Re: [PATCH] mm/highmem: Align-down to page the address for kunmap_flush_on_unmap()

[Al Viro]

On Thu, Jan 26, 2023 at 09:56:07PM +0000, Matthew Wilcox wrote:
> On Thu, Jan 26, 2023 at 09:04:02PM +0000, Al Viro wrote:
> > On Thu, Jan 26, 2023 at 08:48:03PM +0000, Matthew Wilcox wrote:
> > > On Thu, Jan 26, 2023 at 12:38:58PM -0800, Andrew Morton wrote:
> > > > On Thu, 26 Jan 2023 15:33:46 +0100 "Fabio M. De Francesco" <fmdefrancesco@xxxxxxxxx> wrote:
> > > > 
> > > > > If ARCH_HAS_FLUSH_ON_KUNMAP is defined (PA-RISC case), __kunmap_local()
> > > > > calls kunmap_flush_on_unmap(). The latter currently flushes the wrong
> > > > > address (as confirmed by Matthew Wilcox and Helge Deller). Al Viro
> > > > > proposed to call kunmap_flush_on_unmap() on an aligned-down to page
> > > > > address in order to fix this issue. Consensus has been reached on this
> > > > > solution.
> > > > 
> > > > What are the user-visible runtime effects of this flaw?
> > > 
> > > The version of this patch I sent out includes this information,
> > > as well as the missed alignment for kunmap_atomic().
> > 
> > One point: AFAICS, the situation right now is
> > 	* all callers of kunmap_local() pass page-aligned pointers
> 
> Ah, no.  kmap_local_folio() accepts a byte offset within the folio
> and returns a pointer to that byte.  I hadn't noticed the parisc
> case and thought it was already allowed to pass a misaligned pointer
> to kunmap_local() since it is allowed for the highmem case.  It
> simplified the callers, so it looked like a good tradeoff.
> 
> See, eg 338f379cf7c2:
> 
> -               src_addr = kmap_atomic(src_page);
> -               dest_addr = kmap_atomic(dest_page);
> +               src_addr = kmap_local_folio(src_folio,
> +                                       offset_in_folio(src_folio, srcoff));
> +               dst_addr = kmap_local_folio(dst_folio,
> +                                       offset_in_folio(dst_folio, dstoff));
> 
> -               if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len))
> +               if (memcmp(src_addr, dst_addr, cmp_len))
> 
> -               kunmap_atomic(dest_addr);
> -               kunmap_atomic(src_addr);
> +               kunmap_local(dst_addr);
> +               kunmap_local(src_addr);

Argh...  Missed that.

OK, then Fixes: on the patch makes a lot of sense.  And it still
needs a chunk in D/m/highmem.rst...