kernel: fix sysfn 68.30 (unload driver)

- restore the SRV handle after DRV_EXIT (ecx is clobbered by the driver)
- restore esp via ebp: START may be cdecl or stdcall
- pass the SRV to free in eax, not on the stack
- check SRV.base before calling DRV_EXIT

.gitea/ansible/playbooks/deploy.yml

@@ -54,3 +54,27 @@
         src: "{{ stable_artifact_path }}/"
         dest: "{{ storage_path }}/../"
         mode: preserve
+
+    # older builds shrink to compressed images and lose their unpacked tree; the newest few keep both
+    - name: Looking for zstd on the storage server
+      ansible.builtin.shell: command -v zstd
+      register: zstd_present
+      changed_when: false
+      failed_when: false
+
+    - name: Requiring zstd for compaction
+      ansible.builtin.assert:
+        that: zstd_present.rc == 0
+        fail_msg: >-
+          zstd is not installed on the storage server, so older builds cannot
+          be compacted. Install it and re-run this build.
+
+    - name: Compacting older builds
+      ansible.builtin.script:
+        cmd: >-
+          ../../utils/compact-builds.sh
+          {{ storage_path | quote }}
+          {{ compact_keep | default(5) | int }}
+          {{ compact_max | default(10) | int }}
+      register: compaction
+      changed_when: "'compacted 0 of' not in compaction.stdout"

.gitea/utils/compact-builds.sh

@@ -0,0 +1,74 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0-only
+# SPDX-FileCopyrightText: 2026 KolibriOS team
+
+# Shrink older builds on the storage server: compress their images and drop
+# their unpacked tree, which is rebuildable from the commit. sha256sums.txt
+# keeps naming the bare images, so a compacted build has to be decompressed
+# before it can be verified.
+#
+# Layout:  <root>/<version>/<lang>/kolibrios-<descr>-<lang>.{img,iso,raw} -> .zst
+#          <root>/<version>/<lang>/data/                                  -> removed
+#
+# Usage: compact-builds.sh <storage-root> [keep] [max-per-run]
+#        DRY_RUN=1 walks the same selection and changes nothing
+
+set -euo pipefail
+
+root=${1:?storage root required}
+# the build just published is always among these, so an upload in flight is safe
+keep=${2:-5}
+# a build is hundreds of MiB to re-compress: cap the first pass over the backlog
+max=${3:-10}
+dry=${DRY_RUN:-0}
+verb=compacted
+[ "$dry" = 1 ] && verb="would compact"
+
+# --rm drops the source only once the archive is written; the box serves the
+# site over NFS, so compaction yields to it
+compress_image() {
+    nice -n 10 ionice -c3 zstd -12 -T1 -q --rm "$1"
+}
+export -f compress_image
+
+leftovers() {
+    local build=$1
+    find "$build" -mindepth 2 -maxdepth 2 \
+        \( -type d -name data \
+           -o -type f \( -name '*.img' -o -name '*.iso' -o -name '*.raw' \) \) \
+        -print -quit
+}
+
+compacted=0
+pending=0
+
+while IFS= read -r version; do
+    dir="$root/$version"
+    [ -n "$(leftovers "$dir")" ] || continue
+
+    pending=$((pending + 1))
+    [ "$compacted" -lt "$max" ] || continue
+
+    if [ "$dry" = 1 ]; then
+        compacted=$((compacted + 1))
+        echo "$verb $version"
+        continue
+    fi
+
+    # a build holds nine images: one per core, each zstd kept to one thread so
+    # the two levels of parallelism do not fight. Archive before dropping the
+    # tree, so a failure here leaves the build merely uncompacted and the next
+    # run picks it up again
+    find "$dir" -mindepth 2 -maxdepth 2 -type f \
+         \( -name '*.img' -o -name '*.iso' -o -name '*.raw' \) -print0 \
+        | xargs -0 -r -P "$(nproc)" -I{} bash -c 'compress_image "$@"' _ {}
+
+    find "$dir" -mindepth 2 -maxdepth 2 -type d -name data -exec rm -rf {} +
+
+    compacted=$((compacted + 1))
+    echo "$verb $version"
+# <tag>-<commit count>-g<sha>: order by the count, never by mtime
+done < <(find "$root" -mindepth 1 -maxdepth 1 -type d -name '*-*-g*' -printf '%f\n' \
+             | sort -t- -k2,2n | head -n "-$keep")
+
+echo "$verb $compacted of $pending older build(s), $((pending - compacted)) left for the next run"

