Dropping the locks prevents a deadlock when running with -sdl or -vnc
in addition to -spice.
When server calls get_cursor_command, and we have an active ds
cursor related callback in non vga mode, we need to lock to prevent
the iothread (via sdl/vnc gui_update timer) from touching the ds as well.
Currently (-sdl/-vnc) + -spice seems to work, due to dropping the locking in
qxl-render.c:qxl_render_cursor, but this is just waiting to break because of
touching the cursor from two threads without any locking.
---
hw/qxl-render.c | 13 +++++++++----
1 files changed, 9 insertions(+), 4 deletions(-)
diff --git a/hw/qxl-render.c b/hw/qxl-render.c
index 58965e0..1065388 100644
--- a/hw/qxl-render.c
+++ b/hw/qxl-render.c
@@ -209,18 +209,23 @@ void qxl_render_cursor(PCIQXLDevice *qxl, QXLCommandExt
*ext)
if (c == NULL) {
c = cursor_builtin_left_ptr();
}
- qemu_mutex_lock_iothread();
+ /* TODO: move this operation to iothread via pipe
+ * we can't use the global lock here without dropping it
+ * in gui_update (vl.c), or we get a dead lock (gui_update
+ * calls dispatcher, waiting on pipe read, and spice server calls
+ * this function, waiting on the lock that iothread is holding).
+ * But when used with sdl this calls sdl.c:sdl_mouse_define, which
+ * afaict must be locked or called from iothread. Moving to iothread
+ * seems easiest from correctness pov. */
qxl->ssd.ds->cursor_define(c);
qxl->ssd.ds->mouse_set(x, y, 1);
- qemu_mutex_unlock_iothread();
cursor_put(c);
break;
case QXL_CURSOR_MOVE:
x = cmd->u.position.x;
y = cmd->u.position.y;
- qemu_mutex_lock_iothread();
+ /* TODO: move this operation to iothread via pipe. See comment above */
qxl->ssd.ds->mouse_set(x, y, 1);
- qemu_mutex_unlock_iothread();
break;
}
}