* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Eli Cohen @ 2011-10-09 10:30 UTC (permalink / raw)
To: Benjamin Herrenschmidt
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <1318153939.29415.401.camel@pasglop>
On Sun, Oct 09, 2011 at 11:52:19AM +0200, Benjamin Herrenschmidt wrote:
>
> > > To go back to the driver code, the statements that ring a "bell" are:
> > >
> > > *(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn;
> > >
> > > This doesn't look right unless "doorbell_qpn" itself is already somewhat
> > > in the appropriate byte order.
>
> > This is something that supports my claim that the chipset swaps
> > endianess in ppc.
>
> No the chipset doesn't swap
>
> > Look at mlx4_en_activate_tx_ring():
> > ring->doorbell_qpn = swab32(ring->qp.qpn << 8);
>
> That looks gross, I think somebody writing this driver doesn't
> understand endianness.
>
> > so in LE machines it layed as big endian in memory while in BE machines
> > it is layed as little endian in memory.
>
> Yes which is very odd, it should be layed out the same in memory
> regardless of the machine. However in this case, this isn't accessed
> directly via DMA (this field at least), so what you appear to be doing
> here is to artificially "undo" what writel does later on (see below).
>
> > Then we write this value to the device registers which must get it in
> > big endian otherwise it won't work - and we know this works in both
> > ppc and x86. You can ignore the case of blue flame:
>
> Well it works because your device is odd and has BE registers :-) It's
> however not the right way to do and it's broken in your blue flame case
> (for what is now obvious reasons, see below).
>
> What's happening basically here is that you are swapping once in
> swab32 , store that swapped value, then writel will re-swap on power and
> not swap on x86 (because the writel accessor performs swapping). So on
> x86 you basically do LE -> BE and on ppc you do BE -> LE -> BE :-)
> Pretty inefficient.
>
> None of this has anything to do with the chipset which doesn't swap
> anything behind your back.
>
> Ideally you want to avoid that swapping altogether and use the right
> accessor that indicates that your register is BE to start with. IE.
> remove the swab32 completely and then use something like
> iowrite32be() instead of writel().
I agree, this looks better but does it work on memory mapped io or
only on io pci space? All our registers are memory mapped...
>
> Basically, the problem you have is that writel() has an implicit "write
> to LE register" semantic. Your register is BE. the "iomap" variants
> provide you with more fine grained "be" variants to use in that case.
> There's also writel_be() but that one doesn't exist on every
> architecture afaik.
So writel_be is the function I should use for memory mapped io? If it
does not exist for all platforms it's a pitty :-(
>
> Now, once the mmio problem is out of the way, let's look back at how you
> then use that qpn.
>
> With the current code, you've generated something in memory which is
> byte reversed, so essentially "LE" on ppc and "BE" on x86.
>
> Then, this statement:
>
> *(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn;
>
> Will essentially write it out as-is in memory for use by the chip. The chip,
> from what you say, expects BE, so this will be broken on PPC.
I see. So this field is layed in le for ppc and the rest of the
descriptor is be. so I assum that __iowrite64_copy() does not swap
anything but we still have tx_desc->ctrl.vlan_tag in the wrong
endianess.
>
> Here too, the right solution is to instead not do that swab32 to begin
> with (ring->doorbell_qpn remains a native endian value) and instead do,
> in addition to the above mentioned change to the writel:
>
> *(u32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
>
> (Also get rid of that cast and define vlan_tag as a __be32 to start
> with).
>
> Cheers,
> Ben.
Thanks for your review. I will send another patch which should fix the
deficiencies.
^ permalink raw reply
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Benjamin Herrenschmidt @ 2011-10-09 9:52 UTC (permalink / raw)
To: Eli Cohen
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <20111009092102.GK2681@mtldesk30>
> > To go back to the driver code, the statements that ring a "bell" are:
> >
> > *(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn;
> >
> > This doesn't look right unless "doorbell_qpn" itself is already somewhat
> > in the appropriate byte order.
> This is something that supports my claim that the chipset swaps
> endianess in ppc.
No the chipset doesn't swap
> Look at mlx4_en_activate_tx_ring():
> ring->doorbell_qpn = swab32(ring->qp.qpn << 8);
That looks gross, I think somebody writing this driver doesn't
understand endianness.
> so in LE machines it layed as big endian in memory while in BE machines
> it is layed as little endian in memory.
Yes which is very odd, it should be layed out the same in memory
regardless of the machine. However in this case, this isn't accessed
directly via DMA (this field at least), so what you appear to be doing
here is to artificially "undo" what writel does later on (see below).
> Then we write this value to the device registers which must get it in
> big endian otherwise it won't work - and we know this works in both
> ppc and x86. You can ignore the case of blue flame:
Well it works because your device is odd and has BE registers :-) It's
however not the right way to do and it's broken in your blue flame case
(for what is now obvious reasons, see below).
What's happening basically here is that you are swapping once in
swab32 , store that swapped value, then writel will re-swap on power and
not swap on x86 (because the writel accessor performs swapping). So on
x86 you basically do LE -> BE and on ppc you do BE -> LE -> BE :-)
Pretty inefficient.
None of this has anything to do with the chipset which doesn't swap
anything behind your back.
Ideally you want to avoid that swapping altogether and use the right
accessor that indicates that your register is BE to start with. IE.
remove the swab32 completely and then use something like
iowrite32be() instead of writel().
Basically, the problem you have is that writel() has an implicit "write
to LE register" semantic. Your register is BE. the "iomap" variants
provide you with more fine grained "be" variants to use in that case.
There's also writel_be() but that one doesn't exist on every
architecture afaik.
Now, once the mmio problem is out of the way, let's look back at how you
then use that qpn.
With the current code, you've generated something in memory which is
byte reversed, so essentially "LE" on ppc and "BE" on x86.
Then, this statement:
*(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn;
Will essentially write it out as-is in memory for use by the chip. The chip,
from what you say, expects BE, so this will be broken on PPC.
Here too, the right solution is to instead not do that swab32 to begin
with (ring->doorbell_qpn remains a native endian value) and instead do,
in addition to the above mentioned change to the writel:
*(u32 *) (&tx_desc->ctrl.vlan_tag) |= cpu_to_be32(ring->doorbell_qpn);
(Also get rid of that cast and define vlan_tag as a __be32 to start
with).
Cheers,
Ben.
^ permalink raw reply
* Re: [PATCH] gro: refetch inet6_protos[] after pulling ext headers
From: Eric Dumazet @ 2011-10-09 9:33 UTC (permalink / raw)
To: Yan, Zheng; +Cc: netdev@vger.kernel.org, Herbert Xu, davem@davemloft.net
In-Reply-To: <4E915C9B.4000807@intel.com>
Le dimanche 09 octobre 2011 à 16:34 +0800, Yan, Zheng a écrit :
> ipv6_gro_receive() doesn't update the protocol ops after pulling
> the ext headers. It looks like a typo.
>
> Signed-off-by: Zheng Yan <zheng.z.yan@intel.com>
> ---
> diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c
> index 3b5669a..d27c797 100644
> --- a/net/ipv6/af_inet6.c
> +++ b/net/ipv6/af_inet6.c
> @@ -875,6 +875,7 @@ static struct sk_buff **ipv6_gro_receive(struct sk_buff **head,
> skb_reset_transport_header(skb);
> __skb_push(skb, skb_gro_offset(skb));
>
> + ops = rcu_dereference(inet6_protos[proto]);
> if (!ops || !ops->gro_receive)
> goto out_unlock;
Good catch !
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
^ permalink raw reply
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Eli Cohen @ 2011-10-09 9:21 UTC (permalink / raw)
To: Benjamin Herrenschmidt
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <1318149536.29415.384.camel@pasglop>
On Sun, Oct 09, 2011 at 10:38:56AM +0200, Benjamin Herrenschmidt wrote:
> On Sun, 2011-10-09 at 10:07 +0200, Eli Cohen wrote:
>
> > > Well, first, what do you mean by "swapped" ? :-) But no, it won't for
> > > all intend and purpose, this is a copy routine, copy routines never
> > > swap, neither do fifo accesses for example.
> > When I say swapped, I mean not necessairliy by software. I think that
> > the chipset will swap the the data. The reason I think so is that the
> > CPU arch is big endian, while PCI bus is defined as little endian.
