vlib: improve node interrupt handling
- add ability to pass data together with interrupt
- avoid locking for local interrupts (same thread)
Type: improvement
Change-Id: I73a2ab2e716bb887a1f02c87788ae83e329f9b40
Signed-off-by: Damjan Marion <damarion@cisco.com>
diff --git a/src/vlib/main.c b/src/vlib/main.c
index c3ca8b1..2e100b2 100644
--- a/src/vlib/main.c
+++ b/src/vlib/main.c
@@ -1702,6 +1702,26 @@
{
}
+static_always_inline u64
+dispatch_pending_interrupts (vlib_main_t * vm, vlib_node_main_t * nm,
+ u64 cpu_time_now)
+{
+ vlib_node_runtime_t *n;
+
+ for (int i = 0; i < _vec_len (nm->pending_local_interrupts); i++)
+ {
+ vlib_node_interrupt_t *in;
+ in = vec_elt_at_index (nm->pending_local_interrupts, i);
+ n = vec_elt_at_index (nm->nodes_by_type[VLIB_NODE_TYPE_INPUT],
+ in->node_runtime_index);
+ n->interrupt_data = in->data;
+ cpu_time_now = dispatch_node (vm, n, VLIB_NODE_TYPE_INPUT,
+ VLIB_NODE_STATE_INTERRUPT, /* frame */ 0,
+ cpu_time_now);
+ }
+ vec_reset_length (nm->pending_local_interrupts);
+ return cpu_time_now;
+}
static_always_inline void
vlib_main_or_worker_loop (vlib_main_t * vm, int is_main)
@@ -1712,7 +1732,6 @@
u64 cpu_time_now;
f64 now;
vlib_frame_queue_main_t *fqm;
- u32 *last_node_runtime_indices = 0;
u32 frame_queue_check_counter = 0;
/* Initialize pending node vector. */
@@ -1732,10 +1751,9 @@
cpu_time_now = clib_cpu_time_now ();
/* Pre-allocate interupt runtime indices and lock. */
- vec_alloc (nm->pending_interrupt_node_runtime_indices, 32);
- vec_alloc (last_node_runtime_indices, 32);
- if (!is_main)
- clib_spinlock_init (&nm->pending_interrupt_lock);
+ vec_alloc (nm->pending_local_interrupts, 32);
+ vec_alloc (nm->pending_remote_interrupts, 32);
+ clib_spinlock_init (&nm->pending_interrupt_lock);
/* Pre-allocate expired nodes. */
if (!nm->polling_threshold_vector_length)
@@ -1821,40 +1839,27 @@
if (PREDICT_TRUE (is_main && vm->queue_signal_pending == 0))
vm->queue_signal_callback (vm);
- /* Next handle interrupts. */
- {
- /* unlocked read, for performance */
- uword l = _vec_len (nm->pending_interrupt_node_runtime_indices);
- uword i;
- if (PREDICT_FALSE (l > 0))
- {
- u32 *tmp;
- if (!is_main)
- {
- clib_spinlock_lock (&nm->pending_interrupt_lock);
- /* Re-read w/ lock held, in case another thread added an item */
- l = _vec_len (nm->pending_interrupt_node_runtime_indices);
- }
+ /* handle local interruots */
+ if (_vec_len (nm->pending_local_interrupts))
+ cpu_time_now = dispatch_pending_interrupts (vm, nm, cpu_time_now);
- tmp = nm->pending_interrupt_node_runtime_indices;
- nm->pending_interrupt_node_runtime_indices =
- last_node_runtime_indices;
- last_node_runtime_indices = tmp;
- _vec_len (last_node_runtime_indices) = 0;
- if (!is_main)
- clib_spinlock_unlock (&nm->pending_interrupt_lock);
- for (i = 0; i < l; i++)
- {
- n = vec_elt_at_index (nm->nodes_by_type[VLIB_NODE_TYPE_INPUT],
- last_node_runtime_indices[i]);
- cpu_time_now =
- dispatch_node (vm, n, VLIB_NODE_TYPE_INPUT,
- VLIB_NODE_STATE_INTERRUPT,
- /* frame */ 0,
- cpu_time_now);
- }
- }
- }
+ /* handle remote interruots */
+ if (_vec_len (nm->pending_remote_interrupts))
+ {
+ vlib_node_interrupt_t *in;
+
+ /* at this point it is known that
+ * vec_len (nm->pending_local_interrupts) is zero so we quickly swap
+ * local and remote vector under the spinlock */
+ clib_spinlock_lock (&nm->pending_interrupt_lock);
+ in = nm->pending_local_interrupts;
+ nm->pending_local_interrupts = nm->pending_remote_interrupts;
+ nm->pending_remote_interrupts = in;
+ clib_spinlock_unlock (&nm->pending_interrupt_lock);
+
+ cpu_time_now = dispatch_pending_interrupts (vm, nm, cpu_time_now);
+ }
+
/* Input nodes may have added work to the pending vector.
