int
lfs_markv(struct proc *p, fsid_t *fsidp, BLOCK_INFO *blkiov,
int blkcnt)
{
BLOCK_INFO *blkp;
IFILE *ifp;
struct buf *bp;
struct inode *ip = NULL;
struct lfs *fs;
struct mount *mntp;
struct vnode *vp = NULL;
ino_t lastino;
daddr_t b_daddr, v_daddr;
int cnt, error;
int do_again = 0;
int numrefed = 0;
ino_t maxino;
size_t obsize;
/* number of blocks/inodes that we have already bwrite'ed */
int nblkwritten, ninowritten;
if ((mntp = vfs_getvfs(fsidp)) == NULL)
return (ENOENT);
fs = VFSTOULFS(mntp)->um_lfs;
if (fs->lfs_ronly)
return EROFS;
maxino = (lfs_fragstoblks(fs, VTOI(fs->lfs_ivnode)->i_ffs1_blocks) -
fs->lfs_cleansz - fs->lfs_segtabsz) * fs->lfs_ifpb;
cnt = blkcnt;
if ((error = vfs_busy(mntp, NULL)) != 0)
return (error);
/*
* This seglock is just to prevent the fact that we might have to sleep
* from allowing the possibility that our blocks might become
* invalid.
*
* It is also important to note here that unless we specify SEGM_CKP,
* any Ifile blocks that we might be asked to clean will never get
* to the disk.
*/
lfs_seglock(fs, SEGM_CLEAN | SEGM_CKP | SEGM_SYNC);
/* Mark blocks/inodes dirty. */
error = 0;
/* these were inside the initialization for the for loop */
v_daddr = LFS_UNUSED_DADDR;
lastino = LFS_UNUSED_INUM;
nblkwritten = ninowritten = 0;
for (blkp = blkiov; cnt--; ++blkp)
{
/* Bounds-check incoming data, avoid panic for failed VGET */
if (blkp->bi_inode <= 0 || blkp->bi_inode >= maxino) {
error = EINVAL;
goto err3;
}
/*
* Get the IFILE entry (only once) and see if the file still
* exists.
*/
if (lastino != blkp->bi_inode) {
/*
* Finish the old file, if there was one. The presence
* of a usable vnode in vp is signaled by a valid v_daddr.
*/
if (v_daddr != LFS_UNUSED_DADDR) {
lfs_vunref(vp);
numrefed--;
}
/*
* Start a new file
*/
lastino = blkp->bi_inode;
if (blkp->bi_inode == LFS_IFILE_INUM)
v_daddr = fs->lfs_idaddr;
else {
LFS_IENTRY(ifp, fs, blkp->bi_inode, bp);
/* XXX fix for force write */
v_daddr = ifp->if_daddr;
brelse(bp, 0);
}
if (v_daddr == LFS_UNUSED_DADDR)
continue;
/* Get the vnode/inode. */
error = lfs_fastvget(mntp, blkp->bi_inode, v_daddr,
&vp,
(blkp->bi_lbn == LFS_UNUSED_LBN
? blkp->bi_bp
: NULL));
if (!error) {
numrefed++;
}
if (error) {
DLOG((DLOG_CLEAN, "lfs_markv: lfs_fastvget"
" failed with %d (ino %d, segment %d)\n",
error, blkp->bi_inode,
lfs_dtosn(fs, blkp->bi_daddr)));
/*
* If we got EAGAIN, that means that the
* Inode was locked. This is
* recoverable: just clean the rest of
* this segment, and let the cleaner try
* again with another. (When the
* cleaner runs again, this segment will
* sort high on the list, since it is
* now almost entirely empty.) But, we
* still set v_daddr = LFS_UNUSED_ADDR
* so as not to test this over and over
* again.
*/
if (error == EAGAIN) {
error = 0;
do_again++;
}
#ifdef DIAGNOSTIC
else if (error != ENOENT)
panic("lfs_markv VFS_VGET FAILED");
#endif
/* lastino = LFS_UNUSED_INUM; */
v_daddr = LFS_UNUSED_DADDR;
vp = NULL;
ip = NULL;
continue;
}
ip = VTOI(vp);
ninowritten++;
} else if (v_daddr == LFS_UNUSED_DADDR) {
/*
* This can only happen if the vnode is dead (or
* in any case we can't get it...e.g., it is
* inlocked). Keep going.
