map.c

早期freebsd实现

am_node *find_ap(dir)
char *dir;
{
	int i;

	for (i = last_used_map; i >= 0; --i) {
		am_node *mp = exported_ap[i];
		if (mp && (mp->am_flags & AMF_ROOT)) {
			mp = find_ap2(dir, exported_ap[i]);
			if (mp)
				return mp;
		}
	}

	return 0;
}

/*
 * Find the mount node corresponding
 * to the mntfs structure.
 */
am_node *find_mf P((mntfs *mf));
am_node *find_mf(mf)
mntfs *mf;
{
	int i;

	for (i = last_used_map; i >= 0; --i) {
		am_node *mp = exported_ap[i];
		if (mp && mp->am_mnt == mf)
			return mp;
	}
	return 0;
}

/*
 * Get the filehandle for a particular named directory.
 * This is used during the bootstrap to tell the kernel
 * the filehandles of the initial automount points.
 */
nfs_fh *root_fh(dir)
char *dir;
{
	static nfs_fh nfh;
	am_node *mp = root_ap(dir, TRUE);
	if (mp) {
		mp_to_fh(mp, &nfh);
		/*
		 * Patch up PID to match main server...
		 */
		if (!foreground) {
			long pid = getppid();
			((struct am_fh *) &nfh)->fhh_pid = pid;
#ifdef DEBUG
			dlog("root_fh substitutes pid %d", pid);
#endif
		}
		return &nfh;
	}

	/*
	 * Should never get here...
	 */
	plog(XLOG_ERROR, "Can't find root filehandle for %s", dir);
	return 0;
}

am_node *root_ap(dir, path)
char *dir;
int path;
{
	am_node *mp = find_ap(dir);
	if (mp && mp->am_parent == root_node)
		return mp;

	return 0;
}

/*
 * Timeout all nodes waiting on
 * a given Fserver.
 */
void map_flush_srvr P((fserver *fs));
void map_flush_srvr(fs)
fserver *fs;
{
	int i;
	int done = 0;

	for (i = last_used_map; i >= 0; --i) {
		am_node *mp = exported_ap[i];
		if (mp && mp->am_mnt && mp->am_mnt->mf_server == fs) {
			plog(XLOG_INFO, "Flushed %s; dependent on %s", mp->am_path, fs->fs_host);
			mp->am_ttl = clocktime();
			done = 1;
		}
	}
	if (done)
		reschedule_timeout_mp();
}

/*
 * Mount a top level automount node
 * by calling lookup in the parent
 * (root) node which will cause the
 * automount node to be automounted.
 */
int mount_auto_node P((char *dir, voidp arg));
int mount_auto_node(dir, arg)
char *dir;
voidp arg;
{
	int error = 0;
	(void) afs_ops.lookuppn((am_node *) arg, dir, &error, VLOOK_CREATE);
	if (error > 0) {
		errno = error; /* XXX */
		plog(XLOG_ERROR, "Could not mount %s: %m", dir);
	}
	return error;
}

/*
 * Cause all the top-level mount nodes
 * to be automounted
 */
int mount_exported P((void));
int mount_exported()
{
	/*
	 * Iterate over all the nodes to be started
	 */
	return root_keyiter((void (*)P((char*,void*))) mount_auto_node, root_node);
}

/*
 * Construct top-level node
 */
void make_root_node P((void));
void make_root_node()
{
	mntfs *root_mnt;
	char *rootmap = ROOT_MAP;
	root_node = exported_ap_alloc();

	/*
	 * Allocate a new map
	 */
	init_map(root_node, "");
	/*
	 * Allocate a new mounted filesystem
	 */
	root_mnt = find_mntfs(&root_ops, (am_opts *) 0, "", rootmap, "", "", "");
	/*
	 * Replace the initial null reference
	 */
	free_mntfs(root_node->am_mnt);
	root_node->am_mnt = root_mnt;

	/*
	 * Initialise the root
	 */
	if (root_mnt->mf_ops->fs_init)
		(*root_mnt->mf_ops->fs_init)(root_mnt);

	/*
	 * Mount the root
	 */
	root_mnt->mf_error = (*root_mnt->mf_ops->mount_fs)(root_node);
}

