From: Daniel Borkmann <daniel@iogearbox.net> Date: 2016-06-16 14:37:59
On 06/16/2016 03:57 PM, Derek Ditch wrote:
[...]
quoted
Begin forwarded message:
From: bugzilla-daemon@bugzilla.kernel.org
Subject: [Bug 120441] af_packet no longer uses symmetric hashing
Date: June 16, 2016 at 07:35:54 -0500
To: derek.ditch@gmail.com
https://bugzilla.kernel.org/show_bug.cgi?id=120441
Just thinking out loud, perhaps one way would be to implement your own fanout
hashing via eBPF demuxer, f.e. in case CPU selection etc wouldn't work. Another
option could be to have a (2nd) customizable hash demux that would enforce to
always go through the flow-dissector and where you would have a socket option
to control keys for it, but I presume that seems rather less desirable.
quoted
--- Comment #2 from Eric Leblond <eric@regit.org> ---
This patch has been introduced in 4.2.
--
You are receiving this mail because:
You reported the bug.
Begin forwarded message:
From: bugzilla-daemon@bugzilla.kernel.org
Subject: [Bug 120441] af_packet no longer uses symmetric hashing
Date: June 16, 2016 at 06:59:58 -0500
To: derek.ditch@gmail.com
https://bugzilla.kernel.org/show_bug.cgi?id=120441
Eric Leblond [off-list ref] changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |eric@regit.org
--- Comment #1 from Eric Leblond <eric@regit.org> ---
On 16Jun 2016, at 09:37, Daniel Borkmann [off-list ref] wrote:
On 06/16/2016 03:57 PM, Derek Ditch wrote:
[...]
quoted
quoted
Begin forwarded message:
From: bugzilla-daemon@bugzilla.kernel.org
Subject: [Bug 120441] af_packet no longer uses symmetric hashing
Date: June 16, 2016 at 07:35:54 -0500
To: derek.ditch@gmail.com
https://bugzilla.kernel.org/show_bug.cgi?id=120441
Just thinking out loud, perhaps one way would be to implement your own fanout
hashing via eBPF demuxer, f.e. in case CPU selection etc wouldn't work. Another
option could be to have a (2nd) customizable hash demux that would enforce to
always go through the flow-dissector and where you would have a socket option
to control keys for it, but I presume that seems rather less desirable.
It’s less desirable, but at the moment seems like the path of least resistance. I haven’t written an eBPF program before, so I’m reading up on that. Once I get that worked out I can add that to the socket. Have you implemented this in netsniff-ng or something else yet, Daniel?
—
Derek Ditch
derek.ditch@gmail.com
GPG: 0x2543A3B5
quoted
quoted
--- Comment #2 from Eric Leblond <eric@regit.org> ---
This patch has been introduced in 4.2.
--
You are receiving this mail because:
You reported the bug.
Begin forwarded message:
From: bugzilla-daemon@bugzilla.kernel.org
Subject: [Bug 120441] af_packet no longer uses symmetric hashing
Date: June 16, 2016 at 06:59:58 -0500
To: derek.ditch@gmail.com
https://bugzilla.kernel.org/show_bug.cgi?id=120441
Eric Leblond [off-list ref] changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |eric@regit.org
--- Comment #1 from Eric Leblond <eric@regit.org> ---
From: Hannes Frederic Sowa <hidden> Date: 2016-06-16 14:50:07
On 16.06.2016 16:37, Daniel Borkmann wrote:
On 06/16/2016 03:57 PM, Derek Ditch wrote:
[...]
quoted
quoted
Begin forwarded message:
From: bugzilla-daemon@bugzilla.kernel.org
Subject: [Bug 120441] af_packet no longer uses symmetric hashing
Date: June 16, 2016 at 07:35:54 -0500
To: derek.ditch@gmail.com
https://bugzilla.kernel.org/show_bug.cgi?id=120441
Just thinking out loud, perhaps one way would be to implement your own
fanout
hashing via eBPF demuxer, f.e. in case CPU selection etc wouldn't work.
Another
option could be to have a (2nd) customizable hash demux that would
enforce to
always go through the flow-dissector and where you would have a socket
option
to control keys for it, but I presume that seems rather less desirable.
Isn't a symmetric hashing for af-packet something very fundamental?
Looks much more like a bug to me.
Bye,
Hannes
From: Daniel Borkmann <daniel@iogearbox.net> Date: 2016-06-16 15:18:22
On 06/16/2016 04:50 PM, Hannes Frederic Sowa wrote:
On 16.06.2016 16:37, Daniel Borkmann wrote:
quoted
On 06/16/2016 03:57 PM, Derek Ditch wrote:
[...]
quoted
quoted
Begin forwarded message:
From: bugzilla-daemon@bugzilla.kernel.org
Subject: [Bug 120441] af_packet no longer uses symmetric hashing
Date: June 16, 2016 at 07:35:54 -0500
To: derek.ditch@gmail.com
https://bugzilla.kernel.org/show_bug.cgi?id=120441
Just thinking out loud, perhaps one way would be to implement your own
fanout
hashing via eBPF demuxer, f.e. in case CPU selection etc wouldn't work.
Another
option could be to have a (2nd) customizable hash demux that would
enforce to
always go through the flow-dissector and where you would have a socket
option
to control keys for it, but I presume that seems rather less desirable.
Isn't a symmetric hashing for af-packet something very fundamental?
Looks much more like a bug to me.
I agree it would be useful resp. fundamental. I'm not quite sure whether
PACKET_FANOUT_HASH always had this guarantee since the initial implementation
f.e. depending on whether the skb got fanout demuxed from RX or TX paths,
you would either use a hw hash or it would have gone through the kernel's
(predecessor) flow dissector (and therefore potentially to different packet
sockets eventually). It might be the case that this got noticed due to the
flow dissector nowadays collecting more input for building its hash, hmm.
From: Willem de Bruijn <willemdebruijn.kernel@gmail.com> Date: 2016-06-16 15:33:46
quoted
Isn't a symmetric hashing for af-packet something very fundamental?
Looks much more like a bug to me.
I agree it would be useful resp. fundamental. I'm not quite sure whether
PACKET_FANOUT_HASH always had this guarantee since the initial
implementation
It is based on rxhash, which is intended to have this property. See for
instance the commit messages from commit b249dcb82d32 ("rps:
consistent rxhash")
"In case we compute a software skb->rxhash, we can generate a
consistent hash : Its value will be the same in both flow directions."
commit 6862234238e8 ("net: small bug on rxhash calculation") also
explicitly refers to this property.
f.e. depending on whether the skb got fanout demuxed from RX or TX paths,
you would either use a hw hash or it would have gone through the kernel's
(predecessor) flow dissector
It's quite possible that not all hardware hash implementations do the
right thing, indeed.