Kernel/Networking: Modify outgoing packet in place

This avoids creation of new buffers and unnecessary memcpys. The old
interface still exists for UDP but will be removed when I am less lazy.

From testing it does not appear to have any performance improvement but
this is most likely due to other bottlenecks as extra copies should
always be worse.

7 files changed, 126 insertions, 54 deletions

diff --git a/kernel/drivers/rtl8139.c b/kernel/drivers/rtl8139.c
index 669cb6c..946c405 100644
--- a/kernel/drivers/rtl8139.c
+++ b/kernel/drivers/rtl8139.c
@@ -111,6 +111,24 @@ void rtl8139_handler(void *regs) {
   EOI(0xB);
 }
 
+u8 *nic_get_buffer(void) {
+  if (send_buffers_loop > 3) {
+    send_buffers_loop = 0;
+  }
+  return send_buffers[send_buffers_loop];
+}
+
+void nic_send_buffer(u16 data_size) {
+  const struct PCI_DEVICE *device = &rtl8139;
+  outl(device->gen.base_mem_io + 0x20 + send_buffers_loop * 4,
+       (u32)virtual_to_physical(send_buffers[send_buffers_loop], NULL));
+  outl(device->gen.base_mem_io + 0x10 + send_buffers_loop * 4, data_size);
+  send_buffers_loop += 1;
+  if (send_buffers_loop > 3) {
+    send_buffers_loop = 0;
+  }
+}
+
 void rtl8139_send_data(u8 *data, u16 data_size) {
   if (data_size > 0x1000) {
     rtl8139_send_data(data, 0x1000);
diff --git a/kernel/drivers/rtl8139.h b/kernel/drivers/rtl8139.h
index c61e6de..ebefd77 100644
--- a/kernel/drivers/rtl8139.h
+++ b/kernel/drivers/rtl8139.h
@@ -2,3 +2,5 @@
 void get_mac_address(u8 mac[6]);
 void rtl8139_init(void);
 void rtl8139_send_data(u8 *data, u16 data_size);
+u8 *nic_get_buffer(void);
+void nic_send_buffer(u16 data_size);
diff --git a/kernel/network/ethernet.c b/kernel/network/ethernet.c
index dbd1b6d..fd385b3 100644
--- a/kernel/network/ethernet.c
+++ b/kernel/network/ethernet.c
@@ -9,12 +9,6 @@
 #include <stdio.h>
 #include <string.h>
 
-struct EthernetHeader {
-  u8 mac_dst[6];
-  u8 mac_src[6];
-  u16 type;
-};
-
 u32 crc32(const char *buf, size_t len) {
   static u32 table[256];
   static int have_table = 0;
@@ -68,6 +62,27 @@ void handle_ethernet(const u8 *packet, u64 packet_length) {
   }
 }
 
+void send_ethernet_packet2(u8 mac_dst[6], u16 type, u64 payload_length) {
+  u64 buffer_size =
+      sizeof(struct EthernetHeader) + payload_length + sizeof(u32);
+  assert(buffer_size < 0x1000);
+
+  u8 *buffer = nic_get_buffer();
+  struct EthernetHeader *eth_header = (struct EthernetHeader *)buffer;
+  memset(eth_header, 0, sizeof(struct EthernetHeader));
+  buffer += sizeof(struct EthernetHeader);
+  buffer += payload_length;
+
+  memcpy(eth_header->mac_dst, mac_dst, sizeof(u8[6]));
+  get_mac_address(eth_header->mac_src);
+  eth_header->type = htons(type);
+  *(u32 *)(buffer) =
+      htonl(crc32((const char *)nic_get_buffer(), buffer_size - 4));
+
+  nic_send_buffer(buffer_size);
+  //  rtl8139_send_data(ethernet_buffer, buffer_size);
+}
+
 u8 ethernet_buffer[0x1000];
 void send_ethernet_packet(u8 mac_dst[6], u16 type, u8 *payload,
                           u64 payload_length) {
diff --git a/kernel/network/ethernet.h b/kernel/network/ethernet.h
index ffb7893..6534a64 100644
--- a/kernel/network/ethernet.h
+++ b/kernel/network/ethernet.h
@@ -1,5 +1,15 @@
+#ifndef ETHERNET_H
+#define ETHERNET_H
 #include <typedefs.h>
 
