ip: reassembly: send packet out on correct worker
Note which worker received fragment with offset zero and use this worker
to send out the reassembled packet.
Type: fix
Change-Id: I1d3cee16788db3b230682525239c0100d51dc380
Signed-off-by: Klement Sekera <ksekera@cisco.com>
diff --git a/src/vnet/ip/ip4_reassembly.c b/src/vnet/ip/ip4_reassembly.c
index b82bafe..7159b8a 100644
--- a/src/vnet/ip/ip4_reassembly.c
+++ b/src/vnet/ip/ip4_reassembly.c
@@ -25,6 +25,7 @@
#include <vnet/ip/ip.h>
#include <vppinfra/bihash_16_8.h>
#include <vnet/ip/ip4_reassembly.h>
+#include <stddef.h>
#define MSEC_PER_SEC 1000
#define IP4_REASS_TIMEOUT_DEFAULT_MS 100
@@ -61,6 +62,7 @@
IP4_REASS_RC_TOO_MANY_FRAGMENTS,
IP4_REASS_RC_INTERNAL_ERROR,
IP4_REASS_RC_NO_BUF,
+ IP4_REASS_RC_HANDOFF,
} ip4_reass_rc_t;
typedef struct
@@ -85,7 +87,7 @@
struct
{
u32 reass_index;
- u32 thread_index;
+ u32 memory_owner_thread_index;
};
u64 as_u64;
} ip4_reass_val_t;
@@ -139,10 +141,16 @@
u16 min_fragment_length;
// number of fragments in this reassembly
u32 fragments_n;
+ // thread owning memory for this context (whose pool contains this ctx)
+ u32 memory_owner_thread_index;
+ // thread which received fragment with offset 0 and which sends out the
+ // completed reassembly
+ u32 sendout_thread_index;
} ip4_reass_t;
typedef struct
{
+ // pool of reassembly contexts
ip4_reass_t *pool;
u32 reass_n;
u32 id_counter;
@@ -167,7 +175,6 @@
// convenience
vlib_main_t *vlib_main;
- vnet_main_t *vnet_main;
// node index of ip4-drop node
u32 ip4_drop_idx;
@@ -176,7 +183,6 @@
/** Worker handoff */
u32 fq_index;
u32 fq_feature_index;
-
} ip4_reass_main_t;
extern ip4_reass_main_t ip4_reass_main;
@@ -200,6 +206,7 @@
RANGE_DISCARD,
RANGE_OVERLAP,
FINALIZE,
+ HANDOFF,
} ip4_reass_trace_operation_e;
typedef struct
@@ -219,6 +226,8 @@
ip4_reass_range_trace_t trace_range;
u32 size_diff;
u32 op_id;
+ u32 thread_id;
+ u32 thread_id_to;
u32 fragment_first;
u32 fragment_last;
u32 total_data_len;
@@ -256,11 +265,17 @@
CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *);
CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
ip4_reass_trace_t *t = va_arg (*args, ip4_reass_trace_t *);
- s = format (s, "reass id: %u, op id: %u ", t->reass_id, t->op_id);
- u32 indent = format_get_indent (s);
- s = format (s, "first bi: %u, data len: %u, ip/fragment[%u, %u]",
- t->trace_range.first_bi, t->total_data_len, t->fragment_first,
- t->fragment_last);
+ u32 indent = 0;
+ if (~0 != t->reass_id)
+ {
+ s = format (s, "reass id: %u, op id: %u, ", t->reass_id, t->op_id);
+ indent = format_get_indent (s);
+ s =
+ format (s,
+ "first bi: %u, data len: %u, ip/fragment[%u, %u]",
+ t->trace_range.first_bi, t->total_data_len, t->fragment_first,
+ t->fragment_last);
+ }
switch (t->action)
{
case RANGE_SHRINK:
@@ -283,28 +298,36 @@
case FINALIZE:
s = format (s, "\n%Ufinalize reassembly", format_white_space, indent);
break;
+ case HANDOFF:
+ s =
+ format (s, "handoff from thread #%u to thread #%u", t->thread_id,
+ t->thread_id_to);
+ break;
}
return s;
}
static void
ip4_reass_add_trace (vlib_main_t * vm, vlib_node_runtime_t * node,
- ip4_reass_main_t * rm, ip4_reass_t * reass, u32 bi,
- ip4_reass_trace_operation_e action, u32 size_diff)
+ ip4_reass_main_t * rm, u32 reass_id, u32 op_id,
+ u32 bi, u32 first_bi, u32 data_len,
+ ip4_reass_trace_operation_e action, u32 size_diff,
+ u32 thread_id_to)
{
vlib_buffer_t *b = vlib_get_buffer (vm, bi);
vnet_buffer_opaque_t *vnb = vnet_buffer (b);
ip4_reass_trace_t *t = vlib_add_trace (vm, node, b, sizeof (t[0]));
- t->reass_id = reass->id;