> > That's why I think a swap will occur in ppc and not in x86.
>
> No it won't "swap the data". The wiring between PCI and the CPU bus is
> done in a way called "byte address invariant", and there is some kind of
> flip of byte lanes related essentially to ensure that, but for all
> intend and purpose it's transparent.
>
> > It's a special descriptor that resides both in memory and also written
> > to the device's register. An it contains both data and control
> > informartion.
>
> Data should not be swapped then. Only individual bits of control
> information. In any case, the buffer should be generated in the right
> format in memory to start with. The copy routine doesn't need to swap.
>
> To go back to the driver code, the statements that ring a "bell" are:
>
> *(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn;
>
> This doesn't look right unless "doorbell_qpn" itself is already somewhat
> in the appropriate byte order.
This is something that supports my claim that the chipset swaps
endianess in ppc.
Look at mlx4_en_activate_tx_ring():
ring->doorbell_qpn = swab32(ring->qp.qpn << 8);
so in LE machines it layed as big endian in memory while in BE machines
it is layed as little endian in memory.
Then we write this value to the device registers which must get it in
big endian otherwise it won't work - and we know this works in both
ppc and x86. You can ignore the case of blue flame:
} else if (nreq) {
qp->sq.head += nreq;
/*
* Make sure that descriptors are written before
* doorbell record.
*/
wmb();
writel(qp->doorbell_qpn, qp->bf.uar->map +
MLX4_SEND_DOORBELL); <== remember that it is layed in little endian
but the device must get it in big endian.
/*
* Make sure doorbells don't leak out of SQ spinlock
* and reach the HCA out of order.
*/
mmiowb();
}
>
> Is that vlan_tag a big or little endian quantity ? Either way, this
> looks broken in either x86 or ppc unless doorbell_qpn itself is already
> in the right endian.
>
> But since later I see
>
> writel(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL);
>
> That rings a nasty bell, it looks like doorbell_pqn is in CPU (native)
> endian, so it should have to be swapped before being OR'ed into the
> descriptor, either that or the HW does some black magic I don't
> understand.
>
> Cheers,
> Ben.
>
^ permalink raw reply
* [PATCH] r8169: Fix WOL in power down case
From: Marc Ballarin @ 2011-10-09 8:49 UTC (permalink / raw)
To: Francois Romieu; +Cc: hayeswang, netdev
In-Reply-To: <20111007232834.GA29770@electric-eye.fr.zoreil.com>
Commit 92fc43b4159b518f5baae57301f26d770b0834c9 ("r8169: modify the flow of the
hw reset.") breaks WOL on some versions of the hardware.
Commit 106633897e086e1b47126996aac1a427eb80eb1b ("r8169: fix WOL setting for
8105 and 8111evl") tries to fix this, but only does so in the standby case.
This patch applies an analogous fix to the shutdown path.
Signed-off-by: Marc Ballarin <ballarin.marc@gmx.de>
---
diff --git a/drivers/net/r8169.c b/drivers/net/r8169.c
index c236670..a2d6e3a 100644
--- a/drivers/net/r8169.c
+++ b/drivers/net/r8169.c
@@ -5817,6 +5817,15 @@ static void rtl_shutdown(struct pci_dev *pdev)
/* PCI commit */
RTL_R8(ChipCmd);
}
+ /* rtl8105, rtl8111E, and rtl8111evl need the following bits
+ * set for WOL to work */
+ if ((tp->mac_version == RTL_GIGA_MAC_VER_32 ||
+ tp->mac_version == RTL_GIGA_MAC_VER_33 ||
+ tp->mac_version == RTL_GIGA_MAC_VER_34) &&
+ (__rtl8169_get_wol(tp) & WAKE_ANY)) {
+ RTL_W32(RxConfig, RTL_R32(RxConfig) | AcceptBroadcast |
+ AcceptMulticast | AcceptMyPhys);
+ }
pci_wake_from_d3(pdev, true);
pci_set_power_state(pdev, PCI_D3hot);
^ permalink raw reply related
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Benjamin Herrenschmidt @ 2011-10-09 8:38 UTC (permalink / raw)
To: Eli Cohen
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <20111009080747.GJ2681@mtldesk30>
On Sun, 2011-10-09 at 10:07 +0200, Eli Cohen wrote:
> > Well, first, what do you mean by "swapped" ? :-) But no, it won't for
> > all intend and purpose, this is a copy routine, copy routines never
> > swap, neither do fifo accesses for example.
> When I say swapped, I mean not necessairliy by software. I think that
> the chipset will swap the the data. The reason I think so is that the
> CPU arch is big endian, while PCI bus is defined as little endian.
> That's why I think a swap will occur in ppc and not in x86.
No it won't "swap the data". The wiring between PCI and the CPU bus is
done in a way called "byte address invariant", and there is some kind of
flip of byte lanes related essentially to ensure that, but for all
intend and purpose it's transparent.
> It's a special descriptor that resides both in memory and also written
> to the device's register. An it contains both data and control
> informartion.
Data should not be swapped then. Only individual bits of control
information. In any case, the buffer should be generated in the right
format in memory to start with. The copy routine doesn't need to swap.
To go back to the driver code, the statements that ring a "bell" are:
*(u32 *) (&tx_desc->ctrl.vlan_tag) |= ring->doorbell_qpn;
This doesn't look right unless "doorbell_qpn" itself is already somewhat
in the appropriate byte order.
Is that vlan_tag a big or little endian quantity ? Either way, this
looks broken in either x86 or ppc unless doorbell_qpn itself is already
in the right endian.
But since later I see
writel(ring->doorbell_qpn, ring->bf.uar->map + MLX4_SEND_DOORBELL);
That rings a nasty bell, it looks like doorbell_pqn is in CPU (native)
endian, so it should have to be swapped before being OR'ed into the
descriptor, either that or the HW does some black magic I don't
understand.
Cheers,
Ben.
^ permalink raw reply
* [PATCH] gro: refetch inet6_protos[] after pulling ext headers
From: Yan, Zheng @ 2011-10-09 8:34 UTC (permalink / raw)
To: netdev@vger.kernel.org, Herbert Xu, davem@davemloft.net
ipv6_gro_receive() doesn't update the protocol ops after pulling
the ext headers. It looks like a typo.
Signed-off-by: Zheng Yan <zheng.z.yan@intel.com>
---
diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c
index 3b5669a..d27c797 100644
--- a/net/ipv6/af_inet6.c
+++ b/net/ipv6/af_inet6.c
@@ -875,6 +875,7 @@ static struct sk_buff **ipv6_gro_receive(struct sk_buff **head,
skb_reset_transport_header(skb);
__skb_push(skb, skb_gro_offset(skb));
+ ops = rcu_dereference(inet6_protos[proto]);
if (!ops || !ops->gro_receive)
goto out_unlock;
^ permalink raw reply related
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Eli Cohen @ 2011-10-09 8:07 UTC (permalink / raw)
To: Benjamin Herrenschmidt
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <1318147254.29415.377.camel@pasglop>
On Sun, Oct 09, 2011 at 10:00:54AM +0200, Benjamin Herrenschmidt wrote:
> On Sun, 2011-10-09 at 09:35 +0200, Eli Cohen wrote:
> > On Sun, Oct 09, 2011 at 09:25:18AM +0200, Benjamin Herrenschmidt wrote:
> > > On Thu, 2011-10-06 at 15:57 +0200, Eli Cohen wrote:
> > > > On Wed, Oct 05, 2011 at 10:15:02AM +0200, Eli Cohen wrote:
> > > >
> > > > How about this patch - can you give it a try?
> > > >
> > > >
> > > > >From dee60547aa9e35a02835451d9e694cd80dd3072f Mon Sep 17 00:00:00 2001
> > > > From: Eli Cohen <eli@mellanox.co.il>
> > > > Date: Thu, 6 Oct 2011 15:50:02 +0200
> > > > Subject: [PATCH] mlx4_en: Fix blue flame on powerpc
> > > >
> > > > The source buffer used for copying into the blue flame register is already in
> > > > big endian. However, when copying to device on powerpc, the endianess is
> > > > swapped so the data reaches th device in little endian which is wrong. On x86
> > > > based platform no swapping occurs so it reaches the device with the correct
> > > > endianess. Fix this by calling le32_to_cpu() on the buffer. On LE systems there
> > > > is no change; on BE there will be a swap.
