[PATCH] gianfar: Fall back to software tcp/udp checksum on older controllers

[PATCH] gianfar: Fall back to software tcp/udp checksum on older controllers

[Alex Dubov]

As specified by errata eTSEC49 of MPC8548 and errata eTSEC12 of MPC83xx,=0A=
older revisions of gianfar controllers will be unable to calculate a TCP/UD=
P=0Apacket checksum for some aligments of the appropriate FCB. This patch c=
hecks=0Afor FCB alignment on such controllers and falls back to software ch=
ecksumming=0Aif the aligment is known to be bad.=0A=0ASigned-off-by: Alex D=
ubov <oakad@yahoo.com>=0A---=0AThis is my, somewhat different approach to M=
atthew Creech proposed solution.=0A=0A drivers/net/gianfar.c |   21 +++++++=
++++++++++++--=0A drivers/net/gianfar.h |    1 +=0A 2 files changed, 20 ins=
ertions(+), 2 deletions(-)=0A=0Adiff --git a/drivers/net/gianfar.c b/driver=
s/net/gianfar.c=0Aindex 5ed8f9f..b4f0e99 100644=0A--- a/drivers/net/gianfar=
.c=0A+++ b/drivers/net/gianfar.c=0A@@ -950,6 +950,11 @@ static void gfar_de=
tect_errata(struct gfar_private *priv)=0A =09=09=09(pvr =3D=3D 0x80861010 &=
& (mod & 0xfff9) =3D=3D 0x80c0))=0A =09=09priv->errata |=3D GFAR_ERRATA_A00=
2;=0A =0A+=09/* MPC8313 Rev < 2.0, MPC8548 rev 2.0 */=0A+=09if ((pvr =3D=3D=
 0x80850010 && mod =3D=3D 0x80b0 && rev < 0x0020)=0A+=09    || (pvr =3D=3D =
0x80210020 && mod =3D=3D 0x8030 && rev =3D=3D 0x0020))=0A+=09=09priv->errat=
a |=3D GFAR_ERRATA_12;=0A+=0A =09if (priv->errata)=0A =09=09dev_info(dev, "=
enabled errata workarounds, flags: 0x%x\n",=0A =09=09=09 priv->errata);=0A@=
@ -2156,8 +2161,20 @@ static int gfar_start_xmit(struct sk_buff *skb, struc=
t net_device *dev)=0A =09/* Set up checksumming */=0A =09if (CHECKSUM_PARTI=
AL =3D=3D skb->ip_summed) {=0A =09=09fcb =3D gfar_add_fcb(skb);=0A-=09=09ls=
tatus |=3D BD_LFLAG(TXBD_TOE);=0A-=09=09gfar_tx_checksum(skb, fcb);=0A+=09=
=09switch (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))=0A+=09=09=09? 1 =
: 0) {=0A+=09=09case 1:=0A+=09=09=09/* as specified by errata */=0A+=09=09=
=09if (((unsigned long)fcb % 0x20) > 0x18) { =0A+=09=09=09=09__skb_pull(skb=
, GMAC_FCB_LEN);=0A+=09=09=09=09skb_checksum_help(skb);=0A+=09=09=09=09brea=
k;=0A+=09=09=09}=0A+=09=09=09/* otherwise, fall through */=0A+=09=09default=
:=0A+=09=09=09lstatus |=3D BD_LFLAG(TXBD_TOE);=0A+=09=09=09gfar_tx_checksum=
(skb, fcb);=0A+=09=09}=0A =09}=0A =0A =09if (vlan_tx_tag_present(skb)) {=0A=
diff --git a/drivers/net/gianfar.h b/drivers/net/gianfar.h=0Aindex 54de413.=
.ec5d595 100644=0A--- a/drivers/net/gianfar.h=0A+++ b/drivers/net/gianfar.h=
=0A@@ -1039,6 +1039,7 @@ enum gfar_errata {=0A =09GFAR_ERRATA_74=09=09=3D 0=
x01,=0A =09GFAR_ERRATA_76=09=09=3D 0x02,=0A =09GFAR_ERRATA_A002=09=3D 0x04,=
=0A+=09GFAR_ERRATA_12=09=09=3D 0x08, /* a.k.a errata eTSEC49 */=0A };=0A =
=0A /* Struct stolen almost completely (and shamelessly) from the FCC enet =
source=0A-- =0A1.7.3.2=0A=0A=0A=0A=0A      

