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

From: Lars Everbrand
Date: Tue Dec 15 2020 - 16:58:17 EST


On Tue, Dec 08, 2020 at 01:46:09PM -0800, Jay Vosburgh wrote:
>
> Jakub Kicinski <kuba@xxxxxxxxxx> 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@xxxxxxxxxxxxx