On 8/1/23 12:51, Han Zhou wrote:
> On Tue, Jul 18, 2023 at 6:16 AM Dumitru Ceara <[email protected]> wrote:
>>
>> On 7/18/23 12:14, Han Zhou wrote:
>>> On Mon, Jul 17, 2023 at 9:51 PM Dumitru Ceara <[email protected]> wrote:
>>>>
>>>> On 7/14/23 08:41, Ales Musil wrote:
>>>>> On Thu, Jul 13, 2023 at 4:39 PM Dumitru Ceara <[email protected]>
> wrote:
>>>>>
>>>>>> Regular dp-hash is not a canonical L4 hash (at least with the netlink
>>>>>> datapath).  If the datapath supports l4 symmetrical dp-hash use that
>>> one
>>>>>> instead.
>>>>>>
>>>>>> Reported-at: https://github.com/ovn-org/ovn/issues/112
>>>>>> Reported-at: https://bugzilla.redhat.com/show_bug.cgi?id=2188679
>>>>>> Signed-off-by: Dumitru Ceara <[email protected]>
>>>>>>
>>>>>
>>>>> Hi Dumitru,
>>>>>
>>>>>
>>>>>> ---
>>>>>>  include/ovn/features.h |  2 ++
>>>>>>  lib/actions.c          |  6 ++++++
>>>>>>  lib/features.c         | 49
> +++++++++++++++++++++++++++++++++---------
>>>>>>  3 files changed, 47 insertions(+), 10 deletions(-)
>>>>>>
>>>>>> diff --git a/include/ovn/features.h b/include/ovn/features.h
>>>>>> index de8f1f5489..3bf536127f 100644
>>>>>> --- a/include/ovn/features.h
>>>>>> +++ b/include/ovn/features.h
>>>>>> @@ -34,12 +34,14 @@ enum ovs_feature_support_bits {
>>>>>>      OVS_CT_ZERO_SNAT_SUPPORT_BIT,
>>>>>>      OVS_DP_METER_SUPPORT_BIT,
>>>>>>      OVS_CT_TUPLE_FLUSH_BIT,
>>>>>> +    OVS_DP_HASH_L4_SYM_BIT,
>>>>>>  };
>>>>>>
>>>>>>  enum ovs_feature_value {
>>>>>>      OVS_CT_ZERO_SNAT_SUPPORT = (1 << OVS_CT_ZERO_SNAT_SUPPORT_BIT),
>>>>>>      OVS_DP_METER_SUPPORT = (1 << OVS_DP_METER_SUPPORT_BIT),
>>>>>>      OVS_CT_TUPLE_FLUSH_SUPPORT = (1 << OVS_CT_TUPLE_FLUSH_BIT),
>>>>>> +    OVS_DP_HASH_L4_SYM_SUPPORT = (1 << OVS_DP_HASH_L4_SYM_BIT),
>>>>>>  };
>>>>>>
>>>>>>  void ovs_feature_support_destroy(void);
>>>>>> diff --git a/lib/actions.c b/lib/actions.c
>>>>>> index 037172e606..9d52cb75a8 100644
>>>>>> --- a/lib/actions.c
>>>>>> +++ b/lib/actions.c
>>>>>> @@ -1625,6 +1625,12 @@ encode_SELECT(const struct ovnact_select
>>> *select,
>>>>>>      struct ds ds = DS_EMPTY_INITIALIZER;
>>>>>>      ds_put_format(&ds, "type=select,selection_method=dp_hash");
>>>>>>
>>>>>> +    if (ovs_feature_is_supported(OVS_DP_HASH_L4_SYM_SUPPORT)) {
>>>>>> +        /* Select dp-hash l4_symmetric by setting the upper 32bits
> of
>>>>>> +         * selection_method_param to 1: */
>>>>>>
>>>>>
>>>>> This comment is a bit unfortunate, because it reads like you want to
> set
>>>>> all bits of the upper half
>>>>> to 1 e.g. 0xffffffff. Maybe change it to: "selection_method_param to
>>> value
>>>>> 1 (1 << 32)." during merge, wdyt?
>>>>>
>>>>>
>>>>
>>>> Makes sense, thanks for the suggestion!  I changed it accordingly.
