mbox series

[net-next,0/6] ethtool: allow dumping policies to user space

Message ID 20201005155753.2333882-1-kuba@kernel.org
Headers show
Series ethtool: allow dumping policies to user space | expand

Message

Jakub Kicinski Oct. 5, 2020, 3:57 p.m. UTC
Hi!

This series wires up ethtool policies to ops, so they can be
dumped to user space for feature discovery.

First two patches wire up GET, third patch wires up SET.

Next - take care of linking up nested policies for the header
(which is what we actually care about right now). And once header
policy is linked make sure that attribute range validation is
done by policy, not code conditions for flags. New type of
policy is needed to validate masks (patch 5).

Netlink as always staying a step ahead of all the other kernel
API interfaces :)

Jakub Kicinski (6):
  ethtool: wire up get policies to ops
  ethtool: use the attributes parsed by the core in get commands
  ethtool: wire up set policies to ops
  ethtool: link up ethnl_header_policy as a nested policy
  netlink: add mask validation
  ethtool: specify which header flags are supported per command

 include/net/netlink.h        |  11 ++++
 include/uapi/linux/netlink.h |   2 +
 lib/nlattr.c                 |  36 ++++++++++
 net/ethtool/cabletest.c      |  30 +++------
 net/ethtool/channels.c       |  22 +++----
 net/ethtool/coalesce.c       |  22 +++----
 net/ethtool/debug.c          |  20 ++----
 net/ethtool/eee.c            |  21 +++---
 net/ethtool/features.c       |  22 +++----
 net/ethtool/linkinfo.c       |  22 +++----
 net/ethtool/linkmodes.c      |  22 +++----
 net/ethtool/linkstate.c      |   8 +--
 net/ethtool/netlink.c        | 123 +++++++++++++++++++++++++----------
 net/ethtool/netlink.h        |  33 +++++++++-
 net/ethtool/pause.c          |  19 ++----
 net/ethtool/privflags.c      |  22 +++----
 net/ethtool/rings.c          |  20 ++----
 net/ethtool/strset.c         |   6 +-
 net/ethtool/tsinfo.c         |   7 +-
 net/ethtool/tunnels.c        |  42 +++++-------
 net/ethtool/wol.c            |  19 ++----
 net/netlink/policy.c         |   8 +++
 22 files changed, 303 insertions(+), 234 deletions(-)

Comments

Jakub Kicinski Oct. 5, 2020, 7:16 p.m. UTC | #1
On Mon, 05 Oct 2020 20:56:29 +0200 Johannes Berg wrote:
> On Mon, 2020-10-05 at 08:57 -0700, Jakub Kicinski wrote:
> > @@ -783,6 +799,9 @@ static const struct genl_ops ethtool_genl_ops[] = {
> >  		.start	= ethnl_default_start,
> >  		.dumpit	= ethnl_default_dumpit,
> >  		.done	= ethnl_default_done,
> > +		.policy = ethnl_rings_get_policy,
> > +		.maxattr = ARRAY_SIZE(ethnl_rings_get_policy) - 1,
> > +
> >  	},  
> 
> If you find some other reason to respin, perhaps remove that blank line
> :)
> 
> Unrelated to that, it bothers me a bit that you put here the maxattr as
> the ARRAY_SIZE(), which is of course fine, but then still have
> 
> > @@ -127,7 +127,7 @@ const struct ethnl_request_ops ethnl_privflags_request_ops = {
> >  	.max_attr		= ETHTOOL_A_PRIVFLAGS_MAX,  
> 
> max_attr here, using the original define

Ah, another good catch, this is obviously no longer needed. I will
remove those members in v2.

> yes, mostly the policies use
> the define to size them, but they didn't really *need* to, and one might
> make an argument that on the policy arrays the size might as well be
> removed (and it be sized automatically based on the contents) since all
> the unspecified attrs are rejected anyway.
> 
> But with the difference it seems to me that it'd be possible to get this
> mixed up?

Right, I prefer not to have the unnecessary NLA_REJECTS, so my thinking
was - use the format I like for the new code, but leave the existing
rejects for a separate series / discussion.

If we remove the rejects we still need something like

extern struct nla_policy policy[lastattr + 1];

For array_size to work, but I think that's fine. And we'd get a
compiler errors if the sizes don't match up.

