Add minimum burst throughout the scheduler pipeline and a flush counter.
Use a single threaded ring implementation for the reorder buffer free list.
Signed-off-by: Radu Nicolau <radu.nicolau@intel.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
--vdev="event_sw0,credit_quanta=64"
+Scheduler tuning arguments
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The scheduler minimum number of events that are processed can be increased to
+reduce per event overhead and increase internal burst sizes, which can
+improve throughput.
+
+* ``min_burst`` specifies the minimum number of inflight events that can be
+ moved to the next stage in the scheduler. Default value is 1.
+
+* ``refill_once`` is a switch that when set instructs the scheduler to deque
+ the events waiting in the ingress rings only once per call. The default
+ behavior is to dequeue as needed.
+
+* ``deq_burst`` is the burst size used to dequeue from the port rings.
+ Default value is 32, and it should be increased to 64 or 128 when setting
+ ``refill_once=1``.
+
+.. code-block:: console
+
+ --vdev="event_sw0,min_burst=8,deq_burst=64,refill_once=1"
+
Limitations
-----------
See the :doc:`../regexdevs/octeontx2` for more details.
+* **Updated Software Eventdev driver.**
+
+ Added performance tuning arguments to allow tuning the scheduler for
+ better throughtput in high core count use cases.
+
* **Updated ioat rawdev driver**
The ioat rawdev driver has been updated and enhanced. Changes include:
#include <rte_memory.h>
#include <rte_malloc.h>
-#define QE_RING_NAMESIZE 32
-
-struct qe_ring {
- char name[QE_RING_NAMESIZE] __rte_cache_aligned;
- uint32_t ring_size; /* size of memory block allocated to the ring */
- uint32_t mask; /* mask for read/write values == ring_size -1 */
- uint32_t size; /* actual usable space in the ring */
- volatile uint32_t write_idx __rte_cache_aligned;
- volatile uint32_t read_idx __rte_cache_aligned;
-
- struct rte_event ring[0] __rte_cache_aligned;
+/* Custom single threaded ring implementation used for ROB */
+struct rob_ring {
+ uint32_t ring_size;
+ uint32_t mask;
+ uint32_t size;
+ uint32_t write_idx;
+ uint32_t read_idx;
+ void *ring[0] __rte_cache_aligned;
};
-static inline struct qe_ring *
-qe_ring_create(const char *name, unsigned int size, unsigned int socket_id)
+static inline struct rob_ring *
+rob_ring_create(unsigned int size, unsigned int socket_id)
{
- struct qe_ring *retval;
+ struct rob_ring *retval;
const uint32_t ring_size = rte_align32pow2(size + 1);
size_t memsize = sizeof(*retval) +
(ring_size * sizeof(retval->ring[0]));
retval = rte_zmalloc_socket(NULL, memsize, 0, socket_id);
if (retval == NULL)
goto end;
-
- snprintf(retval->name, sizeof(retval->name), "EVDEV_RG_%s", name);
retval->ring_size = ring_size;
retval->mask = ring_size - 1;
retval->size = size;
}
static inline void
-qe_ring_destroy(struct qe_ring *r)
+rob_ring_free(struct rob_ring *r)
{
rte_free(r);
}
static __rte_always_inline unsigned int
-qe_ring_count(const struct qe_ring *r)
+rob_ring_count(const struct rob_ring *r)
{
return r->write_idx - r->read_idx;
}
static __rte_always_inline unsigned int
-qe_ring_free_count(const struct qe_ring *r)
+rob_ring_free_count(const struct rob_ring *r)
{
