lkml.org 
[lkml]   [2010]   [Apr]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/2] mm,migration: Prevent rmap_walk_[anon|ksm] seeing the wrong VMA information
On Tue, Apr 27, 2010 at 05:37:59PM +0200, Andrea Arcangeli wrote:
> On Tue, Apr 27, 2010 at 11:29:05AM +0100, Mel Gorman wrote:
> > It could have been in both but the vma lock should have been held across
> > the rmap_one. It still reproduces but it's still the right thing to do.
> > This is the current version of patch 2/2.
>
> Well, keep in mind I reproduced the swapops bug with 2.6.33 anon-vma
> code, it's unlikely that focusing on patch 2 you'll fix bug in
> swapops.h. If this is a bug in the new anon-vma code, it needs fixing
> of course! But I doubt this bug is related to swapops in execve on the
> bprm->p args.
>

Why do you doubt it's unrelated to execve? From what I've seen during the day,
there is a race in execve where the VMA gets moved (under the anon_vma lock)
before the page-tables are copied with move_ptes (without a lock). In that
case, execve can encounter and copy migration ptes before migration removes
them. This also applies to mainline because it is only taking the RCU lock
and not the anon_vma->lock.

I have a prototype that "handles" the situation with the new anon_vma
code by removing the migration ptes it finds while moving page tables
but it needs more work before releasing.

An alternative would be to split vma_adjust() into locked and unlocked
versions. shift_arg_pages() could then take the anon_vma lock to lock
both the VMA move and the pagetable copy here.

/*
* cover the whole range: [new_start, old_end)
*/
if (vma_adjust(vma, new_start, old_end, vma->vm_pgoff, NULL))
return -ENOMEM;

/*
* move the page tables downwards, on failure we rely on
* process cleanup to remove whatever mess we made.
*/
if (length != move_page_tables(vma, old_start,
vma, new_start, length))
return -ENOMEM;

It'd be messy to split up the locking of vma_adjust like this though and
exec() will hold the anon_vma locks for longer just to guard against
migration. It's not clear this is better than having move_ptes handle
the

> I've yet to check in detail patch 1 sorry, I'll let you know my
> opinion about it as soon as I checked it in detail.
>

No problem. I still need to revisit all of these patches once I am
confident the swapops bug cannot be triggered any more.



\
 
 \ /
  Last update: 2010-04-27 18:39    [W:0.052 / U:0.080 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site