Some code defines the IPv6 wildcard address as a local variable and
use it with memcmp() or ipv6_addr_equal().
Let's use in6addr_any and ipv6_addr_any() instead.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Paul Holzinger reported [0] that commit 5456262d2b ("net: Fix
incorrect address comparison when searching for a bind2 bucket")
introduced a bind() regression. Paul also gave a nice repro that
calls two types of bind() on the same port, both of which now
succeed, but the second call should fail:
bind(fd1, ::, port) + bind(fd2, 127.0.0.1, port)
The cited commit added address family tests in three functions to
fix the uninit-value KMSAN report. [1] However, the test added to
inet_bind2_bucket_match_addr_any() removed a necessary conflict
check; the dual-stack wildcard address no longer conflicts with
an IPv4 non-wildcard address.
If tb->family is AF_INET6 and sk->sk_family is AF_INET in
inet_bind2_bucket_match_addr_any(), we still need to check
if tb has the dual-stack wildcard address.
Note that the IPv4 wildcard address does not conflict with
IPv6 non-wildcard addresses.
[0]: https://lore.kernel.org/netdev/e21bf153-80b0-9ec0-15ba-e04a4ad42c34@redhat.com/
[1]: https://lore.kernel.org/netdev/CAG_fn=Ud3zSW7AZWXc+asfMhZVL5ETnvuY44Pmyv4NPv-ijN-A@mail.gmail.com/
Fixes: 5456262d2b ("net: Fix incorrect address comparison when searching for a bind2 bucket")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reported-by: Paul Holzinger <pholzing@redhat.com>
Link: https://lore.kernel.org/netdev/CAG_fn=Ud3zSW7AZWXc+asfMhZVL5ETnvuY44Pmyv4NPv-ijN-A@mail.gmail.com/
Reviewed-by: Eric Dumazet <edumazet@google.com>
Tested-by: Paul Holzinger <pholzing@redhat.com>
Reviewed-by: Martin KaFai Lau <martin.lau@kernel.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Users who want to share a single public IP address for outgoing connections
between several hosts traditionally reach for SNAT. However, SNAT requires
state keeping on the node(s) performing the NAT.
A stateless alternative exists, where a single IP address used for egress
can be shared between several hosts by partitioning the available ephemeral
port range. In such a setup:
1. Each host gets assigned a disjoint range of ephemeral ports.
2. Applications open connections from the host-assigned port range.
3. Return traffic gets routed to the host based on both, the destination IP
and the destination port.
An application which wants to open an outgoing connection (connect) from a
given port range today can choose between two solutions:
1. Manually pick the source port by bind()'ing to it before connect()'ing
the socket.
This approach has a couple of downsides:
a) Search for a free port has to be implemented in the user-space. If
the chosen 4-tuple happens to be busy, the application needs to retry
from a different local port number.
Detecting if 4-tuple is busy can be either easy (TCP) or hard
(UDP). In TCP case, the application simply has to check if connect()
returned an error (EADDRNOTAVAIL). That is assuming that the local
port sharing was enabled (REUSEADDR) by all the sockets.
# Assume desired local port range is 60_000-60_511
s = socket(AF_INET, SOCK_STREAM)
s.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
s.bind(("192.0.2.1", 60_000))
s.connect(("1.1.1.1", 53))
# Fails only if 192.0.2.1:60000 -> 1.1.1.1:53 is busy
# Application must retry with another local port
In case of UDP, the network stack allows binding more than one socket
to the same 4-tuple, when local port sharing is enabled
(REUSEADDR). Hence detecting the conflict is much harder and involves
querying sock_diag and toggling the REUSEADDR flag [1].
b) For TCP, bind()-ing to a port within the ephemeral port range means
that no connecting sockets, that is those which leave it to the
network stack to find a free local port at connect() time, can use
the this port.
IOW, the bind hash bucket tb->fastreuse will be 0 or 1, and the port
will be skipped during the free port search at connect() time.
2. Isolate the app in a dedicated netns and use the use the per-netns
ip_local_port_range sysctl to adjust the ephemeral port range bounds.
The per-netns setting affects all sockets, so this approach can be used
only if:
- there is just one egress IP address, or
- the desired egress port range is the same for all egress IP addresses
used by the application.
For TCP, this approach avoids the downsides of (1). Free port search and
4-tuple conflict detection is done by the network stack:
system("sysctl -w net.ipv4.ip_local_port_range='60000 60511'")
s = socket(AF_INET, SOCK_STREAM)
s.setsockopt(SOL_IP, IP_BIND_ADDRESS_NO_PORT, 1)
s.bind(("192.0.2.1", 0))
s.connect(("1.1.1.1", 53))
# Fails if all 4-tuples 192.0.2.1:60000-60511 -> 1.1.1.1:53 are busy
For UDP this approach has limited applicability. Setting the
IP_BIND_ADDRESS_NO_PORT socket option does not result in local source
port being shared with other connected UDP sockets.
Hence relying on the network stack to find a free source port, limits the
number of outgoing UDP flows from a single IP address down to the number
of available ephemeral ports.
To put it another way, partitioning the ephemeral port range between hosts
using the existing Linux networking API is cumbersome.
To address this use case, add a new socket option at the SOL_IP level,
named IP_LOCAL_PORT_RANGE. The new option can be used to clamp down the
ephemeral port range for each socket individually.
The option can be used only to narrow down the per-netns local port
range. If the per-socket range lies outside of the per-netns range, the
latter takes precedence.
UAPI-wise, the low and high range bounds are passed to the kernel as a pair
of u16 values in host byte order packed into a u32. This avoids pointer
passing.
PORT_LO = 40_000
PORT_HI = 40_511
s = socket(AF_INET, SOCK_STREAM)
v = struct.pack("I", PORT_HI << 16 | PORT_LO)
s.setsockopt(SOL_IP, IP_LOCAL_PORT_RANGE, v)
s.bind(("127.0.0.1", 0))
s.getsockname()
# Local address between ("127.0.0.1", 40_000) and ("127.0.0.1", 40_511),
# if there is a free port. EADDRINUSE otherwise.
[1] https://github.com/cloudflare/cloudflare-blog/blob/232b432c1d57/2022-02-connectx/connectx.py#L116
Reviewed-by: Marek Majkowski <marek@cloudflare.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
While one cpu is working on looking up the right socket from ehash
table, another cpu is done deleting the request socket and is about
to add (or is adding) the big socket from the table. It means that
we could miss both of them, even though it has little chance.
Let me draw a call trace map of the server side.
