[PATCH 0/3] Enable multiple MSI feature in pSeries

[Michael Ellerman]

On Mon, 2013-02-04 at 11:49 +0800, Mike Qiu wrote:
> > On Tue, 2013-01-15 at 15:38 +0800, Mike Qiu wrote:
> > > Currently, multiple MSI feature hasn't been enabled in pSeries,
> > > These patches try to enbale this feature.
> > Hi Mike,
> > 
> > > These patches have been tested by using ipr driver, and the driver patch
> > > has been made by Wen Xiong <wenxiong at linux.vnet.ibm.com>:
> > So who wrote these patches? Normally we would expect the original author
> > to post the patches if at all possible.
> Hi Michael
> 
> These Multiple MSI patches were wrote by myself, you know this feature
> has not enabled
> and it need device driver to test whether it works suitable. So I test
> my patches use 
> Wen Xiong's ipr patches, which has been send out to the maillinglist.
> 
> I'm the original author :)

Ah OK, sorry, that was more or less clear from your mail but I just
misunderstood.

> > > [PATCH 0/7] Add support for new IBM SAS controllers
> > I would like to see the full series, including the driver enablement.
> Yep, but the driver patches were wrote by Wen Xiong and has been send
> out.

OK, you mean this series?

http://thread.gmane.org/gmane.linux.scsi/79639


> I just use her patches to test my patches. all device support Multiple
> MSI can use my feature not only IBM SAS controllers, I also test my
> patches use the broadcom wireless card tg3, and also works OK.

You mean drivers/net/ethernet/broadcom/tg3.c ? I don't see where it
calls pci_enable_msi_block() ?

All devices /can/ use it, but the driver needs to be updated. Currently
we have two drivers that do so (in Linus' tree), plus the updated IPR.

> > > Test platform: One partition of pSeries with one cpu core(4 SMTs) and 
> > >                RAID bus controller: IBM PCI-E IPR SAS Adapter (ASIC) in POWER7
> > > OS version: SUSE Linux Enterprise Server 11 SP2  (ppc64) with 3.8-rc3 kernel 
> > > 
> > > IRQ 21 and 22 are assigned to the ipr device which support 2 mutiple MSI.
> > > 
> > > The test results is shown by 'cat /proc/interrups':
> > >           CPU0       CPU1       CPU2       CPU3       
> > > 21:          6          5          5          5      XICS Level     host1-0
> > > 22:        817        814        816        813      XICS Level     host1-1
> > This shows that you are correctly configuring two MSIs.
> > 
> > But the key advantage of using multiple interrupts is to distribute load
> > across CPUs and improve performance. So I would like to see some
> > performance numbers that show that there is a real benefit for all the
> > extra complexity in the code.

> Yes, the system just has suport two MSIs. Anyway, I will try to do
> some proformance test, to show the real benefit.
> But actually it needs the driver to do so. As the data show above, it
> seems there is some problems in use the interrupt, the irq 21 use few,
> most use 22, I will discuss with the driver author to see why and if
> she fixed, I will give out the proformance result.

Yeah that would be good.

I really dislike that we have a separate API for multi-MSI vs MSI-X, and
pci_enable_msi_block() also pushes the contiguous power-of-2 allocation
into the irq domain layer, which is unpleasant. So if we really must do
multi-MSI I would like to do it differently.

cheers