.gitea/workflows/submodule-bump.yaml

@@ -34,10 +34,11 @@ jobs:
         uses: actions/checkout@v4
         with:
           submodules: true
+          token: ${{ secrets.BOT_TOKEN }}
 
       - name: Bump ${{ matrix.name }} and open PR
         env:
-          TOKEN: ${{ secrets.GITEA_TOKEN }}
+          TOKEN: ${{ secrets.BOT_TOKEN }}
           API: ${{ github.api_url }}
           REPO: ${{ github.repository }}
           NAME: ${{ matrix.name }}

drivers/ethernet/i8255x.asm

@@ -130,6 +130,8 @@ RU_STATUS_IDLE          = 0000b shl 2
 RU_STATUS_SUSPENDED     = 0001b shl 2
 RU_STATUS_NO_RESOURCES  = 0010b shl 2
 RU_STATUS_READY         = 0100b shl 2
+SCB_STATUS_CUS          = 11000000b     ; CU Status
+CU_STATUS_ACTIVE        = 10b shl 6
 SCB_STATUS_FCP          = 1 shl 8       ; Flow Control Pause
 SCB_STATUS_SWI          = 1 shl 10      ; Software Interrupt
 SCB_STATUS_MDI          = 1 shl 11      ; MDI read/write complete
@@ -183,6 +185,8 @@ struct  txfd
 
 ends
 
+TXFD_STATUS_C           = 1 shl 15
+
 TXFD_CMD_IA             = 1 shl 0
 TXFD_CMD_CFG            = 1 shl 1
 TXFD_CMD_TX             = 1 shl 2
@@ -628,6 +632,9 @@ reset:
         mov     ax, CU_START or INT_MASK
         out     dx, ax
         call    cmd_wait
+; cmd_wait only means the SCB accepted the command. The Configure below
+; reuses confcmd, so a slow chip would read it instead of our MAC.
+        call    confcmd_wait
 
 ;-------------
 ; Configure CU
@@ -652,6 +659,7 @@ reset:
         mov     ax, CU_START                            ; expect Interrupts from now on
         out     dx, ax
         call    cmd_wait
+        call    confcmd_wait
 
 ; Start media check timer
         mov     [ebx + device.state], ETH_LINK_DOWN
@@ -678,6 +686,11 @@ init_rx_ring:
 ;---------------------
 ; build rxfd structure
 
+        cmp     [ebx + device.rx_desc], 0       ; the RFD of an earlier reset
+        je      @f
+        invoke  NetFree, [ebx + device.rx_desc]
+        mov     [ebx + device.rx_desc], 0
+  @@:
         invoke  NetAlloc, 2000
         test    eax, eax
         jz      .out_of_mem
@@ -689,6 +702,7 @@ init_rx_ring:
         mov     [esi + sizeof.NET_BUFF + rxfd.command], RXFD_CMD_EL or RXFD_CMD_SUSPEND
         mov     [esi + sizeof.NET_BUFF + rxfd.link], eax
         mov     [esi + sizeof.NET_BUFF + rxfd.count], 0
+        mov     [esi + sizeof.NET_BUFF + rxfd.rx_buf_addr], 0xffffffff  ; simplified mode, no RBD
         mov     [esi + sizeof.NET_BUFF + rxfd.size], 1528
 
         ret
@@ -709,6 +723,15 @@ init_tx_ring:
         invoke  GetPhysAddr
         mov     ecx, TX_RING_SIZE
   .next_desc:
+; after a reset with frames still unreclaimed: free them, or transmit sees the
+; descriptor busy (virt_addr) and reports TX overrun forever
+        cmp     [esi + txfd.virt_addr], 0
+        je      @f
+        push    eax ecx
+        invoke  NetFree, [esi + txfd.virt_addr]
+        pop     ecx eax
+        mov     [esi + txfd.virt_addr], 0
+  @@:
         mov     [esi + txfd.status], 0
         mov     [esi + txfd.command], 0
         lea     edx, [eax + txfd.buf_addr]
@@ -763,7 +786,7 @@ proc transmit stdcall bufferptr
         lea     edi, [ebx + device.tx_ring + eax]
 
