lkml.org 
[lkml]   [2010]   [Apr]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 0/3] Fix migration races in rmap_walk() V2
On Wed, 28 Apr 2010 00:32:42 +0200
Andrea Arcangeli <aarcange@redhat.com> wrote:

> On Tue, Apr 27, 2010 at 05:27:36PM -0500, Christoph Lameter wrote:
> > Can we simply wait like in the fault path?
>
> There is no bug there, no need to wait either. I already audited it
> before, and I didn't see any bug. Unless you can show a bug with CPU A
> running the rmap_walk on process1 before process2, there is no bug to
> fix there.
>
I think there is no bug, either. But that safety is fragile.


> >
> > > Patch 3 notes that while a VMA is moved under the anon_vma lock, the page
> > > tables are not similarly protected. Where migration PTEs are
> > > encountered, they are cleaned up.
> >
> > This means they are copied / moved etc and "cleaned" up in a state when
> > the page was unlocked. Migration entries are not supposed to exist when
> > a page is not locked.
>
> patch 3 is real, and the first thought I had was to lock down the page
> before running vma_adjust and unlock after move_page_tables. But these
> are virtual addresses. Maybe there's a simpler way to keep migration
> away while we run those two operations.
>

Doing some check in move_ptes() after vma_adjust() is not safe.
IOW, when vma's information and information in page-table is incosistent...objrmap
is broken and migartion will cause panic.

Then...I think there are 2 ways.
1. use seqcounter in "mm_struct" as previous patch and lock it at mremap.
or
2. get_user_pages_fast() when do remap.

Thanks,
-Kame









\
 
 \ /
  Last update: 2010-04-28 02:19    [W:0.171 / U:0.184 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site