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: <20190314111637.20236-1-hans@owltronix.com> <103c19f4-02a2-f54c-13f8-30bc5447426e@intel.com>
 <FE25456B-A63E-4601-960D-96F519250B8C@javigon.com> <CAJbgVnU2RXK4ci79n0szwTpZ+UB2zx977PHo1ie7h5y5kJrbhQ@mail.gmail.com>
In-Reply-To: <CAJbgVnU2RXK4ci79n0szwTpZ+UB2zx977PHo1ie7h5y5kJrbhQ@mail.gmail.com>
From:   Hans Holmberg <hans@owltronix.com>
Date:   Fri, 15 Mar 2019 09:34:14 +0100
Message-ID: <CANr-nt0DHudD9TFBqLzhgBuYz8GqVbwgiU7W2b5OVsLtpMuu_w@mail.gmail.com>
Subject: Re: [PATCH RFC] lightnvm: pblk: fix crash in pblk_end_partial_read
 due to multipage bvecs
To:     Heiner Litz <hlitz@ucsc.edu>
Cc:     =?UTF-8?Q?Javier_Gonz=C3=A1lez?= <javier@javigon.com>,
        =?UTF-8?Q?Matias_Bj=C3=B8rling?= <mb@lightnvm.io>,
        Klaus Birkelund Jensen <klaus.jensen@cnexlabs.com>,
        "Konopko, Igor J" <igor.j.konopko@intel.com>,
        linux-block@vger.kernel.org,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Hans Holmberg <hans.holmberg@cnexlabs.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Mar 14, 2019 at 5:15 PM Heiner Litz <hlitz@ucsc.edu> wrote:
>
> On Thu, Mar 14, 2019 at 7:16 AM Javier Gonz=C3=A1lez <javier@javigon.com>=
 wrote:
> >
> > > On 14 Mar 2019, at 06.49, Igor Konopko <igor.j.konopko@intel.com> wro=
te:
> > >
> > > While reading this patch, idea came to my mind - maybe it would be si=
mply better to get rid of partial read handling from pblk in current form a=
t all and use bio_split() & bio_chain() combination instead?
> > >
> > > This would definitely reduce a number of this 'kind of complicated' c=
ode inside pblk and let block layer help us with that. Even that the perfor=
mance of such a requests could be a little worse (few smaller IOs instead o=
f single vector IO), I believe that partial read is more a corner case, the=
n a typical use case, so maybe this would be a path to go.
> > >
> > > Let me know what you think about such an approach - I can make a patc=
h with that if you want.
>
> I like this idea as it will clean up a lot of code and get rid of the
> read_bitmap.
> Note that a single request can be turned into up to 32 requests if
> cached and non-cached sectors alternate, each one probably requiring
> an l2p lookup. I still think it's worth doing though.
> When you split, note that you have to release all line_refs which were
> acquired during l2p lookup.
>
> >
> > I agree with Igor.
> >

I agree as well, this path deserves some loving care and cleanup.
Let's just try to avoid read tail latency degradation.

Simplifying the code is a great first step, then we can optimize if needed =
be.

Thanks for the feedback y'all!


> > As I mentioned offline, we should fix this in a way that survives
> > further changes in struct bio; either making pblk handling visible to
> > the outside or rethinking the whole thing.
> >
> > Igor: If you can send a patch you mention, it would be great. I have
> > been trying the helper approach for some time, but it is too specific,
> > as fixing holes in the bvec breaks the bio advance-only semantics. Your
> > approach seems much better.
> >
> > Thanks,
> > Javier