On 25-08-08 00:37, Rene Herman wrote:
On 24-08-08 21:48, Krzysztof Helt wrote:
From the Documentation/memory-barriers.txt " (*) inX(), outX(): ... They are guaranteed to be fully ordered with respect to each other.
They are not guaranteed to be fully ordered with respect to othertypes of memory and I/O operation. ... "
No barriers are needed in this case.
On x86 certainly not and I would expect these barriers are there for cases/architectures where inb() and outb() are redefined to memory mapped I/O or there would just be no point at all.
In that case, the first outb() writes an index and we'd better make sure that it hits the hardware before we write the value. Hence my suspicion that if at all, the mb() should be in between.
No idea who added those and why...
Well, thinking about this more, I guess it actually makes some sense.
No need for a wmb() in between because we know that the outbs are fully ordered (supposedly even when redefined) but being a general accessor function we pretend we don't know that the next access is going to be through an inb/outb again -- we are supposedly being careful to allow combining using this accessor with direct MMIO access on other and/or virtualized architectures.
That is... as far as I understand this in the first place. Takashi or Jaroslav, do you perhaps remember barriers being added to cs4231_lib() and why? Could we just get rid of them?
(context was snipped, but this question is about the mb() in snd_cs4231_dout; snd_wss_dout in the new wss_lib form).
Rene.