Broadcom bge(4) panics while shutting down

[Alexander Sack]

On Thu, May 14, 2009 at 4:01 PM, Alexander Sack <pisymbol at gmail.com> wrote:
> On Thu, May 14, 2009 at 3:47 PM, Xin LI <delphij at delphij.net> wrote:
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> Hi, Alexander,
>>
>> Alexander Sack wrote:
>>> Hello:
>>>
>>> Under heavy traffic (100% utilization GIGE on a 2 port BGE card)
>>> running BGE CURRENT driver I see panics on shutdown.  The reason is
>>> because bge_rxeof() while processing its RX ring of BD's drops the
>>> softc lock when it hands it off to its input function.  If bge_stop()
>>> is waiting for it, it will then proceed to acquire lock and then
>>> quiesce the hardware (reseting the card, clearing out BDs etc.).  Once
>>> bge_stop() releases the softc lock, then bge_rxeof() under an
>>> interrupt context (no polling here) will reacquire and continue to
>>> process the ring which is a bad idea.  It should check to see if the
>>> card is still running before continuing processing BDs (i.e. once
>>> IF_DRV_RUNNING has been reset by bge_stop(), bge_rxeof() is done, bail
>>> out).
>>>
>>> Here is my first go around with this patch:
>>>
>>>
>>> -- if_bge.c.CURRENT   2009-05-14 14:39:39.000000000 -0400
>>> +++ if_bge.c  2009-05-14 14:39:24.000000000 -0400
>>> @@ -3081,6 +3081,10 @@
>>>               uint16_t                vlan_tag = 0;
>>>               int                     have_tag = 0;
>>>
>>> +             if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
>>> +                     return;
>>> +             }
>>> +
>>>  #ifdef DEVICE_POLLING
>>>               if (ifp->if_capenable & IFCAP_POLLING) {
>>>                       if (sc->rxcycles <= 0)
>>>
>>>
>>> This prevents any panics during shutdown under heavy load and AS IT
>>> TURNS out (I feel stupid for not looking) that em(4) already had this
>>> check in its em_rxeof() function (right at the top of the loop).  I'm
>>> more than happy changing it to the em style but above seems reasonable
>>> to me though I have to verify there isn't anything missing off the
>>> loop from a hardware standpoint (I don't think so because bge_stop()
>>> did all the dirty work so I believe touching any registers after that
>>> from bge_rxeof() is a bad idea).
>>>
>>> Preliminary testing shows no more panics start and stopping ports
>>> under heavy load (panics were almost immediate otherwise).
>>>
>>> Thoughts?
>>
>> I think this would solve the problem but I'm not sure whether this would
>> increase some overhead on the RX path.  It seems that there is a race
>> between bge_release_resources() and bge_intr(), I mean, it might be a
>> good idea to "drain" bge_intr() instead?
>
> Are you talking about detach time?  Because bge_stop() gets called
> before bge_release_resources() and stops host interrupts so where is
> the race again?  I mean at this point no more interrupts should be
> delivered to bge_intr() (I can confirm from spec since BGE has
> released it in the wild).  So why would you "drain" it at this
> point....(the hardware is down including the firmware).
>
> I agree it adds a little overhead to the standard bge_rxeof() path
> which I agree is very sensitive to change.  However, I think the check
> at top is tolerable since the other recourse is crash.  I mean its
> very easy to reproduce.  Flood a Broadcom card with traffic then stop
> the card and let the race begin...it will go down in bge_rxeof() after
> bge_stop releases the lock.
>
> I actually did not look at changing anything structurally to perhaps
> make this whole predicament better but minimally there should be a
> shield against this no?
>
> -aps
>

http://www.freebsd.org/cgi/query-pr.cgi?pr=134548

To track...with patch (though spacing got killed, my apologies, I
moved the check into the while logic a la em).  I've tested this with
zero issue so far.

-aps