Kmap_atomic vs Kmap

[Phillip Lougher <phillip@lougher.demon.co.uk>]

Andrew Morton wrote in relation to my initial SquashFS patch:
> Phillip Lougher <phillip@lougher.demon.co.uk> wrote:
> 
>>+skip_read:
>>+	memset(pageaddr + bytes, 0, PAGE_CACHE_SIZE - bytes);
>>+	kunmap(page);
>>+	flush_dcache_page(page);
>>+	SetPageUptodate(page);
>>+	unlock_page(page);
>>+
>>+	return 0;
>>+}

> See if you can use kmap_atomic() here - kmap() is slow, theoretically
> deadlocky and is deprecated.
> 

 From some browsing of the kernel source and a useful article on lwn.net 
I think I can replace all the readpage_xxx (symlink, readpage & 
readpage4k) kmap/kunmaps with kmap_atomic(page, KM_USER0)/kunmap(kaddr, 
KM_USER0)...

The lwn.net article mentions that k[un]map_atomic is the new alternative 
to kmap (as Andrew Morton mentioned).

I've noticed in asm-ppc/highmem.h the following comment

/*
  * The use of kmap_atomic/kunmap_atomic is discouraged - kmap/kunmap
  * gives a more generic (and caching) interface. But kmap_atomic can
  * be used in IRQ contexts, so in some (very limited) cases we need
  * it.
  */

Given what Andrew and the lwn.net article says, this comment is rather 
confusing.  Is it wrong?

Regards

Phillip Lougher