Re: [PATCH] change name of rep_nop

David Woodhouse (dwmw2@infradead.org)
Tue, 09 Oct 2001 01:03:22 +0100


torvalds@transmeta.com said:
> There's no way we should implement "simon_says".
> There's a difference between stupiud hardware changing memory from
> under us through mapping tricks, and cache coherency in general.

True. Although for simplicity's sake I wasn't talking about the mapping
tricks, this was just for writing/erasing flash chips without doing any
paging - it you're mapping different chips into the same hole you need the
same cache-flush tricks even for read-only chips.

> What you want is the "memory_went_away_from_us()" kind of cache flush,
> which has nothing at all to do with cache coherency. And it should be
> explicitly and clearly named THAT

As you wish. How about:

void memory_went_away_from_us(void);
and
void memory_range_went_away_from_us(unsigned long start, unsigned long len);

Where 'start' is an ioremap cookie.

> and you should not blame the fact that x86 is always 100% cache coherent
> and that the normal cache coherency routines should _never_ be anything
> but a nop.

Indeed. That's eminently sane - it was just the nomenclature and the
documentation which was less so.

> Also note that wbinvd is known to corrupted the caches and lead to
> lockups on certain x86 cores. You need to be a _lot_ more careful than
> just doing it indiscriminately from a driver.

Yeah - but x86 isn't a particularly interesting architecture in this
context. If you can't selectively flush a range, you'll probably find that
you haven't gained enough from being able to do burst reads to offset the
cost of the complete cache flushes.

--
dwmw2

- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/