CPU 0 CPU 1
----- -----
tcp_v4_rcv() syn_recv_sock()
inet_ehash_insert()
-> sk_nulls_del_node_init_rcu(osk)
__inet_lookup_established()
-> __sk_nulls_add_node_rcu(sk, list)
Notice that the CPU 0 is receiving the data after the final ack
during 3-way shakehands and CPU 1 is still handling the final ack.
Why could this be a real problem?
This case is happening only when the final ack and the first data
receiving by different CPUs. Then the server receiving data with
ACK flag tries to search one proper established socket from ehash
table, but apparently it fails as my map shows above. After that,
the server fetches a listener socket and then sends a RST because
it finds a ACK flag in the skb (data), which obeys RST definition
in RFC 793.
Besides, Eric pointed out there's one more race condition where it
handles tw socket hashdance. Only by adding to the tail of the list
before deleting the old one can we avoid the race if the reader has
already begun the bucket traversal and it would possibly miss the head.
Many thanks to Eric for great help from beginning to end.
Fixes: 5e0724d027 ("tcp/dccp: fix hashdance race for passive sessions")
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jason Xing <kernelxing@tencent.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/lkml/20230112065336.41034-1-kerneljasonxing@gmail.com/
Link: https://lore.kernel.org/r/20230118015941.1313-1-kerneljasonxing@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
__inet_hash_connect() has a fast path taken if sk_head(&tb->owners) is
equal to the sk parameter.
sk_head() returns the hlist_entry() with respect to the sk_node field.
However entries in the tb->owners list are inserted with respect to the
sk_bind_node field with sk_add_bind_node().
Thus the check would never pass and the fast path never execute.
This fast path has never been executed or tested as this bug seems
to be present since commit 1da177e4c3 ("Linux-2.6.12-rc2"), thus
remove it to reduce code complexity.
Signed-off-by: Pietro Borrello <borrello@diag.uniroma1.it>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230112-inet_hash_connect_bind_head-v3-1-b591fd212b93@diag.uniroma1.it
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Jiri Slaby reported regression of bind() with a simple repro. [0]
The repro creates a TIME_WAIT socket and tries to bind() a new socket
with the same local address and port. Before commit 28044fc1d4 ("net:
Add a bhash2 table hashed by port and address"), the bind() failed with
-EADDRINUSE, but now it succeeds.
The cited commit should have put TIME_WAIT sockets into bhash2; otherwise,
inet_bhash2_conflict() misses TIME_WAIT sockets when validating bind()
requests if the address is not a wildcard one.
The straight option is to move sk_bind2_node from struct sock to struct
sock_common to add twsk to bhash2 as implemented as RFC. [1] However, the
binary layout change in the struct sock could affect performances moving
hot fields on different cachelines.
To avoid that, we add another TIME_WAIT list in inet_bind2_bucket and check
it while validating bind().
[0]: https://lore.kernel.org/netdev/6b971a4e-c7d8-411e-1f92-fda29b5b2fb9@kernel.org/
[1]: https://lore.kernel.org/netdev/20221221151258.25748-2-kuniyu@amazon.com/
Fixes: 28044fc1d4 ("net: Add a bhash2 table hashed by port and address")
Reported-by: Jiri Slaby <jirislaby@kernel.org>
Suggested-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Joanne Koong <joannelkoong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
-----BEGIN PGP SIGNATURE-----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=QRhK
-----END PGP SIGNATURE-----
Merge tag 'random-6.2-rc1-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/crng/random
Pull random number generator updates from Jason Donenfeld:
- Replace prandom_u32_max() and various open-coded variants of it,
there is now a new family of functions that uses fast rejection
sampling to choose properly uniformly random numbers within an
interval:
get_random_u32_below(ceil) - [0, ceil)
get_random_u32_above(floor) - (floor, U32_MAX]
get_random_u32_inclusive(floor, ceil) - [floor, ceil]
Coccinelle was used to convert all current users of
prandom_u32_max(), as well as many open-coded patterns, resulting in
improvements throughout the tree.
I'll have a "late" 6.1-rc1 pull for you that removes the now unused
prandom_u32_max() function, just in case any other trees add a new
use case of it that needs to converted. According to linux-next,
there may be two trivial cases of prandom_u32_max() reintroductions
that are fixable with a 's/.../.../'. So I'll have for you a final
conversion patch doing that alongside the removal patch during the
second week.
This is a treewide change that touches many files throughout.
- More consistent use of get_random_canary().
- Updates to comments, documentation, tests, headers, and
simplification in configuration.
- The arch_get_random*_early() abstraction was only used by arm64 and
wasn't entirely useful, so this has been replaced by code that works
in all relevant contexts.
- The kernel will use and manage random seeds in non-volatile EFI
variables, refreshing a variable with a fresh seed when the RNG is
initialized. The RNG GUID namespace is then hidden from efivarfs to
prevent accidental leakage.
These changes are split into random.c infrastructure code used in the
EFI subsystem, in this pull request, and related support inside of
EFISTUB, in Ard's EFI tree. These are co-dependent for full
functionality, but the order of merging doesn't matter.
- Part of the infrastructure added for the EFI support is also used for
an improvement to the way vsprintf initializes its siphash key,
replacing an sleep loop wart.
- The hardware RNG framework now always calls its correct random.c
input function, add_hwgenerator_randomness(), rather than sometimes
going through helpers better suited for other cases.
- The add_latent_entropy() function has long been called from the fork
handler, but is a no-op when the latent entropy gcc plugin isn't
used, which is fine for the purposes of latent entropy.
But it was missing out on the cycle counter that was also being mixed
in beside the latent entropy variable. So now, if the latent entropy
gcc plugin isn't enabled, add_latent_entropy() will expand to a call
to add_device_randomness(NULL, 0), which adds a cycle counter,
without the absent latent entropy variable.
- The RNG is now reseeded from a delayed worker, rather than on demand
when used. Always running from a worker allows it to make use of the
CPU RNG on platforms like S390x, whose instructions are too slow to
do so from interrupts. It also has the effect of adding in new inputs
more frequently with more regularity, amounting to a long term
transcript of random values. Plus, it helps a bit with the upcoming
vDSO implementation (which isn't yet ready for 6.2).
- The jitter entropy algorithm now tries to execute on many different
CPUs, round-robining, in hopes of hitting even more memory latencies
and other unpredictable effects. It also will mix in a cycle counter
when the entropy timer fires, in addition to being mixed in from the
main loop, to account more explicitly for fluctuations in that timer
firing. And the state it touches is now kept within the same cache
line, so that it's assured that the different execution contexts will
cause latencies.
