Re: Speeding up df

[Bernhard Voelker]

@Padraig:
I'll be away from keyboard the next 3 weeks, so please continue
with this one to finally push.


On 07/25/2016 01:21 PM, Philipp Thomas wrote:
> * Pádraig Brady (address@hidden) [20160725 01:20]:
> 
>>> otherwise it looks good to me.
>>
>> Excellent. +1
> 
> I'll pass on your kind words to my collegue.
> 
>> Worth a mention in improvements in NEWS too.
> 
> Like so?
> 
> +  Df now uses a hash table to filter the mount list. This sped up
> +  processing of a 22000 lines mount list by nearly 60%.
> 
> If yes and Bernies explains me why we'd need an additional hash_free call
> I'll post the changed patch.

Thanks.  I wrapped this into a patch with a nice commit message.
I also squashed in the renaming of the 'devlist' variable, as that
many "devlists" were too confusing.

I see this hashing as a starting point, as other functions also need
a better lookup ... and to avoid superfluous stat/statfs calls (see
e.g. "strace -e stat,statfs df -a").

BTW: surprisingly, one can get a HUGE /proc/self/mountinfo very easily:

  # !!! DO NOT DO THIS (with "seq 25") !!!
  mkdir /root/tmp &&
  for f in $(seq 25); do \
    mkdir /root/tmp/$f \
      && mount --bind / /root/tmp/$f
  done

To get the 20.000-30.000 entries, you only need ~"seq 16".

Have a nice day,
Berny

0001-df-improve-performance-with-many-mount-points.patch
Description: Text Data