Re: O_DIRECT patch for processors with VIPT cache for mainlinekernel (specifically arm in our case)

[Ralf Baechle]

On Thu, Nov 20, 2008 at 08:27:19AM -0700, Matthew Wilcox wrote:

> I'm not quite sure why you need kmap_coherent().  If a page is mapped into
> userspace, you can find what address it's mapped to from
> page->mapping->i_mmap and page->index.  OTOH, that's potentially

Even if we know the userspace address of a page we do not necessarily have
a usable mapping for kernel purposes.  The userspace mapping might be r/o
when we need r/w or it might be in another process.  kmap_coherent takes
the job of creating a r/w mapping on a suitable kernel virtual address
that will avoid any aliases.

> page->mapping->i_mmap and page->index.  OTOH, that's potentially
> expensive since you need to grab the spinlock, and unless you have all
> user addresses coherent with each other (like parisc does), you need to
> figure out which process to be coherent with.

Having all userspace addresses of a page across all processes coherent with
each other is the only practicable solution in Linux; at least I don't think
how otherwise and within the currently kernel framework a platform could
sanely handle userspace-userspace aliases.  So we're talking about extending
this to cover userspace-kernelspace aliases.

The original reason for the introduction of kmap_coherent was avoiding
a cache alias in when a multi-threaded process forks.  The issue has been
debated on lkml in 2006 as part of my submission of a patchset under the
subject of "Fix COW D-cache aliasing on fork".  The description is somewhat
lengthy so I omit it here.

One of the ugly parts of kmap_coherent() is that it cannot be used safely
if the page has been marked as dirty by flush_dcache_page(); the callers
know about this and deal with it.

> I know James Bottomley did an experiment (and did an OLS presentation
> ...) on unmapping the entire page cache and greatly expanding the kmap
> area to do just this kind of thing.  I think he even got a speedup.

The speedup is no surprise.

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