DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [PATCH] common/cpt: rework pending queue
@ 2021-09-24 11:20 Anoob Joseph
  2021-09-24 17:45 ` Akhil Goyal
  0 siblings, 1 reply; 2+ messages in thread
From: Anoob Joseph @ 2021-09-24 11:20 UTC (permalink / raw)
  To: Akhil Goyal
  Cc: David George, Jerin Jacob, Ankur Dwivedi, Tejasree Kondoj, dev,
	Anoob Joseph

From: David George <david.george@sophos.com>

Replace pending queue with one that allows concurrent single producer and
single consumer. This relaxes the restriction of only allowing a single
lcore to operate on a given queue pair.

Signed-off-by: David George <david.george@sophos.com>
Signed-off-by: Anoob Joseph <anoobj@marvell.com>
---
 doc/guides/cryptodevs/octeontx.rst                 |   6 --
 doc/guides/cryptodevs/octeontx2.rst                |   6 --
 drivers/common/cpt/cpt_common.h                    |  73 +++++++++++++--
 drivers/crypto/octeontx/otx_cryptodev_hw_access.c  |  19 ++--
 drivers/crypto/octeontx/otx_cryptodev_hw_access.h  |  14 ++-
 drivers/crypto/octeontx/otx_cryptodev_ops.c        |  75 +++++++---------
 .../crypto/octeontx2/otx2_cryptodev_hw_access.h    |   8 +-
 drivers/crypto/octeontx2/otx2_cryptodev_ops.c      | 100 ++++++++++++---------
 8 files changed, 179 insertions(+), 122 deletions(-)

diff --git a/doc/guides/cryptodevs/octeontx.rst b/doc/guides/cryptodevs/octeontx.rst
index 4972274..a39f3f3 100644
--- a/doc/guides/cryptodevs/octeontx.rst
+++ b/doc/guides/cryptodevs/octeontx.rst
@@ -135,9 +135,3 @@ application:
 
         ./dpdk-test
         RTE>>cryptodev_octeontx_asym_autotest
-
-Limitations
------------
-
-Multiple lcores may not operate on the same crypto queue pair. The lcore that
-enqueues to a queue pair is the one that must dequeue from it.
diff --git a/doc/guides/cryptodevs/octeontx2.rst b/doc/guides/cryptodevs/octeontx2.rst
index f0beb92..811e61a 100644
--- a/doc/guides/cryptodevs/octeontx2.rst
+++ b/doc/guides/cryptodevs/octeontx2.rst
@@ -186,9 +186,3 @@ Features supported
 * AES-128/192/256-GCM
 * AES-128/192/256-CBC-SHA1-HMAC
 * AES-128/192/256-CBC-SHA256-128-HMAC
-
-Limitations
------------
-
-Multiple lcores may not operate on the same crypto queue pair. The lcore that
-enqueues to a queue pair is the one that must dequeue from it.
diff --git a/drivers/common/cpt/cpt_common.h b/drivers/common/cpt/cpt_common.h
index 724e5ec..d70668a 100644
--- a/drivers/common/cpt/cpt_common.h
+++ b/drivers/common/cpt/cpt_common.h
@@ -5,6 +5,7 @@
 #ifndef _CPT_COMMON_H_
 #define _CPT_COMMON_H_
 
+#include <rte_prefetch.h>
 #include <rte_mempool.h>
 
 /*
@@ -32,14 +33,12 @@ struct cpt_qp_meta_info {
  *
  */
 struct pending_queue {
-	/** Pending requests count */
-	uint64_t pending_count;
 	/** Array of pending requests */
-	uintptr_t *req_queue;
+	void **rid_queue;
 	/** Tail of queue to be used for enqueue */
-	uint16_t enq_tail;
+	unsigned int tail;
 	/** Head of queue to be used for dequeue */
-	uint16_t deq_head;
+	unsigned int head;
 };
 
 struct cpt_request_info {
@@ -61,4 +60,68 @@ struct cpt_request_info {
 	uint8_t extra_time;
 } __rte_aligned(8);
 
