Re: [PATCH v4 1/7] iio: introduce IIO_DECLARE_BUFFER_WITH_TS macros

[David Lechner <dlechner@baylibre.com>]

On 4/29/25 2:36 PM, Andy Shevchenko wrote:
> On Tue, Apr 29, 2025 at 10:31 PM David Lechner <dlechner@baylibre.com> wrote:
>> On 4/28/25 9:12 PM, David Lechner wrote:
>>> On 4/28/25 3:23 PM, David Lechner wrote:
>>>> Add new macros to help with the common case of declaring a buffer that
>>>> is safe to use with iio_push_to_buffers_with_ts(). This is not trivial
>>>> to do correctly because of the alignment requirements of the timestamp.
>>>> This will make it easier for both authors and reviewers.
>>>>
>>>> To avoid double __align() attributes in cases where we also need DMA
>>>> alignment, add a 2nd variant IIO_DECLARE_DMA_BUFFER_WITH_TS().
> 
> ...
> 
>>>> +/**
>>>> + * IIO_DECLARE_DMA_BUFFER_WITH_TS() - Declare a DMA-aligned buffer with timestamp
>>>> + * @type: element type of the buffer
>>>> + * @name: identifier name of the buffer
>>>> + * @count: number of elements in the buffer
>>>> + *
>>>> + * Same as IIO_DECLARE_BUFFER_WITH_TS(), but is uses __aligned(IIO_DMA_MINALIGN)
>>>> + * to ensure that the buffer doesn't share cachelines with anything that comes
>>>> + * before it in a struct. This should not be used for stack-allocated buffers
>>>> + * as stack memory cannot generally be used for DMA.
>>>> + */
>>>> +#define IIO_DECLARE_DMA_BUFFER_WITH_TS(type, name, count)   \
>>>> +    __IIO_DECLARE_BUFFER_WITH_TS(type, name, count)         \
>>>> +    /* IIO_DMA_MINALIGN may be 4 on some 32-bit arches. */  \
>>>> +    __aligned(MAX(IIO_DMA_MINALIGN, sizeof(s64)))
>>>
>>> I just realized my logic behind this is faulty. It assumes sizeof(s64) ==
>>> __alignof__(s64), but that isn't always true and that is what caused the builds
>>> to hit the static_assert() on v3.
>>>
>>> We should be able to leave this as __aligned(IIO_DMA_MINALIGN)
>>>
>>> And have this (with better error message):
>>>
>>> static assert(IIO_DMA_MINALIGN % __alignof__(s64) == 0);
>>
>> I was working late yesterday and should have saved that reply until morning
>> to think about it more!
>>
>> We do want to align to to sizeof(s64) instead of __alignof__(s64) to avoid
>> issues with, e.g. 32-bit kernel and 64-bit userspace (same reason that
>> aligned_s64 exists and always uses 8-byte alignment).
>>
>> So I think this patch is correct as-is after all.
> 
> I'm wondering, shouldn't it be better just to make sure that
> IIO_DMA_MINALIGN is always bigger or equal to sizeof(s64)?
> 

Sounds reasonable to me. From what I have seen while working on this is that
there are quite a few drivers using IIO_DMA_MINALIGN expecting it to be
sufficient for timestamp alignment, which as it seems is not always the case.

I'll wait for Jonathan to weigh in though before spinning up a new patch.