Hi Olivier, > -----Original Message----- > From: Olivier Matz [mailto:olivier.matz at 6wind.com] > Sent: Thursday, June 9, 2016 3:51 PM > To: Lu, Wenzhuo; Stephen Hemminger > Cc: dev at dpdk.org; Tao, Zhe > Subject: Re: [dpdk-dev] [PATCH 2/8] lib/librte_ether: defind RX/TX lock mode > > Hi, > > On 06/08/2016 09:34 AM, Lu, Wenzhuo wrote: > > Hi Stephen, > > > > > >> -----Original Message----- > >> From: Stephen Hemminger [mailto:stephen at networkplumber.org] > >> Sent: Wednesday, June 8, 2016 10:16 AM > >> To: Lu, Wenzhuo > >> Cc: dev at dpdk.org; Tao, Zhe > >> Subject: Re: [dpdk-dev] [PATCH 2/8] lib/librte_ether: defind RX/TX > >> lock mode > >> > >> On Mon, 6 Jun 2016 13:40:47 +0800 > >> Wenzhuo Lu <wenzhuo.lu at intel.com> wrote: > >> > >>> Define lock mode for RX/TX queue. Because when resetting the device > >>> we want the resetting thread to get the lock of the RX/TX queue to > >>> make sure the RX/TX is stopped. > >>> > >>> Using next ABI macro for this ABI change as it has too much impact. > >>> 7 APIs and 1 global variable are impacted. > >>> > >>> Signed-off-by: Wenzhuo Lu <wenzhuo.lu at intel.com> > >>> Signed-off-by: Zhe Tao <zhe.tao at intel.com> > >> > >> Why does this patch set make a different assumption the rest of the DPDK? > >> > >> The rest of the DPDK operates on the principle that the application > >> is smart enough to stop the device before making changes. There is no > >> equivalent to the Linux kernel RTNL mutex. The API assumes > >> application threads are well behaved and will not try and sabotage each > other. > >> > >> If you restrict the reset operation to only being available when > >> RX/TX is stopped, then no lock is needed. > >> > >> The fact that it requires lots more locking inside each device driver > >> implies to me this is not correct way to architect this. > > +1 > > I'm not sure adding locks is the proper way to do. > This is the application responsibility to ensure that: > - control functions are not called concurrently on the same port > - rx/tx functions are not called when the device is stopped/reset/... > > However, I do think the usage paradigms of the ethdev api should be better > documented in rte_ethdev.h (ex: which functions can be called concurrently). > This would be a first step. > > If we really want a helper API to do that in DPDK, the _next_ step could be to > add them in the ethdev api to achieve this. Maybe something like (the function > names could be better): > > - to be called on one control thread: > > rte_eth_stop_rxtx(port) > rte_eth_start_rxtx(port) > > rte_eth_get_rxtx_state(port) > -> return "running" if at least one core is inside the rx/tx code > -> return "stopped" if all cores are outside the rx/tx code > > - to be called on dataplane cores: > > /* same than rte_eth_rx_burst(), but checks if rx/tx is allowed > * first, else do nothing */ > rte_eth_rx_burst_interruptible() > rte_eth_tx_burst_interruptible() > > > The code of control thread could be: > > rte_eth_stop_rxtx(port); > /* wait that all dataplane cores finished their processing */ > while (rte_eth_get_rxtx_state(port) != stopped) > ; > rte_eth_some_control_operation(port); > rte_eth_start_rxtx(port); > > > I think this could be done without any lock, just with the proper memory > barriers > and a per-core status. > > But this API may impose a paradigm to the application, and I'm not sure the > DPDK should do that. I don't quite catch your point. Seems your solution still need the APP to change the code. I think it's more complex than just letting the APP to stop the rx/tx and reset the port. Our purpose of this patch set is to let APP do less as possible. It's not a good choice if we make it more complex. And seems it's hard to stop and start rx/tx in rte layer. Normally APP should do that. To my opinion, we have to introduce lock in rte to achieve that.
> > Regards, > Olivier
