Is there any reason why bonding should have an operstate of up when none
of its slaves are in an up state? In this particular scenario it seems
like the bonding device should at least assert NO-CARRIER, thoughts?
$ ip -o -d link show | grep "bond5"
2: enp0s31f6: <NO-CARRIER,BROADCAST,MULTICAST,SLAVE,UP> mtu 1500 qdisc
fq_codel master bond5 state DOWN mode DEFAULT group default qlen 1000\
link/ether 8c:8c:aa:f8:62:16 brd ff:ff:ff:ff:ff:ff promiscuity 0
minmtu 68 maxmtu 9000 \ bond_slave state ACTIVE mii_status UP
link_failure_count 0 perm_hwaddr 8c:8c:aa:f8:62:16 queue_id 0
numtxqueues 1 numrxqueues 1 gso_max_size 65536 gso_max_segs 65535
41: bond5: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc
noqueue state UP mode DEFAULT group default qlen 1000\ link/ether
8c:8c:aa:f8:62:16 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu
65535 \ bond mode balance-xor miimon 0 updelay 0 downdelay 0
peer_notify_delay 0 use_carrier 1 arp_interval 0 arp_validate none
arp_all_targets any primary_reselect always fail_over_mac none
xmit_hash_policy layer2 resend_igmp 1 num_grat_arp 1 all_slaves_active 0
min_links 0 lp_interval 1 packets_per_slave 1 lacp_rate slow ad_select
stable tlb_dynamic_lb 1 numtxqueues 16 numrxqueues 16 gso_max_size 65536
gso_max_segs 65535
$ cat /sys/class/net/enp0s31f6/operstate
down
$ cat /sys/class/net/bond5/operstate
up
This is an older kernel (4.18.0-305.7.1.el8_4.x86_64) but I do not see
any changes upstream that would indicate a change in this operation.
Thanks,
-Jon
Jonathan Toppins <[email protected]> wrote:
>Is there any reason why bonding should have an operstate of up when none
>of its slaves are in an up state? In this particular scenario it seems
>like the bonding device should at least assert NO-CARRIER, thoughts?
>
>$ ip -o -d link show | grep "bond5"
>2: enp0s31f6: <NO-CARRIER,BROADCAST,MULTICAST,SLAVE,UP> mtu 1500 qdisc
>fq_codel master bond5 state DOWN mode DEFAULT group default qlen 1000\
>link/ether 8c:8c:aa:f8:62:16 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68
>maxmtu 9000 \ bond_slave state ACTIVE mii_status UP link_failure_count
>0 perm_hwaddr 8c:8c:aa:f8:62:16 queue_id 0 numtxqueues 1 numrxqueues 1
>gso_max_size 65536 gso_max_segs 65535
>41: bond5: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc noqueue
>state UP mode DEFAULT group default qlen 1000\ link/ether
>8c:8c:aa:f8:62:16 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu
>65535 \ bond mode balance-xor miimon 0 updelay 0 downdelay 0
>peer_notify_delay 0 use_carrier 1 arp_interval 0 arp_validate none
I'm going to speculate that your problem is that miimon and
arp_interval are both 0, and the bond then doesn't have any active
mechanism to monitor the link state of its interfaces. There might be a
warning in dmesg to this effect.
Do you see what you'd consider to be correct behavior if miimon
is set to 100?
-J
>arp_all_targets any primary_reselect always fail_over_mac none
>xmit_hash_policy layer2 resend_igmp 1 num_grat_arp 1 all_slaves_active 0
>min_links 0 lp_interval 1 packets_per_slave 1 lacp_rate slow ad_select
>stable tlb_dynamic_lb 1 numtxqueues 16 numrxqueues 16 gso_max_size 65536
>gso_max_segs 65535
>
>$ cat /sys/class/net/enp0s31f6/operstate
>down
>
>$ cat /sys/class/net/bond5/operstate
>up
>
>This is an older kernel (4.18.0-305.7.1.el8_4.x86_64) but I do not see any
>changes upstream that would indicate a change in this operation.
>
>Thanks,
>-Jon
---
-Jay Vosburgh, [email protected]
On 8/7/21 6:42 PM, Jay Vosburgh wrote:
> Jonathan Toppins <[email protected]> wrote:
>
>> Is there any reason why bonding should have an operstate of up when none
>> of its slaves are in an up state? In this particular scenario it seems
>> like the bonding device should at least assert NO-CARRIER, thoughts?
