Re: [PATCH] seqlock: fix raw_read_seqcount_latch()

[Peter Zijlstra]

On Sat, May 21, 2016 at 11:14:49PM +0300, Alexey Dobriyan wrote:
> lockless_dereference() is supposed to take pointer not integer.
> 
> Signed-off-by: Alexey Dobriyan <adobriyan@xxxxxxxxx>
> ---
> 
>  include/linux/seqlock.h |    4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> --- a/include/linux/seqlock.h
> +++ b/include/linux/seqlock.h
> @@ -277,7 +277,7 @@ static inline void raw_write_seqcount_barrier(seqcount_t *s)
>  
>  static inline int raw_read_seqcount_latch(seqcount_t *s)
>  {
> -	return lockless_dereference(s->sequence);
> +	return lockless_dereference(s)->sequence;
>  }
>  
>  /**
> @@ -331,7 +331,7 @@ static inline int raw_read_seqcount_latch(seqcount_t *s)
>   *	unsigned seq, idx;
>   *
>   *	do {
> - *		seq = lockless_dereference(latch->seq);
> + *		seq = lockless_dereference(latch)->seq;
>   *
>   *		idx = seq & 0x01;
>   *		entry = data_query(latch->data[idx], ...);

So while the code was dubious; I it is now wrong, but my head hurts.

I'll queue the below, TJs per-cpu change and the lockless_dereference()
void * cast trick.


---
Subject: seqcount: Re-fix raw_read_seqcount_latch()

Commit 50755bc1c305 ("seqlock: fix raw_read_seqcount_latch()") broke
raw_read_seqcount_latch().

If you look at the comment that was modified; the thing that changes is
the seq count, not the latch pointer.

 * void latch_modify(struct latch_struct *latch, ...)
 * {
 *	smp_wmb();	<- Ensure that the last data[1] update is visible
 *	latch->seq++;
 *	smp_wmb();	<- Ensure that the seqcount update is visible
 *
 *	modify(latch->data[0], ...);
 *
 *	smp_wmb();	<- Ensure that the data[0] update is visible
 *	latch->seq++;
 *	smp_wmb();	<- Ensure that the seqcount update is visible
 *
 *	modify(latch->data[1], ...);
 * }
 *
 * The query will have a form like:
 *
 * struct entry *latch_query(struct latch_struct *latch, ...)
 * {
 *	struct entry *entry;
 *	unsigned seq, idx;
 *
 *	do {
 *		seq = lockless_dereference(latch->seq);

So here we have:

		seq = READ_ONCE(latch->seq);
		smp_read_barrier_depends();

Which is exactly what we want; the new code:

		seq = ({ p = READ_ONCE(latch);
			 smp_read_barrier_depends(); p })->seq;

is just wrong; because it looses the volatile read on seq, which can now
be torn or worse 'optimized'. And the read_depend barrier is also placed
wrong, we want it after the load of seq, to match the above data[]
up-to-date wmb()s.

Such that when we dereference latch->data[] below, we're guaranteed to
observe the right data.

 *
 *		idx = seq & 0x01;
 *		entry = data_query(latch->data[idx], ...);
 *
 *		smp_rmb();
 *	} while (seq != latch->seq);
 *
 *	return entry;
 * }

So yes, not passing a pointer is not pretty, but the code was correct,
and isn't anymore now.

Change to explicit READ_ONCE()+smp_read_barrier_depends() to avoid
confusion and allow strict lockless_dereference() checking.

Fixes: 50755bc1c305 ("seqlock: fix raw_read_seqcount_latch()")
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
 include/linux/seqlock.h | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/linux/seqlock.h b/include/linux/seqlock.h
index 7973a821ac58..f3db247cebc8 100644
--- a/include/linux/seqlock.h
+++ b/include/linux/seqlock.h
@@ -277,7 +277,9 @@ static inline void raw_write_seqcount_barrier(seqcount_t *s)
 
 static inline int raw_read_seqcount_latch(seqcount_t *s)
 {
-	return lockless_dereference(s)->sequence;
+	int seq = READ_ONCE(s->sequence);
+	smp_read_barrier_depends();
+	return seq;
 }
 
 /**
@@ -331,7 +333,7 @@ static inline int raw_read_seqcount_latch(seqcount_t *s)
  *	unsigned seq, idx;
  *
  *	do {
- *		seq = lockless_dereference(latch)->seq;
+ *		seq = raw_read_seqcount_latch(&latch->seq);
  *
  *		idx = seq & 0x01;
  *		entry = data_query(latch->data[idx], ...);