kernel/net: lock the socket in the retransmission timer

- take SOCKET.mutex while rewriting SND_NXT/cwnd/ssthresh; skip if an ACK stopped or re-armed the timer meanwhile - back off from t_rxtcur: the first timeout was 3.2 s, the second 25.6 s
kernel/net: fix TCP RTT measurement
2026-09-30 10:14:17 +03:00 · 2026-09-30 10:14:15 +03:00 · 2026-09-30 10:14:12 +03:00 · 2026-09-30 10:14:12 +03:00
5 changed files with 116 additions and 7 deletions
@@ -63,7 +63,7 @@ TCP_OPT_TIMESTAMP       = 8
 TCP_time_MSL            = 47          ; max segment lifetime (30s)
 TCP_time_re_min         = 2           ; min retransmission (1,28s)
 TCP_time_re_max         = 100         ; max retransmission (64s)
-TCP_time_pers_min       = 8           ; min persist (5,12s)
+TCP_time_pers_min       = 2           ; min persist (1,28s)
 TCP_time_pers_max       = 94          ; max persist (60,16s)
 TCP_time_keep_init      = 118         ; connection establishment (75,52s)
 TCP_time_keep_idle      = 4608        ; idle time before 1st probe (2h)
@@ -525,8 +525,11 @@ endl

        test    [temp_bits], TCP_BIT_TIMESTAMP
        jz      .no_timestamp_rtt
+        cmp     [ebx + TCP_SOCKET.ts_ecr], 0    ; nothing echoed
+        je      .no_timestamp_rtt
        mov     eax, [timestamp]
        sub     eax, [ebx + TCP_SOCKET.ts_ecr]
+        shr     eax, 6                          ; 1/100 s -> 640 ms ticks
        inc     eax
        call    tcp_xmit_timer
        jmp     .rtt_done
@@ -1114,8 +1117,11 @@ endl

        test    [temp_bits], TCP_BIT_TIMESTAMP
        jz      .timestamp_not_present
+        cmp     [ebx + TCP_SOCKET.ts_ecr], 0    ; nothing echoed
+        je      .timestamp_not_present
        mov     eax, [timestamp]
        sub     eax, [ebx + TCP_SOCKET.ts_ecr]
+        shr     eax, 6                          ; 1/100 s -> 640 ms ticks
        inc     eax
        call    tcp_xmit_timer
        jmp     .rtt_done_
@@ -626,7 +626,7 @@ endl
        mov     [eax + TCP_SOCKET.SND_MAX], edx                 ; [eax + TCP_SOCKET.SND_NXT] from before we updated it

        cmp     [eax + TCP_SOCKET.t_rtt], 0                     ; are we currently timing anything?
-        je      @f
+        jne     @f
        mov     [eax + TCP_SOCKET.t_rtt], 1                     ; nope, start transmission timer
        mov     [eax + TCP_SOCKET.t_rtseq], edi
        inc     [TCPS_segstimed]
@@ -96,7 +96,7 @@ macro   tcp_init_socket socket {
        mov     [socket + TCP_SOCKET.t_srtt], TCP_time_srtt_default
        mov     [socket + TCP_SOCKET.t_rttvar], TCP_time_rtt_default * 4
        mov     [socket + TCP_SOCKET.t_rttmin], TCP_time_re_min
-;;; TODO: TCP_time_rangeset
+        mov     [socket + TCP_SOCKET.t_rxtcur], TCP_time_rtt_default

        mov     [socket + TCP_SOCKET.SND_CWND], TCP_max_win shl TCP_max_winshift
        mov     [socket + TCP_SOCKET.SND_SSTHRESH], TCP_max_win shl TCP_max_winshift
@@ -518,7 +518,7 @@ tcp_xmit_timer:

        inc     [TCPS_rttupdated]

-        cmp     [ebx + TCP_SOCKET.t_rtt], 0
+        cmp     [ebx + TCP_SOCKET.t_srtt], 0    ; first sample?
        je      .no_rtt_yet

 ; srtt is stored as a fixed point with 3 bits after the binary point.
@@ -534,7 +534,7 @@ tcp_xmit_timer:
        pop     ecx

        add     [ebx + TCP_SOCKET.t_srtt], eax
-        ja      @f
+        jg      @f                              ; signed: delta may be negative
        mov     [ebx + TCP_SOCKET.t_srtt], 1
  @@:

