Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 14 Feb 2019 17:07:47 -0500
From:   Joel Fernandes <joelaf@google.com>
To:     Todd Kjos <tkjos@google.com>
Cc:     Todd Kjos <tkjos@android.com>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        Arve =?iso-8859-1?B?SGr4bm5lduVn?= <arve@android.com>,
        "open list:ANDROID DRIVERS" <devel@driverdev.osuosl.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Martijn Coenen <maco@google.com>,
        "Joel Fernandes (Google)" <joel@joelfernandes.org>,
        Android Kernel Team <kernel-team@android.com>
Subject: Re: [PATCH v3 1/7] binder: create userspace-to-binder-buffer copy
 function
Message-ID: <20190214220747.GC185075@google.com>
References: <20190208183520.30886-1-tkjos@google.com>
 <20190208183520.30886-2-tkjos@google.com>
 <20190214194540.GA185075@google.com>
 <CAHRSSEw22kuwLqjXHocBQTv4d5hS1tXi4uirjBhgQ=yGbw9J4g@mail.gmail.com>
 <CAJWu+oqq0s3x7j+gPzSRREj2gjZxzAPS_=XVM4uhC9GjSM9zfw@mail.gmail.com>
 <20190214212549.GB185075@google.com>
 <CAHRSSExEc0uF38wwgtF6JeVgcLiDKiPznK+SmX8fFEL8k-gpVA@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAHRSSExEc0uF38wwgtF6JeVgcLiDKiPznK+SmX8fFEL8k-gpVA@mail.gmail.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Feb 14, 2019 at 01:55:19PM -0800, 'Todd Kjos' via kernel-team wrote:
> On Thu, Feb 14, 2019 at 1:25 PM Joel Fernandes <joelaf@google.com> wrote:
> >
> > On Thu, Feb 14, 2019 at 03:53:54PM -0500, Joel Fernandes wrote:
> > > On Thu, Feb 14, 2019 at 3:42 PM Todd Kjos <tkjos@google.com> wrote:
> > > >
> > > > On Thu, Feb 14, 2019 at 11:45 AM Joel Fernandes <joelaf@google.com> wrote:
> > > [snip]
> > > > > > + * check_buffer() - verify that buffer/offset is safe to access
> > > > > > + * @alloc: binder_alloc for this proc
> > > > > > + * @buffer: binder buffer to be accessed
> > > > > > + * @offset: offset into @buffer data
> > > > > > + * @bytes: bytes to access from offset
> > > > > > + *
> > > > > > + * Check that the @offset/@bytes are within the size of the given
> > > > > > + * @buffer and that the buffer is currently active and not freeable.
> > > > > > + * Offsets must also be multiples of sizeof(u32). The kernel is
> > > > >
> > > > > In all callers of binder_alloc_copy_user_to_buffer, the alignment of offsets
> > > > > is set to sizeof(void *). Then shouldn't this function check for sizeof(void *)
> > > > > alignment instead of u32?
> > > >
> > > > But there are other callers of check_buffer() later in the series that
> > > > don't require pointer-size alignment. u32 alignment is consistent with
> > > > the alignment requirements of the binder driver before this change.
> > > > The copy functions don't actually need to insist on alignment, but
> > > > these binder buffer objects have always used u32 alignment which has
> > > > been checked in the driver. If user code misaligned it, then errors
> > > > are returned. The alignment checks are really to be consistent with
> > > > previous binder driver behavior.
> > >
> > > Got it, thanks.
> >
> > One more thing I wanted to ask is, kmap() will now cause global lock
> > contention because of using spin_lock due to kmap_high().
> >
> > Previously the binder driver was made to not use global lock (as you had
> > done). Now these paths will start global locking on 32-bit architectures.
> > Would that degrade performance?
> 
> There was a lot of concern about 32-bit performance both for
> lock-contention and the cost of map/unmap operations. Of course,
> 32-bit systems are also where the primary win is -- namely avoiding
> vmalloc space depletion. So there was a several months-long evaluation
> period on 32-bit devices by a silicon vendor who did a lot of testing
> across a broad set of benchmarks / workloads to verify the performance
> costs are acceptable. We also ran tests to try to exhaust the kmap
> space with multiple large buffers.
> 
> The testing did find that there is some performance degradation for
> large buffer transfers, but there are no cases where this
> significantly impacted a meaningful user workload.
> 
> >
> > Are we not using kmap_atomic() in this patch because of any concern that the
> > kmap fixmap space is limited and may run out?
> 
> We're not using the atomic version here since we can't guarantee that
> the loop will be atomic since we are calling copy_from_user(). Later
> in the series, other cases do use kmap_atomic().

Got it, thanks for all the clarifications,

- Joel