[PATCH v3] locking/rwsem: add acquire barrier to read_slowpath exit when queue is empty

[Jan Stancek]

LTP mtest06 has been observed to rarely hit "still mapped when deleted"
and following BUG_ON on arm64:
  page:ffff7e02fa37e480 refcount:3 mapcount:1 mapping:ffff80be3d678ab0 index:0x0
  xfs_address_space_operations [xfs]
  flags: 0xbfffe000000037(locked|referenced|uptodate|lru|active)
  page dumped because: VM_BUG_ON_PAGE(page_mapped(page))
  ------------[ cut here ]------------
  kernel BUG at mm/filemap.c:171!
  Internal error: Oops - BUG: 0 [#1] SMP
  CPU: 220 PID: 154292 Comm: mmap1 Not tainted 5.2.0-0ecfebd.cki #1
  Hardware name: HPE Apollo 70 /C01_APACHE_MB , BIOS L50_5.13_1.10 05/17/2019
  pstate: 40400089 (nZcv daIf +PAN -UAO)
  pc : unaccount_page_cache_page+0x17c/0x1a0
  lr : unaccount_page_cache_page+0x17c/0x1a0
  Call trace:
  unaccount_page_cache_page+0x17c/0x1a0
  delete_from_page_cache_batch+0xa0/0x300
  truncate_inode_pages_range+0x1b8/0x640
  truncate_inode_pages_final+0x88/0xa8
  evict+0x1a0/0x1d8
  iput+0x150/0x240
  dentry_unlink_inode+0x120/0x130
  __dentry_kill+0xd8/0x1d0
  dentry_kill+0x88/0x248
  dput+0x168/0x1b8
  __fput+0xe8/0x208
  ____fput+0x20/0x30
  task_work_run+0xc0/0xf0
  do_notify_resume+0x2b0/0x328
  work_pending+0x8/0x10

The extra mapcount originated from pagefault handler, which handled
pagefault for vma that has already been detached. vma is detached
under mmap_sem write lock by detach_vmas_to_be_unmapped(), which
also invalidates vmacache.

When pagefault handler (under mmap_sem read lock) called find_vma(),
vmacache_valid() wrongly reported vmacache as valid.

After rwsem down_read() returns via 'queue empty' path (as of v5.2),
it does so without issuing read_acquire on sem->count:
  down_read
    __down_read
      rwsem_down_read_failed
        __rwsem_down_read_failed_common
          raw_spin_lock_irq(&sem->wait_lock);
          if (list_empty(&sem->wait_list)) {
            if (atomic_long_read(&sem->count) >= 0) {
              raw_spin_unlock_irq(&sem->wait_lock);
              return sem;

Suspected problem here is that last *_acquire on down_read() side
happens before write side issues *_release:
  1. writer: has the lock
  2. reader: down_read() issues *read_acquire on entry
  3. writer: mm->vmacache_seqnum++; downgrades lock (*fetch_add_release)
  4. reader: __rwsem_down_read_failed_common() finds it can take lock and returns
  5. reader: observes stale mm->vmacache_seqnum

----------------------------------- 8< ------------------------------------
C rwsem

{
	atomic_t rwsem_count = ATOMIC_INIT(1);
	int vmacache_seqnum = 10;
}

P0(int *vmacache_seqnum, atomic_t *rwsem_count)
{
	r0 = READ_ONCE(*vmacache_seqnum);
	WRITE_ONCE(*vmacache_seqnum, r0 + 1);
	/* downgrade_write */
	r1 = atomic_fetch_add_release(-1+256, rwsem_count);
}

P1(int *vmacache_seqnum, atomic_t *rwsem_count, spinlock_t *sem_wait_lock)
{
	/* rwsem_read_trylock */
	r0 = atomic_add_return_acquire(256, rwsem_count);
	/* rwsem_down_read_slowpath */
	spin_lock(sem_wait_lock);
	r0 = atomic_read(rwsem_count);
	if ((r0 & 1) == 0) {
		// BUG: needs barrier
		spin_unlock(sem_wait_lock);
		r1 = READ_ONCE(*vmacache_seqnum);
	}
}
exists (1:r1=10)
----------------------------------- >8 ------------------------------------

I can reproduce the problem by running LTP mtest06 in a loop and building
kernel (-j $NCPUS) in parallel. It does reproduce since v4.20 up to v5.2
on arm64 HPE Apollo 70 (224 CPUs, 256GB RAM, 2 nodes). It triggers reliably
within ~hour. Patched kernel ran fine for 10+ hours with clean dmesg.
Tests were done against v5.2, since commit cf69482d62d9 ("locking/rwsem:
Enable readers spinning on writer") makes it much harder to reproduce.

v2: Move barrier after test (Waiman Long)
    Use smp_acquire__after_ctrl_dep() (Peter Zijlstra)
v3: Add comment to barrier (Waiman Long, Will Deacon)
    Add litmus test

Related: https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/mem/mtest06/mmap1.c
Related: commit dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")
Fixes: 4b486b535c33 ("locking/rwsem: Exit read lock slowpath if queue empty & no writer")

Signed-off-by: Jan Stancek <jstancek@xxxxxxxxxx>
Reviewed-by: Will Deacon <will@xxxxxxxxxx>
Acked-by: Waiman Long <longman@xxxxxxxxxx>
Cc: stable@xxxxxxxxxxxxxxx # v4.20+
Cc: Waiman Long <longman@xxxxxxxxxx>
Cc: Davidlohr Bueso <dbueso@xxxxxxx>
Cc: Will Deacon <will@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Ingo Molnar <mingo@xxxxxxxxxx>
---
 kernel/locking/rwsem.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
index 37524a47f002..fe02aef39e9d 100644
--- a/kernel/locking/rwsem.c
+++ b/kernel/locking/rwsem.c
@@ -1032,6 +1032,13 @@ static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
 		 */
 		if (adjustment && !(atomic_long_read(&sem->count) &
 		     (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
+			/*
+			 * Add an acquire barrier here to make sure no stale
+			 * data acquired before the above test, where the writer
+			 * may still be holding the lock, will be reused in the
+			 * reader critical section.
+			 */
+			smp_acquire__after_ctrl_dep();
 			raw_spin_unlock_irq(&sem->wait_lock);
 			rwsem_set_reader_owned(sem);
 			lockevent_inc(rwsem_rlock_fast);
-- 
1.8.3.1