Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <572cfcc0-197a-9ead-9cb-3c5bf5e735@google.com> <Y6amzxU7choHAXWi@infradead.org>
 <c67eba0-fc5a-4ad0-971-cf80bc1c6e5a@google.com>
In-Reply-To: <c67eba0-fc5a-4ad0-971-cf80bc1c6e5a@google.com>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Sat, 24 Dec 2022 08:56:23 -0800
Message-ID: <CAHk-=wjVw9=Hio5pRfGW45+yL-geWfNGyPeSFess3FAZQwVJrg@mail.gmail.com>
Subject: Re: 6.2 nvme-pci: something wrong
To:     Hugh Dickins <hughd@google.com>
Cc:     Christoph Hellwig <hch@infradead.org>,
        Jens Axboe <axboe@kernel.dk>, Keith Busch <kbusch@kernel.org>,
        Sagi Grimberg <sagi@grimberg.me>,
        Chaitanya Kulkarni <kch@nvidia.com>,
        Thorsten Leemhuis <regressions@leemhuis.info>,
        linux-block@vger.kernel.org, linux-nvme@lists.infradead.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Sat, Dec 24, 2022 at 2:19 AM Hugh Dickins <hughd@google.com> wrote:
>
> Regarding the awful 0's based queue depth: yes, it just looked to me
> as if the way that got handled in pci.c before differed from the way
> it gets handled in pci.c and core.c now, one too many "+ 1"s or "- 1"s
> somewhere.

The commit in question seems to replace nvme_pci_alloc_tag_set() calls
with nvme_alloc_io_tag_set(), and that has a big difference in how
queue_depth is set.

It used to do (in nnvme_pci_alloc_tag_set()):

        set->queue_depth = min_t(unsigned, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;

but now it does (in nvme_alloc_io_tag_set())

        set->queue_depth = ctrl->sqsize + 1;

instead.

So that "set->queue_depth" _seems_ to have historically had that "-1"
(that "zero means one" that apparently sqsize also has), but the new
code basically undoes it.

I don't know the code at all, but this does all seem to be a change
(and *very* confusing).

The fact that Hugh gets it to work by doint that

        set->queue_depth = ctrl->sqsize;

does seem to match the whole "it used to subtract one" behavior it
had. Which is why I assume Hugh tried that patch in the first place.

             Linus