af_unix: Revert 'lock_interruptible' in stream receive code

With b3ca9b02b00704053a38bfe4c31dbbb9c13595d0, the AF_UNIX SOCK_STREAM receive code was changed from using mutex_lock(&u->readlock) to mutex_lock_interruptible(&u->readlock) to prevent signals from being delayed for an indefinite time if a thread sleeping on the mutex happened to be selected for handling the signal. But this was never a problem with the stream receive code (as opposed to its datagram counterpart) as that never went to sleep waiting for new messages with the mutex held and thus, wouldn't cause secondary readers to block on the mutex waiting for the sleeping primary reader. As the interruptible locking makes the code more complicated in exchange for no benefit, change it back to using mutex_lock.

---

Considering that the datagram receive routine also doesn't go the sleep with the mutex held anymore, the 37ab4fa7844a044dc21fde45e2a0fc2f3c3b6490 change to unix_autobind is now similarly purposeless.

index 1c3c1f3..b1314c0 100644
@@ -2263,14 +2263,7 @@ static int unix_stream_read_generic(struct unix_stream_read_state *state)
 	/* Lock the socket to prevent queue disordering
 	 * while sleeps in memcpy_tomsg
 	 */
-	err = mutex_lock_interruptible(&u->readlock);
-	if (unlikely(err)) {
-		/* recvmsg() in non blocking mode is supposed to return -EAGAIN
-		 * sk_rcvtimeo is not honored by mutex_lock_interruptible()
-		 */
-		err = noblock ? -EAGAIN : -ERESTARTSYS;
-		goto out;
-	}
+	mutex_lock(&u->readlock);
 
 	if (flags & MSG_PEEK)
 		skip = sk_peek_offset(sk, flags);

@@ -2314,12 +2307,12 @@ again:
 			timeo = unix_stream_data_wait(sk, timeo, last,
 						      last_len);
 
-			if (signal_pending(current) ||
-			    mutex_lock_interruptible(&u->readlock)) {
+			if (signal_pending(current)) {
 				err = sock_intr_errno(timeo);
 				goto out;
 			}
 
+			mutex_lock(&u->readlock);
 			continue;
 unlock:
 			unix_state_unlock(sk);

I wouldn't do this conversion, yet. There is still a deadlock lingering around which should be solved earlier:

http://lists.openwall.net/netdev/2015/11/10/4

Unfortunately I haven't found a good way how to solve it, yet.

Thanks,
Hannes

Judging from the link, that's not related to the stream receive code.

No, but to commit 37ab4fa7844a044dc21fde45e2a0fc2f3c3b6490 where the mutexes of unix_bind and unix_autobind got changed.

The unix_stream_read_generic conversion is fine.

Thanks,
Hannes

Hannes Frederic Sowa <[email protected]> writes:

[...]

[...]

Interesting problem. Assuming the description

(a while ago) A: socketpair()

B: splice() from a pipe to /mnt/regular_file
does sb_start_write() on /mnt

C: try to freeze /mnt
wait for B to finish with /mnt

A: bind() try to bind our socket to /mnt/new_socket_name
lock our socket, see it not bound yet decide that it needs to create something in /mnt try to do sb_start_write() on /mnt, block (it's waiting for C).

D: splice() from the same pipe to our socket
lock the pipe, see that socket is connected try to lock the socket, block waiting for A

B: get around to actually feeding a chunk from
pipe to file, try to lock the pipe.

is correct, the sequence of events could be described as

Given
a/b - acquire a block b (eg, get read lock on superblock
rwsem)

b/a - acquire b block a

c - u->readlock

d - pipe lock

[*y] - blocks waiting for y

B a/b

C b/a[*B]

A c
A a/b[*C]

D d
D c[*A]

B d[*D]

considering that C waits for B, the situation is A blocked by B, D blocked by A, B blocked by D. This could be avoided by making A do the a/b[*C] before acquiring c. D then wouldn't end up blocked waiting for A and hence, B would complete after D completed, enabling C to complete and finally, A. The present unix_mknod is

static int unix_mknod(const char *sun_path, umode_t mode, struct path *res)

{
        struct dentry *dentry;
        struct path path;
        int err = 0;
        /*
         * Get the parent directory, calculate the hash for last
         * component.
         */
        dentry = kern_path_create(AT_FDCWD, sun_path, &path, 0);
        err = PTR_ERR(dentry);
        if (IS_ERR(dentry))
                return err;

        /*
         * All right, let's create it.
         */
        err = security_path_mknod(&path, dentry, mode, 0);
        if (!err) {
                err = vfs_mknod(d_inode(path.dentry), dentry, mode, 0);
                if (!err) {
                        res->mnt = mntget(path.mnt);
                        res->dentry = dget(dentry);
                }
        }
        done_path_create(&path, dentry);
        return err;
}

The a/b[*C] is a side-effect of the kern_path_create. unix_mknod is called with u->readlock held because an already bound socket must not be bound (binded?) again. As far as I understand the above, the actual filesystem manipulation is performed by vfs_mknod. It should be possible to split this function in two so that the sequence of 'bind events' becomes

1. kern_path_create (acquires superblock rw sem)

2. lock u->readlock

3. already bound? yes goto 5

4. create directory entry

5. done_path_create ... / unlock u->readlock

Below is a patch changing the code as described. I've tested that creating sockets with names in the filesystem still works but nothing else (At least not systematically. My 'workstation' didn't blow up in the 21 minutes I've been running the modified kernel on it).

