/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Copyright (C) 2004 Cluster File Systems, Inc.
- * Author: Phil Schwan <phil@clusterfs.com>
+ * Copyright (C) 2004, 2005 Cluster File Systems, Inc.
+ *
+ * Author: Phil Schwan <phil@clusterfs.com>
+ * Author: Oleg Drokin <green@clusterfs.com>
+ * Author: Yury Umanets <yury@clusterfs.com>
+ * Review: Nikita Danilov <nikita@clusterfs.com>
*
* This file is part of Lustre, http://www.lustre.org.
*
{
struct l_wait_info lwi;
struct ll_sb_info *sbi;
- int rc;
+ int rc = 0;
ENTRY;
LASSERT(dentry != NULL);
LASSERT(!IS_ERR(dentry));
sbi = ll_s2sbi(dentry->d_sb);
+ lwi = LWI_TIMEOUT(timeout * HZ, NULL, NULL);
for (; !d_mountpoint(dentry) && tries > 0; tries--) {
- lwi = LWI_TIMEOUT(timeout * HZ, NULL, NULL);
l_wait_event(sbi->ll_gns_waitq, d_mountpoint(dentry), &lwi);
- }
-
- if ((rc = d_mountpoint(dentry) ? 1 : 0)) {
- spin_lock(&sbi->ll_gns_lock);
- LASSERT(sbi->ll_gns_state == LL_GNS_MOUNTING);
- sbi->ll_gns_state = LL_GNS_FINISHED;
- spin_unlock(&sbi->ll_gns_lock);
- }
-
- complete(&sbi->ll_gns_mount_finished);
- RETURN(rc);
+ if (signal_pending(current))
+ GOTO(out, rc = -EINTR);
+ }
+
+ if (!d_mountpoint(dentry))
+ rc = -ETIME;
+
+ EXIT;
+out:
+ spin_lock(&sbi->ll_gns_lock);
+ sbi->ll_gns_state = LL_GNS_FINISHED;
+ spin_unlock(&sbi->ll_gns_lock);
+ return rc;
}
/*
* tries to mount the mount object under passed @dentry. In the case of success
- * @dentry will become mount point and 0 will be retuned. Error code will be
+ * @dentry will become mount point and 0 will be returned. Error code will be
* returned otherwise.
*/
-int ll_gns_mount_object(struct dentry *dentry,
- struct vfsmount *mnt)
+int
+ll_gns_mount_object(struct dentry *dentry, struct vfsmount *mnt)
{
- struct ll_dentry_data *lld = dentry->d_fsdata;
- char *p, *path, *pathpage, *argv[4];
+ char *path, *pathpage, *datapage, *argv[4];
struct file *mntinfo_fd = NULL;
- struct address_space *mapping;
int cleanup_phase = 0, rc = 0;
struct ll_sb_info *sbi;
struct dentry *dchild;
- struct page *datapage;
- filler_t *filler;
ENTRY;
+ LASSERT(dentry->d_inode != NULL);
+
+ if (!S_ISDIR(dentry->d_inode->i_mode))
+ RETURN(-EINVAL);
+
+ sbi = ll_i2sbi(dentry->d_inode);
+
if (mnt == NULL) {
CERROR("suid directory found, but no "
"vfsmount available.\n");
RETURN(-EINVAL);
}
- CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
-
- LASSERT(dentry->d_inode != NULL);
- LASSERT(S_ISDIR(dentry->d_inode->i_mode));
- LASSERT(lld != NULL);
-
- sbi = ll_i2sbi(dentry->d_inode);
- LASSERT(sbi != NULL);
+ if (atomic_read(&sbi->ll_gns_enabled) == 0)
+ RETURN(-EINVAL);
- /* another thead is in progress of mouning some entry */
spin_lock(&sbi->ll_gns_lock);
- if (sbi->ll_gns_state == LL_GNS_MOUNTING) {
+
+ /*
+ * another thead is in progress or just finished mounting the
+ * dentry. Handling that.
+ */
+ if (sbi->ll_gns_state != LL_GNS_IDLE) {
+ /*
+ * another thread is trying to mount GNS dentry. We'd like to
+ * handling that.
+ */
spin_unlock(&sbi->ll_gns_lock);
- wait_for_completion(&sbi->ll_gns_mount_finished);
+ restart:
+ /*
+ * check if dentry is mount point already, if so, do not restart
+ * syscal.
+ */
if (d_mountpoint(dentry))
RETURN(0);
- }
- /* another thread mounted it already */
- if (sbi->ll_gns_state == LL_GNS_FINISHED) {
+ spin_lock(&sbi->ll_gns_lock);
+ if (sbi->ll_gns_pending_dentry &&
+ is_subdir(sbi->ll_gns_pending_dentry, dentry)) {
+ spin_unlock(&sbi->ll_gns_lock);
+ RETURN(-EAGAIN);
+ }
spin_unlock(&sbi->ll_gns_lock);
- /* we lost a race; just return */
+ /*
+ * waiting for GNS complete and check dentry again, it may be
+ * mounted already.
