[PATCH RESEND 1/8] netlink: call get_fd_nl_family before decode nlmsghdr

[JingPiao Chen]

On Tue, Aug 08, 2017 at 03:22:31PM +0300, Dmitry V. Levin wrote:
> On Tue, Aug 08, 2017 at 03:07:40PM +0300, Dmitry V. Levin wrote:
> > On Sun, Aug 06, 2017 at 07:52:17AM +0800, JingPiao Chen wrote:
> > > On Sun, Aug 06, 2017 at 02:10:16AM +0300, Dmitry V. Levin wrote:
> > > > On Tue, Aug 01, 2017 at 07:48:41AM +0800, JingPiao Chen wrote:
> > > > > Prepare for NETLINK_KOBJECT_UEVENT decode. The messages
> > > > > of NETLINK_KOBJECT_UEVENT do not contain nlmsghdr.
> > > > [...]
> > > > > (decode_nlmsghdr_with_payload): Skip family specific decoders
> > > > > for type < NLMSG_MIN_TYPE && type != NLMSG_DONE.
> > > > [...]
> > > > > @@ -182,7 +177,7 @@ decode_nlmsg_type(const uint16_t type, const unsigned int family)
> > > > >  	const struct xlat *xlat = netlink_types;
> > > > >  	const char *dflt = "NLMSG_???";
> > > > >  
> > > > > -	if (type != NLMSG_DONE && family < ARRAY_SIZE(nlmsg_types)) {
> > > > > +	if (type >= NLMSG_MIN_TYPE && family < ARRAY_SIZE(nlmsg_types)) {
> > > > 
> > > > Why?  How this is related to the rest of NETLINK_KOBJECT_UEVENT change?
> > > 
> > > Previous code when type < NLMSG_MIN_TYPE && type != NLMSG_DONE,
> > > family = -2 (NL_FAMILY_DEFAULT), family < ARRAY_SIZE(nlmsg_types) filter
> > > this case. Now get family ignore the nlmsg_type.
> > > When type < NLMSG_MIN_TYPE && type != NLMSG_DONE, family is not
> > > a negative. family < ARRAY_SIZE(nlmsg_types) can not filter this case.
> > > 
> > > Related commit:
> > > v4.17-43-g1b63425, v4.17-44-g8700030: These commit introduce get_fd_nl_family.
> > > And get family only when type >= NLMSG_MIN_TYPE.
> > > 
> > > v4.17-123-g54aed21: This commit changed get family when
> > > type >= NLMSG_MIN_TYPE || type == NLMSG_DONE.
> > 
> > Now that you change things to call get_fd_nl_family from decode_netlink
> > unconditionally, family is available unconditionally, too (although it
> > can be -1).  This is all clear enough but doesn't answer my question:
> > 
> > Why do you think it means that decode_nlmsg_type should not use
> > family-specific decoders for message type < NLMSG_MIN_TYPE?
> > Is decoding of these message types 100% family-agnostic?
> > 
> > It's so far from being obvious that if it's true, there must be a comment
> > explaining why it's true.
>
> Now that you change things to call get_fd_nl_family from decode_netlink
> unconditionally, family is available unconditionally, too (although it
> can be -1).  This is all clear enough but doesn't answer my question:
> 
> Why do you think it means that decode_nlmsg_type should not use
> family-specific decoders for message type < NLMSG_MIN_TYPE?
> Is decoding of these message types 100% family-agnostic?
> 
> It's so far from being obvious that if it's true, there must be a comment
> explaining why it's true.

include/linux/netlink.h:
#define NLMSG_MIN_TYPE          0x10    /* < 0x10: reserved control messages */

If use decode_nlmsg_type family-specific decoders for message
type < NLMSG_MIN_TYPE, test will fail:
-sendto(3, {len=16, type=NLMSG_ERROR, ...
+sendto(3, {len=16, type=0x2 /* SOCK_DIAG_??? */, ...

comment:
	/* type < NLMSG_MIN_TYPE are reserved control messages. */

> The same issue is with message type checks in decode_payload:
> there has to be a comment explaining why family-specific netlink decoders
> are not invoked for (nlmsg_type < 0 && nlmsg_type != NLMSG_DONE).

v4.17-123-g54aed21 commit message:
While many NLMSG_DONE messages indeed have payload containing
just one integer, there are exceptions. Handle this by passing
payloads of NLMSG_DONE messages to family specific netlink
payload decoders.

I shorten it as comment:
	/*
	 * While many NLMSG_DONE messages indeed have payload
	 * containing just one integer, there are exceptions.
	 * Passing payloads of NLMSG_DONE messages to family
	 * specific netlink payload decoders.
	 */
This comment only explain why invoke family-specific netlink decoders
for nlmsg_type == NLMSG_DONE.

Are you prefer to handle type < NLMSG_MIN_TYPE in family-specific
netlink decoders? (decode_netlink_sock_diag should make a change,
decode_netlink_selinux and decode_netlink_crypto can handle now.)

Need I send version 2?

--
JingPiao Chen