* tag 'random-6.2-rc1-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/crng/random: (23 commits)
random: include <linux/once.h> in the right header
random: align entropy_timer_state to cache line
random: mix in cycle counter when jitter timer fires
random: spread out jitter callback to different CPUs
random: remove extraneous period and add a missing one in comments
efi: random: refresh non-volatile random seed when RNG is initialized
vsprintf: initialize siphash key using notifier
random: add back async readiness notifier
random: reseed in delayed work rather than on-demand
random: always mix cycle counter in add_latent_entropy()
hw_random: use add_hwgenerator_randomness() for early entropy
random: modernize documentation comment on get_random_bytes()
random: adjust comment to account for removed function
random: remove early archrandom abstraction
random: use random.trust_{bootloader,cpu} command line option only
stackprotector: actually use get_random_canary()
stackprotector: move get_random_canary() into stackprotector.h
treewide: use get_random_u32_inclusive() when possible
treewide: use get_random_u32_{above,below}() instead of manual loop
treewide: use get_random_u32_below() instead of deprecated function
...
When we call connect() for a socket bound to a wildcard address, we update
saddr locklessly. However, it could result in a data race; another thread
iterating over bhash might see a corrupted address.
Let's update saddr under the bhash bucket's lock.
Fixes: 3df80d9320 ("[DCCP]: Introduce DCCPv6")
Fixes: 7c657876b6 ("[DCCP]: Initial implementation")
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Joanne Koong <joannelkoong@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When we call inet_bhash2_update_saddr(), prev_saddr is always non-NULL.
Let's remove the unnecessary test.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Joanne Koong <joannelkoong@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This is a simple mechanical transformation done by:
@@
expression E;
@@
- prandom_u32_max
+ get_random_u32_below
(E)
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Darrick J. Wong <djwong@kernel.org> # for xfs
Reviewed-by: SeongJae Park <sj@kernel.org> # for damon
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com> # for infiniband
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk> # for arm
Acked-by: Ulf Hansson <ulf.hansson@linaro.org> # for mmc
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
On embedded systems with little memory and no relevant
security concerns, it is beneficial to reduce the size
of the table.
Reducing the size from 2^16 to 2^8 saves 255 KiB
of kernel RAM.
Makes the table size configurable as an expert option.
The size was previously increased from 2^8 to 2^16
in commit 4c2c8f03a5 ("tcp: increase source port perturb table to
2^16").
Signed-off-by: Gleb Mazovetskiy <glex.spb@gmail.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rather than incurring a division or requesting too many random bytes for
the given range, use the prandom_u32_max() function, which only takes
the minimum required bytes from the RNG and avoids divisions. This was
done mechanically with this coccinelle script:
@basic@
expression E;
type T;
identifier get_random_u32 =~ "get_random_int|prandom_u32|get_random_u32";
typedef u64;
@@
(
- ((T)get_random_u32() % (E))
+ prandom_u32_max(E)
|
- ((T)get_random_u32() & ((E) - 1))
+ prandom_u32_max(E * XXX_MAKE_SURE_E_IS_POW2)
|
- ((u64)(E) * get_random_u32() >> 32)
+ prandom_u32_max(E)
|
- ((T)get_random_u32() & ~PAGE_MASK)
+ prandom_u32_max(PAGE_SIZE)
)
@multi_line@
identifier get_random_u32 =~ "get_random_int|prandom_u32|get_random_u32";
identifier RAND;
expression E;
@@
- RAND = get_random_u32();
... when != RAND
- RAND %= (E);
+ RAND = prandom_u32_max(E);
// Find a potential literal
@literal_mask@
expression LITERAL;
type T;
identifier get_random_u32 =~ "get_random_int|prandom_u32|get_random_u32";
position p;
@@
((T)get_random_u32()@p & (LITERAL))
// Add one to the literal.
@script:python add_one@
literal << literal_mask.LITERAL;
RESULT;
@@
value = None
if literal.startswith('0x'):
value = int(literal, 16)
elif literal[0] in '123456789':
value = int(literal, 10)
if value is None:
print("I don't know how to handle %s" % (literal))
cocci.include_match(False)
elif value == 2**32 - 1 or value == 2**31 - 1 or value == 2**24 - 1 or value == 2**16 - 1 or value == 2**8 - 1:
print("Skipping 0x%x for cleanup elsewhere" % (value))
cocci.include_match(False)
elif value & (value + 1) != 0:
print("Skipping 0x%x because it's not a power of two minus one" % (value))
cocci.include_match(False)
elif literal.startswith('0x'):
coccinelle.RESULT = cocci.make_expr("0x%x" % (value + 1))
else:
coccinelle.RESULT = cocci.make_expr("%d" % (value + 1))
// Replace the literal mask with the calculated result.
@plus_one@
expression literal_mask.LITERAL;
position literal_mask.p;
expression add_one.RESULT;
identifier FUNC;
@@
- (FUNC()@p & (LITERAL))
+ prandom_u32_max(RESULT)
@collapse_ret@
type T;
identifier VAR;
expression E;
@@
{
- T VAR;
- VAR = (E);
- return VAR;
+ return E;
}
@drop_var@
type T;
identifier VAR;
@@
{
- T VAR;
... when != VAR
}
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Yury Norov <yury.norov@gmail.com>
Reviewed-by: KP Singh <kpsingh@kernel.org>
Reviewed-by: Jan Kara <jack@suse.cz> # for ext4 and sbitmap
Reviewed-by: Christoph Böhmwalder <christoph.boehmwalder@linbit.com> # for drbd
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Heiko Carstens <hca@linux.ibm.com> # for s390
Acked-by: Ulf Hansson <ulf.hansson@linaro.org> # for mmc
Acked-by: Darrick J. Wong <djwong@kernel.org> # for xfs
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
The _SLOW designation wasn't really descriptive of anything. This is
meant to be called from process context when it's possible to sleep. So
name this more aptly _SLEEPABLE, which better fits its intended use.
Fixes: 62c07983be ("once: add DO_ONCE_SLOW() for sleepable contexts")
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20221003181413.1221968-1-Jason@zx2c4.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Christophe Leroy reported a ~80ms latency spike
happening at first TCP connect() time.
This is because __inet_hash_connect() uses get_random_once()
to populate a perturbation table which became quite big
after commit 4c2c8f03a5 ("tcp: increase source port perturb table to 2^16")
get_random_once() uses DO_ONCE(), which block hard irqs for the duration
of the operation.
This patch adds DO_ONCE_SLOW() which uses a mutex instead of a spinlock
for operations where we prefer to stay in process context.
