Re: 3 PCI NIC's on OpenBSD 2.6(x86) sharing IRQ's

[Nick Holland <nhollan@home.com>]

It isn't the days that were the problem, it is the technology used.

You CAN NOT share interrupts on the ISA bus.
Not this today, not fifteen years ago, not the day the IBM PC shipped.
(I suddenly feel old.)

However, you most certainly CAN share interrupts on the PCI bus.  That
was one of the points of the PCI bus (personally, while they were at
it, I wish they would have ganged up another IRQ controller, but they
didn't ask me).

You can not share IRQs between the ISA and the PCI bus.  If IRQ 5 is
being used by a device on the ISA bus, you can not use it on the PCI
bus, and visa versa.

There is nothing wrong with using IRQ 9, as long as nothing else is
using it that can't share (see preceeding paragraph, nothing special
about IRQ 2/9 here).

There is nothing wrong with using IRQ 2, as long as your software was
written in the last 15 years and knows that on an AT (or better) class
machine, hardware that uses IRQ 2 is really "hitting" IRQ 9.  This
*could* create a problem, as on an 8088-based PC, hardware that hit
the IRQ2 line hit IRQ 2.  Same card plugged into a 286 or better
system would have the software experiencing an IRQ 9.  Haven't come
across software that stupid in a very long time (I don't ever recall
having come across software that stupid, actually, but I expect that
pre-AT hardware could well have had -- must have had! -- driver
problems with this).  Being that OpenBSD requires a minimum of a 386
to run, I rather suspect there is no concept of IRQ2 on the ISA bus in
OpenBSD.

I've heard that bull about "can't use 2/9" almost since the AT came
out.  Untrue.  I believe the misinformation came from the fact that
many ISA video cards utilized IRQ 2/9 as the video interrupt, but
virtually no software used it, and none of the software-based
"diagnostic" programs found any software linking to IRQ2/9, so many
people forgot that 2/9 was occupied, even if unutilized, and they
assumed that the cascade interrupt was the problem -- it wasn't.  This
is not an issue with PCI video cards, which register their IRQ usage,
you don't end up with PCI cards sitting on "hidden" resources.

As far as I am aware, there is no significant penalty for sharing
interrupts on the PCI bus.  The code has to be in place to identify
which card fired the IRQ anyway, otherwise the application would be
unable to share IRQs, and thus, be in violation of PCI design
standards.  PCI drivers don't just assume that IRQ was "for them" --
they gotta check and make sure -- even if they are the only device on
the IRQ (for next week, they may not be).  Yeah, you could probably
gain a small amount of performance if you wrote your drivers assuming
non-shared IRQs, but you would probably find yourself being lynched if
you were to try to pull something like that.

NOW...all that being said...Long ago on software platforms we would
rather not think about, there are legends of some PCI card
manufacturers which mearly took source for their ISA card drivers and
slightly modified it for their "new" PCI cards and thus, IRQ sharing
wouldn't work.  A more common issue is that some early PCI machines
were, well, brain dead in hardware.  They didn't do much of anything
right, sharing IRQs was the least of the problems.  Also early on,
supposedly you could have troubles sharing IRQs between different
drivers -- (say) Adaptec's SCSI card driver might battle with (say)
3Com's NIC driver over the IRQs, as the authors never talked to each
other, but that all happened on OSs that are not the subject of this
mail list -- and you can start to see some of the wonders of
Open-Source software (and more importantly, an OS developed by a small
core of very skilled people).

Nick.
(a.k.a., "Sid")


Steve Williams wrote:
> 
> Hi,
> 
> I am adding another(third) NIC into an OpenBSD(x86) box that is
> running 2.6.
> 
> This will make three Intel 10/100 NIC's in this box.  They are all
> PCI.
> 
> When I rebooted after installing the third NIC, dmesg shows that
> two of the NIC's are using the same IRQ.
> 
> Humm... I haven't done hardware since back in the "old" days, when it
> was critical that nothing share IRQ's.
> 
> Is this still the case?
> 
> Is it OK that these two cards are sharing an IRQ?
> 
> fxp0 at pci0 dev 13 function 0 "Intel 82557" rev 0x05: irq 11, address ...
> fxp1 at pci0 dev 15 function 0 "Intel 82557" rev 0x05: irq 9, address ...
> fxp2 at pci0 dev 17 function 0 "Intel 82557" rev 0x08: irq 11, address ...
> 
> Thanks for any assistance!
> --
>         Steve Williams, Calgary, Alberta, Canada
>         Genie Computer Systems Inc.
>         steve@genie96.com
> 
> "A man doesn't begin to attain wisdom until he recognizes that he is
>  no longer indispensable."
> - Admiral Richard E. Byrd ( 1888-1957 )

-- 
http://www.holland-consulting.net/