> > >
> > > That looks wrong.
> > Not sure I understand: are you saying that on ppc, when you call
> > __iowrite64_copy, it will not reach the device swapped?
>
> Well, first, what do you mean by "swapped" ? :-) But no, it won't for
> all intend and purpose, this is a copy routine, copy routines never
> swap, neither do fifo accesses for example.
When I say swapped, I mean not necessairliy by software. I think that
the chipset will swap the the data. The reason I think so is that the
CPU arch is big endian, while PCI bus is defined as little endian.
That's why I think a swap will occur in ppc and not in x86.
>
> > The point is that we must always have the buffer ready in big endian
> > in memory. In the case of blue flame, we must also copy it to the
> > device registers in pci memory space. So if we use the buffer we
> > already prepared, we must have another swap. I can think of a nicer
> > way to implement this functionality but the question is do you think
> > my observation above is wrong and why.
>
> No. If it's in memory BE then the copy routine will keep it BE. A copy
> routine doesn't swap and doesn't affect endianness.
>
> Additionally, a swapping phase like you proposed doing 32-bit swaps
> means that you know for sure that the buffer is made of 32-bit
> quantities, is that the case ?
Yes
> Even if you had needed that swap, if your
> buffer had contained 16-bit or 64-bit quantities, you're toast.
>
> What is this buffer anyway ? A descriptor or a network packet ?
It's a special descriptor that resides both in memory and also written
to the device's register. An it contains both data and control
informartion.
>
> If it's a packet, then it's data, endianness has no meaning (or rather
> it has for individual fields of the packets but they are already in the
> right format and a 32-bit swap will never be right).
>
> It's almost never right to perform swapping when copying data (or
> reading/writing a FIFO).
>
> Cheers,
> Ben.
^ permalink raw reply
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Benjamin Herrenschmidt @ 2011-10-09 8:00 UTC (permalink / raw)
To: Eli Cohen
Cc: netdev@vger.kernel.org, Eli Cohen, linuxppc-dev,
Thadeu Lima de Souza Cascardo, Yevgeny Petrilin
In-Reply-To: <20111009073546.GI2681@mtldesk30>
On Sun, 2011-10-09 at 09:35 +0200, Eli Cohen wrote:
> On Sun, Oct 09, 2011 at 09:25:18AM +0200, Benjamin Herrenschmidt wrote:
> > On Thu, 2011-10-06 at 15:57 +0200, Eli Cohen wrote:
> > > On Wed, Oct 05, 2011 at 10:15:02AM +0200, Eli Cohen wrote:
> > >
> > > How about this patch - can you give it a try?
> > >
> > >
> > > >From dee60547aa9e35a02835451d9e694cd80dd3072f Mon Sep 17 00:00:00 2001
> > > From: Eli Cohen <eli@mellanox.co.il>
> > > Date: Thu, 6 Oct 2011 15:50:02 +0200
> > > Subject: [PATCH] mlx4_en: Fix blue flame on powerpc
> > >
> > > The source buffer used for copying into the blue flame register is already in
> > > big endian. However, when copying to device on powerpc, the endianess is
> > > swapped so the data reaches th device in little endian which is wrong. On x86
> > > based platform no swapping occurs so it reaches the device with the correct
> > > endianess. Fix this by calling le32_to_cpu() on the buffer. On LE systems there
> > > is no change; on BE there will be a swap.
> >
> > That looks wrong.
> Not sure I understand: are you saying that on ppc, when you call
> __iowrite64_copy, it will not reach the device swapped?
Well, first, what do you mean by "swapped" ? :-) But no, it won't for
all intend and purpose, this is a copy routine, copy routines never
swap, neither do fifo accesses for example.
> The point is that we must always have the buffer ready in big endian
> in memory. In the case of blue flame, we must also copy it to the
> device registers in pci memory space. So if we use the buffer we
> already prepared, we must have another swap. I can think of a nicer
> way to implement this functionality but the question is do you think
> my observation above is wrong and why.
No. If it's in memory BE then the copy routine will keep it BE. A copy
routine doesn't swap and doesn't affect endianness.
Additionally, a swapping phase like you proposed doing 32-bit swaps
means that you know for sure that the buffer is made of 32-bit
quantities, is that the case ? Even if you had needed that swap, if your
buffer had contained 16-bit or 64-bit quantities, you're toast.
What is this buffer anyway ? A descriptor or a network packet ?
If it's a packet, then it's data, endianness has no meaning (or rather
it has for individual fields of the packets but they are already in the
right format and a 32-bit swap will never be right).
It's almost never right to perform swapping when copying data (or
reading/writing a FIFO).
Cheers,
Ben.
^ permalink raw reply
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Eli Cohen @ 2011-10-09 7:35 UTC (permalink / raw)
To: Benjamin Herrenschmidt
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <1318145118.29415.371.camel@pasglop>
On Sun, Oct 09, 2011 at 09:25:18AM +0200, Benjamin Herrenschmidt wrote:
> On Thu, 2011-10-06 at 15:57 +0200, Eli Cohen wrote:
> > On Wed, Oct 05, 2011 at 10:15:02AM +0200, Eli Cohen wrote:
> >
> > How about this patch - can you give it a try?
> >
> >
> > >From dee60547aa9e35a02835451d9e694cd80dd3072f Mon Sep 17 00:00:00 2001
> > From: Eli Cohen <eli@mellanox.co.il>
> > Date: Thu, 6 Oct 2011 15:50:02 +0200
> > Subject: [PATCH] mlx4_en: Fix blue flame on powerpc
> >
> > The source buffer used for copying into the blue flame register is already in
> > big endian. However, when copying to device on powerpc, the endianess is
> > swapped so the data reaches th device in little endian which is wrong. On x86
> > based platform no swapping occurs so it reaches the device with the correct
> > endianess. Fix this by calling le32_to_cpu() on the buffer. On LE systems there
> > is no change; on BE there will be a swap.
>
> That looks wrong.
Not sure I understand: are you saying that on ppc, when you call
__iowrite64_copy, it will not reach the device swapped?
The point is that we must always have the buffer ready in big endian
in memory. In the case of blue flame, we must also copy it to the
device registers in pci memory space. So if we use the buffer we
already prepared, we must have another swap. I can think of a nicer
way to implement this functionality but the question is do you think
my observation above is wrong and why.
>
> What is this __iowrite64_copy... oh I see
>
> Nice, somebody _AGAIN_ added a bunch of "generic" IO accessors that are
> utterly wrong on all archs except x86 (ok, -almost-). There isn't a
> single bloody memory barrier in there !
>
> So, __iowrite64_copy is doing raw_writel which will -not- swap, so your
> buffer is going to have the same endianness in the destination than it
> has in the source. This is _NOT_ the right place to do a swap.
>
> It's the original construction of the descriptor that needs change. The
> data itself should never need to be affected accross a copy operation
> (unless your HW is terminally busted).
>
> Cheers,
> Ben.
>
> > Signed-off-by: Eli Cohen <eli@mellanox.co.il>
> > ---
> > drivers/net/mlx4/en_tx.c | 10 ++++++++++
> > 1 files changed, 10 insertions(+), 0 deletions(-)
> >
> > diff --git a/drivers/net/mlx4/en_tx.c b/drivers/net/mlx4/en_tx.c
> > index 16337fb..3743acc 100644
> > --- a/drivers/net/mlx4/en_tx.c
> > +++ b/drivers/net/mlx4/en_tx.c
> > @@ -601,6 +601,16 @@ u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb)
> >
> > static void mlx4_bf_copy(unsigned long *dst, unsigned long *src, unsigned bytecnt)
> > {
> > + int i;
> > + __le32 *psrc = (__le32 *)src;
> > +
> > + /*
> > + * the buffer is already in big endian. For little endian machines that's
> > + * fine. For big endain machines we must swap since the chipset swaps again
> > + */
> > + for (i = 0; i < bytecnt / 4; ++i)
> > + psrc[i] = le32_to_cpu(psrc[i]);
> > +
> > __iowrite64_copy(dst, src, bytecnt / 8);
> > }
> >
> > --
> > 1.7.7.rc0.70.g82660
> >
> >
> >
> > > On Tue, Oct 04, 2011 at 05:26:20PM -0300, Thadeu Lima de Souza Cascardo wrote:
> > >
> > > I believe we have an endianess problem here. The source buffer is in
> > > big endian - in x86 archs, it will rich the pci device unswapped since
> > > both x86 and pci are little endian. In ppc, it wil be swapped by the
> > > chipset so it will reach the device in little endian which is wrong.
