event/octeontx: add multiproducer timer arm and cancel

Signed-off-by: Pavan Nikhilesh <pbhagavatula@caviumnetworks.com>
Acked-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>

drivers/event/octeontx/Makefile

@@ -30,20 +30,25 @@ SRCS-$(CONFIG_RTE_LIBRTE_PMD_OCTEONTX_SSOVF) += ssovf_worker.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_OCTEONTX_SSOVF) += ssovf_evdev.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_OCTEONTX_SSOVF) += ssovf_evdev_selftest.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_OCTEONTX_SSOVF) += ssovf_probe.c
+SRCS-$(CONFIG_RTE_LIBRTE_PMD_OCTEONTX_SSOVF) += timvf_worker.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_OCTEONTX_SSOVF) += timvf_evdev.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_OCTEONTX_SSOVF) += timvf_probe.c
 
 ifeq ($(CONFIG_RTE_TOOLCHAIN_GCC),y)
 CFLAGS_ssovf_worker.o += -fno-prefetch-loop-arrays
+CFLAGS_timvf_worker.o += -fno-prefetch-loop-arrays
 
 ifeq ($(shell test $(GCC_VERSION) -ge 46 && echo 1), 1)
 CFLAGS_ssovf_worker.o += -Ofast
+CFLAGS_timvf_worker.o += -Ofast
 else
 CFLAGS_ssovf_worker.o += -O3 -ffast-math
+CFLAGS_timvf_worker.o += -O3 -ffast-math
 endif
 
 else
 CFLAGS_ssovf_worker.o += -Ofast
+CFLAGS_timvf_worker.o += -Ofast
 endif
 
 include $(RTE_SDK)/mk/rte.lib.mk

drivers/event/octeontx/meson.build

@@ -5,6 +5,7 @@ sources = files('ssovf_worker.c',
 		'ssovf_evdev.c',
 		'ssovf_evdev_selftest.c',
 		'ssovf_probe.c',
+		'timvf_worker.c',
 		'timvf_evdev.c',
 		'timvf_probe.c'
 )

drivers/event/octeontx/timvf_evdev.c

@@ -130,6 +130,7 @@ timvf_ring_start(const struct rte_event_timer_adapter *adptr)
 
 	timvf_write64((uintptr_t)timr->bkt,
 			(uint8_t *)timr->vbar0 + TIM_VRING_BASE);
+	timvf_set_chunk_refill(timr);
 	if (timvf_ring_conf_set(&rctrl, timr->tim_ring_id)) {
 		ret = -EACCES;
 		goto error;
@@ -312,7 +313,13 @@ timvf_timer_adapter_caps_get(const struct rte_eventdev *dev, uint64_t flags,
 		timvf_ops.stats_reset = timvf_stats_reset;
 	}
 
+	if (enable_stats)
+		timvf_ops.arm_burst = timvf_timer_arm_burst_mp_stats;
+	else
+		timvf_ops.arm_burst = timvf_timer_arm_burst_mp;
+
+	timvf_ops.cancel_burst = timvf_timer_cancel_burst;
 	*caps = RTE_EVENT_TIMER_ADAPTER_CAP_INTERNAL_PORT;
 	*ops = &timvf_ops;
-	return -EINVAL;
+	return 0;
 }

drivers/event/octeontx/timvf_evdev.h

@@ -75,6 +75,33 @@
 #define TIM_VRING_AURA				(0x108)
 #define TIM_VRING_REL				(0x110)
 
