On Sun, Dec 18, 2022 at 02:32:09PM +0100, xavie...@mailoo.org wrote:
> I tested, it looks promising:
>
> 0. Rebooted on the new GENERIC.MP built, uname -v
>
> GENERIC.MP#0
>
> 1. Physical plug
>
> ugen0 at uhub0 port 7 "INNO TECH USB to Serial" rev 1.10/0.02 addr 3
>
> 2. Read
>
> tail -f /dev/ugen0.01
>
> 3. Physical unplug
>
> ugen0 detached
>
> 4. Tail reports i/o error as expected
>
> tail: /dev/ugen0.01: Input/output error
> tail: Lost file /dev/ugen0.01: Input/output error
>
> 5. ^C
>
> No panic occurring.
Good, we are on the right track.
Below is a revised patch. Could you replace the initial patch with
this one and test again?
This new patch ensures that klist_invalidate() does get called after
vdevgone(). Making the call from ugen_do_close() might be brittle
if kqueue event registration (ugenkqfilter()) happens to race with
detaching (usbd_is_dying() possibly stops the race, though).
Index: dev/usb/ugen.c
===================================================================
RCS file: src/sys/dev/usb/ugen.c,v
retrieving revision 1.116
diff -u -p -r1.116 ugen.c
--- dev/usb/ugen.c 2 Jul 2022 08:50:42 -0000 1.116
+++ dev/usb/ugen.c 18 Dec 2022 15:48:11 -0000
@@ -798,6 +798,10 @@ ugen_detach(struct device *self, int fla
for (endptno = 0; endptno < USB_MAX_ENDPOINTS; endptno++) {
if (sc->sc_is_open[endptno])
ugen_do_close(sc, endptno, FREAD|FWRITE);
+
+ /* ugenkqfilter() always uses IN. */
+ sce = &sc->sc_endpoints[endptno][IN];
+ klist_invalidate(&sce->rsel.si_note);
}
return (0);
}
