xfs_filestream.c

linux 内核源代码

	/* Finally, free the memory allocated for the item. */
	kmem_zone_free(item_zone, item);
}

/*
 * xfs_filestream_init() is called at xfs initialisation time to set up the
 * memory zone that will be used for filestream data structure allocation.
 */
int
xfs_filestream_init(void)
{
	item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item");
#ifdef XFS_FILESTREAMS_TRACE
	xfs_filestreams_trace_buf = ktrace_alloc(XFS_FSTRM_KTRACE_SIZE, KM_SLEEP);
#endif
	return item_zone ? 0 : -ENOMEM;
}

/*
 * xfs_filestream_uninit() is called at xfs termination time to destroy the
 * memory zone that was used for filestream data structure allocation.
 */
void
xfs_filestream_uninit(void)
{
#ifdef XFS_FILESTREAMS_TRACE
	ktrace_free(xfs_filestreams_trace_buf);
#endif
	kmem_zone_destroy(item_zone);
}

/*
 * xfs_filestream_mount() is called when a file system is mounted with the
 * filestream option.  It is responsible for allocating the data structures
 * needed to track the new file system's file streams.
 */
int
xfs_filestream_mount(
	xfs_mount_t	*mp)
{
	int		err;
	unsigned int	lifetime, grp_count;

	/*
	 * The filestream timer tunable is currently fixed within the range of
	 * one second to four minutes, with five seconds being the default.  The
	 * group count is somewhat arbitrary, but it'd be nice to adhere to the
	 * timer tunable to within about 10 percent.  This requires at least 10
	 * groups.
	 */
	lifetime  = xfs_fstrm_centisecs * 10;
	grp_count = 10;

	err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count,
	                     xfs_fstrm_free_func);

	return err;
}

/*
 * xfs_filestream_unmount() is called when a file system that was mounted with
 * the filestream option is unmounted.  It drains the data structures created
 * to track the file system's file streams and frees all the memory that was
 * allocated.
 */
void
xfs_filestream_unmount(
	xfs_mount_t	*mp)
{
	xfs_mru_cache_destroy(mp->m_filestream);
}

/*
 * If the mount point's m_perag array is going to be reallocated, all
 * outstanding cache entries must be flushed to avoid accessing reference count
 * addresses that have been freed.  The call to xfs_filestream_flush() must be
 * made inside the block that holds the m_peraglock in write mode to do the
 * reallocation.
 */
void
xfs_filestream_flush(
	xfs_mount_t	*mp)
{
	xfs_mru_cache_flush(mp->m_filestream);
}

/*
 * Return the AG of the filestream the file or directory belongs to, or
 * NULLAGNUMBER otherwise.
 */
xfs_agnumber_t
xfs_filestream_lookup_ag(
	xfs_inode_t	*ip)
{
	xfs_mru_cache_t	*cache;
	fstrm_item_t	*item;
	xfs_agnumber_t	ag;
	int		ref;

	if (!(ip->i_d.di_mode & (S_IFREG | S_IFDIR))) {
		ASSERT(0);
		return NULLAGNUMBER;
	}

	cache = ip->i_mount->m_filestream;
	item = xfs_mru_cache_lookup(cache, ip->i_ino);
	if (!item) {
		TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0);
		return NULLAGNUMBER;
	}

	ASSERT(ip == item->ip);
	ag = item->ag;
	ref = xfs_filestream_peek_ag(ip->i_mount, ag);
	xfs_mru_cache_done(cache);

	TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref);
	return ag;
}

/*
 * xfs_filestream_associate() should only be called to associate a regular file
 * with its parent directory.  Calling it with a child directory isn't
 * appropriate because filestreams don't apply to entire directory hierarchies.
 * Creating a file in a child directory of an existing filestream directory
 * starts a new filestream with its own allocation group association.
 *
 * Returns < 0 on error, 0 if successful association occurred, > 0 if
 * we failed to get an association because of locking issues.
 */
int
xfs_filestream_associate(
	xfs_inode_t	*pip,
	xfs_inode_t	*ip)
{
	xfs_mount_t	*mp;
	xfs_mru_cache_t	*cache;
	fstrm_item_t	*item;
	xfs_agnumber_t	ag, rotorstep, startag;
	int		err = 0;