> > > So, in mlx4_bf_copy, you could loop over the buffer and swap32 the all
> > > the dwords before the call to __iowrite64_copy. Of course which should
> > > fix this in an arch independent manner. Let me know this works for
> > > you.
> > >
> > > > On Tue, Oct 04, 2011 at 08:02:12AM +0200, Benjamin Herrenschmidt wrote:
> > > > > On Mon, 2011-10-03 at 17:53 -0300, Thadeu Lima de Souza Cascardo wrote:
> > > > >
> > > > > .../...
> > > > >
> > > > > > > Can you also send me the output of ethtool -i?
> > > > > > > It seems that there is a problem with write combining on Power processors, we will check this issue.
> > > > > > >
> > > > > > > Yevgeny
> > > > > >
> > > > > > Hello, Yevgeny.
> > > > > >
> > > > > > You will find the output of ethtool -i below.
> > > > > >
> > > > > > I am copying Ben and powerpc list, in case this is an issue with Power
> > > > > > processors. They can provide us some more insight into this.
> > > > >
> > > > > May I get some background please ? :-)
> > > > >
> > > > > I'm not aware of a specific issue with write combining but I'd need to
> > > > > know more about what you are doing and the code to do it to comment on
> > > > > whether it should work or not.
> > > > >
> > > > > Cheers,
> > > > > Ben.
> > > > >
> > > > >
> > > >
> > > > Hello, Ben.
> > > >
> > > > Sorry for that. I am testing mlx4_en driver on a POWER. Yevgeny has
> > > > added blue frame support, that does not require writing to the device
> > > > memory to indicate a new packet (the doorbell register as it is called).
> > > >
> > > > Well, the ring is getting full with no interrupt or packets transmitted.
> > > > I simply added a write to the doorbell register and it works for me.
> > > > Yevgeny says this is not the right fix, claiming there is a problem with
> > > > write combining on POWER. The code uses memory barriers, so I don't know
> > > > why there is any problem.
> > > >
> > > > I am posting the code here to show better what the situation is.
> > > > Yevgeny can tell more about the device and the driver.
> > > >
> > > > The code below is the driver as of now, including a diff with what I
> > > > changed and had resulted OK for me. Before the blue frame support, the
> > > > only code executed was the else part. I can't tell much what the device
> > > > should be seeing and doing after the blue frame part of the code is
> > > > executed. But it does send the packet if I write to the doorbell
> > > > register.
> > > >
> > > > Yevgeny, can you tell us what the device should be doing and why you
> > > > think this is a problem on POWER? Is it possible that this is simply a
> > > > problem with the firmware version?
> > > >
> > > > Thanks,
> > > > Cascardo.
> > > >
> > > > ---
> > > > if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
> > > > !vlan_tag) {
> > > > *(u32 *) (&tx_desc->ctrl.vlan_tag) |=
> > > > ring->doorbell_qpn;
> > > > op_own |= htonl((bf_index & 0xffff) << 8);
> > > > /* Ensure new descirptor hits memory
> > > > * before setting ownership of this descriptor to HW */
> > > > wmb();
> > > > tx_desc->ctrl.owner_opcode = op_own;
> > > >
> > > > wmb();
> > > >
> > > > mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned
> > > > long *) &tx_desc->ctrl,
> > > > desc_size);
> > > >
> > > > wmb();
> > > >
> > > > ring->bf.offset ^= ring->bf.buf_size;
> > > > } else {
> > > > /* Ensure new descirptor hits memory
> > > > * before setting ownership of this descriptor to HW */
> > > > wmb();
> > > > tx_desc->ctrl.owner_opcode = op_own;
> > > > - wmb();
> > > > - writel(ring->doorbell_qpn, ring->bf.uar->map +
> > > > MLX4_SEND_DOORBELL);
> > > > }
> > > >
> > > > + wmb();
> > > > + writel(ring->doorbell_qpn, ring->bf.uar->map +
> > > > MLX4_SEND_DOORBELL);
> > > > +
> > > > ---
>
^ permalink raw reply
* RE: [PATCH] mlx4_en: fix transmit of packages when blue frame isenabled
From: Benjamin Herrenschmidt @ 2011-10-09 7:28 UTC (permalink / raw)
To: David Laight
Cc: Eli Cohen, Thadeu Lima de Souza Cascardo, Eli Cohen,
Yevgeny Petrilin, linuxppc-dev, netdev
In-Reply-To: <AE90C24D6B3A694183C094C60CF0A2F6D8AE76@saturn3.aculab.com>
On Thu, 2011-10-06 at 15:10 +0100, David Laight wrote:
> horrid...
> 1) I'm not sure the caller expects the buffer to be corrupted.
> 2) It contains a lot of memory cycles.
> 3) It looked from the calls that this code is copying descriptors,
> so the transfer length is probably 1 or 2 words - so the loop
> is inefficient.
> 4) ppc doesn't have a fast byteswap instruction (very new gcc might
> use the byteswapping memery access for the le32_to_cpu() though),
> so it would be better getting the byteswap done inside
> __iowrite64_copy() - since that is probably requesting a byteswap
> anyway.
> OTOH I'm not at all clear about the 64bit xfers....
And it's just plain wrong anyway. You should never have to byteswap a
copy.
Ben.
^ permalink raw reply
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Benjamin Herrenschmidt @ 2011-10-09 7:25 UTC (permalink / raw)
To: Eli Cohen
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <20111006135759.GH2681@mtldesk30>
On Thu, 2011-10-06 at 15:57 +0200, Eli Cohen wrote:
> On Wed, Oct 05, 2011 at 10:15:02AM +0200, Eli Cohen wrote:
>
> How about this patch - can you give it a try?
>
>
> >From dee60547aa9e35a02835451d9e694cd80dd3072f Mon Sep 17 00:00:00 2001
> From: Eli Cohen <eli@mellanox.co.il>
> Date: Thu, 6 Oct 2011 15:50:02 +0200
> Subject: [PATCH] mlx4_en: Fix blue flame on powerpc
>
> The source buffer used for copying into the blue flame register is already in
> big endian. However, when copying to device on powerpc, the endianess is
> swapped so the data reaches th device in little endian which is wrong. On x86
> based platform no swapping occurs so it reaches the device with the correct
> endianess. Fix this by calling le32_to_cpu() on the buffer. On LE systems there
> is no change; on BE there will be a swap.
That looks wrong.
What is this __iowrite64_copy... oh I see
Nice, somebody _AGAIN_ added a bunch of "generic" IO accessors that are
utterly wrong on all archs except x86 (ok, -almost-). There isn't a
single bloody memory barrier in there !
So, __iowrite64_copy is doing raw_writel which will -not- swap, so your
buffer is going to have the same endianness in the destination than it
has in the source. This is _NOT_ the right place to do a swap.
It's the original construction of the descriptor that needs change. The
data itself should never need to be affected accross a copy operation
(unless your HW is terminally busted).
Cheers,
Ben.
> Signed-off-by: Eli Cohen <eli@mellanox.co.il>
> ---
> drivers/net/mlx4/en_tx.c | 10 ++++++++++
> 1 files changed, 10 insertions(+), 0 deletions(-)
>
> diff --git a/drivers/net/mlx4/en_tx.c b/drivers/net/mlx4/en_tx.c
> index 16337fb..3743acc 100644
> --- a/drivers/net/mlx4/en_tx.c
> +++ b/drivers/net/mlx4/en_tx.c
> @@ -601,6 +601,16 @@ u16 mlx4_en_select_queue(struct net_device *dev, struct sk_buff *skb)
>
> static void mlx4_bf_copy(unsigned long *dst, unsigned long *src, unsigned bytecnt)
> {
> + int i;
> + __le32 *psrc = (__le32 *)src;
> +
> + /*
> + * the buffer is already in big endian. For little endian machines that's
> + * fine. For big endain machines we must swap since the chipset swaps again
> + */
> + for (i = 0; i < bytecnt / 4; ++i)
> + psrc[i] = le32_to_cpu(psrc[i]);
> +
> __iowrite64_copy(dst, src, bytecnt / 8);
> }
>
> --
> 1.7.7.rc0.70.g82660
>
>
>
> > On Tue, Oct 04, 2011 at 05:26:20PM -0300, Thadeu Lima de Souza Cascardo wrote:
> >
> > I believe we have an endianess problem here. The source buffer is in
> > big endian - in x86 archs, it will rich the pci device unswapped since
> > both x86 and pci are little endian. In ppc, it wil be swapped by the
> > chipset so it will reach the device in little endian which is wrong.