Then __inet_hash_connect() can use get_random_slow_once()
to populate its perturbation table.
Fixes: 4c2c8f03a5 ("tcp: increase source port perturb table to 2^16")
Fixes: 190cc82489 ("tcp: change source port randomizarion at connect() time")
Reported-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Link: https://lore.kernel.org/netdev/CANn89iLAEYBaoYajy0Y9UmGFff5GPxDUoG-ErVB2jDdRNQ5Tug@mail.gmail.com/T/#t
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Willy Tarreau <w@1wt.eu>
Tested-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>
The v6_rcv_saddr and rcv_saddr are inside a union in the
'struct inet_bind2_bucket'. When searching a bucket by following the
bhash2 hashtable chain, eg. inet_bind2_bucket_match, it is only using
the sk->sk_family and there is no way to check if the inet_bind2_bucket
has a v6 or v4 address in the union. This leads to an uninit-value
KMSAN report in [0] and also potentially incorrect matches.
This patch fixes it by adding a family member to the inet_bind2_bucket
and then tests 'sk->sk_family != tb->family' before matching
the sk's address to the tb's address.
Cc: Joanne Koong <joannelkoong@gmail.com>
Fixes: 28044fc1d4 ("net: Add a bhash2 table hashed by port and address")
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Tested-by: Alexander Potapenko <glider@google.com>
Link: https://lore.kernel.org/r/20220927002544.3381205-1-kafai@fb.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The more sockets we have in the hash table, the longer we spend looking
up the socket. While running a number of small workloads on the same
host, they penalise each other and cause performance degradation.
The root cause might be a single workload that consumes much more
resources than the others. It often happens on a cloud service where
different workloads share the same computing resource.
On EC2 c5.24xlarge instance (196 GiB memory and 524288 (1Mi / 2) ehash
entries), after running iperf3 in different netns, creating 24Mi sockets
without data transfer in the root netns causes about 10% performance
regression for the iperf3's connection.
thash_entries sockets length Gbps
524288 1 1 50.7
24Mi 48 45.1
It is basically related to the length of the list of each hash bucket.
For testing purposes to see how performance drops along the length,
I set 131072 (1Mi / 8) to thash_entries, and here's the result.
thash_entries sockets length Gbps
131072 1 1 50.7
1Mi 8 49.9
2Mi 16 48.9
4Mi 32 47.3
8Mi 64 44.6
16Mi 128 40.6
24Mi 192 36.3
32Mi 256 32.5
40Mi 320 27.0
48Mi 384 25.0
To resolve the socket lookup degradation, we introduce an optional
per-netns hash table for TCP, but it's just ehash, and we still share
the global bhash, bhash2 and lhash2.
With a smaller ehash, we can look up non-listener sockets faster and
isolate such noisy neighbours. In addition, we can reduce lock contention.
We can control the ehash size by a new sysctl knob. However, depending
on workloads, it will require very sensitive tuning, so we disable the
feature by default (net.ipv4.tcp_child_ehash_entries == 0). Moreover,
we can fall back to using the global ehash in case we fail to allocate
enough memory for a new ehash. The maximum size is 16Mi, which is large
enough that even if we have 48Mi sockets, the average list length is 3,
and regression would be less than 1%.
We can check the current ehash size by another read-only sysctl knob,
net.ipv4.tcp_ehash_entries. A negative value means the netns shares
the global ehash (per-netns ehash is disabled or failed to allocate
memory).
# dmesg | cut -d ' ' -f 5- | grep "established hash"
TCP established hash table entries: 524288 (order: 10, 4194304 bytes, vmalloc hugepage)
# sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = 524288 # can be changed by thash_entries
# sysctl net.ipv4.tcp_child_ehash_entries
net.ipv4.tcp_child_ehash_entries = 0 # disabled by default
# ip netns add test1
# ip netns exec test1 sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = -524288 # share the global ehash
# sysctl -w net.ipv4.tcp_child_ehash_entries=100
net.ipv4.tcp_child_ehash_entries = 100
# ip netns add test2
# ip netns exec test2 sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = 128 # own a per-netns ehash with 2^n buckets
When more than two processes in the same netns create per-netns ehash
concurrently with different sizes, we need to guarantee the size in
one of the following ways:
1) Share the global ehash and create per-netns ehash
First, unshare() with tcp_child_ehash_entries==0. It creates dedicated
netns sysctl knobs where we can safely change tcp_child_ehash_entries
and clone()/unshare() to create a per-netns ehash.
2) Control write on sysctl by BPF
We can use BPF_PROG_TYPE_CGROUP_SYSCTL to allow/deny read/write on
sysctl knobs.
Note that the global ehash allocated at the boot time is spread over
available NUMA nodes, but inet_pernet_hashinfo_alloc() will allocate
pages for each per-netns ehash depending on the current process's NUMA
policy. By default, the allocation is done in the local node only, so
the per-netns hash table could fully reside on a random node. Thus,
depending on the NUMA policy the netns is created with and the CPU the
current thread is running on, we could see some performance differences
for highly optimised networking applications.
Note also that the default values of two sysctl knobs depend on the ehash
size and should be tuned carefully:
tcp_max_tw_buckets : tcp_child_ehash_entries / 2
tcp_max_syn_backlog : max(128, tcp_child_ehash_entries / 128)
As a bonus, we can dismantle netns faster. Currently, while destroying
netns, we call inet_twsk_purge(), which walks through the global ehash.
It can be potentially big because it can have many sockets other than
TIME_WAIT in all netns. Splitting ehash changes that situation, where
it's only necessary for inet_twsk_purge() to clean up TIME_WAIT sockets
in each netns.
With regard to this, we do not free the per-netns ehash in inet_twsk_kill()
to avoid UAF while iterating the per-netns ehash in inet_twsk_purge().
Instead, we do it in tcp_sk_exit_batch() after calling tcp_twsk_purge() to
keep it protocol-family-independent.
In the future, we could optimise ehash lookup/iteration further by removing
netns comparison for the per-netns ehash.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will soon introduce an optional per-netns ehash.
This means we cannot use tcp_hashinfo directly in most places.
Instead, access it via net->ipv4.tcp_death_row.hashinfo.
The access will be valid only while initialising tcp_hashinfo
itself and creating/destroying each netns.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will soon introduce an optional per-netns ehash.
This means we cannot use the global sk->sk_prot->h.hashinfo
to fetch a TCP hashinfo.
Instead, set NULL to sk->sk_prot->h.hashinfo for TCP and get
a proper hashinfo from net->ipv4.tcp_death_row.hashinfo.
