qemu-char: don't issue CHR_EVENT_OPEN in a BH

When CHR_EVENT_OPENED was initially added, it was CHR_EVENT_RESET,
and it was issued as a bottom-half:

86e94dea5b740dad65446c857f6959eae43e0ba6

Which we basically used to print out a greeting/prompt for the
monitor.

AFAICT the only reason this was ever done in a BH was because in
some cases we'd modify the chr_write handler for a new chardev
backend *after* the site where we issued the reset (see:
86e94d:qemu_chr_open_stdio())

At some point this event was renamed to CHR_EVENT_OPENED, and we've
maintained the use of this BH ever since.

However, due to 9f939df955a4152aad69a19a77e0898631bb2c18, we schedule
the BH via g_idle_add(), which is causing events to sometimes be
delivered after we've already begun processing data from backends,
leading to:

 known bugs:

  QMP:
    session negotation resets with OPENED event, in some cases this
    is causing new sessions to get sporadically reset

 potential bugs:

  hw/usb/redirect.c:
    can_read handler checks for dev->parser != NULL, which may be
    true if CLOSED BH has not been executed yet. In the past, OPENED
    quiesced outstanding CLOSED events prior to us reading client
    data. If it's delayed, our check may allow reads to occur even
    though we haven't processed the OPENED event yet, and when we
    do finally get the OPENED event, our state may get reset.

  qtest.c:
    can begin session before OPENED event is processed, leading to
    a spurious reset of the system and irq_levels

  gdbstub.c:
    may start a gdb session prior to the machine being paused

To fix these, let's just drop the BH.

Since the initial reasoning for using it still applies to an extent,
work around that by deferring the delivery of CHR_EVENT_OPENED until
after the chardevs have been fully initialized, toward the end of
qmp_chardev_add() (or some cases, qemu_chr_new_from_opts()). This
defers delivery long enough that we can be assured a CharDriverState
is fully initialized before CHR_EVENT_OPENED is sent.

Also, rather than requiring each chardev to do an explicit open, do it
automatically, and allow the small few who don't desire such behavior to
suppress the OPENED-on-init behavior by setting a 'explicit_be_open'
flag.

We additionally add missing OPENED events for stdio backends on w32,
which were previously not being issued, causing us to not recieve the
banner and initial prompts for qmp/hmp.

Reported-by: Stefan Priebe <s.priebe@profihost.ag>
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Message-id: 1370636393-21044-1-git-send-email-mdroth@linux.vnet.ibm.com
Cc: qemu-stable@nongnu.org
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>

2 files changed, 6 insertions, 0 deletions

diff --git a/ui/console.c b/ui/console.c
index b30853f6ef..28bba6de99 100644
--- a/ui/console.c
+++ b/ui/console.c
@@ -1788,6 +1788,10 @@ static CharDriverState *text_console_init(ChardevVC *vc)
     s->chr = chr;
     chr->opaque = s;
     chr->chr_set_echo = text_console_set_echo;
+    /* console/chardev init sometimes completes elsewhere in a 2nd
+     * stage, so defer OPENED events until they are fully initialized
+     */
+    chr->explicit_be_open = true;
 
     if (display_state) {
         text_console_do_init(chr, display_state);
diff --git a/ui/gtk.c b/ui/gtk.c
index efd7836b29..63f6081913 100644
--- a/ui/gtk.c
+++ b/ui/gtk.c
@@ -1119,6 +1119,8 @@ static CharDriverState *gd_vc_handler(ChardevVC *unused)
 
     chr = g_malloc0(sizeof(*chr));
     chr->chr_write = gd_vc_chr_write;
+    /* defer OPENED events until our vc is fully initialized */
+    chr->explicit_be_open = true;
 
     vcs[nb_vcs++] = chr;