- return r->size - qe_ring_count(r);
+ return r->size - rob_ring_count(r);
}
static __rte_always_inline unsigned int
-qe_ring_enqueue_burst(struct qe_ring *r, const struct rte_event *qes,
- unsigned int nb_qes, uint16_t *free_count)
+rob_ring_enqueue(struct rob_ring *r, void *re)
{
const uint32_t size = r->size;
const uint32_t mask = r->mask;
const uint32_t read = r->read_idx;
uint32_t write = r->write_idx;
const uint32_t space = read + size - write;
- uint32_t i;
-
- if (space < nb_qes)
- nb_qes = space;
-
- for (i = 0; i < nb_qes; i++, write++)
- r->ring[write & mask] = qes[i];
-
- rte_smp_wmb();
-
- if (nb_qes != 0)
- r->write_idx = write;
-
- *free_count = space - nb_qes;
-
- return nb_qes;
+ if (space < 1)
+ return 0;
+ r->ring[write & mask] = re;
+ r->write_idx++;
+ return 1;
}
static __rte_always_inline unsigned int
-qe_ring_enqueue_burst_with_ops(struct qe_ring *r, const struct rte_event *qes,
- unsigned int nb_qes, uint8_t *ops)
-{
- const uint32_t size = r->size;
- const uint32_t mask = r->mask;
- const uint32_t read = r->read_idx;
- uint32_t write = r->write_idx;
- const uint32_t space = read + size - write;
- uint32_t i;
-
- if (space < nb_qes)
- nb_qes = space;
-
- for (i = 0; i < nb_qes; i++, write++) {
- r->ring[write & mask] = qes[i];
- r->ring[write & mask].op = ops[i];
- }
-
- rte_smp_wmb();
-
- if (nb_qes != 0)
- r->write_idx = write;
-
- return nb_qes;
-}
-
-static __rte_always_inline unsigned int
-qe_ring_dequeue_burst(struct qe_ring *r, struct rte_event *qes,
- unsigned int nb_qes)
+rob_ring_dequeue(struct rob_ring *r, void **re)
{
const uint32_t mask = r->mask;
uint32_t read = r->read_idx;
const uint32_t write = r->write_idx;
const uint32_t items = write - read;
- uint32_t i;
-
- if (items < nb_qes)
- nb_qes = items;
-
-
- for (i = 0; i < nb_qes; i++, read++)
- qes[i] = r->ring[read & mask];
-
- rte_smp_rmb();
-
- if (nb_qes != 0)
- r->read_idx += nb_qes;
-
- return nb_qes;
+ if (items < 1)
+ return 0;
+ *re = r->ring[read & mask];
+ r->read_idx++;
+ return 1;
}
#endif
#include "sw_evdev.h"
#include "iq_chunk.h"
+#include "event_ring.h"
#define EVENTDEV_NAME_SW_PMD event_sw
#define NUMA_NODE_ARG "numa_node"
#define SCHED_QUANTA_ARG "sched_quanta"
#define CREDIT_QUANTA_ARG "credit_quanta"
+#define MIN_BURST_SIZE_ARG "min_burst"
+#define DEQ_BURST_SIZE_ARG "deq_burst"
+#define REFIL_ONCE_ARG "refill_once"
static void
sw_info_get(struct rte_eventdev *dev, struct rte_event_dev_info *info);
qid->priority = queue_conf->priority;
if (qid->type == RTE_SCHED_TYPE_ORDERED) {
- char ring_name[RTE_RING_NAMESIZE];
uint32_t window_size;
/* rte_ring and window_size_mask require require window_size to
0,
window_size * sizeof(qid->reorder_buffer[0]));
- snprintf(ring_name, sizeof(ring_name), "sw%d_q%d_freelist",
- dev_id, idx);
-
- /* lookup the ring, and if it already exists, free it */
- struct rte_ring *cleanup = rte_ring_lookup(ring_name);
- if (cleanup)
- rte_ring_free(cleanup);
-
- qid->reorder_buffer_freelist = rte_ring_create(ring_name,
- window_size,
- socket_id,
- RING_F_SP_ENQ | RING_F_SC_DEQ);
+ qid->reorder_buffer_freelist = rob_ring_create(window_size,
+ socket_id);
if (!qid->reorder_buffer_freelist) {
SW_LOG_DBG("freelist ring create failed");
goto cleanup;
* that many.
