ((e2fs_daddr_t *)bp->b_data)[off];
} while (daddr == 0 && ++off < MNINDIR(ump));
nextbn += off * lbn_count(ump, num - 1);
/*
* We need to recompute the LBNs of indirect
* blocks, so restart with the updated block offset.
*/
if (off != ap->in_off)
break;
}
if (num == 0) {
/*
* We found a data block.
*/
bn = nextbn;
break;
}
}
if (bp != NULL)
bqrelse(bp);
if (bn >= numblks)
error = ENXIO;
if (error == 0 && *offp < bn * bsize)
*offp = bn * bsize;
return (error);
}
/*
* Create an array of logical block number/offset pairs which represent the
* path of indirect blocks required to access a data block. The first "pair"

View File

@ -64,6 +64,7 @@ e4fs_daddr_t ext2_blkpref(struct inode *, e2fs_lbn_t, int, e2fs_daddr_t *,
int ext2_bmap(struct vop_bmap_args *);
int ext2_bmaparray(struct vnode *, daddr_t, daddr_t *, int *, int *);
int ext4_bmapext(struct vnode *, int32_t, int64_t *, int *, int *);
int ext2_bmap_seekdata(struct vnode *, off_t *);
void ext2_clusteracct(struct m_ext2fs *, char *, int, e4fs_daddr_t, int);
void ext2_dirbad(struct inode *ip, doff_t offset, char *how);
int ext2_ei2i(struct ext2fs_dinode *, struct inode *);

View File

@ -2157,11 +2157,24 @@ ext2_read(struct vop_read_args *ap)
static int
ext2_ioctl(struct vop_ioctl_args *ap)
{
struct vnode *vp;
int error;
vp = ap->a_vp;
switch (ap->a_command) {
case FIOSEEKDATA:
if (!(VTOI(vp)->i_flag & IN_E4EXTENTS)) {
error = vn_lock(vp, LK_SHARED);
if (error == 0) {
error = ext2_bmap_seekdata(vp,
(off_t *)ap->a_data);
VOP_UNLOCK(vp);
} else
error = EBADF;
return (error);
}
case FIOSEEKHOLE:
return (vn_bmap_seekhole(ap->a_vp, ap->a_command,
return (vn_bmap_seekhole(vp, ap->a_command,
(off_t *)ap->a_data, ap->a_cred));
default:
return (ENOTTY);