Message-ID: <Y4chryL2YvfhgQmk@Air-de-Roger>
Date:   Wed, 30 Nov 2022 10:26:07 +0100
From:   Roger Pau Monné <roger.pau@...rix.com>
To:     Stefano Stabellini <sstabellini@...nel.org>
Cc:     linux-kernel@...r.kernel.org, xen-devel@...ts.xenproject.org,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        Jiri Slaby <jirislaby@...nel.org>,
        Juergen Gross <jgross@...e.com>,
        Boris Ostrovsky <boris.ostrovsky@...cle.com>,
        Jan Beulich <jbeulich@...e.com>,
        Olof Johansson <olof@...om.net>, Ingo Molnar <mingo@...e.hu>,
        Jeremy Fitzhardinge <jeremy@...source.com>,
        Chris Wright <chrisw@...s-sol.org>,
        linuxppc-dev@...ts.ozlabs.org
Subject: Re: [PATCH] hvc/xen: prevent concurrent accesses to the shared ring

On Tue, Nov 29, 2022 at 02:12:10PM -0800, Stefano Stabellini wrote:
> On Tue, 29 Nov 2022, Roger Pau Monne wrote:
> > The hvc machinery registers both a console and a tty device based on
> > the hv ops provided by the specific implementation.  Those two
> > interfaces however have different locks, and there's no single locks
> > that's shared between the tty and the console implementations, hence
> > the driver needs to protect itself against concurrent accesses.
> > Otherwise concurrent calls using the split interfaces are likely to
> > corrupt the ring indexes, leaving the console unusable.
> >
> > Introduce a lock to xencons_info to serialize accesses to the shared
> > ring.  This is only required when using the shared memory console,
> > concurrent accesses to the hypercall based console implementation are
> > not an issue.
> > 
> > Note the conditional logic in domU_read_console() is slightly modified
> > so the notify_daemon() call can be done outside of the locked region:
> > it's an hypercall and there's no need for it to be done with the lock
> > held.
> > 
> > Fixes: b536b4b96230 ('xen: use the hvc console infrastructure for Xen console')
> > Signed-off-by: Roger Pau Monné <roger.pau@...rix.com>
> > ---
> > While the write handler (domU_write_console()) is used by both the
> > console and the tty ops, that's not the case for the read side
> > (domU_read_console()).  It's not obvious to me whether we could get
> > concurrent poll calls from the poll_get_char tty hook, hence stay on
> > the safe side also serialize read accesses in domU_read_console().
> 
> I think domU_read_console doesn't need it. struct hv_ops and struct
> console are both already locked although independently locked.
> 
> I think we shouldn't add an unrequired lock there.

Not all accesses are done using the tty lock.  There's a path using
tty_find_polling_driver() in kgdboc.c that directly calls into the
->poll_get_char() hook without any locks apparently taken.

> 
> > ---
> >  drivers/tty/hvc/hvc_xen.c | 16 ++++++++++++++--
> >  1 file changed, 14 insertions(+), 2 deletions(-)
> > 
> > diff --git a/drivers/tty/hvc/hvc_xen.c b/drivers/tty/hvc/hvc_xen.c
> > index 7c23112dc923..d65741983837 100644
> > --- a/drivers/tty/hvc/hvc_xen.c
> > +++ b/drivers/tty/hvc/hvc_xen.c
> > @@ -43,6 +43,7 @@ struct xencons_info {
> >  	int irq;
> >  	int vtermno;
> >  	grant_ref_t gntref;
> > +	spinlock_t ring_lock;
> >  };
> >  
> >  static LIST_HEAD(xenconsoles);
> > @@ -84,12 +85,15 @@ static int __write_console(struct xencons_info *xencons,
> >  	XENCONS_RING_IDX cons, prod;
> >  	struct xencons_interface *intf = xencons->intf;
> >  	int sent = 0;
> > +	unsigned long flags;
> >  
> > +	spin_lock_irqsave(&xencons->ring_lock, flags);
> >  	cons = intf->out_cons;
> >  	prod = intf->out_prod;
> >  	mb();			/* update queue values before going on */
> >  
> >  	if ((prod - cons) > sizeof(intf->out)) {
> > +		spin_unlock_irqrestore(&xencons->ring_lock, flags);
> >  		pr_err_once("xencons: Illegal ring page indices");
> >  		return -EINVAL;
> >  	}
> > @@ -99,6 +103,7 @@ static int __write_console(struct xencons_info *xencons,
> >  
> >  	wmb();			/* write ring before updating pointer */
> >  	intf->out_prod = prod;
> > +	spin_unlock_irqrestore(&xencons->ring_lock, flags);
> >  
> >  	if (sent)
> >  		notify_daemon(xencons);
> > @@ -141,16 +146,19 @@ static int domU_read_console(uint32_t vtermno, char *buf, int len)
> >  	int recv = 0;
> >  	struct xencons_info *xencons = vtermno_to_xencons(vtermno);
> >  	unsigned int eoiflag = 0;
> > +	unsigned long flags;
> >  
> >  	if (xencons == NULL)
> >  		return -EINVAL;
> >  	intf = xencons->intf;
> >  
> > +	spin_lock_irqsave(&xencons->ring_lock, flags);
> >  	cons = intf->in_cons;
> >  	prod = intf->in_prod;
> >  	mb();			/* get pointers before reading ring */
> >  
> >  	if ((prod - cons) > sizeof(intf->in)) {
> > +		spin_unlock_irqrestore(&xencons->ring_lock, flags);
> >  		pr_err_once("xencons: Illegal ring page indices");
> >  		return -EINVAL;
> >  	}
> > @@ -174,10 +182,13 @@ static int domU_read_console(uint32_t vtermno, char *buf, int len)
> >  		xencons->out_cons = intf->out_cons;
> >  		xencons->out_cons_same = 0;
> >  	}
> > +	if (!recv && xencons->out_cons_same++ > 1) {
> > +		eoiflag = XEN_EOI_FLAG_SPURIOUS;
> > +	}
> > +	spin_unlock_irqrestore(&xencons->ring_lock, flags);
> > +
> >  	if (recv) {
> >  		notify_daemon(xencons);
> > -	} else if (xencons->out_cons_same++ > 1) {
> > -		eoiflag = XEN_EOI_FLAG_SPURIOUS;
> >  	}
> >  
> >  	xen_irq_lateeoi(xencons->irq, eoiflag);
> > @@ -576,6 +587,7 @@ static int __init xen_hvc_init(void)
> >  
> >  		info = vtermno_to_xencons(HVC_COOKIE);
> >  		info->irq = bind_evtchn_to_irq_lateeoi(info->evtchn);
> > +		spin_lock_init(&info->ring_lock);
> 
> Don't we also need a call to spin_lock_init in xencons_connect_backend
> and xen_cons_init and xenboot_console_setup ?

Not in xencons_connect_backend(), as that's called on resume.  Will
fix the missing lock init, didn't realize the console init paths are
so convoluted.

Early PV console on the shared ring worked fine,  I wonder why that
didn't explode.

Thanks, Roger.

Powered by blists - more mailing lists