Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20200904165216.1799796-1-hch@lst.de> <20200904165216.1799796-4-hch@lst.de>
 <20200904180617.GQ1236603@ZenIV.linux.org.uk> <20200904223518.GR1236603@ZenIV.linux.org.uk>
In-Reply-To: <20200904223518.GR1236603@ZenIV.linux.org.uk>
From:   Arnd Bergmann <arnd@arndb.de>
Date:   Mon, 7 Sep 2020 10:07:49 +0200
Message-ID: <CAK8P3a1doHskr=8uPrGetCAtD7QTRS5r=cfy6VJSoDsnE0=Aaw@mail.gmail.com>
Subject: Re: [PATCH 3/8] asm-generic: fix unaligned access hamdling in raw_copy_{from,to}_user
To:     Al Viro <viro@zeniv.linux.org.uk>
Cc:     Christoph Hellwig <hch@lst.de>,
        Paul Walmsley <paul.walmsley@sifive.com>,
        Palmer Dabbelt <palmer@dabbelt.com>,
        linux-riscv <linux-riscv@lists.infradead.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        linux-arch <linux-arch@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Sat, Sep 5, 2020 at 12:35 AM Al Viro <viro@zeniv.linux.org.uk> wrote:
>
> On Fri, Sep 04, 2020 at 07:06:17PM +0100, Al Viro wrote:
> > On Fri, Sep 04, 2020 at 06:52:11PM +0200, Christoph Hellwig wrote:
> > > Use get_unaligned and put_unaligned for the small constant size cases
> > > in the generic uaccess routines.  This ensures they can be used for
> > > architectures that do not support unaligned loads and stores, while
> > > being a no-op for those that do.
> >
> > Frankly, I would rather get rid of those constant-sized cases entirely;
> > sure, we'd need to adjust asm-generic/uaccess.h defaults for __get_user(),
> > but there that kind of stuff would make sense; in raw_copy_from_user()
> > it really doesn't.

Right. When I originally wrote that part of asm-generic/uaccess.h, the
idea was that its __get_user()/__put_user() would end up being used
across most architectures, which then would only have to implement
custom __copy_from_user()/__copy_to_user() with those special cases
to get the optimum behavior. It didn't work out in the end, since
copy_from_user() also has to deal with unaligned or partial copies
that prevent it from degrading into a single instruction on anything
other than the simplest NOMMU architectures.

I'd still hope we can eventually come up with a generic
__get_user()/__put_user() helper that avoids all the common
architecture specific bugs in them, with a simpler way for
an architecture to hook into with a set of inline functions
while leaving the macros to common code, but that can be
done another time.

> IOW, there's a scattering of potentially valid uses that might be better
> off with get_user().  And there's generic non-MMU variant that gets
> used in get_user()/__get_user() on h8300 and riscv.  This one *is*
> valid, but I don't think that raw_copy_from_user() is the right layer
> for that.
>
> For raw_copy_to_user() the picture is similar.  And I'd like to get
> rid of that magical crap.  Let's not make it harder...

Agreed

         Arnd