xfs_qm.c

linux-2.6.15.6

			 */
			xfs_qm_dqflock_pushbuf_wait(dqp);
		}
		/*
		 * Let go of the mplist lock. We don't want to hold it
		 * across a disk write.
		 */
		xfs_qm_mplist_unlock(mp);
		error = xfs_qm_dqflush(dqp, flags);
		xfs_dqunlock(dqp);
		if (error)
			return (error);

		xfs_qm_mplist_lock(mp);
		if (recl != XFS_QI_MPLRECLAIMS(mp)) {
			xfs_qm_mplist_unlock(mp);
			/* XXX restart limit */
			goto again;
		}
	}

	xfs_qm_mplist_unlock(mp);
	/* return ! busy */
	return (0);
}
/*
 * Release the group dquot pointers the user dquots may be
 * carrying around as a hint. mplist is locked on entry and exit.
 */
STATIC void
xfs_qm_detach_gdquots(
	xfs_mount_t	*mp)
{
	xfs_dquot_t	*dqp, *gdqp;
	int		nrecl;

 again:
	ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
	dqp = XFS_QI_MPLNEXT(mp);
	while (dqp) {
		xfs_dqlock(dqp);
		if ((gdqp = dqp->q_gdquot)) {
			xfs_dqlock(gdqp);
			dqp->q_gdquot = NULL;
		}
		xfs_dqunlock(dqp);

		if (gdqp) {
			/*
			 * Can't hold the mplist lock across a dqput.
			 * XXXmust convert to marker based iterations here.
			 */
			nrecl = XFS_QI_MPLRECLAIMS(mp);
			xfs_qm_mplist_unlock(mp);
			xfs_qm_dqput(gdqp);

			xfs_qm_mplist_lock(mp);
			if (nrecl != XFS_QI_MPLRECLAIMS(mp))
				goto again;
		}
		dqp = dqp->MPL_NEXT;
	}
}

/*
 * Go through all the incore dquots of this file system and take them
 * off the mplist and hashlist, if the dquot type matches the dqtype
 * parameter. This is used when turning off quota accounting for
 * users and/or groups, as well as when the filesystem is unmounting.
 */
STATIC int
xfs_qm_dqpurge_int(
	xfs_mount_t	*mp,
	uint		flags) /* QUOTAOFF/UMOUNTING/UQUOTA/PQUOTA/GQUOTA */
{
	xfs_dquot_t	*dqp;
	uint		dqtype;
	int		nrecl;
	xfs_dquot_t	*nextdqp;
	int		nmisses;

	if (mp->m_quotainfo == NULL)
		return (0);

	dqtype = (flags & XFS_QMOPT_UQUOTA) ? XFS_DQ_USER : 0;
	dqtype |= (flags & XFS_QMOPT_PQUOTA) ? XFS_DQ_PROJ : 0;
	dqtype |= (flags & XFS_QMOPT_GQUOTA) ? XFS_DQ_GROUP : 0;

	xfs_qm_mplist_lock(mp);

	/*
	 * In the first pass through all incore dquots of this filesystem,
	 * we release the group dquot pointers the user dquots may be
	 * carrying around as a hint. We need to do this irrespective of
	 * what's being turned off.
	 */
	xfs_qm_detach_gdquots(mp);

      again:
	nmisses = 0;
	ASSERT(XFS_QM_IS_MPLIST_LOCKED(mp));
	/*
	 * Try to get rid of all of the unwanted dquots. The idea is to
	 * get them off mplist and hashlist, but leave them on freelist.
	 */
	dqp = XFS_QI_MPLNEXT(mp);
	while (dqp) {
		/*
		 * It's OK to look at the type without taking dqlock here.
		 * We're holding the mplist lock here, and that's needed for
		 * a dqreclaim.
		 */
		if ((dqp->dq_flags & dqtype) == 0) {
			dqp = dqp->MPL_NEXT;
			continue;
		}

		if (! xfs_qm_dqhashlock_nowait(dqp)) {
			nrecl = XFS_QI_MPLRECLAIMS(mp);
			xfs_qm_mplist_unlock(mp);
			XFS_DQ_HASH_LOCK(dqp->q_hash);
			xfs_qm_mplist_lock(mp);

			/*
			 * XXXTheoretically, we can get into a very long
			 * ping pong game here.
			 * No one can be adding dquots to the mplist at
			 * this point, but somebody might be taking things off.
			 */
			if (nrecl != XFS_QI_MPLRECLAIMS(mp)) {
				XFS_DQ_HASH_UNLOCK(dqp->q_hash);
				goto again;
			}
		}

