Re: [PATCHv7 10/24] zram: add zlib compression backend support

[Zaslonko Mikhail <zaslonko@linux.ibm.com>]

Hello,

On 13.05.2025 07:41, Sergey Senozhatsky wrote:
> Sorry for the delay,
> 
> On (25/05/09 17:18), Zaslonko Mikhail wrote:
>>> When zram transitioned from Crypto API (scomp) to custom compression
>>> API I picked the CryptoAPI deflate DEFLATE_DEF_WINBITS value:
>>>
>>> crypto/deflate.c: DEFLATE_DEF_WINBITS	11
>>>
>>> which is then passed to zlib_deflateInit2() and zlib_inflateInit2().
>>>
>>>> I tried to build the kernel with DEFLATE_DEF_WINBITS set to 15 and
>>>> verified that s390 hardware deflate acceleration works for zram devices
>>>> with a deflate compression.
>>>
>>> If we define it as 15 on non-s390 machines, will there be any
>>> consequences?  Increased memory usage?  By how much?
>>
>> On s390, setting windowBits to 15 would lead to zlib workarea size
>> increased by 120K (0x24dc8 -> 0x42dc8) per compression stream,
>> i.e. per online CPU. 
>> On non-s390 machine, that impact will be about 115K per stream. 
>> Increasing window size should improve deflate compression,
>> although the compression speed might be affected. Couldn't find any
>> relevant zlib benchmarks though.
>>
>> Not sure what other consequences might there be for zram. Do you see any?
> 
> The increased per-CPU memory usage is the only thing I can think of.
> I guess for zram we could turn this into a run-time parameter, but for
> Crypto API compile-time is the only option, I guess.

With 'run-time parameter' you mean adding 'windowBits' as another deflate compression
algorithm parameter for zram? Guess we could do this, using default value of 15 then.

> 
> Can you send a patch series (for zram and Crypto API) that sets
> windowBits to 15?

I can do it for zram. Not sure if Crypto should be changed as well. Or is it
supposed to have the same compression defaults as zram?

Thanks,
Mikhail