Re: [PATCH v3] ad7877: keep dma rx buffers in seperate cache lines

From: Dmitry Torokhov <dmitry.torokhov@gmail.com>
To: Mike Frysinger <vapier.adi@gmail.com>
Cc: "Andrew Morton" <akpm@linux-foundation.org>,
	"Oskar Schirmer" <os@emlix.com>,
	"Michael Hennerich" <Michael.Hennerich@analog.com>,
	linux-input@vger.kernel.org, linux-kernel@vger.kernel.org,
	"Daniel Glöckner" <dg@emlix.com>,
	"Oliver Schneidewind" <osw@emlix.com>,
	"Johannes Weiner" <jw@emlix.com>
Subject: Re: [PATCH v3] ad7877: keep dma rx buffers in seperate cache lines
Date: Mon, 10 May 2010 23:33:09 -0700	[thread overview]
Message-ID: <20100511063309.GC9644@core.coreip.homeip.net> (raw)
In-Reply-To: <AANLkTiktu7N5IXknkkJ9D5L4z5plhuGvD45IVqJm9zG1@mail.gmail.com>

On Tue, May 11, 2010 at 02:23:34AM -0400, Mike Frysinger wrote:
> On Tue, May 11, 2010 at 02:21, Dmitry Torokhov wrote:
> > On Tue, May 11, 2010 at 02:11:41AM -0400, Mike Frysinger wrote:
> >> On Tue, May 11, 2010 at 02:05, Dmitry Torokhov wrote:
> >> > On Mon, May 10, 2010 at 02:22:25PM -0700, Andrew Morton wrote:
> >> >> On Mon, 10 May 2010 12:42:34 +0200 Oskar Schirmer wrote:
> >> >> > With dma based spi transmission, data corruption
> >> >> > is observed occasionally. With dma buffers located
> >> >> > right next to msg and xfer fields, cache lines
> >> >> > correctly flushed in preparation for dma usage
> >> >> > may be polluted again when writing to fields
> >> >> > in the same cache line.
> >> >> >
> >> >> > Make sure cache fields used with dma do not
> >> >> > share cache lines with fields changed during
> >> >> > dma handling. As both fields are part of a
> >> >> > struct that is allocated via kzalloc, thus
> >> >> > cache aligned, moving the fields to the 1st
> >> >> > position and insert padding for alignment
> >> >> > does the job.
> >> >>
> >> >> This sounds odd.  Doesn't it imply that some code somewhere is missing
> >> >> some DMA synchronisation actions?
> >> >>
> >> >> >
> >> >> > v2: add a comment to explain why alignment is needed
> >> >> >
> >> >> > v3: fix the typo in comment and layout (- to end of line)
> >> >> >
> >> >> > diff --git a/drivers/input/touchscreen/ad7877.c b/drivers/input/touchscreen/ad7877.c
> >> >> > index 885354c..9ebb1b4 100644
> >> >> > --- a/drivers/input/touchscreen/ad7877.c
> >> >> > +++ b/drivers/input/touchscreen/ad7877.c
> >> >> > @@ -153,15 +153,29 @@ enum {
> >> >> >   */
> >> >> >
> >> >> >  struct ser_req {
> >> >> > +   u16                     sample;
> >> >> > +   /*
> >> >> > +    * DMA (thus cache coherency maintenance) requires the
> >> >> > +    * transfer buffers to live in their own cache lines.
> >> >> > +    */
> >> >> > +   char                    __padalign[L1_CACHE_BYTES - sizeof(u16)];
> >> >>
> >> >> It would be better to use __cacheline_aligned, rather than open-coding
> >> >> things in this manner.
> >> >>
> >> >
> >> > OK, then I have the following which I will apply unless someone shouts.
> >> >
> >> > --
> >> > Dmitry
> >> >
> >> > Input: ad7877 - keep dma rx buffers in seperate cache lines
> >> >
> >> > From: Oskar Schirmer <os@emlix.com>
> >> >
> >> > With dma based spi transmission, data corruption is observed
> >> > occasionally. With dma buffers located right next to msg and