/*
 * Cause all the nodes to be unmounted by timing
 * them out.
 */
void umount_exported(P_void)
{
	int i;
	for (i = last_used_map; i >= 0; --i) {
		am_node *mp = exported_ap[i];
		if (mp) {
			mntfs *mf = mp->am_mnt;
			if (mf->mf_flags & MFF_UNMOUNTING) {
				/*
				 * If this node is being unmounted then
				 * just ignore it.  However, this could
				 * prevent amd from finishing if the
				 * unmount gets blocked since the am_node
				 * will never be free'd.  am_unmounted needs
				 * telling about this possibility. - XXX
				 */
				continue;
			}
			if (mf && !(mf->mf_ops->fs_flags & FS_DIRECTORY)) {
				/*
				 * When shutting down this had better
				 * look like a directory, otherwise it
				 * can't be unmounted!
				 */
				mk_fattr(mp, NFDIR);
			}
			if ((--immediate_abort < 0 && !(mp->am_flags & AMF_ROOT) && mp->am_parent) ||
			    (mf->mf_flags & MFF_RESTART)) {
				/*
				 * Just throw this node away without
				 * bothering to unmount it.  If the
				 * server is not known to be up then
				 * don't discard the mounted on directory
				 * or Amd might hang...
				 */
				if (mf->mf_server &&
					(mf->mf_server->fs_flags & (FSF_DOWN|FSF_VALID)) != FSF_VALID)
					mf->mf_flags &= ~MFF_MKMNT;
				am_unmounted(mp);
			} else {
				/*
				 * Any other node gets forcibly
				 * timed out
				 */
				mp->am_flags &= ~AMF_NOTIMEOUT;
				mp->am_mnt->mf_flags &= ~MFF_RSTKEEP;
				mp->am_ttl = 0;
				mp->am_timeo = 1;
				mp->am_timeo_w = 0;
			}
		}
	}
}

static int unmount_node P((am_node *mp));
static int unmount_node(mp)
am_node *mp;
{
	mntfs *mf = mp->am_mnt;
	int error;

	if ((mf->mf_flags & MFF_ERROR) || mf->mf_refc > 1) {
		/*
		 * Just unlink
		 */
#ifdef DEBUG
		if (mf->mf_flags & MFF_ERROR)
			dlog("No-op unmount of error node %s", mf->mf_info);
#endif /* DEBUG */
		error = 0;
	} else {
#ifdef DEBUG
		dlog("Unmounting %s (%s)", mf->mf_mount, mf->mf_info);
#endif /* DEBUG */
		error = (*mf->mf_ops->umount_fs)(mp);
	}

	if (error) {
#ifdef DEBUG
		errno = error; /* XXX */
		dlog("%s: unmount: %m", mf->mf_mount);
#endif /* DEBUG */
	}

	return error;
}

#ifdef FLUSH_KERNEL_NAME_CACHE
static void flush_kernel_name_cache P((am_node*));
static void flush_kernel_name_cache(mp)
am_node *mp;
{
	int islink = (mp->am_mnt->mf_fattr.type == NFLNK);
	int isdir = (mp->am_mnt->mf_fattr.type == NFDIR);
	int elog = 0;
	if (islink) {
		if (unlink(mp->am_path) < 0)
			elog = 1;
	} else if (isdir) {
		if (rmdir(mp->am_path) < 0)
			elog = 1;
	}
	if (elog)
		plog(XLOG_WARNING, "failed to clear \"%s\" from dnlc: %m", mp->am_path);
}
#endif /* FLUSH_KERNEL_NAME_CACHE */

static int unmount_node_wrap P((voidp vp));
static int unmount_node_wrap(vp)
voidp vp;
{
#ifndef FLUSH_KERNEL_NAME_CACHE
	return unmount_node((am_node*) vp);
#else /* FLUSH_KERNEL_NAME_CACHE */
	/*
	 * This code should just say:
	 * return unmount_node((am_node *) vp);
	 *
	 * However...
	 * The kernel keeps a cached copy of filehandles,
	 * and doesn't ever uncache them (apparently).  So
	 * when Amd times out a node the kernel will have a
	 * stale filehandle.  When the kernel next uses the
	 * filehandle it gets ESTALE.
	 *
	 * The workaround:
	 * Arrange that when a node is removed an unlink or
	 * rmdir is done on that path so that the kernel
	 * cache is done.  Yes - yuck.
	 *
	 * This can all be removed (and the background
	 * unmount flag in sfs_ops) if/when the kernel does
	 * something smarter.
	 *
	 * If the unlink or rmdir failed then just log a warning,
	 * don't fail the unmount.  This can occur if the kernel
	 * client code decides that the object is still referenced
	 * and should be renamed rather than discarded.
	 *
	 * There is still a race condition here...
	 * if another process is trying to access the same
	 * filesystem at the time we get here, then
	 * it will block, since the MF_UNMOUNTING flag will
	 * be set.  That may, or may not, cause the entire
	 * system to deadlock.  Hmmm...
	 */
	am_node *mp = (am_node *) vp;
	int isauto = mp->am_parent && (mp->am_parent->am_mnt->mf_fattr.type == NFDIR);
	int error = unmount_node(mp);
	if (error)
		return error;
	if (isauto && (int)amd_state < (int)Finishing)
		flush_kernel_name_cache(mp);