		/*
		 * Take the dquot off the mplist and hashlist. It may remain on
		 * freelist in INACTIVE state.
		 */
		nextdqp = dqp->MPL_NEXT;
		nmisses += xfs_qm_dqpurge(dqp, flags);
		dqp = nextdqp;
	}
	xfs_qm_mplist_unlock(mp);
	return nmisses;
}

int
xfs_qm_dqpurge_all(
	xfs_mount_t	*mp,
	uint		flags)
{
	int		ndquots;

	/*
	 * Purge the dquot cache.
	 * None of the dquots should really be busy at this point.
	 */
	if (mp->m_quotainfo) {
		while ((ndquots = xfs_qm_dqpurge_int(mp, flags))) {
			delay(ndquots * 10);
		}
	}
	return 0;
}

STATIC int
xfs_qm_dqattach_one(
	xfs_inode_t	*ip,
	xfs_dqid_t	id,
	uint		type,
	uint		doalloc,
	uint		dolock,
	xfs_dquot_t	*udqhint, /* hint */
	xfs_dquot_t	**IO_idqpp)
{
	xfs_dquot_t	*dqp;
	int		error;

	ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
	error = 0;
	/*
	 * See if we already have it in the inode itself. IO_idqpp is
	 * &i_udquot or &i_gdquot. This made the code look weird, but
	 * made the logic a lot simpler.
	 */
	if ((dqp = *IO_idqpp)) {
		if (dolock)
			xfs_dqlock(dqp);
		xfs_dqtrace_entry(dqp, "DQATTACH: found in ip");
		goto done;
	}

	/*
	 * udqhint is the i_udquot field in inode, and is non-NULL only
	 * when the type arg is group/project. Its purpose is to save a
	 * lookup by dqid (xfs_qm_dqget) by caching a group dquot inside
	 * the user dquot.
	 */
	ASSERT(!udqhint || type == XFS_DQ_GROUP || type == XFS_DQ_PROJ);
	if (udqhint && !dolock)
		xfs_dqlock(udqhint);

	/*
	 * No need to take dqlock to look at the id.
	 * The ID can't change until it gets reclaimed, and it won't
	 * be reclaimed as long as we have a ref from inode and we hold
	 * the ilock.
	 */
	if (udqhint &&
	    (dqp = udqhint->q_gdquot) &&
	    (be32_to_cpu(dqp->q_core.d_id) == id)) {
		ASSERT(XFS_DQ_IS_LOCKED(udqhint));
		xfs_dqlock(dqp);
		XFS_DQHOLD(dqp);
		ASSERT(*IO_idqpp == NULL);
		*IO_idqpp = dqp;
		if (!dolock) {
			xfs_dqunlock(dqp);
			xfs_dqunlock(udqhint);
		}
		goto done;
	}
	/*
	 * We can't hold a dquot lock when we call the dqget code.
	 * We'll deadlock in no time, because of (not conforming to)
	 * lock ordering - the inodelock comes before any dquot lock,
	 * and we may drop and reacquire the ilock in xfs_qm_dqget().
	 */
	if (udqhint)
		xfs_dqunlock(udqhint);
	/*
	 * Find the dquot from somewhere. This bumps the
	 * reference count of dquot and returns it locked.
	 * This can return ENOENT if dquot didn't exist on
	 * disk and we didn't ask it to allocate;
	 * ESRCH if quotas got turned off suddenly.
	 */
	if ((error = xfs_qm_dqget(ip->i_mount, ip, id, type,
				 doalloc|XFS_QMOPT_DOWARN, &dqp))) {
		if (udqhint && dolock)
			xfs_dqlock(udqhint);
		goto done;
	}

	xfs_dqtrace_entry(dqp, "DQATTACH: found by dqget");
	/*
	 * dqget may have dropped and re-acquired the ilock, but it guarantees
	 * that the dquot returned is the one that should go in the inode.
	 */
	*IO_idqpp = dqp;
	ASSERT(dqp);
	ASSERT(XFS_DQ_IS_LOCKED(dqp));
	if (! dolock) {
		xfs_dqunlock(dqp);
		goto done;
	}
	if (! udqhint)
		goto done;

