Re: [dpdk-dev] [PATCH 2/2] bond: add mode 4 support

[Neil Horman <nhorman@tuxdriver.com>]

On Fri, Sep 19, 2014 at 12:47:35PM +0000, Wodkowski, PawelX wrote:
> > -----Original Message-----
> > From: Neil Horman [mailto:nhorman@tuxdriver.com]
> > Sent: Thursday, September 18, 2014 18:03
> > To: Wodkowski, PawelX
> > Cc: dev@dpdk.org; Jastrzebski, MichalX K; Doherty, Declan
> > Subject: Re: [dpdk-dev] [PATCH 2/2] bond: add mode 4 support
> > 
> > On Thu, Sep 18, 2014 at 08:07:31AM +0000, Wodkowski, PawelX wrote:
> > > > > +int
> > > > > +bond_mode_8023ad_deactivate_slave(struct rte_eth_dev *bond_dev,
> > > > > +		uint8_t slave_pos)
> > > > > +{
> > > > > +	struct bond_dev_private *internals = bond_dev->data->dev_private;
> > > > > +	struct mode8023ad_data *data = &internals->mode4;
> > > > > +	struct port *port;
> > > > > +	uint8_t i;
> > > > > +
> > > > > +	bond_mode_8023ad_stop(bond_dev);
> > > > > +
> > > > > +	/* Exclude slave from transmit policy. If this slave is an aggregator
> > > > > +	 * make all aggregated slaves unselected to force sellection logic
> > > > > +	 * to select suitable aggregator for this port	 */
> > > > > +	for (i = 0; i < internals->active_slave_count; i++) {
> > > > > +		port = &data->port_list[slave_pos];
> > > > > +		if (port->used_agregator_idx == slave_pos) {
> > > > > +			port->selected = UNSELECTED;
> > > > > +			port->actor_state &= ~(STATE_SYNCHRONIZATION |
> > > > STATE_DISTRIBUTING |
> > > > > +				STATE_COLLECTING);
> > > > > +
> > > > > +			/* Use default aggregator */
> > > > > +			port->used_agregator_idx = i;
> > > > > +		}
> > > > > +	}
> > > > > +
> > > > > +	port = &data->port_list[slave_pos];
> > > > > +	timer_cancel(&port->current_while_timer);
> > > > > +	timer_cancel(&port->periodic_timer);
> > > > > +	timer_cancel(&port->wait_while_timer);
> > > > > +	timer_cancel(&port->tx_machine_timer);
> > > > > +
> > > > These all seem rather racy.  Alarm callbacks are executed with the alarm list
> > > > locks not held.  So there is every possibility that you could execute these (or
> > > > any timer_cancel calls in this PMD in parallel with the internal state machine
> > > > timer callback, and leave either with a corrupted timer list (resulting from a
> > > > double free between here, and the actual callback site),
> > >
> > > I don't think so. Yes, callbacks are executed with  alarm list locks not held, but
> > > this is not the issue because access to list itself is guarded by lock and
> > > ap->executing variable. So list will not be trashed. Check source of
> > > eal_alarm_callback(), rte_eal_alarm_set() and rte_eal_alarm_cancel().
> > >
> > Yes, you're right, the list is probably safe wht the executing bit.
> > 
> > > > or a timer that is
> > > > actually still pending when a slave is removed.
> > > >
> > > This is not the issue also, but problem might be similar. I assumed that alarms
> > > are atomic but when I looked at rte alarms closer I saw a race condition
> > > between and rte_eal_alarm_cancel() from  bond_mode_8023ad_stop()
> > > and rte_eal_alarm_set() from state machines callback. This need to be
> > > reworked in some way.
> > 
> > Yes, this is what I was referring to:
> > 
> > CPU0				CPU1
> > rte_eal_alarm_callback		bond_8023ad_deactivate_slave
> > -bond_8023_ad_periodic_cb	timer_cancel
> > timer_set
> > 
> > If those timer functions operate on the same timer, the result is that you can
> > leave the stop/deactivate slave paths with a timer function for that slave still
> > pending. The bonding mode needs some internal state to serialize those
> > operations and determine if the timer should be reactivated.
> > 
> > Neil
> 
> I did rethink the issue and problem is much simpler than it looks like. I did the 
> following:
> 1. Change internal state machine alarms to use rte_rdtsc(). This makes all 
>  mode 4 internal timer_*() function not affected by any race condition.
> 2. Do a busy loop when canceling main callback timer until cancel is successfull.
> This should do the trick about race condition. Do you agree?
> 
I think that will work, but I believe you're making it more complicated (and
less reusable) than it needs to be.  What I think you really need to do is
create a new rte api call, rte_eal_alarm_cancel_sync (something like the
equivalent of del_timer_sync in linux, that wraps up the
while(rte_eal_alarm_cancel(...) == 0) {rte_pause} in its own function (so other
call sites can use it, as I don't think this is an uncommon problem), Then just
create a bonding-internal state flag to signal the periodic callback that it
shouldn't re-arm the timer.  That way all you have to do is set the flag, and
call rte_eal_alarm_cancel_sync, and you're done.  And other applications will be
able to handle this common type of operation as well

Neil