[lkml]   [2012]   [May]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [patch 0/5] refault distance-based file cache sizing
On Tue, May 01, 2012 at 02:26:56PM -0700, Andrew Morton wrote:
> On Tue, 01 May 2012 17:19:16 -0400
> Rik van Riel <> wrote:
> > On 05/01/2012 03:08 PM, Andrew Morton wrote:
> > > On Tue, 1 May 2012 10:41:48 +0200
> > > Johannes Weiner<> wrote:
> > >
> > >> This series stores file cache eviction information in the vacated page
> > >> cache radix tree slots and uses it on refault to see if the pages
> > >> currently on the active list need to have their status challenged.
> > >
> > > So we no longer free the radix-tree node when everything under it has
> > > been reclaimed? One could create workloads which would result in a
> > > tremendous amount of memory used by radix_tree_node_cachep objects.
> > >
> > > So I assume these things get thrown away at some point. Some
> > > discussion about the life-cycle here would be useful.
> >
> > I assume that in the current codebase Johannes has, we would
> > have to rely on the inode cache shrinker to reclaim the inode
> > and throw out the radix tree nodes.
> >
> > Having a better way to deal with radix tree nodes that contain
> > stale entries (where the evicted pages would no longer receive
> > special treatment on re-fault, because it has been so long) get
> > reclaimed would be nice for a future version.
> >
> Well, think of a stupid workload which creates a large number of very
> large but sparse files (populated with one page in each 64, for
> example). Get them all in cache, then sit there touching the inodes to
> keep then fresh. What's the worst case here?

With 8G of RAM, it takes a minimally populated file (one page per leaf
node) of 3.5TB to consume all memory for radix tree nodes.

The worst case is going OOM without someone to blame as the objects
are owned by the kernel.

Is this a use case we should worry about? A realistic one, I mean, it
wouldn't be the first one to take down a machine maliciously and could
be prevented by rlimiting the maximum file size.

That aside, entries that are past the point where they would mean
anything, as Rik described above, are a waste of memory, the severity
of which depends on how much of its previously faulted data an inode
has evicted while still being in active use.

For me it's not a question of whether we want a mechanism to reclaim
old shadow pages of inodes that are still in use, but how critical
this is, and then how accurate it needs to be etc.

 \ /
  Last update: 2012-05-03 15:41    [W:0.311 / U:12.608 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site