	ASSERT(udqhint);
	ASSERT(dolock);
	ASSERT(XFS_DQ_IS_LOCKED(dqp));
	if (! xfs_qm_dqlock_nowait(udqhint)) {
		xfs_dqunlock(dqp);
		xfs_dqlock(udqhint);
		xfs_dqlock(dqp);
	}
      done:
#ifdef QUOTADEBUG
	if (udqhint) {
		if (dolock)
			ASSERT(XFS_DQ_IS_LOCKED(udqhint));
	}
	if (! error) {
		if (dolock)
			ASSERT(XFS_DQ_IS_LOCKED(dqp));
	}
#endif
	return (error);
}


/*
 * Given a udquot and gdquot, attach a ptr to the group dquot in the
 * udquot as a hint for future lookups. The idea sounds simple, but the
 * execution isn't, because the udquot might have a group dquot attached
 * already and getting rid of that gets us into lock ordering contraints.
 * The process is complicated more by the fact that the dquots may or may not
 * be locked on entry.
 */
STATIC void
xfs_qm_dqattach_grouphint(
	xfs_dquot_t	*udq,
	xfs_dquot_t	*gdq,
	uint		locked)
{
	xfs_dquot_t	*tmp;

#ifdef QUOTADEBUG
	if (locked) {
		ASSERT(XFS_DQ_IS_LOCKED(udq));
		ASSERT(XFS_DQ_IS_LOCKED(gdq));
	}
#endif
	if (! locked)
		xfs_dqlock(udq);

	if ((tmp = udq->q_gdquot)) {
		if (tmp == gdq) {
			if (! locked)
				xfs_dqunlock(udq);
			return;
		}

		udq->q_gdquot = NULL;
		/*
		 * We can't keep any dqlocks when calling dqrele,
		 * because the freelist lock comes before dqlocks.
		 */
		xfs_dqunlock(udq);
		if (locked)
			xfs_dqunlock(gdq);
		/*
		 * we took a hard reference once upon a time in dqget,
		 * so give it back when the udquot no longer points at it
		 * dqput() does the unlocking of the dquot.
		 */
		xfs_qm_dqrele(tmp);

		xfs_dqlock(udq);
		xfs_dqlock(gdq);

	} else {
		ASSERT(XFS_DQ_IS_LOCKED(udq));
		if (! locked) {
			xfs_dqlock(gdq);
		}
	}

	ASSERT(XFS_DQ_IS_LOCKED(udq));
	ASSERT(XFS_DQ_IS_LOCKED(gdq));
	/*
	 * Somebody could have attached a gdquot here,
	 * when we dropped the uqlock. If so, just do nothing.
	 */
	if (udq->q_gdquot == NULL) {
		XFS_DQHOLD(gdq);
		udq->q_gdquot = gdq;
	}
	if (! locked) {
		xfs_dqunlock(gdq);
		xfs_dqunlock(udq);
	}
}


/*
 * Given a locked inode, attach dquot(s) to it, taking U/G/P-QUOTAON
 * into account.
 * If XFS_QMOPT_DQALLOC, the dquot(s) will be allocated if needed.
 * If XFS_QMOPT_DQLOCK, the dquot(s) will be returned locked. This option pretty
 * much made this code a complete mess, but it has been pretty useful.
 * If XFS_QMOPT_ILOCKED, then inode sent is already locked EXCL.
 * Inode may get unlocked and relocked in here, and the caller must deal with
 * the consequences.
 */
int
xfs_qm_dqattach(
	xfs_inode_t	*ip,
	uint		flags)
{
	xfs_mount_t	*mp = ip->i_mount;
	uint		nquotas = 0;
	int		error = 0;

	if ((! XFS_IS_QUOTA_ON(mp)) ||
	    (! XFS_NOT_DQATTACHED(mp, ip)) ||
	    (ip->i_ino == mp->m_sb.sb_uquotino) ||
	    (ip->i_ino == mp->m_sb.sb_gquotino))
		return (0);