*/
for (i = 0; i < window_size - 1; i++) {
- if (rte_ring_sp_enqueue(qid->reorder_buffer_freelist,
- &qid->reorder_buffer[i]) < 0)
+ if (rob_ring_enqueue(qid->reorder_buffer_freelist,
+ &qid->reorder_buffer[i]) != 1)
goto cleanup;
}
}
if (qid->reorder_buffer_freelist) {
- rte_ring_free(qid->reorder_buffer_freelist);
+ rob_ring_free(qid->reorder_buffer_freelist);
qid->reorder_buffer_freelist = NULL;
}
if (qid->type == RTE_SCHED_TYPE_ORDERED) {
rte_free(qid->reorder_buffer);
- rte_ring_free(qid->reorder_buffer_freelist);
+ rob_ring_free(qid->reorder_buffer_freelist);
}
memset(qid, 0, sizeof(*qid));
}
qid->stats.rx_pkts, qid->stats.rx_dropped,
qid->stats.tx_pkts);
if (qid->type == RTE_SCHED_TYPE_ORDERED) {
- struct rte_ring *rob_buf_free =
+ struct rob_ring *rob_buf_free =
qid->reorder_buffer_freelist;
if (rob_buf_free)
fprintf(f, "\tReorder entries in use: %u\n",
- rte_ring_free_count(rob_buf_free));
+ rob_ring_free_count(rob_buf_free));
else
fprintf(f,
"\tReorder buffer not initialized\n");
return 0;
}
+static int
+set_deq_burst_sz(const char *key __rte_unused, const char *value, void *opaque)
+{
+ int *deq_burst_sz = opaque;
+ *deq_burst_sz = atoi(value);
+ if (*deq_burst_sz < 0 || *deq_burst_sz > SCHED_DEQUEUE_MAX_BURST_SIZE)
+ return -1;
+ return 0;
+}
+
+static int
+set_min_burst_sz(const char *key __rte_unused, const char *value, void *opaque)
+{
+ int *min_burst_sz = opaque;
+ *min_burst_sz = atoi(value);
+ if (*min_burst_sz < 0 || *min_burst_sz > SCHED_DEQUEUE_MAX_BURST_SIZE)
+ return -1;
+ return 0;
+}
+
+static int
+set_refill_once(const char *key __rte_unused, const char *value, void *opaque)
+{
+ int *refill_once_per_call = opaque;
+ *refill_once_per_call = atoi(value);
+ if (*refill_once_per_call < 0 || *refill_once_per_call > 1)
+ return -1;
+ return 0;
+}
static int32_t sw_sched_service_func(void *args)
{
NUMA_NODE_ARG,
SCHED_QUANTA_ARG,
CREDIT_QUANTA_ARG,
+ MIN_BURST_SIZE_ARG,
+ DEQ_BURST_SIZE_ARG,
+ REFIL_ONCE_ARG,
NULL
};
const char *name;
int socket_id = rte_socket_id();
int sched_quanta = SW_DEFAULT_SCHED_QUANTA;
int credit_quanta = SW_DEFAULT_CREDIT_QUANTA;
+ int min_burst_size = 1;
+ int deq_burst_size = SCHED_DEQUEUE_DEFAULT_BURST_SIZE;
+ int refill_once = 0;
name = rte_vdev_device_name(vdev);
params = rte_vdev_device_args(vdev);
return ret;
}
+ ret = rte_kvargs_process(kvlist, MIN_BURST_SIZE_ARG,
+ set_min_burst_sz, &min_burst_size);
+ if (ret != 0) {
+ SW_LOG_ERR(
+ "%s: Error parsing minimum burst size parameter",
+ name);
+ rte_kvargs_free(kvlist);
+ return ret;
+ }
+
+ ret = rte_kvargs_process(kvlist, DEQ_BURST_SIZE_ARG,
+ set_deq_burst_sz, &deq_burst_size);
+ if (ret != 0) {
+ SW_LOG_ERR(
+ "%s: Error parsing dequeue burst size parameter",
+ name);
+ rte_kvargs_free(kvlist);
+ return ret;
+ }
+
+ ret = rte_kvargs_process(kvlist, REFIL_ONCE_ARG,
+ set_refill_once, &refill_once);
+ if (ret != 0) {
+ SW_LOG_ERR(
+ "%s: Error parsing refill once per call switch",
+ name);
+ rte_kvargs_free(kvlist);
+ return ret;
+ }
+
rte_kvargs_free(kvlist);
}
}
SW_LOG_INFO(
- "Creating eventdev sw device %s, numa_node=%d, sched_quanta=%d, credit_quanta=%d\n",
- name, socket_id, sched_quanta, credit_quanta);
+ "Creating eventdev sw device %s, numa_node=%d, "
+ "sched_quanta=%d, credit_quanta=%d "
+ "min_burst=%d, deq_burst=%d, refill_once=%d\n",
