aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/svm/svm_fifo.c150
-rw-r--r--src/svm/svm_fifo.h41
-rw-r--r--src/vnet/session/session.c74
-rw-r--r--src/vnet/session/session.h38
-rwxr-xr-xsrc/vnet/session/session_cli.c266
-rw-r--r--src/vnet/tcp/builtin_client.c2
-rw-r--r--src/vnet/tcp/builtin_server.c7
-rw-r--r--src/vnet/tcp/tcp.c208
-rw-r--r--src/vnet/tcp/tcp.h53
-rw-r--r--src/vnet/tcp/tcp_input.c114
-rw-r--r--src/vnet/tcp/tcp_test.c289
11 files changed, 1111 insertions, 131 deletions
diff --git a/src/svm/svm_fifo.c b/src/svm/svm_fifo.c
index da60fee51ec..fc2189c5176 100644
--- a/src/svm/svm_fifo.c
+++ b/src/svm/svm_fifo.c
@@ -61,6 +61,90 @@ format_ooo_segment (u8 * s, va_list * args)
}
u8 *
+svm_fifo_dump_trace (u8 * s, svm_fifo_t * f)
+{
+#if SVM_FIFO_TRACE
+ svm_fifo_trace_elem_t *seg = 0;
+ int i = 0;
+
+ if (f->trace)
+ {
+ vec_foreach (seg, f->trace)
+ {
+ s = format (s, "{%u, %u, %u}, ", seg->offset, seg->len, seg->action);
+ i++;
+ if (i % 5 == 0)
+ s = format (s, "\n");
+ }
+ s = format (s, "\n");
+ }
+ return s;
+#else
+ return 0;
+#endif
+}
+
+u8 *
+svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose)
+{
+ int i, trace_len;
+ u8 *data = 0;
+ svm_fifo_trace_elem_t *trace;
+ u32 offset;
+ svm_fifo_t *dummy_fifo;
+
+ if (!f)
+ return s;
+
+#if SVM_FIFO_TRACE
+ trace = f->trace;
+ trace_len = vec_len (trace);
+#else
+ trace = 0;
+ trace_len = 0;
+#endif
+
+ dummy_fifo = svm_fifo_create (f->nitems);
+ memset (f->data, 0xFF, f->nitems);
+
+ vec_validate (data, f->nitems);
+ for (i = 0; i < vec_len (data); i++)
+ data[i] = i;
+
+ for (i = 0; i < trace_len; i++)
+ {
+ offset = trace[i].offset;
+ if (trace[i].action == 1)
+ {
+ if (verbose)
+ s = format (s, "adding [%u, %u]:", trace[i].offset,
+ (trace[i].offset +
+ trace[i].len) % dummy_fifo->nitems);
+ svm_fifo_enqueue_with_offset (dummy_fifo, trace[i].offset,
+ trace[i].len, &data[offset]);
+ }
+ else if (trace[i].action == 2)
+ {
+ if (verbose)
+ s = format (s, "adding [%u, %u]:", 0, trace[i].len);
+ svm_fifo_enqueue_nowait (dummy_fifo, trace[i].len, &data[offset]);
+ }
+ else if (!no_read)
+ {
+ if (verbose)
+ s = format (s, "read: %u", trace[i].len);
+ svm_fifo_dequeue_drop (dummy_fifo, trace[i].len);
+ }
+ if (verbose)
+ s = format (s, "%U", format_svm_fifo, dummy_fifo, 1);
+ }
+
+ s = format (s, "result: %U", format_svm_fifo, dummy_fifo, 1);
+
+ return s;
+}
+
+u8 *
format_ooo_list (u8 * s, va_list * args)
{
svm_fifo_t *f = va_arg (*args, svm_fifo_t *);
@@ -73,6 +157,7 @@ format_ooo_list (u8 * s, va_list * args)
s = format (s, " %U\n", format_ooo_segment, seg);
ooo_segment_index = seg->next;
}
+
return s;
}
@@ -94,10 +179,10 @@ format_svm_fifo (u8 * s, va_list * args)
if (verbose)
{
- s = format (s, " ooo pool %d active elts\n",
- pool_elts (f->ooo_segments));
+ s = format (s, " ooo pool %d active elts newest %u\n",
+ pool_elts (f->ooo_segments), f->ooos_newest);
if (svm_fifo_has_ooo_data (f))
- s = format (s, " %U", format_ooo_list, f);
+ s = format (s, " %U", format_ooo_list, f, verbose);
}
return s;
}
@@ -116,7 +201,6 @@ svm_fifo_create (u32 data_size_in_bytes)
memset (f, 0, sizeof (*f));
f->nitems = data_size_in_bytes;
f->ooos_list_head = OOO_SEGMENT_INVALID_INDEX;
-
return (f);
}
@@ -178,6 +262,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
u32 new_index, s_end_pos, s_index;
u32 normalized_position, normalized_end_position;
+ ASSERT (offset + length <= ooo_segment_distance_from_tail (f, f->head));
normalized_position = (f->tail + offset) % f->nitems;
normalized_end_position = (f->tail + offset + length) % f->nitems;
@@ -205,17 +290,9 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
s = prev;
s_end_pos = ooo_segment_end_pos (f, s);
- /* Check head and tail now since segment may be wider at both ends so
- * merge tests lower won't work */
- if (position_lt (f, normalized_position, s->start))
- {
- s->start = normalized_position;
- s->length = position_diff (f, s_end_pos, s->start);
- }
- if (position_gt (f, normalized_end_position, s_end_pos))
- {
- s->length = position_diff (f, normalized_end_position, s->start);
- }
+ /* Since we have previous, normalized start position cannot be smaller
+ * than prev->start. Check tail */
+ ASSERT (position_lt (f, s->start, normalized_position));
goto check_tail;
}
@@ -256,6 +333,7 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
/* Pool might've moved, get segment again */
s = pool_elt_at_index (f->ooo_segments, s_index);
+ /* Needs to be last */
ASSERT (s->next == OOO_SEGMENT_INVALID_INDEX);
new_s->prev = s_index;
@@ -274,32 +352,22 @@ ooo_segment_add (svm_fifo_t * f, u32 offset, u32 length)
{
s->start = normalized_position;
s->length = position_diff (f, s_end_pos, s->start);
- }
- /* Overlapping tail */
- else if (position_gt (f, normalized_end_position, s_end_pos))
- {
- s->length = position_diff (f, normalized_end_position, s->start);
- }
- /* New segment completely covered by current one */
- else
- {
- /* Do Nothing */
- s = 0;
- goto done;
+ f->ooos_newest = s - f->ooo_segments;
}
check_tail:
- /* The new segment's tail may cover multiple smaller ones */
+
+ /* Overlapping tail */
if (position_gt (f, normalized_end_position, s_end_pos))
{
- /* Remove the completely overlapped segments */
- it = (s->next != OOO_SEGMENT_INVALID_INDEX) ?
- pool_elt_at_index (f->ooo_segments, s->next) : 0;
+ s->length = position_diff (f, normalized_end_position, s->start);
+
+ /* Remove the completely overlapped segments in the tail */
+ it = ooo_segment_next (f, s);
while (it && position_leq (f, ooo_segment_end_pos (f, it),
normalized_end_position))
{
- next = (it->next != OOO_SEGMENT_INVALID_INDEX) ?