Note that we need not use sk->sk_prot->h.hashinfo if DCCP is
disabled.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch adds no functional change and cleans up some functions
that the following patches touch around so that we make them tidy
and easy to review/revert. The changes are
- Keep reverse christmas tree order
- Remove unnecessary init of port in inet_csk_find_open_port()
- Use req_to_sk() once in reqsk_queue_unlink()
- Use sock_net(sk) once in tcp_time_wait() and tcp_v[46]_connect()
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The current bind hashtable (bhash) is hashed by port only.
In the socket bind path, we have to check for bind conflicts by
traversing the specified port's inet_bind_bucket while holding the
hashbucket's spinlock (see inet_csk_get_port() and
inet_csk_bind_conflict()). In instances where there are tons of
sockets hashed to the same port at different addresses, the bind
conflict check is time-intensive and can cause softirq cpu lockups,
as well as stops new tcp connections since __inet_inherit_port()
also contests for the spinlock.
This patch adds a second bind table, bhash2, that hashes by
port and sk->sk_rcv_saddr (ipv4) and sk->sk_v6_rcv_saddr (ipv6).
Searching the bhash2 table leads to significantly faster conflict
resolution and less time holding the hashbucket spinlock.
Please note a few things:
* There can be the case where the a socket's address changes after it
has been bound. There are two cases where this happens:
1) The case where there is a bind() call on INADDR_ANY (ipv4) or
IPV6_ADDR_ANY (ipv6) and then a connect() call. The kernel will
assign the socket an address when it handles the connect()
2) In inet_sk_reselect_saddr(), which is called when rebuilding the
sk header and a few pre-conditions are met (eg rerouting fails).
In these two cases, we need to update the bhash2 table by removing the
entry for the old address, and add a new entry reflecting the updated
address.
* The bhash2 table must have its own lock, even though concurrent
accesses on the same port are protected by the bhash lock. Bhash2 must
have its own lock to protect against cases where sockets on different
ports hash to different bhash hashbuckets but to the same bhash2
hashbucket.
This brings up a few stipulations:
1) When acquiring both the bhash and the bhash2 lock, the bhash2 lock
will always be acquired after the bhash lock and released before the
bhash lock is released.
2) There are no nested bhash2 hashbucket locks. A bhash2 lock is always
acquired+released before another bhash2 lock is acquired+released.
* The bhash table cannot be superseded by the bhash2 table because for
bind requests on INADDR_ANY (ipv4) or IPV6_ADDR_ANY (ipv6), every socket
bound to that port must be checked for a potential conflict. The bhash
table is the only source of port->socket associations.
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In our server, there may be no high order (>= 6) memory since we reserve
lots of HugeTLB pages when booting. Then the system panic. So use
alloc_large_system_hash() to allocate table_perturb.
Fixes: e926147618 ("tcp: dynamically allocate the perturb table used by source ports")
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20220607070214.94443-1-songmuchun@bytedance.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We currently have one tcp bind table (bhash) which hashes by port
number only. In the socket bind path, we check for bind conflicts by
traversing the specified port's inet_bind2_bucket while holding the
bucket's spinlock (see inet_csk_get_port() and inet_csk_bind_conflict()).
In instances where there are tons of sockets hashed to the same port
at different addresses, checking for a bind conflict is time-intensive
and can cause softirq cpu lockups, as well as stops new tcp connections
since __inet_inherit_port() also contests for the spinlock.
This patch proposes adding a second bind table, bhash2, that hashes by
port and ip address. Searching the bhash2 table leads to significantly
faster conflict resolution and less time holding the spinlock.
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This is no longer a macro, but an inlined function.
INET_MATCH() -> inet_match()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Suggested-by: Olivier Hartkopp <socketcan@hartkopp.net>
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
INET6_MATCH() runs without holding a lock on the socket.
We probably need to annotate most reads.
This patch makes INET6_MATCH() an inline function
to ease our changes.
v2: inline function only defined if IS_ENABLED(CONFIG_IPV6)
Change the name to inet6_match(), this is no longer a macro.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
INET_MATCH() runs without holding a lock on the socket.
We probably need to annotate most reads.
This patch makes INET_MATCH() an inline function
to ease our changes.
v2:
We remove the 32bit version of it, as modern compilers
should generate the same code really, no need to
try to be smarter.
Also make 'struct net *net' the first argument.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The listen sk is currently stored in two hash tables,
listening_hash (hashed by port) and lhash2 (hashed by port and address).
After commit 0ee58dad5b ("net: tcp6: prefer listeners bound to an address")
and commit d9fbc7f643 ("net: tcp: prefer listeners bound to an address"),
the TCP-SYN lookup fast path does not use listening_hash.
The commit 05c0b35709 ("tcp: seq_file: Replace listening_hash with lhash2")
also moved the seq_file (/proc/net/tcp) iteration usage from
listening_hash to lhash2.
There are still a few listening_hash usages left.
One of them is inet_reuseport_add_sock() which uses the listening_hash
to search a listen sk during the listen() system call. This turns
out to be very slow on use cases that listen on many different
VIPs at a popular port (e.g. 443). [ On top of the slowness in
adding to the tail in the IPv6 case ]. The latter patch has a
selftest to demonstrate this case.
This patch takes this chance to move all remaining listening_hash
usages to lhash2 and then retire listening_hash.
Since most changes need to be done together, it is hard to cut
the listening_hash to lhash2 switch into small patches. The
changes in this patch is highlighted here for the review
purpose.
1. Because of the listening_hash removal, lhash2 can use the
sk->sk_nulls_node instead of the icsk->icsk_listen_portaddr_node.
This will also keep the sk_unhashed() check to work as is
after stop adding sk to listening_hash.
The union is removed from inet_listen_hashbucket because
only nulls_head is needed.
2. icsk->icsk_listen_portaddr_node and its helpers are removed.
3. The current lhash2 users needs to iterate with sk_nulls_node
instead of icsk_listen_portaddr_node.
One case is in the inet[6]_lhash2_lookup().
Another case is the seq_file iterator in tcp_ipv4.c.
One thing to note is sk_nulls_next() is needed
because the old inet_lhash2_for_each_icsk_continue()
does a "next" first before iterating.
4. Move the remaining listening_hash usage to lhash2
inet_reuseport_add_sock() which this series is
trying to improve.
inet_diag.c and mptcp_diag.c are the final two
remaining use cases and is moved to lhash2 now also.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch folds lhash2 related functions into __inet_hash and
inet_unhash. This will make the removal of the listening_hash
in a latter patch easier to review.
First, this patch folds inet_hash2 into __inet_hash.
