Re: [PATCH 3/4] myri10ge - Driver core

[Brice Goglin <brice@myri.com>]

Brice Goglin wrote:
> Andi Kleen wrote:
>   
>> On Friday 19 May 2006 12:00, Arnd Bergmann wrote:
>>   
>>     
>>> On Friday 19 May 2006 04:25, Brice Goglin wrote:
>>>     
>>>       
>>>> dev_mc_upload() from net/core/dev_mcast.c does
>>>>
>>>> spin_lock_bh(&dev->xmit_lock);
>>>> __dev_mc_upload(dev);
>>>>
>>>> which calls dev->set_multicast_list(), which is
>>>> myri10ge_set_multicast_list()
>>>>
>>>> which calls myri10ge_change_promisc
>>>>
>>>> which calls myri10ge_send_cmd
>>>>       
>>>>         
>>> Hmm, if that is the only path where you call it, it may be
>>> helpful to change myri10ge_change_promisc to call a special
>>> atomic version of myri10ge_send_cmd then, while all others
>>> use the regular version, which you can then convert to
>>> do msleep as well.
>>>     
>>>       
>> Or just drop the xmit lock while you do that. As long as you
>> handle races with ->start_xmit yourself it's ok
>>   
>>     
>
> Looks like we can do that.
>   

Forget what I said. It's not that simple. __dev_mc_upload() is also
called from igmp6_group_added() (through dev_mc_add()). In this path,
there are two spin_lock_bh that are held:

igmp6_group_added(...)
    [...]
    spin_lock_bh(&mc->mca_lock);
    [...]
    dev_mc_add(...)

dev_mv_add(..)
    [...]
    spin_lock_bh(&dev->xmit_lock);
    [...]
    __dev_mc_upload(...)

We have actually seen this path being taken by adding a
    if (in_interrupt()) {
       printk("bad bad bad\n");
       dump_stack();
    }
in our myri10ge_send_cmd(...).

We are going to add a "sleepable" parameter to myri10ge_send_cmd() so
that only set_multicast_list() will use it with our previous udelay loop.

Brice