- pool_elt_at_index (f->ooo_segments, it->next) : 0;
+ next = ooo_segment_next (f, it);
ooo_segment_del (f, it - f->ooo_segments);
it = next;
}
@@ -307,16 +375,12 @@ check_tail:
/* If partial overlap with last, merge */
if (it && position_leq (f, it->start, normalized_end_position))
{
- s->length =
- position_diff (f, ooo_segment_end_pos (f, it), s->start);
+ s->length = position_diff (f, ooo_segment_end_pos (f, it),
+ s->start);
ooo_segment_del (f, it - f->ooo_segments);
}
+ f->ooos_newest = s - f->ooo_segments;
}
-
-done:
- /* Most recently updated segment */
- if (s)
- f->ooos_newest = s - f->ooo_segments;
}
/**
@@ -422,6 +486,8 @@ svm_fifo_enqueue_internal (svm_fifo_t * f, u32 max_bytes, u8 * copy_from_here)
total_copy_bytes = max_bytes;
}
+ svm_fifo_trace_add (f, f->head, total_copy_bytes, 2);
+
/* Any out-of-order segments to collect? */
if (PREDICT_FALSE (f->ooos_list_head != OOO_SEGMENT_INVALID_INDEX))
total_copy_bytes += ooo_segment_try_collect (f, total_copy_bytes);
@@ -499,6 +565,8 @@ svm_fifo_enqueue_with_offset_internal (svm_fifo_t * f,
if ((required_bytes + offset_from_tail) > (nitems - cursize))
return -1;
+ svm_fifo_trace_add (f, offset, required_bytes, 1);
+
ooo_segment_add (f, offset, required_bytes);
/* Number of bytes we're going to copy */
@@ -707,6 +775,8 @@ svm_fifo_dequeue_drop (svm_fifo_t * f, u32 max_bytes)
/* Number of bytes we're going to drop */
total_drop_bytes = (cursize < max_bytes) ? cursize : max_bytes;
+ svm_fifo_trace_add (f, f->tail, total_drop_bytes, 3);
+
/* Number of bytes in first copy segment */
first_drop_bytes =
((nitems - f->head) < total_drop_bytes) ?
diff --git a/src/svm/svm_fifo.h b/src/svm/svm_fifo.h
index fe21de476d7..a83cd858f83 100644
--- a/src/svm/svm_fifo.h
+++ b/src/svm/svm_fifo.h
@@ -36,8 +36,16 @@ typedef struct
format_function_t format_ooo_segment;
format_function_t format_ooo_list;
+#define SVM_FIFO_TRACE (0)
#define OOO_SEGMENT_INVALID_INDEX ((u32)~0)
+typedef struct
+{
+ u32 offset;
+ u32 len;
+ u32 action;
+} svm_fifo_trace_elem_t;
+
typedef struct _svm_fifo
{
volatile u32 cursize; /**< current fifo size */
@@ -64,9 +72,28 @@ typedef struct _svm_fifo
u32 ooos_newest; /**< Last segment to have been updated */
struct _svm_fifo *next; /**< next in freelist/active chain */
struct _svm_fifo *prev; /**< prev in active chain */
+#if SVM_FIFO_TRACE
+ svm_fifo_trace_elem_t *trace;
+#endif
CLIB_CACHE_LINE_ALIGN_MARK (data);
} svm_fifo_t;
+#if SVM_FIFO_TRACE
+#define svm_fifo_trace_add(_f, _s, _l, _t) \
+{ \
+ svm_fifo_trace_elem_t *trace_elt; \
+ vec_add2(_f->trace, trace_elt, 1); \
+ trace_elt->offset = _s; \
+ trace_elt->len = _l; \
+ trace_elt->action = _t; \
+}
+#else
+#define svm_fifo_trace_add(_f, _s, _l, _t)
+#endif
+
+u8 *svm_fifo_dump_trace (u8 * s, svm_fifo_t * f);
+u8 *svm_fifo_replay (u8 * s, svm_fifo_t * f, u8 no_read, u8 verbose);
+
static inline u32
svm_fifo_max_dequeue (svm_fifo_t * f)
{
@@ -132,6 +159,12 @@ svm_fifo_newest_ooo_segment (svm_fifo_t * f)
return pool_elt_at_index (f->ooo_segments, f->ooos_newest);
}
+always_inline void
+svm_fifo_newest_ooo_segment_reset (svm_fifo_t * f)
+{
+ f->ooos_newest = OOO_SEGMENT_INVALID_INDEX;
+}
+
always_inline u32
ooo_segment_distance_from_tail (svm_fifo_t * f, u32 pos)
{
@@ -174,6 +207,14 @@ ooo_segment_get_prev (svm_fifo_t * f, ooo_segment_t * s)
return pool_elt_at_index (f->ooo_segments, s->prev);
}
+always_inline ooo_segment_t *
+ooo_segment_next (svm_fifo_t * f, ooo_segment_t * s)
+{
+ if (s->next == OOO_SEGMENT_INVALID_INDEX)
+ return 0;
+ return pool_elt_at_index (f->ooo_segments, s->next);
+}
+
#endif /* __included_ssvm_fifo_h__ */
/*
diff --git a/src/vnet/session/session.c b/src/vnet/session/session.c
index 0a86d563f0e..2c2a27c1af9 100644
--- a/src/vnet/session/session.c
+++ b/src/vnet/session/session.c
@@ -325,9 +325,9 @@ stream_session_half_open_lookup (session_manager_main_t * smm,
}
transport_connection_t *
-stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
- u16 lcl_port, u16 rmt_port, u8 proto,
- u32 my_thread_index)
+stream_session_lookup_transport_wt4 (ip4_address_t * lcl, ip4_address_t * rmt,
+ u16 lcl_port, u16 rmt_port, u8 proto,
+ u32 my_thread_index)
{
session_manager_main_t *smm = &session_manager_main;
session_kv4_t kv4;
@@ -358,9 +358,40 @@ stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
}
transport_connection_t *
-stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
- u16 lcl_port, u16 rmt_port, u8 proto,
- u32 my_thread_index)
+stream_session_lookup_transport4 (ip4_address_t * lcl, ip4_address_t * rmt,
+ u16 lcl_port, u16 rmt_port, u8 proto)
+{
+ session_manager_main_t *smm = &session_manager_main;
+ session_kv4_t kv4;
+ stream_session_t *s;
+ int rv;
+
+ /* Lookup session amongst established ones */
+ make_v4_ss_kv (&kv4, lcl, rmt, lcl_port, rmt_port, proto);
+ rv = clib_bihash_search_inline_16_8 (&smm->v4_session_hash, &kv4);
+ if (rv == 0)
+ {
+ s = stream_session_get_from_handle (kv4.value);
+ return tp_vfts[s->session_type].get_connection (s->connection_index,
+ s->thread_index);
+ }
+
+ /* If nothing is found, check if any listener is available */
+ s = stream_session_lookup_listener4 (lcl, lcl_port, proto);
+ if (s)
+ return tp_vfts[s->session_type].get_listener (s->connection_index);
+
+ /* Finally, try half-open connections */
+ rv = clib_bihash_search_inline_16_8 (&smm->v4_half_open_hash, &kv4);
+ if (rv == 0)
+ return tp_vfts[proto].get_half_open (kv4.value & 0xFFFFFFFF);
+ return 0;
+}
+
+transport_connection_t *
+stream_session_lookup_transport_wt6 (ip6_address_t * lcl, ip6_address_t * rmt,
+ u16 lcl_port, u16 rmt_port, u8 proto,
+ u32 my_thread_index)
{
session_manager_main_t *smm = &session_manager_main;
stream_session_t *s;
@@ -390,6 +421,37 @@ stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
return 0;
}
+transport_connection_t *
+stream_session_lookup_transport6 (ip6_address_t * lcl, ip6_address_t * rmt,
+ u16 lcl_port, u16 rmt_port, u8 proto)
+{
+ session_manager_main_t *smm = &session_manager_main;
+ stream_session_t *s;
+ session_kv6_t kv6;
+ int rv;
+
+ make_v6_ss_kv (&kv6, lcl, rmt, lcl_port, rmt_port, proto);
+ rv = clib_bihash_search_inline_48_8 (&smm->v6_session_hash, &kv6);
+ if (rv == 0)
+ {
+ s = stream_session_get_from_handle (kv6.value);
+ return tp_vfts[s->session_type].get_connection (s->connection_index,
+ s->thread_index);
+ }
+
+ /* If nothing is found, check if any listener is available */
+ s = stream_session_lookup_listener6 (lcl, lcl_port, proto);
+ if (s)
+ return tp_vfts[s->session_type].get_listener (s->connection_index);
+
+ /* Finally, try half-open connections */
+ rv = clib_bihash_search_inline_48_8 (&smm->v6_half_open_hash, &kv6);
+ if (rv == 0)
+ return tp_vfts[proto].get_half_open (kv6.value & 0xFFFFFFFF);
+
+ return 0;
+}
+
int
stream_session_create_i (segment_manager_t * sm, transport_connection_t * tc,
stream_session_t ** ret_s)
diff --git a/src/vnet/session/session.h b/src/vnet/session/session.h
index b4507d4eaed..6069c5740f7 100644
--- a/src/vnet/session/session.h
+++ b/src/vnet/session/session.h
@@ -263,15 +263,30 @@ stream_session_t *stream_session_lookup6 (ip6_address_t * lcl,
ip6_address_t * rmt, u16 lcl_port,
u16 rmt_port, u8 proto);
transport_connection_t
- * stream_session_lookup_transport4 (ip4_address_t * lcl,
- ip4_address_t * rmt, u16 lcl_port,
