[PATCH v3 2/9] net/tcp: fix selective acknowledge
Mikhail Kshevetskiy
mikhail.kshevetskiy at iopsys.eu
Wed Jul 31 11:21:03 CEST 2024
Current code assume that all (except last) packets are of the same size.
This is definitely wrong. Replace SACK code with a new one, that does
not rely on this assumption. Also this code uses less memory.
Signed-off-by: Mikhail Kshevetskiy <mikhail.kshevetskiy at iopsys.eu>
---
net/tcp.c | 200 +++++++++++++++++++++++-------------------------------
1 file changed, 86 insertions(+), 114 deletions(-)
diff --git a/net/tcp.c b/net/tcp.c
index 3e3118de450..724536cb352 100644
--- a/net/tcp.c
+++ b/net/tcp.c
@@ -38,21 +38,6 @@ static u32 tcp_ack_edge;
static int tcp_activity_count;
-/*
- * Search for TCP_SACK and review the comments before the code section
- * TCP_SACK is the number of packets at the front of the stream
- */
-
-enum pkt_state {PKT, NOPKT};
-struct sack_r {
- struct sack_edges se;
- enum pkt_state st;
-};
-
-static struct sack_r edge_a[TCP_SACK];
-static unsigned int sack_idx;
-static unsigned int prev_len;
-
/*
* TCP lengths are stored as a rounded up number of 32 bit words.
* Add 3 to length round up, rounded, then divided into the
@@ -69,6 +54,11 @@ static enum tcp_state current_tcp_state;
/* Current TCP RX packet handler */
static rxhand_tcp *tcp_packet_handler;
+static inline s32 tcp_seq_cmp(u32 a, u32 b)
+{
+ return (s32)(a - b);
+}
+
/**
* tcp_get_tcp_state() - get current TCP state
*
@@ -267,6 +257,7 @@ int tcp_set_tcp_header(uchar *pkt, int dport, int sport, int payload_len,
action = TCP_FIN;
current_tcp_state = TCP_FIN_WAIT_1;
} else {
+ tcp_lost.len = TCP_OPT_LEN_2;
current_tcp_state = TCP_SYN_SENT;
}
break;
@@ -353,6 +344,20 @@ int tcp_set_tcp_header(uchar *pkt, int dport, int sport, int payload_len,
return pkt_hdr_len;
}
+static void tcp_update_ack_edge(void)
+{
+ if (tcp_seq_cmp(tcp_ack_edge, tcp_lost.hill[0].l) >= 0) {
+ tcp_ack_edge = tcp_lost.hill[0].r;
+
+ memmove(&tcp_lost.hill[0], &tcp_lost.hill[1],
+ (TCP_SACK_HILLS - 1) * sizeof(struct sack_edges));
+
+ tcp_lost.len -= TCP_OPT_LEN_8;
+ tcp_lost.hill[TCP_SACK_HILLS - 1].l = TCP_O_NOP;
+ tcp_lost.hill[TCP_SACK_HILLS - 1].r = TCP_O_NOP;
+ }
+}
+
/**
* tcp_hole() - Selective Acknowledgment (Essential for fast stream transfer)
* @tcp_seq_num: TCP sequence start number
@@ -360,106 +365,79 @@ int tcp_set_tcp_header(uchar *pkt, int dport, int sport, int payload_len,
*/
void tcp_hole(u32 tcp_seq_num, u32 len)
{
- u32 idx_sack, sack_in;
- u32 sack_end = TCP_SACK - 1;
- u32 hill = 0;
- enum pkt_state expect = PKT;
- u32 seq = tcp_seq_num - tcp_seq_init;
- u32 hol_l = tcp_ack_edge - tcp_seq_init;
- u32 hol_r = 0;
-
- /* Place new seq number in correct place in receive array */
- if (prev_len == 0)
- prev_len = len;
-
- idx_sack = sack_idx + ((tcp_seq_num - tcp_ack_edge) / prev_len);
- if (idx_sack < TCP_SACK) {
- edge_a[idx_sack].se.l = tcp_seq_num;
- edge_a[idx_sack].se.r = tcp_seq_num + len;
- edge_a[idx_sack].st = PKT;
+ int i, j, cnt, cnt_move;
- /*
- * The fin (last) packet is not the same length as data
- * packets, and if it's length is recorded and used for
- * array index calculation, calculation breaks.