+#define TIM_CTL1_W0_S_BUCKET			20
+#define TIM_CTL1_W0_M_BUCKET			((1ull << (40 - 20)) - 1)
+
+#define TIM_BUCKET_W1_S_NUM_ENTRIES		(0) /*Shift*/
+#define TIM_BUCKET_W1_M_NUM_ENTRIES		((1ull << (32 - 0)) - 1)
+#define TIM_BUCKET_W1_S_SBT			(32)
+#define TIM_BUCKET_W1_M_SBT			((1ull << (33 - 32)) - 1)
+#define TIM_BUCKET_W1_S_HBT			(33)
+#define TIM_BUCKET_W1_M_HBT			((1ull << (34 - 33)) - 1)
+#define TIM_BUCKET_W1_S_BSK			(34)
+#define TIM_BUCKET_W1_M_BSK			((1ull << (35 - 34)) - 1)
+#define TIM_BUCKET_W1_S_LOCK			(40)
+#define TIM_BUCKET_W1_M_LOCK			((1ull << (48 - 40)) - 1)
+#define TIM_BUCKET_W1_S_CHUNK_REMAINDER		(48)
+#define TIM_BUCKET_W1_M_CHUNK_REMAINDER		((1ull << (64 - 48)) - 1)
+
+#define TIM_BUCKET_SEMA	\
+	(TIM_BUCKET_CHUNK_REMAIN)
+
+#define TIM_BUCKET_CHUNK_REMAIN \
+	(TIM_BUCKET_W1_M_CHUNK_REMAINDER << TIM_BUCKET_W1_S_CHUNK_REMAINDER)
+
+#define TIM_BUCKET_LOCK \
+	(TIM_BUCKET_W1_M_LOCK << TIM_BUCKET_W1_S_LOCK)
+
+#define TIM_BUCKET_SEMA_WLOCK \
+	(TIM_BUCKET_CHUNK_REMAIN | (1ull << TIM_BUCKET_W1_S_LOCK))
 
 #define NSEC_PER_SEC 1E9
 #define NSEC2CLK(__ns, __freq) (((__ns) * (__freq)) / NSEC_PER_SEC)
@@ -168,5 +195,13 @@ void *timvf_bar(uint8_t id, uint8_t bar);
 int timvf_timer_adapter_caps_get(const struct rte_eventdev *dev, uint64_t flags,
 		uint32_t *caps, const struct rte_event_timer_adapter_ops **ops,
 		uint8_t enable_stats);
+uint16_t timvf_timer_cancel_burst(const struct rte_event_timer_adapter *adptr,
+		struct rte_event_timer **tim, const uint16_t nb_timers);
+uint16_t timvf_timer_arm_burst_mp(const struct rte_event_timer_adapter *adptr,
+		struct rte_event_timer **tim, const uint16_t nb_timers);
+uint16_t timvf_timer_arm_burst_mp_stats(
+		const struct rte_event_timer_adapter *adptr,
+		struct rte_event_timer **tim, const uint16_t nb_timers);
+void timvf_set_chunk_refill(struct timvf_ring * const timr);
 
 #endif /* __TIMVF_EVDEV_H__ */

drivers/event/octeontx/timvf_worker.c

@@ -0,0 +1,107 @@
+/*
+ * SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2017 Cavium, Inc
+ */
+
+#include "timvf_worker.h"
+
+static inline int
+timvf_timer_reg_checks(const struct timvf_ring * const timr,
+		struct rte_event_timer * const tim)
+{
+	if (unlikely(tim->state)) {
+		tim->state = RTE_EVENT_TIMER_ERROR;
+		rte_errno = EALREADY;
+		goto fail;
+	}
+
+	if (unlikely(!tim->timeout_ticks ||
+				tim->timeout_ticks >= timr->nb_bkts)) {
+		tim->state = tim->timeout_ticks ? RTE_EVENT_TIMER_ERROR_TOOLATE
+			: RTE_EVENT_TIMER_ERROR_TOOEARLY;
+		rte_errno = EINVAL;
+		goto fail;
+	}
+
+	return 0;
+fail:
+	return -EINVAL;
+}
+
+static inline void
+timvf_format_event(const struct rte_event_timer * const tim,
+		struct tim_mem_entry * const entry)
+{
+	entry->w0 = (tim->ev.event & 0xFFC000000000) >> 6 |
+		(tim->ev.event & 0xFFFFFFFFF);
+	entry->wqe = tim->ev.u64;
+}
+
+uint16_t
+timvf_timer_cancel_burst(const struct rte_event_timer_adapter *adptr,
+		struct rte_event_timer **tim, const uint16_t nb_timers)
+{
+	RTE_SET_USED(adptr);
+	int ret;
+	uint16_t index;
+
+	for (index = 0; index < nb_timers; index++) {
+		if (tim[index]->state == RTE_EVENT_TIMER_CANCELED) {
+			rte_errno = EALREADY;
+			break;
+		}
+
+		if (tim[index]->state != RTE_EVENT_TIMER_ARMED) {
+			rte_errno = EINVAL;
+			break;
+		}
+		ret = timvf_rem_entry(tim[index]);
+		if (ret) {
+			rte_errno = -ret;
+			break;
+		}
+	}
+	return index;
+}
+
+uint16_t
+timvf_timer_arm_burst_mp(const struct rte_event_timer_adapter *adptr,
+		struct rte_event_timer **tim, const uint16_t nb_timers)
+{
+	int ret;
+	uint16_t index;
+	struct tim_mem_entry entry;
+	struct timvf_ring *timr = adptr->data->adapter_priv;
+	for (index = 0; index < nb_timers; index++) {
+		if (timvf_timer_reg_checks(timr, tim[index]))
+			break;
+		timvf_format_event(tim[index], &entry);
+		ret = timvf_add_entry_mp(timr, tim[index]->timeout_ticks,
+				tim[index], &entry);
+		if (unlikely(ret)) {
+			rte_errno = -ret;
+			break;
+		}
+	}
+
+	return index;
+}
+
+uint16_t
+timvf_timer_arm_burst_mp_stats(const struct rte_event_timer_adapter *adptr,
+		struct rte_event_timer **tim, const uint16_t nb_timers)
+{
+	uint16_t ret;
+	struct timvf_ring *timr = adptr->data->adapter_priv;
+
+	ret = timvf_timer_arm_burst_mp(adptr, tim, nb_timers);
+	timr->tim_arm_cnt += ret;
+
+	return ret;
+}
+
+void
+timvf_set_chunk_refill(struct timvf_ring * const timr)
+{
+	timr->refill_chunk = timvf_refill_chunk_generic;
+}