- u16 rmt_port, u8 proto,
- u32 thread_index);
-transport_connection_t
- * stream_session_lookup_transport6 (ip6_address_t * lcl,
- ip6_address_t * rmt, u16 lcl_port,
- u16 rmt_port, u8 proto,
- u32 thread_index);
+ * stream_session_lookup_transport_wt4 (ip4_address_t * lcl,
+ ip4_address_t * rmt, u16 lcl_port,
+ u16 rmt_port, u8 proto,
+ u32 thread_index);
+transport_connection_t *stream_session_lookup_transport4 (ip4_address_t * lcl,
+ ip4_address_t * rmt,
+ u16 lcl_port,
+ u16 rmt_port,
+ u8 proto);
+transport_connection_t *stream_session_lookup_transport_wt6 (ip6_address_t *
+ lcl,
+ ip6_address_t *
+ rmt,
+ u16 lcl_port,
+ u16 rmt_port,
+ u8 proto,
+ u32
+ thread_index);
+transport_connection_t *stream_session_lookup_transport6 (ip6_address_t * lcl,
+ ip6_address_t * rmt,
+ u16 lcl_port,
+ u16 rmt_port,
+ u8 proto);
+
stream_session_t *stream_session_lookup_listener (ip46_address_t * lcl,
u16 lcl_port, u8 proto);
void stream_session_table_add_for_tc (transport_connection_t * tc, u64 value);
@@ -415,7 +430,12 @@ void stream_session_cleanup (stream_session_t * s);
void session_send_session_evt_to_thread (u64 session_handle,
fifo_event_type_t evt_type,
u32 thread_index);
+
u8 *format_stream_session (u8 * s, va_list * args);
+uword unformat_stream_session (unformat_input_t * input, va_list * args);
+uword unformat_transport_connection (unformat_input_t * input,
+ va_list * args);
+
int
send_session_connected_callback (u32 app_index, u32 api_context,
stream_session_t * s, u8 is_fail);
diff --git a/src/vnet/session/session_cli.c b/src/vnet/session/session_cli.c
index e06bc586be3..e8e6f99c742 100755
--- a/src/vnet/session/session_cli.c
+++ b/src/vnet/session/session_cli.c
@@ -81,12 +81,141 @@ format_stream_session (u8 * s, va_list * args)
{
clib_warning ("Session in state: %d!", ss->session_state);
}
-
vec_free (str);
return s;
}
+uword
+unformat_stream_session_id (unformat_input_t * input, va_list * args)
+{
+ u8 *proto = va_arg (*args, u8 *);
+ ip46_address_t *lcl = va_arg (*args, ip46_address_t *);
+ ip46_address_t *rmt = va_arg (*args, ip46_address_t *);
+ u16 *lcl_port = va_arg (*args, u16 *);
+ u16 *rmt_port = va_arg (*args, u16 *);
+ u8 *is_ip4 = va_arg (*args, u8 *);
+ u8 tuple_is_set = 0;
+
+ memset (lcl, 0, sizeof (*lcl));
+ memset (rmt, 0, sizeof (*rmt));
+
+ if (unformat (input, "tcp"))
+ {
+ *proto = TRANSPORT_PROTO_TCP;
+ }
+ if (unformat (input, "udp"))
+ {
+ *proto = TRANSPORT_PROTO_UDP;
+ }
+ else if (unformat (input, "%U:%d->%U:%d", unformat_ip4_address, &lcl->ip4,
+ lcl_port, unformat_ip4_address, &rmt->ip4, rmt_port))
+ {
+ *is_ip4 = 1;
+ tuple_is_set = 1;
+ }
+ else if (unformat (input, "%U:%d->%U:%d", unformat_ip6_address, &lcl->ip6,
+ lcl_port, unformat_ip6_address, &rmt->ip6, rmt_port))
+ {
+ *is_ip4 = 0;
+ tuple_is_set = 1;
+ }
+ else
+ return 0;
+
+ if (tuple_is_set)
+ return 1;
+
+ return 0;
+}
+
+uword
+unformat_stream_session (unformat_input_t * input, va_list * args)
+{
+ stream_session_t **result = va_arg (*args, stream_session_t **);
+ stream_session_t *s;
+ u8 proto = ~0;
+ ip46_address_t lcl, rmt;
+ u32 lcl_port = 0, rmt_port = 0;
+ u8 is_ip4 = 0, s_type = ~0, id_is_set = 0;
+
+ if (unformat (input, "%U", unformat_stream_session_id, &proto, &lcl, &rmt,
+ &lcl_port, &rmt_port, &is_ip4))
+ {
+ id_is_set = 1;
+ }
+ else
+ return 0;
+
+ if (!id_is_set)
+ {
+ return 0;
+ }
+
+ s_type = session_type_from_proto_and_ip (proto, is_ip4);
+ if (is_ip4)
+ s = stream_session_lookup4 (&lcl.ip4, &rmt.ip4,
+ clib_host_to_net_u16 (lcl_port),
+ clib_host_to_net_u16 (rmt_port), s_type);
+ else
+ s = stream_session_lookup6 (&lcl.ip6, &rmt.ip6,
+ clib_host_to_net_u16 (lcl_port),
+ clib_host_to_net_u16 (rmt_port), s_type);
+ if (s)
+ {
+ *result = s;
+ return 1;
+ }
+ return 0;
+}
+
+uword
+unformat_transport_connection (unformat_input_t * input, va_list * args)
+{
+ transport_connection_t **result = va_arg (*args, transport_connection_t **);
+ u32 suggested_proto = va_arg (*args, u32);
+ transport_connection_t *tc;
+ u8 proto = ~0;
+ ip46_address_t lcl, rmt;
+ u32 lcl_port = 0, rmt_port = 0;
+ u8 is_ip4 = 0, s_type = ~0, id_is_set = 0;
+
+ if (unformat (input, "%U", unformat_stream_session_id, &proto, &lcl, &rmt,
+ &lcl_port, &rmt_port, &is_ip4))
+ {
+ id_is_set = 1;
+ }
+ else
+ return 0;
+
+ if (!id_is_set)
+ {
+ return 0;
+ }
+
+ proto = (proto == (u8) ~ 0) ? suggested_proto : proto;
+ if (proto == (u8) ~ 0)
+ return 0;
+ s_type = session_type_from_proto_and_ip (proto, is_ip4);
+ if (is_ip4)
+ tc = stream_session_lookup_transport4 (&lcl.ip4, &rmt.ip4,
+ clib_host_to_net_u16 (lcl_port),
+ clib_host_to_net_u16 (rmt_port),
+ s_type);
+ else
+ tc = stream_session_lookup_transport6 (&lcl.ip6, &rmt.ip6,
+ clib_host_to_net_u16 (lcl_port),
+ clib_host_to_net_u16 (rmt_port),
+ s_type);
+
+ if (tc)
+ {
+ *result = tc;
+ return 1;
+ }
+ return 0;
+}
+
static clib_error_t *
show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
vlib_cli_command_t * cmd)
@@ -95,13 +224,7 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
int verbose = 0, i;
stream_session_t *pool;
stream_session_t *s;
- u8 *str = 0, one_session = 0, proto_set = 0, proto = 0;
- u8 is_ip4 = 0, s_type = 0;
- ip4_address_t lcl_ip4, rmt_ip4;
- u32 lcl_port = 0, rmt_port = 0;
-
- memset (&lcl_ip4, 0, sizeof (lcl_ip4));
- memset (&rmt_ip4, 0, sizeof (rmt_ip4));
+ u8 *str = 0, one_session = 0;
if (!smm->is_enabled)
{
@@ -114,40 +237,18 @@ show_session_command_fn (vlib_main_t * vm, unformat_input_t * input,
;
else if (unformat (input, "verbose"))
verbose = 1;
- else if (unformat (input, "tcp"))
- {
- proto_set = 1;
- proto = TRANSPORT_PROTO_TCP;
- }
- else if (unformat (input, "%U:%d->%U:%d",
- unformat_ip4_address, &lcl_ip4, &lcl_port,
- unformat_ip4_address, &rmt_ip4, &rmt_port))
+ else if (unformat (input, "%U", unformat_stream_session, &s))
{
one_session = 1;
- is_ip4 = 1;
}
-
else
- break;
+ return clib_error_return (0, "unknown input `%U'",
+ format_unformat_error, input);
}
if (one_session)
{
- if (!proto_set)
- {
- vlib_cli_output (vm, "proto not set");
- return clib_error_return (0, "proto not set");
- }
-
- s_type = session_type_from_proto_and_ip (proto, is_ip4);
- s = stream_session_lookup4 (&lcl_ip4, &rmt_ip4,
- clib_host_to_net_u16 (lcl_port),
- clib_host_to_net_u16 (rmt_port), s_type);
- if (s)
- vlib_cli_output (vm, "%U", format_stream_session, s, 2);
- else
- vlib_cli_output (vm, "session does not exist");
-
+ vlib_cli_output (vm, "%U", format_stream_session, s, 2);
return 0;
}
@@ -275,6 +376,103 @@ VLIB_CLI_COMMAND (clear_session_command, static) =
/* *INDENT-ON* */
static clib_error_t *
+show_session_fifo_trace_command_fn (vlib_main_t * vm,
+ unformat_input_t * input,
+ vlib_cli_command_t * cmd)
+{
+ stream_session_t *s = 0;
+ u8 is_rx = 0, *str = 0;
+
+ while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+ {
+ if (unformat (input, "%U", unformat_stream_session, &s))
+ ;
+ else if (unformat (input, "rx"))
+ is_rx = 1;
+ else if (unformat (input, "tx"))
+ is_rx = 0;
+ else
+ return clib_error_return (0, "unknown input `%U'",
+ format_unformat_error, input);
+ }
+
+ if (!SVM_FIFO_TRACE)
+ {
+ vlib_cli_output (vm, "fifo tracing not enabled");
+ return 0;
+ }
+
+ if (!s)
+ {
+ vlib_cli_output (vm, "could not find session");
+ return 0;
+ }
+
+ str = is_rx ?
