If request_irq fails in i8042_open it will call serio_unregister_port,
which takes serio_sem. i8042_open may be called:
serio_register_port - serio_find_dev - dev->connect
serio_register_device - dev->connect
Both serio_register_port and serio_register_device take serio_sem as well.
I think that serio_{register|unregister}_port can be converted into
submitting requests to kseriod thus removing deadlock on the serio_sem.
The patch below is on top of serio* patches in Andrew Morton's -mm tree.
Dmitry
===================================================================
[email protected], 2003-11-30 02:42:54-05:00, [email protected]
Input: Use kseriod to register/unregister serio ports
serio.c | 22 ++++++++++++++--------
1 files changed, 14 insertions(+), 8 deletions(-)
===================================================================
diff -Nru a/drivers/input/serio/serio.c b/drivers/input/serio/serio.c
--- a/drivers/input/serio/serio.c Sun Nov 30 03:01:40 2003
+++ b/drivers/input/serio/serio.c Sun Nov 30 03:01:40 2003
@@ -83,8 +83,10 @@
}
}
-#define SERIO_RESCAN 1
-#define SERIO_RECONNECT 2
+#define SERIO_RESCAN 1
+#define SERIO_RECONNECT 2
+#define SERIO_REGISTER_PORT 3
+#define SERIO_UNREGISTER_PORT 4
static DECLARE_WAIT_QUEUE_HEAD(serio_wait);
static DECLARE_COMPLETION(serio_exited);
@@ -111,6 +113,14 @@
goto event_done;
switch (event->type) {
+ case SERIO_REGISTER_PORT :
+ __serio_register_port(event->serio);
+ break;
+
+ case SERIO_UNREGISTER_PORT :
+ __serio_unregister_port(event->serio);
+ break;
+
case SERIO_RECONNECT :
if (event->serio->dev && event->serio->dev->reconnect)
if (event->serio->dev->reconnect(event->serio) == 0)
@@ -192,9 +202,7 @@
void serio_register_port(struct serio *serio)
{
- down(&serio_sem);
- __serio_register_port(serio);
- up(&serio_sem);
+ serio_queue_event(serio, SERIO_REGISTER_PORT);
}
/*
@@ -210,9 +218,7 @@
void serio_unregister_port(struct serio *serio)
{
- down(&serio_sem);
- __serio_unregister_port(serio);
- up(&serio_sem);
+ serio_queue_event(serio, SERIO_UNREGISTER_PORT);
}
/*
On Sun, Nov 30, 2003 at 03:03:57AM -0500, Dmitry Torokhov wrote:
> If request_irq fails in i8042_open it will call serio_unregister_port,
> which takes serio_sem. i8042_open may be called:
>
> serio_register_port - serio_find_dev - dev->connect
> serio_register_device - dev->connect
>
> Both serio_register_port and serio_register_device take serio_sem as well.
>
> I think that serio_{register|unregister}_port can be converted into
> submitting requests to kseriod thus removing deadlock on the serio_sem.
>
> The patch below is on top of serio* patches in Andrew Morton's -mm tree.
It's nice to avoid the deadlock this way, but I think it's not a good
idea to make the register/unregister asynchronous - it could be a nasty
surprise for an unsuspecting driver writer.
> Dmitry
>
> ===================================================================
>
>
> [email protected], 2003-11-30 02:42:54-05:00, [email protected]
> Input: Use kseriod to register/unregister serio ports
>
>
> serio.c | 22 ++++++++++++++--------
> 1 files changed, 14 insertions(+), 8 deletions(-)
>
> ===================================================================
>
> diff -Nru a/drivers/input/serio/serio.c b/drivers/input/serio/serio.c
> --- a/drivers/input/serio/serio.c Sun Nov 30 03:01:40 2003
> +++ b/drivers/input/serio/serio.c Sun Nov 30 03:01:40 2003
> @@ -83,8 +83,10 @@
> }
> }
>
> -#define SERIO_RESCAN 1
> -#define SERIO_RECONNECT 2
> +#define SERIO_RESCAN 1
> +#define SERIO_RECONNECT 2
> +#define SERIO_REGISTER_PORT 3
> +#define SERIO_UNREGISTER_PORT 4
>
> static DECLARE_WAIT_QUEUE_HEAD(serio_wait);
> static DECLARE_COMPLETION(serio_exited);
> @@ -111,6 +113,14 @@
> goto event_done;
>
> switch (event->type) {
> + case SERIO_REGISTER_PORT :
> + __serio_register_port(event->serio);
> + break;
> +
> + case SERIO_UNREGISTER_PORT :
> + __serio_unregister_port(event->serio);
> + break;
> +
> case SERIO_RECONNECT :
> if (event->serio->dev && event->serio->dev->reconnect)
> if (event->serio->dev->reconnect(event->serio) == 0)
> @@ -192,9 +202,7 @@
>
> void serio_register_port(struct serio *serio)
> {
> - down(&serio_sem);
> - __serio_register_port(serio);
> - up(&serio_sem);
> + serio_queue_event(serio, SERIO_REGISTER_PORT);
> }
>
> /*
> @@ -210,9 +218,7 @@
>
> void serio_unregister_port(struct serio *serio)
> {
> - down(&serio_sem);
> - __serio_unregister_port(serio);
> - up(&serio_sem);
> + serio_queue_event(serio, SERIO_UNREGISTER_PORT);
> }
>
> /*
>
--
Vojtech Pavlik
SuSE Labs, SuSE CR
On Sunday 30 November 2003 04:00 am, Vojtech Pavlik wrote:
> On Sun, Nov 30, 2003 at 03:03:57AM -0500, Dmitry Torokhov wrote:
> > If request_irq fails in i8042_open it will call
> > serio_unregister_port, which takes serio_sem. i8042_open may be
> > called:
> >
> > serio_register_port - serio_find_dev - dev->connect
> > serio_register_device - dev->connect
> >
> > Both serio_register_port and serio_register_device take serio_sem as
> > well.
> >
> > I think that serio_{register|unregister}_port can be converted into
> > submitting requests to kseriod thus removing deadlock on the
> > serio_sem.
> >
> > The patch below is on top of serio* patches in Andrew Morton's -mm
> > tree.
>
> It's nice to avoid the deadlock this way, but I think it's not a good
> idea to make the register/unregister asynchronous - it could be a nasty
> surprise for an unsuspecting driver writer.
>
Serio_register_port is not guaranteed to find a driver for the serio anyway
as the driver can be compiled as a module and loaded much much later so it
should not be a concern. It is somewhat asynchronous already.
Dmitry