For unhash, the current call sequence is like
inet_unhash() => __inet_unhash() => inet_unhash2().
The specific testing cases in __inet_unhash() are mostly related
to TCP_LISTEN sk and its caller inet_unhash() already has
the TCP_LISTEN test, so this patch folds both __inet_unhash() and
inet_unhash2() into inet_unhash().
Note that all listening_hash users also have lhash2 initialized,
so the !h->lhash2 check is no longer needed.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
After commit 0ee58dad5b ("net: tcp6: prefer listeners bound to an address")
and commit d9fbc7f643 ("net: tcp: prefer listeners bound to an address"),
the count is no longer used. This patch removes it.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In commit 190cc82489 ("tcp: change source port randomizarion at
connect() time"), the table_perturb[] array was introduced and an
index was taken from the port_offset via hash_32(). But it turns
out that hash_32() performs a multiplication while the input here
comes from the output of SipHash in secure_seq, that is well
distributed enough to avoid the need for yet another hash.
Suggested-by: Amit Klein <aksecurity@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Moshe Kol, Amit Klein, and Yossi Gilad reported being able to accurately
identify a client by forcing it to emit only 40 times more connections
than there are entries in the table_perturb[] table. The previous two
improvements consisting in resalting the secret every 10s and adding
randomness to each port selection only slightly improved the situation,
and the current value of 2^8 was too small as it's not very difficult
to make a client emit 10k connections in less than 10 seconds.
Thus we're increasing the perturb table from 2^8 to 2^16 so that the
same precision now requires 2.6M connections, which is more difficult in
this time frame and harder to hide as a background activity. The impact
is that the table now uses 256 kB instead of 1 kB, which could mostly
affect devices making frequent outgoing connections. However such
components usually target a small set of destinations (load balancers,
database clients, perf assessment tools), and in practice only a few
entries will be visited, like before.
A live test at 1 million connections per second showed no performance
difference from the previous value.
Reported-by: Moshe Kol <moshe.kol@mail.huji.ac.il>
Reported-by: Yossi Gilad <yossi.gilad@mail.huji.ac.il>
Reported-by: Amit Klein <aksecurity@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We'll need to further increase the size of this table and it's likely
that at some point its size will not be suitable anymore for a static
table. Let's allocate it on boot from inet_hashinfo2_init(), which is
called from tcp_init().
Cc: Moshe Kol <moshe.kol@mail.huji.ac.il>
Cc: Yossi Gilad <yossi.gilad@mail.huji.ac.il>
Cc: Amit Klein <aksecurity@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Here we're randomly adding between 0 and 7 random increments to the
selected source port in order to add some noise in the source port
selection that will make the next port less predictable.
With the default port range of 32768-60999 this means a worst case
reuse scenario of 14116/8=1764 connections between two consecutive
uses of the same port, with an average of 14116/4.5=3137. This code
was stressed at more than 800000 connections per second to a fixed
target with all connections closed by the client using RSTs (worst
condition) and only 2 connections failed among 13 billion, despite
the hash being reseeded every 10 seconds, indicating a perfectly
safe situation.
Cc: Moshe Kol <moshe.kol@mail.huji.ac.il>
Cc: Yossi Gilad <yossi.gilad@mail.huji.ac.il>
Cc: Amit Klein <aksecurity@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Amit Klein suggests that we use different parts of port_offset for the
table's index and the port offset so that there is no direct relation
between them.
Cc: Jason A. Donenfeld <Jason@zx2c4.com>
Cc: Moshe Kol <moshe.kol@mail.huji.ac.il>
Cc: Yossi Gilad <yossi.gilad@mail.huji.ac.il>
Cc: Amit Klein <aksecurity@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
SipHash replaced MD5 in secure_ipv{4,6}_port_ephemeral() via commit
7cd23e5300 ("secure_seq: use SipHash in place of MD5"), but the output
remained truncated to 32-bit only. In order to exploit more bits from the
hash, let's make the functions return the full 64-bit of siphash_3u32().
We also make sure the port offset calculation in __inet_hash_connect()
remains done on 32-bit to avoid the need for div_u64_rem() and an extra
cost on 32-bit systems.
Cc: Jason A. Donenfeld <Jason@zx2c4.com>
Cc: Moshe Kol <moshe.kol@mail.huji.ac.il>
Cc: Yossi Gilad <yossi.gilad@mail.huji.ac.il>
Cc: Amit Klein <aksecurity@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit
9652dc2eb9 ("tcp: relax listening_hash operations")
removed the need to disable bottom half while acquiring
listening_hash.lock. There are still two callers left which disable
bottom half before the lock is acquired.
On PREEMPT_RT the softirqs are preemptible and local_bh_disable() acts
as a lock to ensure that resources, that are protected by disabling
bottom halves, remain protected.
This leads to a circular locking dependency if the lock acquired with
disabled bottom halves is also acquired with enabled bottom halves
followed by disabling bottom halves. This is the reverse locking order.
It has been observed with inet_listen_hashbucket:🔒
local_bh_disable() + spin_lock(&ilb->lock):
inet_listen()
inet_csk_listen_start()
sk->sk_prot->hash() := inet_hash()
local_bh_disable()
__inet_hash()
spin_lock(&ilb->lock);
acquire(&ilb->lock);
Reverse order: spin_lock(&ilb2->lock) + local_bh_disable():
tcp_seq_next()
listening_get_next()
spin_lock(&ilb2->lock);
acquire(&ilb2->lock);
tcp4_seq_show()
get_tcp4_sock()
sock_i_ino()
read_lock_bh(&sk->sk_callback_lock);
acquire(softirq_ctrl) // <---- whoops
acquire(&sk->sk_callback_lock)
Drop local_bh_disable() around __inet_hash() which acquires
listening_hash->lock. Split inet_unhash() and acquire the
listen_hashbucket lock without disabling bottom halves; the inet_ehash
lock with disabled bottom halves.
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://lkml.kernel.org/r/12d6f9879a97cd56c09fb53dee343cbb14f7f1f7.camel@gmx.de
Link: https://lkml.kernel.org/r/X9CheYjuXWc75Spa@hirez.programming.kicks-ass.net
Link: https://lore.kernel.org/r/YgQOebeZ10eNx1W6@linutronix.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
It may be helpful to have access to the ifindex during bpf socket
lookup. An example may be to scope certain socket lookup logic to
specific interfaces, i.e. an interface may be made exempt from custom
lookup code.
Add the ifindex of the arriving connection to the bpf_sk_lookup API.