---

index 1c3c1f3..ed3d380 100644
@@ -953,32 +953,30 @@ fail:
 	return NULL;
 }
 
-static int unix_mknod(const char *sun_path, umode_t mode, struct path *res)
+static struct dentry *unix_path_create(const char *sun_path, struct path *path)
 {
-	struct dentry *dentry;
-	struct path path;
-	int err = 0;
 	/*
 	 * Get the parent directory, calculate the hash for last
 	 * component.
 	 */
-	dentry = kern_path_create(AT_FDCWD, sun_path, &path, 0);
-	err = PTR_ERR(dentry);
-	if (IS_ERR(dentry))
-		return err;
 
-	/*
-	 * All right, let's create it.
-	 */
-	err = security_path_mknod(&path, dentry, mode, 0);
+	return kern_path_create(AT_FDCWD, sun_path, path, 0);
+}
+
+static int unix_mknod(struct dentry *dentry, struct path *path, umode_t mode,
+		      struct path *res)
+{
+	int err;
+
+	err = security_path_mknod(path, dentry, mode, 0);
 	if (!err) {
-		err = vfs_mknod(d_inode(path.dentry), dentry, mode, 0);
+		err = vfs_mknod(d_inode(path->dentry), dentry, mode, 0);
 		if (!err) {
-			res->mnt = mntget(path.mnt);
+			res->mnt = mntget(path->mnt);
 			res->dentry = dget(dentry);
 		}
 	}
-	done_path_create(&path, dentry);
+
 	return err;
 }
 

@@ -993,6 +991,8 @@ static int unix_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
 	unsigned int hash;
 	struct unix_address *addr;
 	struct hlist_head *list;
+	struct path parent_path;
+	struct dentry *parent;
 
 	err = -EINVAL;
 	if (sunaddr->sun_family != AF_UNIX)

@@ -1008,9 +1008,18 @@ static int unix_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
 		goto out;
 	addr_len = err;
 
+	parent = NULL;
+	if (sun_path[0]) {
+		parent = unix_path_create(sun_path, &parent_path);
+
+		err = PTR_ERR(parent);
+		if (IS_ERR(parent))
+			goto out;
+	}
+
 	err = mutex_lock_interruptible(&u->readlock);
 	if (err)
-		goto out;
+		goto out_parent;
 
 	err = -EINVAL;
 	if (u->addr)

@@ -1026,11 +1035,11 @@ static int unix_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
 	addr->hash = hash ^ sk->sk_type;
 	atomic_set(&addr->refcnt, 1);
 
-	if (sun_path[0]) {
+	if (parent) {
 		struct path path;
 		umode_t mode = S_IFSOCK |
 		       (SOCK_INODE(sock)->i_mode & ~current_umask());
-		err = unix_mknod(sun_path, mode, &path);
+		err = unix_mknod(parent, &parent_path, mode, &path);
 		if (err) {
 			if (err == -EEXIST)
 				err = -EADDRINUSE;

@@ -1063,6 +1072,10 @@ out_unlock:
 	spin_unlock(&unix_table_lock);
 out_up:
 	mutex_unlock(&u->readlock);
+out_parent:
+	if (parent)
+		done_path_create(&parent_path, parent);
+
 out:
 	return err;
 }

[...]

[from the page]

[...]

Some more explanations on this: There two groups of three in the above (X <- Y supposed to mean 'Y waits for X'), B <- C <- A and A <- D <- B. 'B blocking C blocking A' is really the same as if B was holding an abstract mutex m0 A wants. Likewise, A <- D <- B is equivalent to A holding an abstract mutex m1 B wants. Conceptually, there are two threads and two locks here,

B: acquires m0 then m1
A: acquires m1 then m0

and because of the conflicting locking orders, the whole shoggoth deadlocks sooner or later (Fhtagn!).

The obvious idea to fix this is to reverse either A or B. I think A should be reversed because that's probably easier (unless there's some technical problem with that I don't yet know of) and because this avoids a situation where some other thread which wants the readlock mutex has to wait until some completeld unrelated filesystem operations have completed.

But theory only gets one so far and it would be good if someone capable of reproducing the problem tested this.

From: Rainer Weikusat <[email protected]>
Date: Wed, 16 Dec 2015 20:09:25 +0000

Applied, thanks Rainer.