NFS: Fix a race when updating an existing write

commit 76d2e3890fb169168c73f2e4f8375c7cc24a765e upstream.

After nfs_lock_and_join_requests() tests for whether the request is
still attached to the mapping, nothing prevents a call to
nfs_inode_remove_request() from succeeding until we actually lock the
page group.
The reason is that whoever called nfs_inode_remove_request() doesn't
necessarily have a lock on the page group head.

So in order to avoid races, let's take the page group lock earlier in
nfs_lock_and_join_requests(), and hold it across the removal of the
request in nfs_inode_remove_request().

Reported-by: Jeff Layton <jlayton@kernel.org>
Tested-by: Joe Quanaim <jdq@meta.com>
Tested-by: Andrew Steffen <aksteffen@meta.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Fixes: bd37d6fce1 ("NFSv4: Convert nfs_lock_and_join_requests() to use nfs_page_find_head_request()")
Cc: stable@vger.kernel.org
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/nfs/pagelist.c

@@ -272,13 +272,14 @@ nfs_page_group_unlock(struct nfs_page *req)
 	nfs_page_clear_headlock(req);
 }
 
-/*
- * nfs_page_group_sync_on_bit_locked
+/**
+ * nfs_page_group_sync_on_bit_locked - Test if all requests have @bit set
+ * @req: request in page group
+ * @bit: PG_* bit that is used to sync page group
  *
  * must be called with page group lock held
  */
-static bool
-nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
+bool nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
 {
 	struct nfs_page *head = req->wb_head;
 	struct nfs_page *tmp;

fs/nfs/write.c

@@ -156,20 +156,10 @@ nfs_page_set_inode_ref(struct nfs_page *req, struct inode *inode)
 	}
 }
 
-static int
-nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
+static void nfs_cancel_remove_inode(struct nfs_page *req, struct inode *inode)
 {
-	int ret;
-
-	if (!test_bit(PG_REMOVE, &req->wb_flags))
-		return 0;
-	ret = nfs_page_group_lock(req);
-	if (ret)
-		return ret;
 	if (test_and_clear_bit(PG_REMOVE, &req->wb_flags))
 		nfs_page_set_inode_ref(req, inode);
-	nfs_page_group_unlock(req);
-	return 0;
 }
 
 static struct nfs_page *nfs_folio_private_request(struct folio *folio)
@@ -238,36 +228,6 @@ static struct nfs_page *nfs_folio_find_head_request(struct folio *folio)
 	return req;
 }
 
-static struct nfs_page *nfs_folio_find_and_lock_request(struct folio *folio)
-{
-	struct inode *inode = folio_file_mapping(folio)->host;
-	struct nfs_page *req, *head;
-	int ret;
-
-	for (;;) {
-		req = nfs_folio_find_head_request(folio);
-		if (!req)
-			return req;
-		head = nfs_page_group_lock_head(req);
-		if (head != req)
-			nfs_release_request(req);
-		if (IS_ERR(head))
-			return head;
-		ret = nfs_cancel_remove_inode(head, inode);
-		if (ret < 0) {
-			nfs_unlock_and_release_request(head);
-			return ERR_PTR(ret);
-		}
-		/* Ensure that nobody removed the request before we locked it */
-		if (head == nfs_folio_private_request(folio))
-			break;
-		if (folio_test_swapcache(folio))
-			break;
-		nfs_unlock_and_release_request(head);
-	}
-	return head;
-}
-
 /* Adjust the file length if we're writing beyond the end */
 static void nfs_grow_file(struct folio *folio, unsigned int offset,
 			  unsigned int count)
@@ -621,20 +581,37 @@ static struct nfs_page *nfs_lock_and_join_requests(struct folio *folio)
 	struct nfs_commit_info cinfo;
 	int ret;
 
-	nfs_init_cinfo_from_inode(&cinfo, inode);
 	/*
 	 * A reference is taken only on the head request which acts as a
 	 * reference to the whole page group - the group will not be destroyed
 	 * until the head reference is released.
 	 */
-	head = nfs_folio_find_and_lock_request(folio);
-	if (IS_ERR_OR_NULL(head))
-		return head;
+retry:
+	head = nfs_folio_find_head_request(folio);
+	if (!head)
+		return NULL;
+
+	while (!nfs_lock_request(head)) {
+		ret = nfs_wait_on_request(head);
+		if (ret < 0) {
+			nfs_release_request(head);
+			return ERR_PTR(ret);
+		}
+	}
 
 	ret = nfs_page_group_lock(head);
 	if (ret < 0)
 		goto out_unlock;
 
+	/* Ensure that nobody removed the request before we locked it */
+	if (head != folio->private && !folio_test_swapcache(folio)) {
+		nfs_page_group_unlock(head);
+		nfs_unlock_and_release_request(head);
+		goto retry;
+	}
+
+	nfs_cancel_remove_inode(head, inode);
+
 	/* lock each request in the page group */
 	for (subreq = head->wb_this_page;
 	     subreq != head;
@@ -855,7 +832,8 @@ static void nfs_inode_remove_request(struct nfs_page *req)
 {
 	struct nfs_inode *nfsi = NFS_I(nfs_page_to_inode(req));
 
-	if (nfs_page_group_sync_on_bit(req, PG_REMOVE)) {
+	nfs_page_group_lock(req);
+	if (nfs_page_group_sync_on_bit_locked(req, PG_REMOVE)) {
 		struct folio *folio = nfs_page_to_folio(req->wb_head);
 		struct address_space *mapping = folio_file_mapping(folio);
 
@@ -867,6 +845,7 @@ static void nfs_inode_remove_request(struct nfs_page *req)
 		}
 		spin_unlock(&mapping->private_lock);
 	}
+	nfs_page_group_unlock(req);
 
 	if (test_and_clear_bit(PG_INODE_REF, &req->wb_flags)) {
 		atomic_long_dec(&nfsi->nrequests);

include/linux/nfs_page.h

@@ -162,6 +162,7 @@ extern void nfs_join_page_group(struct nfs_page *head,
 extern int nfs_page_group_lock(struct nfs_page *);
 extern void nfs_page_group_unlock(struct nfs_page *);
 extern bool nfs_page_group_sync_on_bit(struct nfs_page *, unsigned int);
+extern bool nfs_page_group_sync_on_bit_locked(struct nfs_page *, unsigned int);
 extern	int nfs_page_set_headlock(struct nfs_page *req);
 extern void nfs_page_clear_headlock(struct nfs_page *req);
 extern bool nfs_async_iocounter_wait(struct rpc_task *, struct nfs_lock_context *);