+ svm_fifo_dump_trace (str, s->server_rx_fifo) :
+ svm_fifo_dump_trace (str, s->server_tx_fifo);
+
+ vlib_cli_output (vm, "%v", str);
+ return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (show_session_fifo_trace_command, static) =
+{
+ .path = "show session fifo trace",
+ .short_help = "show session fifo trace <session>",
+ .function = show_session_fifo_trace_command_fn,
+};
+/* *INDENT-ON* */
+
+static clib_error_t *
+session_replay_fifo_command_fn (vlib_main_t * vm, unformat_input_t * input,
+ vlib_cli_command_t * cmd)
+{
+ stream_session_t *s = 0;
+ u8 is_rx = 0, *str = 0;
+
+ while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+ {
+ if (unformat (input, "%U", unformat_stream_session, &s))
+ ;
+ else if (unformat (input, "rx"))
+ is_rx = 1;
+ else
+ return clib_error_return (0, "unknown input `%U'",
+ format_unformat_error, input);
+ }
+
+ if (!SVM_FIFO_TRACE)
+ {
+ vlib_cli_output (vm, "fifo tracing not enabled");
+ return 0;
+ }
+
+ if (!s)
+ {
+ vlib_cli_output (vm, "could not find session");
+ return 0;
+ }
+
+ str = is_rx ?
+ svm_fifo_replay (str, s->server_rx_fifo, 0, 1) :
+ svm_fifo_replay (str, s->server_tx_fifo, 0, 1);
+
+ vlib_cli_output (vm, "%v", str);
+ return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (session_replay_fifo_trace_command, static) =
+{
+ .path = "session replay fifo",
+ .short_help = "session replay fifo <session>",
+ .function = session_replay_fifo_command_fn,
+};
+/* *INDENT-ON* */
+
+static clib_error_t *
session_enable_disable_fn (vlib_main_t * vm, unformat_input_t * input,
vlib_cli_command_t * cmd)
{
diff --git a/src/vnet/tcp/builtin_client.c b/src/vnet/tcp/builtin_client.c
index a6c8a23582b..a92bacaa37d 100644
--- a/src/vnet/tcp/builtin_client.c
+++ b/src/vnet/tcp/builtin_client.c
@@ -464,6 +464,8 @@ builtin_session_connected_callback (u32 app_index, u32 api_context,
static void
builtin_session_reset_callback (stream_session_t * s)
{
+ if (s->session_state == SESSION_STATE_READY)
+ clib_warning ("Reset active connection %U", format_stream_session, s, 2);
return;
}
diff --git a/src/vnet/tcp/builtin_server.c b/src/vnet/tcp/builtin_server.c
index 8e958ac0b45..4ecaf56a70c 100644
--- a/src/vnet/tcp/builtin_server.c
+++ b/src/vnet/tcp/builtin_server.c
@@ -99,8 +99,7 @@ builtin_session_disconnect_callback (stream_session_t * s)
void
builtin_session_reset_callback (stream_session_t * s)
{
- clib_warning ("called.. ");
-
+ clib_warning ("Reset session %U", format_stream_session, s, 2);
stream_session_cleanup (s);
}
@@ -224,10 +223,6 @@ builtin_server_rx_callback (stream_session_t * s)
clib_warning ("session stuck: %U", format_stream_session, s, 2);
}
}
- else
- {
- bsm->rx_retries[thread_index][s->session_index] = 0;
- }
return 0;
}
diff --git a/src/vnet/tcp/tcp.c b/src/vnet/tcp/tcp.c
index f379e699839..8ed325d2809 100644
--- a/src/vnet/tcp/tcp.c
+++ b/src/vnet/tcp/tcp.c
@@ -732,6 +732,7 @@ format_tcp_connection (u8 * s, va_list * args)
if (verbose > 1)
s = format (s, " %U\n%U", format_tcp_timers, tc, format_tcp_vars, tc);
}
+
return s;
}
@@ -792,6 +793,30 @@ format_tcp_sacks (u8 * s, va_list * args)
}
u8 *
+format_tcp_rcv_sacks (u8 * s, va_list * args)
+{
+ tcp_connection_t *tc = va_arg (*args, tcp_connection_t *);
+ sack_block_t *sacks = tc->rcv_opts.sacks;
+ sack_block_t *block;
+ int i, len = 0;
+
+ len = vec_len (sacks);
+ for (i = 0; i < len - 1; i++)
+ {
+ block = &sacks[i];
+ s = format (s, " start %u end %u\n", block->start - tc->iss,
+ block->end - tc->iss);
+ }
+ if (len)
+ {
+ block = &sacks[len - 1];
+ s = format (s, " start %u end %u", block->start - tc->iss,
+ block->end - tc->iss);
+ }
+ return s;
+}
+
+u8 *
format_tcp_sack_hole (u8 * s, va_list * args)
{
sack_scoreboard_hole_t *hole = va_arg (*args, sack_scoreboard_hole_t *);
@@ -820,6 +845,7 @@ format_tcp_scoreboard (u8 * s, va_list * args)
s = format (s, "%U", format_tcp_sack_hole, hole);
hole = scoreboard_next_hole (sb, hole);
}
+
return s;
}
@@ -1304,7 +1330,189 @@ VLIB_CLI_COMMAND (tcp_src_address_command, static) =
};
/* *INDENT-ON* */
+static u8 *
+tcp_scoreboard_dump_trace (u8 * s, sack_scoreboard_t * sb)
+{
+#if TCP_SCOREBOARD_TRACE
+
+ scoreboard_trace_elt_t *block;
+ int i = 0;
+
+ if (!sb->trace)
+ return s;
+
+ s = format (s, "scoreboard trace:");
+ vec_foreach (block, sb->trace)
+ {
+ s = format (s, "{%u, %u, %u, %u, %u}, ", block->start, block->end,
+ block->ack, block->snd_una_max, block->group);
+ if ((++i % 3) == 0)
+ s = format (s, "\n");
+ }
+ return s;
+#else
+ return 0;
+#endif
+}
+
+static clib_error_t *
+tcp_show_scoreboard_trace_fn (vlib_main_t * vm, unformat_input_t * input,
+ vlib_cli_command_t * cmd_arg)
+{
+ transport_connection_t *tconn = 0;
+ tcp_connection_t *tc;
+ u8 *s = 0;
+ while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+ {
+ if (unformat (input, "%U", unformat_transport_connection, &tconn,
+ TRANSPORT_PROTO_TCP))
+ ;
+ else
+ return clib_error_return (0, "unknown input `%U'",
+ format_unformat_error, input);
+ }
+
+ if (!TCP_SCOREBOARD_TRACE)
+ {
+ vlib_cli_output (vm, "scoreboard tracing not enabled");
+ return 0;
+ }
+
+ tc = tcp_get_connection_from_transport (tconn);
+ s = tcp_scoreboard_dump_trace (s, &tc->sack_sb);
+ vlib_cli_output (vm, "%v", s);
+ return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (tcp_show_scoreboard_trace_command, static) =
+{
+ .path = "show tcp scoreboard trace",
+ .short_help = "show tcp scoreboard trace <connection>",
+ .function = tcp_show_scoreboard_trace_fn,
+};
+/* *INDENT-ON* */
+u8 *
+tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose)
+{
+ int i, trace_len;
+ scoreboard_trace_elt_t *trace;
+ u32 next_ack, left, group, has_new_ack = 0;
+ tcp_connection_t _dummy_tc, *dummy_tc = &_dummy_tc;
