[PATCH net-next] bonding: correct rr balancing during link failure

[PATCH net-next] bonding: correct rr balancing during link failure

[Lars Everbrand]

This patch updates the sending algorithm for roundrobin to avoid
over-subscribing interface(s) when one or more interfaces in the bond is
not able to send packets. This happened when order was not random and
more than 2 interfaces were used.

Previously the algorithm would find the next available interface
when an interface failed to send by, this means that most often it is
current_interface + 1. The problem is that when the next packet is to be
sent and the "normal" algorithm then continues with interface++ which
then hits that same interface again.

This patch updates the resending algorithm to update the global counter
of the next interface to use.

Example (prior to patch):

Consider 6 x 100 Mbit/s interfaces in a rr bond. The normal order of links
being used to send would look like:
1 2 3 4 5 6  1 2 3 4 5 6  1 2 3 4 5 6 ...

If, for instance, interface 2 where unable to send the order would have been:
1 3 3 4 5 6  1 3 3 4 5 6  1 3 3 4 5 6 ...

The resulting speed (for TCP) would then become:
50 + 0 + 100 + 50 + 50 + 50 = 300 Mbit/s
instead of the expected 500 Mbit/s.

If interface 3 also would fail the resulting speed would be half of the
expected 400 Mbit/s (33 + 0 + 0 + 100 + 33 + 33).

Signed-off-by: Lars Everbrand <lars.everbrand@protonmail.com>
---
 drivers/net/bonding/bond_main.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index e0880a3840d7..e02d9c6d40ee 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -4107,6 +4107,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
 		if (--i < 0) {
 			if (bond_slave_can_tx(slave))
 				return slave;
+			bond->rr_tx_counter++;
 		}
 	}
 
@@ -4117,6 +4118,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
 			break;
 		if (bond_slave_can_tx(slave))
 			return slave;
+		bond->rr_tx_counter++;
 	}
 	/* no slave that can tx has been found */
 	return NULL;
-- 
2.29.2

Re: [PATCH net-next] bonding: correct rr balancing during link failure

[Jakub Kicinski]

On Wed, 02 Dec 2020 20:55:57 +0000 Lars Everbrand wrote:
> This patch updates the sending algorithm for roundrobin to avoid
> over-subscribing interface(s) when one or more interfaces in the bond is
> not able to send packets. This happened when order was not random and
> more than 2 interfaces were used.
> 
> Previously the algorithm would find the next available interface
> when an interface failed to send by, this means that most often it is
> current_interface + 1. The problem is that when the next packet is to be
> sent and the "normal" algorithm then continues with interface++ which
> then hits that same interface again.
> 
> This patch updates the resending algorithm to update the global counter
> of the next interface to use.
> 
> Example (prior to patch):
> 
> Consider 6 x 100 Mbit/s interfaces in a rr bond. The normal order of links
> being used to send would look like:
> 1 2 3 4 5 6  1 2 3 4 5 6  1 2 3 4 5 6 ...
> 
> If, for instance, interface 2 where unable to send the order would have been:
> 1 3 3 4 5 6  1 3 3 4 5 6  1 3 3 4 5 6 ...
> 
> The resulting speed (for TCP) would then become:
> 50 + 0 + 100 + 50 + 50 + 50 = 300 Mbit/s
> instead of the expected 500 Mbit/s.
> 
> If interface 3 also would fail the resulting speed would be half of the
> expected 400 Mbit/s (33 + 0 + 0 + 100 + 33 + 33).
> 
> Signed-off-by: Lars Everbrand <lars.everbrand@protonmail.com>

Thanks for the patch!

Looking at the code in question it feels a little like we're breaking
abstractions if we bump the counter directly in get_slave_by_id.

For one thing when the function is called for IGMP packets the counter
should not be incremented at all. But also if packets_per_slave is not
1 we'd still be hitting the same leg multiple times (packets_per_slave
/ 2). So it seems like we should round the counter up somehow?

For IGMP maybe we don't have to call bond_get_slave_by_id() at all,
IMHO, just find first leg that can TX. Then we can restructure
bond_get_slave_by_id() appropriately for the non-IGMP case.

> diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
> index e0880a3840d7..e02d9c6d40ee 100644
> --- a/drivers/net/bonding/bond_main.c
> +++ b/drivers/net/bonding/bond_main.c
> @@ -4107,6 +4107,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
>  		if (--i < 0) {
>  			if (bond_slave_can_tx(slave))
>  				return slave;
> +			bond->rr_tx_counter++;
>  		}
>  	}
>  
> @@ -4117,6 +4118,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
>  			break;
>  		if (bond_slave_can_tx(slave))
>  			return slave;
> +		bond->rr_tx_counter++;
>  	}
>  	/* no slave that can tx has been found */
>  	return NULL;

Re: [PATCH net-next] bonding: correct rr balancing during link failure

[Jay Vosburgh]

Jakub Kicinski <kuba@kernel.org> wrote:

>On Wed, 02 Dec 2020 20:55:57 +0000 Lars Everbrand wrote:
>> This patch updates the sending algorithm for roundrobin to avoid
>> over-subscribing interface(s) when one or more interfaces in the bond is
>> not able to send packets. This happened when order was not random and
>> more than 2 interfaces were used.
>> 
>> Previously the algorithm would find the next available interface
>> when an interface failed to send by, this means that most often it is
>> current_interface + 1. The problem is that when the next packet is to be
>> sent and the "normal" algorithm then continues with interface++ which
>> then hits that same interface again.
>> 
>> This patch updates the resending algorithm to update the global counter
>> of the next interface to use.
>> 
>> Example (prior to patch):
>> 
>> Consider 6 x 100 Mbit/s interfaces in a rr bond. The normal order of links
>> being used to send would look like:
>> 1 2 3 4 5 6  1 2 3 4 5 6  1 2 3 4 5 6 ...
>> 
>> If, for instance, interface 2 where unable to send the order would have been:
>> 1 3 3 4 5 6  1 3 3 4 5 6  1 3 3 4 5 6 ...
>> 
>> The resulting speed (for TCP) would then become:
>> 50 + 0 + 100 + 50 + 50 + 50 = 300 Mbit/s
>> instead of the expected 500 Mbit/s.
>> 
>> If interface 3 also would fail the resulting speed would be half of the
>> expected 400 Mbit/s (33 + 0 + 0 + 100 + 33 + 33).

	Are these bandwidth numbers from observation of the actual
behavior?  I'm not sure the real system would behave this way; my
suspicion is that it would increase the likelihood of drops on the
overused slave, not that the overall capacity would be limited.

>> Signed-off-by: Lars Everbrand <lars.everbrand@protonmail.com>
>
>Thanks for the patch!
>
>Looking at the code in question it feels a little like we're breaking
>abstractions if we bump the counter directly in get_slave_by_id.

	Agreed; I think a better way to fix this is to enable the slave
array for balance-rr mode, and then use the array to find the right
slave.  This way, we then avoid the problematic "skip unable to tx"
logic for free.

>For one thing when the function is called for IGMP packets the counter
>should not be incremented at all. But also if packets_per_slave is not
>1 we'd still be hitting the same leg multiple times (packets_per_slave
>/ 2). So it seems like we should round the counter up somehow?
>
>For IGMP maybe we don't have to call bond_get_slave_by_id() at all,
>IMHO, just find first leg that can TX. Then we can restructure
>bond_get_slave_by_id() appropriately for the non-IGMP case.

	For IGMP, the theory is to confine that traffic to a single
device.  Normally, this will be curr_active_slave, which is updated even
in balance-rr mode as interfaces are added to or removed from the bond.
The call to bond_get_slave_by_id should be a fallback in case
curr_active_slave is empty, and should be the exception, and may not be
possible at all.

	But either way, the IGMP path shouldn't mess with rr_tx_counter,
it should be out of band of the normal TX packet counting, so to speak.

	-J

>> diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
>> index e0880a3840d7..e02d9c6d40ee 100644
>> --- a/drivers/net/bonding/bond_main.c
>> +++ b/drivers/net/bonding/bond_main.c
>> @@ -4107,6 +4107,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
>>  		if (--i < 0) {
>>  			if (bond_slave_can_tx(slave))
>>  				return slave;
>> +			bond->rr_tx_counter++;
>>  		}
>>  	}
>>  
>> @@ -4117,6 +4118,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
>>  			break;
>>  		if (bond_slave_can_tx(slave))
>>  			return slave;
>> +		bond->rr_tx_counter++;
>>  	}
>>  	/* no slave that can tx has been found */
>>  	return NULL;
>

---
	-Jay Vosburgh, jay.vosburgh@canonical.com

Re: [PATCH net-next] bonding: correct rr balancing during link failure

[Lars Everbrand]

