[PATCH] r8169: fix invalid register dump

[PATCH] r8169: fix invalid register dump

[Peter Wu]

From: Peter Wu <lekensteyn@gmail.com>

For some reason, my PCIe RTL8111E onboard NIC on a GA-Z68X-UD3H-B3
motherboard reads as FFs when reading from MMIO with a block size
larger than 7. Therefore change to reading blocks of four bytes.

Signed-off-by: Peter Wu <lekensteyn@gmail.com>
---
Hi,

Personally I dislike the repetition of this patch, a hypothetical
`memcpy_fromio_block(bytes, tp->mmio_addr, 4)` would look cleaner, but
perhaps I am doing something wrong because apparantly nobody had a need
for reading multiple blocks from IO?

Anyway, this patch (with patched ethtool) allows `ethtool -d` to be more
useful for the mentioned chip.

Regards,
Peter
---
 drivers/net/ethernet/realtek/r8169.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/realtek/r8169.c b/drivers/net/ethernet/realtek/r8169.c
index b5eb419..28af01c 100644
--- a/drivers/net/ethernet/realtek/r8169.c
+++ b/drivers/net/ethernet/realtek/r8169.c
@@ -1897,12 +1897,17 @@ static void rtl8169_get_regs(struct net_device *dev, struct ethtool_regs *regs,
 			     void *p)
 {
 	struct rtl8169_private *tp = netdev_priv(dev);
+	char *bytes = p;
+	int i;
 
 	if (regs->len > R8169_REGS_SIZE)
 		regs->len = R8169_REGS_SIZE;
 
 	rtl_lock_work(tp);
-	memcpy_fromio(p, tp->mmio_addr, regs->len);
+	for (i = 0; i < regs->len - 4; i += 4)
+		memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);
+	if (i < regs->len)
+		memcpy_fromio(bytes + i, tp->mmio_addr + i, regs->len - i);
 	rtl_unlock_work(tp);
 }
 
-- 
1.8.3.4

Re: [PATCH] r8169: fix invalid register dump

[Francois Romieu]

Peter Wu <lekensteyn@gmail.com> :
[...]
> diff --git a/drivers/net/ethernet/realtek/r8169.c b/drivers/net/ethernet/realtek/r8169.c
> index b5eb419..28af01c 100644
> --- a/drivers/net/ethernet/realtek/r8169.c
> +++ b/drivers/net/ethernet/realtek/r8169.c
> @@ -1897,12 +1897,17 @@ static void rtl8169_get_regs(struct net_device *dev, struct ethtool_regs *regs,
[...]
> -	memcpy_fromio(p, tp->mmio_addr, regs->len);
> +	for (i = 0; i < regs->len - 4; i += 4)
> +		memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);

regs->len is u32. regs->len - 4 should not underflow on normal situation but
it's a bit sloppy.

-- 
Ueimor

Re: [PATCH] r8169: fix invalid register dump

[Peter Wu]

On Wednesday 14 August 2013 00:20:38 Francois Romieu wrote:
> > -     memcpy_fromio(p, tp->mmio_addr, regs->len);
> > +     for (i = 0; i < regs->len - 4; i += 4)
> > +             memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);
> 
> regs->len is u32. regs->len - 4 should not underflow on normal situation but
> it's a bit sloppy.

Heh, I forgot about that case. Fixed it in the below patch. By the way,
have you observed similar behavior on other hardware?

Thanks,
Peter
--
From: Peter Wu <lekensteyn@gmail.com>

For some reason, my PCIe RTL8111E onboard NIC on a GA-Z68X-UD3H-B3
motherboard reads as FFs when reading from MMIO with a block size
larger than 7. Therefore change to reading blocks of four bytes.

Signed-off-by: Peter Wu <lekensteyn@gmail.com>
---
 drivers/net/ethernet/realtek/r8169.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/realtek/r8169.c b/drivers/net/ethernet/realtek/r8169.c
index b5eb419..d367763 100644
--- a/drivers/net/ethernet/realtek/r8169.c
+++ b/drivers/net/ethernet/realtek/r8169.c
@@ -1897,12 +1897,19 @@ static void rtl8169_get_regs(struct net_device *dev, struct ethtool_regs *regs,
 			     void *p)
 {
 	struct rtl8169_private *tp = netdev_priv(dev);
+	char *bytes = p;
+	int i;
 
 	if (regs->len > R8169_REGS_SIZE)
 		regs->len = R8169_REGS_SIZE;
 
 	rtl_lock_work(tp);
-	memcpy_fromio(p, tp->mmio_addr, regs->len);
+	if (regs->len >= 4) {
+		for (i = 0; i < regs->len - 4; i += 4)
+			memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);
+	}
+	if (i < regs->len)
+		memcpy_fromio(bytes + i, tp->mmio_addr + i, regs->len - i);
 	rtl_unlock_work(tp);
 }
 
-- 
1.8.3.4

Re: [PATCH] r8169: fix invalid register dump

[Francois Romieu]

Peter Wu <lekensteyn@gmail.com> :
[...]
> Heh, I forgot about that case. Fixed it in the below patch. By the way,
> have you observed similar behavior on other hardware?