+struct EthernetHeader {
+  u8 mac_dst[6];
+  u8 mac_src[6];
+  u16 type;
+};
+
 void handle_ethernet(const u8 *packet, u64 packet_length);
+void send_ethernet_packet2(u8 mac_dst[6], u16 type, u64 payload_length);
 void send_ethernet_packet(u8 mac_dst[6], u16 type, u8 *payload,
                           u64 payload_length);
+#endif //  ETHERNET_H
diff --git a/kernel/network/ipv4.c b/kernel/network/ipv4.c
index 6074141..de6984a 100644
--- a/kernel/network/ipv4.c
+++ b/kernel/network/ipv4.c
@@ -1,5 +1,6 @@
 #include <assert.h>
 #include <drivers/pit.h>
+#include <drivers/rtl8139.h>
 #include <kmalloc.h>
 #include <network/arp.h>
 #include <network/bytes.h>
@@ -26,6 +27,37 @@ static u16 ip_checksum(const u16 *data, u16 length) {
   return htons(~acc);
 }
 
+void send_ipv4_packet2(ipv4_t ip, u8 protocol, u16 length) {
+  u16 header[10];
+  header[0] = (4 /*version*/ << 4) | (5 /*IHL*/);
+
+  header[1] = htons(length + 20);
+
+  header[2] = 0;
+
+  header[3] = 0;
+  header[4] = (protocol << 8) | 0xF8;
+
+  header[5] = 0;
+  header[6] = (ip_address.d >> 0) & 0xFFFF;
+  header[7] = (ip_address.d >> 16) & 0xFFFF;
+
+  header[8] = (ip.d >> 0) & 0xFFFF;
+  header[9] = (ip.d >> 16) & 0xFFFF;
+
+  header[5] = ip_checksum(header, 20);
+  u16 packet_length = length + 20;
+  // TODO
+  //  assert(packet_length < sizeof(ipv4_buffer));
+  u8 *packet = nic_get_buffer() + sizeof(struct EthernetHeader);
+  memcpy(packet, header, 20);
+
+  u8 mac[6];
+  for (; !get_mac_from_ip(ip, mac);)
+    ;
+  send_ethernet_packet2(mac, 0x0800, packet_length);
+}
+
 u8 ipv4_buffer[0x1000];
 void send_ipv4_packet(ipv4_t ip, u8 protocol, const u8 *payload, u16 length) {
   u16 header[10];
diff --git a/kernel/network/ipv4.h b/kernel/network/ipv4.h
index 7b44eb1..4d19ae9 100644
--- a/kernel/network/ipv4.h
+++ b/kernel/network/ipv4.h
@@ -2,3 +2,4 @@
 
 void handle_ipv4(const u8 *payload, u32 packet_length);
 void send_ipv4_packet(ipv4_t ip, u8 protocol, const u8 *payload, u16 length);
+void send_ipv4_packet2(ipv4_t ip, u8 protocol, u16 length);
diff --git a/kernel/network/tcp.c b/kernel/network/tcp.c
index 8717250..e608f6c 100644
--- a/kernel/network/tcp.c
+++ b/kernel/network/tcp.c
@@ -1,11 +1,13 @@
 #include <assert.h>
 #include <cpu/arch_inst.h>
 #include <drivers/pit.h>
+#include <drivers/rtl8139.h>
 #include <fs/vfs.h>
 #include <interrupts.h>
 #include <math.h>
 #include <network/arp.h>
 #include <network/bytes.h>
+#include <network/ethernet.h>
 #include <network/ipv4.h>
 #include <network/tcp.h>
 #include <network/udp.h>
@@ -82,43 +84,37 @@ u16 tcp_calculate_checksum(ipv4_t src_ip, u32 dst_ip, const u8 *payload,
   return htons(tcp_checksum_final(tmp, payload, payload_length));
 }
 
-static void tcp_send(struct TcpConnection *con, u8 *buffer, u16 length,
-                     u32 seq_num, u32 payload_length) {
-  send_ipv4_packet((ipv4_t){.d = con->outgoing_ip}, 6, buffer, length);
+static u8 *tcp_get_buffer(void) {
+  u8 *nb = nic_get_buffer();
+  return nb + sizeof(struct EthernetHeader) + sizeof(u16[10]);
 }
 
-u8 tcp_buffer[0x1000];
 void tcp_send_empty_payload(struct TcpConnection *con, u8 flags) {
-  struct TcpHeader header;
-  memset(&header, 0, sizeof(header));
-  header.src_port = htons(con->incoming_port);
-  header.dst_port = htons(con->outgoing_port);
-  header.seq_num = htonl(con->snd_nxt);
+  struct TcpHeader *header = (struct TcpHeader *)tcp_get_buffer();
+  memset(header, 0, sizeof(struct TcpHeader));
+  header->src_port = htons(con->incoming_port);
+  header->dst_port = htons(con->outgoing_port);
+  header->seq_num = htonl(con->snd_nxt);
   if (flags & ACK) {
     con->should_send_ack = 0;
-    header.ack_num = htonl(con->rcv_nxt);
+    header->ack_num = htonl(con->rcv_nxt);
   } else {
-    header.ack_num = 0;
+    header->ack_num = 0;
   }
-  header.data_offset = 5;
-  header.reserved = 0;
-  header.flags = flags;
-  header.window_size = htons(con->rcv_wnd);
-  header.urgent_pointer = 0;
+  header->data_offset = 5;
+  header->reserved = 0;
+  header->flags = flags;
+  header->window_size = htons(con->rcv_wnd);
+  header->urgent_pointer = 0;
 
