Skip site navigation (1)Skip section navigation (2)
Date:      Tue, 30 Oct 2018 09:46:11 -0700 (PDT)
From:      "Rodney W. Grimes" <freebsd-rwg@pdx.rh.CN85.dnsmgr.net>
To:        "Bjoern A. Zeeb" <bzeeb-lists@lists.zabbadoz.net>
Cc:        Kristof Provost <kp@freebsd.org>, Ernie Luzar <luzar722@gmail.com>, FreeBSD current <freebsd-current@freebsd.org>
Subject:   Re: 12.0-BETA1 vnet with pf firewall
Message-ID:  <201810301646.w9UGkBce062560@pdx.rh.CN85.dnsmgr.net>
In-Reply-To: <9D50D781-73BA-45B0-ADBB-CF01DE587BC5@lists.zabbadoz.net>

next in thread | previous in thread | raw e-mail | index | archive | help
> On 30 Oct 2018, at 14:14, Rodney W. Grimes wrote:
> 
> >> On 30 Oct 2018, at 14:29, Bjoern A. Zeeb wrote:
> >>> On 30 Oct 2018, at 12:23, Kristof Provost wrote:
> >>>> I?m not too familiar with this part of the vnet code, but it looks
> >>>> to me like we?ve got more per-vnet variables that was originally
> >>>> anticipated, so we may need to just increase the allocated space.
> >>>
> >>> Can you elfdump -a the two modules and see how big their set_vnet
> >>> section sizes are?  I see:
> >>>
> >>> pf.ko:  sh_size: 6664
> >>> ipl.ko: sh_size: 2992
> >>>
> >> I see exactly the same numbers.
> >>
> >>> VNET_MODMIN is two pages (8k).  So yes, that would exceed the module
> >>> space.
> >>> Having 6.6k global variable space is a bit excessive?  Where does 
> >>> that
> >>> come from?  multicast used to have a similar problem in the past 
> >>> that
> >>> it could not be loaded as a module as it had a massive array there 
> >>> and
> >>> we changed it to be malloced and that reduced it to a pointer.
> >>>
> >>> 0000000000000f38 l     O set_vnet       0000000000000428
> >>> vnet_entry_pfr_nulltable
> >> That?s a default table. It?s large because it uses MAXPATHLEN for
> >> the pfrt_anchor string.
> >>
> >>> 0000000000000b10 l     O set_vnet       00000000000003d0
> >>> vnet_entry_pf_default_rule
> >> Default rule. Rules potentially contain names, tag names, interface
> >> names, ? so it?s a large structure.
> >>
> >>> 0000000000001370 l     O set_vnet       0000000000000690
> >>> vnet_entry_pf_main_anchor
> >> Anchors use MAXPATHLEN for the anchor path, so that?s 1024 bytes 
> >> right
> >> away.
> >>
> >>> 0000000000000000 l     O set_vnet       0000000000000120
> >>> vnet_entry_pf_status
> >>>
> >> pf status. Mostly counters.
> >>
> >> I?ll see about putting moving those into the heap on my todo list.
> >
> > Though that removes the current situation, it is a partial fix,
> > doesnt this static sized 2 page VNET_MODMIN needs to be fixed in the
> > longer term?
> 
> I think about it the other way round:  we might want to bump it to 4 
> pages in short term for 12.0 maybe?

81MB in your 10000 vnet case, not horrible really as a quick and
slightly dirty fix.

> The problem is that whether or not you use modules these 2/4 pages will 
> be allocated per-vnet, so if you run 50 vnet jails that?s 100/200 
> pages.  And while people might say memory is cheap, I?ve run 10.000 
> vnet jails before on a single machine ? it adds up.  I wonder if we 
> could make it a tunable though..  Let me quickly think about it and come 
> up with a patch.

A boot time tunable that defaulted to 2 pages would be good too.

We are faced with similiar issues in bhyve, where each VM created
gets an allocation based on MAXCPU, I have wip in process to alter
this so that the 4th (and currently hidden) cpu topology option
controls the size of these allocations.

> 
> I?ll also go and see to get better error reporting into the 
> link_elf*.c files for this case.

Great!

-- 
Rod Grimes                                                 rgrimes@freebsd.org



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201810301646.w9UGkBce062560>