Hi, On Fri, 14 Jan 2022 at 13:44, Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote: > > Hi Alexander, > > alex.aring@xxxxxxxxx wrote on Thu, 13 Jan 2022 19:01:56 -0500: > > > Hi, > > > > On Thu, 13 Jan 2022 at 12:07, Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote: > > > > > > Hi Alexander, > > > > > > alex.aring@xxxxxxxxx wrote on Wed, 12 Jan 2022 17:44:02 -0500: > > > > > > > Hi, > > > > > > > > On Wed, 12 Jan 2022 at 12:33, Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote: > > > > ... > > > > > + return 0; > > > > > +} > > > > > diff --git a/net/mac802154/tx.c b/net/mac802154/tx.c > > > > > index c829e4a75325..40656728c624 100644 > > > > > --- a/net/mac802154/tx.c > > > > > +++ b/net/mac802154/tx.c > > > > > @@ -54,6 +54,9 @@ ieee802154_tx(struct ieee802154_local *local, struct sk_buff *skb) > > > > > struct net_device *dev = skb->dev; > > > > > int ret; > > > > > > > > > > + if (unlikely(mac802154_scan_is_ongoing(local))) > > > > > + return NETDEV_TX_BUSY; > > > > > + > > > > > > > > Please look into the functions "ieee802154_wake_queue()" and > > > > "ieee802154_stop_queue()" which prevent this function from being > > > > called. Call stop before starting scanning and wake after scanning is > > > > done or stopped. > > > > > > Mmmh all this is already done, isn't it? > > > - mac802154_trigger_scan_locked() stops the queue before setting the > > > promiscuous mode > > > - mac802154_end_of_scan() wakes the queue after resetting the > > > promiscuous mode to its original state > > > > > > Should I drop the check which stands for an extra precaution? > > > > > > > no, I think then it should be a WARN_ON() more without any return > > (hopefully it will survive). This case should never happen otherwise > > we have a bug that we wake the queue when we "took control about > > transmissions" only. > > Change the name, I think it will be in future not only scan related. > > Maybe "mac802154_queue_stopped()". Everything which is queued from > > socket/upperlayer(6lowpan) goes this way. > > Got it. > > I've changed the name of the helper, and used an atomic variable there > to follow the count. > > > > But overall I think I don't understand well this part. What is > > > a bit foggy to me is why the (async) tx implementation does: > > > > > > *Core* *Driver* > > > > > > stop_queue() > > > drv_async_xmit() ------- > > > \------> do something > > > ------- calls ieee802154_xmit_complete() > > > wakeup_queue() <--------/ > > > > > > So we actually disable the queue for transmitting. Why?? > > > > > > > Because all transceivers have either _one_ transmit framebuffer or one > > framebuffer for transmit and receive one time. We need to report to > > stop giving us more skb's while we are busy with one to transmit. > > This all will/must be changed in future if there is hardware outside > > which is more powerful and the driver needs to control the flow here. > > > > That ieee802154_xmit_complete() calls wakeup_queue need to be > > forbidden when we are in "synchronous transmit mode"/the queue is > > stopped. The synchronous transmit mode is not for any hotpath, it's > > for MLME and I think we also need a per phy lock to avoid multiple > > synchronous transmissions at one time. Please note that I don't think > > here only about scan operation, also for other possible MLME-ops. > > > > First, thank you very much for all your guidance and reviews, I think I > have a much clearer understanding now. > > I've tried to follow your advices, creating: > - a way of tracking ongoing transmissions > - a synchronous API for MLME transfers > Please note that I think we cannot use netif_stop_queue() from context outside of netif xmit() callback. It's because the atomic counter itself is racy in xmit(), we need to be sure xmit() can't occur while stopping the queue. I think maybe "netif_tx_disable()" is the right call to stop from another context, because it holds the tx_lock, which I believe is held while xmit(). Where the wake queue call should be fine to call..., maybe we can remove some EXPORT_SYMBOL() then? I saw that some drivers call "ieee802154_wake_queue()" in error cases, may we introduce a new helper "?ieee802154_xmit_error?" for error cases so you can also catch error cases in your sync tx. See `grep -r "ieee802154_wake_queue" drivers/net/ieee802154`, if we have more information we might add more meaning into the error cases (e.g. proper errno). > I've decided to use the wait_queue + atomic combo which looks nice. > Everything seems to work, I just need a bit of time to clean and rework > a bit the series before sending a v3. > Okay, sounds good to implement both requirements. - Alex
