Re: [PATCH] Remove needless flush_dcache_page call

[Nick Piggin]

On Fri, Jan 16, 2009 at 02:51:19PM +0900, MinChan Kim wrote:
> On Thu, Jan 15, 2009 at 10:33:38PM -0700, Matthew Wilcox wrote:
> > On Fri, Jan 16, 2009 at 02:28:04PM +0900, MinChan Kim wrote:
> > > Now, Anyone don't maintain cramfs.
> > > I don't know who is maintain romfs. so I send this patch to linux-mm, 
> > > lkml, linux-dev. 
> > > 
> > > I am not sure my thought is right. 
> > > 
> > > When readpage is called, page with argument in readpage is just new 
> > > allocated because kernel can't find that page in page cache. 
> > > 
> > > At this time, any user process can't map the page to their address space. 
> > > so, I think D-cache aliasing probelm never occur. 
> > > 
> > > It make sense ?
> > 
> > Sorry, no.  You have to call fluch_dcache_page() in two situations --
> > when the kernel is going to read some data that userspace wrote, *and*
> > when userspace is going to read some data that the kernel wrote.  From a
> > quick look at the patch, this seems to be the second case.  The kernel
> > wrote data to a pagecache page, and userspace should be able to read it.
> > 
> > To understand why this is necessary, consider a processor which is
> > virtually indexed and has a writeback cache.  The kernel writes to a
> > page, then a user process reads from the same page through a different
> > address.  The cache doesn't find the data the kernel wrote because it
> > has a different virtual index, so userspace reads stale data.
> 
> I see. :)
> 
> Thanks for quick reponse and good explaination.
> Hmm,.. one more question. 
> 
> I can't find flush_dcache_page call in mpage_readpage which is 
> generic read function. In case of ext fs, it use mpage_readpage 
> with readpage.
> 
> who and where call flush_dcache_page in mpage_readpage call path?

I think if the page is populated via IO, then it is responsibility of the
IO layers (eg dma API) to ensure caches are consistent. Presumably this
would include calling flush_dcache_page if we CPU is being used for
the copies (eg. see drivers/block/brd.c).

But there are quite possibly holes around here because not as much testing
is done on CPUs with these kinds of caches. Eg. brd probably should be
doing a flush_dcache_page in the rw == WRITE direction AFAIKS, so it picks
up user aliases here.

--
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/