On Tue, Dec 08, 2020 at 01:46:09PM -0800, Jay Vosburgh wrote:
> 
> Jakub Kicinski <kuba@kernel.org> wrote:
> 
> >On Wed, 02 Dec 2020 20:55:57 +0000 Lars Everbrand wrote:
> 	Are these bandwidth numbers from observation of the actual
> behavior?  I'm not sure the real system would behave this way; my
> suspicion is that it would increase the likelihood of drops on the
> overused slave, not that the overall capacity would be limited.
I tested this with with 2 VMs and 5 bridges with bandwidth limitiation
via 'virsh domiftune' to bring the speed down to something similar to 
100Mbit/s.

iperf results:

with patch:
---
working       iperf
interfaces    speed [mbit/s]
5             442
4             363
3             278
2             199
1             107

without patch:
---
working       iperf
interfaces    speed [mbit/s]
5             444
4             226
3             155
2             129
1             107

The speed at 5x100 is not going as high as I expected but the
sub-optimal speed is still visible.

Note that the degradation tested is with downing interfaces sequentially
which is the worst-case for this problem.

> >Looking at the code in question it feels a little like we're breaking
> >abstractions if we bump the counter directly in get_slave_by_id.
> 
> 	Agreed; I think a better way to fix this is to enable the slave
> array for balance-rr mode, and then use the array to find the right
> slave.  This way, we then avoid the problematic "skip unable to tx"
> logic for free.
> 
> >For one thing when the function is called for IGMP packets the counter
> >should not be incremented at all. But also if packets_per_slave is not
> >1 we'd still be hitting the same leg multiple times (packets_per_slave
> >/ 2). So it seems like we should round the counter up somehow?
> >
> >For IGMP maybe we don't have to call bond_get_slave_by_id() at all,
> >IMHO, just find first leg that can TX. Then we can restructure
> >bond_get_slave_by_id() appropriately for the non-IGMP case.
> 
> 	For IGMP, the theory is to confine that traffic to a single
> device.  Normally, this will be curr_active_slave, which is updated even
> in balance-rr mode as interfaces are added to or removed from the bond.
> The call to bond_get_slave_by_id should be a fallback in case
> curr_active_slave is empty, and should be the exception, and may not be
> possible at all.
> 
> 	But either way, the IGMP path shouldn't mess with rr_tx_counter,
> it should be out of band of the normal TX packet counting, so to speak.
> 
> 	-J
> 
> >> diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
> >> index e0880a3840d7..e02d9c6d40ee 100644
> >> --- a/drivers/net/bonding/bond_main.c
> >> +++ b/drivers/net/bonding/bond_main.c
> >> @@ -4107,6 +4107,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
> >>  		if (--i < 0) {
> >>  			if (bond_slave_can_tx(slave))
> >>  				return slave;
> >> +			bond->rr_tx_counter++;
> >>  		}
> >>  	}
> >>
> >> @@ -4117,6 +4118,7 @@ static struct slave *bond_get_slave_by_id(struct bonding *bond,
> >>  			break;
> >>  		if (bond_slave_can_tx(slave))
> >>  			return slave;
> >> +		bond->rr_tx_counter++;
> >>  	}
> >>  	/* no slave that can tx has been found */
> >>  	return NULL;
> >
> 
> ---
> 	-Jay Vosburgh, jay.vosburgh@canonical.com

Re: [PATCH net-next] bonding: correct rr balancing during link failure

[Lars Everbrand]

On Sat, Dec 05, 2020 at 11:45:13AM -0800, Jakub Kicinski wrote:
> Thanks for the patch!
Kind words for my first attempt at this. Sorry for answering a bit late,
proton-bridge is not my best friend lately.
> 
> Looking at the code in question it feels a little like we're breaking
> abstractions if we bump the counter directly in get_slave_by_id.
My intention was to avoid a big change, and this was the easiest way. I
trust your opinion here.
> 
> For one thing when the function is called for IGMP packets the counter
> should not be incremented at all. But also if packets_per_slave is not
> 1 we'd still be hitting the same leg multiple times (packets_per_slave
> / 2). So it seems like we should round the counter up somehow?
I did not consider this case, I only test =1 and random. Yeah, it breaks
if the counter is updated per packet in any >1 case. 
> 
> For IGMP maybe we don't have to call bond_get_slave_by_id() at all,
> IMHO, just find first leg that can TX. Then we can restructure
> bond_get_slave_by_id() appropriately for the non-IGMP case.
I can have another look but my I am not confident that I am skilled
enough in this area to produce a larger overhaul...