race prone. Some just gather statistics, but some are later used in
different calculations.
A real problem was the race provoked underflow of the states_cur counter
on a rule. Once it goes below zero, it wraps to UINT32_MAX. Later this
value is used in pf_state_expires() and any state created by this rule
is immediately expired.
Thus, make fields states_cur, states_tot and src_nodes of struct
pf_rule be counter(9)s.
Thanks to Dennis for providing me shell access to problematic box and
his help with reproducing, debugging and investigating the problem.
Thanks to: Dennis Yusupoff <dyr smartspb.net>
Also reported by: dumbbell, pgj, Rambler
Sponsored by: Nginx, Inc.
LibAliasSetAddress() uses its own mutex to serialize changes.
While here, convert ifp->if_xname access to if_name() function.
MFC after: 2 weeks
Sponsored by: Yandex LLC
re-links dynamic states to default rule instead of
flushing on rule deletion.
This can be useful while performing ruleset reload
(think about `atomic` reload via changing sets).
Currently it is turned off by default.
MFC after: 2 weeks
Sponsored by: Yandex LLC
This lock gets deleted in sys/netpfil/ipfw/ip_fw2.c:vnet_ipfw_uninit().
Therefore, vnet_ipfw_nat_uninit() *must* be called before vnet_ipfw_uninit(),
but this doesn't always happen, because the VNET_SYSINIT order is the same for both functions.
In sys/net/netpfil/ipfw/ip_fw2.c and sys/net/netpfil/ipfw/ip_fw_nat.c,
IPFW_SI_SUB_FIREWALL == IPFW_NAT_SI_SUB_FIREWALL == SI_SUB_PROTO_IFATTACHDOMAIN
and
IPFW_MODULE_ORDER == IPFW_NAT_MODULE_ORDER
Consequently, if VIMAGE is enabled, and jails are created and destroyed,
the system sometimes crashes, because we are trying to use a deleted lock.
To reproduce the problem:
(1) Take a GENERIC kernel config, and add options for: VIMAGE, WITNESS,
INVARIANTS.
(2) Run this command in a loop:
jail -l -u root -c path=/ name=foo persist vnet && jexec foo ifconfig lo0 127.0.0.1/8 && jail -r foo
(see http://lists.freebsd.org/pipermail/freebsd-current/2010-November/021280.html )
Fix the problem by increasing the value of IPFW_NAT_SI_SUB_FIREWALL,
so that vnet_ipfw_nat_uninit() runs after vnet_ipfw_uninit().
where "m" is number of source nodes and "n" is number of states. Thus,
on heavy loaded router its processing consumed a lot of CPU time.
Reimplement it with O(m+n) complexity. We first scan through source
nodes and disconnect matching ones, putting them on the freelist and
marking with a cookie value in their expire field. Then we scan through
the states, detecting references to source nodes with a cookie, and
disconnect them as well. Then the freelist is passed to pf_free_src_nodes().
In collaboration with: Kajetan Staszkiewicz <kajetan.staszkiewicz innogames.de>
PR: kern/176763
Sponsored by: InnoGames GmbH
Sponsored by: Nginx, Inc.
- Removed pf_remove_src_node().
- Introduce pf_unlink_src_node() and pf_unlink_src_node_locked().
These function do not proceed with freeing of a node, just disconnect
it from storage.
- New function pf_free_src_nodes() works on a list of previously
disconnected nodes and frees them.
- Utilize new API in pf_purge_expired_src_nodes().
In collaboration with: Kajetan Staszkiewicz <kajetan.staszkiewicz innogames.de>
Sponsored by: InnoGames GmbH
Sponsored by: Nginx, Inc.
so they can be used in the userspace version of ipfw/dummynet
(normally using netmap for the I/O path).
This is the first of a few commits to ease compiling the
ipfw kernel code in userspace.
- Do not return blindly if proto isn't ICMP.
- The dport is in network order, so fix comparisons.
- Remove ridiculous htonl(arc4random()).
- Push local variable to a narrower block.
default from the very beginning. It was placed in wrong namespace
net.link.ether, originally it had been at another wrong namespace. It was
incorrectly documented at incorrect manual page arp(8). Since new-ARP commit,
the tunable have been consulted only on route addition, and ignored on route
deletion. Behaviour of a system with tunable turned off is not fully correct,
and has no advantages comparing to normal behavior.
Original log:
Make sure pd2 has a pointer to the icmp header in the payload; fixes
panic seen with some some icmp types in icmp error message payloads.
