[PATCH 4.17 095/220] IB/uverbs: Fix ordering of ucontext check in ib_uverbs_write

[Greg Kroah-Hartman]

4.17-stable review patch.  If anyone has any objections, please let me know.

------------------

From: Jason Gunthorpe <jgg@xxxxxxxxxxxx>

commit 1eb9364ce81d9445ad6f9d44921a91d2a6597156 upstream.

During disassociation the ucontext will become NULL, however due to how
the SRCU locking works the ucontext must only be examined after looking
at the ib_dev, which governs the RCU control flow.

With the wrong ordering userspace will see EINVAL instead of EIO for a
disassociated uverbs FD, which breaks rdma-core.

Cc: stable@xxxxxxxxxxxxxxx
Fixes: 491d5c6a3023 ("RDMA/uverbs: Move uncontext check before SRCU read lock")
Reported-by: Mark Bloch <markb@xxxxxxxxxxxx>
Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxxxx>
Reviewed-by: Leon Romanovsky <leonro@xxxxxxxxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>

---
 drivers/infiniband/core/uverbs_main.c |   14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

--- a/drivers/infiniband/core/uverbs_main.c
+++ b/drivers/infiniband/core/uverbs_main.c
@@ -734,10 +734,6 @@ static ssize_t ib_uverbs_write(struct fi
 	if (ret)
 		return ret;
 
-	if (!file->ucontext &&
-	    (command != IB_USER_VERBS_CMD_GET_CONTEXT || extended))
-		return -EINVAL;
-
 	if (extended) {
 		if (count < (sizeof(hdr) + sizeof(ex_hdr)))
 			return -EINVAL;
@@ -757,6 +753,16 @@ static ssize_t ib_uverbs_write(struct fi
 		goto out;
 	}
 
+	/*
+	 * Must be after the ib_dev check, as once the RCU clears ib_dev ==
+	 * NULL means ucontext == NULL
+	 */
+	if (!file->ucontext &&
+	    (command != IB_USER_VERBS_CMD_GET_CONTEXT || extended)) {
+		ret = -EINVAL;
+		goto out;
+	}
+
 	if (!verify_command_mask(ib_dev, command, extended)) {
 		ret = -EOPNOTSUPP;
 		goto out;