[openib-general] [PATCH v3 1/7] AMSO1100 Low Level Driver.

Arjan van de Ven arjan at infradead.org
Wed Jun 21 10:13:33 PDT 2006


>  0;
> > > +
> > > +		__raw_writew(0, elem->hw_desc + C2_RXP_STATUS);
> > > +		__raw_writew(0, elem->hw_desc + C2_RXP_COUNT);
> > > +		__raw_writew(0, elem->hw_desc + C2_RXP_LEN);
> > 
> > you seem to be a fan of the __raw_write() functions... any reason why?
> > __raw_ is not a magic "go faster" prefix....
> > 
> 
> In this particular case, I believe this is done to avoid a swap of '0'
> since its not necessary.  

but.. that should writew() and co just autodetect (or do it at compile
time)...
(maybe it doesn't and we have an optimization opportunity here ;)

> > Also on a related note, have you checked the driver for the needed PCI
> > posting flushes?
> > 
> 
> Um, what's a 'PCI posting flush'?  Can you point me where its
> described/used so I can see if we need it?  Thanx.

ok pci posting...

basically, if you use writel() and co, the PCI bridges in the middle are
allowed (and the more fancy ones do) cache the write, to see if more
writes follow, so that the bridge can do the writes as a single burst to
the device, rather than as individual writes. This is of course great...
... except when you really want the write to hit the device before the
driver continues with other actions. 

Now the PCI spec is set up such that any traffic in the other direction
(basically readl() and co) will first flush the write through the system
before the read is actually sent to the device, so doing a dummy readl()
is a good way to flush any pending posted writes.

Where does this matter? 
it matters most at places such as irq enabling/disabling, IO submission
and possibly IRQ acking, but also often in eeprom-like read/write logic
(where you do manual clocking and need to do delays between the
write()'s). But in general... any place where you do writel() without
doing any readl() before doing nothing to the card for a long time, or
where you are waiting for the card to do something (or want it done NOW,
such as IRQ disabling) you need to issue a (dummy) readl() to flush
pending writes out to the hardware.


does this explanation make any sense? if not please feel free to ask any
questions, I know I'm not always very good at explaining things.

Greetings,
   Arjan van de Ven





More information about the general mailing list