	ASSERT(pip->i_d.di_mode & S_IFDIR);
	ASSERT(ip->i_d.di_mode & S_IFREG);
	if (!(pip->i_d.di_mode & S_IFDIR) || !(ip->i_d.di_mode & S_IFREG))
		return -EINVAL;

	mp = pip->i_mount;
	cache = mp->m_filestream;
	down_read(&mp->m_peraglock);

	/*
	 * We have a problem, Houston.
	 *
	 * Taking the iolock here violates inode locking order - we already
	 * hold the ilock. Hence if we block getting this lock we may never
	 * wake. Unfortunately, that means if we can't get the lock, we're
	 * screwed in terms of getting a stream association - we can't spin
	 * waiting for the lock because someone else is waiting on the lock we
	 * hold and we cannot drop that as we are in a transaction here.
	 *
	 * Lucky for us, this inversion is rarely a problem because it's a
	 * directory inode that we are trying to lock here and that means the
	 * only place that matters is xfs_sync_inodes() and SYNC_DELWRI is
	 * used. i.e. freeze, remount-ro, quotasync or unmount.
	 *
	 * So, if we can't get the iolock without sleeping then just give up
	 */
	if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL)) {
		up_read(&mp->m_peraglock);
		return 1;
	}

	/* If the parent directory is already in the cache, use its AG. */
	item = xfs_mru_cache_lookup(cache, pip->i_ino);
	if (item) {
		ASSERT(item->ip == pip);
		ag = item->ag;
		xfs_mru_cache_done(cache);

		TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag));
		err = _xfs_filestream_update_ag(ip, pip, ag);

		goto exit;
	}

	/*
	 * Set the starting AG using the rotor for inode32, otherwise
	 * use the directory inode's AG.
	 */
	if (mp->m_flags & XFS_MOUNT_32BITINODES) {
		rotorstep = xfs_rotorstep;
		startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
		mp->m_agfrotor = (mp->m_agfrotor + 1) %
		                 (mp->m_sb.sb_agcount * rotorstep);
	} else
		startag = XFS_INO_TO_AGNO(mp, pip->i_ino);

	/* Pick a new AG for the parent inode starting at startag. */
	err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0);
	if (err || ag == NULLAGNUMBER)
		goto exit_did_pick;

	/* Associate the parent inode with the AG. */
	err = _xfs_filestream_update_ag(pip, NULL, ag);
	if (err)
		goto exit_did_pick;

	/* Associate the file inode with the AG. */
	err = _xfs_filestream_update_ag(ip, pip, ag);
	if (err)
		goto exit_did_pick;

	TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag));

exit_did_pick:
	/*
	 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
	 * reference it took on it, since the file and directory will have taken
	 * their own now if they were successfully cached.
	 */
	if (ag != NULLAGNUMBER)
		xfs_filestream_put_ag(mp, ag);

exit:
	xfs_iunlock(pip, XFS_IOLOCK_EXCL);
	up_read(&mp->m_peraglock);
	return -err;
}

/*
 * Pick a new allocation group for the current file and its file stream.  This
 * function is called by xfs_bmap_filestreams() with the mount point's per-ag
 * lock held.
 */
int
xfs_filestream_new_ag(
	xfs_bmalloca_t	*ap,
	xfs_agnumber_t	*agp)
{
	int		flags, err;
	xfs_inode_t	*ip, *pip = NULL;
	xfs_mount_t	*mp;
	xfs_mru_cache_t	*cache;
	xfs_extlen_t	minlen;
	fstrm_item_t	*dir, *file;
	xfs_agnumber_t	ag = NULLAGNUMBER;

	ip = ap->ip;
	mp = ip->i_mount;
	cache = mp->m_filestream;
	minlen = ap->alen;
	*agp = NULLAGNUMBER;

