[PATCH] mm/migrate: fix race between lock page and clear PG_Isolated

[Hugh Dickins]

On Tue, 15 Mar 2022, David Hildenbrand wrote:
> On 15.03.22 05:21, Andrew Morton wrote:
> > On Tue, 15 Mar 2022 11:05:15 +0800 Andrew Yang <andrew.yang at mediatek.com> wrote:
> > 
> >> When memory is tight, system may start to compact memory for large
> >> continuous memory demands. If one process tries to lock a memory page
> >> that is being locked and isolated for compaction, it may wait a long time
> >> or even forever. This is because compaction will perform non-atomic
> >> PG_Isolated clear while holding page lock, this may overwrite PG_waiters
> >> set by the process that can't obtain the page lock and add itself to the
> >> waiting queue to wait for the lock to be unlocked.
> >>
> >> CPU1                            CPU2
> >> lock_page(page); (successful)
> >>                                 lock_page(); (failed)
> >> __ClearPageIsolated(page);      SetPageWaiters(page) (may be overwritten)
> >> unlock_page(page);
> >>
> >> The solution is to not perform non-atomic operation on page flags while
> >> holding page lock.
> > 
> > Sure, the non-atomic bitop optimization is really risky and I suspect
> > we reach for it too often.  Or at least without really clearly
> > demonstrating that it is safe, and documenting our assumptions.
> 
> I agree. IIRC, non-atomic variants are mostly only safe while the
> refcount is 0. Everything else is just absolutely fragile.

It is normal and correct to use __SetPageFlag(page) on a page just allocated
from the buddy, and not yet logically visible to others: that has refcount 1.

Of course, it might have refcount 2 or more, through being speculatively
visible to get_page_unless_zero() users: perhaps through earlier usage of
the same struct page, or by physical scan of memmap.

Those few such others - compaction's isolate_migratepages_block() is the
one I know best - must be very careful in their sequence of operations.

Preliminary read-only checks are usually okay (but some VM_BUG_ON_PGFLAGS
are increasingly problematic: I've had to turn off that CONFIG), then
get_page_unless_zero(), then read-only check that the page is of the
manageable kind (PageLRU in my world), and only then can it be safe to
lock the page - which of course touches page flags, and so would be
problematic for a racing user's __SetPageFlag(page).

But PageMovable and PageIsolated are beyond my ken: I can't say there.

Hugh