         ; Check if current descriptor is free or still in use
-        cmp     [edi + txfd.status], 0
+        cmp     [edi + txfd.virt_addr], 0               ; buffer not reclaimed yet
         jne     .overrun
 
         ; Fill in status and command values
@@ -786,6 +809,22 @@ proc transmit stdcall bufferptr
         mov     eax, edi
         invoke  GetPhysAddr
         set_io  [ebx + device.io_addr], 0
+
+        ; CU_START is only valid while the CU is idle or suspended:
+        ; wait for the previous frame to leave (1.2 ms at 10 Mbit)
+        push    eax ecx
+        set_io  [ebx + device.io_addr], REG_SCB_STATUS
+        mov     ecx, 100000
+  .cu_busy:
+        in      al, dx
+        and     al, SCB_STATUS_CUS
+        cmp     al, CU_STATUS_ACTIVE
+        jne     .cu_free
+        dec     ecx
+        jnz     .cu_busy
+        DEBUGF  2, "CU still active, starting anyway\n"
+  .cu_free:
+        pop     ecx eax
         set_io  [ebx + device.io_addr], REG_SCB_PTR
         out     dx, eax
 
@@ -854,7 +893,7 @@ int_handler:
 
         DEBUGF  1,"Status: %x\n", ax
 
-        test    ax, SCB_STATUS_FR               ; did we receive a frame?
+        test    ax, SCB_STATUS_FR or SCB_STATUS_RNR     ; frame received or receiver stopped?
         jz      .no_rx
 
         push    ax
@@ -901,6 +940,7 @@ int_handler:
         mov     [esi + sizeof.NET_BUFF + rxfd.command], RXFD_CMD_EL or RXFD_CMD_SUSPEND
         mov     [esi + sizeof.NET_BUFF + rxfd.link], eax
         mov     [esi + sizeof.NET_BUFF + rxfd.count], 0
+        mov     [esi + sizeof.NET_BUFF + rxfd.rx_buf_addr], 0xffffffff  ; simplified mode, no RBD
         mov     [esi + sizeof.NET_BUFF + rxfd.size], 1528
 
 ; restart RX
@@ -916,11 +956,22 @@ int_handler:
         mov     ax, RX_START
         out     dx, ax
         call    cmd_wait
-  .out_of_mem:
 
 ; Hand the frame over to the kernel
         jmp     [EthInput]
 
+  .out_of_mem:
+; The buffer is still our only RFD: drop the frame and reuse it
+        mov     esi, [esp]                      ; the buffer
+        mov     ecx, [esi + NET_BUFF.length]
+        dec     [ebx + device.packets_rx]       ; counted above, before the allocation
+        sub     dword [ebx + device.bytes_rx], ecx
+        sbb     dword [ebx + device.bytes_rx + 4], 0
+        inc     [ebx + device.packets_rx_drop]
+        add     esp, 12                         ; buffer, .rx_loop, ebx
+        mov     esi, [ebx + device.rx_desc]
+        jmp     .not_ok
+
   .not_ok:
 ; Reset the FD
         mov     [esi + sizeof.NET_BUFF + rxfd.status], 0
@@ -945,6 +996,27 @@ int_handler:
 