>>>>
>>>>>> +        ds_put_cstr(&ds, ",selection_method_param=0x100000000");
>>>>>> +    }
>>>>>> +
>>>>>>      struct mf_subfield sf = expr_resolve_field(&select->res_field);
>>>>>>
>>>>>>      for (size_t bucket_id = 0; bucket_id < select->n_dsts;
>>> bucket_id++) {
>>>>>> diff --git a/lib/features.c b/lib/features.c
>>>>>> index 97c9c86f00..d24e8f6c5c 100644
>>>>>> --- a/lib/features.c
>>>>>> +++ b/lib/features.c
>>>>>> @@ -21,6 +21,7 @@
>>>>>>  #include "lib/dirs.h"
>>>>>>  #include "socket-util.h"
>>>>>>  #include "lib/vswitch-idl.h"
>>>>>> +#include "odp-netlink.h"
>>>>>>  #include "openvswitch/vlog.h"
>>>>>>  #include "openvswitch/ofpbuf.h"
>>>>>>  #include "openvswitch/rconn.h"
>>>>>> @@ -33,20 +34,48 @@ VLOG_DEFINE_THIS_MODULE(features);
>>>>>>
>>>>>>  #define FEATURES_DEFAULT_PROBE_INTERVAL_SEC 5
>>>>>>
>>>>>> +/* Parses 'cap_name' from 'ovs_capabilities' and returns whether the
>>>>>> + * type of capability is supported or not. */
>>>>>> +typedef bool ovs_feature_parse_func(const struct smap
>>> *ovs_capabilities,
>>>>>> +                                    const char *cap_name);
>>>>>> +
>>>>>>  struct ovs_feature {
>>>>>>      enum ovs_feature_value value;
>>>>>>      const char *name;
>>>>>> +    ovs_feature_parse_func *parse;
>>>>>>  };
>>>>>>
>>>>>> +static bool
>>>>>> +bool_parser(const struct smap *ovs_capabilities, const char
> *cap_name)
>>>>>> +{
>>>>>> +    return smap_get_bool(ovs_capabilities, cap_name, false);
>>>>>> +}
>>>>>> +
>>>>>> +static bool
>>>>>> +dp_hash_l4_sym_support_parser(const struct smap *ovs_capabilities,
>>>>>> +                              const char *cap_name OVS_UNUSED)
>>>>>> +{
>>>>>> +    int max_hash_alg = smap_get_int(ovs_capabilities,
> "max_hash_alg",
>>> 0);
>>>>>> +
>>>>>> +    return max_hash_alg == OVS_HASH_ALG_SYM_L4;
>>>>>> +}
>>>>>> +
>>>>>>  static struct ovs_feature all_ovs_features[] = {
>>>>>>      {
>>>>>>          .value = OVS_CT_ZERO_SNAT_SUPPORT,
>>>>>> -        .name = "ct_zero_snat"
>>>>>> +        .name = "ct_zero_snat",
>>>>>> +        .parse = bool_parser,
>>>>>>      },
>>>>>>      {
>>>>>>          .value = OVS_CT_TUPLE_FLUSH_SUPPORT,
>>>>>> -        .name = "ct_flush"
>>>>>> -    }
>>>>>> +        .name = "ct_flush",
>>>>>> +        .parse = bool_parser,
>>>>>> +    },
>>>>>> +    {
>>>>>> +        .value = OVS_DP_HASH_L4_SYM_SUPPORT,
>>>>>> +        .name = "dp_hash_l4_sym_support",
>>>>>> +        .parse = dp_hash_l4_sym_support_parser,
>>>>>> +    },
>>>>>>  };
>>>>>>
>>>>>>  /* A bitmap of OVS features that have been detected as 'supported'.
> */
>>>>>> @@ -65,6 +94,7 @@ ovs_feature_is_valid(enum ovs_feature_value
> feature)
>>>>>>      case OVS_CT_ZERO_SNAT_SUPPORT:
>>>>>>      case OVS_DP_METER_SUPPORT:
>>>>>>      case OVS_CT_TUPLE_FLUSH_SUPPORT:
>>>>>> +    case OVS_DP_HASH_L4_SYM_SUPPORT:
>>>>>>          return true;
>>>>>>      default:
>>>>>>          return false;
>>>>>> @@ -183,18 +213,17 @@ ovs_feature_support_run(const struct smap
>>>>>> *ovs_capabilities,
>>>>>>      }
>>>>>>
>>>>>>      for (size_t i = 0; i < ARRAY_SIZE(all_ovs_features); i++) {
>>>>>> -        enum ovs_feature_value value = all_ovs_features[i].value;
>>>>>> -        const char *name = all_ovs_features[i].name;
>>>>>> -        bool old_state = supported_ovs_features & value;
>>>>>> -        bool new_state = smap_get_bool(ovs_capabilities, name,
> false);
>>>>>> +        struct ovs_feature *feature = &all_ovs_features[i];
>>>>>> +        bool old_state = supported_ovs_features & feature->value;
>>>>>> +        bool new_state = feature->parse(ovs_capabilities,
>>> feature->name);
>>>>>>          if (new_state != old_state) {
>>>>>>              updated = true;
>>>>>>              if (new_state) {
>>>>>> -                supported_ovs_features |= value;
>>>>>> +                supported_ovs_features |= feature->value;
>>>>>>              } else {
>>>>>> -                supported_ovs_features &= ~value;
>>>>>> +                supported_ovs_features &= ~feature->value;
>>>>>>              }
>>>>>> -            VLOG_INFO_RL(&rl, "OVS Feature: %s, state: %s", name,
>>>>>> +            VLOG_INFO_RL(&rl, "OVS Feature: %s, state: %s",
>>> feature->name,
>>>>>>                           new_state ? "supported" : "not supported");
>>>>>>          }
>>>>>>      }
>>>>>> --
>>>>>> 2.31.1
>>>>>>
>>>>>> _______________________________________________
>>>>>> dev mailing list
>>>>>> [email protected]
>>>>>> https://mail.openvswitch.org/mailman/listinfo/ovs-dev
>>>>>>
>>>>>>
>>>>> Other than that it looks good.
