Re: [PATCH] 100% CPU usage on keyboard disconnect

[Steven Newbury <steve@snewbury.org.uk>]

On Thu, 2020-10-15 at 10:12 -0700, Luiz Augusto von Dentz wrote:
> Hi Steven,
> 
> On Thu, Oct 15, 2020 at 4:13 AM Steven Newbury <steve@snewbury.org.uk
> > wrote:
> > There are a couple of issues with g_io_channel usage in bluez which
> > cause CPUs to spin on half-closed channels.
> > 
> > This patch fixes bugs where bluetooth keyboards fail to work on
> > initial
> > connection, and cause 100% cpu on disconnect.
> > 
> > Also fix bug with similar symptoms triggered by some other HID
> > devices
> > such as Sony PS3 BD Remotes.
> > 
> > In the previous discussion on the kernel bugzilla below, it was
> > suggested to remove sec_watch, and I attached a follow-up patch to
> > do
> > so, however that change causes problems with current bluez-5
> > releases
> > where a fd is used after being closed.
> > 
> > See https://bugzilla.kernel.org/show_bug.cgi?id=204275
> > 
> > Signed-off-by: Steven Newbury <steve@snewbury.org.uk>
> > ---
> >  attrib/interactive.c    | 4 +++-
> >  profiles/input/device.c | 3 ++-
> >  2 files changed, 6 insertions(+), 3 deletions(-)
> > 
> > diff --git a/attrib/interactive.c b/attrib/interactive.c
> > index 9a7976d34..453ff064e 100644
> > --- a/attrib/interactive.c
> > +++ b/attrib/interactive.c
> > @@ -64,6 +64,7 @@ static int opt_psm = 0;
> >  static int opt_mtu = 0;
> >  static int start;
> >  static int end;
> > +static guint gsrc;
> > 
> >  static void cmd_help(int argcp, char **argvp);
> > 
> > @@ -193,6 +194,7 @@ static void disconnect_io()
> >         attrib = NULL;
> >         opt_mtu = 0;
> > 
> > +       g_source_remove(gsrc);
> >         g_io_channel_shutdown(iochannel, FALSE, NULL);
> >         g_io_channel_unref(iochannel);
> >         iochannel = NULL;
> > @@ -415,7 +417,7 @@ static void cmd_connect(int argcp, char
> > **argvp)
> >                 error("%s\n", gerr->message);
> >                 g_error_free(gerr);
> >         } else
> > -               g_io_add_watch(iochannel, G_IO_HUP,
> > channel_watcher, NULL);
> > +               gsrc = g_io_add_watch(iochannel, G_IO_HUP,
> > channel_watcher, NULL);
> >  }
> 
> I wouldn't bother with the fix above since the attrib part will be
> going away soon.
> 
> >  static void cmd_disconnect(int argcp, char **argvp)
> > diff --git a/profiles/input/device.c b/profiles/input/device.c
> > index a711ef527..9abf595f6 100644
> > --- a/profiles/input/device.c
> > +++ b/profiles/input/device.c
> > @@ -982,7 +982,8 @@ static int hidp_add_connection(struct
> > input_device *idev)
> >                 }
> > 
> >                 idev->req = req;
> > -               idev->sec_watch = g_io_add_watch(idev->intr_io,
> > G_IO_OUT,
> > +               if (!idev->sec_watch)
> > +                       idev->sec_watch = g_io_add_watch(idev-
> > >intr_io, G_IO_IN,
> >                                                         encrypt_not
> > ify, idev);
> 
> If this is happening isn't there a idev->req already set and we are
> overwriting it?
> 
It was definitely causing a problem, but I can't remember exactly what
occurred, I wrote the patch originally years ago!  I think it leaked
the watch.