afs: Don't use mutex for I/O operation lock

Don't use the standard mutex for the I/O operation lock, but rather
implement our own as the standard mutex must be released in the same thread
as locked it.  This is a problem when it comes to doing async FetchData
where the lock will be dropped from the workqueue that processed the
incoming data and not from the issuing thread.

Signed-off-by: David Howells <dhowells@redhat.com>
Link: https://lore.kernel.org/r/20241216204124.3752367-12-dhowells@redhat.com
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
Signed-off-by: Christian Brauner <brauner@kernel.org>

fs/afs/fs_operation.c

@@ -49,6 +49,105 @@ struct afs_operation *afs_alloc_operation(struct key *key, struct afs_volume *vo
 	return op;
 }
 
+struct afs_io_locker {
+	struct list_head	link;
+	struct task_struct	*task;
+	unsigned long		have_lock;
+};
+
+/*
+ * Unlock the I/O lock on a vnode.
+ */
+static void afs_unlock_for_io(struct afs_vnode *vnode)
+{
+	struct afs_io_locker *locker;
+
+	spin_lock(&vnode->lock);
+	locker = list_first_entry_or_null(&vnode->io_lock_waiters,
+					  struct afs_io_locker, link);
+	if (locker) {
+		list_del(&locker->link);
+		smp_store_release(&locker->have_lock, 1); /* The unlock barrier. */
+		smp_mb__after_atomic(); /* Store have_lock before task state */
+		wake_up_process(locker->task);
+	} else {
+		clear_bit(AFS_VNODE_IO_LOCK, &vnode->flags);
+	}
+	spin_unlock(&vnode->lock);
+}
+
+/*
+ * Lock the I/O lock on a vnode uninterruptibly.  We can't use an ordinary
+ * mutex as lockdep will complain if we unlock it in the wrong thread.
+ */
+static void afs_lock_for_io(struct afs_vnode *vnode)
+{
+	struct afs_io_locker myself = { .task = current, };
+
+	spin_lock(&vnode->lock);
+
+	if (!test_and_set_bit(AFS_VNODE_IO_LOCK, &vnode->flags)) {
+		spin_unlock(&vnode->lock);
+		return;
+	}
+
+	list_add_tail(&myself.link, &vnode->io_lock_waiters);
+	spin_unlock(&vnode->lock);
+
+	for (;;) {
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		if (smp_load_acquire(&myself.have_lock)) /* The lock barrier */
+			break;
+		schedule();
+	}
+	__set_current_state(TASK_RUNNING);
+}
+
+/*
+ * Lock the I/O lock on a vnode interruptibly.  We can't use an ordinary mutex
+ * as lockdep will complain if we unlock it in the wrong thread.
+ */
+static int afs_lock_for_io_interruptible(struct afs_vnode *vnode)
+{
+	struct afs_io_locker myself = { .task = current, };
+	int ret = 0;
+
+	spin_lock(&vnode->lock);
+
+	if (!test_and_set_bit(AFS_VNODE_IO_LOCK, &vnode->flags)) {
+		spin_unlock(&vnode->lock);
+		return 0;
+	}
+
+	list_add_tail(&myself.link, &vnode->io_lock_waiters);
+	spin_unlock(&vnode->lock);
+
+	for (;;) {
+		set_current_state(TASK_INTERRUPTIBLE);
+		if (smp_load_acquire(&myself.have_lock) || /* The lock barrier */
+		    signal_pending(current))
+			break;
+		schedule();
+	}
+	__set_current_state(TASK_RUNNING);
+
+	/* If we got a signal, try to transfer the lock onto the next
+	 * waiter.
+	 */
+	if (unlikely(signal_pending(current))) {
+		spin_lock(&vnode->lock);
+		if (myself.have_lock) {
+			spin_unlock(&vnode->lock);
+			afs_unlock_for_io(vnode);
+		} else {
+			list_del(&myself.link);
+			spin_unlock(&vnode->lock);
+		}
+		ret = -ERESTARTSYS;
+	}
+	return ret;
+}
+
 /*
  * Lock the vnode(s) being operated upon.
  */
@@ -60,7 +159,7 @@ static bool afs_get_io_locks(struct afs_operation *op)
 	_enter("");
 