Obtained from: OpenBSD
Stricter state checking for ICMP and ICMPv6 packets: include the ICMP type
in one port of the state key, using the type to determine which
side should be the id, and which should be the type. Also:
- Handle ICMP6 messages which are typically sent to multicast
addresses but recieve unicast replies, by doing fallthrough lookups
against the correct multicast address. - Clear up some mistaken
assumptions in the PF code:
- Not all ICMP packets have an icmp_id, so simulate
one based on other data if we can, otherwise set it to 0.
- Don't modify the icmp id field in NAT unless it's echo
- Use the full range of possible id's when NATing icmp6 echoy
Difference with OpenBSD version:
- C99ify the new code
- WITHOUT_INET6 safe
Reviewed by: glebius
Obtained from: OpenBSD
in net, to avoid compatibility breakage for no sake.
The future plan is to split most of non-kernel parts of
pfvar.h into pf.h, and then make pfvar.h a kernel only
include breaking compatibility.
Discussed with: bz
to this event, adding if_var.h to files that do need it. Also, include
all includes that now are included due to implicit pollution via if_var.h
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
date: 2010/02/04 14:10:12; author: sthen; state: Exp; lines: +24 -19;
pf_get_sport() picks a random port from the port range specified in a
nat rule. It should check to see if it's in-use (i.e. matches an existing
PF state), if it is, it cycles sequentially through other ports until
it finds a free one. However the check was being done with the state
keys the wrong way round so it was never actually finding the state
to be in-use.
- switch the keys to correct this, avoiding random state collisions
with nat. Fixes PR 6300 and problems reported by robert@ and viq.
- check pf_get_sport() return code in pf_test(); if port allocation
fails the packet should be dropped rather than sent out untranslated.
Help/ok claudio@.
Some additional changes to 1.12:
- We also need to bzero() the key to zero padding, otherwise key
won't match.
- Collapse two if blocks into one with ||, since both conditions
lead to the same processing.
- Only naddr changes in the cycle, so move initialization of other
fields above the cycle.
- s/u_intXX_t/uintXX_t/g
PR: kern/181690
Submitted by: Olivier Cochard-Labbé <olivier cochard.me>
Sponsored by: Nginx, Inc.
thing done by the dummynet handler is taskqueue_enqueue() call, it doesn't
need extra switch to the clock SWI context.
On idle system this change in half reduces number of active CPU cycles and
wakes up only one CPU from sleep instead of two.
I was going to make this change much earlier as part of calloutng project,
but waited for better solution with skipping idle ticks to be implemented.
Unfortunately with 10.0 release coming it is better get at least this.
* Do per vnet instance cleanup (previously it was only for vnet0 on
module unload, and led to libalias leaks and possible panics due to
stale pointer dereferences).
* Instead of protecting ipfw hooks registering/deregistering by only
vnet0 lock (which does not prevent pointers access from another
vnets), introduce per vnet ipfw_nat_loaded variable. The variable is
set after hooks are registered and unset before they are deregistered.
* Devirtualize ifaddr_event_tag as we run only one event handler for
all vnets.
* It is supposed that ifaddr_change event handler is called in the
interface vnet context, so add an assertion.
Reviewed by: zec
MFC after: 2 weeks
Before this change state creating sequence was:
1) lock wire key hash
2) link state's wire key
3) unlock wire key hash
4) lock stack key hash
5) link state's stack key
6) unlock stack key hash
7) lock ID hash
8) link into ID hash
9) unlock ID hash
What could happen here is that other thread finds the state via key
hash lookup after 6), locks ID hash and does some processing of the
state. When the thread creating state unblocks, it finds the state
it was inserting already non-virgin.
Now we perform proper interlocking between key hash locks and ID hash
lock:
1) lock wire & stack hashes
2) link state's keys
3) lock ID hash
4) unlock wire & stack hashes
5) link into ID hash
6) unlock ID hash
To achieve that, the following hacking was performed in pf_state_key_attach():
- Key hash mutex is marked with MTX_DUPOK.
- To avoid deadlock on 2 key hash mutexes, we lock them in order determined
by their address value.
- pf_state_key_attach() had a magic to reuse a > FIN_WAIT_2 state. It unlinked
the conflicting state synchronously. In theory this could require locking
a third key hash, which we can't do now.
Now we do not remove the state immediately, instead we leave this task to
the purge thread. To avoid conflicts in a short period before state is
purged, we push to the very end of the TAILQ.
- On success, before dropping key hash locks, pf_state_key_attach() locks
ID hash and returns.
Tested by: Ian FREISLICH <ianf clue.co.za>