>>
>> $ ip -o -d link show | grep "bond5"
>> 2: enp0s31f6: <NO-CARRIER,BROADCAST,MULTICAST,SLAVE,UP> mtu 1500 qdisc
>> fq_codel master bond5 state DOWN mode DEFAULT group default qlen 1000\
>> link/ether 8c:8c:aa:f8:62:16 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68
>> maxmtu 9000 \ bond_slave state ACTIVE mii_status UP link_failure_count
>> 0 perm_hwaddr 8c:8c:aa:f8:62:16 queue_id 0 numtxqueues 1 numrxqueues 1
>> gso_max_size 65536 gso_max_segs 65535
>> 41: bond5: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc noqueue
>> state UP mode DEFAULT group default qlen 1000\ link/ether
>> 8c:8c:aa:f8:62:16 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu
>> 65535 \ bond mode balance-xor miimon 0 updelay 0 downdelay 0
>> peer_notify_delay 0 use_carrier 1 arp_interval 0 arp_validate none
>
> I'm going to speculate that your problem is that miimon and
> arp_interval are both 0, and the bond then doesn't have any active
> mechanism to monitor the link state of its interfaces. There might be a
> warning in dmesg to this effect.
>
> Do you see what you'd consider to be correct behavior if miimon
> is set to 100?
>
setting miimon = 100 does appear to fix it.
It is interesting that there is no link monitor on by default. For
example when I enslave enp0s31f6 to a new bond with miimon == 0,
enp0s31f6 starts admin down and will never de-assert NO-CARRIER the bond
always results in an operstate of up. It seems like miimon = 100 should
be the default since some modes cannot use arpmon.
Thank you for the discussion, see below for the steps taken.
$ sudo ip link set dev enp0s31f6 nomaster
$ sudo ip link add dev bond6 type bond mode balance-xor
$ sudo ip -o -d link set dev bond6 up
$ ip -o -d link show dev bond6
62: bond6: <NO-CARRIER,BROADCAST,MULTICAST,MASTER,UP> mtu 1500 qdisc
noqueue state DOWN mode DEFAULT group default qlen 1000\ link/ether
3e:12:01:8a:ed:b1 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu
65535 \ bond mode balance-xor miimon 0 updelay 0 downdelay 0
peer_notify_delay 0 use_carrier 1 arp_interval 0 arp_validate none
arp_all_targets any primary_reselect always fail_over_mac none
xmit_hash_policy layer2 resend_igmp 1 num_grat_arp 1 all_slaves_active 0
min_links 0 lp_interval 1 packets_per_slave 1 lacp_rate slow ad_select
stable tlb_dynamic_lb 1 numtxqueues 16 numrxqueues 16 gso_max_size 65536
gso_max_segs 65535
$ ip -o -d link show dev enp0s31f6
2: enp0s31f6: <BROADCAST,MULTICAST> mtu 1500 qdisc fq_codel state DOWN
mode DEFAULT group default qlen 1000\ link/ether 8c:8c:aa:f8:62:16
brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu 9000 numtxqueues 1
numrxqueues 1 gso_max_size 65536 gso_max_segs 65535
$ sudo ip -o -d link set dev enp0s31f6 master bond6
$ ip -o -d link show dev bond6
62: bond6: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc
noqueue state UP mode DEFAULT group default qlen 1000\ link/ether
8c:8c:aa:f8:62:16 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu
65535 \ bond mode balance-xor miimon 0 updelay 0 downdelay 0
peer_notify_delay 0 use_carrier 1 arp_interval 0 arp_validate none
arp_all_targets any primary_reselect always fail_over_mac none
xmit_hash_policy layer2 resend_igmp 1 num_grat_arp 1 all_slaves_active 0
min_links 0 lp_interval 1 packets_per_slave 1 lacp_rate slow ad_select
stable tlb_dynamic_lb 1 numtxqueues 16 numrxqueues 16 gso_max_size 65536
gso_max_segs 65535
$ sudo ip link set dev enp0s31f6 nomaster
$ ip -o -d link show dev bond6
62: bond6: <NO-CARRIER,BROADCAST,MULTICAST,MASTER,UP> mtu 1500 qdisc
noqueue state DOWN mode DEFAULT group default qlen 1000\ link/ether
ae:b8:6e:b3:ca:3f brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu
65535 \ bond mode balance-xor miimon 0 updelay 0 downdelay 0
peer_notify_delay 0 use_carrier 1 arp_interval 0 arp_validate none
arp_all_targets any primary_reselect always fail_over_mac none
xmit_hash_policy layer2 resend_igmp 1 num_grat_arp 1 all_slaves_active 0
min_links 0 lp_interval 1 packets_per_slave 1 lacp_rate slow ad_select
stable tlb_dynamic_lb 1 numtxqueues 16 numrxqueues 16 gso_max_size 65536
gso_max_segs 65535
On Sat, Aug 07, 2021 at 08:09:31PM -0400, Jonathan Toppins wrote:
> setting miimon = 100 does appear to fix it.
