Re: [PATCH v2 5/5] kcsan, seqlock: Fix incorrect assumption in read_seqbegin()

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Nov 04, 2024 at 04:43:09PM +0100, Marco Elver wrote:
> During testing of the preceding changes, I noticed that in some cases,
> current->kcsan_ctx.in_flat_atomic remained true until task exit. This is
> obviously wrong, because _all_ accesses for the given task will be
> treated as atomic, resulting in false negatives i.e. missed data races.
> 
> Debugging led to fs/dcache.c, where we can see this usage of seqlock:
> 
> 	struct dentry *d_lookup(const struct dentry *parent, const struct qstr *name)
> 	{
> 		struct dentry *dentry;
> 		unsigned seq;
> 
> 		do {
> 			seq = read_seqbegin(&rename_lock);
> 			dentry = __d_lookup(parent, name);
> 			if (dentry)
> 				break;
> 		} while (read_seqretry(&rename_lock, seq));
> 	[...]


How's something like this completely untested hack?


	struct dentry *dentry;

	read_seqcount_scope (&rename_lock) {
		dentry = __d_lookup(parent, name);
		if (dentry)
			break;
	}


But perhaps naming isn't right, s/_scope/_loop/ ?


--- a/include/linux/seqlock.h
+++ b/include/linux/seqlock.h
@@ -829,6 +829,33 @@ static inline unsigned read_seqretry(con
 	return read_seqcount_retry(&sl->seqcount, start);
 }
 
+
+static inline unsigned read_seq_scope_begin(const struct seqlock_t *sl)
+{
+	unsigned ret = read_seqcount_begin(&sl->seqcount);
+	kcsan_atomic_next(0);
+	kcsan_flat_atomic_begin();
+	return ret;
+}
+
+static inline void read_seq_scope_end(unsigned *seq)
+{
+	kcsan_flat_atomic_end();
+}
+
+static inline bool read_seq_scope_retry(const struct seqlock_t *sl, unsigned *seq)
+{
+	bool done = !read_seqcount_retry(&sl->seqcount, *seq);
+	if (!done)
+		*seq = read_seqcount_begin(&sl->seqcount);
+	return done;
+}
+
+#define read_seqcount_scope(sl) \
+	for (unsigned seq __cleanup(read_seq_scope_end) =		\
+			read_seq_scope_begin(sl), done = 0;		\
+	     !done; done = read_seq_scope_retry(sl, &seq))
+
 /*
  * For all seqlock_t write side functions, use the internal
  * do_write_seqcount_begin() instead of generic write_seqcount_begin().