Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
MIME-Version: 1.0
References: <70cda2a3-f246-d45b-f600-1f9d15ba22ff@gmail.com>
 <87eflmpqkb.fsf@notabene.neil.brown.name> <20190222211006.GA10987@redhat.com>
 <7f0aeb7b-fdaa-0625-f785-05c342047550@kernel.dk> <20190222235459.GA11726@redhat.com>
In-Reply-To: <20190222235459.GA11726@redhat.com>
From:   John Dorminy <jdorminy@redhat.com>
Date:   Fri, 22 Feb 2019 21:02:06 -0500
Message-ID: <CAMeeMh-2ANOr_Sb66EyA_HULkVRudD7fyOZsDbpRpDrshwnR2w@mail.gmail.com>
Subject: Re: block: be more careful about status in __bio_chain_endio
To:     Mike Snitzer <snitzer@redhat.com>
Cc:     Jens Axboe <axboe@kernel.dk>, NeilBrown <neilb@suse.com>,
        linux-block@vger.kernel.org,
        device-mapper development <dm-devel@redhat.com>,
        Milan Broz <gmazyland@gmail.com>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

I am perhaps not understanding the intricacies here, or not seeing a
barrier protecting it, so forgive me if I'm off base. I think reading
parent->bi_status here is unsafe.
Consider the following sequence of events on two threads.

Thread 0                                 Thread 1
In __bio_chain_endio:                    In __bio_chain_endio:
[A] Child 0 reads parent->bi_status,
    no error.
                                         Child bio 1 reads parent, no error seen
                                         It sets parent->bi_status to an error
                                         It calls bio_put.
Child bio 0 calls bio_put
[end __bio_chain_endio]                  [end __bio_chain_endio]
                                         In bio_chain_endio(), bio_endio(parent)
                                         is called, calling bio_remaining_done()
                                         which decrements __bi_remaining to 1
                                         and returns false, so no further endio
                                         stuff is done.
In bio_chain_endio(), bio_endio(parent)
is called, calling bio_remaining_done(),
decrementing parent->__bi_remaining to
 0, and continuing to finish parent.
Either for block tracing or for parent's
bi_end_io(), this thread tries to read
parent->bi_status again.

The compiler or the CPU may cache the read from [A], and since there
are no intervening barriers, parent->bi_status is still believed on
thread 0 to be success. Thus the bio may still be falsely believed to
have completed successfully, even though child 1 set an error in it.

Am I missing a subtlety here?