Date:   Mon, 20 Feb 2023 20:33:35 +0000
From:   Matthew Wilcox <willy@infradead.org>
To:     Danilo Krummrich <dakr@redhat.com>
Cc:     matthew.brost@intel.com, dri-devel@lists.freedesktop.org,
        corbet@lwn.net, nouveau@lists.freedesktop.org, ogabbay@kernel.org,
        linux-doc@vger.kernel.org, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org, boris.brezillon@collabora.com,
        bskeggs@redhat.com, tzimmermann@suse.de, Liam.Howlett@oracle.com,
        bagasdotme@gmail.com, christian.koenig@amd.com,
        jason@jlekstrand.net
Subject: Re: [PATCH drm-next v2 04/16] maple_tree: add flag MT_FLAGS_LOCK_NONE
Message-ID: <Y/PZH/q2Xsr3od9m@casper.infradead.org>
References: <20230217134422.14116-1-dakr@redhat.com>
 <20230217134422.14116-5-dakr@redhat.com>
 <Y+/Xn11dfdn7SfBD@casper.infradead.org>
 <3bb02ec3-4d19-9135-cabc-26ed210f7396@redhat.com>
 <Y/ONYhyDCPEYH1ml@casper.infradead.org>
 <e43f6acc-175d-1031-c4a2-67a6f1741866@redhat.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <e43f6acc-175d-1031-c4a2-67a6f1741866@redhat.com>
Precedence: bulk

On Mon, Feb 20, 2023 at 06:06:03PM +0100, Danilo Krummrich wrote:
> On 2/20/23 16:10, Matthew Wilcox wrote:
> > This is why we like people to use the spinlock embedded in the tree.
> > There's nothing for the user to care about.  If the access really is
> > serialised, acquiring/releasing the uncontended spinlock is a minimal
> > cost compared to all the other things that will happen while modifying
> > the tree.
> 
> I think as for the users of the GPUVA manager we'd have two cases:
> 
> 1) Accesses to the manager (and hence the tree) are serialized, no lock
> needed.
> 
> 2) Multiple operations on the tree must be locked in order to make them
> appear atomic.

Could you give an example here of what you'd like to do?  Ideally
something complicated so I don't say "Oh, you can just do this" when
there's a more complex example for which "this" won't work.  I'm sure
that's embedded somewhere in the next 20-odd patches, but it's probably
quicker for you to describe in terms of tree operations that have to
appear atomic than for me to try to figure it out.

> In either case the embedded spinlock wouldn't be useful, we'd either need an
> external lock or no lock at all.
> 
> If there are any internal reasons why specific tree operations must be
> mutually excluded (such as those you explain below), wouldn't it make more
> sense to always have the internal lock and, optionally, allow users to
> specify an external lock additionally?

So the way this works for the XArray, which is a little older than the
Maple tree, is that we always use the internal spinlock for
modifications (possibly BH or IRQ safe), and if someone wants to
use an external mutex to make some callers atomic with respect to each
other, they're free to do so.  In that case, the XArray doesn't check
the user's external locking at all, because it really can't know.

I'd advise taking that approach; if there's really no way to use the
internal spinlock to make your complicated updates appear atomic
then just let the maple tree use its internal spinlock, and you can
also use your external mutex however you like.