[Intel-gfx] [PATCH RFC PKS/PMEM 22/58] fs/f2fs: Utilize new kmap_thread()

[James Bottomley]

On Fri, 2020-10-09 at 14:34 -0700, Eric Biggers wrote:
> On Fri, Oct 09, 2020 at 12:49:57PM -0700, ira.weiny at intel.com wrote:
> > From: Ira Weiny <ira.weiny at intel.com>
> > 
> > The kmap() calls in this FS are localized to a single thread.  To
> > avoid the over head of global PKRS updates use the new
> > kmap_thread() call.
> > 
> > Cc: Jaegeuk Kim <jaegeuk at kernel.org>
> > Cc: Chao Yu <chao at kernel.org>
> > Signed-off-by: Ira Weiny <ira.weiny at intel.com>
> > ---
> >  fs/f2fs/f2fs.h | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> > 
> > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> > index d9e52a7f3702..ff72a45a577e 100644
> > --- a/fs/f2fs/f2fs.h
> > +++ b/fs/f2fs/f2fs.h
> > @@ -2410,12 +2410,12 @@ static inline struct page
> > *f2fs_pagecache_get_page(
> >  
> >  static inline void f2fs_copy_page(struct page *src, struct page
> > *dst)
> >  {
> > -	char *src_kaddr = kmap(src);
> > -	char *dst_kaddr = kmap(dst);
> > +	char *src_kaddr = kmap_thread(src);
> > +	char *dst_kaddr = kmap_thread(dst);
> >  
> >  	memcpy(dst_kaddr, src_kaddr, PAGE_SIZE);
> > -	kunmap(dst);
> > -	kunmap(src);
> > +	kunmap_thread(dst);
> > +	kunmap_thread(src);
> >  }
> 
> Wouldn't it make more sense to switch cases like this to
> kmap_atomic()?
> The pages are only mapped to do a memcpy(), then they're immediately
> unmapped.

On a VIPT/VIVT architecture, this is horrendously wasteful.  You're
taking something that was mapped at colour c_src mapping it to a new
address src_kaddr, which is likely a different colour and necessitates
flushing the original c_src, then you copy it to dst_kaddr, which is
also likely a different colour from c_dst, so dst_kaddr has to be
flushed on kunmap and c_dst has to be invalidated on kmap.  What we
should have is an architectural primitive for doing this, something
like kmemcopy_arch(dst, src).  PIPT architectures can implement it as
the above (possibly losing kmap if they don't need it) but VIPT/VIVT
architectures can set up a correctly coloured mapping so they can
simply copy from c_src to c_dst without any need to flush and the data
arrives cache hot at c_dst.

James