Notwithstanding really old PCI adapters - sometimes - none that I remember of.

[...]
> diff --git a/drivers/net/ethernet/realtek/r8169.c b/drivers/net/ethernet/realtek/r8169.c
> index b5eb419..d367763 100644
> --- a/drivers/net/ethernet/realtek/r8169.c
> +++ b/drivers/net/ethernet/realtek/r8169.c
> @@ -1897,12 +1897,19 @@ static void rtl8169_get_regs(struct net_device *dev, struct ethtool_regs *regs,
>  			     void *p)
>  {
>  	struct rtl8169_private *tp = netdev_priv(dev);
> +	char *bytes = p;
> +	int i;
>  
>  	if (regs->len > R8169_REGS_SIZE)
>  		regs->len = R8169_REGS_SIZE;
>  
>  	rtl_lock_work(tp);
> -	memcpy_fromio(p, tp->mmio_addr, regs->len);
> +	if (regs->len >= 4) {
> +		for (i = 0; i < regs->len - 4; i += 4)
> +			memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);
> +	}
> +	if (i < regs->len)

Comparison with random stack stuff when regs->len < 4. :o/

-- 
Ueimor

Re: [PATCH] r8169: fix invalid register dump

[Peter Wu]

On Wednesday 14 August 2013 21:58:29 Francois Romieu wrote:
> > -     memcpy_fromio(p, tp->mmio_addr, regs->len);
> > +     if (regs->len >= 4) {
> > +             for (i = 0; i < regs->len - 4; i += 4)
> > +                     memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);
> > +     }
> > +     if (i < regs->len)
> 
> Comparison with random stack stuff when regs->len < 4. :o/

Right, let's rm $OLD_PATCH and consider this one.

Checklist:
1. super large regs->len: won't be greater than R8169_REGS_SIZE (256)
2. regs->len == 0: 0 < 0 is false, nothing is copied
3. regs->len is 1, 2 or 3: i = 0, at most 3 bytes will be copied
4. regs->len is 4, i < 4 - 4, skip loop, 0 < regs->len, copy 4
5. regs->len is 5, i < 5 - 4, copy; 4 < regs->len, copy 1

With this I can now say with confidence that I haven't overlooked something
related to integer overflow. You have a very sharp eye, thanks for
catching my mistakes.

Regards,
Peter
---
From: Peter Wu <lekensteyn@gmail.com>

For some reason, my PCIe RTL8111E onboard NIC on a GA-Z68X-UD3H-B3
motherboard reads as FFs when reading from MMIO with a block size
larger than 7. Therefore change to reading blocks of four bytes.

Signed-off-by: Peter Wu <lekensteyn@gmail.com>
---
 drivers/net/ethernet/realtek/r8169.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/realtek/r8169.c b/drivers/net/ethernet/realtek/r8169.c
index b5eb419..19524c0 100644
--- a/drivers/net/ethernet/realtek/r8169.c
+++ b/drivers/net/ethernet/realtek/r8169.c
@@ -1897,12 +1897,19 @@ static void rtl8169_get_regs(struct net_device *dev, struct ethtool_regs *regs,
 			     void *p)
 {
 	struct rtl8169_private *tp = netdev_priv(dev);
+	char *bytes = p;
+	int i = 0;
 
 	if (regs->len > R8169_REGS_SIZE)
 		regs->len = R8169_REGS_SIZE;
 
 	rtl_lock_work(tp);
-	memcpy_fromio(p, tp->mmio_addr, regs->len);
+	if (regs->len >= 4) {
+		for (; i < regs->len - 4; i += 4)
+			memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);
+	}
+	if (i < regs->len)
+		memcpy_fromio(bytes + i, tp->mmio_addr + i, regs->len - i);
 	rtl_unlock_work(tp);
 }
 
-- 
1.8.3.4

Re: [PATCH] r8169: fix invalid register dump

[Francois Romieu]

Peter Wu <lekensteyn@gmail.com> :
[...]
> Checklist:
> 1. super large regs->len: won't be greater than R8169_REGS_SIZE (256)
> 2. regs->len == 0: 0 < 0 is false, nothing is copied
> 3. regs->len is 1, 2 or 3: i = 0, at most 3 bytes will be copied
> 4. regs->len is 4, i < 4 - 4, skip loop, 0 < regs->len, copy 4
> 5. regs->len is 5, i < 5 - 4, copy; 4 < regs->len, copy 1

