Re: [PATCH 8/28] rt2x00: optimize mac/bssid writing

[Ivo van Doorn <ivdoorn@gmail.com>]

On Wednesday 28 February 2007 19:24, Michael Buesch wrote:
> On Wednesday 28 February 2007 19:15, Ivo van Doorn wrote:
> > On Wednesday 28 February 2007 18:36, Michael Buesch wrote:
> > > On Wednesday 28 February 2007 15:07, Ivo van Doorn wrote:
> > > > Handling the mac and bssid configuration can be done much easier
> > > > by writing the passed data directly into the register instead
> > > > of moving it to a local variable first.
> > > > 
> > > > Signed-off-by: Ivo van Doorn <IvDoorn@gmail.com>
> > > > 
> > > > ---
> > > > 
> > > > diff --git a/drivers/net/wireless/mac80211/rt2x00/rt2400pci.c b/drivers/net/wireless/mac80211/rt2x00/rt2400pci.c
> > > > index 27e151d..b6bf9f3 100644
> > > > --- a/drivers/net/wireless/mac80211/rt2x00/rt2400pci.c
> > > > +++ b/drivers/net/wireless/mac80211/rt2x00/rt2400pci.c
> > > > @@ -319,14 +319,11 @@ static inline void rt2400pci_close_debugfs(struct rt2x00_dev *rt2x00dev){}
> > > >   */
> > > >  static void rt2400pci_config_bssid(struct rt2x00_dev *rt2x00dev, u8 *bssid)
> > > >  {
> > > > -	u32 reg[2] = { 0, 0 };
> > > > -
> > > >  	/*
> > > >  	 * The BSSID is passed to us as an array of bytes,
> > > >  	 * that array is little endian, so no need for byte ordering.
> > >            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > >  	 */
> > > > -	memcpy(&reg, bssid, ETH_ALEN);
> > > > -	rt2x00_register_multiwrite(rt2x00dev, CSR5, &reg[0], sizeof(reg));
> > > > +	rt2x00_register_multiwrite(rt2x00dev, CSR5, (u32*)bssid, ETH_ALEN);
> > >                                                     ^^^^^^
> > > 
> > > This doesn't break on BE machines?
> > 
> > No, the multiwrite (just like multiread) does not perform byteordering,
> > so the device will receive each byte in the correct order.
> 
> No wait. It's not about performing byteordering somewhere.
> It's that casting a bytearray (which is little endian) into an
> u32 array (which is CPU endian) is almost always wrong.
> 
> Do you writel this array in a loop inside of register_multiwrite?
> If yes, it's broken. writel expects values in CPU endianess.
> This array will always be little endian.

register_multiread uses memcpy_fromio(), and
register_multiwrite uses memcpy_toio().
So that should be safe to use right?
Otherwise would making register_multiwrite  accept a void* pointer be better?

Ivo