@@ -556,10 +556,10 @@ tcp_xmit_timer:
        pop     edx

        add     [ebx + TCP_SOCKET.t_rttvar], eax
-        ja      @f
+        jg      @f
        mov     [ebx + TCP_SOCKET.t_rttvar], 1
  @@:
-        ret
+        jmp     .rto


  .no_rtt_yet:
@@ -572,6 +572,16 @@ tcp_xmit_timer:
        mov     [ebx + TCP_SOCKET.t_rttvar], eax
        pop     ecx

+  .rto:
+; Retransmit timeout = srtt + 4*rttvar, reset the backoff
+        push    ecx
+        mov     ecx, [ebx + TCP_SOCKET.t_srtt]
+        shr     ecx, TCP_RTT_SHIFT
+        add     ecx, [ebx + TCP_SOCKET.t_rttvar]
+        tcpt_rangeset [ebx + TCP_SOCKET.t_rxtcur], ecx, TCP_time_re_min, TCP_time_re_max
+        pop     ecx
+        mov     [ebx + TCP_SOCKET.t_rxtshift], 0
+        mov     [ebx + TCP_SOCKET.t_rtt], 0     ; the timed segment is done
        ret


@@ -91,6 +91,10 @@ proc tcp_timer_640ms
        jne     .loop

        inc     [eax + TCP_SOCKET.t_idle]
+        cmp     [eax + TCP_SOCKET.t_rtt], 0     ; timing a segment?
+        je      @f
+        inc     [eax + TCP_SOCKET.t_rtt]
+  @@:

        test    [eax + TCP_SOCKET.timer_flags], timer_flag_retransmission
        jz      .check_more2
@@ -99,6 +103,95 @@ proc tcp_timer_640ms