+ */
+ wait_for_completion(&sbi->ll_gns_mount_finished);
if (d_mountpoint(dentry))
RETURN(0);
+
+ /*
+ * check for he case when there are few waiters and all they are
+ * awakened, but only one will find GNS state LL_GNS_IDLE, and
+ * the rest will face with LL_GNS_MOUNTING. --umka
+ */
+ spin_lock(&sbi->ll_gns_lock);
+ if (sbi->ll_gns_state != LL_GNS_IDLE) {
+ spin_unlock(&sbi->ll_gns_lock);
+ goto restart;
+ }
+ spin_unlock(&sbi->ll_gns_lock);
}
LASSERT(sbi->ll_gns_state == LL_GNS_IDLE);
+ CDEBUG(D_INODE, "mounting dentry %p\n", dentry);
- spin_lock(&dentry->d_lock);
- dentry->d_flags |= DCACHE_GNS_MOUNTING;
- spin_unlock(&dentry->d_lock);
-
/* mounting started */
sbi->ll_gns_state = LL_GNS_MOUNTING;
+ sbi->ll_gns_pending_dentry = dentry;
spin_unlock(&sbi->ll_gns_lock);
/* we need to build an absolute pathname to pass to mount */
GOTO(cleanup, rc = PTR_ERR(dchild));
}
- /* sychronizing with possible /proc/fs/...write */
+ /* synchronizing with possible /proc/fs/...write */
down(&sbi->ll_gns_sem);
/*
* mount object name is taken from sbi, where it is set in mount time or
* via /proc/fs... tunable. It may be ".mntinfo" or so.
*/
- dchild = ll_d_lookup(sbi->ll_gns_oname, dentry,
- strlen(sbi->ll_gns_oname));
+
+ /*
+ * recursive lookup with trying to mount SUID bit marked directories on
+ * the way is not possible here, as lookup_one_len() does not pass @nd
+ * to ->lookup() and this is checked in ll_lookup_it().
+ */
+ dchild = ll_lookup_one_len(sbi->ll_gns_oname, dentry,
+ strlen(sbi->ll_gns_oname));
up(&sbi->ll_gns_sem);
- if (!dchild)
- GOTO(cleanup, rc = -ENOENT);
+ cleanup_phase = 2;
if (IS_ERR(dchild)) {
+ rc = PTR_ERR(dchild);
CERROR("can't find mount object %*s/%*s err = %d.\n",
(int)dentry->d_name.len, dentry->d_name.name,
- (int)dchild->d_name.len, dchild->d_name.name,
- (int)PTR_ERR(dchild));
- GOTO(cleanup, rc = PTR_ERR(dchild));
+ strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
+ rc);
+ GOTO(cleanup, rc);
}
+ /* mount object is not found */
+ if (!dchild->d_inode)
+ GOTO(cleanup, rc = -ENOENT);
+
+ /* check if found child is regular file */
+ if (!S_ISREG(dchild->d_inode->i_mode))
+ GOTO(cleanup, rc = -EBADF);
+
mntget(mnt);
/* ok, mount object if found, opening it. */
if (IS_ERR(mntinfo_fd)) {
CERROR("can't open mount object %*s/%*s err = %d.\n",
(int)dentry->d_name.len, dentry->d_name.name,
- (int)dchild->d_name.len, dchild->d_name.name,
+ strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
(int)PTR_ERR(mntinfo_fd));
- dput(dchild);
mntput(mnt);
GOTO(cleanup, rc = PTR_ERR(mntinfo_fd));
}
- cleanup_phase = 2;
+ cleanup_phase = 3;
- if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE) {
+ if (mntinfo_fd->f_dentry->d_inode->i_size > PAGE_SIZE - 1) {
CERROR("mount object %*s/%*s is too big (%Ld)\n",
(int)dentry->d_name.len, dentry->d_name.name,
- (int)dchild->d_name.len, dchild->d_name.name,
+ strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
mntinfo_fd->f_dentry->d_inode->i_size);
GOTO(cleanup, rc = -EFBIG);
}
+ datapage = (char *)__get_free_page(GFP_KERNEL);
+ if (!datapage)
+ GOTO(cleanup, rc = -ENOMEM);
+
+ cleanup_phase = 4;
+
/* read data from mount object. */
- mapping = mntinfo_fd->f_dentry->d_inode->i_mapping;
- filler = (filler_t *)mapping->a_ops->readpage;
- datapage = read_cache_page(mapping, 0, filler,
- mntinfo_fd);
- if (IS_ERR(datapage)) {
- CERROR("can't read data from mount object %*s/%*s\n",
+ rc = kernel_read(mntinfo_fd, 0, datapage, PAGE_SIZE - 1);
+ if (rc < 0) {
+ CERROR("can't read mount object %*s/%*s data, err %d\n",
(int)dentry->d_name.len, dentry->d_name.name,
- (int)dchild->d_name.len, dchild->d_name.name);
