Re: [PATCH 01/17] mm/highmem: Lift memcpy_[to|from]_page and memset_page to core

[Ira Weiny <ira.weiny@intel.com>]

On Mon, Nov 30, 2020 at 11:22:32AM +0200, Joonas Lahtinen wrote:
> Quoting Matthew Wilcox (2020-11-27 15:20:06)
> > On Fri, Nov 27, 2020 at 03:06:24PM +0200, Joonas Lahtinen wrote:
> > > Quoting ira.weiny@intel.com (2020-11-24 08:07:39)
> > > > From: Ira Weiny <ira.weiny@intel.com>
> > > > 
> > > > Working through a conversion to a call such as kmap_thread() revealed
> > > > many places where the pattern kmap/memcpy/kunmap occurred.
> > > > 
> > > > Eric Biggers, Matthew Wilcox, Christoph Hellwig, Dan Williams, and Al
> > > > Viro all suggested putting this code into helper functions.  Al Viro
> > > > further pointed out that these functions already existed in the iov_iter
> > > > code.[1]
> > > > 
> > > > Placing these functions in 'highmem.h' is suboptimal especially with the
> > > > changes being proposed in the functionality of kmap.  From a caller
> > > > perspective including/using 'highmem.h' implies that the functions
> > > > defined in that header are only required when highmem is in use which is
> > > > increasingly not the case with modern processors.  Some headers like
> > > > mm.h or string.h seem ok but don't really portray the functionality
> > > > well.  'pagemap.h', on the other hand, makes sense and is already
> > > > included in many of the places we want to convert.
> > > > 
> > > > Another alternative would be to create a new header for the promoted
> > > > memcpy functions, but it masks the fact that these are designed to copy
> > > > to/from pages using the kernel direct mappings and complicates matters
> > > > with a new header.
> > > > 
> > > > Lift memcpy_to_page(), memcpy_from_page(), and memzero_page() to
> > > > pagemap.h.
> > > > 
> > > > Also, add a memcpy_page(), memmove_page, and memset_page() to cover more
> > > > kmap/mem*/kunmap. patterns.
> > > > 
> > > > [1] https://lore.kernel.org/lkml/20201013200149.GI3576660@ZenIV.linux.org.uk/
> > > >     https://lore.kernel.org/lkml/20201013112544.GA5249@infradead.org/
> > > > 
> > > > Cc: Dave Hansen <dave.hansen@intel.com>
> > > > Suggested-by: Matthew Wilcox <willy@infradead.org>
> > > > Suggested-by: Christoph Hellwig <hch@infradead.org>
> > > > Suggested-by: Dan Williams <dan.j.williams@intel.com>
> > > > Suggested-by: Al Viro <viro@zeniv.linux.org.uk>
> > > > Suggested-by: Eric Biggers <ebiggers@kernel.org>
> > > > Signed-off-by: Ira Weiny <ira.weiny@intel.com>
> > > 
> > > <SNIP>
> > > 
> > > > +static inline void memset_page(struct page *page, int val, size_t offset, size_t len)
> > > > +{
> > > > +       char *addr = kmap_atomic(page);
> > > > +       memset(addr + offset, val, len);
> > > > +       kunmap_atomic(addr);
> > > > +}
> > > 
> > > Other functions have (page, offset) pair. Insertion of 'val' in the middle here required
> > > to take a double look during review.
> > 
> > Let's be explicit here.  Your suggested order is:
> > 
> >         (page, offset, val, len)
> > 
> > right?  I think I would prefer that to (page, val, offset, len).
> 
> Yeah, I think that would be most consistent order.

Yes as I have been reworking these I have found it odd as well.  I'm going to
swap it around.  Been learning Coccinelle which has helped find other
instances...  So V2 is taking a bit of time.

Thanks,
Ira