> I do see that you still need this to size the attrs for parsing them
> even after patch 2 where this:
> 
> >  	.req_info_size		= sizeof(struct privflags_req_info),
> >  	.reply_data_size	= sizeof(struct privflags_reply_data),
> > -	.request_policy		= privflags_get_policy,
> > +	.request_policy		= ethnl_privflags_get_policy,  
> 
> gets removed completely.
> 
> 
> Perhaps we can look up the genl_ops pointer, or add the ops pointer to
> struct genl_info (could point to the temporary full struct that gets
> populated, size of genl_info itself doesn't matter much since it's on
> the stack and temporary), and then use ops->maxattr instead of
> request_ops->max_attr in ethnl_default_parse()?

Hm, maybe my split of patches 1 and 2 hurts more than it helps.
Let me merge the two in v2.
Johannes Berg Oct. 5, 2020, 7:21 p.m. UTC | #2
On Mon, 2020-10-05 at 12:16 -0700, Jakub Kicinski wrote:
> On Mon, 05 Oct 2020 20:56:29 +0200 Johannes Berg wrote:
> > On Mon, 2020-10-05 at 08:57 -0700, Jakub Kicinski wrote:
> > > @@ -783,6 +799,9 @@ static const struct genl_ops ethtool_genl_ops[] = {
> > >  		.start	= ethnl_default_start,
> > >  		.dumpit	= ethnl_default_dumpit,
> > >  		.done	= ethnl_default_done,
> > > +		.policy = ethnl_rings_get_policy,
> > > +		.maxattr = ARRAY_SIZE(ethnl_rings_get_policy) - 1,
> > > +
> > >  	},  
> > 
> > If you find some other reason to respin, perhaps remove that blank line
> > :)
> > 
> > Unrelated to that, it bothers me a bit that you put here the maxattr as
> > the ARRAY_SIZE(), which is of course fine, but then still have
> > 
> > > @@ -127,7 +127,7 @@ const struct ethnl_request_ops ethnl_privflags_request_ops = {
> > >  	.max_attr		= ETHTOOL_A_PRIVFLAGS_MAX,  
> > 
> > max_attr here, using the original define
> 
> Ah, another good catch, this is obviously no longer needed. I will
> remove those members in v2.

Good point, I misread/misunderstood the code and thought it was still
being used to size the parsing array, but that's of course no longer
there since the genl core now does it.

> > But with the difference it seems to me that it'd be possible to get this
> > mixed up?
> 
> Right, I prefer not to have the unnecessary NLA_REJECTS, so my thinking
> was - use the format I like for the new code, but leave the existing
> rejects for a separate series / discussion.
> 
> If we remove the rejects we still need something like
> 
> extern struct nla_policy policy[lastattr + 1];

Not sure I understand? You're using strict validation (I think), so
attrs that are out of range will be rejected same as NLA_REJECT (well,
with a different message) in __nla_validate_parse():

        nla_for_each_attr(nla, head, len, rem) {
                u16 type = nla_type(nla);

                if (type == 0 || type > maxtype) {
                        if (validate & NL_VALIDATE_MAXTYPE) {
                                NL_SET_ERR_MSG_ATTR(extack, nla,
                                                    "Unknown attribute type");
                                return -EINVAL;
                        }


In fact, if you're using strict validation even the default
(0==NLA_UNSPEC) will be rejected, just like NLA_REJECT.


Or am I confused somewhere?

johannes
Johannes Berg Oct. 5, 2020, 7:28 p.m. UTC | #3
On Mon, 2020-10-05 at 12:25 -0700, Jakub Kicinski wrote:
> On Mon, 05 Oct 2020 20:58:57 +0200 Johannes Berg wrote:
> > On Mon, 2020-10-05 at 08:57 -0700, Jakub Kicinski wrote:
> > > @@ -47,19 +61,16 @@ int ethnl_parse_header_dev_get(struct ethnl_req_info *req_info,
> > >  		NL_SET_ERR_MSG(extack, "request header missing");
> > >  		return -EINVAL;
> > >  	}
> > > +	/* Use most permissive header policy here, ops should specify their
> > > +	 * actual header policy via NLA_POLICY_NESTED(), and the real
> > > +	 * validation will happen in genetlink code.
> > > +	 */
> > >  	ret = nla_parse_nested(tb, ETHTOOL_A_HEADER_MAX, header,
> > > -			       ethnl_header_policy, extack);
> > > +			       ethnl_header_policy_stats, extack);  
> > 
> > Would it make sense to just remove the validation here? It's already
> > done, so it just costs extra cycles and can't really fail, and if there
> > are more diverse policies in the future this might also very quickly get
> > out of hand?
> 
> I was slightly worried I missed a command and need last line of defence,

Ah. I was just about to suggest to put it into the family policy/maxattr
but that won't work of course since this is nested.