+ name, socket_id, sched_quanta, credit_quanta,
+ min_burst_size, deq_burst_size, refill_once);
dev = rte_event_pmd_vdev_init(name,
sizeof(struct sw_evdev), socket_id);
/* copy values passed from vdev command line to instance */
sw->credit_update_quanta = credit_quanta;
sw->sched_quanta = sched_quanta;
+ sw->sched_min_burst_size = min_burst_size;
+ sw->sched_deq_burst_size = deq_burst_size;
+ sw->refill_once_per_iter = refill_once;
/* register service with EAL */
struct rte_service_spec service;
RTE_PMD_REGISTER_VDEV(EVENTDEV_NAME_SW_PMD, evdev_sw_pmd_drv);
RTE_PMD_REGISTER_PARAM_STRING(event_sw, NUMA_NODE_ARG "=<int> "
- SCHED_QUANTA_ARG "=<int>" CREDIT_QUANTA_ARG "=<int>");
+ SCHED_QUANTA_ARG "=<int>" CREDIT_QUANTA_ARG "=<int>"
+ MIN_BURST_SIZE_ARG "=<int>" DEQ_BURST_SIZE_ARG "=<int>"
+ REFIL_ONCE_ARG "=<int>");
RTE_LOG_REGISTER(eventdev_sw_log_level, pmd.event.sw, NOTICE);
/* report dequeue burst sizes in buckets */
#define SW_DEQ_STAT_BUCKET_SHIFT 2
/* how many packets pulled from port by sched */
-#define SCHED_DEQUEUE_BURST_SIZE 32
+#define SCHED_DEQUEUE_DEFAULT_BURST_SIZE 32
+/* max buffer size */
+#define SCHED_DEQUEUE_MAX_BURST_SIZE 256
+
+/* Flush the pipeline after this many no enq to cq */
+#define SCHED_NO_ENQ_CYCLE_FLUSH 256
+
#define SW_PORT_HIST_LIST (MAX_SW_PROD_Q_DEPTH) /* size of our history list */
#define NUM_SAMPLES 64 /* how many data points use for average stats */
/* Track packet order for reordering when needed */
struct reorder_buffer_entry *reorder_buffer; /*< pkts await reorder */
- struct rte_ring *reorder_buffer_freelist; /* available reorder slots */
+ struct rob_ring *reorder_buffer_freelist; /* available reorder slots */
uint32_t reorder_buffer_index; /* oldest valid reorder buffer entry */
uint32_t window_size; /* Used to wrap reorder_buffer_index */
uint32_t pp_buf_start;
uint32_t pp_buf_count;
uint16_t cq_buf_count;
- struct rte_event pp_buf[SCHED_DEQUEUE_BURST_SIZE];
+ struct rte_event pp_buf[SCHED_DEQUEUE_MAX_BURST_SIZE];
struct rte_event cq_buf[MAX_SW_CONS_Q_DEPTH];
uint8_t num_qids_mapped;
uint32_t xstats_count_mode_port;
uint32_t xstats_count_mode_queue;
+ /* Minimum burst size*/
+ uint32_t sched_min_burst_size __rte_cache_aligned;
+ /* Port dequeue burst size*/
+ uint32_t sched_deq_burst_size;
+ /* Refill pp buffers only once per scheduler call*/
+ uint32_t refill_once_per_iter;
+ /* Current values */
+ uint32_t sched_flush_count;
+ uint32_t sched_min_burst;
+
/* Contains all ports - load balanced and directed */
struct sw_port ports[SW_PORTS_MAX] __rte_cache_aligned;
#include <rte_event_ring.h>
#include "sw_evdev.h"
#include "iq_chunk.h"
+#include "event_ring.h"
#define SW_IQS_MASK (SW_IQS_MAX-1)
/* use cheap bit mixing, we only need to lose a few bits */
#define SW_HASH_FLOWID(f) (((f) ^ (f >> 10)) & FLOWID_MASK)
+
static inline uint32_t
sw_schedule_atomic_to_cq(struct sw_evdev *sw, struct sw_qid * const qid,
uint32_t iq_num, unsigned int count)
if (keep_order)
/* only schedule as many as we have reorder buffer entries */
count = RTE_MIN(count,
- rte_ring_count(qid->reorder_buffer_freelist));
+ rob_ring_count(qid->reorder_buffer_freelist));
for (i = 0; i < count; i++) {
const struct rte_event *qe = iq_peek(&qid->iq[iq_num]);
cq_idx = 0;
cq = qid->cq_map[cq_idx++];
- } while (rte_event_ring_free_count(
- sw->ports[cq].cq_worker_ring) == 0 ||