+static __rte_always_inline void
+pending_queue_push(struct pending_queue *q, void *rid, unsigned int off,
+			const int qsize)
+{
+	/* NOTE: no free space check, but it is expected that one is made */
+	q->rid_queue[(q->tail + off) & (qsize - 1)] = rid;
+}
+
+static __rte_always_inline void
+pending_queue_commit(struct pending_queue *q, unsigned int cnt,
+			const unsigned int qsize)
+{
+	/* Ensure ordering between setting the entry and updating the tail */
+	rte_atomic_thread_fence(__ATOMIC_RELEASE);
+
+	q->tail = (q->tail + cnt) & (qsize - 1);
+}
+
+static __rte_always_inline void
+pending_queue_pop(struct pending_queue *q, const int qsize)
+{
+	/* NOTE: no empty check, but it is expected that one is made prior */
+
+	q->head = (q->head + 1) & (qsize - 1);
+}
+
+static __rte_always_inline void
+pending_queue_peek(struct pending_queue *q, void **rid, const int qsize,
+			int prefetch_next)
+{
+	void *next_rid;
+	/* NOTE: no empty check, but it is expected that one is made */
+
+	*rid = q->rid_queue[q->head];
+
+	if (likely(prefetch_next)) {
+		next_rid = q->rid_queue[(q->head + 1) & (qsize - 1)];
+		rte_prefetch_non_temporal((void *)next_rid);
+	}
+}
+
+static __rte_always_inline unsigned int
+pending_queue_level(struct pending_queue *q, const int qsize)
+{
+	return (q->tail - q->head) & (qsize - 1);
+}
+
+static __rte_always_inline unsigned int
+pending_queue_free_slots(struct pending_queue *q, const int qsize,
+		const int reserved_slots)
+{
+	int free_slots;
+
+	free_slots = qsize - pending_queue_level(q, qsize);
+
+	/* Use only use qsize - 1 */
+	free_slots -= 1 + reserved_slots;
+
+	if (unlikely(free_slots < 0))
+		return 0;
+
+	return free_slots;
+}
+
 #endif /* _CPT_COMMON_H_ */
diff --git a/drivers/crypto/octeontx/otx_cryptodev_hw_access.c b/drivers/crypto/octeontx/otx_cryptodev_hw_access.c
index ab335c6..7b89a62 100644
--- a/drivers/crypto/octeontx/otx_cryptodev_hw_access.c
+++ b/drivers/crypto/octeontx/otx_cryptodev_hw_access.c
@@ -527,10 +527,10 @@ otx_cpt_get_resource(const struct rte_cryptodev *dev, uint8_t group,
 	memset(&cptvf->pqueue, 0, sizeof(cptvf->pqueue));
 
 	/* Chunks are of fixed size buffers */
+
+	qlen = DEFAULT_CMD_QLEN;
 	chunks = DEFAULT_CMD_QCHUNKS;
 	chunk_len = DEFAULT_CMD_QCHUNK_SIZE;
-
-	qlen = chunks * chunk_len;
 	/* Chunk size includes 8 bytes of next chunk ptr */
 	chunk_size = chunk_len * CPT_INST_SIZE + CPT_NEXT_CHUNK_PTR_SIZE;
 
@@ -538,7 +538,7 @@ otx_cpt_get_resource(const struct rte_cryptodev *dev, uint8_t group,
 	len = chunks * RTE_ALIGN(sizeof(struct command_chunk), 8);
 
 	/* For pending queue */
-	len += qlen * sizeof(uintptr_t);
+	len += qlen * RTE_ALIGN(sizeof(cptvf->pqueue.rid_queue[0]), 8);
 
