[ Upstream commit 38d7e44433 ]
tcp_measure_rcv_mss() is used to update icsk->icsk_ack.rcv_mss
(tcpi_rcv_mss in tcp_info) and tp->scaling_ratio.
Calling it from tcp_data_queue_ofo() makes sure these
fields are updated, and permits a better tuning
of sk->sk_rcvbuf, in the case a new flow receives many ooo
packets.
Fixes: dfa2f04833 ("tcp: get rid of sysctl_tcp_adv_win_scale")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250711114006.480026-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit a041f70e57 ]
If the new coming segment covers more than one skbs in the ofo queue,
and which seq is equal to rcv_nxt, then the sequence range
that is duplicated will be sent as DUP SACK, the detail as below,
in step6, the {501,2001} range is clearly including too much
DUP SACK range, in violation of RFC 2883 rules.
1. client > server: Flags [.], seq 501:1001, ack 1325288529, win 20000, length 500
2. server > client: Flags [.], ack 1, [nop,nop,sack 1 {501:1001}], length 0
3. client > server: Flags [.], seq 1501:2001, ack 1325288529, win 20000, length 500
4. server > client: Flags [.], ack 1, [nop,nop,sack 2 {1501:2001} {501:1001}], length 0
5. client > server: Flags [.], seq 1:2001, ack 1325288529, win 20000, length 2000
6. server > client: Flags [.], ack 2001, [nop,nop,sack 1 {501:2001}], length 0
After this fix, the final ACK is as below:
6. server > client: Flags [.], ack 2001, options [nop,nop,sack 1 {501:1001}], length 0
[edumazet] added a new packetdrill test in the following patch.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: xin.guo <guoxin0309@gmail.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250626123420.1933835-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 3ac9e29211 ]
The referenced commit replaced a call to __xfrm4|6_udp_encap_rcv() with
a custom check for non-ESP markers. But what the called function also
did was setting the transport header to the ESP header. The function
that follows, esp4|6_gro_receive(), relies on that being set when it calls
xfrm_parse_spi(). We have to set the full offset as the skb's head was
not moved yet so adding just the UDP header length won't work.
Fixes: e3fd057776 ("xfrm: Fix UDP GRO handling for some corner cases")
Signed-off-by: Tobias Brunner <tobias@strongswan.org>
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9f735b6f8a ]
Since "net: gro: use cb instead of skb->network_header", the skb network
header is no longer set in the GRO path.
This breaks fraglist segmentation, which relies on ip_hdr()/tcp_hdr()
to check for address/port changes.
Fix this regression by selectively setting the network header for merged
segment skbs.
Fixes: 186b1ea73a ("net: gro: use cb instead of skb->network_header")
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20250705150622.10699-1-nbd@nbd.name
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit d3a5f2871a ]
Syzkaller reported a bug [1] where sk->sk_forward_alloc can overflow.
When we send data, if an skb exists at the tail of the write queue, the
kernel will attempt to append the new data to that skb. However, the code
that checks for available space in the skb is flawed:
'''
copy = size_goal - skb->len
'''
The types of the variables involved are:
'''
copy: ssize_t (s64 on 64-bit systems)
size_goal: int
skb->len: unsigned int
'''
Due to C's type promotion rules, the signed size_goal is converted to an
unsigned int to match skb->len before the subtraction. The result is an
unsigned int.
When this unsigned int result is then assigned to the s64 copy variable,
it is zero-extended, preserving its non-negative value. Consequently, copy
is always >= 0.
Assume we are sending 2GB of data and size_goal has been adjusted to a
value smaller than skb->len. The subtraction will result in copy holding a
very large positive integer. In the subsequent logic, this large value is
used to update sk->sk_forward_alloc, which can easily cause it to overflow.
The syzkaller reproducer uses TCP_REPAIR to reliably create this
condition. However, this can also occur in real-world scenarios. The
tcp_bound_to_half_wnd() function can also reduce size_goal to a small
value. This would cause the subsequent tcp_wmem_schedule() to set
sk->sk_forward_alloc to a value close to INT_MAX. Further memory
allocation requests would then cause sk_forward_alloc to wrap around and
become negative.
[1]: https://syzkaller.appspot.com/bug?extid=de6565462ab540f50e47
Reported-by: syzbot+de6565462ab540f50e47@syzkaller.appspotmail.com
Fixes: 270a1c3de4 ("tcp: Support MSG_SPLICE_PAGES")
Signed-off-by: Jiayuan Chen <jiayuan.chen@linux.dev>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Howells <dhowells@redhat.com>
Link: https://patch.msgid.link/20250707054112.101081-1-jiayuan.chen@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit dbe0ca8da1 ]
There is a bug with passive TFO sockets returning an invalid NAPI ID 0
from SO_INCOMING_NAPI_ID. Normally this is not an issue, but zero copy
receive relies on a correct NAPI ID to process sockets on the right
queue.
Fix by adding a sk_mark_napi_id_set().
Fixes: e5907459ce ("tcp: Record Rx hash and NAPI ID in tcp_child_process")
Signed-off-by: David Wei <dw@davidwei.uk>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250617212102.175711-5-dw@davidwei.uk
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit d0fa59897e ]
After the following commit from 2024:
commit e37ab73736 ("tcp: fix to allow timestamp undo if no retransmits were sent")
...there was buggy behavior where TCP connections without SACK support
could easily see erroneous undo events at the end of fast recovery or
RTO recovery episodes. The erroneous undo events could cause those
connections to suffer repeated loss recovery episodes and high
retransmit rates.
The problem was an interaction between the non-SACK behavior on these
connections and the undo logic. The problem is that, for non-SACK
connections at the end of a loss recovery episode, if snd_una ==
high_seq, then tcp_is_non_sack_preventing_reopen() holds steady in
CA_Recovery or CA_Loss, but clears tp->retrans_stamp to 0. Then upon
the next ACK the "tcp: fix to allow timestamp undo if no retransmits
were sent" logic saw the tp->retrans_stamp at 0 and erroneously
concluded that no data was retransmitted, and erroneously performed an
undo of the cwnd reduction, restoring cwnd immediately to the value it
had before loss recovery. This caused an immediate burst of traffic
and build-up of queues and likely another immediate loss recovery
episode.
