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: <87edkce118.wl-tiwai@suse.de> <20230809143801.GA693@lst.de>
 <CAHk-=wiyWOaPtOJ1PTdERswXV9m7W_UkPV-HE0kbpr48mbnrEA@mail.gmail.com> <87wmy4ciap.wl-tiwai@suse.de>
In-Reply-To: <87wmy4ciap.wl-tiwai@suse.de>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Wed, 9 Aug 2023 10:01:50 -0700
Message-ID: <CAHk-=wh-mUL6mp4chAc6E_UjwpPLyCPRCJK+iB4ZMD2BqjwGHA@mail.gmail.com>
Subject: Re: [PATCH RFC] Introduce uniptr_t as a generic "universal" pointer
To:     Takashi Iwai <tiwai@suse.de>
Cc:     Christoph Hellwig <hch@lst.de>, linux-kernel@vger.kernel.org,
        Andy Shevchenko <andriy.shevchenko@linux.intel.com>,
        Mark Brown <broonie@kernel.org>, netdev@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, 9 Aug 2023 at 09:05, Takashi Iwai <tiwai@suse.de> wrote:
>
> OTOH, it simplifies the code well for us; as of now, we have two
> callbacks for copying PCM memory from/to the device, distinct for
> kernel and user pointers.  It's basically either copy_from_user() or
> memcpy() of the given size depending on the caller.  The sockptr_t or
> its variant would allow us to unify those to a single callback.

I didn't see the follow-up patches that use this, but...

> (And yeah, iov_iter is there, but it's definitely overkill for the
> purpose.)

You can actually use a "simplified form" of iov_iter, and it's not all that bad.

If the actual copying operation is just a memcpy, you're all set: just
do copy_to/from_iter(), and it's a really nice interface, and you
don't have to carry "ptr+size" things around.

And we now have a simple way to generate simple iov_iter's, so
*creating* the iter is trivial too:

        struct iov_iter iter;
        int ret = import_ubuf(ITER_SRC/DEST, uptr, len, &iter);

        if (unlikely(ret < 0))
                return ret;

and you're all done. You can now pass '&iter' around, and it has a
nice user pointer and a range in it, and copying that thing is easy.

Perhaps somewhat strangely (*) we don't have the same for a simple
kernel buffer, but adding that wouldn't be hard. You either end up
using a 'kvec', or we could even add something like ITER_KBUF if it
really matters.

Right now the kernel buffer init is a *bit* more involved than the
above ubuf case:

        struct iov_iter iter;
        struct kvec kvec = { kptr, len};

        iov_iter_kvec(&iter, ITER_SRC/DEST, &kvec, 1, len);

and that's maybe a *bit* annoying, but we could maybe simplify this
with some helper macros even without ITER_KBUF.

So yes, iov_iter does have some abstraction overhead, but it really
isn't that bad. And it *does* allow you to do a lot of things, and can
actually simplify the users quite a bit, exactly because it allows you
to just pass that single iter pointer around, and you automatically
have not just the user/kernel distinction, you have the buffer size,
and you have a lot of helper functions to use it.

I really think that if you want a user-or-kernel buffer interface, you
should use these things.

Please? At least look into it.

                 Linus

(*) Well, not so strange - we've just never needed it.