Re: [PATCH] gianfar: Fall back to software tcp/udp checksum on older controllers

[Anton Vorontsov]

Hello Alex,

On Thu, Jan 27, 2011 at 12:14:21AM -0800, Alex Dubov wrote:
> As specified by errata eTSEC49 of MPC8548 and errata eTSEC12 of MPC83xx,
> older revisions of gianfar controllers will be unable to calculate a TCP/UDP
> packet checksum for some aligments of the appropriate FCB. This patch checks
> for FCB alignment on such controllers and falls back to software checksumming
> if the aligment is known to be bad.
> 
> Signed-off-by: Alex Dubov <oakad@yahoo.com>
> ---
> This is my, somewhat different approach to Matthew Creech proposed solution.
> 
>  drivers/net/gianfar.c |   21 +++++++++++++++++++--
>  drivers/net/gianfar.h |    1 +
>  2 files changed, 20 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c
> index 5ed8f9f..b4f0e99 100644
> --- a/drivers/net/gianfar.c
> +++ b/drivers/net/gianfar.c
> @@ -950,6 +950,11 @@ static void gfar_detect_errata(struct gfar_private *priv)
>  			(pvr == 0x80861010 && (mod & 0xfff9) == 0x80c0))
>  		priv->errata |= GFAR_ERRATA_A002;
>  
> +	/* MPC8313 Rev < 2.0, MPC8548 rev 2.0 */
> +	if ((pvr == 0x80850010 && mod == 0x80b0 && rev < 0x0020)
> +	    || (pvr == 0x80210020 && mod == 0x8030 && rev == 0x0020))

Please indent it like the above: with two tabs. This is
to keep things consistent.

> +		priv->errata |= GFAR_ERRATA_12;
> +
>  	if (priv->errata)
>  		dev_info(dev, "enabled errata workarounds, flags: 0x%x\n",
>  			 priv->errata);
> @@ -2156,8 +2161,20 @@ static int gfar_start_xmit(struct sk_buff *skb, struct net_device *dev)
>  	/* Set up checksumming */
>  	if (CHECKSUM_PARTIAL == skb->ip_summed) {
>  		fcb = gfar_add_fcb(skb);
> -		lstatus |= BD_LFLAG(TXBD_TOE);
> -		gfar_tx_checksum(skb, fcb);
> +		switch (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))
> +			? 1 : 0) {

The switch construction is quite bizarre. ;-) Why not

if (gfar_has_errata() && (ulong)fcb % 0x20 > 18) {
	csum_help();
} else {
	lstatus |=...
	tx_csum();
}

?

Thanks,

> +		case 1:
> +			/* as specified by errata */
> +			if (((unsigned long)fcb % 0x20) > 0x18) { 
> +				__skb_pull(skb, GMAC_FCB_LEN);
> +				skb_checksum_help(skb);
> +				break;
> +			}
> +			/* otherwise, fall through */
> +		default:
> +			lstatus |= BD_LFLAG(TXBD_TOE);
> +			gfar_tx_checksum(skb, fcb);
> +		}
>  	}

-- 
Anton Vorontsov
Email: cbouatmailru@gmail.com

Re: [PATCH] gianfar: Fall back to software tcp/udp checksum on older controllers

[David Miller]

Please CC: netdev on future submissions of this patch, otherwise I
won't see it in my queue.