- GOTO(cleanup, rc = PTR_ERR(datapage));
+ strlen(sbi->ll_gns_oname), sbi->ll_gns_oname,
+ rc);
+ GOTO(cleanup, rc);
}
- p = kmap(datapage);
- LASSERT(p != NULL);
- p[PAGE_SIZE - 1] = '\0';
- cleanup_phase = 3;
+ /* no data in mount object? */
+ if (rc == 0) {
+ CERROR("mount object %*s/%*s is empty?\n",
+ (int)dentry->d_name.len, dentry->d_name.name,
+ strlen(sbi->ll_gns_oname), sbi->ll_gns_oname);
+ GOTO(cleanup, rc);
+ }
+ datapage[rc] = '\0';
fput(mntinfo_fd);
mntinfo_fd = NULL;
+ dchild = NULL;
- /* sychronizing with possible /proc/fs/...write */
+ /* synchronizing with possible /proc/fs/...write */
down(&sbi->ll_gns_sem);
/*
* may be /usr/lib/lustre/gns-upcall.sh
*/
argv[0] = sbi->ll_gns_upcall;
- argv[1] = p;
+ argv[1] = datapage;
argv[2] = path;
argv[3] = NULL;
up(&sbi->ll_gns_sem);
- rc = USERMODEHELPER(argv[0], argv, NULL);
+ /* do not wait for helper complete here. */
+ rc = call_usermodehelper(argv[0], argv, NULL, 1);
if (rc) {
- CERROR("failed to call GNS upcall %s, err = %d\n",
- sbi->ll_gns_upcall, rc);
- GOTO(cleanup, rc);
+ CWARN("failed to call GNS upcall %s, err = %d, "
+ "checking for mount anyway\n", sbi->ll_gns_upcall, rc);
}
/*
- * wait for mount completion. This is actually not need, because
- * USERMODEHELPER() returns only when usermode process finishes. But we
- * doing this just for case USERMODEHELPER() semanthics will be changed
- * or usermode upcall program will start mounting in backgound and
- * return instantly. --umka
+ * waiting for dentry become mount point GNS_WAIT_ATTEMPTS times by 1
+ * second.
*/
- if (ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS)) {
+ rc = ll_gns_wait_for_mount(dentry, 1, GNS_WAIT_ATTEMPTS);
+ LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
+
+ /* checking for mount point anyway to not loss mounts */
+ if (d_mountpoint(dentry)) {
struct dentry *rdentry;
struct vfsmount *rmnt;
-
- /* mount is successful */
- LASSERT(sbi->ll_gns_state == LL_GNS_FINISHED);
-
+
rmnt = mntget(mnt);
rdentry = dget(dentry);
mntput(mnt);
dput(dentry);
}
- spin_lock(&dentry->d_lock);
- dentry->d_flags &= ~DCACHE_GNS_PENDING;
- spin_unlock(&dentry->d_lock);
+
+ rc = 0;
} else {
- CERROR("usermode upcall %s failed to mount %s\n",
- sbi->ll_gns_upcall, path);
- rc = -ETIME;
+ CERROR("usermode upcall %s failed to mount %s, err %d\n",
+ sbi->ll_gns_upcall, path, rc);
}
-
+
EXIT;
cleanup:
switch (cleanup_phase) {
+ case 4:
+ free_page((unsigned long)datapage);
case 3:
- kunmap(datapage);
- page_cache_release(datapage);
- case 2:
if (mntinfo_fd != NULL)
fput(mntinfo_fd);
+ case 2:
+ if (dchild != NULL)
+ dput(dchild);
case 1:
free_page((unsigned long)pathpage);
case 0:
spin_lock(&sbi->ll_gns_lock);
sbi->ll_gns_state = LL_GNS_IDLE;
+ sbi->ll_gns_pending_dentry = NULL;
spin_unlock(&sbi->ll_gns_lock);
- spin_lock(&dentry->d_lock);
- dentry->d_flags &= ~DCACHE_GNS_MOUNTING;
- spin_unlock(&dentry->d_lock);
+ /* waking up all waiters after GNS state is LL_GNS_IDLE */
+ complete_all(&sbi->ll_gns_mount_finished);
+ init_completion(&sbi->ll_gns_mount_finished);
}
return rc;
}
}
/*
- * GNS timer callback function. It restarts gns timer and wakes up GNS cvontrol
+ * GNS timer callback function. It restarts gns timer and wakes up GNS control
* thread to process mounts list.
*/
void ll_gns_timer_callback(unsigned long data)
return rc;
}
-/* should we staop GNS control thread? */
+/* should we stop GNS control thread? */
static int inline ll_gns_check_stop(void)
{
mb();
spin_unlock(&gns_lock);
}
- /*
- * letting know stop function know that thread is stoped and it may
- * return.
- */
EXIT;
gns_thread.t_flags = SVC_STOPPED;