On 1/12/26 5:27 PM, Eelco Chaudron wrote:
> 
> 
> On 6 Jan 2026, at 12:55, Ilya Maximets wrote:
> 
>> There are multiple logical and coding issues in this functionality.
>> They are semi-related, so it's easier to fix them all at once instead
>> of trying to untangle the changes into separate commits.
>>
>> The issues are:
>>
>> - The implementation of a non-masked variant of the action is putting
>> the base flow values into the action instead of the updated key, so
>> the packet never gets actually updated.
>>
>> - But this is OK, because the non-masked variant is mostly a dead
>> code, since 'fully_masked' is always false, as mdtype and np are always
>> zeroed out before calling commit_nsh().  So, the only way to hit this
>> code is to manually turn off support for masked set.
>>
>> - The put_nsh_key() function can be called for a mask, but it checks
>> the mdtype value as if it was a key regardless.  And since mdtype is
>> always zeroed out in the mask, the match on the context is never
>> propagated to the datapath flow, even if added during the action
>> commit.
>>
>> - The key and the base are compared for being the same twice, once
>> in the commit_set_nsh_action() and again in commit_nsh().
>>
>> - commit_set_nsh_action() is using an assumption that mdtype is
>> changing in the mask whenever the mask is changing.  This is fine in
>> the current code, but isn't how it supposed to work in general.
>>
>> - The masked variant of the action is generating a set with an empty
>> mask in case the context didn't change.  Same for the base header.
>> This works in userspace, but kernel generally fails the validation on
>> set() with an empty mask.  The attribute should just not be provided
>> in this case.
>>
>> - For some reason the printing function initializes the mask to an
>> exact match, printing out missing attributes in full as all-zero,
>> while they should not be printed if not present.
>>
>> - The masked variant duplicates the part of commit_masked_set_action()
>> twice, while this code can be re-used without much trouble.
>>
>> Fixing all the issues above and making the code look more like it
>> looks for other fields and header types.  Completely removing the
>> get/put_nsh_key() functions as we should not be clearing the context
>> anyway.  If we somehow end up with the MD1 context on an MD2 packet,
>> it means that we have a bug somewhere else that should be fixed.
>>
>> The tests updated to cover the problematic functionality.
>>
>> There are still more issues here.  The main one is that we need to
>> actually probe the datapath before using this action, as not all the
>> kernel versions support NSH and support for set(nsh) was recently
>> removed entirely from the Linux kernel due to being completely wrong.
>> This is the reason for not adding any system tests here, as they
>> would fail or crash the kernel.
>>
>> Will work on probing and the re-introduction of the support in the
>> kernel separately.
>>
>> Fixes: 3d2fbd70bda5 ("userspace: Add support for NSH MD1 match fields")
>> Fixes: f59cb331c481 ("nsh: rework NSH netlink keys and actions")
>> Fixes: 81fdabb94dd7 ("nsh: fix nested mask for OVS_KEY_ATTR_NSH")
>> Signed-off-by: Ilya Maximets <[email protected]>\
> 
> Thanks Ilya for fixing this mess. I have one comment below, as I got confused 
> :)
> 
> //Eelco
> 
>> ---
>>  lib/odp-util.c | 185 ++++++++++++++++++-------------------------------
>>  tests/nsh.at   |  61 +++++++++++++---
>>  2 files changed, 117 insertions(+), 129 deletions(-)
>>
>> diff --git a/lib/odp-util.c b/lib/odp-util.c
>> index ee1868202..fbbfa426f 100644
>> --- a/lib/odp-util.c
>> +++ b/lib/odp-util.c
>> @@ -1050,7 +1050,7 @@ format_odp_set_nsh(struct ds *ds, const struct nlattr 
>> *attr)
>>      struct ovs_key_nsh nsh_mask;
>>
>>      memset(&nsh, 0, sizeof nsh);
>> -    memset(&nsh_mask, 0xff, sizeof nsh_mask);
>> +    memset(&nsh_mask, 0, sizeof nsh_mask);
>>
>>      NL_NESTED_FOR_EACH (a, left, attr) {
>>          enum ovs_nsh_key_attr type = nl_attr_type(a);
>> @@ -6333,10 +6333,6 @@ static void get_arp_key(const struct flow *, struct 
>> ovs_key_arp *);