This commit fixes tcp_packet_delayed() to ignore zero retrans_stamp
values for non-SACK connections when snd_una is at or above high_seq,
because tcp_is_non_sack_preventing_reopen() clears retrans_stamp in
this case, so it's not a valid signal that we can undo.
Note that the commit named in the Fixes footer restored long-present
behavior from roughly 2005-2019, so apparently this bug was present
for a while during that era, and this was simply not caught.
Fixes: e37ab73736 ("tcp: fix to allow timestamp undo if no retransmits were sent")
Reported-by: Eric Wheeler <netdev@lists.ewheeler.net>
Closes: https://lore.kernel.org/netdev/64ea9333-e7f9-0df-b0f2-8d566143acab@ewheeler.net/
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Co-developed-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ea33537d82 ]
If the application can not drain fast enough a TCP socket queue,
tcp_rcv_space_adjust() can overestimate tp->rcvq_space.space.
Then sk->sk_rcvbuf can grow and hit tcp_rmem[2] for no good reason.
Fix this by taking into acount the number of available bytes.
Keeping sk->sk_rcvbuf at the right size allows better cache efficiency.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Wei Wang <weiwan@google.com>
Link: https://patch.msgid.link/20250513193919.1089692-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit b879dcb1ae ]
tcp_rcv_rtt_update() goal is to maintain an estimation of the RTT
in tp->rcv_rtt_est.rtt_us, used by tcp_rcv_space_adjust()
When TCP TS are enabled, tcp_rcv_rtt_update() is using
EWMA to smooth the samples.
Change this to immediately latch the incoming value if it
is lower than tp->rcv_rtt_est.rtt_us, so that tcp_rcv_space_adjust()
does not overshoot tp->rcvq_space.space and sk->sk_rcvbuf.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250513193919.1089692-8-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 3382a1ed7f ]
Commit a1e40ac5b5 ("net: gso: fix udp gso fraglist segmentation after
pull from frag_list") detected invalid geometry in frag_list skbs and
redirects them from skb_segment_list to more robust skb_segment. But some
packets with modified geometry can also hit bugs in that code. We don't
know how many such cases exist. Addressing each one by one also requires
touching the complex skb_segment code, which risks introducing bugs for
other types of skbs. Instead, linearize all these packets that fail the
basic invariants on gso fraglist skbs. That is more robust.
If only part of the fraglist payload is pulled into head_skb, it will
always cause exception when splitting skbs by skb_segment. For detailed
call stack information, see below.
Valid SKB_GSO_FRAGLIST skbs
- consist of two or more segments
- the head_skb holds the protocol headers plus first gso_size
- one or more frag_list skbs hold exactly one segment
- all but the last must be gso_size
Optional datapath hooks such as NAT and BPF (bpf_skb_pull_data) can
modify fraglist skbs, breaking these invariants.
In extreme cases they pull one part of data into skb linear. For UDP,
this causes three payloads with lengths of (11,11,10) bytes were
pulled tail to become (12,10,10) bytes.
The skbs no longer meets the above SKB_GSO_FRAGLIST conditions because
payload was pulled into head_skb, it needs to be linearized before pass
to regular skb_segment.
skb_segment+0xcd0/0xd14
__udp_gso_segment+0x334/0x5f4
udp4_ufo_fragment+0x118/0x15c
inet_gso_segment+0x164/0x338
skb_mac_gso_segment+0xc4/0x13c
__skb_gso_segment+0xc4/0x124
validate_xmit_skb+0x9c/0x2c0
validate_xmit_skb_list+0x4c/0x80
sch_direct_xmit+0x70/0x404
__dev_queue_xmit+0x64c/0xe5c
neigh_resolve_output+0x178/0x1c4
ip_finish_output2+0x37c/0x47c
__ip_finish_output+0x194/0x240
ip_finish_output+0x20/0xf4
ip_output+0x100/0x1a0
NF_HOOK+0xc4/0x16c
ip_forward+0x314/0x32c
ip_rcv+0x90/0x118
__netif_receive_skb+0x74/0x124
process_backlog+0xe8/0x1a4
__napi_poll+0x5c/0x1f8
net_rx_action+0x154/0x314
handle_softirqs+0x154/0x4b8
[118.376811] [C201134] rxq0_pus: [name:bug&]kernel BUG at net/core/skbuff.c:4278!
[118.376829] [C201134] rxq0_pus: [name:traps&]Internal error: Oops - BUG: 00000000f2000800 [#1] PREEMPT SMP
[118.470774] [C201134] rxq0_pus: [name:mrdump&]Kernel Offset: 0x178cc00000 from 0xffffffc008000000
[118.470810] [C201134] rxq0_pus: [name:mrdump&]PHYS_OFFSET: 0x40000000
[118.470827] [C201134] rxq0_pus: [name:mrdump&]pstate: 60400005 (nZCv daif +PAN -UAO)
[118.470848] [C201134] rxq0_pus: [name:mrdump&]pc : [0xffffffd79598aefc] skb_segment+0xcd0/0xd14
[118.470900] [C201134] rxq0_pus: [name:mrdump&]lr : [0xffffffd79598a5e8] skb_segment+0x3bc/0xd14
[118.470928] [C201134] rxq0_pus: [name:mrdump&]sp : ffffffc008013770
Fixes: a1e40ac5b5 ("gso: fix udp gso fraglist segmentation after pull from frag_list")
Signed-off-by: Shiming Cheng <shiming.cheng@mediatek.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9a119669fb ]
fib has two modes:
1. Obtain output device according to source or destination address
2. Obtain the type of the address, e.g. local, unicast, multicast.
'fib daddr type' should return 'local' if the address is configured
in this netns or unicast otherwise.
'fib daddr . iif type' should return 'local' if the address is configured
on the input interface or unicast otherwise, i.e. more restrictive.
However, if the interface is part of a VRF, then 'fib daddr type'
returns unicast even if the address is configured on the incoming
interface.