Process pending vector until there is nothing left.
All pending vectors will be processed from input -> output. */
diff --git a/src/vlib/node.h b/src/vlib/node.h
index 1bdb3bb..ca7564a 100644
--- a/src/vlib/node.h
+++ b/src/vlib/node.h
@@ -471,10 +471,6 @@
vlib_error_t *errors; /**< Vector of errors for this node. */
-#if __SIZEOF_POINTER__ == 4
- u8 pad[8];
-#endif
-
u32 clocks_since_last_overflow; /**< Number of clock cycles. */
u32 max_clock; /**< Maximum clock cycle for an
@@ -512,6 +508,10 @@
u16 state; /**< Input node state. */
+ u32 interrupt_data; /**< Data passed together with interrupt.
+ Valid only when state is
+ VLIB_NODE_STATE_INTERRUPT */
+
u16 n_next_nodes;
u16 cached_next_index; /**< Next frame index that vector
@@ -676,6 +676,12 @@
typedef struct
{
+ u32 node_runtime_index;
+ u32 data;
+} vlib_node_interrupt_t;
+
+typedef struct
+{
/* Public nodes. */
vlib_node_t **nodes;
@@ -690,7 +696,8 @@
vlib_node_runtime_t *nodes_by_type[VLIB_N_NODE_TYPE];
/* Node runtime indices for input nodes with pending interrupts. */
- u32 *pending_interrupt_node_runtime_indices;
+ vlib_node_interrupt_t *pending_local_interrupts;
+ vlib_node_interrupt_t *pending_remote_interrupts;
clib_spinlock_t pending_interrupt_lock;
/* Input nodes are switched from/to interrupt to/from polling mode
diff --git a/src/vlib/node_funcs.h b/src/vlib/node_funcs.h
index d6d04fb..263017d 100644
--- a/src/vlib/node_funcs.h
+++ b/src/vlib/node_funcs.h
@@ -194,14 +194,36 @@
}
always_inline void
-vlib_node_set_interrupt_pending (vlib_main_t * vm, u32 node_index)
+vlib_node_set_interrupt_pending_with_data (vlib_main_t * vm, u32 node_index,
+ u32 data)
{
vlib_node_main_t *nm = &vm->node_main;
vlib_node_t *n = vec_elt (nm->nodes, node_index);
+ vlib_node_interrupt_t *i;
ASSERT (n->type == VLIB_NODE_TYPE_INPUT);
- clib_spinlock_lock_if_init (&nm->pending_interrupt_lock);
- vec_add1 (nm->pending_interrupt_node_runtime_indices, n->runtime_index);
- clib_spinlock_unlock_if_init (&nm->pending_interrupt_lock);
+
+ if (vm == vlib_get_main ())
+ {
+ /* local thread */
+ vec_add2 (nm->pending_local_interrupts, i, 1);
+ i->node_runtime_index = n->runtime_index;
+ i->data = data;
+ }
+ else
+ {
+ /* remote thread */
+ clib_spinlock_lock (&nm->pending_interrupt_lock);
+ vec_add2 (nm->pending_remote_interrupts, i, 1);
+ i->node_runtime_index = n->runtime_index;
+ i->data = data;
+ clib_spinlock_unlock (&nm->pending_interrupt_lock);
+ }
+}
+
+always_inline void
+vlib_node_set_interrupt_pending (vlib_main_t * vm, u32 node_index)
+{
+ vlib_node_set_interrupt_pending_with_data (vm, node_index, 0);
}
always_inline vlib_process_t *