> > So, in mlx4_bf_copy, you could loop over the buffer and swap32 the all
> > the dwords before the call to __iowrite64_copy. Of course which should
> > fix this in an arch independent manner. Let me know this works for
> > you.
> >
> > > On Tue, Oct 04, 2011 at 08:02:12AM +0200, Benjamin Herrenschmidt wrote:
> > > > On Mon, 2011-10-03 at 17:53 -0300, Thadeu Lima de Souza Cascardo wrote:
> > > >
> > > > .../...
> > > >
> > > > > > Can you also send me the output of ethtool -i?
> > > > > > It seems that there is a problem with write combining on Power processors, we will check this issue.
> > > > > >
> > > > > > Yevgeny
> > > > >
> > > > > Hello, Yevgeny.
> > > > >
> > > > > You will find the output of ethtool -i below.
> > > > >
> > > > > I am copying Ben and powerpc list, in case this is an issue with Power
> > > > > processors. They can provide us some more insight into this.
> > > >
> > > > May I get some background please ? :-)
> > > >
> > > > I'm not aware of a specific issue with write combining but I'd need to
> > > > know more about what you are doing and the code to do it to comment on
> > > > whether it should work or not.
> > > >
> > > > Cheers,
> > > > Ben.
> > > >
> > > >
> > >
> > > Hello, Ben.
> > >
> > > Sorry for that. I am testing mlx4_en driver on a POWER. Yevgeny has
> > > added blue frame support, that does not require writing to the device
> > > memory to indicate a new packet (the doorbell register as it is called).
> > >
> > > Well, the ring is getting full with no interrupt or packets transmitted.
> > > I simply added a write to the doorbell register and it works for me.
> > > Yevgeny says this is not the right fix, claiming there is a problem with
> > > write combining on POWER. The code uses memory barriers, so I don't know
> > > why there is any problem.
> > >
> > > I am posting the code here to show better what the situation is.
> > > Yevgeny can tell more about the device and the driver.
> > >
> > > The code below is the driver as of now, including a diff with what I
> > > changed and had resulted OK for me. Before the blue frame support, the
> > > only code executed was the else part. I can't tell much what the device
> > > should be seeing and doing after the blue frame part of the code is
> > > executed. But it does send the packet if I write to the doorbell
> > > register.
> > >
> > > Yevgeny, can you tell us what the device should be doing and why you
> > > think this is a problem on POWER? Is it possible that this is simply a
> > > problem with the firmware version?
> > >
> > > Thanks,
> > > Cascardo.
> > >
> > > ---
> > > if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
> > > !vlan_tag) {
> > > *(u32 *) (&tx_desc->ctrl.vlan_tag) |=
> > > ring->doorbell_qpn;
> > > op_own |= htonl((bf_index & 0xffff) << 8);
> > > /* Ensure new descirptor hits memory
> > > * before setting ownership of this descriptor to HW */
> > > wmb();
> > > tx_desc->ctrl.owner_opcode = op_own;
> > >
> > > wmb();
> > >
> > > mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned
> > > long *) &tx_desc->ctrl,
> > > desc_size);
> > >
> > > wmb();
> > >
> > > ring->bf.offset ^= ring->bf.buf_size;
> > > } else {
> > > /* Ensure new descirptor hits memory
> > > * before setting ownership of this descriptor to HW */
> > > wmb();
> > > tx_desc->ctrl.owner_opcode = op_own;
> > > - wmb();
> > > - writel(ring->doorbell_qpn, ring->bf.uar->map +
> > > MLX4_SEND_DOORBELL);
> > > }
> > >
> > > + wmb();
> > > + writel(ring->doorbell_qpn, ring->bf.uar->map +
> > > MLX4_SEND_DOORBELL);
> > > +
> > > ---
^ permalink raw reply
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Benjamin Herrenschmidt @ 2011-10-09 7:21 UTC (permalink / raw)
To: Eli Cohen
Cc: Thadeu Lima de Souza Cascardo, Yevgeny Petrilin,
netdev@vger.kernel.org, Eli Cohen, linuxppc-dev
In-Reply-To: <20111005081502.GB2681@mtldesk30>
On Wed, 2011-10-05 at 10:15 +0200, Eli Cohen wrote:
> On Tue, Oct 04, 2011 at 05:26:20PM -0300, Thadeu Lima de Souza Cascardo wrote:
>
> I believe we have an endianess problem here. The source buffer is in
> big endian - in x86 archs, it will rich the pci device unswapped since
> both x86 and pci are little endian. In ppc, it wil be swapped by the
> chipset
^^^^^^^ ugh ?
> so it will reach the device in little endian which is wrong.
> So, in mlx4_bf_copy, you could loop over the buffer and swap32 the all
> the dwords before the call to __iowrite64_copy. Of course which should
> fix this in an arch independent manner. Let me know this works for
> you.
That's generally the wrong way to handle endian (to swap32 a whole
buffer).
But I need to know more about the structure of those descriptors etc...
to judge properly.
Cheers,
Ben.
> > On Tue, Oct 04, 2011 at 08:02:12AM +0200, Benjamin Herrenschmidt wrote:
> > > On Mon, 2011-10-03 at 17:53 -0300, Thadeu Lima de Souza Cascardo wrote:
> > >
> > > .../...
> > >
> > > > > Can you also send me the output of ethtool -i?
> > > > > It seems that there is a problem with write combining on Power processors, we will check this issue.
> > > > >
> > > > > Yevgeny
> > > >
> > > > Hello, Yevgeny.
> > > >
> > > > You will find the output of ethtool -i below.
> > > >
> > > > I am copying Ben and powerpc list, in case this is an issue with Power
> > > > processors. They can provide us some more insight into this.
> > >
> > > May I get some background please ? :-)
> > >
> > > I'm not aware of a specific issue with write combining but I'd need to
> > > know more about what you are doing and the code to do it to comment on
> > > whether it should work or not.
> > >
> > > Cheers,
> > > Ben.
> > >
> > >
> >
> > Hello, Ben.
> >
> > Sorry for that. I am testing mlx4_en driver on a POWER. Yevgeny has
> > added blue frame support, that does not require writing to the device
> > memory to indicate a new packet (the doorbell register as it is called).
> >
> > Well, the ring is getting full with no interrupt or packets transmitted.
> > I simply added a write to the doorbell register and it works for me.
> > Yevgeny says this is not the right fix, claiming there is a problem with
> > write combining on POWER. The code uses memory barriers, so I don't know
> > why there is any problem.
> >
> > I am posting the code here to show better what the situation is.
> > Yevgeny can tell more about the device and the driver.
> >
> > The code below is the driver as of now, including a diff with what I
> > changed and had resulted OK for me. Before the blue frame support, the
> > only code executed was the else part. I can't tell much what the device
> > should be seeing and doing after the blue frame part of the code is
> > executed. But it does send the packet if I write to the doorbell
> > register.
> >
> > Yevgeny, can you tell us what the device should be doing and why you
> > think this is a problem on POWER? Is it possible that this is simply a
> > problem with the firmware version?
> >
> > Thanks,
> > Cascardo.