+ sack_block_t *block;
+
+ if (!tc)
+ return s;
+
+ memset (dummy_tc, 0, sizeof (*dummy_tc));
+ tcp_connection_timers_init (dummy_tc);
+ scoreboard_init (&dummy_tc->sack_sb);
+ dummy_tc->rcv_opts.flags |= TCP_OPTS_FLAG_SACK;
+
+#if TCP_SCOREBOARD_TRACE
+ trace = tc->sack_sb.trace;
+ trace_len = vec_len (tc->sack_sb.trace);
+#else
+ trace = 0;
+ trace_len = 0;
+#endif
+
+ for (i = 0; i < trace_len; i++)
+ {
+ if (trace[i].ack != 0)
+ {
+ dummy_tc->snd_una = trace[i].ack - 1448;
+ dummy_tc->snd_una_max = trace[i].ack;
+ }
+ }
+
+ left = 0;
+ while (left < trace_len)
+ {
+ group = trace[left].group;
+ vec_reset_length (dummy_tc->rcv_opts.sacks);
+ has_new_ack = 0;
+ while (trace[left].group == group)
+ {
+ if (trace[left].ack != 0)
+ {
+ if (verbose)
+ s = format (s, "Adding ack %u, snd_una_max %u, segs: ",
+ trace[left].ack, trace[left].snd_una_max);
+ dummy_tc->snd_una_max = trace[left].snd_una_max;
+ next_ack = trace[left].ack;
+ has_new_ack = 1;
+ }
+ else
+ {
+ if (verbose)
+ s = format (s, "[%u, %u], ", trace[left].start,
+ trace[left].end);
+ vec_add2 (dummy_tc->rcv_opts.sacks, block, 1);
+ block->start = trace[left].start;
+ block->end = trace[left].end;
+ }
+ left++;
+ }
+
+ /* Push segments */
+ tcp_rcv_sacks (dummy_tc, next_ack);
+ if (has_new_ack)
+ dummy_tc->snd_una = next_ack + dummy_tc->sack_sb.snd_una_adv;
+
+ if (verbose)
+ s = format (s, "result: %U", format_tcp_scoreboard,
+ &dummy_tc->sack_sb);
+
+ }
+ s = format (s, "result: %U", format_tcp_scoreboard, &dummy_tc->sack_sb);
+
+ return s;
+}
+
+static clib_error_t *
+tcp_scoreboard_trace_fn (vlib_main_t * vm, unformat_input_t * input,
+ vlib_cli_command_t * cmd_arg)
+{
+ transport_connection_t *tconn = 0;
+ tcp_connection_t *tc = 0;
+ u8 *str = 0;
+ while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+ {
+ if (unformat (input, "%U", unformat_transport_connection, &tconn,
+ TRANSPORT_PROTO_TCP))
+ ;
+ else
+ return clib_error_return (0, "unknown input `%U'",
+ format_unformat_error, input);
+ }
+
+ if (!TCP_SCOREBOARD_TRACE)
+ {
+ vlib_cli_output (vm, "scoreboard tracing not enabled");
+ return 0;
+ }
+
+ tc = tcp_get_connection_from_transport (tconn);
+ if (!tc)
+ {
+ vlib_cli_output (vm, "connection not found");
+ return 0;
+ }
+ str = tcp_scoreboard_replay (str, tc, 1);
+ vlib_cli_output (vm, "%v", str);
+ return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (tcp_replay_scoreboard_command, static) =
+{
+ .path = "tcp replay scoreboard",
+ .short_help = "tcp replay scoreboard <connection>",
+ .function = tcp_scoreboard_trace_fn,
+};
+/* *INDENT-ON* */
/*
* fd.io coding-style-patch-verification: ON
diff --git a/src/vnet/tcp/tcp.h b/src/vnet/tcp/tcp.h
index 37b10fd4753..fd0d02b99d6 100644
--- a/src/vnet/tcp/tcp.h
+++ b/src/vnet/tcp/tcp.h
@@ -62,6 +62,7 @@ typedef enum _tcp_state
format_function_t format_tcp_state;
format_function_t format_tcp_flags;
format_function_t format_tcp_sacks;
+format_function_t format_tcp_rcv_sacks;
/** TCP timers */
#define foreach_tcp_timer \
@@ -151,9 +152,19 @@ enum
#undef _
};
+#define TCP_SCOREBOARD_TRACE (0)
#define TCP_MAX_SACK_BLOCKS 15 /**< Max number of SACK blocks stored */
#define TCP_INVALID_SACK_HOLE_INDEX ((u32)~0)
+typedef struct _scoreboard_trace_elt
+{
+ u32 start;
+ u32 end;
+ u32 ack;
+ u32 snd_una_max;
+ u32 group;
+} scoreboard_trace_elt_t;
+
typedef struct _sack_scoreboard_hole
{
u32 next; /**< Index for next entry in linked list */
@@ -177,8 +188,38 @@ typedef struct _sack_scoreboard
u32 rescue_rxt; /**< Rescue sequence number */
u32 lost_bytes; /**< Bytes lost as per RFC6675 */
u32 cur_rxt_hole; /**< Retransmitting from this hole */
+
+#if TCP_SCOREBOARD_TRACE
+ scoreboard_trace_elt_t *trace;
+#endif
+
} sack_scoreboard_t;
+#if TCP_SCOREBOARD_TRACE
+#define tcp_scoreboard_trace_add(_tc, _ack) \
+{ \
+ static u64 _group = 0; \
+ sack_scoreboard_t *_sb = &_tc->sack_sb; \
+ sack_block_t *_sack, *_sacks; \
+ scoreboard_trace_elt_t *_elt; \
+ int i; \
+ _group++; \
+ _sacks = _tc->rcv_opts.sacks; \
+ for (i = 0; i < vec_len (_sacks); i++) \
+ { \
+ _sack = &_sacks[i]; \
+ vec_add2 (_sb->trace, _elt, 1); \
+ _elt->start = _sack->start; \
+ _elt->end = _sack->end; \
+ _elt->ack = _elt->end == _ack ? _ack : 0; \
+ _elt->snd_una_max = _elt->end == _ack ? _tc->snd_una_max : 0; \
+ _elt->group = _group; \
+ } \
+}
+#else
+#define tcp_scoreboard_trace_add(_tc, _ack)
+#endif
+
typedef enum _tcp_cc_algorithm_type
{
TCP_CC_NEWRENO,
@@ -405,6 +446,12 @@ tcp_connection_get_if_valid (u32 conn_index, u32 thread_index)
return pool_elt_at_index (tcp_main.connections[thread_index], conn_index);
}
+always_inline tcp_connection_t *
+tcp_get_connection_from_transport (transport_connection_t * tconn)
+{
+ return (tcp_connection_t *) tconn;
+}
+
void tcp_connection_close (tcp_connection_t * tc);
void tcp_connection_cleanup (tcp_connection_t * tc);
void tcp_connection_del (tcp_connection_t * tc);
@@ -414,6 +461,8 @@ u8 *format_tcp_connection_id (u8 * s, va_list * args);
u8 *format_tcp_connection (u8 * s, va_list * args);
u8 *format_tcp_scoreboard (u8 * s, va_list * args);
+u8 *tcp_scoreboard_replay (u8 * s, tcp_connection_t * tc, u8 verbose);
+
always_inline tcp_connection_t *
tcp_listener_get (u32 tli)
{
@@ -689,7 +738,7 @@ sack_scoreboard_hole_t *scoreboard_next_rxt_hole (sack_scoreboard_t * sb,
start, u8 have_sent_1_smss,
u8 * can_rescue,
u8 * snd_limited);
-void scoreboard_init_high_rxt (sack_scoreboard_t * sb);
+void scoreboard_init_high_rxt (sack_scoreboard_t * sb, u32 seq);
always_inline sack_scoreboard_hole_t *
scoreboard_get_hole (sack_scoreboard_t * sb, u32 index)
@@ -740,6 +789,7 @@ scoreboard_clear (sack_scoreboard_t * sb)
scoreboard_remove_hole (sb, hole);
}
ASSERT (sb->head == sb->tail && sb->head == TCP_INVALID_SACK_HOLE_INDEX);