 	if (op->flags & AFS_OPERATION_UNINTR) {
-		mutex_lock(&vnode->io_lock);
+		afs_lock_for_io(vnode);
 		op->flags |= AFS_OPERATION_LOCK_0;
 		_leave(" = t [1]");
 		return true;
@@ -72,7 +171,7 @@ static bool afs_get_io_locks(struct afs_operation *op)
 	if (vnode2 > vnode)
 		swap(vnode, vnode2);
 
-	if (mutex_lock_interruptible(&vnode->io_lock) < 0) {
+	if (afs_lock_for_io_interruptible(vnode) < 0) {
 		afs_op_set_error(op, -ERESTARTSYS);
 		op->flags |= AFS_OPERATION_STOP;
 		_leave(" = f [I 0]");
@@ -81,10 +180,10 @@ static bool afs_get_io_locks(struct afs_operation *op)
 	op->flags |= AFS_OPERATION_LOCK_0;
 
 	if (vnode2) {
-		if (mutex_lock_interruptible_nested(&vnode2->io_lock, 1) < 0) {
+		if (afs_lock_for_io_interruptible(vnode2) < 0) {
 			afs_op_set_error(op, -ERESTARTSYS);
 			op->flags |= AFS_OPERATION_STOP;
-			mutex_unlock(&vnode->io_lock);
+			afs_unlock_for_io(vnode);
 			op->flags &= ~AFS_OPERATION_LOCK_0;
 			_leave(" = f [I 1]");
 			return false;
@@ -104,9 +203,9 @@ static void afs_drop_io_locks(struct afs_operation *op)
 	_enter("");
 
 	if (op->flags & AFS_OPERATION_LOCK_1)
-		mutex_unlock(&vnode2->io_lock);
+		afs_unlock_for_io(vnode2);
 	if (op->flags & AFS_OPERATION_LOCK_0)
-		mutex_unlock(&vnode->io_lock);
+		afs_unlock_for_io(vnode);
 }
 
 static void afs_prepare_vnode(struct afs_operation *op, struct afs_vnode_param *vp,

fs/afs/internal.h

@@ -702,13 +702,14 @@ struct afs_vnode {
 	struct afs_file_status	status;		/* AFS status info for this file */
 	afs_dataversion_t	invalid_before;	/* Child dentries are invalid before this */
 	struct afs_permits __rcu *permit_cache;	/* cache of permits so far obtained */
-	struct mutex		io_lock;	/* Lock for serialising I/O on this mutex */
+	struct list_head	io_lock_waiters; /* Threads waiting for the I/O lock */
 	struct rw_semaphore	validate_lock;	/* lock for validating this vnode */
 	struct rw_semaphore	rmdir_lock;	/* Lock for rmdir vs sillyrename */
 	struct key		*silly_key;	/* Silly rename key */
 	spinlock_t		wb_lock;	/* lock for wb_keys */
 	spinlock_t		lock;		/* waitqueue/flags lock */
 	unsigned long		flags;
+#define AFS_VNODE_IO_LOCK	0		/* Set if the I/O serialisation lock is held */
 #define AFS_VNODE_UNSET		1		/* set if vnode attributes not yet set */
 #define AFS_VNODE_DIR_VALID	2		/* Set if dir contents are valid */
 #define AFS_VNODE_ZAP_DATA	3		/* set if vnode's data should be invalidated */

fs/afs/super.c

@@ -663,7 +663,7 @@ static void afs_i_init_once(void *_vnode)
 
 	memset(vnode, 0, sizeof(*vnode));
 	inode_init_once(&vnode->netfs.inode);
-	mutex_init(&vnode->io_lock);
+	INIT_LIST_HEAD(&vnode->io_lock_waiters);
 	init_rwsem(&vnode->validate_lock);
 	spin_lock_init(&vnode->wb_lock);
 	spin_lock_init(&vnode->lock);