> >
> > ---
> > if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
> > !vlan_tag) {
> > *(u32 *) (&tx_desc->ctrl.vlan_tag) |=
> > ring->doorbell_qpn;
> > op_own |= htonl((bf_index & 0xffff) << 8);
> > /* Ensure new descirptor hits memory
> > * before setting ownership of this descriptor to HW */
> > wmb();
> > tx_desc->ctrl.owner_opcode = op_own;
> >
> > wmb();
> >
> > mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned
> > long *) &tx_desc->ctrl,
> > desc_size);
> >
> > wmb();
> >
> > ring->bf.offset ^= ring->bf.buf_size;
> > } else {
> > /* Ensure new descirptor hits memory
> > * before setting ownership of this descriptor to HW */
> > wmb();
> > tx_desc->ctrl.owner_opcode = op_own;
> > - wmb();
> > - writel(ring->doorbell_qpn, ring->bf.uar->map +
> > MLX4_SEND_DOORBELL);
> > }
> >
> > + wmb();
> > + writel(ring->doorbell_qpn, ring->bf.uar->map +
> > MLX4_SEND_DOORBELL);
> > +
> > ---
^ permalink raw reply
* Re: [PATCH] mlx4_en: fix transmit of packages when blue frame is enabled
From: Benjamin Herrenschmidt @ 2011-10-09 7:19 UTC (permalink / raw)
To: Thadeu Lima de Souza Cascardo
Cc: netdev@vger.kernel.org, Eli Cohen, eli@dev.mellanox.co.il,
linuxppc-dev, Yevgeny Petrilin
In-Reply-To: <20111004202620.GA3455@oc1711230544.ibm.com>
On Tue, 2011-10-04 at 17:26 -0300, Thadeu Lima de Souza Cascardo wrote:
> if (ring->bf_enabled && desc_size <= MAX_BF && !bounce &&
> !vlan_tag) {
> *(u32 *) (&tx_desc->ctrl.vlan_tag) |=
> ring->doorbell_qpn;
Could this have endianness problems ?
> op_own |= htonl((bf_index & 0xffff) << 8);
Probably better to use cpu_to_be/lexx() here. Also make sure your
descriptors are defined with the appropriate __beXX or __leXX types
so sparse can tell you when you get this wrong
> /* Ensure new descirptor hits memory
> * before setting ownership of this descriptor to HW */
> wmb();
> tx_desc->ctrl.owner_opcode = op_own;
> wmb();
>
> mlx4_bf_copy(ring->bf.reg + ring->bf.offset, (unsigned
> long *) &tx_desc->ctrl,
> desc_size);
What does the above do ?
> wmb();
>
> ring->bf.offset ^= ring->bf.buf_size;
Is this HW visible ? It's a bit odd, are you doing double bufferring and
trying to flip a bit ? If it's HW visible, is endian correct ?
} else {
> /* Ensure new descirptor hits memory
> * before setting ownership of this descriptor to HW */
> wmb();
> tx_desc->ctrl.owner_opcode = op_own;
> - wmb();
> - writel(ring->doorbell_qpn, ring->bf.uar->map +
> MLX4_SEND_DOORBELL);
> }
>
> + wmb();
> + writel(ring->doorbell_qpn, ring->bf.uar->map +
> MLX4_SEND_DOORBELL);
> +
> ---
Cheers,
Ben.
^ permalink raw reply
* raw ethenet socket question
From: Ronny Meeus @ 2011-10-08 20:20 UTC (permalink / raw)
To: netdev
Hello
I'm doing some tests on a raw Ethernet socket and I observe some
strange behavior.
My system is a FreeScale P4040, running Linux 2.6.36.4 compiled for
SMP mode but only 1 core activated.
[ 0.000000] Using P4080 DS machine description
[ 0.000000] Memory CAM mapping: 256/256/256 Mb, residual: 1248Mb
[ 0.000000] Linux version 2.6.36.4 (meeusr@devws108) (gcc version
4.4.6 (Buildroot 2011.08-hgc574811c3fc5) ) #1 SMP Sat Oct 8 20:40:40
CEST 2011
I basically have one of the Ethernet ports in a loopmode, which means
that all packets sent to the port will be received on the same port.
I want to check the amount of packets I can send/receive on my board,
so there a large loop to send/receive packets (see code below).
To stress the system a bit more, I send a number of packets before I
start to receive packets.
If I send 10 or less packets at a time, I do not see issues (see for
(j=0;j<10;j++)), but if I send for example 16 packets before starting
to receive, I always see packet loss.
For example if I send 1000 times 16 packets, I only receive 12000
packets, if I send 1000 times 20 packets, I receive 16000 packets.
I also checked the ifconfig statistics and the counters for Rx and Tx
are always identical and equal to the amount of packets that were
sent. So it looks like the packets are getting lost in the linux
kernel of in my test application.
Can somebody explain this behavior?
Is there a fixed amount of Ethernet packets that are queue in a socket?
Test code:
for (i=0;i<packets_to_send;i++)
{
fd_set readset;
struct timeval tv;
FD_ZERO(&readset);
FD_SET(eth_sock,&readset);
for (j=0;j<10;j++)
{
if (sendto(eth_sock,packet,packet_size, 0,(struct
sockaddr*)eth_dest, sizeof(struct sockaddr_ll)) > 0)
packet_count++;
}
tv.tv_sec = 0;
tv.tv_usec = 500;
while (select(eth_sock+1,&readset,0,0,&tv) == 1) {
int len = recvfrom(eth_sock,recvdata,sizeof(recvdata),0,NULL,NULL);
if (len>0) rcv_count++;
tv.tv_usec = 500;
}
printf("packet sent: %10ld packets received:
%10ld\r",packet_count,rcv_count);
}
This is how I create my socket:
int create_raw_eth_socket(void)
{
int eth_sock;
eth_dest.sll_family = PF_PACKET;
eth_dest.sll_protocol = htons(ETH_P_ALL);
eth_dest.sll_ifindex = itf_ifindex;
eth_dest.sll_hatype = ARPHRD_ETHER;
eth_dest.sll_pkttype = PACKET_OTHERHOST;
eth_dest.sll_halen = ETH_ALEN;
eth_sock = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
bind(eth_sock,(struct sockaddr*)ð_dest,sizeof(eth_dest));
return eth_sock;
}
Regards,
Ronny
^ permalink raw reply
* Re: [BUG] igb : rtnl assert on resume
From: Eric Dumazet @ 2011-10-08 20:16 UTC (permalink / raw)
To: jeffrey.t.kirsher; +Cc: David Miller, netdev
In-Reply-To: <1318102192.5276.27.camel@edumazet-laptop>
Le samedi 08 octobre 2011 à 21:29 +0200, Eric Dumazet a écrit :
> Le samedi 08 octobre 2011 à 11:26 -0700, Jeff Kirsher a écrit :
>
> > Any chance you can pull my net-next tree Eric to verify if the issue
> > still exists?
>
> Sure, I'll do that.
>
Same error with your changes as well.
No hurry, maybe the RTNL assert can be relaxed.
^ permalink raw reply
* Re: [net-next 02/11] igb: Use node specific allocations for the q_vectors and rings
From: David Miller @ 2011-10-08 19:51 UTC (permalink / raw)
To: jeffrey.t.kirsher; +Cc: alexander.h.duyck, netdev, gospo, sassmann
In-Reply-To: <1318056461-19562-3-git-send-email-jeffrey.t.kirsher@intel.com>
From: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Date: Fri, 7 Oct 2011 23:47:32 -0700
> From: Alexander Duyck <alexander.h.duyck@intel.com>
>
> This change is meant to update the ring and vector allocations so that they
> are per node instead of allocating everything on the node that
> ifconfig/modprobe is called on. By doing this we can cut down
> significantly on cross node traffic.
>
> Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
> Tested-by: Aaron Brown <aaron.f.brown@intel.com>
> Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
adapter->node seems superfluous.
It's always "-1" when we enter the allocation functions, and we
always restore it to it's original value upon exit from such
functions.
Just get rid of it and use a local variable in these functions
to keep track of the current allocation node.
Also, what ensures that MSI-X interrupts are targetted to a cpu
on the the node where you've made these allocations? I was
pretty sure Ben Hutchings added infrastructure that's usable
to ensure this, but I can't see where you're using it.
^ permalink raw reply
* Re: [BUG] igb : rtnl assert on resume
From: Eric Dumazet @ 2011-10-08 19:29 UTC (permalink / raw)
To: jeffrey.t.kirsher; +Cc: David Miller, netdev
In-Reply-To: <1318098396.23688.169.camel@jtkirshe-mobl>
Le samedi 08 octobre 2011 à 11:26 -0700, Jeff Kirsher a écrit :
> Any chance you can pull my net-next tree Eric to verify if the issue
> still exists?