+ t->reass_id = reass_id;
t->action = action;
ip4_reass_trace_details (vm, bi, &t->trace_range);
t->size_diff = size_diff;
- t->op_id = reass->trace_op_counter;
- ++reass->trace_op_counter;
+ t->op_id = op_id;
+ t->thread_id = vm->thread_index;
+ t->thread_id_to = thread_id_to;
t->fragment_first = vnb->ip.reass.fragment_first;
t->fragment_last = vnb->ip.reass.fragment_last;
- t->trace_range.first_bi = reass->first_bi;
- t->total_data_len = reass->data_len;
+ t->trace_range.first_bi = first_bi;
+ t->total_data_len = data_len;
#if 0
static u8 *s = NULL;
s = format (s, "%U", format_ip4_reass_trace, NULL, NULL, t);
@@ -314,17 +337,22 @@
#endif
}
+always_inline void
+ip4_reass_free_ctx (ip4_reass_per_thread_t * rt, ip4_reass_t * reass)
+{
+ pool_put (rt->pool, reass);
+ --rt->reass_n;
+}
always_inline void
-ip4_reass_free (ip4_reass_main_t * rm, ip4_reass_per_thread_t * rt,
- ip4_reass_t * reass)
+ip4_reass_free (vlib_main_t * vm, ip4_reass_main_t * rm,
+ ip4_reass_per_thread_t * rt, ip4_reass_t * reass)
{
clib_bihash_kv_16_8_t kv;
kv.key[0] = reass->key.as_u64[0];
kv.key[1] = reass->key.as_u64[1];
clib_bihash_add_del_16_8 (&rm->hash, &kv, 0);
- pool_put (rt->pool, reass);
- --rt->reass_n;
+ return ip4_reass_free_ctx (rt, reass);
}
always_inline void
@@ -396,23 +424,30 @@
ip4_reass_main_t * rm, ip4_reass_per_thread_t * rt,
ip4_reass_kv_t * kv, u8 * do_handoff)
{
- ip4_reass_t *reass = NULL;
- f64 now = vlib_time_now (rm->vlib_main);
+ ip4_reass_t *reass;
+ f64 now;
+again:
+
+ reass = NULL;
+ now = vlib_time_now (vm);
if (!clib_bihash_search_16_8
(&rm->hash, (clib_bihash_kv_16_8_t *) kv, (clib_bihash_kv_16_8_t *) kv))
{
- if (vm->thread_index != kv->v.thread_index)
+ reass =
+ pool_elt_at_index (rm->per_thread_data
+ [kv->v.memory_owner_thread_index].pool,
+ kv->v.reass_index);
+ if (vm->thread_index != reass->memory_owner_thread_index)
{
*do_handoff = 1;
- return NULL;
+ return reass;
}
- reass = pool_elt_at_index (rt->pool, kv->v.reass_index);
if (now > reass->last_heard + rm->timeout)
{
ip4_reass_drop_all (vm, node, rm, reass);
- ip4_reass_free (rm, rt, reass);
+ ip4_reass_free (vm, rm, rt, reass);
reass = NULL;
}
}
@@ -433,6 +468,7 @@
pool_get (rt->pool, reass);
clib_memset (reass, 0, sizeof (*reass));
reass->id = ((u64) vm->thread_index * 1000000000) + rt->id_counter;
+ reass->memory_owner_thread_index = vm->thread_index;
++rt->id_counter;
reass->first_bi = ~0;
reass->last_packet_octet = ~0;
@@ -445,13 +481,18 @@
reass->key.as_u64[0] = ((clib_bihash_kv_16_8_t *) kv)->key[0];
reass->key.as_u64[1] = ((clib_bihash_kv_16_8_t *) kv)->key[1];
kv->v.reass_index = (reass - rt->pool);
- kv->v.thread_index = vm->thread_index;
+ kv->v.memory_owner_thread_index = vm->thread_index;
reass->last_heard = now;
- if (clib_bihash_add_del_16_8 (&rm->hash, (clib_bihash_kv_16_8_t *) kv, 1))
+ int rv =
+ clib_bihash_add_del_16_8 (&rm->hash, (clib_bihash_kv_16_8_t *) kv, 2);
+ if (rv)
{
- ip4_reass_free (rm, rt, reass);
+ ip4_reass_free_ctx (rt, reass);
reass = NULL;
+ // if other worker created a context already work with the other copy
+ if (-2 == rv)
+ goto again;
}
return reass;
@@ -615,7 +656,10 @@
first_b->flags &= ~VLIB_BUFFER_EXT_HDR_VALID;
if (PREDICT_FALSE (first_b->flags & VLIB_BUFFER_IS_TRACED))
{
- ip4_reass_add_trace (vm, node, rm, reass, reass->first_bi, FINALIZE, 0);
+ ip4_reass_add_trace (vm, node, rm, reass->id, reass->trace_op_counter,
+ reass->first_bi, reass->first_bi, reass->data_len,
+ FINALIZE, 0, ~0);
+ ++reass->trace_op_counter;
#if 0
// following code does a hexdump of packet fragments to stdout ...