	ASSERT((flags & XFS_QMOPT_ILOCKED) == 0 ||
	       XFS_ISLOCKED_INODE_EXCL(ip));

	if (! (flags & XFS_QMOPT_ILOCKED))
		xfs_ilock(ip, XFS_ILOCK_EXCL);

	if (XFS_IS_UQUOTA_ON(mp)) {
		error = xfs_qm_dqattach_one(ip, ip->i_d.di_uid, XFS_DQ_USER,
						flags & XFS_QMOPT_DQALLOC,
						flags & XFS_QMOPT_DQLOCK,
						NULL, &ip->i_udquot);
		if (error)
			goto done;
		nquotas++;
	}
	ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
	if (XFS_IS_OQUOTA_ON(mp)) {
		error = XFS_IS_GQUOTA_ON(mp) ?
			xfs_qm_dqattach_one(ip, ip->i_d.di_gid, XFS_DQ_GROUP,
						flags & XFS_QMOPT_DQALLOC,
						flags & XFS_QMOPT_DQLOCK,
						ip->i_udquot, &ip->i_gdquot) :
			xfs_qm_dqattach_one(ip, ip->i_d.di_projid, XFS_DQ_PROJ,
						flags & XFS_QMOPT_DQALLOC,
						flags & XFS_QMOPT_DQLOCK,
						ip->i_udquot, &ip->i_gdquot);
		/*
		 * Don't worry about the udquot that we may have
		 * attached above. It'll get detached, if not already.
		 */
		if (error)
			goto done;
		nquotas++;
	}

	/*
	 * Attach this group quota to the user quota as a hint.
	 * This WON'T, in general, result in a thrash.
	 */
	if (nquotas == 2) {
		ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
		ASSERT(ip->i_udquot);
		ASSERT(ip->i_gdquot);

		/*
		 * We may or may not have the i_udquot locked at this point,
		 * but this check is OK since we don't depend on the i_gdquot to
		 * be accurate 100% all the time. It is just a hint, and this
		 * will succeed in general.
		 */
		if (ip->i_udquot->q_gdquot == ip->i_gdquot)
			goto done;
		/*
		 * Attach i_gdquot to the gdquot hint inside the i_udquot.
		 */
		xfs_qm_dqattach_grouphint(ip->i_udquot, ip->i_gdquot,
					 flags & XFS_QMOPT_DQLOCK);
	}

      done:

#ifdef QUOTADEBUG
	if (! error) {
		if (ip->i_udquot) {
			if (flags & XFS_QMOPT_DQLOCK)
				ASSERT(XFS_DQ_IS_LOCKED(ip->i_udquot));
		}
		if (ip->i_gdquot) {
			if (flags & XFS_QMOPT_DQLOCK)
				ASSERT(XFS_DQ_IS_LOCKED(ip->i_gdquot));
		}
		if (XFS_IS_UQUOTA_ON(mp))
			ASSERT(ip->i_udquot);
		if (XFS_IS_OQUOTA_ON(mp))
			ASSERT(ip->i_gdquot);
	}
#endif

	if (! (flags & XFS_QMOPT_ILOCKED))
		xfs_iunlock(ip, XFS_ILOCK_EXCL);

#ifdef QUOTADEBUG
	else
		ASSERT(XFS_ISLOCKED_INODE_EXCL(ip));
#endif
	return (error);
}

/*
 * Release dquots (and their references) if any.
 * The inode should be locked EXCL except when this's called by
 * xfs_ireclaim.
 */
void
xfs_qm_dqdetach(
	xfs_inode_t	*ip)
{
	if (!(ip->i_udquot || ip->i_gdquot))
		return;

	ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_uquotino);
	ASSERT(ip->i_ino != ip->i_mount->m_sb.sb_gquotino);
	if (ip->i_udquot) {
		xfs_dqtrace_entry_ino(ip->i_udquot, "DQDETTACH", ip);
		xfs_qm_dqrele(ip->i_udquot);
		ip->i_udquot = NULL;
	}
	if (ip->i_gdquot) {
		xfs_dqtrace_entry_ino(ip->i_gdquot, "DQDETTACH", ip);
		xfs_qm_dqrele(ip->i_gdquot);
		ip->i_gdquot = NULL;
	}
}

/*
 * This is called by VFS_SYNC and flags arg determines the caller,
 * and its motives, as done in xfs_sync.
 *
 * vfs_sync: SYNC_FSDATA|SYNC_ATTR|SYNC_BDFLUSH 0x31
 * syscall sync: SYNC_FSDATA|SYNC_ATTR|SYNC_DELWRI 0x25
 * umountroot : SYNC_WAIT | SYNC_CLOSE | SYNC_ATTR | SYNC_FSDATA
 */

int
xfs_qm_sync(
	xfs_mount_t	*mp,
	short		flags)
{
	int		recl, restarts;
	xfs_dquot_t	*dqp;
	uint		flush_flags;
	boolean_t	nowait;
	int		error;

	restarts = 0;
	/*
	 * We won't block unless we are asked to.