*/
continue;
}
/* Past this point we are guaranteed that vp, ip are valid. */
/* Can't clean VU_DIROP directories in case of truncation */
/* XXX - maybe we should mark removed dirs specially? */
if (vp->v_type == VDIR && (vp->v_uflag & VU_DIROP)) {
do_again++;
continue;
}
/* If this BLOCK_INFO didn't contain a block, keep going. */
if (blkp->bi_lbn == LFS_UNUSED_LBN) {
/* XXX need to make sure that the inode gets written in this case */
/* XXX but only write the inode if it's the right one */
if (blkp->bi_inode != LFS_IFILE_INUM) {
LFS_IENTRY(ifp, fs, blkp->bi_inode, bp);
if (ifp->if_daddr == blkp->bi_daddr) {
mutex_enter(&lfs_lock);
LFS_SET_UINO(ip, IN_CLEANING);
mutex_exit(&lfs_lock);
}
brelse(bp, 0);
}
continue;
}
b_daddr = 0;
if (VOP_BMAP(vp, blkp->bi_lbn, NULL, &b_daddr, NULL) ||
LFS_DBTOFSB(fs, b_daddr) != blkp->bi_daddr)
{
if (lfs_dtosn(fs, LFS_DBTOFSB(fs, b_daddr)) ==
lfs_dtosn(fs, blkp->bi_daddr))
{
DLOG((DLOG_CLEAN, "lfs_markv: wrong da same seg: %llx vs %llx\n",
(long long)blkp->bi_daddr, (long long)LFS_DBTOFSB(fs, b_daddr)));
}
do_again++;
continue;
}
/*
* Check block sizes. The blocks being cleaned come from
* disk, so they should have the same size as their on-disk
* counterparts.
*/
if (blkp->bi_lbn >= 0)
obsize = lfs_blksize(fs, ip, blkp->bi_lbn);
else
obsize = fs->lfs_bsize;
/* Check for fragment size change */
if (blkp->bi_lbn >= 0 && blkp->bi_lbn < ULFS_NDADDR) {
obsize = ip->i_lfs_fragsize[blkp->bi_lbn];
}
if (obsize != blkp->bi_size) {
DLOG((DLOG_CLEAN, "lfs_markv: ino %d lbn %lld wrong"
" size (%ld != %d), try again\n",
blkp->bi_inode, (long long)blkp->bi_lbn,
(long) obsize, blkp->bi_size));
do_again++;
continue;
}
/*
* If we get to here, then we are keeping the block. If
* it is an indirect block, we want to actually put it
* in the buffer cache so that it can be updated in the
* finish_meta section. If it's not, we need to
* allocate a fake buffer so that writeseg can perform
* the copyin and write the buffer.
*/
if (ip->i_number != LFS_IFILE_INUM && blkp->bi_lbn >= 0) {
/* Data Block */
bp = lfs_fakebuf(fs, vp, blkp->bi_lbn,
blkp->bi_size, blkp->bi_bp);
/* Pretend we used bread() to get it */
bp->b_blkno = LFS_FSBTODB(fs, blkp->bi_daddr);
} else {
/* Indirect block or ifile */
if (blkp->bi_size != fs->lfs_bsize &&
ip->i_number != LFS_IFILE_INUM)
panic("lfs_markv: partial indirect block?"
" size=%d\n", blkp->bi_size);
bp = getblk(vp, blkp->bi_lbn, blkp->bi_size, 0, 0);
if (!(bp->b_oflags & (BO_DONE|BO_DELWRI))) {
/*
* The block in question was not found
* in the cache; i.e., the block that
* getblk() returned is empty. So, we
* can (and should) copy in the
* contents, because we've already
* determined that this was the right
* version of this block on disk.
*
* And, it can't have changed underneath
* us, because we have the segment lock.
*/
error = copyin(blkp->bi_bp, bp->b_data, blkp->bi_size);
if (error)
goto err2;
}
}
if ((error = lfs_bwrite_ext(bp, BW_CLEAN)) != 0)
goto err2;
nblkwritten++;
/*
* XXX should account indirect blocks and ifile pages as well
*/
if (nblkwritten + lfs_lblkno(fs, ninowritten * sizeof (struct ulfs1_dinode))
> LFS_MARKV_MAX_BLOCKS) {
DLOG((DLOG_CLEAN, "lfs_markv: writing %d blks %d inos\n",
nblkwritten, ninowritten));
lfs_segwrite(mntp, SEGM_CLEAN);
nblkwritten = ninowritten = 0;
}
}
/*
* Finish the old file, if there was one
*/
if (v_daddr != LFS_UNUSED_DADDR) {
lfs_vunref(vp);
numrefed--;
}
#ifdef DIAGNOSTIC
if (numrefed != 0)
panic("lfs_markv: numrefed=%d", numrefed);
#endif
DLOG((DLOG_CLEAN, "lfs_markv: writing %d blks %d inos (check point)\n",
nblkwritten, ninowritten));
/*
* The last write has to be SEGM_SYNC, because of calling semantics.
* It also has to be SEGM_CKP, because otherwise we could write
* over the newly cleaned data contained in a checkpoint, and then
* we'd be unhappy at recovery time.