This is broken for both ipv4 and ipv6.
In the ipv4 case, inet_dev_addr_type must only be used if the
'iif' or 'oif' (strict mode) was requested.
Else inet_addr_type_dev_table() needs to be used and the correct
dev argument must be passed as well so the correct fib (vrf) table
is used.
In the ipv6 case, the bug is similar, without strict mode, dev is NULL
so .flowi6_l3mdev will be set to 0.
Add a new 'nft_fib_l3mdev_master_ifindex_rcu()' helper and use that
to init the .l3mdev structure member.
For ipv6, use it from nft_fib6_flowi_init() which gets called from
both the 'type' and the 'route' mode eval functions.
This provides consistent behaviour for all modes for both ipv4 and ipv6:
If strict matching is requested, the input respectively output device
of the netfilter hooks is used.
Otherwise, use skb->dev to obtain the l3mdev ifindex.
Without this, most type checks in updated nft_fib.sh selftest fail:
FAIL: did not find veth0 . 10.9.9.1 . local in fibtype4
FAIL: did not find veth0 . dead:1::1 . local in fibtype6
FAIL: did not find veth0 . dead:9::1 . local in fibtype6
FAIL: did not find tvrf . 10.0.1.1 . local in fibtype4
FAIL: did not find tvrf . 10.9.9.1 . local in fibtype4
FAIL: did not find tvrf . dead:1::1 . local in fibtype6
FAIL: did not find tvrf . dead:9::1 . local in fibtype6
FAIL: fib expression address types match (iif in vrf)
(fib errounously returns 'unicast' for all of them, even
though all of these addresses are local to the vrf).
Fixes: f6d0cbcf09 ("netfilter: nf_tables: add fib expression")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit e3fd057776 ]
This fixes an issue that's caused if there is a mismatch between the data
offset in the GRO header and the length fields in the regular sk_buff due
to the pskb_pull()/skb_push() calls. That's because the UDP GRO layer
stripped off the UDP header via skb_gro_pull() already while the UDP
header was explicitly not pulled/pushed in this function.
For example, an IKE packet that triggered this had len=data_len=1268 and
the data_offset in the GRO header was 28 (IPv4 + UDP). So pskb_pull()
was called with an offset of 28-8=20, which reduced len to 1248 and via
pskb_may_pull() and __pskb_pull_tail() it also set data_len to 1248.
As the ESP offload module was not loaded, the function bailed out and
called skb_push(), which restored len to 1268, however, data_len remained
at 1248.
So while skb_headlen() was 0 before, it was now 20. The latter caused a
difference of 8 instead of 28 (or 0 if pskb_pull()/skb_push() was called
with the complete GRO data_offset) in gro_try_pull_from_frag0() that
triggered a call to gro_pull_from_frag0() that corrupted the packet.
This change uses a more GRO-like approach seen in other GRO receivers
via skb_gro_header() to just read the actual data we are interested in
and does not try to "restore" the UDP header at this point to call the
existing function. If the offload module is not loaded, it immediately
bails out, otherwise, it only does a quick check to see if the packet
is an IKE or keepalive packet instead of calling the existing function.
Fixes: 172bf009c1 ("xfrm: Support GRO for IPv4 ESP in UDP encapsulation")
Fixes: 221ddb723d ("xfrm: Support GRO for IPv6 ESP in UDP encapsulation")
Signed-off-by: Tobias Brunner <tobias@strongswan.org>
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 028363685b ]
The current scheme for caching the encap socket can lead to reference
leaks when we try to delete the netns.
The reference chain is: xfrm_state -> enacp_sk -> netns
Since the encap socket is a userspace socket, it holds a reference on
the netns. If we delete the espintcp state (through flush or
individual delete) before removing the netns, the reference on the
socket is dropped and the netns is correctly deleted. Otherwise, the
netns may not be reachable anymore (if all processes within the ns
have terminated), so we cannot delete the xfrm state to drop its
reference on the socket.
This patch results in a small (~2% in my tests) performance
regression.
A GC-type mechanism could be added for the socket cache, to clear
references if the state hasn't been used "recently", but it's a lot
more complex than just not caching the socket.
Fixes: e27cca96cd ("xfrm: add espintcp (RFC 8229)")
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 254ba7e603 ]
fib_valid_key_len() is called in the beginning of fib_table_insert()
or fib_table_delete() to check if the prefix length is valid.
fib_table_insert() and fib_table_delete() are called from 3 paths
- ip_rt_ioctl()
- inet_rtm_newroute() / inet_rtm_delroute()
- fib_magic()
In the first ioctl() path, rtentry_to_fib_config() checks the prefix
length with bad_mask(). Also, fib_magic() always passes the correct
prefix: 32 or ifa->ifa_prefixlen, which is already validated.
Let's move fib_valid_key_len() to the rtnetlink path, rtm_to_fib_config().
While at it, 2 direct returns in rtm_to_fib_config() are changed to
goto to match other places in the same function
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20250228042328.96624-12-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 149dfb3161 ]
- Move tcp_count_delivered() earlier and split tcp_count_delivered_ce()
out of it
- Move tcp_in_ack_event() later
- While at it, remove the inline from tcp_in_ack_event() and let
the compiler to decide
Accurate ECN's heuristics does not know if there is going
to be ACE field based CE counter increase or not until after
rtx queue has been processed. Only then the number of ACKed
bytes/pkts is available. As CE or not affects presence of
FLAG_ECE, that information for tcp_in_ack_event is not yet
available in the old location of the call to tcp_in_ack_event().
Signed-off-by: Ilpo Järvinen <ij@kernel.org>
Signed-off-by: Chia-Yu Chang <chia-yu.chang@nokia-bell-labs.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit eaaff9b670 upstream.
In case the fib match is used from the input hook we can avoid the fib
lookup if early demux assigned a socket for us: check that the input
interface matches sk-cached one.
Rework the existing 'lo bypass' logic to first check sk, then
for loopback interface type to elide the fib lookup.
