The call to flush_work before tearing down a table from the netlink
notifier was supposed to make sure that all earlier updates (e.g. rule
add) that might reference that table have been processed.
Unfortunately, flush_work() waits for the last queued instance.
This could be an instance that is different from the one that we must
wait for.
This is because transactions are protected with a pernet mutex, but the
work item is global, so holding the transaction mutex doesn't prevent
another netns from queueing more work.
Make the work item pernet so that flush_work() will wait for all
transactions queued from this netns.
A welcome side effect is that we no longer need to wait for transaction
objects from foreign netns.
The gc work queue is still global. This seems to be ok because nft_set
structures are reference counted and each container structure owns a
reference on the net namespace.
The destroy_list is still protected by a global spinlock rather than
pernet one but the hold time is very short anyway.
v2: call cancel_work_sync before reaping the remaining tables (Pablo).
Fixes: 9f6958ba2e ("netfilter: nf_tables: unconditionally flush pending work before notifier")
Reported-by: syzbot+5d8c5789c8cb076b2c25@syzkaller.appspotmail.com
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
tcp rst/fin packet triggers an immediate teardown of the flow which
results in sending flows back to the classic forwarding path.
This behaviour was introduced by:
da5984e510 ("netfilter: nf_flow_table: add support for sending flows back to the slow path")
b6f27d322a ("netfilter: nf_flow_table: tear down TCP flows if RST or FIN was seen")
whose goal is to expedite removal of flow entries from the hardware
table. Before these patches, the flow was released after the flow entry
timed out.
However, this approach leads to packet races when restoring the
conntrack state as well as late flow re-offload situations when the TCP
connection is ending.
This patch adds a new CLOSING state that is is entered when tcp rst/fin
packet is seen. This allows for an early removal of the flow entry from
the hardware table. But the flow entry still remains in software, so tcp
packets to shut down the flow are not sent back to slow path.
If syn packet is seen from this new CLOSING state, then this flow enters
teardown state, ct state is set to TCP_CONNTRACK_CLOSE state and packet
is sent to slow path, so this TCP reopen scenario can be handled by
conntrack. TCP_CONNTRACK_CLOSE provides a small timeout that aims at
quickly releasing this stale entry from the conntrack table.
Moreover, skip hardware re-offload from flowtable software packet if the
flow is in CLOSING state.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Offload nf_conn entries may not see traffic for a very long time.
To prevent incorrect 'ct is stale' checks during nf_conntrack table
lookup, the gc worker extends the timeout nf_conn entries marked for
offload to a large value.
The existing logic suffers from a few problems.
Garbage collection runs without locks, its unlikely but possible
that @ct is removed right after the 'offload' bit test.
In that case, the timeout of a new/reallocated nf_conn entry will
be increased.
Prevent this by obtaining a reference count on the ct object and
re-check of the confirmed and offload bits.
If those are not set, the ct is being removed, skip the timeout
extension in this case.
Parallel teardown is also problematic:
cpu1 cpu2
gc_worker
calls flow_offload_teardown()
tests OFFLOAD bit, set
clear OFFLOAD bit
ct->timeout is repaired (e.g. set to timeout[UDP_CT_REPLIED])
nf_ct_offload_timeout() called
expire value is fetched
<INTERRUPT>
-> NF_CT_DAY timeout for flow that isn't offloaded
(and might not see any further packets).
Use cmpxchg: if ct->timeout was repaired after the 2nd 'offload bit' test
passed, then ct->timeout will only be updated of ct->timeout was not
altered in between.
As we already have a gc worker for flowtable entries, ct->timeout repair
can be handled from the flowtable gc worker.
This avoids having flowtable specific logic in the conntrack core
and avoids checking entries that were never offloaded.
This allows to remove the nf_ct_offload_timeout helper.
Its safe to use in the add case, but not on teardown.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Its not used (and could be NULL), so remove it.
This allows to use nf_ct_refresh in places where we don't have
an skb without having to double-check that skb == NULL would be safe.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Do not drop a netdev-family chain if the last interface it is registered
for vanishes. Users dumping and storing the ruleset upon shutdown to
restore it upon next boot may otherwise lose the chain and all contained
rules. They will still lose the list of devices, a later patch will fix
that. For now, this aligns the event handler's behaviour with that for
flowtables.
The controversal situation at netns exit should be no problem here:
event handler will unregister the hooks, core nftables cleanup code will
drop the chain itself.
Signed-off-by: Phil Sutter <phil@nwl.cc>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Prepare for hooks with NULL ops.dev pointer (due to non-existent device)
and store the interface name and length as specified by the user upon
creation. No functional change intended.
Signed-off-by: Phil Sutter <phil@nwl.cc>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
The existing rbtree implementation uses singleton elements to represent
ranges, however, userspace provides a set size according to the number
of ranges in the set.
