Re: [PATCH] vhost: fix external message handlers

[Maxime Coquelin]

On 3/8/22 09:37, David Marchand wrote:

On Tue, Mar 8, 2022 at 9:34 AM Maxime Coquelin
<maxime.coque...@redhat.com> wrote:

On 3/7/22 19:11, David Marchand wrote:

Following a rework, external message handlers were receiving a pointer
to a vhost_user message (as stated in the API), but lost the ability to
interact with fds attached to the message.
Restore the original layout and put a build check and reminders.

Bugzilla ID: 953
Fixes: 5e0099dc709e ("vhost: remove payload size limitation")

Signed-off-by: David Marchand <david.march...@redhat.com>
---
This patch is untested, but sending quickly to get feedback from the
reporter and comments from author and maintainers.


---
   lib/vhost/vhost_user.c | 8 ++++----
   lib/vhost/vhost_user.h | 7 +++++--
   2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
index 723c6890c3..589b950458 100644
--- a/lib/vhost/vhost_user.c
+++ b/lib/vhost/vhost_user.c
@@ -3023,8 +3023,8 @@ vhost_user_msg_handler(int vid, int fd)

       handled = false;
       if (dev->extern_ops.pre_msg_handle) {
-             ret = (*dev->extern_ops.pre_msg_handle)(dev->vid,
-                             (void *)&ctx.msg);
+             RTE_BUILD_BUG_ON(offsetof(struct vhu_msg_context, msg) != 0);
+             ret = (*dev->extern_ops.pre_msg_handle)(dev->vid, &ctx);
               switch (ret) {
               case RTE_VHOST_MSG_RESULT_REPLY:
                       send_vhost_reply(dev, fd, &ctx);
@@ -3069,8 +3069,8 @@ vhost_user_msg_handler(int vid, int fd)
   skip_to_post_handle:
       if (ret != RTE_VHOST_MSG_RESULT_ERR &&
                       dev->extern_ops.post_msg_handle) {
-             ret = (*dev->extern_ops.post_msg_handle)(dev->vid,
-                             (void *)&ctx.msg);
+             RTE_BUILD_BUG_ON(offsetof(struct vhu_msg_context, msg) != 0);
+             ret = (*dev->extern_ops.post_msg_handle)(dev->vid, &ctx);
               switch (ret) {
               case RTE_VHOST_MSG_RESULT_REPLY:
                       send_vhost_reply(dev, fd, &ctx);
diff --git a/lib/vhost/vhost_user.h b/lib/vhost/vhost_user.h
index be53669f3b..555f89c97a 100644
--- a/lib/vhost/vhost_user.h
+++ b/lib/vhost/vhost_user.h
@@ -152,10 +152,13 @@ typedef struct VhostUserMsg {
       /* Nothing should be added after the payload */
   } __rte_packed VhostUserMsg;

-struct vhu_msg_context {
+/* Note: this structure and VhostUserMsg can't be changed carelessly as
+ * external message handlers rely on them.
+ */
+__rte_packed struct vhu_msg_context {
+     VhostUserMsg msg;
       int fds[VHOST_MEMORY_MAX_NREGIONS];
       int fd_num;
-     VhostUserMsg msg;
   };

   #define VHOST_USER_HDR_SIZE offsetof(VhostUserMsg, payload.u64)

We should revisit the callbacks prototype when we'll be allowed to break
API, passing the message pointer as void * is definitely not a good
idea.

Indeed.

In the mean time, I agree with your fix:

Reviewed-by: Maxime Coquelin <maxime.coque...@redhat.com>

We may also add Reported-by tag when applying.

There is an issue with clang, I'll send a v2.

Ok.
Fan, can you have a try and confirm it fixes the issue on your side?

Thanks,
Maxime