do
@@ -654,7 +698,7 @@
}
vnet_buffer (first_b)->ip.reass.estimated_mtu = reass->min_fragment_length;
*error0 = IP4_ERROR_NONE;
- ip4_reass_free (rm, rt, reass);
+ ip4_reass_free (vm, rm, rt, reass);
reass = NULL;
return IP4_REASS_RC_OK;
}
@@ -728,8 +772,11 @@
u32 to_be_freed_bi = discard_bi;
if (PREDICT_FALSE (discard_b->flags & VLIB_BUFFER_IS_TRACED))
{
- ip4_reass_add_trace (vm, node, rm, reass, discard_bi, RANGE_DISCARD,
- 0);
+ ip4_reass_add_trace (vm, node, rm, reass->id,
+ reass->trace_op_counter, discard_bi,
+ reass->first_bi, reass->data_len,
+ RANGE_DISCARD, 0, ~0);
+ ++reass->trace_op_counter;
}
if (discard_b->flags & VLIB_BUFFER_NEXT_PRESENT)
{
@@ -753,7 +800,7 @@
ip4_reass_update (vlib_main_t * vm, vlib_node_runtime_t * node,
ip4_reass_main_t * rm, ip4_reass_per_thread_t * rt,
ip4_reass_t * reass, u32 * bi0, u32 * next0, u32 * error0,
- bool is_custom_app)
+ bool is_custom_app, u32 * handoff_thread_idx)
{
ip4_reass_rc_t rc = IP4_REASS_RC_OK;
int consumed = 0;
@@ -794,7 +841,10 @@
}
if (PREDICT_FALSE (fb->flags & VLIB_BUFFER_IS_TRACED))
{
- ip4_reass_add_trace (vm, node, rm, reass, *bi0, RANGE_NEW, 0);
+ ip4_reass_add_trace (vm, node, rm, reass->id,
+ reass->trace_op_counter, *bi0, reass->first_bi,
+ reass->data_len, RANGE_NEW, 0, ~0);
+ ++reass->trace_op_counter;
}
*bi0 = ~0;
reass->min_fragment_length = clib_net_to_host_u16 (fip->length);
@@ -848,8 +898,11 @@
// this fragment is a (sub)part of existing range, ignore it
if (PREDICT_FALSE (fb->flags & VLIB_BUFFER_IS_TRACED))
{
- ip4_reass_add_trace (vm, node, rm, reass, *bi0,
- RANGE_OVERLAP, 0);
+ ip4_reass_add_trace (vm, node, rm, reass->id,
+ reass->trace_op_counter, *bi0,
+ reass->first_bi, reass->data_len,
+ RANGE_OVERLAP, 0, ~0);
+ ++reass->trace_op_counter;
}
break;
}
@@ -868,9 +921,12 @@
reass->data_len -= overlap;
if (PREDICT_FALSE (fb->flags & VLIB_BUFFER_IS_TRACED))
{
- ip4_reass_add_trace (vm, node, rm, reass,
- candidate_range_bi, RANGE_SHRINK,
- overlap);
+ ip4_reass_add_trace (vm, node, rm, reass->id,
+ reass->trace_op_counter,
+ candidate_range_bi,
+ reass->first_bi, reass->data_len,
+ RANGE_SHRINK, 0, ~0);
+ ++reass->trace_op_counter;
}
rc =
ip4_reass_insert_range_in_chain (vm, rm, rt, reass,
@@ -961,14 +1017,24 @@
{
if (PREDICT_FALSE (fb->flags & VLIB_BUFFER_IS_TRACED))
{
- ip4_reass_add_trace (vm, node, rm, reass, *bi0, RANGE_NEW, 0);
+ ip4_reass_add_trace (vm, node, rm, reass->id,
+ reass->trace_op_counter, *bi0, reass->first_bi,
+ reass->data_len, RANGE_NEW, 0, ~0);
+ ++reass->trace_op_counter;
}
}
if (~0 != reass->last_packet_octet &&
reass->data_len == reass->last_packet_octet + 1)
{
- return ip4_reass_finalize (vm, node, rm, rt, reass, bi0, next0, error0,
- is_custom_app);
+ *handoff_thread_idx = reass->sendout_thread_index;
+ rc =
+ ip4_reass_finalize (vm, node, rm, rt, reass, bi0, next0, error0,
+ is_custom_app);
+ if (IP4_REASS_RC_OK == rc
+ && reass->memory_owner_thread_index != reass->sendout_thread_index)
+ {
+ rc = IP4_REASS_RC_HANDOFF;
+ }
}
else
{
@@ -1056,33 +1122,53 @@
ip4_reass_t *reass =
ip4_reass_find_or_create (vm, node, rm, rt, &kv,
&do_handoff);
-
+ if (reass)
+ {
+ const u32 fragment_first =
+ ip4_get_fragment_offset_bytes (ip0);
+ if (0 == fragment_first)
+ {
+ reass->sendout_thread_index = vm->thread_index;
+ }
+ }
if (PREDICT_FALSE (do_handoff))
{
next0 = IP4_REASSEMBLY_NEXT_HANDOFF;
if (is_feature)
vnet_buffer (b0)->ip.
