[RHEL7,COMMIT] proc/mounts: add cursor

Submitted by Vasily Averin on March 18, 2021, 8:12 a.m.

Details

Message ID 202103180812.12I8CGDH027998@vz7build.vvs.sw.ru
State New
Series "proc/mounts: add cursor"
Headers show

Commit Message

Vasily Averin March 18, 2021, 8:12 a.m.
The commit is pushed to "branch-rh7-3.10.0-1160.15.2.vz7.173.x-ovz" and will appear at https://src.openvz.org/scm/ovz/vzkernel.git
after rh7-3.10.0-1160.15.2.vz7.173.7
------>
commit 3be1860090531e002a3188a9492f9539f14e6738
Author: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com>
Date:   Thu Mar 18 11:12:16 2021 +0300

    proc/mounts: add cursor
    
    Patch-set description:
    We see a race when reading mountinfo from criu line-by-line. When we
    read mountinfo line-by-line and some mounts which were already printed
    were umounted before printing next mounts we can see some of next mounts
    skipped from output (consequent cat /proc/<pid>/mountinfo shows them).
    This leads to situations where the mount tree is inconsistend without
    those missing mounts and criu fails.
    
    So mountinfo printing based on numeric position in mounts list is racy.
    We can switch to new mainstream approach witch adds mountinfo "cursor"
    to mount namespace mounts list so that even if some mounts are deleted
    from the list, we can still continue reading next mounts after "cursor"
    consistently.
    
    Original patch description:
    From: Miklos Szeredi <mszeredi@redhat.com>
    
    If mounts are deleted after a read(2) call on /proc/self/mounts (or its
    kin), the subsequent read(2) could miss a mount that comes after the
    deleted one in the list.  This is because the file position is interpreted
    as the number mount entries from the start of the list.
    
    E.g. first read gets entries #0 to #9; the seq file index will be 10.  Then
    entry #5 is deleted, resulting in #10 becoming #9 and #11 becoming #10,
    etc...  The next read will continue from entry #10, and #9 is missed.
    
    Solve this by adding a cursor entry for each open instance.  Taking the
    global namespace_sem for write seems excessive, since we are only dealing
    with a per-namespace list.  Instead add a per-namespace spinlock and use
    that together with namespace_sem taken for read to protect against
    concurrent modification of the mount list.  This may reduce parallelism of
    is_local_mountpoint(), but it's hardly a big contention point.  We could
    also use RCU freeing of cursors to make traversal not need additional
    locks, if that turns out to be neceesary.
    
    Only move the cursor once for each read (cursor is not added on open) to
    minimize cacheline invalidation.  When EOF is reached, the cursor is taken
    off the list, in order to prevent an excessive number of cursors due to
    inactive open file descriptors.
    
    Reported-by: Karel Zak <kzak@redhat.com>
    Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
    
    Changes:
    1) make mnt_is_cursor, lock_ns_list, unlock_ns_list externally visible
    2) use it in VZ specific function sync_collect_filesystems to skip
    cursor mounts from walk over mounts of mount namespace.
    
    Note: This patch can be a bit risky because we add empty "cursor" mount
    entries to the mount namespace list. This way we need to be very careful
    when iterating the list (list_for_each_entry or list_first_entry), we
    need to always skip those "cursor" mounts everywhere where we can access
    it's content, or we can get crash. But there are several places where
    mounts are connected with this mnt_list but not connected to namespace
    and it means that they can't be mixed with "cursor" entries, so
    mainstream kernel just relies on this fact and does not explicitly skip
    "cursor" for them.
    
