[dpdk-dev] [PATCH v9 1/8] ethdev: use locks to protect Rx/Tx callback lists

[Thomas Monjalon]

2016-06-15 09:54, Ananyev, Konstantin:
> 
> > -----Original Message-----
> > From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com]
> > Sent: Wednesday, June 15, 2016 9:49 AM
> > To: Ananyev, Konstantin
> > Cc: Pattan, Reshma; dev at dpdk.org
> > Subject: Re: [dpdk-dev] [PATCH v9 1/8] ethdev: use locks to protect Rx/Tx callback lists
> > 
> > 2016-06-15 08:37, Ananyev, Konstantin:
> > > From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com]
> > > > 2016-06-15 05:30, Pattan, Reshma:
> > > > > From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com]
> > > > > > 2016-06-14 10:38, Reshma Pattan:
> > > > > > > Added spinlocks around add/remove logic of Rx and Tx callbacks to
> > > > > > > avoid corruption of callback lists in multithreaded context.
> > > > > > >
> > > > > > > Signed-off-by: Reshma Pattan <reshma.pattan at intel.com>
> > > > > >
> > > > > > Why cb->next is not locked in burst functions?
> > > > > It is safe to do "read access" here and doesn't require any locking as rx/tx burst is initiated  by only local user(control plane)
> > thread.
> > > > >
> > > > > > Just protecting add/remove but not its usage seems useless.
> > > > > Here locks were required  around add/remove to protect "write access"  because write to callback list is now done from 2
> > threads
> > > > > i.e. one from local user thread(control plane) and another from pdump control thread(initiated by remote pdump request).
> > > >
> > > > So read and write can be done by different threads.
> > >
> > > Yes, and this is possible even in current DPDK version (16.04).
> > > What is added by Reshma's patch - now it is possible to have concurrent write
> > > from 2 different thread to that list.
> > >
> > > > I think the read access would need locking but we do not want it
> > > > in fast path.
> > >
> > > I don't think it would be needed.
> > > As I said - read/write interaction didn't change from what we have right now.
> > > But if you have some particular scenario in mind that you believe would cause
> > > a race condition - please speak up.
> > 
> > If we add/remove a callback during a burst? Is it possible that the next
> > pointer would have a wrong value leading to a crash?
> > Maybe we need a comment to state that we should not alter burst
> > callbacks while running burst functions.
> 
> Current status (16.04):
> It is safe to add/remove RX/TX callbacks while 
> another thread is doing simultaneously RX/TX burst over same queue.
> I.E: it is supposed to be safe to invoke
> rte_eth_add(/remove)_rx(/tx)_callback() and rte_eth_rx_burst()/rte_eth_tx_burst()
> from different threads simultaneously.
> Though it is not safe to free/modify that rte_eth_rxtx_callback while current
> rte_eth_rx_burst()/rte_eth_tx_burst() are still active.
> That exactly what comments for rte_eth_remove_rx_callback() say:
> 
> * Note: the callback is removed from the callback list but it isn't freed
>  * since the it may still be in use. The memory for the callback can be
>  * subsequently freed back by the application by calling rte_free():
>  *
>  * - Immediately - if the port is stopped, or the user knows that no
>  *   callbacks are in flight e.g. if called from the thread doing RX/TX
>  *   on that queue.
>  *
>  * - After a short delay - where the delay is sufficient to allow any
>  *   in-flight callbacks to complete.
> 
> In other words, right now there only way to know for sure that it is safe
> to free the removed callback - is to stop the port.
> 
> Does it need to be changed, so when rte_eth_remove_rx_callback() returns
> user can safely free the callback (or even better rte_eth_remove_rx_callback free the callback for us)?
> In my opinion - yes.
> Though, I think, it has nothing to do with pdump patches, and I think should be a matter
> for separate a patch/discussion.
> 
> Now with pdump library introduction - there is possibility that 2 different threads
> can try to  add/remove callbacks for the same queue simultaneously.
> First one - thread executing control requests from local user,
> second  one - pdump control thread executing pdump requests from pdump client.
> That lock is introduced to avoid race condition between such 2 threads:
> i.e. to prevent multiple threads to modify same list simultaneously.   
> It is not intended to synchronise read/write accesses to the list, see above. 

OK thanks for the explanations