+ ASSERT (pool_elts (sb->holes) == 0);
sb->sacked_bytes = 0;
sb->last_sacked_bytes = 0;
sb->last_bytes_delivered = 0;
@@ -759,6 +809,7 @@ scoreboard_hole_bytes (sack_scoreboard_hole_t * hole)
always_inline u32
scoreboard_hole_index (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
{
+ ASSERT (!pool_is_free_index (sb->holes, hole - sb->holes));
return hole - sb->holes;
}
diff --git a/src/vnet/tcp/tcp_input.c b/src/vnet/tcp/tcp_input.c
index 45db0da69c6..bc7d9015789 100644
--- a/src/vnet/tcp/tcp_input.c
+++ b/src/vnet/tcp/tcp_input.c
@@ -206,8 +206,8 @@ tcp_options_parse (tcp_header_t * th, tcp_options_t * to)
vec_reset_length (to->sacks);
for (j = 0; j < to->n_sack_blocks; j++)
{
- b.start = clib_net_to_host_u32 (*(u32 *) (data + 2 + 4 * j));
- b.end = clib_net_to_host_u32 (*(u32 *) (data + 6 + 4 * j));
+ b.start = clib_net_to_host_u32 (*(u32 *) (data + 2 + 8 * j));
+ b.end = clib_net_to_host_u32 (*(u32 *) (data + 6 + 8 * j));
vec_add1 (to->sacks, b);
}
break;
@@ -540,6 +540,10 @@ scoreboard_remove_hole (sack_scoreboard_t * sb, sack_scoreboard_hole_t * hole)
if (scoreboard_hole_index (sb, hole) == sb->cur_rxt_hole)
sb->cur_rxt_hole = TCP_INVALID_SACK_HOLE_INDEX;
+ /* Poison the entry */
+ if (CLIB_DEBUG > 0)
+ memset (hole, 0xfe, sizeof (*hole));
+
pool_put (sb->holes, hole);
}
@@ -555,7 +559,7 @@ scoreboard_insert_hole (sack_scoreboard_t * sb, u32 prev_index,
hole->start = start;
hole->end = end;
- hole_index = hole - sb->holes;
+ hole_index = scoreboard_hole_index (sb, hole);
prev = scoreboard_get_hole (sb, prev_index);
if (prev)
@@ -680,12 +684,30 @@ scoreboard_next_rxt_hole (sack_scoreboard_t * sb,
}
void
-scoreboard_init_high_rxt (sack_scoreboard_t * sb)
+scoreboard_init_high_rxt (sack_scoreboard_t * sb, u32 seq)
{
sack_scoreboard_hole_t *hole;
hole = scoreboard_first_hole (sb);
- sb->high_rxt = hole->start;
- sb->cur_rxt_hole = sb->head;
+ if (hole)
+ {
+ seq = seq_gt (seq, hole->start) ? seq : hole->start;
+ sb->cur_rxt_hole = sb->head;
+ }
+ sb->high_rxt = seq;
+}
+
+/**
+ * Test that scoreboard is sane after recovery
+ *
+ * Returns 1 if scoreboard is empty or if first hole beyond
+ * snd_una.
+ */
+u8
+tcp_scoreboard_is_sane_post_recovery (tcp_connection_t * tc)
+{
+ sack_scoreboard_hole_t *hole;
+ hole = scoreboard_first_hole (&tc->sack_sb);
+ return (!hole || seq_geq (hole->start, tc->snd_una));
}
void
@@ -712,7 +734,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
{
if (seq_lt (blk->start, blk->end)
&& seq_gt (blk->start, tc->snd_una)
- && seq_gt (blk->start, ack) && seq_leq (blk->end, tc->snd_nxt))
+ && seq_gt (blk->start, ack) && seq_leq (blk->end, tc->snd_una_max))
{
blk++;
continue;
@@ -731,6 +753,8 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
if (vec_len (tc->rcv_opts.sacks) == 0)
return;
+ tcp_scoreboard_trace_add (tc, ack);
+
/* Make sure blocks are ordered */
for (i = 0; i < vec_len (tc->rcv_opts.sacks); i++)
for (j = i + 1; j < vec_len (tc->rcv_opts.sacks); j++)
@@ -797,7 +821,7 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
sb->last_bytes_delivered +=
next_hole->start - hole->end;
}
- else if (!next_hole)
+ else
{
ASSERT (seq_geq (sb->high_sacked, ack));
sb->snd_una_adv = sb->high_sacked - ack;
@@ -824,12 +848,14 @@ tcp_rcv_sacks (tcp_connection_t * tc, u32 ack)
if (seq_lt (blk->end, hole->end))
{
hole_index = scoreboard_hole_index (sb, hole);
- scoreboard_insert_hole (sb, hole_index, blk->end, hole->end);
+ next_hole = scoreboard_insert_hole (sb, hole_index, blk->end,
+ hole->end);
/* Pool might've moved */
hole = scoreboard_get_hole (sb, hole_index);
hole->end = blk->start;
blk_index++;
+ ASSERT (hole->next == scoreboard_hole_index (sb, next_hole));
}
else if (seq_lt (blk->start, hole->end))
{
@@ -957,7 +983,7 @@ tcp_cc_recover (tcp_connection_t * tc)
ASSERT (tc->rto_boff == 0);
ASSERT (!tcp_in_cong_recovery (tc));
-
+ ASSERT (tcp_scoreboard_is_sane_post_recovery (tc));
TCP_EVT_DBG (TCP_EVT_CC_EVT, tc, 3);
return 0;
}
@@ -965,7 +991,7 @@ tcp_cc_recover (tcp_connection_t * tc)
static void
tcp_cc_update (tcp_connection_t * tc, vlib_buffer_t * b)
{
- ASSERT (!tcp_in_cong_recovery (tc));
+ ASSERT (!tcp_in_cong_recovery (tc) || tcp_is_lost_fin (tc));
/* Congestion avoidance */
tc->cc_algo->rcv_ack (tc);
@@ -1064,10 +1090,10 @@ tcp_cc_handle_event (tcp_connection_t * tc, u32 is_dack)
ASSERT (tc->cwnd >= tc->snd_mss);
/* If cwnd allows, send more data */
- if (tcp_opts_sack_permitted (&tc->rcv_opts)
- && scoreboard_first_hole (&tc->sack_sb))
+ if (tcp_opts_sack_permitted (&tc->rcv_opts))
{
- scoreboard_init_high_rxt (&tc->sack_sb);
+ scoreboard_init_high_rxt (&tc->sack_sb,
+ tc->snd_una + tc->snd_mss);
tcp_fast_retransmit_sack (tc);
}
else
@@ -1134,12 +1160,13 @@ partial_ack:
/* Remove retransmitted bytes that have been delivered */
ASSERT (tc->bytes_acked + tc->sack_sb.snd_una_adv
>= tc->sack_sb.last_bytes_delivered);
- rxt_delivered = tc->bytes_acked + tc->sack_sb.snd_una_adv
- - tc->sack_sb.last_bytes_delivered;
- if (0 && rxt_delivered && seq_gt (tc->sack_sb.high_rxt, tc->snd_una))
+
+ if (seq_lt (tc->snd_una, tc->sack_sb.high_rxt))
{
/* If we have sacks and we haven't gotten an ack beyond high_rxt,
* remove sacked bytes delivered */
+ rxt_delivered = tc->bytes_acked + tc->sack_sb.snd_una_adv
+ - tc->sack_sb.last_bytes_delivered;
ASSERT (tc->snd_rxt_bytes >= rxt_delivered);
tc->snd_rxt_bytes -= rxt_delivered;
}
@@ -1256,6 +1283,18 @@ tcp_rcv_ack (tcp_connection_t * tc, vlib_buffer_t * b,
return 0;
}
+static u8
+tcp_sack_vector_is_sane (sack_block_t * sacks)
+{
+ int i;
+ for (i = 1; i < vec_len (sacks); i++)
+ {
+ if (sacks[i - 1].end == sacks[i].start)
+ return 0;
+ }
+ return 1;
+}
+
/**
* Build SACK list as per RFC2018.