Adjust provided userspace set size to the number of singleton elements
in the kernel by multiplying the range by two.
Check if the no-match all-zero element is already in the set, in such
case release one slot in the set size.
Fixes: 0ed6389c48 ("netfilter: nf_tables: rename set implementations")
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Nadia Pinaeva writes:
I am working on a tool that allows collecting network performance
metrics by using conntrack events.
Start time of a conntrack entry is used to evaluate seen_reply
latency, therefore the sooner it is timestamped, the better the
precision is.
In particular, when using this tool to compare the performance of the
same feature implemented using iptables/nftables/OVS it is crucial
to have the entry timestamped earlier to see any difference.
At this time, conntrack events can only get timestamped at recv time in
userspace, so there can be some delay between the event being generated
and the userspace process consuming the message.
There is sys/net/netfilter/nf_conntrack_timestamp, which adds a
64bit timestamp (ns resolution) that records start and stop times,
but its not suited for this either, start time is the 'hashtable insertion
time', not 'conntrack allocation time'.
There is concern that moving the start-time moment to conntrack
allocation will add overhead in case of flooding, where conntrack
entries are allocated and released right away without getting inserted
into the hashtable.
Also, even if this was changed it would not with events other than
new (start time) and destroy (stop time).
Pablo suggested to add new CTA_TIMESTAMP_EVENT, this adds this feature.
The timestamp is recorded in case both events are requested and the
sys/net/netfilter/nf_conntrack_timestamp toggle is enabled.
Reported-by: Nadia Pinaeva <n.m.pinaeva@gmail.com>
Suggested-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
nf_tables_chain_destroy can sleep, it can't be used from call_rcu
callbacks.
Moreover, nf_tables_rule_release() is only safe for error unwinding,
while transaction mutex is held and the to-be-desroyed rule was not
exposed to either dataplane or dumps, as it deactives+frees without
the required synchronize_rcu() in-between.
nft_rule_expr_deactivate() callbacks will change ->use counters
of other chains/sets, see e.g. nft_lookup .deactivate callback, these
must be serialized via transaction mutex.
Also add a few lockdep asserts to make this more explicit.
Calling synchronize_rcu() isn't ideal, but fixing this without is hard
and way more intrusive. As-is, we can get:
WARNING: .. net/netfilter/nf_tables_api.c:5515 nft_set_destroy+0x..
Workqueue: events nf_tables_trans_destroy_work
RIP: 0010:nft_set_destroy+0x3fe/0x5c0
Call Trace:
<TASK>
nf_tables_trans_destroy_work+0x6b7/0xad0
process_one_work+0x64a/0xce0
worker_thread+0x613/0x10d0
In case the synchronize_rcu becomes an issue, we can explore alternatives.
One way would be to allocate nft_trans_rule objects + one nft_trans_chain
object, deactivate the rules + the chain and then defer the freeing to the
nft destroy workqueue. We'd still need to keep the synchronize_rcu path as
a fallback to handle -ENOMEM corner cases though.
Reported-by: syzbot+b26935466701e56cfdc2@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/67478d92.050a0220.253251.0062.GAE@google.com/T/
Fixes: c03d278fdf ("netfilter: nf_tables: wait for rcu grace period on net_device removal")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Softirq can interrupt ongoing packet from process context that is
walking over the percpu area that contains inner header offsets.
Disable bh and perform three checks before restoring the percpu inner
header offsets to validate that the percpu area is valid for this
skbuff:
1) If the NFT_PKTINFO_INNER_FULL flag is set on, then this skbuff
has already been parsed before for inner header fetching to
register.
2) Validate that the percpu area refers to this skbuff using the
skbuff pointer as a cookie. If there is a cookie mismatch, then
this skbuff needs to be parsed again.
3) Finally, validate if the percpu area refers to this tunnel type.
Only after these three checks the percpu area is restored to a on-stack
copy and bh is enabled again.
After inner header fetching, the on-stack copy is stored back to the
percpu area.
Fixes: 3a07327d10 ("netfilter: nft_inner: support for inner tunnel header matching")
Reported-by: syzbot+84d0441b9860f0d63285@syzkaller.appspotmail.com
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Move the timeout/expire/flag members from nft_trans_one_elem struct into
a dybamically allocated structure, only needed when timeout update was
requested.
This halves size of nft_trans_one_elem struct and allows to compact up to
124 elements in one transaction container rather than 62.
This halves memory requirements for a large flush or insert transaction,
where ->update remains NULL.
Care has to be taken to release the extra data in all spots, including
abort path.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Add helpers to release the individual elements contained in the
trans_elem container structure.
No functional change intended.
Followup patch will add 'nelems' member and will turn 'priv' into
a flexible array.