- sw->ports[cq].inflights == SW_PORT_HIST_LIST);
+ } while (sw->ports[cq].inflights == SW_PORT_HIST_LIST ||
+ rte_event_ring_free_count(
+ sw->ports[cq].cq_worker_ring) == 0);
struct sw_port *p = &sw->ports[cq];
if (sw->cq_ring_space[cq] == 0 ||
p->hist_list[head].qid = qid_id;
if (keep_order)
- rte_ring_sc_dequeue(qid->reorder_buffer_freelist,
+ rob_ring_dequeue(qid->reorder_buffer_freelist,
(void *)&p->hist_list[head].rob_entry);
sw->ports[cq].cq_buf[sw->ports[cq].cq_buf_count++] = *qe;
uint32_t pkts_done = 0;
uint32_t count = iq_count(&qid->iq[iq_num]);
- if (count > 0) {
+ if (count >= sw->sched_min_burst) {
if (type == SW_SCHED_TYPE_DIRECT)
pkts_done += sw_schedule_dir_to_cq(sw, qid,
iq_num, count);
for (; qid_start < qid_end; qid_start++) {
struct sw_qid *qid = &sw->qids[qid_start];
- int i, num_entries_in_use;
+ unsigned int i, num_entries_in_use;
if (qid->type != RTE_SCHED_TYPE_ORDERED)
continue;
- num_entries_in_use = rte_ring_free_count(
+ num_entries_in_use = rob_ring_free_count(
qid->reorder_buffer_freelist);
+ if (num_entries_in_use < sw->sched_min_burst)
+ num_entries_in_use = 0;
+
for (i = 0; i < num_entries_in_use; i++) {
struct reorder_buffer_entry *entry;
int j;
if (!entry->ready) {
entry->fragment_index = 0;
- rte_ring_sp_enqueue(
+ rob_ring_enqueue(
qid->reorder_buffer_freelist,
entry);
struct rte_event_ring *worker = port->rx_worker_ring;
port->pp_buf_start = 0;
port->pp_buf_count = rte_event_ring_dequeue_burst(worker, port->pp_buf,
- RTE_DIM(port->pp_buf), NULL);
+ sw->sched_deq_burst_size, NULL);
}
static __rte_always_inline uint32_t
struct sw_port *port = &sw->ports[port_id];
/* If shadow ring has 0 pkts, pull from worker ring */
- if (port->pp_buf_count == 0)
+ if (!sw->refill_once_per_iter && port->pp_buf_count == 0)
sw_refill_pp_buf(sw, port);
while (port->pp_buf_count) {
struct sw_port *port = &sw->ports[port_id];
/* If shadow ring has 0 pkts, pull from worker ring */
- if (port->pp_buf_count == 0)
+ if (!sw->refill_once_per_iter && port->pp_buf_count == 0)
sw_refill_pp_buf(sw, port);
while (port->pp_buf_count) {
/* push all the internal buffered QEs in port->cq_ring to the
* worker cores: aka, do the ring transfers batched.
*/
+ int no_enq = 1;
for (i = 0; i < sw->port_count; i++) {
- struct rte_event_ring *worker = sw->ports[i].cq_worker_ring;
- rte_event_ring_enqueue_burst(worker, sw->ports[i].cq_buf,
- sw->ports[i].cq_buf_count,
- &sw->cq_ring_space[i]);
- sw->ports[i].cq_buf_count = 0;
+ struct sw_port *port = &sw->ports[i];
+ struct rte_event_ring *worker = port->cq_worker_ring;
+
+ /* If shadow ring has 0 pkts, pull from worker ring */
+ if (sw->refill_once_per_iter && port->pp_buf_count == 0)
+ sw_refill_pp_buf(sw, port);
+
+ if (port->cq_buf_count >= sw->sched_min_burst) {
+ rte_event_ring_enqueue_burst(worker,
+ port->cq_buf,
+ port->cq_buf_count,
+ &sw->cq_ring_space[i]);
+ port->cq_buf_count = 0;
+ no_enq = 0;
+ } else {
+ sw->cq_ring_space[i] =
+ rte_event_ring_free_count(worker) -
+ port->cq_buf_count;
+ }
+ }
+
+ if (no_enq) {
+ if (unlikely(sw->sched_flush_count > SCHED_NO_ENQ_CYCLE_FLUSH))
+ sw->sched_min_burst = 1;
+ else
+ sw->sched_flush_count++;
+ } else {
+ if (sw->sched_flush_count)
+ sw->sched_flush_count--;
+ else
+ sw->sched_min_burst = sw->sched_min_burst_size;
}
}
git.droids-corp.org / dpdk.git / commitdiff