drivers/event/octeontx/timvf_worker.h

@@ -0,0 +1,279 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2017 Cavium, Inc
+ */
+
+#include <rte_common.h>
+#include <rte_branch_prediction.h>
+
+#include "timvf_evdev.h"
+
+static inline int16_t
+timr_bkt_fetch_rem(uint64_t w1)
+{
+	return (w1 >> TIM_BUCKET_W1_S_CHUNK_REMAINDER) &
+		TIM_BUCKET_W1_M_CHUNK_REMAINDER;
+}
+
+static inline int16_t
+timr_bkt_get_rem(struct tim_mem_bucket *bktp)
+{
+	return __atomic_load_n(&bktp->chunk_remainder,
+			__ATOMIC_ACQUIRE);
+}
+
+static inline void
+timr_bkt_set_rem(struct tim_mem_bucket *bktp, uint16_t v)
+{
+	__atomic_store_n(&bktp->chunk_remainder, v,
+			__ATOMIC_RELEASE);
+}
+
+static inline void
+timr_bkt_sub_rem(struct tim_mem_bucket *bktp, uint16_t v)
+{
+	__atomic_fetch_sub(&bktp->chunk_remainder, v,
+			__ATOMIC_RELEASE);
+}
+
+static inline uint8_t
+timr_bkt_get_sbt(uint64_t w1)
+{
+	return (w1 >> TIM_BUCKET_W1_S_SBT) & TIM_BUCKET_W1_M_SBT;
+}
+
+static inline uint64_t
+timr_bkt_set_sbt(struct tim_mem_bucket *bktp)
+{
+	const uint64_t v = TIM_BUCKET_W1_M_SBT << TIM_BUCKET_W1_S_SBT;
+	return __atomic_fetch_or(&bktp->w1, v, __ATOMIC_ACQ_REL);
+}
+
+static inline uint64_t
+timr_bkt_clr_sbt(struct tim_mem_bucket *bktp)
+{
+	const uint64_t v = ~(TIM_BUCKET_W1_M_SBT << TIM_BUCKET_W1_S_SBT);
+	return __atomic_fetch_and(&bktp->w1, v, __ATOMIC_ACQ_REL);
+}
+
+static inline uint8_t
+timr_bkt_get_shbt(uint64_t w1)
+{
+	return ((w1 >> TIM_BUCKET_W1_S_HBT) & TIM_BUCKET_W1_M_HBT) |
+		((w1 >> TIM_BUCKET_W1_S_SBT) & TIM_BUCKET_W1_M_SBT);
+}
+
+static inline uint8_t
+timr_bkt_get_hbt(uint64_t w1)
+{
+	return (w1 >> TIM_BUCKET_W1_S_HBT) & TIM_BUCKET_W1_M_HBT;
+}
+
+static inline uint8_t
+timr_bkt_get_bsk(uint64_t w1)
+{
+	return (w1 >> TIM_BUCKET_W1_S_BSK) & TIM_BUCKET_W1_M_BSK;
+}
+
+static inline uint64_t
+timr_bkt_clr_bsk(struct tim_mem_bucket *bktp)
+{
+	/*Clear everything except lock. */
+	const uint64_t v = TIM_BUCKET_W1_M_LOCK << TIM_BUCKET_W1_S_LOCK;
+	return __atomic_fetch_and(&bktp->w1, v, __ATOMIC_ACQ_REL);
+}
+
+static inline uint64_t
+timr_bkt_fetch_sema_lock(struct tim_mem_bucket *bktp)
+{
+	return __atomic_fetch_add(&bktp->w1, TIM_BUCKET_SEMA_WLOCK,
+			__ATOMIC_ACQ_REL);
+}
+
+static inline uint64_t
+timr_bkt_fetch_sema(struct tim_mem_bucket *bktp)
+{
+	return __atomic_fetch_add(&bktp->w1, TIM_BUCKET_SEMA,
+			__ATOMIC_RELAXED);
+}
+
+static inline uint64_t
+timr_bkt_inc_lock(struct tim_mem_bucket *bktp)