Sure, I'll do that.
In the meantime, I upgraded my (slow, old) machine to ubuntu 11.11, and
reproduced the problem on their 3.0.something kernel.
>
> git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/net-next.git
>
> I currently have 11 igb patches on that tree (which I recently
> submitted) that are not in Dave's tree yet.
>
> There have been a number of cleanup's recently, so while this is not
> expected, I am not surprised at the same time since I am in the middle
> of the update to igb.
>
> We have not seen this on my local tree (which has the 11 patches plus
> 4-5 more patches for igb). Almost all of them are
> cleanup's/re-organizing of igb, so nothing specific to the issue you
> have seen.
Thanks
^ permalink raw reply
* Re: e100 + VLANs?
From: Jeff Kirsher @ 2011-10-08 18:34 UTC (permalink / raw)
To: Eric Dumazet; +Cc: Michael Tokarev, netdev
In-Reply-To: <1318091046.5276.22.camel@edumazet-laptop>
[-- Attachment #1: Type: text/plain, Size: 1968 bytes --]
On 10/08/2011 09:24 AM, Eric Dumazet wrote:
> Le samedi 08 octobre 2011 à 14:08 +0400, Michael Tokarev a écrit :
>> > Yesterday I tried to use 802.1Q VLAN tagging with an (oldish)
>> > e100-driven network card, identified by lspci like this:
>> >
>> > 00:12.0 Ethernet controller: Intel Corporation 82557/8/9/0/1 Ethernet Pro 100 (rev 02)
>> >
>> > Just to discover that it does not quite work: packets of
>> > size 1497+ bytes gets lost.
>> >
>> > This appears to be a classical problems in this case -
>> > something forgot to allocate extra 4 bytes for the
>> > packets.
>> >
>> > There's at least one bugreport from 2008 (!) about this
>> > very issue: http://bugs.centos.org/view.php?id=2719
>> > which is still open.
>> >
>> > The kernel I tried this on was 2.6.32, I checked git log
>> > for drivers/net/e100.c - there was no changes up to
>> > current version which may be related to this issue.
>> >
>> > The question: is this a driver problem or hardware? If
>> > it's the driver, can it be fixed? And if it's hardware,
>> > can the driver notify the user somehow - like, by refusing
>> > to enable VLAN (sub)devices maybe?
>> >
>> > Yesterday it was actually a bit more complicated for me,
>> > since the card in question was used to connect to our
>> > ISP, and they use fixed MAC address per port, so I had
>> > to find another NIC which is a) able to work with VLAN
>> > tags properly, and b) is able to change its mac address.
>> > Lucky I had a VIA RhineIII which does both :)
>> >
> Since you have two cards (and probably two machines), maybe you could
> try to track if the problem is a bad transmit or a bad receive ?
>
> tcpdump on both machines, and ping -s 2000 from both sides...
>
> e100 driver seems VLAN enabled at a first glance.
Eric is correct, that e100 does support VLANs.
In addition to Eric's suggestion, can you also provide all the output of
lspci -vvv for the network card?
[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 900 bytes --]
^ permalink raw reply
* Re: [BUG] igb : rtnl assert on resume
From: Jeff Kirsher @ 2011-10-08 18:26 UTC (permalink / raw)
To: Eric Dumazet; +Cc: David Miller, netdev
In-Reply-To: <1318096284.5276.25.camel@edumazet-laptop>
[-- Attachment #1: Type: text/plain, Size: 2309 bytes --]
On Sat, 2011-10-08 at 10:51 -0700, Eric Dumazet wrote:
> Hi guys
>
> Not sure if its a regression (it seems linux-3.0 probably should trigger
> same trace) but I caught this on latest net-next when suspend/resume
> cycle was done :
>
> [ 5279.891512] RTNL: assertion failed at net/core/dev.c (1665)
> [ 5279.891515] Pid: 2785, comm: kworker/u:9 Not tainted 3.1.0-rc9+ #81
> [ 5279.891517] Call Trace:
> [ 5279.891523] [<ffffffff814802ec>] netif_set_real_num_tx_queues+0x1ac/0x1c0
> [ 5279.891540] [<ffffffffa00244d1>] igb_init_interrupt_scheme+0x151/0x4b0 [igb]
> [ 5279.891546] [<ffffffffa002984d>] igb_resume+0x9d/0x160 [igb]
> [ 5279.891550] [<ffffffff812e0b91>] pci_legacy_resume+0x41/0x60
> [ 5279.891552] [<ffffffff812e17c0>] pci_pm_resume+0x90/0xd0
> [ 5279.891555] [<ffffffff813896b7>] pm_op+0xe7/0x1c0
> [ 5279.891558] [<ffffffff81389ba7>] device_resume+0xf7/0x1b0
> [ 5279.891561] [<ffffffff8108a520>] ? async_schedule+0x20/0x20
> [ 5279.891564] [<ffffffff81389c81>] async_resume+0x21/0x50
> [ 5279.891566] [<ffffffff8108a59f>] async_run_entry_fn+0x7f/0x180
> [ 5279.891569] [<ffffffff8107ca73>] process_one_work+0x123/0x480
> [ 5279.891572] [<ffffffff8107d638>] ? manage_workers.isra.30+0x1e8/0x230
> [ 5279.891574] [<ffffffff8107d7de>] worker_thread+0x15e/0x360
> [ 5279.891576] [<ffffffff8107d680>] ? manage_workers.isra.30+0x230/0x230
> [ 5279.891579] [<ffffffff8108225c>] kthread+0x8c/0xa0
> [ 5279.891583] [<ffffffff8157fc74>] kernel_thread_helper+0x4/0x10
> [ 5279.891586] [<ffffffff810821d0>] ? flush_kthread_worker+0xa0/0xa0
> [ 5279.891588] [<ffffffff8157fc70>] ? gs_change+0x13/0x13
>
> Thanks
Any chance you can pull my net-next tree Eric to verify if the issue
still exists?
git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/net-next.git
I currently have 11 igb patches on that tree (which I recently
submitted) that are not in Dave's tree yet.
There have been a number of cleanup's recently, so while this is not
expected, I am not surprised at the same time since I am in the middle
of the update to igb.
We have not seen this on my local tree (which has the 11 patches plus
4-5 more patches for igb). Almost all of them are
cleanup's/re-organizing of igb, so nothing specific to the issue you
have seen.
[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 836 bytes --]
^ permalink raw reply
* [BUG] igb : rtnl assert on resume
From: Eric Dumazet @ 2011-10-08 17:51 UTC (permalink / raw)
To: Kirsher, Jeffrey T, David Miller; +Cc: netdev
Hi guys
Not sure if its a regression (it seems linux-3.0 probably should trigger
same trace) but I caught this on latest net-next when suspend/resume
cycle was done :
[ 5279.891512] RTNL: assertion failed at net/core/dev.c (1665)
[ 5279.891515] Pid: 2785, comm: kworker/u:9 Not tainted 3.1.0-rc9+ #81
[ 5279.891517] Call Trace:
[ 5279.891523] [<ffffffff814802ec>] netif_set_real_num_tx_queues+0x1ac/0x1c0
[ 5279.891540] [<ffffffffa00244d1>] igb_init_interrupt_scheme+0x151/0x4b0 [igb]
[ 5279.891546] [<ffffffffa002984d>] igb_resume+0x9d/0x160 [igb]
[ 5279.891550] [<ffffffff812e0b91>] pci_legacy_resume+0x41/0x60
[ 5279.891552] [<ffffffff812e17c0>] pci_pm_resume+0x90/0xd0
[ 5279.891555] [<ffffffff813896b7>] pm_op+0xe7/0x1c0
[ 5279.891558] [<ffffffff81389ba7>] device_resume+0xf7/0x1b0
[ 5279.891561] [<ffffffff8108a520>] ? async_schedule+0x20/0x20
[ 5279.891564] [<ffffffff81389c81>] async_resume+0x21/0x50
[ 5279.891566] [<ffffffff8108a59f>] async_run_entry_fn+0x7f/0x180
[ 5279.891569] [<ffffffff8107ca73>] process_one_work+0x123/0x480
[ 5279.891572] [<ffffffff8107d638>] ? manage_workers.isra.30+0x1e8/0x230
[ 5279.891574] [<ffffffff8107d7de>] worker_thread+0x15e/0x360
[ 5279.891576] [<ffffffff8107d680>] ? manage_workers.isra.30+0x230/0x230
[ 5279.891579] [<ffffffff8108225c>] kthread+0x8c/0xa0
[ 5279.891583] [<ffffffff8157fc74>] kernel_thread_helper+0x4/0x10
[ 5279.891586] [<ffffffff810821d0>] ? flush_kthread_worker+0xa0/0xa0
[ 5279.891588] [<ffffffff8157fc70>] ? gs_change+0x13/0x13
Thanks
^ permalink raw reply
* Re: e100 + VLANs?