Signed-off-by: Mark Pashmfouroush <markpash@cloudflare.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20211110111016.5670-2-markpash@cloudflare.com
Use of percpu_counter structure to track count of orphaned
sockets is causing problems on modern hosts with 256 cpus
or more.
Stefan Bach reported a serious spinlock contention in real workloads,
that I was able to reproduce with a netfilter rule dropping
incoming FIN packets.
53.56% server [kernel.kallsyms] [k] queued_spin_lock_slowpath
|
---queued_spin_lock_slowpath
|
--53.51%--_raw_spin_lock_irqsave
|
--53.51%--__percpu_counter_sum
tcp_check_oom
|
|--39.03%--__tcp_close
| tcp_close
| inet_release
| inet6_release
| sock_close
| __fput
| ____fput
| task_work_run
| exit_to_usermode_loop
| do_syscall_64
| entry_SYSCALL_64_after_hwframe
| __GI___libc_close
|
--14.48%--tcp_out_of_resources
tcp_write_timeout
tcp_retransmit_timer
tcp_write_timer_handler
tcp_write_timer
call_timer_fn
expire_timers
__run_timers
run_timer_softirq
__softirqentry_text_start
As explained in commit cf86a086a1 ("net/dst: use a smaller percpu_counter
batch for dst entries accounting"), default batch size is too big
for the default value of tcp_max_orphans (262144).
But even if we reduce batch sizes, there would still be cases
where the estimated count of orphans is beyond the limit,
and where tcp_too_many_orphans() has to call the expensive
percpu_counter_sum_positive().
One solution is to use plain per-cpu counters, and have
a timer to periodically refresh this cache.
Updating this cache every 100ms seems about right, tcp pressure
state is not radically changing over shorter periods.
percpu_counter was nice 15 years ago while hosts had less
than 16 cpus, not anymore by current standards.
v2: Fix the build issue for CONFIG_CRYPTO_DEV_CHELSIO_TLS=m,
reported by kernel test robot <lkp@intel.com>
Remove unused socket argument from tcp_too_many_orphans()
Fixes: dd24c00191 ("net: Use a percpu_counter for orphan_count")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Stefan Bach <sfb@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The commit 6da5b0f027 ("net: ensure unbound datagram socket to be
chosen when not in a VRF") modified compute_score() so that a device
match is always made, not just in the case of an l3mdev skb, then
increments the score also for unbound sockets. This ensures that
sockets bound to an l3mdev are never selected when not in a VRF.
But as unbound and bound sockets are now scored equally, this results
in the last opened socket being selected if there are matches in the
default VRF for an unbound socket and a socket bound to a dev that is
not an l3mdev. However, handling prior to this commit was to always
select the bound socket in this case. Reinstate this handling by
incrementing the score only for bound sockets. The required isolation
due to choosing between an unbound socket and a socket bound to an
l3mdev remains in place due to the device match always being made.
The same approach is taken for compute_score() for stream sockets.
Fixes: 6da5b0f027 ("net: ensure unbound datagram socket to be chosen when not in a VRF")
Fixes: e78190581a ("net: ensure unbound stream socket to be chosen when not in a VRF")
Signed-off-by: Mike Manning <mmanning@vyatta.att-mail.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/cf0a8523-b362-1edf-ee78-eef63cbbb428@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When we close a listening socket, to migrate its connections to another
listener in the same reuseport group, we have to handle two kinds of child
sockets. One is that a listening socket has a reference to, and the other
is not.
The former is the TCP_ESTABLISHED/TCP_SYN_RECV sockets, and they are in the
accept queue of their listening socket. So we can pop them out and push
them into another listener's queue at close() or shutdown() syscalls. On
the other hand, the latter, the TCP_NEW_SYN_RECV socket is during the
three-way handshake and not in the accept queue. Thus, we cannot access
such sockets at close() or shutdown() syscalls. Accordingly, we have to
migrate immature sockets after their listening socket has been closed.
Currently, if their listening socket has been closed, TCP_NEW_SYN_RECV
sockets are freed at receiving the final ACK or retransmitting SYN+ACKs. At
that time, if we could select a new listener from the same reuseport group,
no connection would be aborted. However, we cannot do that because
reuseport_detach_sock() sets NULL to sk_reuseport_cb and forbids access to
the reuseport group from closed sockets.
This patch allows TCP_CLOSE sockets to remain in the reuseport group and
access it while any child socket references them. The point is that
reuseport_detach_sock() was called twice from inet_unhash() and
sk_destruct(). This patch replaces the first reuseport_detach_sock() with
reuseport_stop_listen_sock(), which checks if the reuseport group is
capable of migration. If capable, it decrements num_socks, moves the socket
backwards in socks[] and increments num_closed_socks. When all connections
are migrated, sk_destruct() calls reuseport_detach_sock() to remove the
socket from socks[], decrement num_closed_socks, and set NULL to
sk_reuseport_cb.
By this change, closed or shutdowned sockets can keep sk_reuseport_cb.
Consequently, calling listen() after shutdown() can cause EADDRINUSE or
EBUSY in inet_csk_bind_conflict() or reuseport_add_sock() which expects
such sockets not to have the reuseport group. Therefore, this patch also
loosens such validation rules so that a socket can listen again if it has a
reuseport group with num_closed_socks more than 0.
When such sockets listen again, we handle them in reuseport_resurrect(). If
there is an existing reuseport group (reuseport_add_sock() path), we move
the socket from the old group to the new one and free the old one if
necessary. If there is no existing group (reuseport_alloc() path), we
allocate a new reuseport group, detach sk from the old one, and free it if
necessary, not to break the current shutdown behaviour:
- we cannot carry over the eBPF prog of shutdowned sockets
- we cannot attach/detach an eBPF prog to/from listening sockets via
shutdowned sockets
Note that when the number of sockets gets over U16_MAX, we try to detach a
closed socket randomly to make room for the new listening socket in
reuseport_grow().
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.co.jp>
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/bpf/20210612123224.12525-4-kuniyu@amazon.co.jp
Even when implementing RFC 6056 3.3.4 (Algorithm 4: Double-Hash
Port Selection Algorithm), a patient attacker could still be able
to collect enough state from an otherwise idle host.
Idea of this patch is to inject some noise, in the
cases __inet_hash_connect() found a candidate in the first
attempt.
This noise should not significantly reduce the collision
avoidance, and should be zero if connection table
is already well used.
Note that this is not implementing RFC 6056 3.3.5
because we think Algorithm 5 could hurt typical
workloads.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: David Dworken <ddworken@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
RFC 6056 (Recommendations for Transport-Protocol Port Randomization)
provides good summary of why source selection needs extra care.