    --
    
    This commit was reverted previously due to
    https://jira.sw.ru/browse/PSBM-126568
    
    Originally, it introduces mounts iterator implementation which is incompatible
    with old seq_file semantics. Corresponding patch from ms
    3a0cfe5da871 ("fs/seq_file.c: simplify seq_file iteration code and interface")
    needs to be ported to overcome this issue. But this patch introduces many
    other unpredictable issues. So, it was easier to adopt current patch to make
    it work fine with old semantics.
    
    https://jira.sw.ru/browse/PSBM-125812
    (cherry-picked from 9f6c61f96f2d97cbb5f7fa85607bc398f843ff0f)
    Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com>
    Reviewed-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
---
 fs/mount.h            |  18 +++++--
 fs/namespace.c        | 144 ++++++++++++++++++++++++++++++++++++++++++++------
 fs/proc_namespace.c   |   6 ++-
 fs/sync.c             |   5 +-
 include/linux/fs.h    |   3 ++
 include/linux/mount.h |   4 +-
 6 files changed, 158 insertions(+), 22 deletions(-)

Patch hide | download patch | download mbox

diff --git a/fs/mount.h b/fs/mount.h
index eab4a7a..b07b393 100644
--- a/fs/mount.h
+++ b/fs/mount.h
@@ -9,7 +9,13 @@  struct mnt_namespace {
 	atomic_t		count;
 	struct ns_common	ns;
 	struct mount *	root;
+	/*
+	 * Traversal and modification of .list is protected by either
+	 * - taking namespace_sem for write, OR
+	 * - taking namespace_sem for read AND taking .ns_lock.
+	 */
 	struct list_head	list;
+	spinlock_t		ns_lock;
 	struct user_namespace	*user_ns;
 	struct ucounts		*ucounts;
 	u64			seq;	/* Sequence number to prevent loops */
@@ -141,15 +147,19 @@  struct proc_mounts {
 	struct mnt_namespace *ns;
 	struct path root;
 	int (*show)(struct seq_file *, struct vfsmount *);
-	void *cached_mount;
-	u64 cached_event;
-	loff_t cached_index;
+	struct mount cursor;
+	struct mount *last_mntpos;
+	loff_t last_pos;
 };
 
 #define proc_mounts(p) (container_of((p), struct proc_mounts, m))
 
 extern const struct seq_operations mounts_op;
 
+extern inline void lock_ns_list(struct mnt_namespace *ns);
+extern inline void unlock_ns_list(struct mnt_namespace *ns);
+extern inline bool mnt_is_cursor(struct mount *mnt);
+
 extern bool __is_local_mountpoint(struct dentry *dentry);
 static inline bool is_local_mountpoint(struct dentry *dentry)
 {
@@ -166,3 +176,5 @@  static inline bool is_anon_ns(struct mnt_namespace *ns)
 {
 	return ns->seq == 0;
 }
+
+extern void mnt_cursor_del(struct mnt_namespace *ns, struct mount *cursor);
diff --git a/fs/namespace.c b/fs/namespace.c
index 2139d38..042ede7 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -744,6 +744,21 @@  struct vfsmount *lookup_mnt(const struct path *path)
 	return m;
 }
 
+inline void lock_ns_list(struct mnt_namespace *ns)
+{
+	spin_lock(&ns->ns_lock);
+}
+
+inline void unlock_ns_list(struct mnt_namespace *ns)
+{
+	spin_unlock(&ns->ns_lock);
+}
+
+inline bool mnt_is_cursor(struct mount *mnt)
+{
+	return mnt->mnt.mnt_flags & MNT_CURSOR;
+}
+
 /*
  * __is_local_mountpoint - Test to see if dentry is a mountpoint in the
  *                         current mount namespace.
@@ -769,11 +784,15 @@  bool __is_local_mountpoint(struct dentry *dentry)
 		goto out;
 
 	down_read(&namespace_sem);
+	lock_ns_list(ns);
 	list_for_each_entry(mnt, &ns->list, mnt_list) {
+		if (mnt_is_cursor(mnt))
+			continue;
 		is_covered = (mnt->mnt_mountpoint == dentry);
 		if (is_covered)
 			break;
 	}
+	unlock_ns_list(ns);
 	up_read(&namespace_sem);
 out:
 	return is_covered;
@@ -1408,47 +1427,125 @@  void replace_mount_options(struct super_block *sb, char *options)
 }
 EXPORT_SYMBOL(replace_mount_options);
 