*
@@ -1316,6 +1355,9 @@ tcp_update_sack_list (tcp_connection_t * tc, u32 start, u32 end)
/* Replace old vector with new one */
vec_free (tc->snd_sacks);
tc->snd_sacks = new_list;
+
+ /* Segments should not 'touch' */
+ ASSERT (tcp_sack_vector_is_sane (tc->snd_sacks));
}
/** Enqueue data for delivery to application */
@@ -1330,7 +1372,6 @@ tcp_session_enqueue_data (tcp_connection_t * tc, vlib_buffer_t * b,
/* Pure ACK. Update rcv_nxt and be done. */
if (PREDICT_FALSE (data_len == 0))
{
- tc->rcv_nxt = vnet_buffer (b)->tcp.seq_end;
return TCP_ERROR_PURE_ACK;
}
@@ -1385,7 +1426,7 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
u16 data_len)
{
stream_session_t *s0;
- int rv;
+ int rv, offset;
ASSERT (seq_gt (vnet_buffer (b)->tcp.seq_number, tc->rcv_nxt));
@@ -1421,12 +1462,12 @@ tcp_session_enqueue_ooo (tcp_connection_t * tc, vlib_buffer_t * b,
newest = svm_fifo_newest_ooo_segment (s0->server_rx_fifo);
if (newest)
{
- start =
- tc->rcv_nxt + ooo_segment_offset (s0->server_rx_fifo, newest);
+ offset = ooo_segment_offset (s0->server_rx_fifo, newest);
+ ASSERT (offset <= vnet_buffer (b)->tcp.seq_number - tc->rcv_nxt);
+ start = tc->rcv_nxt + offset;
end = start + ooo_segment_length (s0->server_rx_fifo, newest);
tcp_update_sack_list (tc, start, end);
-
- ASSERT (seq_gt (start, tc->rcv_nxt));
+ svm_fifo_newest_ooo_segment_reset (s0->server_rx_fifo);
}
}
@@ -2736,12 +2777,12 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
/* lookup session */
tc0 =
(tcp_connection_t *)
- stream_session_lookup_transport4 (&ip40->dst_address,
- &ip40->src_address,
- tcp0->dst_port,
- tcp0->src_port,
- SESSION_TYPE_IP4_TCP,
- my_thread_index);
+ stream_session_lookup_transport_wt4 (&ip40->dst_address,
+ &ip40->src_address,
+ tcp0->dst_port,
+ tcp0->src_port,
+ SESSION_TYPE_IP4_TCP,
+ my_thread_index);
}
else
{
@@ -2754,12 +2795,12 @@ tcp46_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
tc0 =
(tcp_connection_t *)
- stream_session_lookup_transport6 (&ip60->src_address,
- &ip60->dst_address,
- tcp0->src_port,
- tcp0->dst_port,
- SESSION_TYPE_IP6_TCP,
- my_thread_index);
+ stream_session_lookup_transport_wt6 (&ip60->src_address,
+ &ip60->dst_address,
+ tcp0->src_port,
+ tcp0->dst_port,
+ SESSION_TYPE_IP6_TCP,
+ my_thread_index);
}
/* Length check */
@@ -2931,6 +2972,8 @@ do { \
_(ESTABLISHED, TCP_FLAG_RST | TCP_FLAG_ACK, TCP_INPUT_NEXT_ESTABLISHED,
TCP_ERROR_NONE);
_(ESTABLISHED, TCP_FLAG_SYN, TCP_INPUT_NEXT_ESTABLISHED, TCP_ERROR_NONE);
+ _(ESTABLISHED, TCP_FLAG_SYN | TCP_FLAG_ACK, TCP_INPUT_NEXT_ESTABLISHED,
+ TCP_ERROR_NONE);
/* ACK or FIN-ACK to our FIN */
_(FIN_WAIT_1, TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
_(FIN_WAIT_1, TCP_FLAG_ACK | TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS,
@@ -2954,6 +2997,7 @@ do { \
_(TIME_WAIT, TCP_FLAG_FIN, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
_(TIME_WAIT, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RCV_PROCESS,
TCP_ERROR_NONE);
+ _(TIME_WAIT, TCP_FLAG_RST, TCP_INPUT_NEXT_RCV_PROCESS, TCP_ERROR_NONE);
_(CLOSED, TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET, TCP_ERROR_CONNECTION_CLOSED);
_(CLOSED, TCP_FLAG_RST, TCP_INPUT_NEXT_DROP, TCP_ERROR_CONNECTION_CLOSED);
_(CLOSED, TCP_FLAG_FIN | TCP_FLAG_ACK, TCP_INPUT_NEXT_RESET,
diff --git a/src/vnet/tcp/tcp_test.c b/src/vnet/tcp/tcp_test.c
index 510deb4fec0..f37ba96dab8 100644
--- a/src/vnet/tcp/tcp_test.c
+++ b/src/vnet/tcp/tcp_test.c
@@ -34,6 +34,38 @@
} \
}
+/* *INDENT-OFF* */
+scoreboard_trace_elt_t sb_trace[] = {};
+/* *INDENT-ON* */
+
+static int
+tcp_test_scoreboard_replay (vlib_main_t * vm, unformat_input_t * input)
+{
+ int verbose = 0;
+ tcp_connection_t _tc, *tc = &_tc;
+ u8 *s = 0;
+
+ while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+ {
+ if (unformat (input, "detail"))
+ verbose = 1;
+ else
+ {
+ clib_error_t *e = clib_error_return
+ (0, "unknown input `%U'", format_unformat_error, input);
+ clib_error_report (e);
+ return -1;
+ }
+ }
+
+#if TCP_SCOREBOARD_TRACE
+ tc->sack_sb.trace = sb_trace;
+#endif
+ s = tcp_scoreboard_replay (s, tc, verbose);
+ vlib_cli_output (vm, "%v", s);
+ return 0;
+}
+
static int
tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
{
@@ -47,6 +79,8 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
{
if (unformat (input, "verbose"))
verbose = 1;
+ else if (unformat (input, "replay"))
+ return tcp_test_scoreboard_replay (vm, input);
}
memset (tc, 0, sizeof (*tc));
@@ -282,6 +316,44 @@ tcp_test_sack_rx (vlib_main_t * vm, unformat_input_t * input)
TCP_TEST ((sb->last_bytes_delivered == 400),
"last bytes delivered %d", sb->last_bytes_delivered);
+ /*
+ * One hole close to head, patch head, split in two and start acking
+ * the lowest part
+ */
+ scoreboard_clear (sb);
+ tc->snd_una = 0;
+ tc->snd_una_max = 1000;
+ tc->snd_nxt = 1000;
+
+ block.start = 500;
+ block.end = 1000;
+ vec_add1 (tc->rcv_opts.sacks, block);
+ tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
+
+ tcp_rcv_sacks (tc, 0);
+ if (verbose)
+ vlib_cli_output (vm, "sb added [500, 1000]:\n%U",
+ format_tcp_scoreboard, sb);
+
+ vec_reset_length (tc->rcv_opts.sacks);
+ block.start = 300;
+ block.end = 400;
+ vec_add1 (tc->rcv_opts.sacks, block);
+ tc->rcv_opts.n_sack_blocks = vec_len (tc->rcv_opts.sacks);
+ tcp_rcv_sacks (tc, 100);
+ if (verbose)
+ vlib_cli_output (vm, "sb added [0, 100] [300, 400]:\n%U",
+ format_tcp_scoreboard, sb);
+ TCP_TEST ((pool_elts (sb->holes) == 2),
+ "scoreboard has %d elements", pool_elts (sb->holes));
+
+ tc->snd_una = 100;
+ tcp_rcv_sacks (tc, 200);
+ tcp_rcv_sacks (tc, 300);
+ if (verbose)
+ vlib_cli_output (vm, "sb added [0, 300]:\n%U", format_tcp_scoreboard, sb);
+ TCP_TEST ((sb->sacked_bytes == 500), "sacked bytes %d", sb->sacked_bytes);
+
return 0;
}
@@ -390,6 +462,37 @@ tcp_test_sack_tx (vlib_main_t * vm, unformat_input_t * input)
vlib_cli_output (vm, "advance rcv_nxt to 1200\n%U", format_tcp_sacks, tc);
TCP_TEST ((vec_len (tc->snd_sacks) == 0), "sack blocks %d expected %d",
vec_len (tc->snd_sacks), 0);
+
+
+ /*
+ * Add 2 blocks, overwrite first and update rcv_nxt to also remove it
+ */
+
+ vec_reset_length (tc->snd_sacks);
+ tc->rcv_nxt = 0;
+
+ tcp_update_sack_list (tc, 100, 200);
+ tcp_update_sack_list (tc, 300, 400);
+
+ if (verbose)
+ vlib_cli_output (vm, "add [100, 200] [300, 400]\n%U",
+ format_tcp_sacks, tc);
+ TCP_TEST ((vec_len (tc->snd_sacks) == 2),