David Dworken reminded us that linux implements Algorithm 3
as described in RFC 6056 3.3.3
Quoting David :
In the context of the web, this creates an interesting info leak where
websites can count how many TCP connections a user's computer is
establishing over time. For example, this allows a website to count
exactly how many subresources a third party website loaded.
This also allows:
- Distinguishing between different users behind a VPN based on
distinct source port ranges.
- Tracking users over time across multiple networks.
- Covert communication channels between different browsers/browser
profiles running on the same computer
- Tracking what applications are running on a computer based on
the pattern of how fast source ports are getting incremented.
Section 3.3.4 describes an enhancement, that reduces
attackers ability to use the basic information currently
stored into the shared 'u32 hint'.
This change also decreases collision rate when
multiple applications need to connect() to
different destinations.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: David Dworken <ddworken@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When the TCP stack is in SYN flood mode, the server child socket is
created from the SYN cookie received in a TCP packet with the ACK flag
set.
The child socket is created when the server receives the first TCP
packet with a valid SYN cookie from the client. Usually, this packet
corresponds to the final step of the TCP 3-way handshake, the ACK
packet. But is also possible to receive a valid SYN cookie from the
first TCP data packet sent by the client, and thus create a child socket
from that SYN cookie.
Since a client socket is ready to send data as soon as it receives the
SYN+ACK packet from the server, the client can send the ACK packet (sent
by the TCP stack code), and the first data packet (sent by the userspace
program) almost at the same time, and thus the server will equally
receive the two TCP packets with valid SYN cookies almost at the same
instant.
When such event happens, the TCP stack code has a race condition that
occurs between the momement a lookup is done to the established
connections hashtable to check for the existence of a connection for the
same client, and the moment that the child socket is added to the
established connections hashtable. As a consequence, this race condition
can lead to a situation where we add two child sockets to the
established connections hashtable and deliver two sockets to the
userspace program to the same client.
This patch fixes the race condition by checking if an existing child
socket exists for the same client when we are adding the second child
socket to the established connections socket. If an existing child
socket exists, we drop the packet and discard the second child socket
to the same client.
Signed-off-by: Ricardo Dias <rdias@singlestore.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20201120111133.GA67501@rdias-suse-pc.lan
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The arg exact_dif is not used anymore, remove it. inet_exact_dif_match()
is no longer needed after the above is removed, so remove it too.
Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In the case of TPROXY, bind_conflict optimizations for SO_REUSEADDR or
SO_REUSEPORT are broken, possibly resulting in O(n) instead of O(1) bind
behaviour or in the incorrect reuse of a bind.
the kernel keeps track for each bind_bucket if all sockets in the
bind_bucket support SO_REUSEADDR or SO_REUSEPORT in two fastreuse flags.
These flags allow skipping the costly bind_conflict check when possible
(meaning when all sockets have the proper SO_REUSE option).
For every socket added to a bind_bucket, these flags need to be updated.
As soon as a socket that does not support reuse is added, the flag is
set to false and will never go back to true, unless the bind_bucket is
deleted.
Note that there is no mechanism to re-evaluate these flags when a socket
is removed (this might make sense when removing a socket that would not
allow reuse; this leaves room for a future patch).
For this optimization to work, it is mandatory that these flags are
properly initialized and updated.
When a child socket is created from a listen socket in
__inet_inherit_port, the TPROXY case could create a new bind bucket
without properly initializing these flags, thus preventing the
optimization to work. Alternatively, a socket not allowing reuse could
be added to an existing bind bucket without updating the flags, causing
bind_conflict to never be called as it should.
Call inet_csk_update_fastreuse when __inet_inherit_port decides to create
a new bind_bucket or use a different bind_bucket than the one of the
listen socket.
Fixes: 093d282321 ("tproxy: fix hash locking issue when using port redirection in __inet_inherit_port()")
Acked-by: Matthieu Baerts <matthieu.baerts@tessares.net>
Signed-off-by: Tim Froidcoeur <tim.froidcoeur@tessares.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Run a BPF program before looking up a listening socket on the receive path.
Program selects a listening socket to yield as result of socket lookup by
calling bpf_sk_assign() helper and returning SK_PASS code. Program can
revert its decision by assigning a NULL socket with bpf_sk_assign().
Alternatively, BPF program can also fail the lookup by returning with
SK_DROP, or let the lookup continue as usual with SK_PASS on return, when
no socket has been selected with bpf_sk_assign().
This lets the user match packets with listening sockets freely at the last
possible point on the receive path, where we know that packets are destined
for local delivery after undergoing policing, filtering, and routing.
With BPF code selecting the socket, directing packets destined to an IP
range or to a port range to a single socket becomes possible.
In case multiple programs are attached, they are run in series in the order
in which they were attached. The end result is determined from return codes
of all the programs according to following rules:
1. If any program returned SK_PASS and selected a valid socket, the socket
is used as result of socket lookup.
2. If more than one program returned SK_PASS and selected a socket,
last selection takes effect.
3. If any program returned SK_DROP, and no program returned SK_PASS and
selected a socket, socket lookup fails with -ECONNREFUSED.
4. If all programs returned SK_PASS and none of them selected a socket,
socket lookup continues to htable-based lookup.
Suggested-by: Marek Majkowski <marek@cloudflare.com>
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/20200717103536.397595-5-jakub@cloudflare.com
Prepare for calling into reuseport from __inet_lookup_listener as well.
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Link: https://lore.kernel.org/bpf/20200717103536.397595-4-jakub@cloudflare.com
Michal Kubecek and Firo Yang did a very nice analysis of crashes
happening in __inet_lookup_established().
Since a TCP socket can go from TCP_ESTABLISH to TCP_LISTEN
(via a close()/socket()/listen() cycle) without a RCU grace period,
I should not have changed listeners linkage in their hash table.
They must use the nulls protocol (Documentation/RCU/rculist_nulls.txt),
so that a lookup can detect a socket in a hash list was moved in
another one.
Since we added code in commit d296ba60d8 ("soreuseport: Resolve
merge conflict for v4/v6 ordering fix"), we have to add
hlist_nulls_add_tail_rcu() helper.
Fixes: 3b24d854cb ("tcp/dccp: do not touch listener sk_refcnt under synflood")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Michal Kubecek <mkubecek@suse.cz>
Reported-by: Firo Yang <firo.yang@suse.com>
Reviewed-by: Michal Kubecek <mkubecek@suse.cz>
Link: https://lore.kernel.org/netdev/20191120083919.GH27852@unicorn.suse.cz/
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Kuniyuki Iwashima