   .no_rx_:
         DEBUGF  1, "no more data\n"
+
+; With a single RFD the receiver sits in No Resources/Suspended whenever a
+; frame arrived before we re-armed it. Nothing restarts it but us.
+        set_io  [ebx + device.io_addr], 0
+        set_io  [ebx + device.io_addr], REG_SCB_STATUS
+        in      al, dx
+        and     al, SCB_STATUS_RUS
+        cmp     al, RU_STATUS_READY
+        je      .ru_ready
+        test    byte[esp], 1                    ; bit 0 of the saved status is reserved,
+        jnz     .ru_ready                       ; we mark it: one restart per IRQ
+        or      byte[esp], 1
+        movzx   eax, al
+        DEBUGF  1, "Restarting receiver, RU status %x\n", eax:2
+        mov     esi, [ebx + device.rx_desc]
+        push    ebx
+        test    [esi + sizeof.NET_BUFF + rxfd.status], RXFD_STATUS_C
+        jnz     .rx_loop                        ; a frame completed meanwhile
+        add     esp, 4
+        jmp     .not_ok                         ; re-arm the empty RFD
+  .ru_ready:
         pop     ax
 
   .no_rx:
@@ -955,13 +1027,13 @@ int_handler:
 
         push    eax
   .loop_tx:
-        mov     edi, [ebx + device.last_tx]
-        mov     eax, sizeof.txfd
-        mul     eax
+        mov     eax, [ebx + device.last_tx]
+        mov     edx, sizeof.txfd
+        mul     edx
         lea     edi, [ebx + device.tx_ring + eax]
 
-        cmp     [edi + txfd.status], 0
-        je      .tx_done
+        test    [edi + txfd.status], TXFD_STATUS_C      ; sent by the device?
+        jz      .tx_done
 
         cmp     [edi + txfd.virt_addr], 0
         je      .tx_done
@@ -970,6 +1042,7 @@ int_handler:
 
         push    [edi + txfd.virt_addr]
         mov     [edi + txfd.virt_addr], 0
+        mov     [edi + txfd.status], 0                  ; free for transmit again
         invoke  NetFree
 
         inc     [ebx + device.last_tx]
@@ -980,12 +1053,6 @@ int_handler:
         pop     eax
   .no_tx:
 
-        test    ax, RU_STATUS_NO_RESOURCES
-        jz      .not_out_of_resources
-
-        DEBUGF  2, "Out of resources!\n"
-
-  .not_out_of_resources:
         pop     edi esi ebx
         xor     eax, eax
         inc     eax
@@ -1079,6 +1146,25 @@ proc check_media_mii stdcall dev:dword
 endp
 
 
+; Wait (up to 1 s) until the device has executed the command in confcmd.
+align 4
+confcmd_wait:
+        mov     ecx, 1000                       ; udelay = Sleep(1), 1 ms
+  .loop:
+        test    [ebx + device.confcmd.status], TXFD_STATUS_C
+        jnz     .done
+        call    udelay
+        dec     ecx
+        jnz     .loop
+        movzx   eax, [ebx + device.confcmd.command]
+        DEBUGF  2, "Command 0x%x not completed\n", eax:4
+        ret
+  .done:
+        movzx   eax, [ebx + device.confcmd.command]
+        movzx   ecx, [ebx + device.confcmd.status]
+        DEBUGF  1, "Command 0x%x done, status 0x%x\n", eax:4, ecx:4
+        ret
+
 align 4
 cmd_wait:
 

kernel/trunk/core/memory.inc

@@ -1233,25 +1233,27 @@ f68:
         mov     eax, [ecx + SRV.entry]
         test    eax, eax
         jz      .fail
-        push    ecx    ;save handl
+        mov     eax, -2 ;error free RAM
+        cmp     [ecx + SRV.base], 0
+        je      .fail
 
+        push    ecx         ;save handl
+        mov     ebp, esp    ;START may be cdecl or stdcall
         push    edx         ;cmdline
         push    DRV_EXIT
-        call    eax   ;the result is not checked
-        lea     esp, [esp+8]
-        push    ecx
+        call    [ecx + SRV.entry]   ;the result is not checked
+        mov     esp, ebp
+        pop     esi
 
