@@ -140,6 +140,15 @@ vcs_poll_data_get(struct file *file)
poll->cons_num = console(file_inode(file));
init_waitqueue_head(&poll->waitq);
poll->notifier.notifier_call = vcs_notifier;
+ /*
+ * In order not to lose any update event, we must pretend one might
+ * have occurred before we have a chance to register our notifier.
+ * This is also how user space has come to detect which kernels
+ * support POLLPRI on /dev/vcs* devices i.e. using poll() with
+ * POLLPRI and a zero timeout.
+ */
+ poll->event = VT_UPDATE;
+
if (register_vt_notifier(&poll->notifier) != 0) {
kfree(poll);
return NULL;
Restore and document the forced initial POLLPRI event reporting when poll() is used for the first time. This used to be the implemented behavior before recent changes. Because of the way poll() is implemented, this prevents losing an event happening between the last read() and the first poll() invocation. Since poll() for /dev/vcs* was not always supported, user space probes for its availability as follows: int fd = open("/dev/vcsa", O_RDONLY); struct pollfd p = { .fd = fd, .events = POLLPRI }; available = (poll(&p, 1, 0) == 1); Semantically, it makes sense to signal the first event as such even if it might be spurious. The screen could be modified, and modified back to its initial state before we get to read it, so users must be prepared for that anyway. Signed-off-by: Nicolas Pitre <nico@linaro.org>