kernel/net: allocate socket rings from non-contiguous pages

kernel/trunk/core/heap.inc

@@ -463,7 +463,7 @@ proc kernel_alloc stdcall, size:dword
         and     eax, -PAGE_SIZE;
         mov     [size], eax
         and     eax, eax
-        jz      .err
+        jz      .err0
         mov     ebx, eax
         shr     ebx, 12
         mov     [pages_count], ebx
@@ -509,6 +509,13 @@ proc kernel_alloc stdcall, size:dword
         pop     ebx
         ret
 .err:
+        mov     eax, [lin_addr]
+        test    eax, eax
+        jz      .err0
+        mov     ecx, [pages_count]
+        call    release_pages
+        stdcall free_kernel_space, [lin_addr]
+.err0:
         xor     eax, eax
         pop     edi
         pop     ebx

kernel/trunk/core/memory.inc

@@ -1397,7 +1397,9 @@ proc create_ring_buffer stdcall, size:dword, flags:dword
         mov     [page_tabs + edi], eax
         mov     [page_tabs + edi + edx], eax
         invlpg  [ebx]
-        invlpg  [ebx+0x10000]
+        add     ebx, [size]
+        invlpg  [ebx]                   ; the mirror half
+        sub     ebx, [size]
         add     eax, 0x1000
         add     ebx, 0x1000
         add     edi, 4
@@ -1416,6 +1418,62 @@ proc create_ring_buffer stdcall, size:dword, flags:dword
         ret
 endp
 
+; Same as create_ring_buffer, but from non-contiguous pages.
+; Not for DMA. Free with kernel_free.
+align 4
+proc create_ring_buffer_sparse stdcall, size:dword, flags:dword
+           locals
+             buf_ptr  dd ?
+           endl
+
+        mov     eax, [size]
+        test    eax, eax
+        jz      .fail
+
+        add     eax, eax
+        stdcall alloc_kernel_space, eax
+        test    eax, eax
+        jz      .fail
+        mov     [buf_ptr], eax
+
+        push    ebx esi edi
+        mov     ebx, eax                ; page being mapped
+        mov     esi, [size]             ; distance to its mirror
+        mov     edi, esi
+        shr     edi, 12                 ; pages left
+  .page:
+        call    alloc_page
+        test    eax, eax
+        jz      .mm_fail
+        or      eax, [flags]
+        mov     edx, ebx
+        shr     edx, 12
+        mov     [page_tabs + edx*4], eax
+        invlpg  [ebx]
+        lea     edx, [ebx + esi]
+        shr     edx, 12
+        mov     [page_tabs + edx*4], eax
+        invlpg  [ebx + esi]
+        add     ebx, PAGE_SIZE
+        dec     edi
+        jnz     .page
+
+        mov     eax, [buf_ptr]
+        pop     edi esi ebx
+        ret
+
+  .mm_fail:
+        mov     eax, [buf_ptr]
+        mov     ecx, [size]
+        shr     ecx, 11                 ; both halves, in pages
+        call    release_pages
+        stdcall free_kernel_space, [buf_ptr]
+        pop     edi esi ebx
+        xor     eax, eax
+  .fail:
+        ret
+endp
+
 
 align 4
 proc print_mem

kernel/trunk/network/socket.inc

@@ -198,6 +198,10 @@ ends
 
 SOCKET_STRUCT_SIZE      = 4096          ; in bytes
 
+if sizeof.STREAM_SOCKET > SOCKET_STRUCT_SIZE
+  err "STREAM_SOCKET does not fit in SOCKET_STRUCT_SIZE"
+end if
+
 SOCKET_QUEUE_SIZE       = 10            ; maximum number of incoming packets queued for 1 socket
 ; the incoming packet queue for sockets is placed in the socket struct itself, at this location from start
 SOCKET_QUEUE_LOCATION   = (SOCKET_STRUCT_SIZE - SOCKET_QUEUE_SIZE*sizeof.socket_queue_entry - sizeof.queue)
@@ -1647,7 +1651,7 @@ socket_ring_create:
         mov     esi, eax
 
         push    edx
-        stdcall create_ring_buffer, SOCKET_BUFFER_SIZE, PG_SWR
+        stdcall create_ring_buffer_sparse, SOCKET_BUFFER_SIZE, PG_SWR
         pop     edx
         test    eax, eax
         jz      .fail

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