But actually what you _could_ put there is a dummy policy (non-NULL
pointer) with a maxattr of 0, and then all attrs will be completely
rejected for a command where the policy was missed.

Not if you missed the NLA_POLICY_NESTED() link, though.

johannes
Jakub Kicinski Oct. 5, 2020, 7:31 p.m. UTC | #4
On Mon, 05 Oct 2020 21:21:36 +0200 Johannes Berg wrote:
> > > But with the difference it seems to me that it'd be possible to get this
> > > mixed up?  
> > 
> > Right, I prefer not to have the unnecessary NLA_REJECTS, so my thinking
> > was - use the format I like for the new code, but leave the existing
> > rejects for a separate series / discussion.
> > 
> > If we remove the rejects we still need something like
> > 
> > extern struct nla_policy policy[lastattr + 1];  
> 
> Not sure I understand? You're using strict validation (I think), so
> attrs that are out of range will be rejected same as NLA_REJECT (well,
> with a different message) in __nla_validate_parse():
> 
>         nla_for_each_attr(nla, head, len, rem) {
>                 u16 type = nla_type(nla);
> 
>                 if (type == 0 || type > maxtype) {
>                         if (validate & NL_VALIDATE_MAXTYPE) {
>                                 NL_SET_ERR_MSG_ATTR(extack, nla,
>                                                     "Unknown attribute type");
>                                 return -EINVAL;
>                         }
> 
> 
> In fact, if you're using strict validation even the default
> (0==NLA_UNSPEC) will be rejected, just like NLA_REJECT.
> 
> 
> Or am I confused somewhere?

Yea, I think we're both confused. Agreed with the above.

Are you suggesting:

const struct nla_policy policy[/* no size */] = {
	[HEADER]	= NLA_POLICY(...)
	[OTHER_ATTR]	= NLA_POLICY(...)
};

extern const struct nla_policy policy[/* no size */];

op = {
	.policy = policy,
	.max_attr = OTHER_ATTR,
}

?

What I'm saying is that my preference would be:

const struct nla_policy policy[OTHER_ATTR + 1] = {
	[HEADER]	= NLA_POLICY(...)
	[OTHER_ATTR]	= NLA_POLICY(...)
};

extern const struct nla_policy policy[OTHER_ATTR + 1];

op = {
	.policy = policy,
	.max_attr = ARRAY_SIZE(policy) - 1,
}

Since it's harder to forget to update the op (you don't have to update
op, and compiler will complain about the extern out of sync).
Johannes Berg Oct. 5, 2020, 7:33 p.m. UTC | #5
On Mon, 2020-10-05 at 12:31 -0700, Jakub Kicinski wrote:

> Yea, I think we're both confused. Agreed with the above.
> 
> Are you suggesting:
> 
> const struct nla_policy policy[/* no size */] = {
> 	[HEADER]	= NLA_POLICY(...)
> 	[OTHER_ATTR]	= NLA_POLICY(...)
> };
> 
> extern const struct nla_policy policy[/* no size */];
> 
> op = {
> 	.policy = policy,
> 	.max_attr = OTHER_ATTR,
> }

No, that'd be awkward, for the reason you stated below.

> What I'm saying is that my preference would be:
> 
> const struct nla_policy policy[OTHER_ATTR + 1] = {
> 	[HEADER]	= NLA_POLICY(...)
> 	[OTHER_ATTR]	= NLA_POLICY(...)
> };
> 
> extern const struct nla_policy policy[OTHER_ATTR + 1];
> 
> op = {
> 	.policy = policy,
> 	.max_attr = ARRAY_SIZE(policy) - 1,
> }
> 
> Since it's harder to forget to update the op (you don't have to update
> op, and compiler will complain about the extern out of sync).

Yeah.