Re: [PATCH] gianfar: Fall back to software tcp/udp checksum on older controllers

[Alex Dubov]

> > -=A0=A0=A0 =A0=A0=A0=0A> gfar_tx_checksum(skb, fcb);=0A> > +=A0=A0=A0 =
=A0=A0=A0 switch=0A> (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))=0A> >=
 +=A0=A0=A0 =A0=A0=A0=0A> =A0=A0=A0 ? 1 : 0) {=0A> =0A> The switch construc=
tion is quite bizarre. ;-) Why not=0A> =0A=0AMy testing shows that even on =
broken chips unaligned packets are quite rare=0A(only some SYN packets seem=
 to be affected). So I tried to leave the=0Achecksum offload path as the mo=
st preferred one.=0A=0A=0A=0A      

[PATCH v2] gianfar: Fall back to software tcp/udp checksum on older controllers

[Alex Dubov]

As specified by errata eTSEC49 of MPC8548 and errata eTSEC12 of MPC83xx,
older revisions of gianfar controllers will be unable to calculate a TCP/UDP
packet checksum for some alignments of the appropriate FCB. This patch checks
for FCB alignment on such controllers and falls back to software checksumming
if the alignment is known to be bad.

Signed-off-by: Alex Dubov <oakad@yahoo.com>
---
Changes for v2:
   - Make indentation slightly more consistent.
   - Replace bizarre switch-based condition with plain boring one.

 drivers/net/gianfar.c |   16 ++++++++++++++--
 drivers/net/gianfar.h |    1 +
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c
index 5ed8f9f..3da19a5 100644
--- a/drivers/net/gianfar.c
+++ b/drivers/net/gianfar.c
@@ -950,6 +950,11 @@ static void gfar_detect_errata(struct gfar_private *priv)
 			(pvr == 0x80861010 && (mod & 0xfff9) == 0x80c0))
 		priv->errata |= GFAR_ERRATA_A002;
 
+	/* MPC8313 Rev < 2.0, MPC8548 rev 2.0 */
+	if ((pvr == 0x80850010 && mod == 0x80b0 && rev < 0x0020) ||
+			(pvr == 0x80210020 && mod == 0x8030 && rev == 0x0020))
+		priv->errata |= GFAR_ERRATA_12;
+
 	if (priv->errata)
 		dev_info(dev, "enabled errata workarounds, flags: 0x%x\n",
 			 priv->errata);
@@ -2156,8 +2161,15 @@ static int gfar_start_xmit(struct sk_buff *skb, struct net_device *dev)
 	/* Set up checksumming */
 	if (CHECKSUM_PARTIAL == skb->ip_summed) {
 		fcb = gfar_add_fcb(skb);
-		lstatus |= BD_LFLAG(TXBD_TOE);
-		gfar_tx_checksum(skb, fcb);
+		/* as specified by errata */
+		if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)
+			     && ((unsigned long)fcb % 0x20) > 0x18)) {
+			__skb_pull(skb, GMAC_FCB_LEN);
+			skb_checksum_help(skb);
+		} else {
+			lstatus |= BD_LFLAG(TXBD_TOE);
+			gfar_tx_checksum(skb, fcb);
+		}
 	}
 
 	if (vlan_tx_tag_present(skb)) {
diff --git a/drivers/net/gianfar.h b/drivers/net/gianfar.h
index 54de413..ec5d595 100644
--- a/drivers/net/gianfar.h
+++ b/drivers/net/gianfar.h
@@ -1039,6 +1039,7 @@ enum gfar_errata {
 	GFAR_ERRATA_74		= 0x01,
 	GFAR_ERRATA_76		= 0x02,
 	GFAR_ERRATA_A002	= 0x04,
+	GFAR_ERRATA_12		= 0x08, /* a.k.a errata eTSEC49 */
 };
 
 /* Struct stolen almost completely (and shamelessly) from the FCC enet source
-- 
1.7.3.2




      

[PATCH v2] gianfar: Fall back to software tcp/udp checksum on older controllers

[Alex Dubov]

