Re: [PATCH] ext4: convert to DIV_ROUND_UP() in mpage_process_page_bufs()

[Al Viro]

On Fri, Mar 10, 2023 at 07:05:27AM +0000, Al Viro wrote:
> On Thu, Mar 09, 2023 at 10:54:38PM -0800, Eric Biggers wrote:
> > On Fri, Mar 10, 2023 at 06:46:12AM +0000, Al Viro wrote:
> > > On Thu, Mar 09, 2023 at 10:43:55PM -0800, Eric Biggers wrote:
> > > > On Fri, Mar 10, 2023 at 06:37:29AM +0000, Al Viro wrote:
> > > > > On Thu, Mar 09, 2023 at 10:17:16PM -0800, Eric Biggers wrote:
> > > > > > On Fri, Mar 10, 2023 at 02:07:34PM +0800, Yangtao Li wrote:
> > > > > > > Just for better readability, no code logic change.
> > > > > > > 
> > > > > > > Signed-off-by: Yangtao Li <frank.li@xxxxxxxx>
> > > > > > > ---
> > > > > > >  fs/ext4/inode.c | 3 +--
> > > > > > >  1 file changed, 1 insertion(+), 2 deletions(-)
> > > > > > > 
> > > > > > > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > > > > > > index d251d705c276..d121cde74522 100644
> > > > > > > --- a/fs/ext4/inode.c
> > > > > > > +++ b/fs/ext4/inode.c
> > > > > > > @@ -2218,8 +2218,7 @@ static int mpage_process_page_bufs(struct mpage_da_data *mpd,
> > > > > > >  {
> > > > > > >  	struct inode *inode = mpd->inode;
> > > > > > >  	int err;
> > > > > > > -	ext4_lblk_t blocks = (i_size_read(inode) + i_blocksize(inode) - 1)
> > > > > > > -							>> inode->i_blkbits;
> > > > > > > +	ext4_lblk_t blocks = DIV_ROUND_UP(i_size_read(inode), i_blocksize(inode));
> > > > > > >  
> > > > > > 
> > > > > > Please don't do this.  This makes the code compile down to a division, which is
> > > > > > far less efficient.  I've verified this by checking the assembly generated.
> > > > > 
> > > > > Which compiler is doing that?
> > > > 
> > > > $ gcc --version
> > > > gcc (GCC) 12.2.1 20230201
> > > > 
> > > > i_blocksize(inode) is not a constant, so this should not be particularly
> > > > surprising.  One might hope that a / (1 << b) would be optimized into a >> b,
> > > > but that doesn't seem to happen.
> > > 
> > > It really ought to be a / (1u << b), though...
> > 
> > Sure, that does better:
> > 
> > uint64_t f(uint64_t a, int b)
> > {
> >         return a / (1U << b);
> > }
> > 
> > gcc:
> > 	0000000000000000 <f>:
> > 	   0:	48 89 f8             	mov    %rdi,%rax
> > 	   3:	89 f1                	mov    %esi,%ecx
> > 	   5:	48 d3 e8             	shr    %cl,%rax
> > 	   8:	c3                   	ret
> > 
> > clang:
> > 	0000000000000000 <f>:
> > 	   0:	89 f1                	mov    %esi,%ecx
> > 	   2:	48 89 f8             	mov    %rdi,%rax
> > 	   5:	48 d3 e8             	shr    %cl,%rax
> > 	   8:	c3                   	ret
> > 
> > But with a signed dividend (which is the case here) it gets messed up:
> > 
> > int64_t f(int64_t a, int b)
> > {
> >         return a / (1U << b);
> > }
> 
> *ow*
> 
> And i_size_read() is long long ;-/  Right.

Out of curiosity (and that's already too brittle for practical use) -
does DIV_ROUND_UP_ULL() do any better on full example?