Re: [PATCH V6 2/2 RESEND] ksm: replace jhash2 with faster hash

[Mike Rapoport <rppt@linux.vnet.ibm.com>]

(updated Pasha's e-mail)

Thu, Sep 13, 2018 at 01:35:20PM +0300, Timofey Titovets wrote:
> D?D 1/2 , 25 D,N?D 1/2 . 2018 D3. D2 11:48, Mike Rapoport <rppt@linux.vnet.ibm.com>:
> >
> > On Thu, Jun 07, 2018 at 09:29:49PM -0400, Pavel Tatashin wrote:
> > > > With CONFIG_SYSFS=n there is nothing that will set ksm_run to anything but
> > > > zero and ksm_do_scan will never be called.
> > > >
> > >
> > > Unfortunatly, this is not so:
> > >
> > > In: /linux-master/mm/ksm.c
> > >
> > > 3143#else
> > > 3144 ksm_run = KSM_RUN_MERGE; /* no way for user to start it */
> > > 3145
> > > 3146#endif /* CONFIG_SYSFS */
> > >
> > > So, we do set ksm_run to run right from ksm_init() when CONFIG_SYSFS=n.
> > >
> > > I wonder if this is acceptible to only use xxhash when CONFIG_SYSFS=n ?
> >
> > BTW, with CONFIG_SYSFS=n KSM may start running before hardware acceleration
> > for crc32c is initialized...
> >
> > > Thank you,
> > > Pavel
> > >
> >
> > --
> > Sincerely yours,
> > Mike.
> >
> 
> Little thread bump.
> That patchset can't move forward already for about ~8 month.
> As i see main question in thread: that we have a race with ksm
> initialization and availability of crypto api.
> Maybe we then can fall back to simple plan, and just replace old good
> buddy jhash by just more fast xxhash?
> That allow move question with crypto api & crc32 to background, and
> make things better for now, in 2-3 times.
> 
> What you all think about that?

Sounds reasonable to me

> > crc32c_intel: 1084.10ns
> > crc32c (no hardware acceleration): 7012.51ns
> > xxhash32: 2227.75ns
> > xxhash64: 1413.16ns
> > jhash2: 5128.30ns
> 
> -- 
> Have a nice day,
> Timofey.
> 

-- 
Sincerely yours,
Mike.