This speeds up fib matching a little, before:
93.08 GBit/s (no rules at all)
75.1 GBit/s ("fib saddr . iif oif missing drop" in prerouting)
75.62 GBit/s ("fib saddr . iif oif missing drop" in input)
After:
92.48 GBit/s (no rules at all)
75.62 GBit/s (fib rule in prerouting)
90.37 GBit/s (fib rule in input).
Numbers for the 'no rules' and 'prerouting' are expected to
closely match in-between runs, the 3rd/input test case exercises the
the 'avoid lookup if cached ifindex in sk matches' case.
Test used iperf3 via veth interface, lo can't be used due to existing
loopback test.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 8930424777 ]
Because skb_tunnel_check_pmtu() doesn't handle PACKET_HOST packets,
commit 30a92c9e3d ("openvswitch: Set the skbuff pkt_type for proper
pmtud support.") forced skb->pkt_type to PACKET_OUTGOING for
openvswitch packets that are sent using the OVS_ACTION_ATTR_OUTPUT
action. This allowed such packets to invoke the
iptunnel_pmtud_check_icmp() or iptunnel_pmtud_check_icmpv6() helpers
and thus trigger PMTU update on the input device.
However, this also broke other parts of PMTU discovery. Since these
packets don't have the PACKET_HOST type anymore, they won't trigger the
sending of ICMP Fragmentation Needed or Packet Too Big messages to
remote hosts when oversized (see the skb_in->pkt_type condition in
__icmp_send() for example).
These two skb->pkt_type checks are therefore incompatible as one
requires skb->pkt_type to be PACKET_HOST, while the other requires it
to be anything but PACKET_HOST.
It makes sense to not trigger ICMP messages for non-PACKET_HOST packets
as these messages should be generated only for incoming l2-unicast
packets. However there doesn't seem to be any reason for
skb_tunnel_check_pmtu() to ignore PACKET_HOST packets.
Allow both cases to work by allowing skb_tunnel_check_pmtu() to work on
PACKET_HOST packets and not overriding skb->pkt_type in openvswitch
anymore.
Fixes: 30a92c9e3d ("openvswitch: Set the skbuff pkt_type for proper pmtud support.")
Fixes: 4cb47a8644 ("tunnels: PMTU discovery support for directly bridged IP packets")
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Reviewed-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: Aaron Conole <aconole@redhat.com>
Tested-by: Aaron Conole <aconole@redhat.com>
Link: https://patch.msgid.link/eac941652b86fddf8909df9b3bf0d97bc9444793.1743208264.git.gnault@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit df207de9d9 ]
Matt Dowling reported a weird UDP memory usage issue.
Under normal operation, the UDP memory usage reported in /proc/net/sockstat
remains close to zero. However, it occasionally spiked to 524,288 pages
and never dropped. Moreover, the value doubled when the application was
terminated. Finally, it caused intermittent packet drops.
We can reproduce the issue with the script below [0]:
1. /proc/net/sockstat reports 0 pages
# cat /proc/net/sockstat | grep UDP:
UDP: inuse 1 mem 0
2. Run the script till the report reaches 524,288
# python3 test.py & sleep 5
# cat /proc/net/sockstat | grep UDP:
UDP: inuse 3 mem 524288 <-- (INT_MAX + 1) >> PAGE_SHIFT
3. Kill the socket and confirm the number never drops
# pkill python3 && sleep 5
# cat /proc/net/sockstat | grep UDP:
UDP: inuse 1 mem 524288
4. (necessary since v6.0) Trigger proto_memory_pcpu_drain()
# python3 test.py & sleep 1 && pkill python3
5. The number doubles
# cat /proc/net/sockstat | grep UDP:
UDP: inuse 1 mem 1048577
The application set INT_MAX to SO_RCVBUF, which triggered an integer
overflow in udp_rmem_release().
When a socket is close()d, udp_destruct_common() purges its receive
queue and sums up skb->truesize in the queue. This total is calculated
and stored in a local unsigned integer variable.
The total size is then passed to udp_rmem_release() to adjust memory
accounting. However, because the function takes a signed integer
argument, the total size can wrap around, causing an overflow.
Then, the released amount is calculated as follows:
1) Add size to sk->sk_forward_alloc.
2) Round down sk->sk_forward_alloc to the nearest lower multiple of
PAGE_SIZE and assign it to amount.
3) Subtract amount from sk->sk_forward_alloc.
4) Pass amount >> PAGE_SHIFT to __sk_mem_reduce_allocated().
When the issue occurred, the total in udp_destruct_common() was 2147484480
(INT_MAX + 833), which was cast to -2147482816 in udp_rmem_release().
At 1) sk->sk_forward_alloc is changed from 3264 to -2147479552, and
2) sets -2147479552 to amount. 3) reverts the wraparound, so we don't
see a warning in inet_sock_destruct(). However, udp_memory_allocated
ends up doubling at 4).
Since commit 3cd3399dd7 ("net: implement per-cpu reserves for
memory_allocated"), memory usage no longer doubles immediately after
a socket is close()d because __sk_mem_reduce_allocated() caches the
amount in udp_memory_per_cpu_fw_alloc. However, the next time a UDP
socket receives a packet, the subtraction takes effect, causing UDP
memory usage to double.
This issue makes further memory allocation fail once the socket's
sk->sk_rmem_alloc exceeds net.ipv4.udp_rmem_min, resulting in packet
drops.
To prevent this issue, let's use unsigned int for the calculation and
call sk_forward_alloc_add() only once for the small delta.
Note that first_packet_length() also potentially has the same problem.
[0]:
from socket import *
SO_RCVBUFFORCE = 33
INT_MAX = (2 ** 31) - 1
s = socket(AF_INET, SOCK_DGRAM)
s.bind(('', 0))
s.setsockopt(SOL_SOCKET, SO_RCVBUFFORCE, INT_MAX)
c = socket(AF_INET, SOCK_DGRAM)
c.connect(s.getsockname())
data = b'a' * 100
while True:
c.send(data)
Fixes: f970bd9e3a ("udp: implement memory accounting helpers")
Reported-by: Matt Dowling <madowlin@amazon.com>
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20250401184501.67377-3-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 5a465a0da1 ]
__udp_enqueue_schedule_skb() has the following condition:
if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
goto drop;
sk->sk_rcvbuf is initialised by net.core.rmem_default and later can
be configured by SO_RCVBUF, which is limited by net.core.rmem_max,
or SO_RCVBUFFORCE.