+struct mount *mnt_list_next(struct mnt_namespace *ns,
+			    struct list_head *p)
+{
+	struct mount *mnt, *ret = NULL;
+
+	lock_ns_list(ns);
+	list_for_each_continue(p, &ns->list) {
+		mnt = list_entry(p, typeof(*mnt), mnt_list);
+		if (!mnt_is_cursor(mnt)) {
+			ret = mnt;
+			break;
+		}
+	}
+	unlock_ns_list(ns);
+
+	return ret;
+}
+EXPORT_SYMBOL(mnt_list_next);
+
 #ifdef CONFIG_PROC_FS
 /* iterator; we want it to have access to namespace_sem, thus here... */
 static void *m_start(struct seq_file *m, loff_t *pos)
 {
 	struct proc_mounts *p = proc_mounts(m);
+	struct list_head *prev;
+	struct mount *mnt;
 
 	down_read(&namespace_sem);
-	if (p->cached_event == p->ns->event) {
-		void *v = p->cached_mount;
-		if (*pos == p->cached_index)
-			return v;
-		if (*pos == p->cached_index + 1) {
-			v = seq_list_next(v, &p->ns->list, &p->cached_index);
-			return p->cached_mount = v;
-		}
+	if (!*pos) {
+		prev = &p->ns->list;
+	} else {
+		prev = &p->cursor.mnt_list;
+
+		/* Read after we'd reached the end? */
+		if (list_empty(prev))
+			return NULL;
 	}
 
-	p->cached_event = p->ns->event;
-	p->cached_mount = seq_list_start(&p->ns->list, *pos);
-	p->cached_index = *pos;
-	return p->cached_mount;
+	mnt = mnt_list_next(p->ns, prev);
+
+	/*
+	 * We should be careful here. seq_file iterator
+	 * semantics was changed in ms kernel since
+	 * 3a0cfe5da871 ("fs/seq_file.c: simplify seq_file
+	 *               iteration code and interface")
+	 *
+	 * Starting from this patch generic seq_file code
+	 * not increments position counter of iterator by self
+	 * (except one place to handle buggy ->next() callbacks).
+	 * So, now when user buffer which was provided for read()
+	 * syscall exhausted seq_file code not increments seq_file
+	 * iterator position, but calls user provided ->next()
+	 * callback which must increment position AND switch current
+	 * mount to the next. Call sequence looks like:
+	 *                   ->next() pos = 1, index = 1
+	 *                   ->show() pos = 1, index = 1
+	 *                   ->next() pos = 2, index = 2
+	 * ->show() <- after that we realized that we should stop because
+	 *             user provided read() buffer is full
+	 *             (notice it's not about seq_file overflow!)
+	 *     OUR KERNEL               |       MS KERNEL
+	 *          pos = 2, index = 2  | ->next() pos = 3, index = 3
+	 * ->stop() pos = 2, index = 2  | ->stop() pos = 3, index = 3
+	 *                      next read() syscall
+	 * ->start() pos = 2, index = 3 | ->start() pos = 3, index = 3
+	 * ->show()  pos = 2, index = 3 | ->show() pos = 3, index = 3
+	 *                   ->next()
+	 * As we can see in both cases index was incremented. But in
+	 * MS kernel it was increment by ->next() callback. In our
+	 * kernel it was incremented manually in seq_file code.
+	 *
+	 * We should take that into account in ->start() implementation
+	 * and detect case when index was changed bypassing ->next() callback.
+	 * Ultra-important here is to detect case when mount in front of which
+	 * mount cursor was is unmounted. In this case we should *not* move forward
+	 * mount position because it is, tecnically, already moved forward :)
+	 *
+	 * After we step on kernels with that patch we should
+	 * follow ms implementation 9f6c61f96f2d ("proc/mounts: add cursor")
+	 * here and just remove following lines and p->last_pos field.
+	 */
+	if (mnt && (*pos == p->last_pos + 1) &&
+	    !(p->last_mntpos && (p->last_mntpos != mnt)))
+		mnt = mnt_list_next(p->ns, &mnt->mnt_list);
+
+	p->last_mntpos = mnt;
+
+	return mnt;
 }
 
