Re: [BUG] TCP connection deadlock under simultaneous bidirectional ICSK_ACK_NOMEM (OOM)

[xietangxin <xietangxin@yeah.net>]

On 6/8/2026 7:55 PM, Menglong Dong wrote:
> On 2026/6/4 16:22 xietangxin <xietangxin@yeah.net> write:
>> Hi all,
>>
>> We have observed a TCP connection deadlock on stable 6.6 under heavy stress testing.
>>
>> 1.Both Peer A and Peer B enter the ICSK_ACK_NOMEM branch in tcp_select_window().
>> After commit 8c670bdfa58e ("tcp: correct handling of extreme memory squeeze"),
>> Both peers freeze their rcv_nxt and set rcv_wnd = 0.
>>
>> 2.Prior to freezing, both sides had already sent out flight data.
>> Since both sides are dropping incoming data packets due to OOM, rcv_nxt stops advancing,
>> but the peer's seq of subsequent packets continues to grow.
>>
>> 3.When Peer A receives Peer B's Zero Window ACK,
>> the packet's seq is far ahead of Peer A's frozen rcv_nxt.
>> Both peers drop each other's packet, also no Zero Window Probes are triggered
>> because snd_wnd is never updated to 0.
>>
> 
> Hi,
> 
> The problem you addressed is already fixed in this commit:
> 0e24d17bd966 ("tcp: implement RFC 7323 window retraction receiver requirements"),
> which hasn't been picked to the 6.6 branch.
> 
> That patch doesn't have the Fix tag, so I'm not sure if it will be picked
> to the 6.6 branch. Just CC the linux-stable :)
> 
> Thanks!
> Menglong Dong
> 
>>
>> Simplified Packet Trace:
>>
>> Assume Peer A's rcv_nxt = 1000, and Peer B's rcv_nxt = 5000 initially.
>>
>> Time  Dir      Type        Seq   Ack   Win  Len  Status
>> ------------------------------------------------------------------------
>> T1:   B -> A   [PSH, ACK]  1000  5000  3000 100  (A hits OOM, rcv_nxt=1000)
>> T2:   B -> A   [ACK]       1100  5000  3000 200  (Dropped due to A's OOM)
>> T3:   B -> A   [PSH, ACK]  1300  5000  3000 200  (Dropped due to A's OOM)
>>
>> T4:   A -> B   [PSH, ACK]  5000  1000  3000 100  (B hits OOM, rcv_nxt=5000)
>> T5:   A -> B   [ACK]       5100  1000  3000 200  (Dropped due to B's OOM)
>> T6:   A -> B   [PSH, ACK]  5300  1000  3000 200  (Dropped due to B's OOM)
>>
>> -- Both sides are now in OOM. B's Seq is 1500; A's Seq is 5500 --
>>
>> T7:   B -> A   [ZeroWin]   1500  5000  0    0    (Dropped: Seq 1500 != 1000)
>> T8:   A -> B   [ZeroWin]   5500  1000  0    0    (Dropped: Seq 5500 != 5000)
>> T9:   A -> B   [WinUpdate] 5500  1000  20   0    (Dropped: Seq 5500 != 5000)
>>
>> Should we relax the sequence check in tcp_sequence() for zero window ACK?
>>
>> Any feedback or guidance would be greatly appreciated.
>>
>> -- 
>> Best regards,
>> Tangxin Xie
>>
>>
>>
> 
> 
> 

Hi,

We observed a throughput regression (dropping from ~1GB/s to 100MB/s)
in our test environment after commit 0e24d17bd966
("tcp: implement RFC 7323 window retraction receiver requirements").

When the rcv_buf reaches the pressure triggers tcp_clamp_window().
then rcv_ssthresh is strictly capped to 2 * advmss.
Subsequently, even after the user completely consumes the data and releases
a massive amount of free_space, tcp_select_window() is still heavily
suppressed by the clamped rcv_ssthresh. As a result, the receiver advertises
an extremely small window (Win=23) to the peer.

The sender cannot transmit any new data segments, until the sender's RTO timer
expires and triggers a slow-start recovery. This 200ms silence window slashes
our bandwidth by 90%.


No.   Time           Source       Destination  Info
-----------------------------------------------------------------------------------------------
1045  08:16:06.8005  192.168.1.9  192.168.1.10  [TCP ZeroWindow] 57334 -> 6666 [PSH, ACK] Win=0
1052  08:16:06.8013  192.168.1.9  192.168.1.10  [TCP Window Update] 57334 -> 6666 [ACK] Win=23
1055  08:16:06.8036  192.168.1.10  192.168.1.9  6666 -> 57334 [ACK] Seq=2999704568 Ack=2416286095
=========================== 200ms  SILENCE (RTO WAITING) ===================================
1088  08:16:07.0056  192.168.1.10  192.168.1.9  [TCP Retransmission] 6666 -> 57334 Len=1448
1090  08:16:07.0060  192.168.1.10  192.168.1.9  [TCP Retransmission] Len=2896

-- 
Best regards,
Tangxin Xie