- */
- if (prev_len < len)
- prev_len = len;
- }
+ cnt = (tcp_lost.len - TCP_OPT_LEN_2) / TCP_OPT_LEN_8;
+ for (i = 0; i < cnt; i++) {
+ if (tcp_seq_cmp(tcp_lost.hill[i].r, tcp_seq_num) < 0)
+ continue;
+ if (tcp_seq_cmp(tcp_lost.hill[i].l, tcp_seq_num + len) > 0)
+ break;
- debug_cond(DEBUG_DEV_PKT,
- "TCP 1 seq %d, edg %d, len %d, sack_idx %d, sack_end %d\n",
- seq, hol_l, len, sack_idx, sack_end);
+ if (tcp_seq_cmp(tcp_lost.hill[i].l, tcp_seq_num) > 0)
+ tcp_lost.hill[i].l = tcp_seq_num;
+ if (tcp_seq_cmp(tcp_lost.hill[i].l, tcp_seq_num) < 0) {
+ len += tcp_seq_num - tcp_lost.hill[i].l;
+ tcp_seq_num = tcp_lost.hill[i].l;
+ }
+ if (tcp_seq_cmp(tcp_lost.hill[i].r, tcp_seq_num + len) >= 0) {
+ tcp_update_ack_edge();
+ return;
+ }
- /* Right edge of contiguous stream, is the left edge of first hill */
- hol_l = tcp_seq_num - tcp_seq_init;
- hol_r = hol_l + len;
+ /* check overlapping with next hills */
+ cnt_move = 0;
+ tcp_lost.hill[i].r = tcp_seq_num + len;
+ for (j = i + 1; j < cnt; j++) {
+ if (tcp_seq_cmp(tcp_lost.hill[j].l, tcp_lost.hill[i].r) > 0)
+ break;
- if (IS_ENABLED(CONFIG_PROT_TCP_SACK))
- tcp_lost.len = TCP_OPT_LEN_2;
+ tcp_lost.hill[i].r = tcp_lost.hill[j].r;
+ cnt_move++;
+ }
- debug_cond(DEBUG_DEV_PKT,
- "TCP 1 in %d, seq %d, pkt_l %d, pkt_r %d, sack_idx %d, sack_end %d\n",
- idx_sack, seq, hol_l, hol_r, sack_idx, sack_end);
-
- for (sack_in = sack_idx; sack_in < sack_end && hill < TCP_SACK_HILLS;
- sack_in++) {
- switch (expect) {
- case NOPKT:
- switch (edge_a[sack_in].st) {
- case NOPKT:
- debug_cond(DEBUG_INT_STATE, "N");
- break;
- case PKT:
- debug_cond(DEBUG_INT_STATE, "n");
- if (IS_ENABLED(CONFIG_PROT_TCP_SACK)) {
- tcp_lost.hill[hill].l =
- edge_a[sack_in].se.l;
- tcp_lost.hill[hill].r =
- edge_a[sack_in].se.r;
- }
- expect = PKT;
- break;
- }
- break;
- case PKT:
- switch (edge_a[sack_in].st) {
- case NOPKT:
- debug_cond(DEBUG_INT_STATE, "p");
- if (sack_in > sack_idx &&
- hill < TCP_SACK_HILLS) {
- hill++;
- if (IS_ENABLED(CONFIG_PROT_TCP_SACK))
- tcp_lost.len += TCP_OPT_LEN_8;
- }
- expect = NOPKT;
- break;
- case PKT:
- debug_cond(DEBUG_INT_STATE, "P");
-
- if (tcp_ack_edge == edge_a[sack_in].se.l) {
- tcp_ack_edge = edge_a[sack_in].se.r;
- edge_a[sack_in].st = NOPKT;
- sack_idx++;
- } else {
- if (IS_ENABLED(CONFIG_PROT_TCP_SACK) &&