        DEBUGF  DEBUG_NETWORK_VERBOSE, "socket %x: Retransmission timer expired\n", eax

+; Lock socket, an ACK may have stopped or restarted the timer meanwhile
+        pusha
+        lea     ecx, [eax + SOCKET.mutex]
+        call    mutex_lock
+        popa
+        test    [eax + TCP_SOCKET.timer_flags], timer_flag_retransmission
+        jz      .rexmt_cancelled
+        cmp     [eax + TCP_SOCKET.timer_retransmission], 0
+        jne     .rexmt_cancelled
+
+; Too many retransmissions? Drop the connection
+        inc     [eax + TCP_SOCKET.t_rxtshift]
+        cmp     [eax + TCP_SOCKET.t_rxtshift], TCP_max_rxtshift
+        jbe     .rexmt
+
+        pusha
+        lea     ecx, [eax + SOCKET.mutex]
+        call    mutex_unlock
+        popa
+
+        DEBUGF  DEBUG_NETWORK_VERBOSE, "socket %x: too many retransmissions, dropping\n", eax
+
+        push    [eax + SOCKET.NextPtr]
+        mov     ebx, ETIMEDOUT
+        call    tcp_drop
+        pop     eax
+        jmp     .check_only
+
+  .rexmt_cancelled:
+        pusha
+        lea     ecx, [eax + SOCKET.mutex]
+        call    mutex_unlock
+        popa
+        jmp     .check_more2
+
+  .rexmt:
+        push    ebx ecx edx
+
+; Restart timer with backoff: t_rxtcur << min(t_rxtshift, 6)
+        mov     ebx, [eax + TCP_SOCKET.t_rxtcur]
+        mov     cl, [eax + TCP_SOCKET.t_rxtshift]
+        cmp     cl, 6
+        jbe     @f
+        mov     cl, 6
+       @@:
+        shl     ebx, cl
+        cmp     ebx, TCP_time_re_min
+        jae     @f
+        mov     ebx, TCP_time_re_min
+       @@:
+        cmp     ebx, TCP_time_re_max
+        jbe     @f
+        mov     ebx, TCP_time_re_max
+       @@:
+        mov     [eax + TCP_SOCKET.timer_retransmission], ebx
+
+; Resend from the last acknowledged byte, don't time it
+        push    [eax + TCP_SOCKET.SND_UNA]
+        pop     [eax + TCP_SOCKET.SND_NXT]
+        mov     [eax + TCP_SOCKET.t_rtt], 0
+        mov     [eax + TCP_SOCKET.t_dupacks], 0
+
+; Slow start: ssthresh = max(2, min(wnd, cwnd) / 2 / mss) * mss, cwnd = mss
+        mov     ecx, [eax + TCP_SOCKET.t_maxseg]
+        mov     edx, [eax + TCP_SOCKET.SND_WND]
+        cmp     edx, [eax + TCP_SOCKET.SND_CWND]
+        jbe     @f
+        mov     edx, [eax + TCP_SOCKET.SND_CWND]
+       @@:
+        mov     ebx, eax                                ; socket ptr
+        mov     eax, edx
+        shr     eax, 1
+        xor     edx, edx
+        div     ecx
+        cmp     eax, 2
+        jae     @f
+        mov     eax, 2
+       @@:
+        mul     ecx
+        mov     [ebx + TCP_SOCKET.SND_SSTHRESH], eax
+        mov     [ebx + TCP_SOCKET.SND_CWND], ecx
+        mov     eax, ebx
+        pop     edx ecx ebx
+
+        pusha
+        lea     ecx, [eax + SOCKET.mutex]
+        call    mutex_unlock
+        popa
+
        push    eax
        call    tcp_output
        pop     eax
Author	SHA1	Message	Date
Leency	a8f285ec86	kernel/net: lock the socket in the retransmission timer Check kernel codestyle / Check kernel codestyle (pull_request) Successful in 37s Details Test PR / Build (en_US) (pull_request) Successful in 2m47s Details Test PR / Build (es_ES) (pull_request) Successful in 2m4s Details Test PR / Build (ru_RU) (pull_request) Successful in 1m58s Details - take SOCKET.mutex while rewriting SND_NXT/cwnd/ssthresh; skip if an ACK stopped or re-armed the timer meanwhile - back off from t_rxtcur: the first timeout was 3.2 s, the second 25.6 s	2026-09-30 10:14:17 +03:00
Leency	772031e6b6	kernel/net: fix TCP RTT measurement - tcp_output: the "timing anything?" test was inverted, no segment was timed - tcp_timer: t_rtt was never incremented - tcp_xmit_timer: first sample keyed on t_srtt, not t_rtt; signed jg on srtt/rttvar updates (ja reset them to 1 whenever RTT dropped); t_rtt = 0 - timestamp RTT: 1/100 s -> 640 ms ticks (RTO came out 64x too long); skip it when TSecr is 0	2026-09-30 10:14:15 +03:00
Leency	b1110dc03b	kernel/net: start persist probes at the retransmission floor BSD waits 5.12s before the first zero-window probe and doubles from there. A peer that closes its window and never volunteers an update -- QEMU 0.10's slirp, or any stack applying silly-window avoidance to its updates -- is rediscovered only by our probe, so 5, 10, 20s of silence swallowed most of a 15s upload stage. Start at TCP_time_re_min (1.28s), as Linux starts at the RTO; still exponential, still capped.	2026-09-30 10:14:12 +03:00
Leency	6c30c66aa3	kernel/net: implement the TCP retransmission timeout t_rxtcur, the timeout the retransmission timer is armed with, was never assigned anywhere: not at socket creation, not in tcp_xmit_timer. Every socket armed the timer with 0, the 640 ms tick took it to 0xFFFFFFFF, and it never expired. And when it did (in principle) expire, the handler was a bare call tcp_output, which sends from SND_NXT: the unacknowledged segment before it was never resent, nothing was backed off, the timer was not re-armed. A segment the peer dropped -- a window that shrank under data in flight, any loss on the uplink -- stalled the connection until the application gave up. 23 duplicate ACKs from the peer did not help either. Observed as every upload through QEMU's slirp freezing the moment the host side paused: pcap of an 8 MB POST to a sink that stops reading for 12 s showed the peer's window go 328, the guest's 1400-byte segment beyond it dropped, then nothing but empty ACKs, and after the window reopened the guest sent the NEXT 32 KB with the hole still there and sat. speedtest.net's upload stage lost 3 of 4 connections this way on the 2009 QEMU. - tcp_init_socket: t_rxtcur = TCP_time_rtt_default (BSD TCPTV_RTOBASE) until the first RTT sample. - tcp_xmit_timer: t_rxtcur = srtt + 4*rttvar (BSD TCP_REXMTVAL) within [re_min, re_max]; a fresh sample resets t_rxtshift. - tcp_timer, retransmission expiry: BSD TCPT_REXMT -- drop the connection with ETIMEDOUT past TCP_max_rxtshift; re-arm with the backed-off timeout (shift capped at 6); SND_NXT = SND_UNA; no RTT sample from the resent segment; ssthresh = max(2, min(wnd,cwnd)/2/mss) segments, cwnd = one segment, dupacks = 0; then tcp_output. With the patch the same fixture completes: 8 MB, 12.36 s of which 12 s is the sink's deliberate pause, 205 Mbps once it reads again.	2026-09-30 10:14:12 +03:00