>
> It is interesting that there is no link monitor on by default. For example
> when I enslave enp0s31f6 to a new bond with miimon == 0, enp0s31f6 starts
> admin down and will never de-assert NO-CARRIER the bond always results in an
> operstate of up. It seems like miimon = 100 should be the default since some
> modes cannot use arpmon.
Historically when miimon was implemented, not all NICs nor drivers had
support for link state checking at all! In addition, there are certain
deployments where you could rely on many devices by having a bond device
on top of a vlan or similar device, and where monitoring could cost a
lot of resources and you'd prefer to rely on external monitoring to set
all of them up or down at once.
I do think however that there remains a case with a missing state
transition in the driver: on my laptop I have a bond interface attached
to eth0, and I noticed that if I suspend the laptop with the link up,
when I wake it up with no interface connected, the bond will not turn
down, regardless of miimon. I have not looked closer yet, but I
suspect that we're relying too much on a state change between previous
and current and that one historically impossible transition does not
exist there and/or used to work because it was handled as part of
another change. I'll eventually have a look.
Willy
On 8/8/21 12:49 AM, Willy Tarreau wrote:
> On Sat, Aug 07, 2021 at 08:09:31PM -0400, Jonathan Toppins wrote:
>> setting miimon = 100 does appear to fix it.
>>
>> It is interesting that there is no link monitor on by default. For example
>> when I enslave enp0s31f6 to a new bond with miimon == 0, enp0s31f6 starts
>> admin down and will never de-assert NO-CARRIER the bond always results in an
>> operstate of up. It seems like miimon = 100 should be the default since some
>> modes cannot use arpmon.
>
> Historically when miimon was implemented, not all NICs nor drivers had
> support for link state checking at all! In addition, there are certain
> deployments where you could rely on many devices by having a bond device
> on top of a vlan or similar device, and where monitoring could cost a
> lot of resources and you'd prefer to rely on external monitoring to set
> all of them up or down at once.
>
> I do think however that there remains a case with a missing state
> transition in the driver: on my laptop I have a bond interface attached
> to eth0, and I noticed that if I suspend the laptop with the link up,
> when I wake it up with no interface connected, the bond will not turn
> down, regardless of miimon. I have not looked closer yet, but I
> suspect that we're relying too much on a state change between previous
> and current and that one historically impossible transition does not
> exist there and/or used to work because it was handled as part of
> another change. I'll eventually have a look.
>
> Willy
>
I am likely very wrong but the lack of a recalculation of the bond
carrier state after a lower notifies of an up/down event seemed
incorrect. Maybe a place to start?
diff --git i/drivers/net/bonding/bond_main.c
w/drivers/net/bonding/bond_main.c
index 9018fcc59f78..2b2c4b937142 100644
--- i/drivers/net/bonding/bond_main.c
+++ w/drivers/net/bonding/bond_main.c
@@ -3308,6 +3308,7 @@ static int bond_slave_netdev_event(unsigned long
event,
*/
if (bond_mode_can_use_xmit_hash(bond))
bond_update_slave_arr(bond, NULL);
+ bond_set_carrier(bond);
break;
case NETDEV_CHANGEMTU:
/* TODO: Should slaves be allowed to
Hi Jonathan,
On Sun, Aug 08, 2021 at 09:31:39PM -0400, Jonathan Toppins wrote:
> I am likely very wrong but the lack of a recalculation of the bond carrier
> state after a lower notifies of an up/down event seemed incorrect. Maybe a
> place to start?
Thanks for the test, it could have been a good candidate but it does
not work :-)
That's what I have after the following sequence:
- link is up
- suspend-to-ram
- unplug the cable
- resume
$ ip -br li
eth0 DOWN e8:6a:64:5d:19:ed <NO-CARRIER,BROADCAST,MULTICAST,SLAVE,UP>
eth0.2@eth0 UP e8:6a:64:5d:19:ed <BROADCAST,MULTICAST,SLAVE,UP,LOWER_UP>
bond0 UP e8:6a:64:5d:19:ed <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP>
My bond interface uses eth0 and eth0.2 in active-backup scenario allowing
me to instantly switch between tagged/untagged network depending on the
port I'm connecting to.
I just figured the problem. It's not the bonding driver which is causing
this issue, the issue is with the VLAN interface which incorrectly shows
up while it ought not to, as can be seen above, and the bond naturally
selected it:
Primary Slave: eth0 (primary_reselect always)
Currently Active Slave: eth0.2
MII Status: up
MII Polling Interval (ms): 200
Up Delay (ms): 0
Down Delay (ms): 0
Peer Notification Delay (ms): 0
So the bond driver works well, I'll have to dig into the 802.1q code
and/or see how the no-carrier state is propagated upstream. So you were
not very wrong at all and put me on the right track :-)
Cheers,
Willy