- hill < TCP_SACK_HILLS)
- tcp_lost.hill[hill].r =
- edge_a[sack_in].se.r;
- if (IS_ENABLED(CONFIG_PROT_TCP_SACK) &&
- sack_in == sack_end - 1)
- tcp_lost.hill[hill].r =
- edge_a[sack_in].se.r;
- }
- break;
+ if (cnt_move > 0) {
+ if (cnt > i + cnt_move + 1)
+ memmove(&tcp_lost.hill[i + 1],
+ &tcp_lost.hill[i + cnt_move + 1],
+ cnt_move * sizeof(struct sack_edges));
+
+ cnt -= cnt_move;
+ tcp_lost.len = TCP_OPT_LEN_2 + cnt * TCP_OPT_LEN_8;
+ for (j = cnt; j < TCP_SACK_HILLS; j++) {
+ tcp_lost.hill[j].l = TCP_O_NOP;
+ tcp_lost.hill[j].r = TCP_O_NOP;
}
- break;
}
+
+ tcp_update_ack_edge();
+ return;
}
- debug_cond(DEBUG_INT_STATE, "\n");
- if (!IS_ENABLED(CONFIG_PROT_TCP_SACK) || tcp_lost.len <= TCP_OPT_LEN_2)
- sack_idx = 0;
-}
+
+ if (i == TCP_SACK_HILLS) {
+ tcp_update_ack_edge();
+ return;
+ }
+
+ if (cnt < TCP_SACK_HILLS) {
+ cnt_move = cnt - i;
+ cnt++;
+ } else {
+ cnt = TCP_SACK_HILLS;
+ cnt_move = TCP_SACK_HILLS - i;
+ }
+
+ if (cnt_move > 0)
+ memmove(&tcp_lost.hill[i + 1],
+ &tcp_lost.hill[i],
+ cnt_move * sizeof(struct sack_edges));
+
+ tcp_lost.hill[i].l = tcp_seq_num;
+ tcp_lost.hill[i].r = tcp_seq_num + len;
+ tcp_lost.len = TCP_OPT_LEN_2 + cnt * TCP_OPT_LEN_8;
+
+ tcp_update_ack_edge();
+};
/**
* tcp_parse_options() - parsing TCP options
@@ -509,7 +487,6 @@ static u8 tcp_state_machine(u8 tcp_flags, u32 tcp_seq_num, int payload_len)
u8 tcp_push = tcp_flags & TCP_PUSH;
u8 tcp_ack = tcp_flags & TCP_ACK;
u8 action = TCP_DATA;
- int i;
/*
* tcp_flags are examined to determine TX action in a given state
@@ -536,6 +513,7 @@ static u8 tcp_state_machine(u8 tcp_flags, u32 tcp_seq_num, int payload_len)
action = TCP_SYN | TCP_ACK;
tcp_seq_init = tcp_seq_num;
tcp_ack_edge = tcp_seq_num + 1;
+ tcp_lost.len = TCP_OPT_LEN_2;
current_tcp_state = TCP_SYN_RECEIVED;
} else if (tcp_ack || tcp_fin) {
action = TCP_DATA;
@@ -552,13 +530,7 @@ static u8 tcp_state_machine(u8 tcp_flags, u32 tcp_seq_num, int payload_len)
action |= TCP_ACK;
tcp_seq_init = tcp_seq_num;
tcp_ack_edge = tcp_seq_num + 1;
- sack_idx = 0;
- edge_a[sack_idx].se.l = tcp_ack_edge;
- edge_a[sack_idx].se.r = tcp_ack_edge;
- prev_len = 0;
current_tcp_state = TCP_ESTABLISHED;
- for (i = 0; i < TCP_SACK; i++)
- edge_a[i].st = NOPKT;
if (tcp_syn && tcp_ack)
action |= TCP_PUSH;
--
2.39.2
More information about the U-Boot
mailing list