Re: [PATCH v2 1/4] sha1sum: use AF_ALG when available

[Matteo Croce]

On Wed, Apr 25, 2018 at 7:41 PM, Bernhard Voelker
<address@hidden> wrote:
> On 04/25/2018 01:26 PM, Matteo Croce wrote:
>> +#define BLOCKSIZE 32768
>
> Did you try different buffer sizes?
>
> In coreutils, we have a little script to determine the optimal BLOCKSIZE
> - which is 128*1024 since 2014:
>
> https://git.sv.gnu.org/cgit/coreutils.git/tree/src/ioblksize.h?id=22424dde74d0
>

Hi Bernhard,

I knew it and I tried different blocksizes too, but for some odd
reason there are no improvements with block size past 16k, so I left
the old value which was 32k. These are
the timings of sha1sum on a 16g file:

    1024  22.991s
    2048  18.609s
    4096  16.397s
    8192  15.255s
   16384  14.544s
   32768  14.126s
   65536  15.396s
  131072  15.247s
  262144  13.713s
  524288  15.141s
 1048576  15.139s

I have an Intel i7 i7-6700K CPU and a recent kernel with mitigations
enabled for Meltdown, which should give a penality on every context
switch, this is why I say "for some odd reason" :)

Keep in mind that with files smaller than 2 GB the whole hash will be
calculated with a single call to sendfile()

>
>> +#ifdef AF_ALG
>> +  int ret;
>> +
>> +  ret = afalg_stream(stream, resblock, "sha1", SHA1_DIGEST_SIZE);
>> +  if (!ret)
>> +      return 0;
>> +
>> +  if (ret == -EIO)
>> +      return 1;
>> +#endif
>
> afalg_stream also returns other errors, namely EAFNOSUPPORT and EINVAL,
> which should be handled as well, right?
>

EAFNOSUPPORT means that af_alg is not supported, hence the generic C
code is used as fallback.
EINVAL means that an string longer than
sizeof(sockaddr_alg.sockaddr_alg) is passed to afalg_stream().
This is impossible as we pass the constant value "sha1" to afalg_stream().

> Thanks & have a nice day,
> Berny

Thanks, have a nice day too.
-- 
Matteo Croce
per aspera ad upstream