NFS: Fix a race when updating an existing write

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 <[e-mail]>
Tested-by: Joe Quanaim <[e-mail]>
Tested-by: Andrew Steffen <[e-mail]>
Reviewed-by: Jeff Layton <[e-mail]>
Fixes: bd37d6fce184 ("NFSv4: Convert nfs_lock_and_join_requests() to use nfs_page_find_head_request()")
Cc: [e-mail]
Signed-off-by: Trond Myklebust <[e-mail]>

fs/nfs/pagelist.c

···
 	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

···
 	}
 }
 
-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;
 }
 
 /**
···
 		}
 	}
 
+	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) {
+		nfs_page_group_unlock(head);
 		nfs_unlock_and_release_request(head);
 		goto retry;
 	}
 
-	ret = nfs_cancel_remove_inode(head, inode);
-	if (ret < 0)
-		goto out_unlock;
-
-	ret = nfs_page_group_lock(head);
-	if (ret < 0)
-		goto out_unlock;
+	nfs_cancel_remove_inode(head, inode);
 
 	/* lock each request in the page group */
 	for (subreq = head->wb_this_page;
···
 {
 	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->mapping;
 
···
 		}
 		spin_unlock(&mapping->i_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

···
 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 *);