[ofa-general] spin_lock_irqsave in ib_send_mad

[Suresh Shelvapille]

Sean:

Do you see a problem if the piece of code in ib_send_mad is re-written as follows:

----------------------------------------------------  
if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
                ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
                                   &bad_send_wr);
                spin_lock_irqsave(&qp_info->send_queue.lock, flags);
                list = &qp_info->send_queue.list;
} else {
                ret = 0;
                spin_lock_irqsave(&qp_info->send_queue.lock, flags);
                list = &qp_info->overflow_list;
}
--------------------------------------------------------

I tested it in my environment and it seemed to work OK. 

Thanks,
Suri

> -----Original Message-----
> From: Sean Hefty [mailto:sean.hefty at intel.com]
> Sent: Wednesday, June 10, 2009 1:01 PM
> To: 'Suresh Shelvapille'; general at lists.openfabrics.org
> Cc: 'Roland Dreier'
> Subject: RE: [ofa-general] spin_lock_irqsave in ib_send_mad
> 
> >mad.c:ib_send_mad() calls ib_post_send() after taking spin_lock_irqsave().
> >
> >Is it really necessary to take the spinlock during the entire time of
> >ib_post_send()? It appears like it is only necessary for list manipulation!
> 
> It protects the list and the counters.  It's technically not needed around
> ib_post_send.
> 
> - Sean