Date:	Sat, 19 Mar 2011 01:12:00 +0200 (EET)
From:	Julian Anastasov <ja@....bg>
To:	netdev@...r.kernel.org
cc:	David Miller <davem@...emloft.net>
Subject: [PATCH 0/4] ipv4: fix problems with same IPs in many subnets


 	Here are some changes for the IPv4 addresses and
the routes we delete for setups using same primary or
secondary IPs in many subnets on device. They are made for
net-next from March 13.

 	Attached is a test script that tests the kernel
for the observed problems and validates some operations.
The same script should be available here:

http://www.ssi.bg/~ja/test_ipdel.sh

 	Can be run before and after patching to compare results.

 	Usage:

# Tune script: device, etc

# Run test with generating sorted listings, useful for diff -u
./test_ipdel.sh list > out1

# Run test with brief output
./test_ipdel.sh

 	As there some questions on using same IPs in
different subnets, here is some info and the current
problems I'm seeing.

 	__inet_insert_ifa() considers IPs as secondary when
they use same prefixlen (mask) as existing primary IP. So, it is
perfectly valid for some device:

- to have two primary IFAs for same IP "P1" but with different
masks "M1" and "M2", probably with different scope (visibility
for the defined subnet):
 	IFA1: P1/M1
 	IFA2: P1/M2

- when adding secondary IPs the same rule applies:
 	IFA3: S1/M1, can use primary P1 from IFA1
 	IFA4: S1/M2, can use primary P1 from IFA2

 	inet_ifa_byprefix() can correctly match IFA1 or IFA2
by address (P1) and mask (M1 or M2).

 	To summarize for IPs: the combination of device, IP address
and prefixlen (mask) should be unique. Same IPs can be used
in many overlapping subnets. The IPs for device are always
saved in sorted form: all primaries before all secondaries, the
primaries are sorted by decreasing scope, primaries with
same scope are in order of adding, the secondaries are in
random (adding) order but with same scope as their primary.

 	When addresses are converted to local or
broadcast routes we have:

- All routes are added with prefsrc. The prefsrc is the primary
address. As the combination of local IP and its primary address
is unique only for its subnet if we use the same IPs in another
subnet it is possible to try to add the same routes twice
resulting in duplication. But only one route is created in such case
and on deletion we should be careful not to delete routes that
are also required for the same IPs in different subnet.

- When creating local or broadcast route for the added address we
ignore the prefixlen. As result, the combination of address
(FIB node=IP/32), primary IP (fib_prefsrc) and device (fib_info, NH)
should be unique. The above IFA1 and IFA2 will try to create same
local route:

local P1/32 via dev DEV ... src P1

 	and the result will be single route.

 	IFA3 and IFA4 will add single route too:

local S1/32 via dev DEV ... src P1

 	When addresses are deleted we should remove exactly the
routes that were created for our IFA but with checks for
duplication in another subnet.

 	The problems in current code:

- fib_del_ifaddr() unconditionally tries to remove the subnet
route when deleting primary IFA. The problem is that if
prefixlen is 32 we need more checks, eg. if this IP is
still used in another subnet.

- fib_del_ifaddr() forgets to match the primary address for
every IFA that is checked. As result, routes for deleted
secondary IFAs can remain until their primary IFA is deleted
just because the same address is used in another subnet
but with different primary IP.

- fib_table_delete() does not match the prefsrc, so if there
are many routes that differ in prefsrc, always the first
route is deleted. By this way, we can not remove the right
route for the deleted IFAs.

- On secondary address promotion, when deleting the primary
address all routes with such prefsrc are currently expected to
be deleted by fib_sync_down_addr without sending events for
the remaining secondaries. But as fib_sync_down_addr calling
depends on other subnets and even on other devices we can leak
the routes.

 	The following patches fix the above problems.
I don't try to forbid same IPs to be used in multiple
subnets because it is perfectly valid subnets to overlap.

 	As patch 2 needs to know the primary IP (but not its IFA)
for every secondary address, another implementation is possible
where new field ifa_primary can cache the primary IP used
for the IFA. Then we will avoid the inet_ifa_byprefix calls
when we walk secondaries from different subnets.

 	For now I'm using a 'same_prefsrc' flag that detects if
multiple subnets use same primary IP, so such lookups are
avoided later for the secondaries for the common case when
same IPs are not used in more than one subnet.

Regards

--
Julian Anastasov <ja@....bg>
Download attachment "test_ipdel.sh" of type "APPLICATION/x-sh" (3655 bytes)

Powered by blists - more mailing lists