tcp: fast retransmit pacing
Force pacing for fast retransmit to avoid bursts of retransmitted
packets.
Change-Id: I2ff42c328899b36322c4de557b1f7d853dba8fe2
Signed-off-by: Florin Coras <fcoras@cisco.com>
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index 8c0c5da..bdf9c7a 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -558,10 +558,11 @@
void
tcp_enable_pacing (tcp_connection_t * tc)
{
- u32 max_burst, byte_rate;
- max_burst = 16 * tc->snd_mss;
+ u32 initial_bucket, byte_rate;
+ initial_bucket = 16 * tc->snd_mss;
byte_rate = 2 << 16;
- transport_connection_tx_pacer_init (&tc->connection, byte_rate, max_burst);
+ transport_connection_tx_pacer_init (&tc->connection, byte_rate,
+ initial_bucket);
tc->mrtt_us = (u32) ~ 0;
}
@@ -1318,6 +1319,7 @@
num_threads = 1 /* main thread */ + vtm->n_threads;
vec_validate (tm->connections, num_threads - 1);
+ vec_validate (tm->wrk_ctx, num_threads - 1);
/*
* Preallocate connections. Assume that thread 0 won't
@@ -1339,6 +1341,13 @@
if (preallocated_connections_per_thread)
pool_init_fixed (tm->connections[thread],
preallocated_connections_per_thread);
+ vec_validate (tm->wrk_ctx[thread].pending_fast_rxt, 0);
+ vec_validate (tm->wrk_ctx[thread].ongoing_fast_rxt, 0);
+ vec_validate (tm->wrk_ctx[thread].postponed_fast_rxt, 0);
+ vec_reset_length (tm->wrk_ctx[thread].pending_fast_rxt);
+ vec_reset_length (tm->wrk_ctx[thread].ongoing_fast_rxt);
+ vec_reset_length (tm->wrk_ctx[thread].postponed_fast_rxt);
+ tm->wrk_ctx[thread].vm = vlib_mains[thread];
}
/*
@@ -1358,7 +1367,6 @@
clib_spinlock_init (&tm->half_open_lock);
}
- vec_validate (tm->wrk_ctx, num_threads - 1);
tcp_initialize_timer_wheels (tm);
tm->bytes_per_buffer = vlib_buffer_free_list_buffer_size
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 4ba3d5e..f7424c3 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -160,7 +160,7 @@
};
#define TCP_SCOREBOARD_TRACE (0)
-#define TCP_MAX_SACK_BLOCKS 32 /**< Max number of SACK blocks stored */
+#define TCP_MAX_SACK_BLOCKS 256 /**< Max number of SACK blocks stored */
#define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
typedef struct _scoreboard_trace_elt
@@ -390,6 +390,9 @@
needing fast rxt */
u32 *ongoing_fast_rxt; /**< vector of connections
now doing fast rxt */
+ u32 *postponed_fast_rxt; /**< vector of connections
+ that will do fast rxt */
+ vlib_main_t *vm; /**< pointer to vm */
CLIB_CACHE_LINE_ALIGN_MARK (cacheline1);
u8 cached_opts[40]; /**< cached 'on the wire'
@@ -722,8 +725,8 @@
tcp_cc_rcv_ack (tcp_connection_t * tc)
{
tc->cc_algo->rcv_ack (tc);
- tcp_update_pacer (tc);
tc->tsecr_last_ack = tc->rcv_opts.tsecr;
+ tcp_update_pacer (tc);
}
always_inline void
diff --git a/src/vnet/tcp/tcp_debug.h b/src/vnet/tcp/tcp_debug.h
index 8f626b1..cd4a6f0 100755
--- a/src/vnet/tcp/tcp_debug.h
+++ b/src/vnet/tcp/tcp_debug.h
@@ -627,10 +627,8 @@
#if TCP_DEBUG_CC
-#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, ...) \
+#define TCP_EVT_CC_EVT_PRINT(_tc, _sub_evt) \
{ \
- if (_tc->snd_una != _tc->iss) \
- TCP_EVT_CC_STAT_PRINT (_tc); \
ELOG_TYPE_DECLARE (_e) = \
{ \
.format = "cc: %s snd_space %u snd_una %u out %u flight %u", \
@@ -638,8 +636,8 @@
.n_enum_strings = 7, \
.enum_strings = { \
"fast-rxt", \
- "rxt-timeout", \
"first-rxt", \
+ "rxt-timeout", \
"recovered", \
"congestion", \
"undo", \
@@ -653,8 +651,18 @@
ed->data[3] = tcp_bytes_out(_tc); \
ed->data[4] = tcp_flight_size (_tc); \
}
+
+#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, ...) \
+{ \
+ if (_tc->snd_una != _tc->iss) \
+ TCP_EVT_CC_STAT_PRINT (_tc); \
+ if ((_sub_evt <= 1 && TCP_DEBUG_CC > 1) \
+ || (_sub_evt > 1 && TCP_DEBUG_CC > 0)) \
+ TCP_EVT_CC_EVT_PRINT (_tc, _sub_evt); \
+}
#else
-#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, ...)