+{
+	const uint64_t v = 1ull << TIM_BUCKET_W1_S_LOCK;
+	return __atomic_fetch_add(&bktp->w1, v, __ATOMIC_ACQ_REL);
+}
+
+static inline void
+timr_bkt_dec_lock(struct tim_mem_bucket *bktp)
+{
+	__atomic_add_fetch(&bktp->lock, 0xff, __ATOMIC_ACQ_REL);
+}
+
+static inline uint32_t
+timr_bkt_get_nent(uint64_t w1)
+{
+	return (w1 >> TIM_BUCKET_W1_S_NUM_ENTRIES) &
+		TIM_BUCKET_W1_M_NUM_ENTRIES;
+}
+
+static inline void
+timr_bkt_inc_nent(struct tim_mem_bucket *bktp)
+{
+	__atomic_add_fetch(&bktp->nb_entry, 1, __ATOMIC_RELAXED);
+}
+
+static inline void
+timr_bkt_add_nent(struct tim_mem_bucket *bktp, uint32_t v)
+{
+	__atomic_add_fetch(&bktp->nb_entry, v, __ATOMIC_RELAXED);
+}
+
+static inline uint64_t
+timr_bkt_clr_nent(struct tim_mem_bucket *bktp)
+{
+	const uint64_t v = ~(TIM_BUCKET_W1_M_NUM_ENTRIES <<
+			TIM_BUCKET_W1_S_NUM_ENTRIES);
+	return __atomic_and_fetch(&bktp->w1, v, __ATOMIC_ACQ_REL);
+}
+
+static inline struct tim_mem_entry *
+timr_clr_bkt(struct timvf_ring * const timr, struct tim_mem_bucket * const bkt)
+{
+	struct tim_mem_entry *chunk;
+	struct tim_mem_entry *pnext;
+	chunk = ((struct tim_mem_entry *)(uintptr_t)bkt->first_chunk);
+	chunk = (struct tim_mem_entry *)(uintptr_t)(chunk + nb_chunk_slots)->w0;
+
+	while (chunk) {
+		pnext = (struct tim_mem_entry *)(uintptr_t)
+			((chunk + nb_chunk_slots)->w0);
+		rte_mempool_put(timr->chunk_pool, chunk);
+		chunk = pnext;
+	}
+	return (struct tim_mem_entry *)(uintptr_t)bkt->first_chunk;
+}
+
+static inline int
+timvf_rem_entry(struct rte_event_timer *tim)
+{
+	uint64_t lock_sema;
+	struct tim_mem_entry *entry;
+	struct tim_mem_bucket *bkt;
+	if (tim->impl_opaque[1] == 0 ||
+			tim->impl_opaque[0] == 0)
+		return -ENOENT;
+
+	entry = (struct tim_mem_entry *)(uintptr_t)tim->impl_opaque[0];
+	if (entry->wqe != tim->ev.u64) {
+		tim->impl_opaque[1] = tim->impl_opaque[0] = 0;
+		return -ENOENT;
+	}
+	bkt = (struct tim_mem_bucket *)(uintptr_t)tim->impl_opaque[1];
+	lock_sema = timr_bkt_inc_lock(bkt);
+	if (timr_bkt_get_shbt(lock_sema)
+			|| !timr_bkt_get_nent(lock_sema)) {
+		timr_bkt_dec_lock(bkt);
+		tim->impl_opaque[1] = tim->impl_opaque[0] = 0;
+		return -ENOENT;
+	}
+
+	entry->w0 = entry->wqe = 0;
+	timr_bkt_dec_lock(bkt);
+
+	tim->state = RTE_EVENT_TIMER_CANCELED;
+	tim->impl_opaque[1] = tim->impl_opaque[0] = 0;
+	return 0;
+}
+
+static inline struct tim_mem_entry *
+timvf_refill_chunk_generic(struct tim_mem_bucket * const bkt,