 	/* So that instruction queues start as pg size aligned */
 	len = RTE_ALIGN(len, pg_sz);
@@ -573,14 +573,11 @@ otx_cpt_get_resource(const struct rte_cryptodev *dev, uint8_t group,
 	}
 
 	/* Pending queue setup */
-	cptvf->pqueue.req_queue = (uintptr_t *)mem;
-	cptvf->pqueue.enq_tail = 0;
-	cptvf->pqueue.deq_head = 0;
-	cptvf->pqueue.pending_count = 0;
-
-	mem +=  qlen * sizeof(uintptr_t);
-	len -=  qlen * sizeof(uintptr_t);
-	dma_addr += qlen * sizeof(uintptr_t);
+	cptvf->pqueue.rid_queue = (void **)mem;
+
+	mem +=  qlen * RTE_ALIGN(sizeof(cptvf->pqueue.rid_queue[0]), 8);
+	len -=  qlen * RTE_ALIGN(sizeof(cptvf->pqueue.rid_queue[0]), 8);
+	dma_addr += qlen * RTE_ALIGN(sizeof(cptvf->pqueue.rid_queue[0]), 8);
 
 	/* Alignment wastage */
 	used_len = alloc_len - len;
diff --git a/drivers/crypto/octeontx/otx_cryptodev_hw_access.h b/drivers/crypto/octeontx/otx_cryptodev_hw_access.h
index f7b1e93..7c6b1e4 100644
--- a/drivers/crypto/octeontx/otx_cryptodev_hw_access.h
+++ b/drivers/crypto/octeontx/otx_cryptodev_hw_access.h
@@ -23,10 +23,16 @@
 #define CPT_INTR_POLL_INTERVAL_MS	(50)
 
 /* Default command queue length */
-#define DEFAULT_CMD_QCHUNKS		2
-#define DEFAULT_CMD_QCHUNK_SIZE		1023
-#define DEFAULT_CMD_QLEN \
-		(DEFAULT_CMD_QCHUNK_SIZE * DEFAULT_CMD_QCHUNKS)
+#define DEFAULT_CMD_QLEN	2048
+#define DEFAULT_CMD_QCHUNKS	2
+
+/* Instruction memory benefits from being 1023, so introduce
+ * reserved entries so we can't overrun the instruction queue
+ */
+#define DEFAULT_CMD_QRSVD_SLOTS DEFAULT_CMD_QCHUNKS
+#define DEFAULT_CMD_QCHUNK_SIZE \
+		((DEFAULT_CMD_QLEN - DEFAULT_CMD_QRSVD_SLOTS) / \
+		DEFAULT_CMD_QCHUNKS)
 
 #define CPT_CSR_REG_BASE(cpt)		((cpt)->reg_base)
 
diff --git a/drivers/crypto/octeontx/otx_cryptodev_ops.c b/drivers/crypto/octeontx/otx_cryptodev_ops.c
index c51be63..76ea1f8 100644
--- a/drivers/crypto/octeontx/otx_cryptodev_ops.c
+++ b/drivers/crypto/octeontx/otx_cryptodev_ops.c
@@ -431,16 +431,10 @@ otx_cpt_asym_session_clear(struct rte_cryptodev *dev,
 
 static __rte_always_inline void * __rte_hot
 otx_cpt_request_enqueue(struct cpt_instance *instance,
-			struct pending_queue *pqueue,
 			void *req, uint64_t cpt_inst_w7)
 {
 	struct cpt_request_info *user_req = (struct cpt_request_info *)req;
 
-	if (unlikely(pqueue->pending_count >= DEFAULT_CMD_QLEN)) {
-		rte_errno = EAGAIN;
-		return NULL;
-	}
-
 	fill_cpt_inst(instance, req, cpt_inst_w7);
 
