Re: [PATCH RFC] mm: migrate: Fix races of __find_get_block() and page migration

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, Jul 12, 2019 at 11:17:46AM +0200, Jan Kara wrote:
> On Thu 11-07-19 17:04:55, Andrew Morton wrote:
> > On Thu, 11 Jul 2019 14:58:38 +0200 Jan Kara <jack@xxxxxxx> wrote:
> > 
> > > buffer_migrate_page_norefs() can race with bh users in a following way:
> > > 
> > > CPU1					CPU2
> > > buffer_migrate_page_norefs()
> > >   buffer_migrate_lock_buffers()
> > >   checks bh refs
> > >   spin_unlock(&mapping->private_lock)
> > > 					__find_get_block()
> > > 					  spin_lock(&mapping->private_lock)
> > > 					  grab bh ref
> > > 					  spin_unlock(&mapping->private_lock)
> > >   move page				  do bh work
> > > 
> > > This can result in various issues like lost updates to buffers (i.e.
> > > metadata corruption) or use after free issues for the old page.
> > > 
> > > Closing this race window is relatively difficult. We could hold
> > > mapping->private_lock in buffer_migrate_page_norefs() until we are
> > > finished with migrating the page but the lock hold times would be rather
> > > big. So let's revert to a more careful variant of page migration requiring
> > > eviction of buffers on migrated page. This is effectively
> > > fallback_migrate_page() that additionally invalidates bh LRUs in case
> > > try_to_free_buffers() failed.
> > 
> > Is this premature optimization?  Holding ->private_lock while messing
> > with the buffers would be the standard way of addressing this.  The
> > longer hold times *might* be an issue, but we don't know this, do we? 
> > If there are indeed such problems then they could be improved by, say,
> > doing more of the newpage preparation prior to taking ->private_lock.
> 
> I didn't check how long the private_lock hold times would actually be, it
> just seems there's a lot of work done before the page is fully migrated a
> we could release the lock. And since the lock blocks bh lookup,
> set_page_dirty(), etc. for the whole device, it just seemed as a bad idea.
> I don't think much of a newpage setup can be moved outside of private_lock
> - in particular page cache replacement, page copying, page state migration
> all need to be there so that bh code doesn't get confused.
> 
> But I guess it's fair to measure at least ballpark numbers of what the lock
> hold times would be to get idea whether the contention concern is
> substantiated or not.
> 

I think it would be tricky to measure and quantify how much the contention
is an issue. While it would be possible to construct a microbenchmark that
should illustrate the problem, it would tell us relatively little about
how much of a problem it is generally. It would be relatively difficult
to detect the contention and stalls in block lookups due to migration
would be tricky to spot. Careful use of lock_stat might help but
enabling that has consequences of its own.

However, a rise in allocation failures due to dirty pages not being
migrated is relatively easy to detect and the consequences are relatively
benign -- failed high-order allocation that is usually ok versus a stall
on block lookups that could have a wider general impact.

On that basis, I think the patch you proposed is the more appropriate as
a fix for the race which has the potential for data corruption. So;

Acked-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>

> Finally, I guess I should mention there's one more approach to the problem
> I was considering: Modify bh code to fully rely on page lock instead of
> private_lock for bh lookup. That would make sense scalability-wise on its
> own. The problem with it is that __find_get_block() would become a sleeping
> function. There aren't that many places calling the function and most of
> them seem fine with it but still it is non-trivial amount of work to do the
> conversion and it can have some fallout so it didn't seem like a good
> solution for a data-corruption issue that needs to go to stable...
> 

Maybe *if* it's shown there is a major issue with increased high-order
allocation failures, it would be worth looking into but right now, I
think it's overkill with relatively high risk and closing the potential
race is more important.

-- 
Mel Gorman
SUSE Labs




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux