svn commit: r281239 - head/sys/netinet

[Adrian Chadd]

hi gleb,

        /* Skip initialization of globals for non-default instances. */
        if (!IS_DEFAULT_VNET(curvnet))
                return;

.. ie, no. It doesn't double-init the mutexes.

And the purging from bucket behaviour should have been precisely what
the previous behaviour was.

               /*
                 * drop something from the tail of the current queue
                 * before proceeding further
                 */

.. if you'd like to fix that then cool, but it's orthogonal to what I
was trying to do. I didn't want to go and do multiple things in one
patch set - this was /just/ to unroll the per-bucket locking. If you'd
like to fix the existing potentially risky bucket purging code in the
ip reassembly code then cool, let's do it - but in a separate
patchset.

V_nipq leaking is something we both missed during that review cycle.
I'm glad you caught it now.


-adrian

On 8 April 2015 at 01:48, Gleb Smirnoff <glebius at freebsd.org> wrote:
> On Tue, Apr 07, 2015 at 11:09:35PM +0000, Adrian Chadd wrote:
> A> Author: adrian
> A> Date: Tue Apr  7 23:09:34 2015
> A> New Revision: 281239
> A> URL: https://svnweb.freebsd.org/changeset/base/281239
> A>
> A> Log:
> A>   Move the IPv4 reassembly queue locking from a single lock to be per-bucket (global).
> A>
> A>   This significantly improves performance on multi-core servers where there
> A>   is any kind of IPv4 reassembly going on.
> A>
> A>   glebius@ would like to see the locking moved to be attached to the reassembly
> A>   bucket, which would make it per-bucket + per-VNET, instead of being global.
> A>   I decided to keep it global for now as it's the minimal useful change;
> A>   if people agree / wish to migrate it to be per-bucket / per-VNET then please
> A>   do feel free to do so.  I won't complain.
> A>
> A>   Thanks to Norse Corp for giving me access to much larger servers
> A>   to test this at across the 4 core boxes I have at home.
> A>
> A>   Differential Revision:     https://reviews.freebsd.org/D2095
> A>   Reviewed by:       glebius (initial comments incorporated into this patch)
>
> No, they are not.
>
> Also, I just noticed another issue with the patch. You initialize the mutexes,
> which are global, in the ip_init() which is called for every VNET. That would
> lead to panic of initializing already initialized mutex. If you followed my advice,
> that would not happen. Now VIMAGE+INVARIANTS is broken instantly, and w/o
> INVARIANTS may panic if VIMAGE created under IP traffic.
>
> My comment that purging from a bucket that is determined by source IP has
> security implications was also ignored.
>
> I will fix this all myself.
>
> --
> Totus tuus, Glebius.