>>  static void put_arp_key(const struct ovs_key_arp *, struct flow *);
>>  static void get_nd_key(const struct flow *, struct ovs_key_nd *);
>>  static void put_nd_key(const struct ovs_key_nd *, struct flow *);
>> -static void get_nsh_key(const struct flow *flow, struct ovs_key_nsh *nsh,
>> -                        bool is_mask);
>> -static void put_nsh_key(const struct ovs_key_nsh *nsh, struct flow *flow,
>> -                        bool is_mask);
>>
>>  /* These share the same layout. */
>>  union ovs_key_tp {
>> @@ -7936,16 +7932,13 @@ commit_set_action(struct ofpbuf *odp_actions, enum 
>> ovs_key_attr key_type,
>>      nl_msg_end_nested(odp_actions, offset);
>>  }
>>
>> -/* Masked set actions have a mask following the data within the netlink
>> - * attribute.  The unmasked bits in the data will be cleared as the data
>> - * is copied to the action. */
>> -void
>> -commit_masked_set_action(struct ofpbuf *odp_actions,
>> -                         enum ovs_key_attr key_type,
>> -                         const void *key_, const void *mask_, size_t 
>> key_size)
>> +/* Commit one attribute for a masked set action.  Masked set actions have
>> + * a mask following the data within the netlink attribute.  The unmasked 
>> bits
>> + * in the data will be cleared as the data is copied to the attribute. */
>> +static void
>> +commit_masked_attribute(struct ofpbuf *odp_actions, int key_type,
>> +                        const void *key_, const void *mask_, size_t 
>> key_size)
>>  {
>> -    size_t offset = nl_msg_start_nested(odp_actions,
>> -                                        OVS_ACTION_ATTR_SET_MASKED);
>>      char *data = nl_msg_put_unspec_uninit(odp_actions, key_type, key_size * 
>> 2);
>>      const char *key = key_, *mask = mask_;
>>
>> @@ -7954,6 +7947,17 @@ commit_masked_set_action(struct ofpbuf *odp_actions,
>>      while (key_size--) {
>>          *data++ = *key++ & *mask++;
>>      }
>> +}
>> +
>> +/* A helper to commit a masked set action for a single attribute. */
>> +void
>> +commit_masked_set_action(struct ofpbuf *odp_actions,
>> +                         enum ovs_key_attr key_type,
>> +                         const void *key_, const void *mask_, size_t 
>> key_size)
>> +{
>> +    size_t offset = nl_msg_start_nested(odp_actions,
>> +                                        OVS_ACTION_ATTR_SET_MASKED);
>> +    commit_masked_attribute(odp_actions, key_type, key_, mask_, key_size);
>>      nl_msg_end_nested(odp_actions, offset);
>>  }
>>
>> @@ -8541,116 +8545,65 @@ commit_set_nw_action(const struct flow *flow, 
>> struct flow *base,
>>      return 0;
>>  }
>>
>> -static inline void
>> -get_nsh_key(const struct flow *flow, struct ovs_key_nsh *nsh, bool is_mask)
>> -{
>> -    *nsh = flow->nsh;
>> -    if (!is_mask) {
>> -        if (nsh->mdtype != NSH_M_TYPE1) {
>> -            memset(nsh->context, 0, sizeof(nsh->context));
>> -        }
>> -    }
>> -}
>> -
>> -static inline void
>> -put_nsh_key(const struct ovs_key_nsh *nsh, struct flow *flow,
>> -            bool is_mask OVS_UNUSED)
>> -{
>> -    flow->nsh = *nsh;
>> -    if (flow->nsh.mdtype != NSH_M_TYPE1) {
>> -        memset(flow->nsh.context, 0, sizeof(flow->nsh.context));
>> -    }
>> -}
>> -
>>  static bool
>> -commit_nsh(const struct ovs_key_nsh * flow_nsh, bool use_masked_set,
>> -           const struct ovs_key_nsh *key, struct ovs_key_nsh *base,
>> -           struct ovs_key_nsh *mask, size_t size,
>> +commit_nsh(bool use_masked_set, const struct ovs_key_nsh *key,
>> +           struct ovs_key_nsh *base, struct ovs_key_nsh *mask, size_t size,
>>             struct ofpbuf *odp_actions)
>>  {
>> -    enum ovs_key_attr attr = OVS_KEY_ATTR_NSH;
>> -
>> -    if (memcmp(key, base, size)  == 0) {
>> +    if (!memcmp(key, base, size)) {
>>          /* Mask bits are set when we have either read or set the 
>> corresponding
>>           * values.  Masked bits will be exact-matched, no need to set them
>>           * if the value did not actually change. */
>>          return false;
>>      }
>>
>> -    bool fully_masked = odp_mask_is_exact(attr, mask, size);
>> +    bool fully_masked = odp_mask_is_exact(OVS_KEY_ATTR_NSH, mask, size);
