[Libguestfs] [libnbd PATCH v3 06/22] states: Break deadlock if server goofs on extended replies

[Eric Blake]

One of the benefits of extended replies is that we can do a
fixed-length read for the entire header of every server reply, which
is fewer syscalls than the split-read approach required by structured
replies.  But one of the drawbacks of doing a large read is that if
the server is non-compliant (not a problem for normal servers, but
something I hit rather more than I'd like to admit while developing
extended header support in servers), nbd_pwrite() and friends will
deadlock if the server replies with the wrong header.  Add in some
code to catch that failure mode and move the state machine to DEAD
sooner, to make it easier to diagnose the fault in the server.

Unlike in the case of an unexpected simply reply from a structured
server (where we never over-read, and can therefore commit b31e7bac
can merely fail the command with EPROTO and successfully move on to
the next server reply), in this case we really do have to move to
DEAD: in addition to having already read the 16 or 20 bytes that the
server sent in its (short) reply for this command, we may have already
read the initial bytes of the server's next reply, but we have no way
to push those extra bytes back onto our read stream for parsing on our
next pass through the state machine.

Signed-off-by: Eric Blake <eblake at redhat.com>
---
 generator/states-reply.c | 23 ++++++++++++++++++++++-
 1 file changed, 22 insertions(+), 1 deletion(-)

diff --git a/generator/states-reply.c b/generator/states-reply.c
index 4e9f2dde..d4710d91 100644
--- a/generator/states-reply.c
+++ b/generator/states-reply.c
@@ -109,7 +109,28 @@  REPLY.START:
  REPLY.RECV_REPLY:
   switch (recv_into_rbuf (h)) {
   case -1: SET_NEXT_STATE (%.DEAD); return 0;
-  case 1: SET_NEXT_STATE (%.READY); return 0;
+  case 1: SET_NEXT_STATE (%.READY);
+    /* Special case: if we have a short read, but got at least far
+     * enough to decode the magic number, we can check if the server
+     * is matching our expectations. This lets us avoid deadlocking if
+     * a buggy server sends only 16 bytes of a simple reply, and is
+     * waiting for our next command, while we are blocked waiting for
+     * the server to send 32 bytes of an extended reply.
+     */
+    if (h->extended_headers &&
+        (char *)h->rbuf >= (char *)&h->sbuf.reply.hdr.extended.flags) {
+      uint32_t magic = be32toh (h->sbuf.reply.hdr.extended.magic);
+      if (magic != NBD_EXTENDED_REPLY_MAGIC) {
+        SET_NEXT_STATE (%.DEAD); /* We've probably lost synchronization. */
+        set_error (0, "invalid or unexpected reply magic 0x%" PRIx32, magic);
+#if 0 /* uncomment to see desynchronized data */
+        nbd_internal_hexdump (&h->sbuf.reply.hdr.extended.flags,
+                              sizeof (h->sbuf.reply.hdr.extended.flags),
+                              stderr);
+#endif
+      }
+    }
+    return 0;
   case 0: SET_NEXT_STATE (%CHECK_SIMPLE_OR_STRUCTURED_REPLY);
   }
   return 0;
-- 
2.40.1