-        mov     ebx, [ecx + SRV.base]
-        mov     eax, -2 ;error free RAM
-        test    ebx, ebx
-        jz      .fail
-        stdcall kernel_free, ebx ;del driver
-
-        mov     eax, [ecx + SRV.fd]
-        mov     edx, [ecx + SRV.bk]
+        mov     eax, [esi + SRV.fd]
+        mov     edx, [esi + SRV.bk]
         mov     [edx + SRV.fd], eax
         mov     [eax + SRV.bk], edx
-        stdcall free, ecx
+        push    eax
+        stdcall kernel_free, [esi + SRV.base] ;del driver
+        mov     eax, esi
+        call    free        ;eax = mem
+        pop     eax
         ;dec     [count_services]
 
         mov     [esp + SYSCALL_STACK.eax], eax

kernel/trunk/network/ARP.inc

@@ -83,6 +83,31 @@ macro arp_init {
 
 }
 
+;-----------------------------------------------------------------;
+;                                                                 ;
+; arp_clear_device: Empty the ARP table and counters of a removed ;
+;                   device.                                       ;
+;                                                                 ;
+;  IN:  edi = device number * 4                                   ;
+;                                                                 ;
+;-----------------------------------------------------------------;
+macro arp_clear_device {
+
+        xor     eax, eax
+        mov     [ARP_entries + edi], eax
+        mov     [ARP_packets_tx + edi], eax
+        mov     [ARP_packets_rx + edi], eax
+        mov     [ARP_conflicts + edi], eax
+
+        push    edi
+        imul    edi, (ARP_TABLE_SIZE * sizeof.ARP_entry)/4
+        add     edi, ARP_table
+        mov     ecx, (ARP_TABLE_SIZE * sizeof.ARP_entry)/2
+        rep stosw
+        pop     edi
+
+}
+
 ;-----------------------------------------------------------------;
 ;                                                                 ;
 ; arp_decrease_entry_ttls                                         ;
@@ -313,6 +338,9 @@ arp_output_request:
         DEBUGF  DEBUG_NETWORK_VERBOSE, "ARP_output_request: ip=%u.%u.%u.%u device=0x%x\n",\
         [esp]:1, [esp + 1]:1, [esp + 2]:1, [esp + 3]:1, ebx
 
+        test    ebx, ebx                        ; device is gone
+        jz      .exit
+
         mov     ax, ETHER_PROTO_ARP
         mov     ecx, sizeof.ARP_header
         mov     edx, ETH_BROADCAST              ; broadcast mac

kernel/trunk/network/IPv4.inc

@@ -105,6 +105,29 @@ macro   ipv4_init {
 }
 
 
+;-----------------------------------------------------------------;
+;                                                                 ;
+; ipv4_clear_device: Reset the IPv4 settings of a removed device, ;
+;                    so no route points to it anymore.            ;
+;                                                                 ;
+;  IN:  edi = device number * 4                                   ;
+;                                                                 ;
+;-----------------------------------------------------------------;
+macro   ipv4_clear_device {
+
+        xor     eax, eax
+        mov     [IPv4_address + edi], eax
+        mov     [IPv4_subnet + edi], eax
+        mov     [IPv4_nameserver + edi], eax
+        mov     [IPv4_gateway + edi], eax
+        mov     [IPv4_broadcast + edi], eax
+        mov     [IPv4_packets_tx + edi], eax
+        mov     [IPv4_packets_rx + edi], eax
+        mov     [IPv4_packets_dumped + edi], eax
+
+}
+
+
 ;-----------------------------------------------------------------;
 ;                                                                 ;
 ; Decrease TimeToLive of all fragment slots                       ;
@@ -757,6 +780,8 @@ ipv4_output_raw:
         push    esi eax
 
         call    ipv4_route
+        test    eax, eax
+        jz      .arp_error
         call    arp_ip_to_mac
 
         test    eax, 0xffff0000         ; error bits
@@ -967,6 +992,8 @@ ipv4_route:
 ; Check for on-link
         xor     edi, edi
   .loop:
+        cmp     [net_device_list + edi], 0      ; skip slots without a device
+        je      .next
         mov     ebx, [IPv4_address + edi]
         and     ebx, [IPv4_subnet + edi]
         jz      .next
@@ -983,7 +1010,10 @@ ipv4_route:
         mov     edi, 4                  ; skip loopback device
   .loop_gw:
         cmp     [IPv4_gateway + edi], 0
+        je      .next_gw
+        cmp     [net_device_list + edi], 0
         jne     .found_gw
+  .next_gw:
         add     edi, 4
         cmp     edi, 4*NET_DEVICES_MAX
         jb      .loop_gw
@@ -1049,14 +1079,23 @@ ipv4_route:
         cmp     eax, 0xffffffff
         je      @f
 