*/
lfs_segwrite(mntp, SEGM_CLEAN | SEGM_CKP | SEGM_SYNC);
lfs_segunlock(fs);
vfs_unbusy(mntp, false, NULL);
if (error)
return (error);
else if (do_again)
return EAGAIN;
return 0;
err2:
DLOG((DLOG_CLEAN, "lfs_markv err2\n"));
/*
* XXX we're here because copyin() failed.
* XXX it means that we can't trust the cleanerd. too bad.
* XXX how can we recover from this?
*/
err3:
/*
* XXX should do segwrite here anyway?
*/
if (v_daddr != LFS_UNUSED_DADDR) {
lfs_vunref(vp);
--numrefed;
}
lfs_segunlock(fs);
vfs_unbusy(mntp, false, NULL);
#ifdef DIAGNOSTIC
if (numrefed != 0)
panic("lfs_markv: numrefed=%d", numrefed);
#endif
return (error);
}
int
lfs_bmapv(struct lwp *l, fsid_t *fsidp, BLOCK_INFO *blkiov, int blkcnt)
{
BLOCK_INFO *blkp;
IFILE *ifp;
struct buf *bp;
struct inode *ip = NULL;
struct lfs *fs;
struct mount *mntp;
struct ulfsmount *ump;
struct vnode *vp;
ino_t lastino;
daddr_t v_daddr;
int cnt, error;
int numrefed = 0;
error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_LFS,
KAUTH_REQ_SYSTEM_LFS_BMAPV, NULL, NULL, NULL);
if (error)
return (error);
if ((mntp = vfs_getvfs(fsidp)) == NULL)
return (ENOENT);
if ((error = vfs_busy(mntp, NULL)) != 0)
return (error);
ump = VFSTOULFS(mntp);
fs = ump->um_lfs;
if (fs->lfs_cleaner_thread == NULL)
fs->lfs_cleaner_thread = curlwp;
KASSERT(fs->lfs_cleaner_thread == curlwp);
cnt = blkcnt;
error = 0;
/* these were inside the initialization for the for loop */
vp = NULL;
v_daddr = LFS_UNUSED_DADDR;
lastino = LFS_UNUSED_INUM;
for (blkp = blkiov; cnt--; ++blkp)
{
/*
* Get the IFILE entry (only once) and see if the file still
* exists.
*/
if (lastino != blkp->bi_inode) {
/*
* Finish the old file, if there was one.
*/
if (vp != NULL) {
vput(vp);
vp = NULL;
numrefed--;
}
/*
* Start a new file
*/
lastino = blkp->bi_inode;
if (blkp->bi_inode == LFS_IFILE_INUM)
v_daddr = lfs_sb_getidaddr(fs);
else {
LFS_IENTRY(ifp, fs, blkp->bi_inode, bp);
v_daddr = lfs_if_getdaddr(fs, ifp);
brelse(bp, 0);
}
if (v_daddr == LFS_UNUSED_DADDR) {
blkp->bi_daddr = LFS_UNUSED_DADDR;
continue;
}
error = lfs_fastvget(mntp, blkp->bi_inode, NULL,
LK_SHARED, &vp);
if (error) {
DLOG((DLOG_CLEAN, "lfs_bmapv: lfs_fastvget ino"
"%d failed with %d",
blkp->bi_inode,error));
KASSERT(vp == NULL);
continue;
} else {
KASSERT(VOP_ISLOCKED(vp));
numrefed++;
}
ip = VTOI(vp);
} else if (vp == NULL) {
/*
* This can only happen if the vnode is dead.
* Keep going. Note that we DO NOT set the
* bi_addr to anything -- if we failed to get
* the vnode, for example, we want to assume
* conservatively that all of its blocks *are*
* located in the segment in question.
* lfs_markv will throw them out if we are
* wrong.
*/
continue;
}
/* Past this point we are guaranteed that vp, ip are valid. */
if (blkp->bi_lbn == LFS_UNUSED_LBN) {
/*
* We just want the inode address, which is
* conveniently in v_daddr.
*/
blkp->bi_daddr = v_daddr;
} else {
daddr_t bi_daddr;
error = VOP_BMAP(vp, blkp->bi_lbn, NULL,
&bi_daddr, NULL);
if (error)
{
blkp->bi_daddr = LFS_UNUSED_DADDR;
continue;
}
blkp->bi_daddr = LFS_DBTOFSB(fs, bi_daddr);
/* Fill in the block size, too */
if (blkp->bi_lbn >= 0)
blkp->bi_size = lfs_blksize(fs, ip, blkp->bi_lbn);
else
blkp->bi_size = lfs_sb_getbsize(fs);
}
}
/*
* Finish the old file, if there was one.
*/
if (vp != NULL) {
vput(vp);
vp = NULL;
numrefed--;
}
#ifdef DIAGNOSTIC
if (numrefed != 0)
panic("lfs_bmapv: numrefed=%d", numrefed);
#endif
vfs_unbusy(mntp, false, NULL);
return 0;
}
