I'm also worried about the other two versions, though: memory-barriers.txt#1724:
1724 (*) The compiler is within its rights to invent stores to a variable, i.e. the compiler is free to decide __bio_chain_endio looks like this: static struct bio *__bio_chain_endio(struct bio *bio) { struct bio *parent = bio->bi_private; blk_status_t tmp = parent->bi_status; parent->bi_status = bio->bi_status; if (!bio->bi_status) parent->bi_status = tmp; bio_put(bio); return parent; } In which case, the read and later store on the two different threads may overlap in such a way that bio_endio sometimes sees success, even if one child had an error. As a result, I believe the setting of parent->bi_status needs to be a WRITE_ONCE() and the later reading needs to be a READ_ONCE() [although, since the later reading happens in many different functions, perhaps some other barrier to make sure all readers get the correct value is in order.]