> >> > xfer fields, cache lines correctly flushed in preparation for
> >> > dma usage may be polluted again when writing to fields in the
> >> > same cache line.
> >> >
> >> > Make sure cache fields used with dma do not share cache lines
> >> > with fields changed during dma handling. As both fields are part
> >> > of a struct that is allocated via kzalloc, thus cache aligned,
> >> > moving the fields to the 1st position and insert padding for
> >> > alignment does the job.
> >> >
> >> > Signed-off-by: Oskar Schirmer <os@emlix.com>
> >> > Signed-off-by: Daniel Glöckner <dg@emlix.com>
> >> > Signed-off-by: Oliver Schneidewind <osw@emlix.com>
> >> > Signed-off-by: Johannes Weiner <jw@emlix.com>
> >> > Acked-by: Mike Frysinger <vapier@gentoo.org>
> >> > [dtor@mail.ru - changed to use ___cacheline_aligned at suggestion
> >> >  of akpm]
> >> > Signed-off-by: Dmitry Torokhov <dtor@mail.ru>
> >> > ---
> >> >
> >> >  drivers/input/touchscreen/ad7877.c |   15 ++++++++++++---
> >> >  1 files changed, 12 insertions(+), 3 deletions(-)
> >> >
> >> >
> >> > diff --git a/drivers/input/touchscreen/ad7877.c b/drivers/input/touchscreen/ad7877.c
> >> > index e019d53..0d2d7e5 100644
> >> > --- a/drivers/input/touchscreen/ad7877.c
> >> > +++ b/drivers/input/touchscreen/ad7877.c
> >> > @@ -156,9 +156,14 @@ struct ser_req {
> >> >        u16                     reset;
> >> >        u16                     ref_on;
> >> >        u16                     command;
> >> > -       u16                     sample;
> >> >        struct spi_message      msg;
> >> >        struct spi_transfer     xfer[6];
> >> > +
> >> > +       /*
> >> > +        * DMA (thus cache coherency maintenance) requires the
> >> > +        * transfer buffers to live in their own cache lines.
> >> > +        */
> >> > +       u16 sample ____cacheline_aligned;
> >> >  };
> >> >
> >> >  struct ad7877 {
> >> > @@ -182,8 +187,6 @@ struct ad7877 {
> >> >        u8                      averaging;
> >> >        u8                      pen_down_acc_interval;
> >> >
> >> > -       u16                     conversion_data[AD7877_NR_SENSE];
> >> > -
> >> >        struct spi_transfer     xfer[AD7877_NR_SENSE + 2];
> >> >        struct spi_message      msg;
> >> >
> >> > @@ -195,6 +198,12 @@ struct ad7877 {
> >> >        spinlock_t              lock;
> >> >        struct timer_list       timer;          /* P: lock */
> >> >        unsigned                pending:1;      /* P: lock */
> >> > +
> >> > +       /*
> >> > +        * DMA (thus cache coherency maintenance) requires the
> >> > +        * transfer buffers to live in their own cache lines.
> >> > +        */
> >> > +       u16 conversion_data[AD7877_NR_SENSE] ____cacheline_aligned;
> >> >  };
> >>
> >> i'm not sure this is correct.  the cached_aligned attribute makes sure
> >> it starts on a cache boundary, but it doesnt make sure it pads out to
> >> one.  so it might work more of the time, but i dont think it's
> >> guaranteed.
> >
> > The buffers are moved to the end of the structure - there is nothing
> > else there.
> 
> what guarantee exactly do you have for that statement ?

The data is kmalloced, kmalloc aligns on cacheline boundary AFAIK which
means that next kmalloc data chunk will not share "our" cacheline.

-- 
Dmitry