As specified by errata eTSEC49 of MPC8548 and errata eTSEC12 of MPC83xx,=0A=
older revisions of gianfar controllers will be unable to calculate a TCP/UD=
P=0Apacket checksum for some alignments of the appropriate FCB. This patch =
checks=0Afor FCB alignment on such controllers and falls back to software c=
hecksumming=0Aif the alignment is known to be bad.=0A=0ASigned-off-by: Alex=
 Dubov <oakad@yahoo.com>=0A---=0AChanges for v2:=0A   - Make indentation sl=
ightly more consistent.=0A   - Replace bizarre switch-based condition with =
plain boring one.=0A=0A drivers/net/gianfar.c |   16 ++++++++++++++--=0A dr=
ivers/net/gianfar.h |    1 +=0A 2 files changed, 15 insertions(+), 2 deleti=
ons(-)=0A=0Adiff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c=0Ain=
dex 5ed8f9f..3da19a5 100644=0A--- a/drivers/net/gianfar.c=0A+++ b/drivers/n=
et/gianfar.c=0A@@ -950,6 +950,11 @@ static void gfar_detect_errata(struct g=
far_private *priv)=0A =09=09=09(pvr =3D=3D 0x80861010 && (mod & 0xfff9) =3D=
=3D 0x80c0))=0A =09=09priv->errata |=3D GFAR_ERRATA_A002;=0A =0A+=09/* MPC8=
313 Rev < 2.0, MPC8548 rev 2.0 */=0A+=09if ((pvr =3D=3D 0x80850010 && mod =
=3D=3D 0x80b0 && rev < 0x0020) ||=0A+=09=09=09(pvr =3D=3D 0x80210020 && mod=
 =3D=3D 0x8030 && rev =3D=3D 0x0020))=0A+=09=09priv->errata |=3D GFAR_ERRAT=
A_12;=0A+=0A =09if (priv->errata)=0A =09=09dev_info(dev, "enabled errata wo=
rkarounds, flags: 0x%x\n",=0A =09=09=09 priv->errata);=0A@@ -2156,8 +2161,1=
5 @@ static int gfar_start_xmit(struct sk_buff *skb, struct net_device *dev=
)=0A =09/* Set up checksumming */=0A =09if (CHECKSUM_PARTIAL =3D=3D skb->ip=
_summed) {=0A =09=09fcb =3D gfar_add_fcb(skb);=0A-=09=09lstatus |=3D BD_LFL=
AG(TXBD_TOE);=0A-=09=09gfar_tx_checksum(skb, fcb);=0A+=09=09/* as specified=
 by errata */=0A+=09=09if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)=
=0A+=09=09=09     && ((unsigned long)fcb % 0x20) > 0x18)) {=0A+=09=09=09__s=
kb_pull(skb, GMAC_FCB_LEN);=0A+=09=09=09skb_checksum_help(skb);=0A+=09=09} =
else {=0A+=09=09=09lstatus |=3D BD_LFLAG(TXBD_TOE);=0A+=09=09=09gfar_tx_che=
cksum(skb, fcb);=0A+=09=09}=0A =09}=0A =0A =09if (vlan_tx_tag_present(skb))=
 {=0Adiff --git a/drivers/net/gianfar.h b/drivers/net/gianfar.h=0Aindex 54d=
e413..ec5d595 100644=0A--- a/drivers/net/gianfar.h=0A+++ b/drivers/net/gian=
far.h=0A@@ -1039,6 +1039,7 @@ enum gfar_errata {=0A =09GFAR_ERRATA_74=09=09=
=3D 0x01,=0A =09GFAR_ERRATA_76=09=09=3D 0x02,=0A =09GFAR_ERRATA_A002=09=3D =
0x04,=0A+=09GFAR_ERRATA_12=09=09=3D 0x08, /* a.k.a errata eTSEC49 */=0A };=
=0A =0A /* Struct stolen almost completely (and shamelessly) from the FCC e=
net source=0A-- =0A1.7.3.2=0A=0A=0A=0A=0A      

RE: [PATCH v2] gianfar: Fall back to software tcp/udp checksum on oldercontrollers

[David Laight]

 
> +		if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)
> +			     && ((unsigned long)fcb % 0x20) > 0x18)) {

You need to check the generated code, but I think you need:

    if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))
	     && unlikely(((unsigned long)fcb % 0x20) > 0x18))

ie unlikely() around both the primitive comparisons.