+; An interface without an address (DHCP not done yet, or the lease
+; dropped when the link went down) can send nothing but broadcasts: with
+; address and mask both zero every destination compared as on-link, and
+; the stack went asking ARP for internet addresses on behalf of 0.0.0.0.
+        test    edx, edx
+        jz      .fail
+
 ; Check if we should route to gateway or not
-        mov     ebx, [IPv4_address + edi]
+        mov     ebx, edx
         and     ebx, [IPv4_subnet + edi]
         mov     ecx, eax
         and     ecx, [IPv4_subnet + edi]
         cmp     ecx, ebx
         je      @f
         mov     eax, [IPv4_gateway + edi]
+        test    eax, eax
+        jz      .fail                                ; off-link and no gateway
   @@:
         DEBUGF  DEBUG_NETWORK_VERBOSE, "IPv4_route: %u\n", edi
         ret

kernel/trunk/network/ethernet.inc

@@ -61,6 +61,41 @@ macro   eth_init {
 
 }
 
+;-----------------------------------------------------------------;
+;                                                                 ;
+; eth_clear_device: Drop the frames of a removed device that are  ;
+;                   still waiting in the input queue, their       ;
+;                   NET_BUFF.device will not be valid anymore.    ;
+;                                                                 ;
+;  IN:  ebx = device ptr                                          ;
+;                                                                 ;
+;-----------------------------------------------------------------;
+macro   eth_clear_device {
+
+local   .loop, .done
+
+        spin_lock_irqsave
+        mov     esi, [ETH_frame_head]
+  .loop:
+        cmp     esi, ETH_frame_head
+        je      .done
+        mov     eax, esi
+        mov     esi, [esi + NET_BUFF.NextPtr]
+        cmp     [eax + NET_BUFF.device], ebx
+        jne     .loop
+
+; unlink it, ETH_frame_head/tail double as the NextPtr/PrevPtr of the list head
+        mov     ecx, [eax + NET_BUFF.PrevPtr]
+        mov     [ecx + NET_BUFF.NextPtr], esi
+        mov     [esi + NET_BUFF.PrevPtr], ecx
+        dec     [ETH_frame_queued]
+        stdcall net_buff_free, eax
+        jmp     .loop
+  .done:
+        spin_unlock_irqrestore
+
+}
+
 align 4
 ; This function is called by ethernet drivers.
 ; Push the received ethernet packet onto the ethernet input queue.
@@ -245,6 +280,9 @@ eth_output:
 
         DEBUGF  DEBUG_NETWORK_VERBOSE, "ETH_output: size=%u device=%x\n", ecx, ebx
 
+        test    ebx, ebx
+        jz      .no_device
+
         cmp     ecx, [ebx + ETH_DEVICE.mtu]
         ja      .too_large
 
@@ -295,11 +333,16 @@ eth_output:
         ret
 
   .too_large:
-        inc     [eax + NET_DEVICE.packets_tx_err]
+        inc     [ebx + NET_DEVICE.packets_tx_err]
         DEBUGF  DEBUG_NETWORK_VERBOSE, "ETH_output: Packet too large!\n"
         xor     eax, eax
         ret
 
+  .no_device:
+        DEBUGF  DEBUG_NETWORK_ERROR, "ETH_output: no device!\n"
+        xor     eax, eax
+        ret
+
 
 
 ;-----------------------------------------------------------------;

kernel/trunk/network/stack.inc

@@ -597,6 +597,16 @@ net_remove_device:
         mov     dword [edi-4], eax
         dec     [net_device_count]
 
