* Re: patch to kaweth.c to align IP header
@ 2002-09-05 8:23 ` Kevin D. Kissell
0 siblings, 0 replies; 6+ messages in thread
From: Kevin D. Kissell @ 2002-09-05 8:23 UTC (permalink / raw)
To: Alan Cox, Quinn Jensen; +Cc: linux-mips
----- Original Message -----
From: "Alan Cox" <alan@lxorguk.ukuu.org.uk>
To: "Quinn Jensen" <jensenq@lineo.com>
Cc: <linux-mips@linux-mips.org>
Sent: Thursday, September 05, 2002 1:34 AM
Subject: Re: patch to kaweth.c to align IP header
> On Wed, 2002-09-04 at 19:26, Quinn Jensen wrote:
> > All,
> >
> > The Kawasaki LSI USB Ethernet driver was causing a crash
> > in ipt_do_table() on mips because the address fields in
> > the IP header were not word aligned. Many (all?) other
>
> You -must- handle alignment traps in the kernel for networking. The
> network code assumes and relies on this property and there are plenty of
> other ways to get misaligned datagrams through things like ip in ip.
>
> > ethernet drivers do an skb_reserve of 2 to word align
> > the address fields, and doing this in kaweth.c fixed
> > my crash.
>
> Its not the crash fix, its however right in the sense its a good
> performance optimisation for most platforms
It is true that, due to the unfortunate lack of foresight in the
design of IP, no pre-alignment of buffers will *guarantee*
that the address or other fields of IP headers will be aligned.
But I note that a design which assumes, for non x86 CPUs,
that unaligned references will be handled by a kernel trap
handler had darn well better assure itself that the misaligned
case is extemely infrequent. Otherwise, it would be a distinctly
better design to extract all multi-byte IP header values using
a macro which could map to a direct, possibly unaligned,
load for CISC architectures, and to appropriate sequences
of instructions for RISC architectures. I haven't measured
the alignment fault path for MIPS/Linux (in any case, MIPS
isn't the only architecture affected here), but if we assume for
the sake of the argument that it's 50 instructions, and that an
unaligned halfword costs 4 inline instructions (lbu,lbu,sll,or),
then using the unaligned reference trap as a crutch is a win
only if the fields are correctly aligned roughtly 94% of the time.
If full 32-bit or 64-bit words are being extracted, a MIPS CPU
can do the unaligned accesses in only 2 in-line instructions,
which would push the breakeven point out even further.
Does anyone have any actual statistical data on the cost
and frequency of this use of the unaligned access fixup?
Regards,
Kevin K.
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: patch to kaweth.c to align IP header
2002-09-05 8:23 ` Kevin D. Kissell
(?)
@ 2002-09-05 8:57 ` Carsten Langgaard
-1 siblings, 0 replies; 6+ messages in thread
From: Carsten Langgaard @ 2002-09-05 8:57 UTC (permalink / raw)
To: Kevin D. Kissell; +Cc: Alan Cox, Quinn Jensen, linux-mips
"Kevin D. Kissell" wrote:
> ----- Original Message -----
> From: "Alan Cox" <alan@lxorguk.ukuu.org.uk>
> To: "Quinn Jensen" <jensenq@lineo.com>
> Cc: <linux-mips@linux-mips.org>
> Sent: Thursday, September 05, 2002 1:34 AM
> Subject: Re: patch to kaweth.c to align IP header
>
> > On Wed, 2002-09-04 at 19:26, Quinn Jensen wrote:
> > > All,
> > >
> > > The Kawasaki LSI USB Ethernet driver was causing a crash
> > > in ipt_do_table() on mips because the address fields in
> > > the IP header were not word aligned. Many (all?) other
> >
> > You -must- handle alignment traps in the kernel for networking. The
> > network code assumes and relies on this property and there are plenty of
> > other ways to get misaligned datagrams through things like ip in ip.
> >
> > > ethernet drivers do an skb_reserve of 2 to word align
> > > the address fields, and doing this in kaweth.c fixed
> > > my crash.
> >
> > Its not the crash fix, its however right in the sense its a good
> > performance optimisation for most platforms
>
> It is true that, due to the unfortunate lack of foresight in the
> design of IP, no pre-alignment of buffers will *guarantee*
> that the address or other fields of IP headers will be aligned.
>
> But I note that a design which assumes, for non x86 CPUs,
> that unaligned references will be handled by a kernel trap
> handler had darn well better assure itself that the misaligned
> case is extemely infrequent. Otherwise, it would be a distinctly
> better design to extract all multi-byte IP header values using
> a macro which could map to a direct, possibly unaligned,
> load for CISC architectures, and to appropriate sequences
> of instructions for RISC architectures. I haven't measured
> the alignment fault path for MIPS/Linux (in any case, MIPS
> isn't the only architecture affected here), but if we assume for
> the sake of the argument that it's 50 instructions, and that an
> unaligned halfword costs 4 inline instructions (lbu,lbu,sll,or),
> then using the unaligned reference trap as a crutch is a win
> only if the fields are correctly aligned roughtly 94% of the time.
>
> If full 32-bit or 64-bit words are being extracted, a MIPS CPU
> can do the unaligned accesses in only 2 in-line instructions,
> which would push the breakeven point out even further.
>
> Does anyone have any actual statistical data on the cost
> and frequency of this use of the unaligned access fixup?
>
There is implemented a counter in the unaligned exception handler, you used
to get hold of the value through /proc/cpuinfo, but this has apparently been
removed from the latest kernels.
>
> Regards,
>
> Kevin K.
--
_ _ ____ ___ Carsten Langgaard Mailto:carstenl@mips.com
|\ /|||___)(___ MIPS Denmark Direct: +45 4486 5527
| \/ ||| ____) Lautrupvang 4B Switch: +45 4486 5555
TECHNOLOGIES 2750 Ballerup Fax...: +45 4486 5556
Denmark http://www.mips.com
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: patch to kaweth.c to align IP header
2002-09-05 8:23 ` Kevin D. Kissell
(?)
(?)
@ 2002-09-05 11:22 ` Alan Cox
-1 siblings, 0 replies; 6+ messages in thread
From: Alan Cox @ 2002-09-05 11:22 UTC (permalink / raw)
To: Kevin D. Kissell; +Cc: Quinn Jensen, linux-mips
On Thu, 2002-09-05 at 09:23, Kevin D. Kissell wrote:
> It is true that, due to the unfortunate lack of foresight in the
> design of IP, no pre-alignment of buffers will *guarantee*
> that the address or other fields of IP headers will be aligned.
It was not done for lack of foresight. It was carefully instrumented
measured and assessed.
> But I note that a design which assumes, for non x86 CPUs,
> that unaligned references will be handled by a kernel trap
> handler had darn well better assure itself that the misaligned
> case is extemely infrequent. Otherwise, it would be a distinctly
It does
> then using the unaligned reference trap as a crutch is a win
> only if the fields are correctly aligned roughtly 94% of the time.
With properly set up network cards (and that can be a problem some can't
DMA to half word start points) the benched numbers I got for normal use
back when we decided to go this way were that no packet appeared
unaligned unless deeply weird stuff like IPX over 802.2 without SNAP was
being used. Nevertheless there are several way users can trigger such
alignment so your code must handle them. Another case it can occur is
PPP.
Hitting 94% aligned is trivially the norm.
^ permalink raw reply [flat|nested] 6+ messages in thread