+ "sack blocks %d expected %d", vec_len (tc->snd_sacks), 2);
+ TCP_TEST ((tc->snd_sacks[0].start == 300),
+ "first sack block start %u expected %u", tc->snd_sacks[0].start,
+ 300);
+
+ tc->rcv_nxt = 100;
+ tcp_update_sack_list (tc, 100, 100);
+ if (verbose)
+ vlib_cli_output (vm, "add [100, 200] rcv_nxt = 100\n%U",
+ format_tcp_sacks, tc);
+ TCP_TEST ((vec_len (tc->snd_sacks) == 1),
+ "sack blocks %d expected %d", vec_len (tc->snd_sacks), 1);
+ TCP_TEST ((tc->snd_sacks[0].start == 300),
+ "first sack block start %u expected %u", tc->snd_sacks[0].start,
+ 300);
return 0;
}
@@ -1188,6 +1291,176 @@ tcp_test_fifo4 (vlib_main_t * vm, unformat_input_t * input)
return 0;
}
+static u32
+fifo_pos (svm_fifo_t * f, u32 pos)
+{
+ return pos % f->nitems;
+}
+
+static int
+tcp_test_fifo5 (vlib_main_t * vm, unformat_input_t * input)
+{
+ svm_fifo_t *f;
+ u32 fifo_size = 400, j = 0, offset = 200;
+ int i, rv, verbose = 0;
+ u8 *test_data = 0, *data_buf = 0;
+ ooo_segment_t *ooo_seg;
+
+ while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+ {
+ if (unformat (input, "verbose"))
+ verbose = 1;
+ else
+ {
+ clib_error_t *e = clib_error_return
+ (0, "unknown input `%U'", format_unformat_error, input);
+ clib_error_report (e);
+ return -1;
+ }
+ }
+
+ f = fifo_prepare (fifo_size);
+ svm_fifo_init_pointers (f, offset);
+
+ vec_validate (test_data, 399);
+ for (i = 0; i < vec_len (test_data); i++)
+ test_data[i] = i % 0xff;
+
+ /*
+ * Start with [100, 200] and [300, 400]
+ */
+ svm_fifo_enqueue_with_offset (f, 100, 100, &test_data[100]);
+ svm_fifo_enqueue_with_offset (f, 300, 100, &test_data[300]);
+
+ TCP_TEST ((svm_fifo_number_ooo_segments (f) == 2),
+ "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+ TCP_TEST ((f->ooos_newest == 1), "newest %u", f->ooos_newest);
+ if (verbose)
+ vlib_cli_output (vm, "fifo after [100, 200] and [300, 400] : %U",
+ format_svm_fifo, f, 2 /* verbose */ );
+
+ /*
+ * Add [225, 275]
+ */
+
+ rv = svm_fifo_enqueue_with_offset (f, 225, 50, &test_data[200]);
+ if (verbose)
+ vlib_cli_output (vm, "fifo after [225, 275] : %U",
+ format_svm_fifo, f, 2 /* verbose */ );
+ TCP_TEST ((svm_fifo_number_ooo_segments (f) == 3),
+ "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+ ooo_seg = svm_fifo_first_ooo_segment (f);
+ TCP_TEST ((ooo_seg->start == fifo_pos (f, 100 + offset)),
+ "first seg start %u expected %u", ooo_seg->start,
+ fifo_pos (f, 100 + offset));
+ TCP_TEST ((ooo_seg->length == 100), "first seg length %u expected %u",
+ ooo_seg->length, 100);
+ ooo_seg = ooo_segment_next (f, ooo_seg);
+ TCP_TEST ((ooo_seg->start == fifo_pos (f, 225 + offset)),
+ "second seg start %u expected %u",
+ ooo_seg->start, fifo_pos (f, 225 + offset));
+ TCP_TEST ((ooo_seg->length == 50), "second seg length %u expected %u",
+ ooo_seg->length, 50);
+ ooo_seg = ooo_segment_next (f, ooo_seg);
+ TCP_TEST ((ooo_seg->start == fifo_pos (f, 300 + offset)),
+ "third seg start %u expected %u",
+ ooo_seg->start, fifo_pos (f, 300 + offset));
+ TCP_TEST ((ooo_seg->length == 100), "third seg length %u expected %u",
+ ooo_seg->length, 100);
+ TCP_TEST ((f->ooos_newest == 2), "newest %u", f->ooos_newest);
+ /*
+ * Add [190, 310]
+ */
+ rv = svm_fifo_enqueue_with_offset (f, 190, 120, &test_data[190]);
+ if (verbose)
+ vlib_cli_output (vm, "fifo after [190, 310] : %U",
+ format_svm_fifo, f, 1 /* verbose */ );
+ TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+ "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+ ooo_seg = svm_fifo_first_ooo_segment (f);
+ TCP_TEST ((ooo_seg->start == fifo_pos (f, offset + 100)),
+ "first seg start %u expected %u",
+ ooo_seg->start, fifo_pos (f, offset + 100));
+ TCP_TEST ((ooo_seg->length == 300), "first seg length %u expected %u",
+ ooo_seg->length, 300);
+
+ /*
+ * Add [0, 150]
+ */
+ rv = svm_fifo_enqueue_nowait (f, 150, test_data);
+
+ if (verbose)
+ vlib_cli_output (vm, "fifo after [0 150] : %U", format_svm_fifo, f,
+ 2 /* verbose */ );
+
+ TCP_TEST ((rv == 400), "managed to enqueue %u expected %u", rv, 400);
+ TCP_TEST ((svm_fifo_number_ooo_segments (f) == 0),
+ "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+
+ vec_validate (data_buf, 399);
+ svm_fifo_peek (f, 0, 400, data_buf);
+ if (compare_data (data_buf, test_data, 0, 400, &j))
+ {
+ TCP_TEST (0, "[%d] peeked %u expected %u", j, data_buf[j],
+ test_data[j]);
+ }
+
+ /*
+ * Add [100 200] and overlap it with [50 250]
+ */
+ svm_fifo_free (f);
+ f = fifo_prepare (fifo_size);
+
+ svm_fifo_enqueue_with_offset (f, 100, 100, &test_data[100]);
+ svm_fifo_enqueue_with_offset (f, 50, 200, &test_data[50]);
+ TCP_TEST ((svm_fifo_number_ooo_segments (f) == 1),
+ "number of ooo segments %u", svm_fifo_number_ooo_segments (f));
+ ooo_seg = svm_fifo_first_ooo_segment (f);
+ TCP_TEST ((ooo_seg->start == 50), "first seg start %u expected %u",
+ ooo_seg->start, 50);
+ TCP_TEST ((ooo_seg->length == 200), "first seg length %u expected %u",
+ ooo_seg->length, 200);
+
+ svm_fifo_free (f);
+ vec_free (test_data);
+ return 0;
+}
+
+/* *INDENT-OFF* */
+svm_fifo_trace_elem_t fifo_trace[] = {};
+/* *INDENT-ON* */
+
+static int
+tcp_test_fifo_replay (vlib_main_t * vm, unformat_input_t * input)
+{
+ svm_fifo_t f;
+ int verbose = 0;
+ u8 no_read = 0, *str = 0;
+
+ while (unformat_check_input (input) != UNFORMAT_END_OF_INPUT)
+ {
+ if (unformat (input, "verbose"))
+ verbose = 1;
+ else if (unformat (input, "no-read"))
+ no_read = 1;
+ else
+ {
+ clib_error_t *e = clib_error_return
+ (0, "unknown input `%U'", format_unformat_error, input);
+ clib_error_report (e);
+ return -1;
+ }
+ }
+
+#if SVMF_FIFO_TRACE
+ f.trace = fifo_trace;
+#endif
+
+ str = svm_fifo_replay (str, &f, no_read, verbose);
+ vlib_cli_output (vm, "%v", str);
+ return 0;
+}
+
static int
tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
{
@@ -1237,6 +1510,14 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
if (tcp_test_fifo3 (vm, input))
return -1;
unformat_free (input);
+
+ res = tcp_test_fifo4 (vm, input);
+ if (res)
+ return res;
+
+ res = tcp_test_fifo5 (vm, input);
+ if (res)
+ return res;
}
else
{
@@ -1256,6 +1537,14 @@ tcp_test_fifo (vlib_main_t * vm, unformat_input_t * input)
{
res = tcp_test_fifo4 (vm, input);
}
+ else if (unformat (input, "fifo5"))
+ {
+ res = tcp_test_fifo5 (vm, input);
+ }
+ else if (unformat (input, "replay"))
+ {
+ res = tcp_test_fifo_replay (vm, input);
+ }
}
return res;