Spartan implementation:

	u32 __iomem *ioaddr = tp->mmio_addr;
	u32 *dw = p;
	int i;

	...

	for (i = regs->len; i > 0; i -= 4)
		memcpy_fromio(dw++, ioaddr++, min(i, 4));

or (min() verges on gluttony):

	for (i = regs->len; i >= 4; i -= 4)
		memcpy_fromio(dw++, ioaddr++, 4);

	if (i > 0)
		memcpy_fromio(dw, ioaddr, i);

-- 
Ueimor

Re: [PATCH] r8169: fix invalid register dump

[Ben Hutchings]

On Wed, 2013-08-14 at 23:31 +0200, Peter Wu wrote:
> On Wednesday 14 August 2013 21:58:29 Francois Romieu wrote:
> > > -     memcpy_fromio(p, tp->mmio_addr, regs->len);
> > > +     if (regs->len >= 4) {
> > > +             for (i = 0; i < regs->len - 4; i += 4)
> > > +                     memcpy_fromio(bytes + i, tp->mmio_addr + i, 4);
> > > +     }
> > > +     if (i < regs->len)
> > 
> > Comparison with random stack stuff when regs->len < 4. :o/
> 
> Right, let's rm $OLD_PATCH and consider this one.
> 
> Checklist:
> 1. super large regs->len: won't be greater than R8169_REGS_SIZE (256)
> 2. regs->len == 0: 0 < 0 is false, nothing is copied
> 3. regs->len is 1, 2 or 3: i = 0, at most 3 bytes will be copied
> 4. regs->len is 4, i < 4 - 4, skip loop, 0 < regs->len, copy 4
> 5. regs->len is 5, i < 5 - 4, copy; 4 < regs->len, copy 1
[...]

The kernel buffer size is max(regs->len,
dev->ethtool_ops->get_regs_len()).  So you can safely ignore regs->len
and always read all your registers.

Ben.

-- 
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: [PATCH] r8169: fix invalid register dump

[Peter Wu]

On Thursday 15 August 2013 00:47:36 Francois Romieu wrote:
> Spartan implementation:
> 
>         u32 __iomem *ioaddr = tp->mmio_addr;
>         u32 *dw = p;
>         int i;
> 
>         ...
> 
>         for (i = regs->len; i > 0; i -= 4)
>                 memcpy_fromio(dw++, ioaddr++, min(i, 4));

I like this one, simple and readable. I was worried about endianness
issues, but that does not apply here as the source and destination
are presumably in byte order.

I have tested this patch with a modified ethtool[1].

 [1]: http://www.spinics.net/lists/netdev/msg246364.html
--
From: Peter Wu <lekensteyn@gmail.com>

For some reason, my PCIe RTL8111E onboard NIC on a GA-Z68X-UD3H-B3
motherboard reads as FFs when reading from MMIO with a block size
larger than 7. Therefore change to reading blocks of four bytes.

Signed-off-by: Peter Wu <lekensteyn@gmail.com>
---
 drivers/net/ethernet/realtek/r8169.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/realtek/r8169.c 
b/drivers/net/ethernet/realtek/r8169.c
index b5eb419..2943916 100644
--- a/drivers/net/ethernet/realtek/r8169.c
+++ b/drivers/net/ethernet/realtek/r8169.c
@@ -1897,12 +1897,16 @@ static void rtl8169_get_regs(struct net_device *dev, 
struct ethtool_regs *regs,
 			     void *p)
 {
 	struct rtl8169_private *tp = netdev_priv(dev);
+	u32 __iomem *data = tp->mmio_addr;
+	u32 *dw = p;
+	int i;
 
 	if (regs->len > R8169_REGS_SIZE)
 		regs->len = R8169_REGS_SIZE;
 
 	rtl_lock_work(tp);
-	memcpy_fromio(p, tp->mmio_addr, regs->len);
+	for (i = regs->len; i > 0; i -= 4)
+		memcpy_fromio(dw++, data++, min(4, i));
 	rtl_unlock_work(tp);
 }
 
-- 
1.8.3.4

Re: [PATCH] r8169: fix invalid register dump

[Peter Wu]

On Thursday 15 August 2013 22:33:15 Ben Hutchings wrote:
> The kernel buffer size is max(regs->len,
> dev->ethtool_ops->get_regs_len()).  So you can safely ignore regs->len
> and always read all your registers.

I see, that is something for a different patch though. While I am at it,
I checked all users of get_regs_len in drivers/net and found only one
other user that also checked the length. Patch will follow shortly.