If we set INT_MAX to sk->sk_rcvbuf, the condition is always false
as sk->sk_rmem_alloc is also signed int.
Then, the size of the incoming skb is added to sk->sk_rmem_alloc
unconditionally.
This results in integer overflow (possibly multiple times) on
sk->sk_rmem_alloc and allows a single socket to have skb up to
net.core.udp_mem[1].
For example, if we set a large value to udp_mem[1] and INT_MAX to
sk->sk_rcvbuf and flood packets to the socket, we can see multiple
overflows:
# cat /proc/net/sockstat | grep UDP:
UDP: inuse 3 mem 7956736 <-- (7956736 << 12) bytes > INT_MAX * 15
^- PAGE_SHIFT
# ss -uam
State Recv-Q ...
UNCONN -1757018048 ... <-- flipping the sign repeatedly
skmem:(r2537949248,rb2147483646,t0,tb212992,f1984,w0,o0,bl0,d0)
Previously, we had a boundary check for INT_MAX, which was removed by
commit 6a1f12dd85 ("udp: relax atomic operation on sk->sk_rmem_alloc").
A complete fix would be to revert it and cap the right operand by
INT_MAX:
rmem = atomic_add_return(size, &sk->sk_rmem_alloc);
if (rmem > min(size + (unsigned int)sk->sk_rcvbuf, INT_MAX))
goto uncharge_drop;
but we do not want to add the expensive atomic_add_return() back just
for the corner case.
Casting rmem to unsigned int prevents multiple wraparounds, but we still
allow a single wraparound.
# cat /proc/net/sockstat | grep UDP:
UDP: inuse 3 mem 524288 <-- (INT_MAX + 1) >> 12
# ss -uam
State Recv-Q ...
UNCONN -2147482816 ... <-- INT_MAX + 831 bytes
skmem:(r2147484480,rb2147483646,t0,tb212992,f3264,w0,o0,bl0,d14468947)
So, let's define rmem and rcvbuf as unsigned int and check skb->truesize
only when rcvbuf is large enough to lower the overflow possibility.
Note that we still have a small chance to see overflow if multiple skbs
to the same socket are processed on different core at the same time and
each size does not exceed the limit but the total size does.
Note also that we must ignore skb->truesize for a small buffer as
explained in commit 363dc73aca ("udp: be less conservative with
sock rmem accounting").
Fixes: 6a1f12dd85 ("udp: relax atomic operation on sk->sk_rmem_alloc")
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20250401184501.67377-2-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 3c9231ea64 ]
When I read through the TSO codes, I found out that we probably
miss initializing the tx_flags of last seg when TSO is turned
off, which means at the following points no more timestamp
(for this last one) will be generated. There are three flags
to be handled in this patch:
1. SKBTX_HW_TSTAMP
2. SKBTX_BPF
3. SKBTX_SCHED_TSTAMP
Note that SKBTX_BPF[1] was added in 6.14.0-rc2 by commit
6b98ec7e88 ("bpf: Add BPF_SOCK_OPS_TSTAMP_SCHED_CB callback")
and only belongs to net-next branch material for now. The common
issue of the above three flags can be fixed by this single patch.
This patch initializes the tx_flags to SKBTX_ANY_TSTAMP like what
the UDP GSO does to make the newly segmented last skb inherit the
tx_flags so that requested timestamp will be generated in each
certain layer, or else that last one has zero value of tx_flags
which leads to no timestamp at all.
Fixes: 4ed2d765df ("net-timestamp: TCP timestamping")
Signed-off-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ee01b2f2d7 ]
In __udp_gso_segment the skb destructor is removed before segmenting the
skb but the socket reference is kept as-is. This is an issue if the
original skb is later orphaned as we can hit the following bug:
kernel BUG at ./include/linux/skbuff.h:3312! (skb_orphan)
RIP: 0010:ip_rcv_core+0x8b2/0xca0
Call Trace:
ip_rcv+0xab/0x6e0
__netif_receive_skb_one_core+0x168/0x1b0
process_backlog+0x384/0x1100
__napi_poll.constprop.0+0xa1/0x370
net_rx_action+0x925/0xe50
The above can happen following a sequence of events when using
OpenVSwitch, when an OVS_ACTION_ATTR_USERSPACE action precedes an
OVS_ACTION_ATTR_OUTPUT action:
1. OVS_ACTION_ATTR_USERSPACE is handled (in do_execute_actions): the skb
goes through queue_gso_packets and then __udp_gso_segment, where its
destructor is removed.
2. The segments' data are copied and sent to userspace.
3. OVS_ACTION_ATTR_OUTPUT is handled (in do_execute_actions) and the
same original skb is sent to its path.
4. If it later hits skb_orphan, we hit the bug.
Fix this by also removing the reference to the socket in
__udp_gso_segment.
Fixes: ad405857b1 ("udp: better wmem accounting on gso")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Link: https://patch.msgid.link/20250226171352.258045-1-atenart@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 8d52da23b6 ]
Recently a bug was discovered where the server had entered TCP_ESTABLISHED
state, but the upper layers were not notified.
The same 5-tuple packet may be processed by different CPUSs, so two
CPUs may receive different ack packets at the same time when the
state is TCP_NEW_SYN_RECV.
In that case, req->ts_recent in tcp_check_req may be changed concurrently,
which will probably cause the newsk's ts_recent to be incorrectly large.
