[PATCH 04/47] block: provide a new BLK_EH_QUIESCED timeout return value

[Jeff Moyer]

Christoph Hellwig <hch at lst.de> writes:

> We always set REQ_ATOM_COMPLETE before calling into ->timeout and thus
> even having a chance to REQ_ATOM_QUIESCED.  Maybe we're talking past
> each other, so if it feels like I'm off track try to explain the
> race in a bit more detail.

Sure.

CPU 0 executes the following:

blk_rq_check_expired():
        if (!blk_mark_rq_complete(rq))  // this succeeds, so we call blk_rq_timed_out
                blk_rq_timed_out(rq);
blk_rq_timed_out():
        if (q->rq_timed_out_fn)
                ret = q->rq_timed_out_fn(req);
        switch (ret) {  // QUIESCED is returned
...
        case BLK_EH_QUIESCED:
                set_bit(REQ_ATOM_QUIESCED, &req->atomic_flags);
                break;

CPU 1 takes an interrupt for the completion of the same request:

blk_complete_request():
        if (!blk_mark_rq_complete(req) ||  // this fails, as it's already marked complete
            test_and_clear_bit(REQ_ATOM_QUIESCED, &req->atomic_flags))  // this succeeds
                __blk_complete_request(req); // so we complete the request

Later, after the adapter reset is finished, you mentioned that all
requests will be completed.  My question is: will this result in a
double completion for this particular request?

I hope that's more clear.

-Jeff