Re: [PATCH v2] locking/rwsem: add acquire barrier to read_slowpath exit when queue is empty

[Will Deacon]

On Wed, Jul 17, 2019 at 02:02:20PM +0200, Jan Stancek wrote:
> 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
> 
> 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)
> 
> 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>
> 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 | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
> index 37524a47f002..5ac72b60608b 100644
> --- a/kernel/locking/rwsem.c
> +++ b/kernel/locking/rwsem.c
> @@ -1032,6 +1032,7 @@ static inline bool rwsem_reader_phase_trylock(struct rw_semaphore *sem,
>  		 */
>  		if (adjustment && !(atomic_long_read(&sem->count) &
>  		     (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
> +			smp_acquire__after_ctrl_dep();
>  			raw_spin_unlock_irq(&sem->wait_lock);
>  			rwsem_set_reader_owned(sem);
>  			lockevent_inc(rwsem_rlock_fast);

If you add a comment to the code outlining the issue (preferably as a litmus
test involving sem->count and some shared data which happens to be
vmacache_seqnum in your test)), then:

Reviewed-by: Will Deacon <will@xxxxxxxxxx>

Thanks,

Will