kernel/net: recheck the ARP entry after sleeping for a reply

arp_del_entry compacts the table, so the entry could move while an app
waited and esi then pointed at another IP's entry - a valid one sent the
packet to the wrong MAC. Look the IP up again when it no longer matches.

kernel/trunk/network/ARP.inc

@@ -23,6 +23,7 @@ ARP_AWAITING_RESPONSE   = 2
 ARP_RESPONSE_TIMEOUT    = 3
 
 ARP_REQUEST_TTL         = 31          ; 20 s
+ARP_WAIT_TICKS          = 100         ; 1 s
 ARP_ENTRY_TTL           = 937         ; 600 s
 ARP_STATIC_ENTRY        = -1
 
@@ -565,10 +566,10 @@ arp_ip_to_mac:
 ; Now send a request packet on the network
         pop     edi eax                 ; IP in eax, device number in ebx, for ARP_output_request
 
-        push    esi edi
+        push    eax esi edi             ; arp_output_request trashes eax
         mov     ebx, [net_device_list + edi]
         call    arp_output_request
-        pop     edi esi
+        pop     edi esi eax
   .found_it:
         cmp     [esi + ARP_entry.Status], ARP_VALID_MAPPING             ; Does it have a MAC assigned?
         je      .valid
@@ -577,11 +578,50 @@ if ARP_BLOCK
 
         cmp     [esi + ARP_entry.Status], ARP_AWAITING_RESPONSE         ; Are we waiting for reply from remote end?
         jne     .give_up
-        push    esi
-        mov     esi, 10                 ; wait 10 ms
-        call    delay_ms
-        pop     esi
-        jmp     .found_it               ; now check again
+
+; Only applications may wait for the reply, kernel threads give up at once
+
+        push    ecx edx
+        mov     edx, [current_slot]
+        cmp     [edx + APPDATA.priority], MAX_PRIORITY
+        je      .wait_over
+        mov     edx, [timer_ticks]
+        add     edx, ARP_WAIT_TICKS
+  .wait:
+        push    ebx
+        mov     ebx, 1                  ; sleep 10 ms
+        call    delay_hs
+        pop     ebx
+
+; The entry may have moved (arp_del_entry), find it again
+        cmp     [esi + ARP_entry.IP], eax
+        je      .same_entry
+        mov     ecx, [ARP_entries + edi]
+        test    ecx, ecx
+        jz      .wait_over
+        mov     esi, edi
+        imul    esi, (sizeof.ARP_entry * ARP_TABLE_SIZE)/4
+        add     esi, ARP_table + ARP_entry.IP
+  .rescan:
+        cmp     [esi], eax
+        je      .same_entry
+        add     esi, sizeof.ARP_entry
+        dec     ecx
+        jnz     .rescan
+        jmp     .wait_over
+  .same_entry:
+        cmp     [esi + ARP_entry.Status], ARP_VALID_MAPPING
+        je      .wait_done
+        cmp     [esi + ARP_entry.Status], ARP_AWAITING_RESPONSE
+        jne     .wait_over
+        mov     ecx, [timer_ticks]
+        sub     ecx, edx
+        js      .wait
+  .wait_over:
+        pop     edx ecx
+        jmp     .give_up
+  .wait_done:
+        pop     edx ecx
 
 else
 

kernel/trunk/network/tcp.inc

@@ -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       = 2           ; min persist (1,28s)
+TCP_time_pers_min       = 8           ; min persist (5,12s)
 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)

kernel/trunk/network/tcp_input.inc

@@ -525,11 +525,8 @@ 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
@@ -1117,11 +1114,8 @@ 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_

kernel/trunk/network/tcp_output.inc

@@ -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?
-        jne     @f
+        je      @f
         mov     [eax + TCP_SOCKET.t_rtt], 1                     ; nope, start transmission timer
         mov     [eax + TCP_SOCKET.t_rtseq], edi
         inc     [TCPS_segstimed]

kernel/trunk/network/tcp_subr.inc

@@ -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
-        mov     [socket + TCP_SOCKET.t_rxtcur], TCP_time_rtt_default
+;;; TODO: TCP_time_rangeset
 
         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_srtt], 0    ; first sample?
+        cmp     [ebx + TCP_SOCKET.t_rtt], 0
         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
-        jg      @f                              ; signed: delta may be negative
+        ja      @f
         mov     [ebx + TCP_SOCKET.t_srtt], 1
   @@:
 
@@ -556,10 +556,10 @@ tcp_xmit_timer:
         pop     edx
 
         add     [ebx + TCP_SOCKET.t_rttvar], eax
-        jg      @f
+        ja      @f
         mov     [ebx + TCP_SOCKET.t_rttvar], 1
   @@:
-        jmp     .rto
+        ret
 
 
   .no_rtt_yet:
@@ -572,16 +572,6 @@ 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
 
 

kernel/trunk/network/tcp_timer.inc

@@ -91,10 +91,6 @@ 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
@@ -103,95 +99,6 @@ 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