I was thinking the third way ;-)

const struct nla_policy policy[] = {
	[HEADER]	= NLA_POLICY(...)
	[OTHER_ATTR]	= NLA_POLICY(...)
};

op = {
	.policy = policy,
	.maxattr = ARRAY_SIZE(policy) - 1,
};


Now you can freely add any attributes, and, due to strict validation,
anything not specified in the policy will be rejected, whether by being
out of range (> maxattr) or not specified (NLA_UNSPEC).

johannes
Jakub Kicinski Oct. 5, 2020, 7:41 p.m. UTC | #6
On Mon, 05 Oct 2020 21:33:55 +0200 Johannes Berg wrote:
> > What I'm saying is that my preference would be:
> > 
> > const struct nla_policy policy[OTHER_ATTR + 1] = {
> > 	[HEADER]	= NLA_POLICY(...)
> > 	[OTHER_ATTR]	= NLA_POLICY(...)
> > };
> > 
> > extern const struct nla_policy policy[OTHER_ATTR + 1];
> > 
> > op = {
> > 	.policy = policy,
> > 	.max_attr = ARRAY_SIZE(policy) - 1,
> > }
> > 
> > Since it's harder to forget to update the op (you don't have to update
> > op, and compiler will complain about the extern out of sync).  
> 
> Yeah.
> 
> I was thinking the third way ;-)
> 
> const struct nla_policy policy[] = {
> 	[HEADER]	= NLA_POLICY(...)
> 	[OTHER_ATTR]	= NLA_POLICY(...)
> };
> 
> op = {
> 	.policy = policy,
> 	.maxattr = ARRAY_SIZE(policy) - 1,
> };
> 
> 
> Now you can freely add any attributes, and, due to strict validation,
> anything not specified in the policy will be rejected, whether by being
> out of range (> maxattr) or not specified (NLA_UNSPEC).

100%, but in ethtool policy is defined in a different compilation unit
than the op array.
Johannes Berg Oct. 5, 2020, 7:46 p.m. UTC | #7
On Mon, 2020-10-05 at 12:41 -0700, Jakub Kicinski wrote:

> > Now you can freely add any attributes, and, due to strict validation,
> > anything not specified in the policy will be rejected, whether by being
> > out of range (> maxattr) or not specified (NLA_UNSPEC).
> 
> 100%, but in ethtool policy is defined in a different compilation unit
> than the op array.

Ah. OK, then that won't work, of course, never mind.

I'd probably go with your preference then, but perhaps drop the actual
size definition:

const struct nla_policy policy[] = {
...
};

extern const struct nla_policy policy[OTHER_ATTR + 1];

op = {
        .policy = policy,
        .max_attr = ARRAY_SIZE(policy) - 1,
}


But that'd really just be to save typing copying it if it ever changes,
since it's compiler checked for consistency.

johannes
Jakub Kicinski Oct. 5, 2020, 7:51 p.m. UTC | #8
On Mon, 05 Oct 2020 21:46:02 +0200 Johannes Berg wrote:
> On Mon, 2020-10-05 at 12:41 -0700, Jakub Kicinski wrote:
> 
> > > Now you can freely add any attributes, and, due to strict validation,
> > > anything not specified in the policy will be rejected, whether by being
> > > out of range (> maxattr) or not specified (NLA_UNSPEC).  
> > 
> > 100%, but in ethtool policy is defined in a different compilation unit
> > than the op array.  
> 
> Ah. OK, then that won't work, of course, never mind.
> 
> I'd probably go with your preference then, but perhaps drop the actual
> size definition:
> 
> const struct nla_policy policy[] = {
> ...
> };
> 
> extern const struct nla_policy policy[OTHER_ATTR + 1];
> 
> op = {
>         .policy = policy,
>         .max_attr = ARRAY_SIZE(policy) - 1,
> }
> 
> 
> But that'd really just be to save typing copying it if it ever changes,
> since it's compiler checked for consistency.