	/*
	 * Look for the file in the cache, removing it if it's found.  Doing
	 * this allows it to be held across the dir lookup that follows.
	 */
	file = xfs_mru_cache_remove(cache, ip->i_ino);
	if (file) {
		ASSERT(ip == file->ip);

		/* Save the file's parent inode and old AG number for later. */
		pip = file->pip;
		ag = file->ag;

		/* Look for the file's directory in the cache. */
		dir = xfs_mru_cache_lookup(cache, pip->i_ino);
		if (dir) {
			ASSERT(pip == dir->ip);

			/*
			 * If the directory has already moved on to a new AG,
			 * use that AG as the new AG for the file. Don't
			 * forget to twiddle the AG refcounts to match the
			 * movement.
			 */
			if (dir->ag != file->ag) {
				xfs_filestream_put_ag(mp, file->ag);
				xfs_filestream_get_ag(mp, dir->ag);
				*agp = file->ag = dir->ag;
			}

			xfs_mru_cache_done(cache);
		}

		/*
		 * Put the file back in the cache.  If this fails, the free
		 * function needs to be called to tidy up in the same way as if
		 * the item had simply expired from the cache.
		 */
		err = xfs_mru_cache_insert(cache, ip->i_ino, file);
		if (err) {
			xfs_fstrm_free_func(ip->i_ino, file);
			return err;
		}

		/*
		 * If the file's AG was moved to the directory's new AG, there's
		 * nothing more to be done.
		 */
		if (*agp != NULLAGNUMBER) {
			TRACE_MOVEAG(mp, ip, pip,
					ag, xfs_filestream_peek_ag(mp, ag),
					*agp, xfs_filestream_peek_ag(mp, *agp));
			return 0;
		}
	}

	/*
	 * If the file's parent directory is known, take its iolock in exclusive
	 * mode to prevent two sibling files from racing each other to migrate
	 * themselves and their parent to different AGs.
	 */
	if (pip)
		xfs_ilock(pip, XFS_IOLOCK_EXCL);

	/*
	 * A new AG needs to be found for the file.  If the file's parent
	 * directory is also known, it will be moved to the new AG as well to
	 * ensure that files created inside it in future use the new AG.
	 */
	ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount;
	flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
	        (ap->low ? XFS_PICK_LOWSPACE : 0);

	err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen);
	if (err || *agp == NULLAGNUMBER)
		goto exit;

	/*
	 * If the file wasn't found in the file cache, then its parent directory
	 * inode isn't known.  For this to have happened, the file must either
	 * be pre-existing, or it was created long enough ago that its cache
	 * entry has expired.  This isn't the sort of usage that the filestreams
	 * allocator is trying to optimise, so there's no point trying to track
	 * its new AG somehow in the filestream data structures.
	 */
	if (!pip) {
		TRACE_ORPHAN(mp, ip, *agp);
		goto exit;
	}

	/* Associate the parent inode with the AG. */
	err = _xfs_filestream_update_ag(pip, NULL, *agp);
	if (err)
		goto exit;

	/* Associate the file inode with the AG. */
	err = _xfs_filestream_update_ag(ip, pip, *agp);
	if (err)
		goto exit;

	TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0,
			*agp, xfs_filestream_peek_ag(mp, *agp));

exit:
	/*
	 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
	 * reference it took on it, since the file and directory will have taken
	 * their own now if they were successfully cached.
	 */
	if (*agp != NULLAGNUMBER)
		xfs_filestream_put_ag(mp, *agp);
	else
		*agp = 0;

	if (pip)
		xfs_iunlock(pip, XFS_IOLOCK_EXCL);

	return err;
}

/*
 * Remove an association between an inode and a filestream object.
 * Typically this is done on last close of an unlinked file.
 */
void
xfs_filestream_deassociate(
	xfs_inode_t	*ip)
{
	xfs_mru_cache_t	*cache = ip->i_mount->m_filestream;

	xfs_mru_cache_delete(cache, ip->i_ino);
}