Hi Justin, > >> >>> > Subject: bluetooth: Fix serialization when adding/deleting connections in hci_sysfs > >> >>> > > >> >>> > add_conn and del_conn should be serialized, but flush_workqueue can't be used > >> >>> > by the worker thread on it's own queue, so use flush_work to serialize add_conn > >> >>> > and del_conn against each other. > >> >>> > > >> >>> > Signed-off-by: Marc Pignat <marc.pignat@xxxxxxx> > >> >>> > >> >>> Acked-by: Jiri Kosina <jkosina@xxxxxxx> > >> >>> > >> >>> FWIW. > >> >> > >> >>nak from my side since I think it is the wrong fix. We really wanna wait > >> >>for all works to finish here. This includes work from other connection > >> >>attempts or terminations. > >> > > >> > IMHO, there is no need to wait for work currently running, since this is a > >> > singlethread workqueue. > >> > >> Yes, sounds right. > >> > >> > > >> > But it is perhaps simpler to use a lock (mutex or watherver locking primitive). > >> > >> I'm here a little bit late. Marcel, I'm quite busy recently, I just > >> see the commit and then this thread. > >> > >> Let me explain why I add two workqueue originally, because workqueue > >> will be defered, so we must guarantee "connection deletion" finished > >> before "connection adding with same bt addr", or the "connection > >> adding" will fail. > >> > >> On the other hand flush "adding" workqueue in "connection deletion" > >> function is not necessary. > >> > >> To fix this bug, I think we can just use the two work struct for > >> add/del, at the same time keeping the original two workqueue. > >> > >> Please see following patch for this, (building-test only, I have no > >> bluetooth device at hand, I can test this the day after tommorrow) > > > > so I spent the whole day figuring out what is going on here and we keep > > making the wrong assumptions over and over again. > > > > First of all, we only add the sysfs device when we have a successful > > connection. And we identify it with the handle. This means that we can > > NOT have any name clashes anymore since the controller has to make sure > > a handle is only assigned once. Previously we did this on the BD_ADDR > > value and that lead to it. That is no longer the case. > > > > Second of all the two work queues introduces way too much complexity for > > a really simple task of adding and removing a sysfs device entry. > > > > The real problem we have right now are that we are not initializing the > > sysfs device when creating the hci_conn. This is just wrong and can lead > > to all kinds of weird invalid data access. And as a result the adding of > > the sysfs device should only set the name and add it. > > > > We also check device_registered before making sure that device_add has > > been run. And instead of adding more locking or crazy work queue > > dependencies, we should use the single thread work queue to ensure the > > correct order of things. > > > > The attached patch introduces a hci_conn_init_sysfs step to make sure we > > setup the sysfs device correctly. I left the flush_work calls, but I > > think they are not needed since a del_conn before add_conn is no longer > > possible now. > > well it seems your not the only one > with a broken bluetooth(latest git pull > this morning) can you try the patch that I attached to the previous email. It should fix exactly this. Regards Marcel -- To unsubscribe from this list: send the line "unsubscribe linux-bluetooth" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
