Subject: Re: [PATCH] ring_buffer: Ensure that buffer page data is aligned.
From: Steven Rostedt <rostedt@goodmis.org>
To: Will Newton <will.newton@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>, fweisbec@gmail.com,
        Linux Kernel list <linux-kernel@vger.kernel.org>
In-Reply-To: <BANLkTinO85w4G9nFjio4HtOcrS0u-Ezz1Q@mail.gmail.com>
References: <BANLkTimG_oBX1bDTEVfZ81_uqXOahqR1ww@mail.gmail.com>
	 <1308230239.9218.137.camel@gandalf.stny.rr.com>
	 <BANLkTinO85w4G9nFjio4HtOcrS0u-Ezz1Q@mail.gmail.com>
Content-Type: text/plain; charset="ISO-8859-15"
Date: Thu, 16 Jun 2011 13:24:29 -0400
Message-ID: <1308245069.9218.148.camel@gandalf.stny.rr.com>
Mime-Version: 1.0
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3112
Lines: 78

On Thu, 2011-06-16 at 14:32 +0100, Will Newton wrote:
> On Thu, Jun 16, 2011 at 2:17 PM, Steven Rostedt <rostedt@goodmis.org> wrote:
> > On Thu, 2011-06-16 at 14:03 +0100, Will Newton wrote:
> >> Explicitly align the start of the buffer page data array to the
> >> required arch alignment. This is required for architectures that
> >> require 8 byte alignment but do not have a 8 byte local_t.
> >
> > What arch does that? A 32bit arch that forces 8 byte alignment?
> 
> Yes, it's likely for all 64bit arches will be aligned correctly.

What arch is this exactly. Is it because 64bit variables must be aligned
on 8byte boundaries?

> 
> >>
> >> Signed-off-by: Will Newton <will.newton@imgtec.com>
> >> ---
> >>  kernel/trace/ring_buffer.c |    4 +++-
> >>  1 files changed, 3 insertions(+), 1 deletions(-)
> >>
> >> I don't believe that any currently in-tree architecture is affected by
> >> this, but it can be an issue on 32bit architectures that require an
> >> 8 byte aligment but only have a 32bit local_t. I think it's potentially
> >> cleaner to make the alignment explicit anyway.
> >>
> >> diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c
> >> index 0ef7b4b..36d5699 100644
> >> --- a/kernel/trace/ring_buffer.c
> >> +++ b/kernel/trace/ring_buffer.c
> >> @@ -215,6 +215,8 @@ EXPORT_SYMBOL_GPL(tracing_is_on);
> >>  # define RB_ARCH_ALIGNMENT           8U
> >>  #endif
> >>
> >> +#define RB_ALIGN_DATA                __aligned(RB_ARCH_ALIGNMENT)
> >> +
> >
> > Note, the code above this is:
> >
> > #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
> > #define RB_ALIGNMENT            4U
> > #define RB_MAX_SMALL_DATA       (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
> > #define RB_EVNT_MIN_SIZE        8U      /* two 32bit words */
> >
> > #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
> > # define RB_FORCE_8BYTE_ALIGNMENT       0
> > # define RB_ARCH_ALIGNMENT              RB_ALIGNMENT
> > #else
> > # define RB_FORCE_8BYTE_ALIGNMENT       1
> > # define RB_ARCH_ALIGNMENT              8U
> > #endif
> >
> >
> > Which means that when CONFIG_64BIT is not set, RB_ARCH_ALIGNMENT is 4.
> > This means that this patch is really a nop and doesn't do anything for
> > your arch.
> 
> I have modified that conditional in my local tree so the patch is not
> a nop there. Obviously it doesn't make any sense to merge that change
> yet. I thought this part of the patch was a reasonable cleanup to make
> the code a bit more self documenting though.

Unfortunately, if anything it confuses the code more. If I were
reviewing it and saw that we are aligning something that is already
aligned with the same value, I would question that code.

Now if/when your arch is merged, we can add this with the other changes
you have, which will make reviewing the code not as confusing.

-- Steve


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/