+		struct timvf_ring * const timr)
+{
+	struct tim_mem_entry *chunk;
+
+	if (bkt->nb_entry || !bkt->first_chunk) {
+		if (unlikely(rte_mempool_get(timr->chunk_pool,
+						(void **)&chunk))) {
+			return NULL;
+		}
+		if (bkt->nb_entry) {
+			*(uint64_t *)(((struct tim_mem_entry *)(uintptr_t)
+					bkt->current_chunk) +
+					nb_chunk_slots) =
+				(uintptr_t) chunk;
+		} else {
+			bkt->first_chunk = (uintptr_t) chunk;
+		}
+	} else {
+		chunk = timr_clr_bkt(timr, bkt);
+		bkt->first_chunk = (uintptr_t)chunk;
+	}
+	*(uint64_t *)(chunk + nb_chunk_slots) = 0;
+
+	return chunk;
+}
+
+static inline struct tim_mem_bucket *
+timvf_get_target_bucket(struct timvf_ring * const timr, const uint32_t rel_bkt)
+{
+	const uint64_t bkt_cyc = rte_rdtsc() - timr->ring_start_cyc;
+	const uint32_t bucket = rte_reciprocal_divide_u64(bkt_cyc,
+			&timr->fast_div) + rel_bkt;
+	const uint32_t tbkt_id = timr->get_target_bkt(bucket,
+			timr->nb_bkts);
+	return &timr->bkt[tbkt_id];
+}
+
+/* Multi producer functions. */
+static inline int
+timvf_add_entry_mp(struct timvf_ring * const timr, const uint32_t rel_bkt,
+		struct rte_event_timer * const tim,
+		const struct tim_mem_entry * const pent)
+{
+	int16_t rem;
+	uint64_t lock_sema;
+	struct tim_mem_bucket *bkt;
+	struct tim_mem_entry *chunk;
+
+__retry:
+	bkt = timvf_get_target_bucket(timr, rel_bkt);
+	/* Bucket related checks. */
+	/*Get Bucket sema*/
+	lock_sema = timr_bkt_fetch_sema_lock(bkt);
+	if (unlikely(timr_bkt_get_shbt(lock_sema))) {
+		timr_bkt_dec_lock(bkt);
+		goto __retry;
+	}
+
+	rem = timr_bkt_fetch_rem(lock_sema);
+
+	if (rem < 0) {
+		/* goto diff bucket. */
+		timr_bkt_dec_lock(bkt);
+		goto __retry;
+	} else if (!rem) {
+		/*Only one thread can be here*/
+		chunk = timr->refill_chunk(bkt, timr);
+		if (unlikely(chunk == NULL)) {
+			timr_bkt_set_rem(bkt, 0);
+			timr_bkt_dec_lock(bkt);
+			tim->impl_opaque[0] = tim->impl_opaque[1] = 0;
+			tim->state = RTE_EVENT_TIMER_ERROR;
+			return -ENOMEM;
+		}
+		bkt->current_chunk = (uintptr_t) chunk;
+		timr_bkt_set_rem(bkt, nb_chunk_slots - 1);
+	} else {
+		chunk = (struct tim_mem_entry *)(uintptr_t)bkt->current_chunk;
+		chunk += nb_chunk_slots - rem;
+	}
+	/* Copy work entry. */
+	*chunk = *pent;
+	timr_bkt_inc_nent(bkt);
+	timr_bkt_dec_lock(bkt);
+
+	tim->impl_opaque[0] = (uintptr_t)chunk;
+	tim->impl_opaque[1] = (uintptr_t)bkt;
+	tim->state = RTE_EVENT_TIMER_ARMED;
+	return 0;
+}