mm: pin address_space before dereferencing it while isolating an LRU page
authorMel Gorman <mgorman@techsingularity.net>
Thu, 1 Feb 2018 00:19:52 +0000 (16:19 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 1 Feb 2018 01:18:39 +0000 (17:18 -0800)
commit69d763fc6d3aee787a3e8c8c35092b4f4960fa5d
treed4ef12c68e16fb5ef0febc788d50b6c0e51a2bb6
parent284cd241a18ee6d999296f8ff3104eb6d2fc898f
mm: pin address_space before dereferencing it while isolating an LRU page

Minchan Kim asked the following question -- what locks protects
address_space destroying when race happens between inode trauncation and
__isolate_lru_page? Jan Kara clarified by describing the race as follows

CPU1                                            CPU2

truncate(inode)                                 __isolate_lru_page()
  ...
  truncate_inode_page(mapping, page);
    delete_from_page_cache(page)
      spin_lock_irqsave(&mapping->tree_lock, flags);
        __delete_from_page_cache(page, NULL)
          page_cache_tree_delete(..)
            ...                                   mapping = page_mapping(page);
            page->mapping = NULL;
            ...
      spin_unlock_irqrestore(&mapping->tree_lock, flags);
      page_cache_free_page(mapping, page)
        put_page(page)
          if (put_page_testzero(page)) -> false
- inode now has no pages and can be freed including embedded address_space

                                                  if (mapping && !mapping->a_ops->migratepage)
- we've dereferenced mapping which is potentially already free.

The race is theoretically possible but unlikely.  Before the
delete_from_page_cache, truncate_cleanup_page is called so the page is
likely to be !PageDirty or PageWriteback which gets skipped by the only
caller that checks the mappping in __isolate_lru_page.  Even if the race
occurs, a substantial amount of work has to happen during a tiny window
with no preemption but it could potentially be done using a virtual
machine to artifically slow one CPU or halt it during the critical
window.

This patch should eliminate the race with truncation by try-locking the
page before derefencing mapping and aborting if the lock was not
acquired.  There was a suggestion from Huang Ying to use RCU as a
side-effect to prevent mapping being freed.  However, I do not like the
solution as it's an unconventional means of preserving a mapping and
it's not a context where rcu_read_lock is obviously protecting rcu data.

Link: http://lkml.kernel.org/r/20180104102512.2qos3h5vqzeisrek@techsingularity.net
Fixes: c82449352854 ("mm: compaction: make isolate_lru_page() filter-aware again")
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Acked-by: Minchan Kim <minchan@kernel.org>
Cc: "Huang, Ying" <ying.huang@intel.com>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/vmscan.c