[PATCH v6 05/24] erofs: add inode operations

Gao Xiang gaoxiang25 at huawei.com
Sat Aug 31 04:46:06 AEST 2019


Hi Christoph,

On Fri, Aug 30, 2019 at 09:42:05AM -0700, Christoph Hellwig wrote:
> On Thu, Aug 29, 2019 at 07:59:22PM +0800, Gao Xiang wrote:
> > On Thu, Aug 29, 2019 at 03:24:26AM -0700, Christoph Hellwig wrote:
> > 
> > []
> > 
> > > 
> > > > +
> > > > +		/* fill last page if inline data is available */
> > > > +		err = fill_inline_data(inode, data, ofs);
> > > 
> > > Well, I think you should move the is_inode_flat_inline and
> > > (S_ISLNK(inode->i_mode) && inode->i_size < PAGE_SIZE) checks from that
> > > helper here, as otherwise you make everyone wonder why you'd always
> > > fill out the inline data.
> > 
> > Currently, fill_inline_data() only fills for fast symlink,
> > later we can fill any tail-end block (such as dir block)
> > for our requirements.
> 
> So change it when that later changes actually come in.  And even then
> having the checks outside the function is a lot more obvious.

Okay.

> 
> > And I think that is minor.
> 
> The problem is that each of these issues might appear minor on their
> own.  But combined a lot of the coding style choices lead to code that
> is more suitable an obsfucated code contest than the Linux kernel as
> trying to understand even just a few places requires jumping through
> tons of helpers with misleading names and spread over various files.
> 
> > The consideration is simply because iget_locked performs better
> > than iget5_locked.
> 
> In what benchmark do the differences show up?

In a word, no benchmark here, just because
"unsigned long on 32-bit platforms is 4 bytes."
but erofs nid is a 64-bit number.

iget_locked will do find_inode_fast (no callback at all)
rather than iget5_locked --> find_inode (test callback) ->
            inode_insert5(set callback) for each new inode.

For most 64-bit platforms, iget_locked is enough,
32-bit platforms become rare...

Thanks,
Gao Xiang



More information about the Linux-erofs mailing list