reass.owner_feature_thread_index =
- kv.v.thread_index;
+ kv.v.memory_owner_thread_index;
else
vnet_buffer (b0)->ip.reass.owner_thread_index =
- kv.v.thread_index;
+ kv.v.memory_owner_thread_index;
}
else if (reass)
{
+ u32 handoff_thread_idx;
switch (ip4_reass_update
(vm, node, rm, rt, reass, &bi0, &next0,
- &error0, is_custom_app))
+ &error0, is_custom_app, &handoff_thread_idx))
{
case IP4_REASS_RC_OK:
/* nothing to do here */
break;
+ case IP4_REASS_RC_HANDOFF:
+ next0 = IP4_REASSEMBLY_NEXT_HANDOFF;
+ b0 = vlib_get_buffer (vm, bi0);
+ if (is_feature)
+ vnet_buffer (b0)->ip.
+ reass.owner_feature_thread_index =
+ handoff_thread_idx;
+ else
+ vnet_buffer (b0)->ip.reass.owner_thread_index =
+ handoff_thread_idx;
+ break;
case IP4_REASS_RC_TOO_MANY_FRAGMENTS:
vlib_node_increment_counter (vm, node->node_index,
IP4_ERROR_REASS_FRAGMENT_CHAIN_TOO_LONG,
1);
ip4_reass_drop_all (vm, node, rm, reass);
- ip4_reass_free (rm, rt, reass);
+ ip4_reass_free (vm, rm, rt, reass);
goto next_packet;
break;
case IP4_REASS_RC_NO_BUF:
@@ -1090,7 +1176,7 @@
IP4_ERROR_REASS_NO_BUF,
1);
ip4_reass_drop_all (vm, node, rm, reass);
- ip4_reass_free (rm, rt, reass);
+ ip4_reass_free (vm, rm, rt, reass);
goto next_packet;
break;
case IP4_REASS_RC_INTERNAL_ERROR:
@@ -1099,7 +1185,7 @@
IP4_ERROR_REASS_INTERNAL_ERROR,
1);
ip4_reass_drop_all (vm, node, rm, reass);
- ip4_reass_free (rm, rt, reass);
+ ip4_reass_free (vm, rm, rt, reass);
goto next_packet;
break;
}
@@ -1119,7 +1205,24 @@
to_next[0] = bi0;
to_next += 1;
n_left_to_next -= 1;
- if (is_feature && IP4_ERROR_NONE == error0)
+ if (next0 == IP4_REASSEMBLY_NEXT_HANDOFF)
+ {
+ if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
+ {
+ if (is_feature)
+ ip4_reass_add_trace (vm, node, rm, ~0,
+ ~0,
+ bi0, ~0, ~0, HANDOFF, 0,
+ vnet_buffer (b0)->ip.
+ reass.owner_feature_thread_index);
+ else
+ ip4_reass_add_trace (vm, node, rm, ~0, ~0, bi0,
+ ~0, ~0, HANDOFF, 0,
+ vnet_buffer (b0)->ip.
+ reass.owner_thread_index);
+ }
+ }
+ else if (is_feature && IP4_ERROR_NONE == error0)
{
b0 = vlib_get_buffer (vm, bi0);
vnet_feature_next (&next0, b0);
@@ -1318,7 +1421,6 @@
vlib_node_t *node;
rm->vlib_main = vm;
- rm->vnet_main = vnet_get_main ();
vec_validate (rm->per_thread_data, vlib_num_workers ());
ip4_reass_per_thread_t *rt;
@@ -1348,7 +1450,6 @@
rm->fq_feature_index =
vlib_frame_queue_main_init (ip4_reass_node_feature.index, 0);
-
return error;
}
@@ -1410,7 +1511,7 @@
{
ip4_reass_t *reass = pool_elt_at_index (rt->pool, i[0]);
ip4_reass_drop_all (vm, node, rm, reass);
- ip4_reass_free (rm, rt, reass);
+ ip4_reass_free (vm, rm, rt, reass);
}
/* *INDENT-ON* */