	return 0;
#endif /* FLUSH_KERNEL_NAME_CACHE */
}

static void free_map_if_success(rc, term, closure)
int rc;
int term;
voidp closure;
{
	am_node *mp = (am_node *) closure;
	mntfs *mf = mp->am_mnt;

	/*
	 * Not unmounting any more
	 */
	mf->mf_flags &= ~MFF_UNMOUNTING;

	/*
	 * If a timeout was defered because the underlying filesystem
	 * was busy then arrange for a timeout as soon as possible.
	 */
	if (mf->mf_flags & MFF_WANTTIMO) {
		mf->mf_flags &= ~MFF_WANTTIMO;
		reschedule_timeout_mp();
	}

	if (term) {
		plog(XLOG_ERROR, "unmount for %s got signal %d", mp->am_path, term);
#if defined(DEBUG) && defined(SIGTRAP)
		/*
		 * dbx likes to put a trap on exit().
		 * Pretend it succeeded for now...
		 */
		if (term == SIGTRAP) {
			am_unmounted(mp);
		}
#endif /* DEBUG */
		amd_stats.d_uerr++;
	} else if (rc) {
		if (rc == EBUSY) {
			plog(XLOG_STATS, "\"%s\" on %s still active", mp->am_path, mf->mf_mount);
		} else {
			errno = rc;	/* XXX */
			plog(XLOG_ERROR, "%s: unmount: %m", mp->am_path);
		}
		amd_stats.d_uerr++;
	} else {
		am_unmounted(mp);
	}

	/*
	 * Wakeup anything waiting for this mount
	 */
	wakeup((voidp) mf);
}

static int unmount_mp(mp)
am_node *mp;
{
	int was_backgrounded = 0;
	mntfs *mf = mp->am_mnt;

#ifdef notdef
	plog(XLOG_INFO, "\"%s\" on %s timed out", mp->am_path, mp->am_mnt->mf_mount);
#endif /* notdef */

	if ((mf->mf_ops->fs_flags & FS_UBACKGROUND) &&
			(mf->mf_flags & MFF_MOUNTED)) {
		if (mf->mf_refc == 1 && !FSRV_ISUP(mf->mf_server)) {
			/*
			 * Don't try to unmount from a server that is known to be down
			 */
			if (!(mf->mf_flags & MFF_LOGDOWN)) {
				/* Only log this once, otherwise gets a bit boring */
				plog(XLOG_STATS, "file server %s is down - timeout of \"%s\" ignored", mf->mf_server->fs_host, mp->am_path);
				mf->mf_flags |= MFF_LOGDOWN;
			}
		} else {
			/* Clear logdown flag - since the server must be up */
			mf->mf_flags &= ~MFF_LOGDOWN;
#ifdef DEBUG
			dlog("\"%s\" on %s timed out", mp->am_path, mp->am_mnt->mf_mount);
			/*dlog("Will background the unmount attempt");*/
#endif /* DEBUG */
			/*
			 * Note that we are unmounting this node
			 */
			mf->mf_flags |= MFF_UNMOUNTING;
			run_task(unmount_node_wrap, (voidp) mp,
				 free_map_if_success, (voidp) mp);
			was_backgrounded = 1;
#ifdef DEBUG
			dlog("unmount attempt backgrounded");
#endif /* DEBUG */
		}
	} else {
#ifdef DEBUG
		dlog("\"%s\" on %s timed out", mp->am_path, mp->am_mnt->mf_mount);
		dlog("Trying unmount in foreground");
#endif
		mf->mf_flags |= MFF_UNMOUNTING;
		free_map_if_success(unmount_node(mp), 0, (voidp) mp);
#ifdef DEBUG
		dlog("unmount attempt done");
#endif /* DEBUG */
	}