+;-----------------------------------------------------------------
+; Forget everything the protocols still know about this interface.
+; Otherwise its IP address keeps matching in ipv4_route, and output
+; ends up calling through the now empty net_device_list slot.
+
+        sub     edi, net_device_list + 4        ; device number * 4
+        ipv4_clear_device
+        arp_clear_device
+        eth_clear_device
+
         call    net_send_event
 
         xor     eax, eax

kernel/trunk/network/tcp_input.inc

@@ -1422,10 +1422,14 @@ endl
         test    [edx + TCP_header.Flags], TH_ACK
         jz      .drop
 
-        mov     eax, ebx
+        push    ebx
+        lea     ecx, [ebx + SOCKET.mutex]
+        call    mutex_unlock
+        pop     eax
+
         mov     ebx, ECONNREFUSED
         call    tcp_drop
-        jmp     .drop
+        jmp     .drop_no_socket
        @@:
 
 ;-----------------------------------------------------------------------------------

kernel/trunk/network/tcp_output.inc

@@ -234,23 +234,25 @@ endl
         mov     ebx, TCP_max_win
         shl     ebx, cl
         pop     ecx
-        sub     ebx, [eax + TCP_SOCKET.RCV_ADV]
-        add     ebx, [eax + TCP_SOCKET.RCV_NXT]
+
+; Clamp to available window first, then subtract what the peer already knows (as in BSD)
 
         cmp     ebx, ecx
-        jl      @f
+        jbe     @f
         mov     ebx, ecx
        @@:
+        sub     ebx, [eax + TCP_SOCKET.RCV_ADV]
+        add     ebx, [eax + TCP_SOCKET.RCV_NXT]
 
         DEBUGF  DEBUG_NETWORK_VERBOSE, "TCP_output: we can increase window by %d bytes\n", ebx
 
         mov     edi, [eax + TCP_SOCKET.t_maxseg]
         shl     edi, 1
-        cmp     ebx, edi
-        jae     .send
+        cmp     ebx, edi                                ; signed: may be negative if window shrank
+        jge     .send
 
         cmp     ebx, SOCKET_BUFFER_SIZE/2
-        jae     .send
+        jge     .send
 
   .no_window:
 

kernel/trunk/network/tcp_usreq.inc

@@ -180,6 +180,8 @@ tcp_connect:
 
   .nomem:
         pop     edx eax
+        lea     ecx, [eax + SOCKET.mutex]
+        call    mutex_unlock
         xor     eax, eax
         dec     eax
         mov     ebx, ENOMEM
@@ -194,6 +196,8 @@ tcp_connect:
   .enoroute:
         pop     eax
         popa
+        lea     ecx, [eax + SOCKET.mutex]
+        call    mutex_unlock
         xor     eax, eax
         dec     eax
         mov     ebx, EADDRNOTAVAIL

programs/develop/libraries/http/http.asm

@@ -1618,21 +1618,23 @@ endl
 
 ; Open a new TCP socket
         mcall   socket, AF_INET4, SOCK_STREAM, 0
-        test    eax, eax
-        jz      .error3
+        cmp     eax, -1
+        je      .error3
         mov     [socketnum], eax
         DEBUGF  1, "Socket: 0x%x\n", eax
 
 ; Connect to the server
         mcall   connect, [socketnum], [sockaddr], 18
         test    eax, eax
-        jnz     .error3
+        jnz     .error4
         DEBUGF  1, "Socket is now connected.\n"
 
         invoke  freeaddrinfo            ; Free allocated memory
         mov     eax, [socketnum]
         ret
 
+  .error4:
+        mcall   close, [socketnum]
   .error3:
         DEBUGF  2, "Could not connect to the remote server\n"
         invoke  freeaddrinfo            ; Free allocated memory

programs/develop/libraries/libs-dev/libimg/jpeg/jpeg.asm

@@ -501,7 +501,8 @@ img.decode.jpg:
 	jnz	.no_create_palette
 	push	ecx edi
 	mov	edi, [eax + Image.Palette]
-	xor	eax, eax
+	mov	eax, 0xFF000000	; opaque, like the palettes of the other decoders:
+				; converted to 32 bpp, alpha 0 made the picture invisible
 	mov	ecx, 256
 @@:
 	stosd