[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240309181942.5a60c3a7@jic23-huawei>
Date: Sat, 9 Mar 2024 18:19:42 +0000
From: Jonathan Cameron <jic23@...nel.org>
To: Vasileios Amoiridis <vassilisamir@...il.com>
Cc: Andy Shevchenko <andriy.shevchenko@...ux.intel.com>, lars@...afoo.de,
ang.iglesiasg@...il.com, mazziesaccount@...il.com, ak@...klinger.de,
petre.rodan@...dimension.ro, phil@...pberrypi.com, 579lpy@...il.com,
linux-iio@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 4/4] iio: pressure: Add triggered buffer support for
BMP280 driver
On Mon, 4 Mar 2024 21:05:47 +0100
Vasileios Amoiridis <vassilisamir@...il.com> wrote:
> On Mon, Mar 04, 2024 at 09:18:27PM +0200, Andy Shevchenko wrote:
> > On Mon, Mar 04, 2024 at 08:08:38PM +0100, Vasileios Amoiridis wrote:
> > > On Mon, Mar 04, 2024 at 01:52:05PM +0200, Andy Shevchenko wrote:
> > > > On Sun, Mar 03, 2024 at 05:53:00PM +0100, Vasileios Amoiridis wrote:
> >
> > ...
> >
> > > > > + struct {
> > > > > + s32 temperature;
> > > > > + u32 pressure;
> > > > > + u32 humidity;
> > > >
> > > > > + s64 timestamp;
> > > >
> > > > Shouldn't this be aligned properly?
> > >
> > > I saw that in some drivers it was added and in some it was not. What is the
> > > difference of aligning just the timestamp of the kernel?
> >
> > You can count yourself. With provided structure as above there is a high
> > probability of misaligned timeout field. The latter has to be aligned on
> > 8 bytes.
> >
>
> I was unaware, but now I am not. Thank you very much for the feedback.
Fun bit of C is that you aren't actually aligning just the timestamp.
A C structure is aligned to the alignment of the maximum element within it.
So by specifying that timestamp is aligned to 8 bytes, you also force the
alignment of the whole structure to 8 bytes.
When you see the outer buffer aligned as well (typically the potentially larger
__aligned (IIO_DMA_MINALIGN)) that's for a different reason. Used on a trailing
element of a structure via iio_priv() that ensures there is nothing else in the
cacheline (maximum one in the system) on systems where this matters due to non
coherent DMA. Still need the __aligned(8) on the timestamp though as otherwise
the internal padding may be wrong (like here).
On some architectures small buffers are always bounced - if that were true on
all of them we could get rid of the complexity of IIO_DMA_MINALIGN.
Alignment is so much fun - particularly with x86_32 which does 8 byte values aligned
to 4 bytes. We had a massive set of patches fixing subtle issues around that a
few years ago.
Jonathan
> > > > > + } iio_buffer;
> >
> > --
> > With Best Regards,
> > Andy Shevchenko
> >
> >
Powered by blists - more mailing lists