+#define TCP_EVT_CC_EVT_HANDLER(_tc, _sub_evt, ...) \
+
#endif
#if TCP_DEBUG_CC > 1
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index ac0e996..154b9ac 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -1199,67 +1199,60 @@
tcp_do_fastretransmits (u32 thread_index)
{
tcp_worker_ctx_t *wrk = &tcp_main.wrk_ctx[thread_index];
- u32 max_burst_size, burst_size, n_segs = 0;
+ u32 max_burst_size, burst_size, n_segs = 0, n_segs_now;
+ u32 *ongoing_fast_rxt, burst_bytes, sent_bytes;
tcp_connection_t *tc;
+ u64 last_cpu_time;
int i;
- if (vec_len (wrk->pending_fast_rxt) == 0)
+ if (vec_len (wrk->pending_fast_rxt) == 0
+ && vec_len (wrk->postponed_fast_rxt) == 0)
return;
- vec_append (wrk->ongoing_fast_rxt, wrk->pending_fast_rxt);
- vec_reset_length (wrk->pending_fast_rxt);
+ last_cpu_time = wrk->vm->clib_time.last_cpu_time;
+ ongoing_fast_rxt = wrk->ongoing_fast_rxt;
+ vec_append (ongoing_fast_rxt, wrk->postponed_fast_rxt);
+ vec_append (ongoing_fast_rxt, wrk->pending_fast_rxt);
+
+ _vec_len (wrk->postponed_fast_rxt) = 0;
+ _vec_len (wrk->pending_fast_rxt) = 0;
max_burst_size = VLIB_FRAME_SIZE / vec_len (wrk->ongoing_fast_rxt);
max_burst_size = clib_max (max_burst_size, 1);
- for (i = 0; i < vec_len (wrk->ongoing_fast_rxt); i++)
+ for (i = 0; i < vec_len (ongoing_fast_rxt); i++)
{
- tc = tcp_connection_get (wrk->ongoing_fast_rxt[i], thread_index);
+ if (n_segs >= VLIB_FRAME_SIZE)
+ {
+ vec_add1 (wrk->postponed_fast_rxt, ongoing_fast_rxt[i]);
+ continue;
+ }
+
+ tc = tcp_connection_get (ongoing_fast_rxt[i], thread_index);
tc->flags &= ~TCP_CONN_FRXT_PENDING;
if (!tcp_in_fastrecovery (tc))
continue;
- /* TODO tx pacer instead of this */
- if (n_segs >= VLIB_FRAME_SIZE)
+ burst_size = clib_min (max_burst_size, VLIB_FRAME_SIZE - n_segs);
+ burst_bytes = transport_connection_tx_pacer_burst (&tc->connection,
+ last_cpu_time);
+ burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
+ if (!burst_size)
{
tcp_program_fastretransmit (tc);
continue;
}
- burst_size = clib_min (max_burst_size, VLIB_FRAME_SIZE - n_segs);
+ n_segs_now = tcp_fast_retransmit (tc, burst_size);
+ sent_bytes = clib_min (n_segs_now * tc->snd_mss, burst_bytes);
+ transport_connection_tx_pacer_update_bytes (&tc->connection,
+ sent_bytes);
- if (tc->cwnd > tc->ssthresh + 3 * tc->snd_mss)
- {
- /* The first segment MUST be retransmitted */
- if (tcp_retransmit_first_unacked (tc))
- {
- tcp_program_fastretransmit (tc);
- continue;
- }
-
- /* Post retransmit update cwnd to ssthresh and account for the
- * three segments that have left the network and should've been
- * buffered at the receiver XXX */
- tc->cwnd = tc->ssthresh + 3 * tc->snd_mss;