So that tcp_validate_incoming will fail. At this point, newsk will not be
able to enter the TCP_ESTABLISHED.
cpu1 cpu2
tcp_check_req
tcp_check_req
req->ts_recent = rcv_tsval = t1
req->ts_recent = rcv_tsval = t2
syn_recv_sock
tcp_sk(child)->rx_opt.ts_recent = req->ts_recent = t2 // t1 < t2
tcp_child_process
tcp_rcv_state_process
tcp_validate_incoming
tcp_paws_check
if ((s32)(rx_opt->ts_recent - rx_opt->rcv_tsval) <= paws_win)
// t2 - t1 > paws_win, failed
tcp_v4_do_rcv
tcp_rcv_state_process
// TCP_ESTABLISHED
The cpu2's skb or a newly received skb will call tcp_v4_do_rcv to get
the newsk into the TCP_ESTABLISHED state, but at this point it is no
longer possible to notify the upper layer application. A notification
mechanism could be added here, but the fix is more complex, so the
current fix is used.
In tcp_check_req, req->ts_recent is used to assign a value to
tcp_sk(child)->rx_opt.ts_recent, so removing the change in req->ts_recent
and changing tcp_sk(child)->rx_opt.ts_recent directly after owning the
req fixes this bug.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Wang Hai <wanghai38@huawei.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 18912c5206 ]
Currently, we report -ETOOSMALL (err) only on the first iteration
(!sent). When we get put_cmsg error after a bunch of successful
put_cmsg calls, we don't signal the error at all. This might be
confusing on the userspace side which will see truncated CMSGs
but no MSG_CTRUNC signal.
Consider the following case:
- sizeof(struct cmsghdr) = 16
- sizeof(struct dmabuf_cmsg) = 24
- total cmsg size (CMSG_LEN) = 40 (16+24)
When calling recvmsg with msg_controllen=60, the userspace
will receive two(!) dmabuf_cmsg(s), the first one will
be a valid one and the second one will be silently truncated. There is no
easy way to discover the truncation besides doing something like
"cm->cmsg_len != CMSG_LEN(sizeof(dmabuf_cmsg))".
Introduce new put_devmem_cmsg wrapper that reports an error instead
of doing the truncation. Mina suggests that it's the intended way
this API should work.
Note that we might now report MSG_CTRUNC when the users (incorrectly)
call us with msg_control == NULL.
Fixes: 8f0b3cc9a4 ("tcp: RX path for devmem TCP")
Reviewed-by: Mina Almasry <almasrymina@google.com>
Signed-off-by: Stanislav Fomichev <sdf@fomichev.me>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250224174401.3582695-1-sdf@fomichev.me
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 7e863e5db6 ]
Pass a dscp_t variable to ip_route_input(), instead of a plain u8, to
prevent accidental setting of ECN bits in ->flowi4_tos.
Callers of ip_route_input() to consider are:
* input_action_end_dx4_finish() and input_action_end_dt4() in
net/ipv6/seg6_local.c. These functions set the tos parameter to 0,
which is already a valid dscp_t value, so they don't need to be
adjusted for the new prototype.
* icmp_route_lookup(), which already has a dscp_t variable to pass as
parameter. We just need to remove the inet_dscp_to_dsfield()
conversion.
* br_nf_pre_routing_finish(), ip_options_rcv_srr() and ip4ip6_err(),
which get the DSCP directly from IPv4 headers. Define a helper to
read the .tos field of struct iphdr as dscp_t, so that these
function don't have to do the conversion manually.
While there, declare *iph as const in br_nf_pre_routing_finish(),
declare its local variables in reverse-christmas-tree order and move
the "err = ip_route_input()" assignment out of the conditional to avoid
checkpatch warning.
Signed-off-by: Guillaume Nault <gnault@redhat.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/e9d40781d64d3d69f4c79ac8a008b8d67a033e8d.1727807926.git.gnault@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Stable-dep-of: 27843ce6ba ("ipvlan: ensure network headers are in skb linear part")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 36b62df568 ]
'sk->copied_seq' was updated in the tcp_eat_skb() function when the action
of a BPF program was SK_REDIRECT. For other actions, like SK_PASS, the
update logic for 'sk->copied_seq' was moved to tcp_bpf_recvmsg_parser()
to ensure the accuracy of the 'fionread' feature.
It works for a single stream_verdict scenario, as it also modified
sk_data_ready->sk_psock_verdict_data_ready->tcp_read_skb
to remove updating 'sk->copied_seq'.
However, for programs where both stream_parser and stream_verdict are
active (strparser purpose), tcp_read_sock() was used instead of
tcp_read_skb() (sk_data_ready->strp_data_ready->tcp_read_sock).
tcp_read_sock() now still updates 'sk->copied_seq', leading to duplicate
updates.
In summary, for strparser + SK_PASS, copied_seq is redundantly calculated
in both tcp_read_sock() and tcp_bpf_recvmsg_parser().
The issue causes incorrect copied_seq calculations, which prevent
correct data reads from the recv() interface in user-land.
We do not want to add new proto_ops to implement a new version of
tcp_read_sock, as this would introduce code complexity [1].
We could have added noack and copied_seq to desc, and then called
ops->read_sock. However, unfortunately, other modules didn’t fully
initialize desc to zero. So, for now, we are directly calling
tcp_read_sock_noack() in tcp_bpf.c.
[1]: https://lore.kernel.org/bpf/20241218053408.437295-1-mrpre@163.com
Fixes: e5c6de5fa0 ("bpf, sockmap: Incorrectly handling copied_seq")
Suggested-by: Jakub Sitnicki <jakub@cloudflare.com>
Signed-off-by: Jiayuan Chen <mrpre@163.com>
Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>
Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://patch.msgid.link/20250122100917.49845-3-mrpre@163.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9b6412e697 ]
Xiumei reported hitting the WARN in xfrm6_tunnel_net_exit while
running tests that boil down to:
- create a pair of netns
- run a basic TCP test over ipcomp6
- delete the pair of netns
The xfrm_state found on spi_byaddr was not deleted at the time we
delete the netns, because we still have a reference on it. This
lingering reference comes from a secpath (which holds a ref on the
xfrm_state), which is still attached to an skb. This skb is not
leaked, it ends up on sk_receive_queue and then gets defer-free'd by
skb_attempt_defer_free.
The problem happens when we defer freeing an skb (push it on one CPU's
defer_list), and don't flush that list before the netns is deleted. In
that case, we still have a reference on the xfrm_state that we don't
expect at this point.
