[RFC] mm:change meminfo cached calculation

[Hugh Dickins]

On Tue, 6 Jan 2015, Andrew Morton wrote:
> On Tue, 6 Jan 2015 17:04:33 -0800 (PST) Hugh Dickins <hughd at google.com> wrote:
> > On Tue, 6 Jan 2015, Andrew Morton wrote:
> > > On Fri, 26 Dec 2014 19:56:49 +0800 "Wang, Yalin" <Yalin.Wang at sonymobile.com> wrote:
> > > 
> > > > This patch subtract sharedram from cached,
> > > > sharedram can only be swap into swap partitions,
> > > > they should be treated as swap pages, not as cached pages.
> > > > 
> > > > ...
> > > >
> > > > --- a/fs/proc/meminfo.c
> > > > +++ b/fs/proc/meminfo.c
> > > > @@ -45,7 +45,7 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
> > > >  	committed = percpu_counter_read_positive(&vm_committed_as);
> > > >  
> > > >  	cached = global_page_state(NR_FILE_PAGES) -
> > > > -			total_swapcache_pages() - i.bufferram;
> > > > +			total_swapcache_pages() - i.bufferram - i.sharedram;
> > > >  	if (cached < 0)
> > > >  		cached = 0;
> > > 
> > > Documentation/filesystems/proc.txt says
> > > 
> > > :      Cached: in-memory cache for files read from the disk (the
> > > :              pagecache).  Doesn't include SwapCached
> > > 
> > > So yes, I guess it should not include shmem.
> > > 
> > > And why not do this as well?
> > > 
> > > 
> > > --- a/Documentation/filesystems/proc.txt~mm-change-meminfo-cached-calculation-fix
> > > +++ a/Documentation/filesystems/proc.txt
> > > @@ -811,7 +811,7 @@ MemAvailable: An estimate of how much me
> > >       Buffers: Relatively temporary storage for raw disk blocks
> > >                shouldn't get tremendously large (20MB or so)
> > >        Cached: in-memory cache for files read from the disk (the
> > > -              pagecache).  Doesn't include SwapCached
> > > +              pagecache).  Doesn't include SwapCached or Shmem.
> > >    SwapCached: Memory that once was swapped out, is swapped back in but
> > >                still also is in the swapfile (if memory is needed it
> > >                doesn't need to be swapped out AGAIN because it is already
> > 
> > Whoa.  Changes of this kind would have made good sense about 14 years ago.
> > And there's plenty more which would benefit from having anon/shmem/file
> > properly distinguished.  But how can we make such a change now,
> > breaking everything that has made its own sense of these counts?
> 
> That's what I was wondering, but I was having some trouble picking a
> situation where it mattered much.

If it doesn't matter, then we don't need to change it.

> What's the problematic scenario
> here?  Userspace that is taking Cached, saying "that was silly" and
> subtracting Shmem from it by hand?

Someone a long time ago saw "that was silly", worked out it was because
of Shmem, adjusted their scripts or whatever accordingly, and has run
happily ever since.

> 
> I suppose that as nobody knows we should err on the side of caution and
> leave this alone.  But the situation is pretty sad - it would be nice
> to make the code agree with the documentation at least.

By all means fix the documentation.  And work on a /proc/meminfo.2015
which has sensibly differentiated counts (and probably omits that
wonderful Linux 2.2-compatible "Buffers").

But there's more to do than I can think of.  Cc'ing Jerome who has a
particular interest in this (no, I haven't forgotten his patches,
but nor have I had a moment to reconsider them).

Hugh