	return was_backgrounded;
}

void timeout_mp()
{
#define NEVER (time_t) 0
#define	smallest_t(t1, t2) \
	(t1 != NEVER ? (t2 != NEVER ? (t1 < t2 ? t1 : t2) : t1) : t2)
#define IGNORE_FLAGS (MFF_MOUNTING|MFF_UNMOUNTING|MFF_RESTART)

	int i;
	time_t t = NEVER;
	time_t now = clocktime();
	int backoff = 0;

#ifdef DEBUG
	dlog("Timing out automount points...");
#endif /* DEBUG */
	for (i = last_used_map; i >= 0; --i) {
		am_node *mp = exported_ap[i];
		mntfs *mf;
		/*
		 * Just continue if nothing mounted, or can't be timed out.
		 */
		if (!mp || (mp->am_flags & AMF_NOTIMEOUT))
			continue;
		/*
		 * Pick up mounted filesystem
		 */
		mf = mp->am_mnt;
		if (!mf)
			continue;
		/*
		 * Don't delete last reference to a restarted filesystem.
		 */
		if ((mf->mf_flags & MFF_RSTKEEP) && mf->mf_refc == 1)
			continue;
		/*
		 * If there is action on this filesystem then ignore it
		 */
		if (!(mf->mf_flags & IGNORE_FLAGS)) {
			int expired = 0;
			mf->mf_flags &= ~MFF_WANTTIMO;
#ifdef DEBUG
			/*dlog("t is initially @%d, zero in %d secs", t, t - now);*/
#endif /* DEBUG */
			if (now >= mp->am_ttl) {
				if (!backoff) {
					expired = 1;
					/*
					 * Move the ttl forward to avoid thrashing effects
					 * on the next call to timeout!
					 */
					/* sun's -tw option */
					if (mp->am_timeo_w < 4 * am_timeo_w)
						mp->am_timeo_w += am_timeo_w;
					mp->am_ttl = now + mp->am_timeo_w;
				} else {
					/*
					 * Just backoff this unmount for
					 * a couple of seconds to avoid
					 * many multiple unmounts being
					 * started in parallel.
					 */
					mp->am_ttl = now + backoff + 1;
				}
			}
			/*
			 * If the next ttl is smallest, use that
			 */
			t = smallest_t(t, mp->am_ttl);

#ifdef DEBUG
			/*dlog("after ttl t is @%d, zero in %d secs", t, t - now);*/
#endif /* DEBUG */

			if (!mp->am_child && mf->mf_error >= 0 && expired) {
				/*
				 * If the unmount was backgrounded then
				 * bump the backoff counter.
				 */
				if (unmount_mp(mp)) {
					backoff = 2;
#ifdef DEBUG
					/*dlog("backing off subsequent unmounts by at least %d seconds", backoff);*/
#endif
				}
			}
		} else if (mf->mf_flags & MFF_UNMOUNTING) {
			mf->mf_flags |= MFF_WANTTIMO;
		}
	}

	if (t == NEVER) {
#ifdef DEBUG
		dlog("No further timeouts");
#endif /* DEBUG */
		t = now + ONE_HOUR;
	}

	/*
	 * Sanity check to avoid runaways.
	 * Absolutely should never get this but
	 * if you do without this trap amd will thrash.
	 */
	if (t <= now) {
		t = now + 6;	/* XXX */
		plog(XLOG_ERROR, "Got a zero interval in timeout_mp()!");
	}
	/*
	 * XXX - when shutting down, make things happen faster
	 */
	if ((int)amd_state >= (int)Finishing)
		t = now + 1;
#ifdef DEBUG
	dlog("Next mount timeout in %ds", t - now);
#endif /* DEBUG */

	timeout_mp_id = timeout(t - now, timeout_mp, 0);

#undef NEVER
#undef smallest_t
#undef IGNORE_FLAGS
}

/*
 * Cause timeout_mp to be called soonest
 */
void reschedule_timeout_mp()
{
	if (timeout_mp_id)
		untimeout(timeout_mp_id);
	timeout_mp_id = timeout(0, timeout_mp, 0);
}