 static void *m_next(struct seq_file *m, void *v, loff_t *pos)
 {
 	struct proc_mounts *p = proc_mounts(m);
+	struct mount *mnt = v;
 
-	p->cached_mount = seq_list_next(v, &p->ns->list, pos);
-	p->cached_index = *pos;
-	return p->cached_mount;
+	++*pos;
+	p->last_pos = *pos;
+	p->last_mntpos = mnt_list_next(p->ns, &mnt->mnt_list);
+	return p->last_mntpos;
 }
 
 static void m_stop(struct seq_file *m, void *v)
 {
+	struct proc_mounts *p = proc_mounts(m);
+	struct mount *mnt = v;
+
+	lock_ns_list(p->ns);
+	if (mnt)
+		list_move_tail(&p->cursor.mnt_list, &mnt->mnt_list);
+	else
+		list_del_init(&p->cursor.mnt_list);
+	unlock_ns_list(p->ns);
 	up_read(&namespace_sem);
 }
 
 static int m_show(struct seq_file *m, void *v)
 {
 	struct proc_mounts *p = proc_mounts(m);
-	struct mount *r = list_entry(v, struct mount, mnt_list);
+	struct mount *r = v;
 	return p->show(m, &r->mnt);
 }
 
@@ -1458,6 +1555,15 @@  const struct seq_operations mounts_op = {
 	.stop	= m_stop,
 	.show	= m_show,
 };
+
+void mnt_cursor_del(struct mnt_namespace *ns, struct mount *cursor)
+{
+	down_read(&namespace_sem);
+	lock_ns_list(ns);
+	list_del(&cursor->mnt_list);
+	unlock_ns_list(ns);
+	up_read(&namespace_sem);
+}
 #endif  /* CONFIG_PROC_FS */
 
 /**
@@ -3474,6 +3580,7 @@  static struct mnt_namespace *alloc_mnt_ns(struct user_namespace *user_ns, bool a
 	INIT_LIST_HEAD(&new_ns->list);
 	INIT_LIST_HEAD(&new_ns->mntns_list);
 	init_waitqueue_head(&new_ns->poll);
+	spin_lock_init(&new_ns->ns_lock);
 	new_ns->user_ns = get_user_ns(user_ns);
 	new_ns->ucounts = ucounts;
 	return new_ns;
@@ -3989,10 +4096,14 @@  static bool mnt_already_visible(struct mnt_namespace *ns, struct vfsmount *new,
 	bool visible = false;
 
 	down_read(&namespace_sem);
+	lock_ns_list(ns);
 	list_for_each_entry(mnt, &ns->list, mnt_list) {
 		struct mount *child;
 		int mnt_flags;
 
+		if (mnt_is_cursor(mnt))
+			continue;
+
 		if (mnt->mnt.mnt_sb->s_type != new->mnt_sb->s_type)
 			continue;
 
