Re: [PATCH v1] clk: Convert managed get functions to devm_add_action API

From: Marc Gonzalez
Date: Wed Dec 11 2019 - 11:17:45 EST


On 02/12/2019 14:51, Robin Murphy wrote:

> On 02/12/2019 9:25 am, Marc Gonzalez wrote:
>
>> On 02/12/2019 02:42, Dmitry Torokhov wrote:
>>
>>> On Thu, Nov 28, 2019 at 10:56:30AM -0800, Bjorn Andersson wrote:
>>>
>>>> On Tue 26 Nov 08:13 PST 2019, Marc Gonzalez wrote:
>>>>
>>>>> Date: Tue, 26 Nov 2019 13:56:53 +0100
>>>>>
>>>>> Using devm_add_action_or_reset() produces simpler code and smaller
>>>>> object size:
>>>>>
>>>>> 1 file changed, 16 insertions(+), 46 deletions(-)
>>>>>
>>>>> text data bss dec hex filename
>>>>> - 1797 80 0 1877 755 drivers/clk/clk-devres.o
>>>>> + 1499 56 0 1555 613 drivers/clk/clk-devres.o
>>>>>
>>>>> Signed-off-by: Marc Gonzalez <marc.w.gonzalez@xxxxxxx>
>>>>
>>>> Looks neat
>>>>
>>>> Reviewed-by: Bjorn Andersson <bjorn.andersson@xxxxxxxxxx>
>>>
>>> This however increases the runtime costs as each custom action cost us
>>> an extra pointer. Given that in a system we likely have many clocks
>>> managed by devres, I am not sure that this code savings is actually
>>> gives us overall win. It might still, I just want to understand how we
>>> are allocating/packing devres structures.
>>
>> I'm not 100% sure what you are saying.
>
> You reduce the text size by a constant amount, at the cost of allocating
> twice as much runtime data per clock (struct action_devres vs. void*).
> Assuming 64-bit pointers, that means that in principle your ~320-byte
> saving would be cancelled out at ~40 managed clocks. However, that's
> also assuming that the minimum allocation granularity is no larger than
> a single pointer, which generally isn't true, so in reality it depends
> on whether the difference in data pushes the total struct devres
> allocation over the next ARCH_KMALLOC_MINALIGN boundary - if it doesn't,
> the difference comes entirely for free; if it does, the memory cost
> tradeoff gets even worse.

Aaah... memory overhead. Thanks for pointing it out.

BEFORE

devm_clk_get()
-> devres_alloc(devm_clk_release, sizeof(*ptr), GFP_KERNEL);
allocates space for a struct devres + a pointer

struct devres {
struct devres_node node;
/*
* Some archs want to perform DMA into kmalloc caches
* and need a guaranteed alignment larger than
* the alignment of a 64-bit integer.
* Thus we use ARCH_KMALLOC_MINALIGN here and get exactly the same
* buffer alignment as if it was allocated by plain kmalloc().
*/
u8 __aligned(ARCH_KMALLOC_MINALIGN) data[];
};

Not sure what it means for a flexible array member to be X-aligned...

(Since the field's address depends on the start address, which is only
determined at run-time...)

For example, on arm64, ARCH_KMALLOC_MINALIGN appears to be 128 (sometimes).

/*
* Memory returned by kmalloc() may be used for DMA, so we must make
* sure that all such allocations are cache aligned. Otherwise,
* unrelated code may cause parts of the buffer to be read into the
* cache before the transfer is done, causing old data to be seen by
* the CPU.
*/
#define ARCH_DMA_MINALIGN (128)


Unless the strict alignment is also imposed on kmalloc?

So basically, a struct devres starts on a multiple-of-128 address,
first the devres_node member, then padding to the next 128, then the
data member?


/*
* Some archs want to perform DMA into kmalloc caches and need a guaranteed
* alignment larger than the alignment of a 64-bit integer.
* Setting ARCH_KMALLOC_MINALIGN in arch headers allows that.
*/
#if defined(ARCH_DMA_MINALIGN) && ARCH_DMA_MINALIGN > 8
#define ARCH_KMALLOC_MINALIGN ARCH_DMA_MINALIGN
#define KMALLOC_MIN_SIZE ARCH_DMA_MINALIGN
#define KMALLOC_SHIFT_LOW ilog2(ARCH_DMA_MINALIGN)
#else
#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
#endif


A devres_node boils down to 2 object pointers + 1 function pointer.

Are there architectures supported by Linux where a function pointer
is not the same size as an object pointer? (ia64 maybe?)



OK, I will give this patch some more thought.

But I need to ask: what is the rationale for the devm_add_action API?

Regards.