lkml.org 
[lkml]   [2017]   [Dec]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH -mm] mm, swap: Fix race between swapoff and some swap operations
On Fri, Dec 08, 2017 at 04:41:38PM +0800, Huang, Ying wrote:
> Minchan Kim <minchan@kernel.org> writes:
>
> > On Fri, Dec 08, 2017 at 01:41:10PM +0800, Huang, Ying wrote:
> >> Minchan Kim <minchan@kernel.org> writes:
> >>
> >> > On Thu, Dec 07, 2017 at 04:29:37PM -0800, Andrew Morton wrote:
> >> >> On Thu, 7 Dec 2017 09:14:26 +0800 "Huang, Ying" <ying.huang@intel.com> wrote:
> >> >>
> >> >> > When the swapin is performed, after getting the swap entry information
> >> >> > from the page table, the PTL (page table lock) will be released, then
> >> >> > system will go to swap in the swap entry, without any lock held to
> >> >> > prevent the swap device from being swapoff. This may cause the race
> >> >> > like below,
> >> >> >
> >> >> > CPU 1 CPU 2
> >> >> > ----- -----
> >> >> > do_swap_page
> >> >> > swapin_readahead
> >> >> > __read_swap_cache_async
> >> >> > swapoff swapcache_prepare
> >> >> > p->swap_map = NULL __swap_duplicate
> >> >> > p->swap_map[?] /* !!! NULL pointer access */
> >> >> >
> >> >> > Because swap off is usually done when system shutdown only, the race
> >> >> > may not hit many people in practice. But it is still a race need to
> >> >> > be fixed.
> >> >>
> >> >> swapoff is so rare that it's hard to get motivated about any fix which
> >> >> adds overhead to the regular codepaths.
> >> >
> >> > That was my concern, too when I see this patch.
> >> >
> >> >>
> >> >> Is there something we can do to ensure that all the overhead of this
> >> >> fix is placed into the swapoff side? stop_machine() may be a bit
> >> >> brutal, but a surprising amount of code uses it. Any other ideas?
> >> >
> >> > How about this?
> >> >
> >> > I think It's same approach with old where we uses si->lock everywhere
> >> > instead of more fine-grained cluster lock.
> >> >
> >> > The reason I repeated to reset p->max to zero in the loop is to avoid
> >> > using lockdep annotation(maybe, spin_lock_nested(something) to prevent
> >> > false positive.
> >> >
> >> > diff --git a/mm/swapfile.c b/mm/swapfile.c
> >> > index 42fe5653814a..9ce007a42bbc 100644
> >> > --- a/mm/swapfile.c
> >> > +++ b/mm/swapfile.c
> >> > @@ -2644,6 +2644,19 @@ SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
> >> > swap_file = p->swap_file;
> >> > old_block_size = p->old_block_size;
> >> > p->swap_file = NULL;
> >> > +
> >> > + if (p->flags & SWP_SOLIDSTATE) {
> >> > + unsigned long ci, nr_cluster;
> >> > +
> >> > + nr_cluster = DIV_ROUND_UP(p->max, SWAPFILE_CLUSTER);
> >> > + for (ci = 0; ci < nr_cluster; ci++) {
> >> > + struct swap_cluster_info *sci;
> >> > +
> >> > + sci = lock_cluster(p, ci * SWAPFILE_CLUSTER);
> >> > + p->max = 0;
> >> > + unlock_cluster(sci);
> >> > + }
> >> > + }
> >> > p->max = 0;
> >> > swap_map = p->swap_map;
> >> > p->swap_map = NULL;
> >> > @@ -3369,10 +3382,10 @@ static int __swap_duplicate(swp_entry_t entry, unsigned char usage)
> >> > goto bad_file;
> >> > p = swap_info[type];
> >> > offset = swp_offset(entry);
> >> > - if (unlikely(offset >= p->max))
> >> > - goto out;
> >> >
> >> > ci = lock_cluster_or_swap_info(p, offset);
> >> > + if (unlikely(offset >= p->max))
> >> > + goto unlock_out;
> >> >
> >> > count = p->swap_map[offset];
> >> >
> >>
> >> Sorry, this doesn't work, because
> >>
> >> lock_cluster_or_swap_info()
> >>
> >> Need to read p->cluster_info, which may be freed during swapoff too.
> >>
> >>
> >> To reduce the added overhead in regular code path, Maybe we can use SRCU
> >> to implement get_swap_device() and put_swap_device()? There is only
> >> increment/decrement on CPU local variable in srcu_read_lock/unlock().
> >> Should be acceptable in not so hot swap path?
> >>
> >> This needs to select CONFIG_SRCU if CONFIG_SWAP is enabled. But I guess
> >> that should be acceptable too?
> >>
> >
> > Why do we need srcu here? Is it enough with rcu like below?
> >
> > It might have a bug/room to be optimized about performance/naming.
> > I just wanted to show my intention.
>
> Yes. rcu should work too. But if we use rcu, it may need to be called
> several times to make sure the swap device under us doesn't go away, for
> example, when checking si->max in __swp_swapcount() and

I think it's not a big concern performance pov and benefit is good
abstraction through current locking function so we don't need much churn.

> add_swap_count_continuation(). And I found we need rcu to protect swap
> cache radix tree array too. So I think it may be better to use one

Could you elaborate it more about swap cache arrary problem?

> calling to srcu_read_lock/unlock() instead of multiple callings to
> rcu_read_lock/unlock().
>
> Best Regards,
> Huang, Ying
>
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@kvack.org. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

\
 
 \ /
  Last update: 2017-12-08 10:11    [W:0.172 / U:0.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site