	David

RE: [PATCH v2] gianfar: Fall back to software tcp/udp checksum on oldercontrollers

[David Laight]

=20
> +		if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)
> +			     && ((unsigned long)fcb % 0x20) > 0x18)) {

You need to check the generated code, but I think you need:

    if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))
	     && unlikely(((unsigned long)fcb % 0x20) > 0x18))

ie unlikely() around both the primitive comparisons.

	David

Re: [PATCH v2] gianfar: Fall back to software tcp/udp checksum on oldercontrollers

[Scott Wood]

On Fri, 28 Jan 2011 09:10:46 +0000
David Laight <David.Laight@ACULAB.COM> wrote:

>  
> > +		if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)
> > +			     && ((unsigned long)fcb % 0x20) > 0x18)) {
> 
> You need to check the generated code, but I think you need:
> 
>     if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))
> 	     && unlikely(((unsigned long)fcb % 0x20) > 0x18))
> 
> ie unlikely() around both the primitive comparisons.

Is the first condition actually unlikely?  If you've got affected
hardware, you'll hit it every time.

If packets with the problematic alignment are rare, seems like it'd be
better to check that first.

-Scott

Re: [PATCH v2] gianfar: Fall back to software tcp/udp checksum on oldercontrollers

[Scott Wood]

On Fri, 28 Jan 2011 09:10:46 +0000
David Laight <David.Laight@ACULAB.COM> wrote:

>  
> > +		if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)
> > +			     && ((unsigned long)fcb % 0x20) > 0x18)) {
> 
> You need to check the generated code, but I think you need:
> 
>     if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))
> 	     && unlikely(((unsigned long)fcb % 0x20) > 0x18))
> 
> ie unlikely() around both the primitive comparisons.

Is the first condition actually unlikely?  If you've got affected
hardware, you'll hit it every time.

If packets with the problematic alignment are rare, seems like it'd be
better to check that first.

-Scott

Re: [PATCH v2] gianfar: Fall back to software tcp/udp checksum on oldercontrollers

[David Miller]

From: Scott Wood <scottwood@freescale.com>
Date: Fri, 28 Jan 2011 10:56:10 -0600

> On Fri, 28 Jan 2011 09:10:46 +0000
> David Laight <David.Laight@ACULAB.COM> wrote:
> 
>>  
>> > +		if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)
>> > +			     && ((unsigned long)fcb % 0x20) > 0x18)) {
>> 
>> You need to check the generated code, but I think you need:
>> 
>>     if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))
>> 	     && unlikely(((unsigned long)fcb % 0x20) > 0x18))
>> 
>> ie unlikely() around both the primitive comparisons.
> 
> Is the first condition actually unlikely?  If you've got affected
> hardware, you'll hit it every time.
> 
> If packets with the problematic alignment are rare, seems like it'd be
> better to check that first.

In cases like this gfar_has_errata() case, better to leave it's
likelyhood unmarked.

And yes, since it's cheaper, checking the alignment should be done
first.

Re: [PATCH v2] gianfar: Fall back to software tcp/udp checksum on oldercontrollers

[David Miller]

From: Scott Wood <scottwood@freescale.com>
Date: Fri, 28 Jan 2011 10:56:10 -0600

> On Fri, 28 Jan 2011 09:10:46 +0000
> David Laight <David.Laight@ACULAB.COM> wrote:
> 
>>  
>> > +		if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12)
>> > +			     && ((unsigned long)fcb % 0x20) > 0x18)) {
>> 
>> You need to check the generated code, but I think you need:
>> 
>>     if (unlikely(gfar_has_errata(priv, GFAR_ERRATA_12))
>> 	     && unlikely(((unsigned long)fcb % 0x20) > 0x18))
>> 
>> ie unlikely() around both the primitive comparisons.
> 
> Is the first condition actually unlikely?  If you've got affected
> hardware, you'll hit it every time.
> 
> If packets with the problematic alignment are rare, seems like it'd be
> better to check that first.

In cases like this gfar_has_errata() case, better to leave it's
likelyhood unmarked.

And yes, since it's cheaper, checking the alignment should be done
first.