These helpers can then loop over all elements.
Care needs to be taken to handle a mix of new elements and existing
elements that are being updated (e.g. timeout refresh).
Before this patch, NEWSETELEM transaction with update is released
early so nft_trans_set_elem_destroy() won't get called, so we need
to skip elements marked as update.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
8c873e2199 ("netfilter: core: free hooks with call_rcu") removed
synchronize_net() call when unregistering basechain hook, however,
net_device removal event handler for the NFPROTO_NETDEV was not updated
to wait for RCU grace period.
Note that 835b803377 ("netfilter: nf_tables_netdev: unregister hooks
on net_device removal") does not remove basechain rules on device
removal, I was hinted to remove rules on net_device removal later, see
5ebe0b0eec ("netfilter: nf_tables: destroy basechain and rules on
netdevice removal").
Although NETDEV_UNREGISTER event is guaranteed to be handled after
synchronize_net() call, this path needs to wait for rcu grace period via
rcu callback to release basechain hooks if netns is alive because an
ongoing netlink dump could be in progress (sockets hold a reference on
the netns).
Note that nf_tables_pre_exit_net() unregisters and releases basechain
hooks but it is possible to see NETDEV_UNREGISTER at a later stage in
the netns exit path, eg. veth peer device in another netns:
cleanup_net()
default_device_exit_batch()
unregister_netdevice_many_notify()
notifier_call_chain()
nf_tables_netdev_event()
__nft_release_basechain()
In this particular case, same rule of thumb applies: if netns is alive,
then wait for rcu grace period because netlink dump in the other netns
could be in progress. Otherwise, if the other netns is going away then
no netlink dump can be in progress and basechain hooks can be released
inmediately.
While at it, turn WARN_ON() into WARN_ON_ONCE() for the basechain
validation, which should not ever happen.
Fixes: 835b803377 ("netfilter: nf_tables_netdev: unregister hooks on net_device removal")
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
The transaction mutex prevents concurrent add/delete, its ok to iterate
those lists outside of rcu read side critical sections.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
asm/unaligned.h is always an include of asm-generic/unaligned.h;
might as well move that thing to linux/unaligned.h and include
that - there's nothing arch-specific in that header.
auto-generated by the following:
for i in `git grep -l -w asm/unaligned.h`; do
sed -i -e "s/asm\/unaligned.h/linux\/unaligned.h/" $i
done
for i in `git grep -l -w asm-generic/unaligned.h`; do
sed -i -e "s/asm-generic\/unaligned.h/linux\/unaligned.h/" $i
done
git mv include/asm-generic/unaligned.h include/linux/unaligned.h
git mv tools/include/asm-generic/unaligned.h tools/include/linux/unaligned.h
sed -i -e "/unaligned.h/d" include/asm-generic/Kbuild
sed -i -e "s/__ASM_GENERIC/__LINUX/" include/linux/unaligned.h tools/include/linux/unaligned.h
Store new timeout and expiration in transaction object, use them to
update elements from .commit path. Otherwise, discard update if .abort
path is exercised.
Use update_flags in the transaction to note whether the timeout,
expiration, or both need to be updated.
Annotate access to timeout extension now that it can be updated while
lockless read access is possible.
Reject timeout updates on elements with no timeout extension.
Element transaction remains in the 96 bytes kmalloc slab on x86_64 after
this update.
This patch requires ("netfilter: nf_tables: use timestamp to check for
set element timeout") to make sure an element does not expire while
transaction is ongoing.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
This patch uses zero as timeout marker for those elements that never expire
when the element is created.
If userspace provides no timeout for an element, then the default set
timeout applies. However, if no default set timeout is specified and
timeout flag is set on, then timeout extension is allocated and timeout
is set to zero to allow for future updates.
Use of zero a never timeout marker has been suggested by Phil Sutter.
Note that, in older kernels, it is already possible to define elements
that never expire by declaring a set with the set timeout flag set on
and no global set timeout, in this case, new element with no explicit
timeout never expire do not allocate the timeout extension, hence, they
never expire. This approach makes it complicated to accomodate element
timeout update, because element extensions do not support reallocations.
Therefore, allocate the timeout extension and use the new marker for
this case, but do not expose it to userspace to retain backward
compatibility in the set listing.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Expiration and timeout are stored in separated set element extensions,
but they are tightly coupled. Consolidate them in a single extension to
simplify and prepare for set element updates.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
element expiration can be read-write locklessly, it can be written by
dynset and read from netlink dump, add annotation.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
- Add missing documentation of struct field and enum items.
- Add missing documentation of function parameter.
Flagged by ./scripts/kernel-doc -none.
No functional change intended.
Compile tested only.
Signed-off-by: Simon Horman <horms@kernel.org>
Reviewed-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