Sounds good, will do (unless Michal speaks up and prefers otherwise :)).
Keller, Jacob E Oct. 5, 2020, 9:33 p.m. UTC | #9
On 10/5/2020 12:31 PM, Jakub Kicinski wrote:
> On Mon, 05 Oct 2020 21:21:36 +0200 Johannes Berg wrote:
>>>> But with the difference it seems to me that it'd be possible to get this
>>>> mixed up?  
>>>
>>> Right, I prefer not to have the unnecessary NLA_REJECTS, so my thinking
>>> was - use the format I like for the new code, but leave the existing
>>> rejects for a separate series / discussion.
>>>
>>> If we remove the rejects we still need something like
>>>
>>> extern struct nla_policy policy[lastattr + 1];  
>>
>> Not sure I understand? You're using strict validation (I think), so
>> attrs that are out of range will be rejected same as NLA_REJECT (well,
>> with a different message) in __nla_validate_parse():
>>
>>         nla_for_each_attr(nla, head, len, rem) {
>>                 u16 type = nla_type(nla);
>>
>>                 if (type == 0 || type > maxtype) {
>>                         if (validate & NL_VALIDATE_MAXTYPE) {
>>                                 NL_SET_ERR_MSG_ATTR(extack, nla,
>>                                                     "Unknown attribute type");
>>                                 return -EINVAL;
>>                         }
>>
>>
>> In fact, if you're using strict validation even the default
>> (0==NLA_UNSPEC) will be rejected, just like NLA_REJECT.
>>
>>
>> Or am I confused somewhere?
> 
> Yea, I think we're both confused. Agreed with the above.
> 
> Are you suggesting:
> 
> const struct nla_policy policy[/* no size */] = {
> 	[HEADER]	= NLA_POLICY(...)
> 	[OTHER_ATTR]	= NLA_POLICY(...)
> };
> 
> extern const struct nla_policy policy[/* no size */];
> 
> op = {
> 	.policy = policy,
> 	.max_attr = OTHER_ATTR,
> }
> 

Why can't .max_attr here just be derived from ARRAY_SIZE? In this
example, ARRAY_SIZE should return 2, so max_attr would be ARRAY_SIZE - 1.

Well, I guess HEADER/OTHER_ATTR could make this a sparse array... in
which case it might not work?

> ?
> 
> What I'm saying is that my preference would be:
> 
> const struct nla_policy policy[OTHER_ATTR + 1] = {
> 	[HEADER]	= NLA_POLICY(...)
> 	[OTHER_ATTR]	= NLA_POLICY(...)
> };
> 
> extern const struct nla_policy policy[OTHER_ATTR + 1];
> 
> op = {
> 	.policy = policy,
> 	.max_attr = ARRAY_SIZE(policy) - 1,
> }
> 
> Since it's harder to forget to update the op (you don't have to update
> op, and compiler will complain about the extern out of sync).
> 
Ahh.. Ok so I guess what you're doing here is defining the array size as
OTHER_ATTR + 1, so that ARRAY_SIZE() - 1 guarantees to equal
OTHER_ATTR.. so as long as OTHER_ATTR is the largest element in the array...

But wouldn't that be the case in the previous example where array size
is automatically determined... as long as you keep the index ordering in
ascending order, then the size of the array would be 1 larger than the
last element...
Keller, Jacob E Oct. 5, 2020, 9:52 p.m. UTC | #10
On 10/5/2020 12:33 PM, Johannes Berg wrote:
> On Mon, 2020-10-05 at 12:31 -0700, Jakub Kicinski wrote:
> 
>> Yea, I think we're both confused. Agreed with the above.
>>
>> Are you suggesting:
>>
>> const struct nla_policy policy[/* no size */] = {
>> 	[HEADER]	= NLA_POLICY(...)
>> 	[OTHER_ATTR]	= NLA_POLICY(...)
>> };
>>
>> extern const struct nla_policy policy[/* no size */];
>>
>> op = {
>> 	.policy = policy,
>> 	.max_attr = OTHER_ATTR,
>> }
> 
> No, that'd be awkward, for the reason you stated below.
> 
>> What I'm saying is that my preference would be:
>>
>> const struct nla_policy policy[OTHER_ATTR + 1] = {
>> 	[HEADER]	= NLA_POLICY(...)
>> 	[OTHER_ATTR]	= NLA_POLICY(...)
>> };
>>
>> extern const struct nla_policy policy[OTHER_ATTR + 1];
>>
>> op = {
>> 	.policy = policy,
>> 	.max_attr = ARRAY_SIZE(policy) - 1,
>> }
>>
>> Since it's harder to forget to update the op (you don't have to update
>> op, and compiler will complain about the extern out of sync).
> 
> Yeah.
> 
> I was thinking the third way ;-)
> 
> const struct nla_policy policy[] = {
> 	[HEADER]	= NLA_POLICY(...)
> 	[OTHER_ATTR]	= NLA_POLICY(...)
> };
> 
> op = {
> 	.policy = policy,
> 	.maxattr = ARRAY_SIZE(policy) - 1,
> };
> 
> 
> Now you can freely add any attributes, and, due to strict validation,
> anything not specified in the policy will be rejected, whether by being
> out of range (> maxattr) or not specified (NLA_UNSPEC).
> 
> johannes
> 

This is what I was thinking of as well.