[PATCH 1/1] net/mana: add 32 bit short doorbell

[Long Li]

> > Subject: RE: [PATCH 1/1] net/mana: add 32 bit short doorbell
> >
> > > > > +#ifdef RTE_ARCH_32
> > > > > +	uint16_t cqe_incr =
> > > > > +(uint16_t)rxq->gdma_cq.head_incr_to_short_db;
> > > >
> > > > How do you make sure head_incr_to_short_db doesn't overflow?
> > > >
> > >
> > > I have checked this with hardware team. In my opinion it would be
> > > easily overflown.
> > > The hw team seems suggesting the hw will take care of this.
> > >
> > > Thanks,
> > > Wei
> >
> > I'm not sure how HW can take care of this when it overflows. When it
> > happens, the HW will miss a doorbell and CQ queue will get full. And
> > eventually you'll lose completions for TX/RX.
> >
> > In mana_alloc_and_post_rx_wqes() and mana_rx_burst(), the code has
> > check for RX/TX_WQE_SHORT_DB_THRESHOLD to make sure tail_incr doesn't
> > overflow when ringing the doorbell.
> >
> > In gdma_poll_completion_queue(), you need to have a similar mechanism
> > to not overflow tail_incr  when ringing the doorbell.
> >
> I am not sure what can be done here. Applications could run in poll mode without
> need to ring cq doorbell, or it could take very long time before it change the
> interrupt state. What we can do when cq->head_incr_to_short_db reaches 0xffff
> in gdma_poll_completion_queue()?
> 
> If it breaks out the loop and return, the next time it enters it may still  at0xffff
> because it has not rung doorbell the have it reset yet.
> 
> If  just resetting the value to 0 and let it keep going in the loop, it is no difference
> than casting it to 16 bit unsigned int, which would be done in
> mana_arm_cq() if it is eventually called.
> 
> Anyway, ringing cq doorbell has not been tested as the driver doesn't support
> interrupts.

If 32bit doorbell doesn't support interrupts, you can remove this feature from 32 bits.

Long