   u8 payload[] = {0};
   u16 payload_length = 0;
-  header.checksum = tcp_calculate_checksum(
-      ip_address, con->outgoing_ip, (const u8 *)payload, payload_length,
-      &header, sizeof(struct TcpHeader) + payload_length);
-  u32 send_len = sizeof(header) + payload_length;
+  header->checksum = tcp_calculate_checksum(
+      ip_address, con->outgoing_ip, (const u8 *)payload, payload_length, header,
+      sizeof(struct TcpHeader) + payload_length);
+  u32 send_len = sizeof(struct TcpHeader) + payload_length;
 
-  assert(send_len < sizeof(tcp_buffer));
-  u8 *send_buffer = tcp_buffer;
-  memcpy(send_buffer, &header, sizeof(header));
-  memcpy(send_buffer + sizeof(header), payload, payload_length);
-
-  tcp_send(con, send_buffer, send_len, con->snd_nxt, 0);
+  send_ipv4_packet2((ipv4_t){.d = con->outgoing_ip}, 6, send_len);
 
   con->snd_nxt += (flags & SYN) ? 1 : 0;
   con->snd_nxt += (flags & FIN) ? 1 : 0;
@@ -144,7 +140,7 @@ void tcp_close_connection(struct TcpConnection *con) {
     // FIXME:
     // Book says it should be FIN but observed network traffic says it
     // should be FIN|ACK?
-    tcp_send_empty_payload(con, FIN);
+    tcp_send_empty_payload(con, FIN | ACK);
     con->state = TCP_STATE_FIN_WAIT1;
     return;
   }
@@ -181,30 +177,28 @@ int send_tcp_packet(struct TcpConnection *con, const u8 *payload,
     payload += len;
     return send_tcp_packet(con, payload, payload_length);
   }
-  struct TcpHeader header = {0};
-  header.src_port = htons(con->incoming_port);
-  header.dst_port = htons(con->outgoing_port);
-  header.seq_num = htonl(con->snd_nxt);
-  header.ack_num = htonl(con->rcv_nxt);
-  header.data_offset = 5;
-  header.reserved = 0;
-  header.flags = PSH | ACK;
-  header.window_size = htons(con->rcv_wnd);
-  header.urgent_pointer = 0;
-  header.checksum = tcp_calculate_checksum(
-      ip_address, con->outgoing_ip, (const u8 *)payload, payload_length,
-      &header, sizeof(struct TcpHeader) + payload_length);
-  u32 send_len = sizeof(header) + payload_length;
-  assert(send_len < sizeof(tcp_buffer));
-  u8 *send_buffer = tcp_buffer;
-  memcpy(send_buffer, &header, sizeof(header));
-  memcpy(send_buffer + sizeof(header), payload, payload_length);
-
-  if (header.flags & ACK) {
+  struct TcpHeader *header = (struct TcpHeader *)tcp_get_buffer();
+  header->src_port = htons(con->incoming_port);
+  header->dst_port = htons(con->outgoing_port);
+  header->seq_num = htonl(con->snd_nxt);
+  header->ack_num = htonl(con->rcv_nxt);
+  header->data_offset = 5;
+  header->reserved = 0;
+  header->flags = PSH | ACK;
+  header->window_size = htons(con->rcv_wnd);
+  header->urgent_pointer = 0;
+  header->checksum = tcp_calculate_checksum(
+      ip_address, con->outgoing_ip, (const u8 *)payload, payload_length, header,
+      sizeof(struct TcpHeader) + payload_length);
+  u32 send_len = sizeof(struct TcpHeader) + payload_length;
+  u8 *send_buffer = tcp_get_buffer();
+  memcpy(send_buffer + sizeof(struct TcpHeader), payload, payload_length);
+
+  if (header->flags & ACK) {
     con->should_send_ack = 0;
   }
 
-  tcp_send(con, send_buffer, send_len, con->snd_nxt, payload_length);
+  send_ipv4_packet2((ipv4_t){.d = con->outgoing_ip}, 6, send_len);
 
   con->snd_nxt += payload_length;
   con->snd_max = max(con->snd_nxt, con->snd_max);