Messages in this thread | | | Subject | Re: [patch 17/20] non-reclaimable mlocked pages | From | Lee Schermerhorn <> | Date | Wed, 19 Dec 2007 11:08:38 -0500 |
| |
On Wed, 2007-12-19 at 09:53 -0500, Rik van Riel wrote: > On Wed, 19 Dec 2007 15:24:07 +0100 > Peter Zijlstra <peterz@infradead.org> wrote: > > > I thought Lee had patches that moved pages with long rmap chains (both > > anon and file) out onto the non-reclaim list, for those a slow > > background scan does make sense. > > I suspect we won't be needing that code. The SEQ replacement for > swap backed pages might reduce the number of pages that need to > be scanned to a reasonable number. > > Remember, steady states are not a big problem with the current VM. > It's the sudden burst of scanning that happens when the VM decides > that it should start swapping (and every anonymous page is referenced) > that kills large systems.
Yes, I still have the patch [for long anon_vma lists--not for excessively mapped file, yet] and I'm keeping it up to date and tested. I do see softlockups on the anon_vma and i_mmap_locks under stress, even with the reader/writer lock patches. I'll be trying the workloads on Rik's latest patches to see if they address these lockups.
Lee
| |