From: Eric Dumazet @ 2011-10-08 16:24 UTC (permalink / raw)
To: Michael Tokarev; +Cc: netdev
In-Reply-To: <4E90212D.8030009@msgid.tls.msk.ru>
Le samedi 08 octobre 2011 à 14:08 +0400, Michael Tokarev a écrit :
> Yesterday I tried to use 802.1Q VLAN tagging with an (oldish)
> e100-driven network card, identified by lspci like this:
>
> 00:12.0 Ethernet controller: Intel Corporation 82557/8/9/0/1 Ethernet Pro 100 (rev 02)
>
> Just to discover that it does not quite work: packets of
> size 1497+ bytes gets lost.
>
> This appears to be a classical problems in this case -
> something forgot to allocate extra 4 bytes for the
> packets.
>
> There's at least one bugreport from 2008 (!) about this
> very issue: http://bugs.centos.org/view.php?id=2719
> which is still open.
>
> The kernel I tried this on was 2.6.32, I checked git log
> for drivers/net/e100.c - there was no changes up to
> current version which may be related to this issue.
>
> The question: is this a driver problem or hardware? If
> it's the driver, can it be fixed? And if it's hardware,
> can the driver notify the user somehow - like, by refusing
> to enable VLAN (sub)devices maybe?
>
> Yesterday it was actually a bit more complicated for me,
> since the card in question was used to connect to our
> ISP, and they use fixed MAC address per port, so I had
> to find another NIC which is a) able to work with VLAN
> tags properly, and b) is able to change its mac address.
> Lucky I had a VIA RhineIII which does both :)
>
Since you have two cards (and probably two machines), maybe you could
try to track if the problem is a bad transmit or a bad receive ?
tcpdump on both machines, and ping -s 2000 from both sides...
e100 driver seems VLAN enabled at a first glance.
^ permalink raw reply
* 群发软件+买家搜索机+最新广交会买家、海关数据,B2B询盘买家500万。
From: 仅10元每天 @ 2011-10-08 16:09 UTC (permalink / raw)
To: nessen, net, netamieya, netdev, nethashem, netmarkc, netrangr,
nette
群发软件+买家搜索机+109届广交会买家、展会买家、海关数据,B2B询盘买家500万。
一共8个包(数据是全行业的，按照行业分好类，并且可以按照关键词查询的)：
1，2011春季109届广交会买家数据库新鲜出炉，超级新鲜买家，新鲜数据，容易成单！
2，最新全球买家库,共451660条数据。
3，2008年,2009年,2010年 春季+秋季广交会买家名录，103 104 105 106 107 108 共六届 共120.6万数据。
4，2010年国际促销协会（PPAI）成员名单 PPAI Members Directory，非常重要的大买家。
5，2010年到香港采购的国外客人名录(香港贸发局提供)，共7.2万数据，超级重要的买家。
6，60.8万条最新国外B2B买家询盘。
7，2009年海关提单数据piers版数据 1千万。
8，群发软件，群发软件的部署与安装。
共 500万个买家，每个均有Email.
保证每天都有买家回复。
保证每天都有买家回复。
要的抓紧联系QQ: 1339625218 或者立即回复邮箱: 1339625218@qq.com
要的抓紧联系QQ: 1339625218 或者立即回复邮箱: 1339625218@qq.com
要的抓紧联系QQ: 1339625218 或者立即回复邮箱: 1339625218@qq.com
诚信为本，如果不信任本人,可以走淘宝交易,收货验证后再付款,这是对您最好的保障了。
保证每天都有买家回复。
保证每天都有买家回复。
保证每天都有买家回复。
广交会买家按产品类别分类，分为以下几类：
1 办公设备
2 编织及藤铁工艺品
3 玻璃
4 餐厨用具
5 车辆
6 大型机械及设备
7 电子电气
8 电子消费品
9 纺织
10 服装
11 个人护理
12 工程机械
13 工具
14 化工
15 计算机及通讯
16 家居用品
17 家居装饰
18 家具
19 家用电器
20 建筑及装饰材料
21 节日用品
22 礼品及赠品
23 摩托车
24 汽车配件
25 食品
26 陶瓷
27 铁石
28 玩具
29 卫浴
30 五金
31 小型机械
32 鞋
33 休闲用品
34 医疗
35 浴室产品
36 园林
37 照明产品
38 钟表眼镜
39 自行车
40 包
保证每天都有买家回复。
保证每天都有买家回复。
保证每天都有买家回复。
保证每天都有买家回复。
保证每天都有买家回复。
^ permalink raw reply
* Re: SOMAXCONN = 128, but max defaults to 2048
From: Olaf van der Spek @ 2011-10-08 12:44 UTC (permalink / raw)
To: Eric Dumazet; +Cc: netdev
In-Reply-To: <1318077339.5276.17.camel@edumazet-laptop>
On Sat, Oct 8, 2011 at 2:35 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> Le samedi 08 octobre 2011 à 13:51 +0200, Olaf van der Spek a écrit :
>> Is the name of SOMAXCONN wrong and is it actually defining the default?
>> Would there be a disadvantage to defining SOMAXCONN as INT_MAX and
>> letting the kernel control the actual max?
>>
>
> You mean : remove somaxconn tunable ?
No, I mean basically removing the compile-time value and only using
the run-time value (from the kernel).
> accept()/listen() is not bound to TCP only.
What else is that value used for?
> We had a recent discussion on the matter lately, but Hagen Paul Pfeifer
> did not polish his patches enough :
>
> http://lists.openwall.net/netdev/2011/03/20/3
Isn't it a single-line change?
Olaf
^ permalink raw reply
* Re: SOMAXCONN = 128, but max defaults to 2048
From: Eric Dumazet @ 2011-10-08 12:35 UTC (permalink / raw)
To: Olaf van der Spek; +Cc: netdev
In-Reply-To: <CAGVGHmt1OVOj0NSEo3qHCh1CR10eZTNjxuUh_mB6sXxcx5gKmA@mail.gmail.com>
Le samedi 08 octobre 2011 à 13:51 +0200, Olaf van der Spek a écrit :
> Hi,
>
> SOMAXCONN is defined as 128, but the max
> ("/proc/sys/net/ipv4/tcp_max_syn_backlog") appears to be 2048 by
> default.
It depends on memory size and TCP hash table size :
dmesg | grep "TCP established hash table entries"
-> TCP established hash table entries: 524288 (order: 11, 8388608 bytes)
sysctl_max_syn_backlog = max(128, cnt / 256);
cnt = 524288
cnt/256 -> 2048
> Is the name of SOMAXCONN wrong and is it actually defining the default?
> Would there be a disadvantage to defining SOMAXCONN as INT_MAX and
> letting the kernel control the actual max?
>
You mean : remove somaxconn tunable ?
accept()/listen() is not bound to TCP only.
But yes, 128 default is a bit old today, given that Apache uses a
listen(fd, 511) default value itself...
I routinely set net.core.somaxconn to 1024 on my servers.
We had a recent discussion on the matter lately, but Hagen Paul Pfeifer
did not polish his patches enough :
http://lists.openwall.net/netdev/2011/03/20/3
FreeBSD has a separate mechanism for not yet established sockets, called
syncache, that allows a low somaxconn per listener.
The socket queue holds only fully established (but not yet accept()ed)
sockets, while syncache holds all the SYN_RECV ones.
^ permalink raw reply
page: next (older) | prev (newer) | latest
- recent:[subjects (threaded)|topics (new)|topics (active)]
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox