Hello Krzyztof,
I am looking at your hdlc_ppp code and I don't understand: why is there
not the equivalent of fr_detach() in there?
pc8300_drv:cpc_remove_one() frees netdevs quite confidently but I wonder
how it can be so sure that there are not skbs in hdlc_ppp's tx_queue
associated with those devices before freeing them....
Even if you wanted to switch a device from PPP to Frame Relay, I don't
see the method right now. If I may ask, please, what am I missing?
If you agree there is a need for detach(), I would be happy to work on
it and make a submission.
thanks for your time,
--
Michael Barkowski
RuggedCom, Inc.
Michael Barkowski wrote:
> Hello Krzyztof,
>
> I am looking at your hdlc_ppp code and I don't understand: why is there
> not the equivalent of fr_detach() in there?
>
> pc8300_drv:cpc_remove_one() frees netdevs quite confidently but I wonder
> how it can be so sure that there are not skbs in hdlc_ppp's tx_queue
> associated with those devices before freeing them....
>
the above is the real danger I see - free the netdev, then ppp's timer
comes along and dequeues from tx_queue an skb with invalid device.
> Even if you wanted to switch a device from PPP to Frame Relay, I don't
> see the method right now. If I may ask, please, what am I missing?
>
Ok - this part was a momentary lapse on my part - please strike from
the record :)
> If you agree there is a need for detach(), I would be happy to work on
> it and make a submission.
>
> thanks for your time,
>
--
Michael Barkowski
905-482-4577
Hello Michael,
Michael Barkowski <[email protected]> writes:
> I am looking at your hdlc_ppp code and I don't understand: why is there
> not the equivalent of fr_detach() in there?
I assume you mean .detach = fr_destroy(). It's used only to kill
subdevices, i.e. it has nothing to do with the interface being up/down.
> pc8300_drv:cpc_remove_one() frees netdevs quite confidently but I wonder
> how it can be so sure that there are not skbs in hdlc_ppp's tx_queue
> associated with those devices before freeing them....q
Theoretically all paths adding skbs to the tx_queue should send them out
before returning (possibly also on behalf of other devices). However I
wonder if it's the case. Let's see: Only ppp_tx_cp() adds to the queue
directly:
- ppp_rx() (calls ppp_tx_flush())
- ppp_timer (calls ppp_tx_flush())
- ppp_cp_event():
- ppp_cp_parse_cr() (calls ppp_tx_flush())
- ppp_stop() calls ppp_cp_event(), but it won't queue any skb, it only
marks the connection as closed and does the same to IPCP and IPV6CP.
This means the problematic part is ppp_start() which calls
ppp_cp_event(LCP, START) = IRC | SCR | 3 meaning
Initialize-Restart-Count, Send-Configure-Request and change state to
REQ_SENT. This causes two problems:
1. The SCR packet will be delayed by 2 seconds (both first and second
SCR will be sent the same time). Perhaps we delay only a little
(instead of full 2 seconds) and only then send the initial packet.
2. (as you noted) the skb will be added to tx_queue and left there. If
we happen to "ifconfig up" and "rmmod driver" before receiving any
packet and before ppp->req_timeout (2 seconds) and before any other
PPP interface does the same, we will eventually get skb with invalid
->dev. This is simple to drain in .close (detach is a wrong place
since it may be called long after the interface is deactivated, there
is no need to delay it past .close). The fix for #1 will already fix
#2, but the redundant safety doesn't cost us anything.
Thanks for noting the problem, I'll post a patch shortly.
Also it seems the timeouts etc. should be configurable. ATM we're only
fixing bugs, good.
--
Krzysztof Halasa
> 1. The SCR packet will be delayed by 2 seconds (both first and second
> SCR will be sent the same time). Perhaps we delay only a little
> (instead of full 2 seconds) and only then send the initial packet.
BTW can I sleep in a netdevice_notifier callback, no?
--
Krzysztof Halasa