Received-SPF: pass (google.com: domain of linux-kernel+bounces-45027-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Date: Tue, 30 Jan 2024 17:15:48 +0000
In-Reply-To: <20240130154632.GB772725@cmpxchg.org>
Precedence: bulk
Mime-Version: 1.0
References: <20240130014208.565554-1-hannes@cmpxchg.org> <ZbiwW5BJhFeGc2Bd@google.com>
 <20240130154632.GB772725@cmpxchg.org>
Message-ID: <ZbkuxJpoHrUF-ULa@google.com>
Subject: Re: [PATCH 00/20] mm: zswap: cleanups
From: Yosry Ahmed <yosryahmed@google.com>
To: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrew Morton <akpm@linux-foundation.org>, Nhat Pham <nphamcs@gmail.com>, 
	Chengming Zhou <zhouchengming@bytedance.com>, linux-mm@kvack.org, 
	linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="us-ascii"

On Tue, Jan 30, 2024 at 10:46:32AM -0500, Johannes Weiner wrote:
> On Tue, Jan 30, 2024 at 08:16:27AM +0000, Yosry Ahmed wrote:
> > Hey Johannes,
> > 
> > On Mon, Jan 29, 2024 at 08:36:36PM -0500, Johannes Weiner wrote:
> > > Cleanups and maintenance items that accumulated while reviewing zswap
> > > patches. Based on akpm/mm-unstable + the UAF fix I sent just now.
> > 
> > Patches 1 to 9 LGTM, thanks for the great cleanups!
> > 
> > I am less excited about patches 10 to 20 though. Don't get me wrong, I
> > am all of logically ordering the code. However, it feels like in this
> > case, we will introduce unnecessary layers in the git history in a lot
> > of places where I find myself checking the history regularly.
> > Personally, I tend to jump around the file using vim search or using a
> > cscope extension to find references/definitions, so I don't feel a need
> > for such reordering.
> 
> I agree that sweeping non-functional changes can be somewhat
> painful. However, moves are among the easiest of those because the
> code itself doesn't change. git log is not really affected, git blame
> <ref-just-before-move> mm/zswap.c works as well.

IIUC with git blame -L <lines> <ref-just-before-move> won't really work
because the lines will have changed significantly. In the future, going
through several layers of git blame will be less convenient because the
function moving will invalidate the lines.

> Backports are easy to
> adjust and verify - mostly, patch will just warn about line offsets.

I usually use git cherry-pick, and it sometimes gets confused if things
are moved far enough IIRC.

> 
> What motivated this was figuring out the writeback/swapoff race. I
> also use search and multiple buffers in my editor, but keeping track
> of the dependencies between shrink_memcg_cb, zswap_writeback_entry and
> third places like load and invalidate became quite unwieldy. There is
> also the search in the logical direction not finding things, or mostly
> unrelated stuff. It's just less error prone to read existing code and
> write new code if related layers are grouped together and the order is
> logical, despite having those tools.
> 
> The problem will also only get worse if there are no natural anchor
> points for new code, and the layering isn't clear. The new shrinker
> code is a case in point. We missed the opportunity in the memcg
> codebase, and I've regretted it for years. It just resulted in more
> fragile, less readable and debuggable code. The zswap code has been
> stagnant in the last few years, and there are relatively few commits
> behind us (git log --pretty=format:"%as %h %s" mm/zswap.c). I figure
> now is a good chance, before the more major changes we have planned.

I agree that if we want to do it, it's much better now than later, just
presenting a differet perspective. No strong feelings.