We already drop the skb's dst in the TCP receive path when it's no
longer needed, so let's also drop the secpath. At this point,
tcp_filter has already called into the LSM hooks that may require the
secpath, so it should not be needed anymore. However, in some of those
places, the MPTCP extension has just been attached to the skb, so we
cannot simply drop all extensions.
Fixes: 68822bdf76 ("net: generalize skb freeing deferral to per-cpu lists")
Reported-by: Xiumei Mu <xmu@redhat.com>
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/5055ba8f8f72bdcb602faa299faca73c280b7735.1739743613.git.sd@queasysnail.net
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 4eae0ee0f1 ]
The arp_req_set_public() function is called with the rtnl lock held,
which provides enough synchronization protection. This makes the RCU
variant of dev_getbyhwaddr() unnecessary. Switch to using the simpler
dev_getbyhwaddr() function since we already have the required rtnl
locking.
This change helps maintain consistency in the networking code by using
the appropriate helper function for the existing locking context.
Since we're not holding the RCU read lock in arp_req_set_public()
existing code could trigger false positive locking warnings.
Fixes: 941666c2e3 ("net: RCU conversion of dev_getbyhwaddr() and arp_ioctl()")
Suggested-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://patch.msgid.link/20250218-arm_fix_selftest-v5-2-d3d6892db9e1@debian.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit f5da7c4518 ]
Since commit under Fixes we set the window clamp in accordance
to newly measured rcvbuf scaling_ratio. If the scaling_ratio
decreased significantly we may put ourselves in a situation
where windows become smaller than rcvq_space, preventing
tcp_rcv_space_adjust() from increasing rcvbuf.
The significant decrease of scaling_ratio is far more likely
since commit 697a6c8cec ("tcp: increase the default TCP scaling ratio"),
which increased the "default" scaling ratio from ~30% to 50%.
Hitting the bad condition depends a lot on TCP tuning, and
drivers at play. One of Meta's workloads hits it reliably
under following conditions:
- default rcvbuf of 125k
- sender MTU 1500, receiver MTU 5000
- driver settles on scaling_ratio of 78 for the config above.
Initial rcvq_space gets calculated as TCP_INIT_CWND * tp->advmss
(10 * 5k = 50k). Once we find out the true scaling ratio and
MSS we clamp the windows to 38k. Triggering the condition also
depends on the message sequence of this workload. I can't repro
the problem with simple iperf or TCP_RR-style tests.
Fixes: a2cbb16039 ("tcp: Update window clamping condition")
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://patch.msgid.link/20250217232905.3162187-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 7d3f3b4367 ]
Check number of paths by fib_info_num_path(),
and update_or_create_fnhe() for every path.
Problem is that pmtu is cached only for the oif
that has received icmp message "need to frag",
other oifs will still try to use "default" iface mtu.
An example topology showing the problem:
| host1
+---------+
| dummy0 | 10.179.20.18/32 mtu9000
+---------+
+-----------+----------------+
+---------+ +---------+
| ens17f0 | 10.179.2.141/31 | ens17f1 | 10.179.2.13/31
+---------+ +---------+
| (all here have mtu 9000) |
+------+ +------+
| ro1 | 10.179.2.140/31 | ro2 | 10.179.2.12/31
+------+ +------+
| |
---------+------------+-------------------+------
|
+-----+
| ro3 | 10.10.10.10 mtu1500
+-----+
|
========================================
some networks
========================================
|
+-----+
| eth0| 10.10.30.30 mtu9000
+-----+
| host2
host1 have enabled multipath and
sysctl net.ipv4.fib_multipath_hash_policy = 1:
default proto static src 10.179.20.18
nexthop via 10.179.2.12 dev ens17f1 weight 1
nexthop via 10.179.2.140 dev ens17f0 weight 1
When host1 tries to do pmtud from 10.179.20.18/32 to host2,
host1 receives at ens17f1 iface an icmp packet from ro3 that ro3 mtu=1500.
And host1 caches it in nexthop exceptions cache.
Problem is that it is cached only for the iface that has received icmp,
and there is no way that ro3 will send icmp msg to host1 via another path.
Host1 now have this routes to host2:
ip r g 10.10.30.30 sport 30000 dport 443
10.10.30.30 via 10.179.2.12 dev ens17f1 src 10.179.20.18 uid 0
cache expires 521sec mtu 1500
ip r g 10.10.30.30 sport 30033 dport 443
10.10.30.30 via 10.179.2.140 dev ens17f0 src 10.179.20.18 uid 0
cache
So when host1 tries again to reach host2 with mtu>1500,
if packet flow is lucky enough to be hashed with oif=ens17f1 its ok,
if oif=ens17f0 it blackholes and still gets icmp msgs from ro3 to ens17f1,
until lucky day when ro3 will send it through another flow to ens17f0.
Signed-off-by: Vladimir Vdovin <deliran@verdict.gg>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20241108093427.317942-1-deliran@verdict.gg
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Stable-dep-of: 139512191b ("ipv4: use RCU protection in __ip_rt_update_pmtu()")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 235174b2be ]
Commit 4094871db1 ("udp: only do GSO if # of segs > 1") avoided GSO
for small packets. But the kernel currently dismisses GSO requests only
after checking MTU/PMTU on gso_size. This means any packets, regardless
of their payload sizes, could be dropped when PMTU becomes smaller than
requested gso_size. We encountered this issue in production and it
caused a reliability problem that new QUIC connection cannot be
established before PMTU cache expired, while non GSO sockets still
worked fine at the same time.
Ideally, do not check any GSO related constraints when payload size is
smaller than requested gso_size, and return EMSGSIZE instead of EINVAL
on MTU/PMTU check failure to be more specific on the error cause.
Fixes: 4094871db1 ("udp: only do GSO if # of segs > 1")
Signed-off-by: Yan Zhai <yan@cloudflare.com>
Suggested-by: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 8c670bdfa5 ]
Testing with iperf3 using the "pasta" protocol splicer has revealed
a problem in the way tcp handles window advertising in extreme memory
squeeze situations.