>>>>>
>>>>> Acked-by: Ales Musil <[email protected]>
>>>>>
>>>>
>>>> Thanks, I applied this to main.
>>>>
>>>> I was wondering if we should backport this though.  It's not exactly a
>>>> bug fix but it does avoid using a broken datapath feature (when
>>>> possible) so it avoids a broken behavior.
>>>>
>>>> The changes are quite contained.
>>>>
>>>> CC-ing some of the maintainers explicitly to get some more eyes on
> this.
>>>>
>>> Thanks Dumitru. Sorry that I was following the reported issue. I tried
> to
>>> look for some more details, but I am not authorized to access bug
> 2175716.
>>
>> Sorry about that, that bug used to be public but some internal policies
>> made it private.
>>
>>> It is not quite clear to me what the impact of this problem is. The OVN
>>> issue (112) seems to be a very common configuration of OVN ECMP and I
>>> assume the problem is not easily reproduced because I have been
>>> testing/deploying the feature and it worked as expected. Do you have
> more
>>> details about the trigger? Are there any related fixes in OVS and
> kernel?
>>>
>>
>> The original bug report came from OpenShift:
>> https://issues.redhat.com/browse/OCPBUGS-7406
>>
>> CC-ing Patryk as he root caused the original issue.
>>
>> The simplified topology is:
>>                               +--- GW1 ---+
>> client ---- cluster-router ---+           +--- (out of OVN) --- server
>>                               +--- GW2 ---+
>>
>> On cluster-router there's an ECMP route to reach 'server' via:
>> - GW1
>> - GW2
>>
>> On GW1 traffic going out of OVN is SNATed to IP1.
>> On GW2 traffic going out of OVN is SNATed to IP2.
>>
>> The client initiates a TCP session towards the server; a path is chosen
>> based on dp-hash, e.g. GW2.
>>
>> From this point on, during normal operation, traffic on this session
>> flows fine via GW2.
>>
>> If the client closes the connection, then for the closing connection
>> there's an exception when computing the dp-hash if the socket is in
>> state TIME_WAIT, substate TCP_FIN_WAIT_2:
>> https://elixir.bootlin.com/linux/v6.2.1/source/net/ipv4/tcp_ipv4.c#L2218
>> In this last case we call:
>>
>> tcp_v4_timewait_ack(sk, skb)
>> |
>> tcp_v4_send_ack(sk, skb, ...)
>>
>> The latter uses a global, per-cpu "ctl_sk" to build the last ACK packet:
>> https://elixir.bootlin.com/linux/v6.2.1/source/net/ipv4/tcp_ipv4.c#L925
>>
>>         ctl_sk = this_cpu_read(ipv4_tcp_sk);
>>         sock_net_set(ctl_sk, net);
>>         ctl_sk->sk_mark = (sk->sk_state == TCP_TIME_WAIT) ?
>>                            inet_twsk(sk)->tw_mark : sk->sk_mark;
>>         ctl_sk->sk_priority = (sk->sk_state == TCP_TIME_WAIT) ?
>>                            inet_twsk(sk)->tw_priority : sk->sk_priority;
>>         transmit_time = tcp_transmit_time(sk);
>>         ip_send_unicast_reply(ctl_sk,
>>                               skb, &TCP_SKB_CB(skb)->header.h4.opt,
>>                               ip_hdr(skb)->saddr, ip_hdr(skb)->daddr,
>>                               &arg, arg.iov[0].iov_len,
>>                               transmit_time);
>>
>> As ctl_sk's sk_txhash is 0, when transmitting this last ack, its hash is
>> computed from the packet contents:
>>
> https://elixir.bootlin.com/linux/v6.2.1/source/include/linux/skbuff.h#L1540
>>
>> static inline __u32 skb_get_hash(struct sk_buff *skb)
>> {
>>         if (!skb->l4_hash && !skb->sw_hash)
>>                 __skb_get_hash(skb);
>>
>>         return skb->hash;
>> }
>> [...]
