[dpdk-dev] [RFC PATCH v3 2/3] eal: add mask and unmask interrupt APIs

[Hyong Youb Kim (hyonkim)]

> -----Original Message-----
> From: Jerin Jacob Kollanukkaran <jerinj at marvell.com>
> Sent: Wednesday, July 17, 2019 6:21 PM
> To: Hyong Youb Kim (hyonkim) <hyonkim at cisco.com>; Nithin Kumar
> Dabilpuram <ndabilpuram at marvell.com>; David Marchand
> <david.marchand at redhat.com>; Thomas Monjalon
> <thomas at monjalon.net>; Ferruh Yigit <ferruh.yigit at intel.com>; Bruce
> Richardson <bruce.richardson at intel.com>
> Cc: John Daley (johndale) <johndale at cisco.com>; Shahed Shaikh
> <shshaikh at marvell.com>; dev at dpdk.org
> Subject: RE: [RFC PATCH v3 2/3] eal: add mask and unmask interrupt APIs
> 
> > > > Not sure. I do not have a good suggestion here :-) Like to hear from
> > > > David when he comes back, as he spent most time on this issue..
> > >
> > > Sure. He is on vacation.
> > > Any reason for thinking, rte_intr_ack()  may not be semantically correct?
> > > I think, it is very much correct if there are no better suggestions.
> > > Anyway it the name, From VFIO perspective, we know what is expected
> so
> > > I think it is fine.
> > >
> > > >
> > > > Why not return -1 and let the caller deal with it?
> > >
> > > If we make it as rte_intr_ack() no need to return -1 for
> > > MSIX-VFIO+Linux as it is semantically correct.
> > >
> >
> > Ack can be ambiguous. For INTx, ack usually means PIO to a NIC register,
> > saying "I got your interrupt, please de-assert irq".
> 
> I think, it vary from the perspective of IRQ Chip(or controller) vs NIC
> register(Source) PoV.
> Since the API starts from rte_intr_* it is more of controller so _ack_ make
> sense
> Other reason for ack:
> 1) It will enforce that it needs to be called form ISR
> 2) It would be have been really correct to unmask if VFIO+MSIx+Linux
> supports
> it
> 3) if it is ack, no need to add unmask counterpart, the _mask_ API
> 

Just curious, what you mean by irq controller? Ack/mask/unmask PIOs
all go to the NIC. It is the NIC that asserts/de-asserts irq..

> >
> > Besides the name, are we agreeing that we want these?
> > - Unmask if INTx
> 
> Yes
> 
> > - Nothing if MSI/MSI-X
> Yes for MSI over VFIO
> No for MSI over UIO/igb_uio
> 

I guess I was not clear. For MSI/MSI-X, we do not want to do
mask/unmask regardless of vfio-pci/igb_uio.  Below is my comment about
linux/windows/freebsd from an earlier email. Do you disagree? I am
sure there are plenty of kernel NIC driver guys here. Please correct
me if I am mistaken...

---
Masking is useful only for INTx, IMO...

Masking MSI/MSI-X via PCI-defined mechanisms (e.g. Mask bit in MSI-X
Table) has no practical use for drivers. Handshaking/masking/unmasking
is done via device/vendor specific ways, as needed. See all those
ack/block/unblock/credit/... mechanisms used in various drivers/NICs
to control interrupts their own way.

A long time ago in early PCIe days, the linux kernel did auto-masking
for MSI/MSI-X (i.e. mask before calling netdev irq handler). It was
soon removed as it was unnecessary overhead (expensive PIOs to NIC for
every interrupt). Windows and FreeBSD do not do auto-masking either.
---

Most drivers have a single irq callback.

handler() {
  do_action()
  rte_intr_umask/ack()
}

Suppose MSI/MSI-X is used (super likely since it is the default).
With igb_uio, rte_intr_umask/ack() will actually do PIO writes to the
NIC to unmask. This is unnecessary overhead.

> I don't  have very strong opinion unmask vs ack. I prefer to have ack due the
> reasons stated above.
> If you really have strong opinion on using unmask, we will stick with that to
> make forward progress.
> Let us know.
> 

I have no strong opinion either.

Thanks..
-Hyong