Under memory pressure, a socket endpoint may temporarily advertise
a zero-sized window, but this is not stored as part of the socket data.
The reasoning behind this is that it is considered a temporary setting
which shouldn't influence any further calculations.
However, if we happen to stall at an unfortunate value of the current
window size, the algorithm selecting a new value will consistently fail
to advertise a non-zero window once we have freed up enough memory.
This means that this side's notion of the current window size is
different from the one last advertised to the peer, causing the latter
to not send any data to resolve the sitution.
The problem occurs on the iperf3 server side, and the socket in question
is a completely regular socket with the default settings for the
fedora40 kernel. We do not use SO_PEEK or SO_RCVBUF on the socket.
The following excerpt of a logging session, with own comments added,
shows more in detail what is happening:
// tcp_v4_rcv(->)
// tcp_rcv_established(->)
[5201<->39222]: ==== Activating log @ net/ipv4/tcp_input.c/tcp_data_queue()/5257 ====
[5201<->39222]: tcp_data_queue(->)
[5201<->39222]: DROPPING skb [265600160..265665640], reason: SKB_DROP_REASON_PROTO_MEM
[rcv_nxt 265600160, rcv_wnd 262144, snt_ack 265469200, win_now 131184]
[copied_seq 259909392->260034360 (124968), unread 5565800, qlen 85, ofoq 0]
[OFO queue: gap: 65480, len: 0]
[5201<->39222]: tcp_data_queue(<-)
[5201<->39222]: __tcp_transmit_skb(->)
[tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160]
[5201<->39222]: tcp_select_window(->)
[5201<->39222]: (inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM) ? --> TRUE
[tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160]
returning 0
[5201<->39222]: tcp_select_window(<-)
[5201<->39222]: ADVERTISING WIN 0, ACK_SEQ: 265600160
[5201<->39222]: [__tcp_transmit_skb(<-)
[5201<->39222]: tcp_rcv_established(<-)
[5201<->39222]: tcp_v4_rcv(<-)
// Receive queue is at 85 buffers and we are out of memory.
// We drop the incoming buffer, although it is in sequence, and decide
// to send an advertisement with a window of zero.
// We don't update tp->rcv_wnd and tp->rcv_wup accordingly, which means
// we unconditionally shrink the window.
[5201<->39222]: tcp_recvmsg_locked(->)
[5201<->39222]: __tcp_cleanup_rbuf(->) tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160
[5201<->39222]: [new_win = 0, win_now = 131184, 2 * win_now = 262368]
[5201<->39222]: [new_win >= (2 * win_now) ? --> time_to_ack = 0]
[5201<->39222]: NOT calling tcp_send_ack()
[tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160]
[5201<->39222]: __tcp_cleanup_rbuf(<-)
[rcv_nxt 265600160, rcv_wnd 262144, snt_ack 265469200, win_now 131184]
[copied_seq 260040464->260040464 (0), unread 5559696, qlen 85, ofoq 0]
returning 6104 bytes
[5201<->39222]: tcp_recvmsg_locked(<-)
// After each read, the algorithm for calculating the new receive
// window in __tcp_cleanup_rbuf() finds it is too small to advertise
// or to update tp->rcv_wnd.
// Meanwhile, the peer thinks the window is zero, and will not send
// any more data to trigger an update from the interrupt mode side.
[5201<->39222]: tcp_recvmsg_locked(->)
[5201<->39222]: __tcp_cleanup_rbuf(->) tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160
[5201<->39222]: [new_win = 262144, win_now = 131184, 2 * win_now = 262368]
[5201<->39222]: [new_win >= (2 * win_now) ? --> time_to_ack = 0]
[5201<->39222]: NOT calling tcp_send_ack()
[tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160]
[5201<->39222]: __tcp_cleanup_rbuf(<-)
[rcv_nxt 265600160, rcv_wnd 262144, snt_ack 265469200, win_now 131184]
[copied_seq 260099840->260171536 (71696), unread 5428624, qlen 83, ofoq 0]
returning 131072 bytes
[5201<->39222]: tcp_recvmsg_locked(<-)
// The above pattern repeats again and again, since nothing changes
// between the reads.
[...]
[5201<->39222]: tcp_recvmsg_locked(->)
[5201<->39222]: __tcp_cleanup_rbuf(->) tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160
[5201<->39222]: [new_win = 262144, win_now = 131184, 2 * win_now = 262368]
[5201<->39222]: [new_win >= (2 * win_now) ? --> time_to_ack = 0]
[5201<->39222]: NOT calling tcp_send_ack()
[tp->rcv_wup: 265469200, tp->rcv_wnd: 262144, tp->rcv_nxt 265600160]
[5201<->39222]: __tcp_cleanup_rbuf(<-)
[rcv_nxt 265600160, rcv_wnd 262144, snt_ack 265469200, win_now 131184]
[copied_seq 265600160->265600160 (0), unread 0, qlen 0, ofoq 0]
returning 54672 bytes
[5201<->39222]: tcp_recvmsg_locked(<-)
// The receive queue is empty, but no new advertisement has been sent.
// The peer still thinks the receive window is zero, and sends nothing.
// We have ended up in a deadlock situation.
Note that well behaved endpoints will send win0 probes, so the problem
will not occur.
Furthermore, we have observed that in these situations this side may
send out an updated 'th->ack_seq´ which is not stored in tp->rcv_wup
as it should be. Backing ack_seq seems to be harmless, but is of
course still wrong from a protocol viewpoint.
We fix this by updating the socket state correctly when a packet has
been dropped because of memory exhaustion and we have to advertize
a zero window.
Further testing shows that the connection recovers neatly from the
squeeze situation, and traffic can continue indefinitely.
Fixes: e2142825c1 ("net: tcp: send zero-window ACK when no memory")
Cc: Menglong Dong <menglong8.dong@gmail.com>
Reviewed-by: Stefano Brivio <sbrivio@redhat.com>
Signed-off-by: Jon Maloy <jmaloy@redhat.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://patch.msgid.link/20250127231304.1465565-1-jmaloy@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