>> void __skb_get_hash(struct sk_buff *skb)
>> {
>>         struct flow_keys keys;
>>         u32 hash;
>>
>>         __flow_hash_secret_init();
>>
>>         hash = ___skb_get_hash(skb, &keys, &hashrnd);
>>
>>         __skb_set_sw_hash(skb, hash, flow_keys_have_l4(&keys));
>> }
>>
>> But the original socket's txhash was actually a random hash:
>> https://elixir.bootlin.com/linux/v6.2.1/source/include/net/sock.h#L2128
>>
>> static inline void sk_set_txhash(struct sock *sk)
>> {
>>         /* This pairs with READ_ONCE() in skb_set_hash_from_sk() */
>>         WRITE_ONCE(sk->sk_txhash, net_tx_rndhash());
>> }
>>
>> This means that the last ack might enter the ovs datapath with a
>> different dp_hash than all other packets sent on that session in that
>> direction. It's possible that this last ack gets hashed to a different
>> path of the ECMP route.
>>
>> This specific issue was fixed in the kernel through:
>>
>>
> https://lore.kernel.org/all/[email protected]/T/#mc84997c4cc755354e5527b41b7a15869536a21f4
>>
>> However, while debugging we realized the problem is actually visible
>> with _any_ retransmit.  That's because a retransmit will trigger a rehash:
>>
>> https://elixir.bootlin.com/linux/latest/source/net/ipv4/tcp_output.c#L4124
>>
>> int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
>> {
>>         const struct tcp_request_sock_ops *af_ops =
> tcp_rsk(req)->af_specific;
>>         struct flowi fl;
>>         int res;
>>
>>         /* Paired with WRITE_ONCE() in sock_setsockopt() */
>>         if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
>>                 tcp_rsk(req)->txhash = net_tx_rndhash();
>>         res = af_ops->send_synack(sk, NULL, &fl, req, NULL,
> TCP_SYNACK_NORMAL,
>>                                   NULL);
>> [...]
>>
>> So suddenly all subsequent packets (skbs) will have a different 'hash'
>> value and might get hashed to a different path.
>>
>> This last part was not fixed.  After a discussion on the netdev mailing
>> list the conclusion was to not change the hash behavior for locally
>> terminated TCP sessions:
>>
>>
> https://lore.kernel.org/all/[email protected]/T/#m1fd8da6d912d5acead66c4c397396e6ac4b160dd
>>
>> Eventually the documentation change patch was dropped and a v2 of the
>> other fixes was accepted:
>> https://www.spinics.net/lists/netdev/msg906252.html
>>
>> The rehash on retransmit problem is still present though making the
>> kernel's dp-hash implementation unusable.
>>
>> During some internal discussions Aaron (in CC) pointed out that OVS
>> supports L4 symmetric hashing and that should be a stable L4 hash.  OVS
>> already reports the maximum supported hashing algorithm through the
>> Datapath.Capabilities.max_hash_alg ovsdb field.  The only problem was
>> that the kernel datapath didn't support L4 symmetric hashing.  But Aaron
>> posted a patch to implement that and the patch got accepted:
>>
>> https://www.spinics.net/lists/netdev/msg911045.html
>>
>> Then we moved OVN to use L4 symmetric hashing if available.  That's to
>> ensure correctness of the packet processing.  We did run performance
>> tests and the hit due to the extra hashing was acceptable, max 3%
>> throughput degradation in OVN scenarios that simulate common traffic
>> patterns (like ovn-kubernetes ones) - in most cases the performance
>> degradation was around 2%.
>>
>> Sorry for the long story but I hope it provides all the required
>> context.  Thinking more about it I should've probably added some of this
>> to the original commit log, I apologize.
>>
>> Regards,
>> Dumitru
>>
> 
> Thanks Dumitru for the details! It is now quite clear and I am not against
> backporting it.
> But I have a question regarding upgrading. For the existing traffic, does
> this mean the packets may choose a different path after OVN upgrading?
> Should this be called out?
> 

Hi Han,

That's a good question.  They might and this affects all traffic that's
not locally originated for the kernel datapath and (potentially) all
traffic for the userspace datapath.  I can post a follow up patch to
update the NEWS file.  Did you have anything else in mind?

Regards,
Dumitru

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to