>> +    size_t set_ofs;
>>
>>      if (use_masked_set && !fully_masked) {
>> -        size_t nsh_key_ofs;
>> -        struct ovs_nsh_key_base nsh_base;
>> -        struct ovs_nsh_key_base nsh_base_mask;
>> -        struct ovs_nsh_key_md1 md1;
>> -        struct ovs_nsh_key_md1 md1_mask;
>> -        size_t offset = nl_msg_start_nested(odp_actions,
>> -                                            OVS_ACTION_ATTR_SET_MASKED);
>> -
>> -        nsh_base.flags = key->flags;
>> -        nsh_base.ttl = key->ttl;
>> -        nsh_base.mdtype = key->mdtype;
>> -        nsh_base.np = key->np;
>> -        nsh_base.path_hdr = key->path_hdr;
>> -
>> -        nsh_base_mask.flags = mask->flags;
>> -        nsh_base_mask.ttl = mask->ttl;
>> -        nsh_base_mask.mdtype = mask->mdtype;
>> -        nsh_base_mask.np = mask->np;
>> -        nsh_base_mask.path_hdr = mask->path_hdr;
>> -
>> -        /* OVS_KEY_ATTR_NSH keys */
>> -        nsh_key_ofs = nl_msg_start_nested(odp_actions, OVS_KEY_ATTR_NSH);
>> -
>> -        /* put value and mask for OVS_NSH_KEY_ATTR_BASE */
>> -        char *data = nl_msg_put_unspec_uninit(odp_actions,
>> -                                              OVS_NSH_KEY_ATTR_BASE,
>> -                                              2 * sizeof(nsh_base));
>> -        const char *lkey = (char *)&nsh_base, *lmask = (char 
>> *)&nsh_base_mask;
>> -        size_t lkey_size = sizeof(nsh_base);
>> -
>> -        while (lkey_size--) {
>> -            *data++ = *lkey++ & *lmask++;
>> -        }
>> -        lmask = (char *)&nsh_base_mask;
>> -        memcpy(data, lmask, sizeof(nsh_base_mask));
>> -
>> -        switch (key->mdtype) {
>> -        case NSH_M_TYPE1:
>> -            memcpy(md1.context, key->context, sizeof key->context);
>> -            memcpy(md1_mask.context, mask->context, sizeof mask->context);
>> -
>> -            /* put value and mask for OVS_NSH_KEY_ATTR_MD1 */
>> -            data = nl_msg_put_unspec_uninit(odp_actions,
>> -                                            OVS_NSH_KEY_ATTR_MD1,
>> -                                            2 * sizeof(md1));
>> -            lkey = (char *)&md1;
>> -            lmask = (char *)&md1_mask;
>> -            lkey_size = sizeof(md1);
>> -
>> -            while (lkey_size--) {
>> -                *data++ = *lkey++ & *lmask++;
>> -            }
>> -            lmask = (char *)&md1_mask;
>> -            memcpy(data, lmask, sizeof(md1_mask));
>> -            break;
>> -        case NSH_M_TYPE2:
>> -        default:
>> -            /* No match support for other MD formats yet. */
>> -            break;
>> +        struct ovs_nsh_key_base nsh_base = {
>> +            .flags = key->flags,
>> +            .ttl = key->ttl,
>> +            /* 'mdtype' and 'np' are not writable. */
>> +            .path_hdr = key->path_hdr,
>> +        };
>> +        struct ovs_nsh_key_base nsh_base_mask = {
>> +            .flags = mask->flags,
>> +            .ttl = mask->ttl,
>> +            /* 'mdtype' and 'np' are not writable. */
>> +            .path_hdr = mask->path_hdr,
>> +        };
>> +        size_t nsh_ofs;
>> +
>> +        set_ofs = nl_msg_start_nested(odp_actions, 
>> OVS_ACTION_ATTR_SET_MASKED);
>> +        nsh_ofs = nl_msg_start_nested(odp_actions, OVS_KEY_ATTR_NSH);
>> +
>> +        if (!is_all_zeros(&nsh_base_mask, sizeof nsh_base_mask)) {
>> +            commit_masked_attribute(odp_actions, OVS_NSH_KEY_ATTR_BASE,
>> +                                    &nsh_base, &nsh_base_mask,
>> +                                    sizeof nsh_base);
>>          }
>>
>> -        nl_msg_end_nested(odp_actions, nsh_key_ofs);
>> +        /* No match support for other MD formats yet. */
>> +        if (key->mdtype == NSH_M_TYPE1
>> +            && !is_all_zeros(mask->context, sizeof *mask->context)) {
> 
> Is the sizeof correct here? This points to an array element, so it would
> only give 4 bytes, not 4×4. I think this is what the old code uses?

Yeah, you're right.  Should be without the star, since it's an array.

Best regards, Ilya Maximets.
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to