@@ -4040,6 +4151,7 @@  static bool mnt_already_visible(struct mnt_namespace *ns, struct vfsmount *new,
 	next:	;
 	}
 found:
+	unlock_ns_list(ns);
 	up_read(&namespace_sem);
 	return visible;
 }
diff --git a/fs/proc_namespace.c b/fs/proc_namespace.c
index 3352373..6da491b 100644
--- a/fs/proc_namespace.c
+++ b/fs/proc_namespace.c
@@ -277,7 +277,10 @@  static int mounts_open_common(struct inode *inode, struct file *file,
 	p->root = root;
 	p->m.poll_event = ns->event;
 	p->show = show;
-	p->cached_event = ~0ULL;
+	INIT_LIST_HEAD(&p->cursor.mnt_list);
+	p->cursor.mnt.mnt_flags = MNT_CURSOR;
+	p->last_mntpos = NULL;
+	p->last_pos = 0;
 
 	return 0;
 
@@ -295,6 +298,7 @@  static int mounts_release(struct inode *inode, struct file *file)
 {
 	struct proc_mounts *p = proc_mounts(file->private_data);
 	path_put(&p->root);
+	mnt_cursor_del(p->ns, &p->cursor);
 	put_mnt_ns(p->ns);
 	return seq_release(inode, file);
 }
diff --git a/fs/sync.c b/fs/sync.c
index 07676a8..d994e5b 100644
--- a/fs/sync.c
+++ b/fs/sync.c
@@ -142,7 +142,8 @@  static int sync_collect_filesystems(struct ve_struct *ve, struct list_head *sync
 	BUG_ON(!list_empty(sync_list));
 
 	down_read(&namespace_sem);
-	list_for_each_entry(mnt, &mnt_ns->list, mnt_list) {
+	mnt = mnt_list_next(mnt_ns, &mnt_ns->list);
+	while (mnt) {
 		if (sync_filesystem_collected(sync_list, mnt->mnt.mnt_sb))
 			continue;
 
@@ -161,6 +162,8 @@  static int sync_collect_filesystems(struct ve_struct *ve, struct list_head *sync
 		ss->sb->s_count++;
 		spin_unlock(&sb_lock);
 		list_add_tail(&ss->list, sync_list);
+
+		mnt = mnt_list_next(mnt_ns, &mnt->mnt_list);
 	}
 	up_read(&namespace_sem);
 	return ret;
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 521ff1a..d225c6c 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -53,6 +53,7 @@  struct cred;
 struct swap_info_struct;
 struct seq_file;
 struct workqueue_struct;
+struct mnt_namespace;
 
 extern void __init inode_init(void);
 extern void __init inode_init_early(void);
@@ -3488,6 +3489,8 @@  extern int file_update_time(struct file *file);
 extern int generic_show_options(struct seq_file *m, struct dentry *root);
 extern void save_mount_options(struct super_block *sb, char *options);
 extern void replace_mount_options(struct super_block *sb, char *options);
+extern struct mount *mnt_list_next(struct mnt_namespace *ns,
+				   struct list_head *p);
 
 static inline bool io_is_direct(struct file *filp)
 {
diff --git a/include/linux/mount.h b/include/linux/mount.h
index 8e0352a..21c97ab 100644
--- a/include/linux/mount.h
+++ b/include/linux/mount.h
@@ -48,7 +48,8 @@  struct mnt_namespace;
 #define MNT_ATIME_MASK (MNT_NOATIME | MNT_NODIRATIME | MNT_RELATIME )
 
 #define MNT_INTERNAL_FLAGS (MNT_SHARED | MNT_WRITE_HOLD | MNT_INTERNAL | \
-			    MNT_DOOMED | MNT_SYNC_UMOUNT | MNT_MARKED)
+			    MNT_DOOMED | MNT_SYNC_UMOUNT | MNT_MARKED | \
+			    MNT_CURSOR)
 
 #define MNT_INTERNAL	0x4000
 
@@ -62,6 +63,7 @@  struct mnt_namespace;
 #define MNT_SYNC_UMOUNT		0x2000000
 #define MNT_MARKED		0x4000000
 #define MNT_UMOUNT		0x8000000
+#define MNT_CURSOR		0x10000000
 
 struct vfsmount {
 	struct dentry *mnt_root;	/* root of the mounted tree */