Eric, Thank you for your patient apply. There is still a question, In receive path, driver does not call dev_hold(), when skb goes to host stack, skb->dev is likely to be used. If device is destroyed before that, it seems dangerous.
Thank you! Zhangjie -----Original Message----- From: Eric Dumazet [mailto:eric.duma...@gmail.com] Sent: Tuesday, August 25, 2015 10:25 PM To: Zhangjie (HZ) Cc: Jason Wang; netdev@vger.kernel.org; Qinchuanyu; Yewudi; liuyongan 00175866; Wangbicheng; Yan Chen Subject: Re: [Question] Usage of dev_hold()/dev_put() On Tue, 2015-08-25 at 08:28 +0000, Zhangjie (HZ) wrote: > Hi, > > The comment of function dev_hold() and dev_put() are really simple. > > Actually, I can’t find a rule to follow. > > Where should I call dev_hold()/dev_put()? > > Are they necessary during xmit or receive or only in register/release? > > I find that, for tap, it calls dev_hold()/dev_put() for every > tun_sendmsg/tun_recvmsg: > > So, should I call dev_hold()/dev_put() for each quote, such like: > “skb->dev = dev” ? > > But, for physical nic, I can’t find dev_hold()/dev_put() during xmit > or receive. > > What kind of scene is it necessary to call dev_hold()/dev_put()? > > > > Look forward for your feedback. > > Thank you ! J > > Zhangjie > > Please do not send HTML mails, otherwise they do not reach netdev mailing list In general, you need to use dev_hold() for every reference on 'dev' stored in the object. However they are some paths were you do not need that : For example in transmit path, skb are stored either in a qdisc or device driver internal queue/ring. At device dismantle we properly delete all these skb, so we do not have to worry about used dev_hold()/dev_put() in transmit. In drivers/net/tun.c, the dev_hold()/dev_put() only is required so that device doesnt disappear between tun_get() and actual use of it after a potentially long copy from user space (that might trigger page faults)