-
- /* If cwnd allows, send more data */
- if (tcp_opts_sack_permitted (&tc->rcv_opts))
- {
- scoreboard_init_high_rxt (&tc->sack_sb,
- tc->snd_una + tc->snd_mss);
- tc->sack_sb.rescue_rxt = tc->snd_una - 1;
- n_segs += tcp_fast_retransmit_sack (tc, burst_size);
- }
- else
- {
- n_segs += tcp_fast_retransmit_no_sack (tc, burst_size);
- }
- }
- else
- n_segs += tcp_fast_retransmit (tc, burst_size);
+ n_segs += n_segs_now;
}
- vec_reset_length (wrk->ongoing_fast_rxt);
+ _vec_len (ongoing_fast_rxt) = 0;
+ wrk->ongoing_fast_rxt = ongoing_fast_rxt;
}
/**
@@ -1298,6 +1291,7 @@
}
else if (tcp_should_fastrecover (tc))
{
+ u32 byte_rate;
ASSERT (!tcp_in_fastrecovery (tc));
/* Heuristic to catch potential late dupacks
@@ -1313,8 +1307,21 @@
tc->cc_algo->rcv_cong_ack (tc, TCP_CC_DUPACK);
if (tcp_opts_sack_permitted (&tc->rcv_opts))
- tc->sack_sb.high_rxt = tc->snd_una;
+ {
+ tc->cwnd = tc->ssthresh;
+ scoreboard_init_high_rxt (&tc->sack_sb, tc->snd_una);
+ tc->sack_sb.rescue_rxt = tc->snd_una - 1;
+ }
+ else
+ {
+ /* Post retransmit update cwnd to ssthresh and account for the
+ * three segments that have left the network and should've been
+ * buffered at the receiver XXX */
+ tc->cwnd = tc->ssthresh + 3 * tc->snd_mss;
+ }
+ byte_rate = (0.3 * tc->cwnd) / ((f64) TCP_TICK * tc->srtt);
+ transport_connection_tx_pacer_init (&tc->connection, byte_rate, 0);
tcp_program_fastretransmit (tc);
return;
}
diff --git a/src/vnet/tcp/tcp_output.c b/src/vnet/tcp/tcp_output.c
index f14a612..81579ef 100644
--- a/src/vnet/tcp/tcp_output.c
+++ b/src/vnet/tcp/tcp_output.c
@@ -1452,10 +1452,10 @@
tc->timers[TCP_TIMER_RETRANSMIT] = TCP_TIMER_HANDLE_INVALID;
}
- TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 1);
-
if (tc->state >= TCP_STATE_ESTABLISHED)
{
+ TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 2);
+
/* Lost FIN, retransmit and return */
if (tcp_is_lost_fin (tc))
{
@@ -1536,6 +1536,8 @@
return;
}
+ TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 2);
+
/* Try without increasing RTO a number of times. If this fails,
* start growing RTO exponentially */
tc->rto_boff += 1;
@@ -1562,6 +1564,8 @@
/* Retransmit SYN-ACK */
else if (tc->state == TCP_STATE_SYN_RCVD)
{
+ TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 2);
+
tc->rto_boff += 1;
if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
@@ -1693,7 +1697,7 @@
old_snd_nxt = tc->snd_nxt;
tc->snd_nxt = tc->snd_una;
- TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 2);
+ TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 1);
n_bytes = tcp_prepare_retransmit_segment (tc, 0, tc->snd_mss, &b);
if (!n_bytes)