 	CPT_LOG_DP_DEBUG("req: %p op: %p ", req, user_req->op);
@@ -460,8 +454,7 @@ otx_cpt_request_enqueue(struct cpt_instance *instance,
 
 static __rte_always_inline void * __rte_hot
 otx_cpt_enq_single_asym(struct cpt_instance *instance,
-			struct rte_crypto_op *op,
-			struct pending_queue *pqueue)
+			struct rte_crypto_op *op)
 {
 	struct cpt_qp_meta_info *minfo = &instance->meta_info;
 	struct rte_crypto_asym_op *asym_op = op->asym;
@@ -525,8 +518,7 @@ otx_cpt_enq_single_asym(struct cpt_instance *instance,
 		goto req_fail;
 	}
 
-	req = otx_cpt_request_enqueue(instance, pqueue, params.req,
-				      sess->cpt_inst_w7);
+	req = otx_cpt_request_enqueue(instance, params.req, sess->cpt_inst_w7);
 	if (unlikely(req == NULL)) {
 		CPT_LOG_DP_ERR("Could not enqueue crypto req");
 		goto req_fail;
@@ -542,8 +534,7 @@ otx_cpt_enq_single_asym(struct cpt_instance *instance,
 
 static __rte_always_inline void * __rte_hot
 otx_cpt_enq_single_sym(struct cpt_instance *instance,
-		       struct rte_crypto_op *op,
-		       struct pending_queue *pqueue)
+		       struct rte_crypto_op *op)
 {
 	struct cpt_sess_misc *sess;
 	struct rte_crypto_sym_op *sym_op = op->sym;
@@ -573,8 +564,7 @@ otx_cpt_enq_single_sym(struct cpt_instance *instance,
 	}
 
 	/* Enqueue prepared instruction to h/w */
-	req = otx_cpt_request_enqueue(instance, pqueue, prep_req,
-				      sess->cpt_inst_w7);
+	req = otx_cpt_request_enqueue(instance, prep_req, sess->cpt_inst_w7);
 	if (unlikely(req == NULL))
 		/* Buffer allocated for request preparation need to be freed */
 		free_op_meta(mdata, instance->meta_info.pool);
@@ -584,8 +574,7 @@ otx_cpt_enq_single_sym(struct cpt_instance *instance,
 
 static __rte_always_inline void * __rte_hot
 otx_cpt_enq_single_sym_sessless(struct cpt_instance *instance,
-				struct rte_crypto_op *op,
-				struct pending_queue *pend_q)
+				struct rte_crypto_op *op)
 {
 	const int driver_id = otx_cryptodev_driver_id;
 	struct rte_crypto_sym_op *sym_op = op->sym;
@@ -607,8 +596,8 @@ otx_cpt_enq_single_sym_sessless(struct cpt_instance *instance,
 
 	sym_op->session = sess;
 
-	req = otx_cpt_enq_single_sym(instance, op, pend_q);
-
+	/* Enqueue op with the tmp session set */
+	req = otx_cpt_enq_single_sym(instance, op);
 	if (unlikely(req == NULL))
 		goto priv_put;
 
@@ -627,22 +616,20 @@ otx_cpt_enq_single_sym_sessless(struct cpt_instance *instance,
 static __rte_always_inline void *__rte_hot
 otx_cpt_enq_single(struct cpt_instance *inst,
 		   struct rte_crypto_op *op,
-		   struct pending_queue *pqueue,
 		   const uint8_t op_type)
 {
 	/* Check for the type */
 
 	if (op_type == OP_TYPE_SYM) {
 		if (op->sess_type == RTE_CRYPTO_OP_WITH_SESSION)
-			return otx_cpt_enq_single_sym(inst, op, pqueue);
+			return otx_cpt_enq_single_sym(inst, op);
 		else
-			return otx_cpt_enq_single_sym_sessless(inst, op,
-							       pqueue);
+			return otx_cpt_enq_single_sym_sessless(inst, op);
 	}
 
 	if (op_type == OP_TYPE_ASYM) {
 		if (op->sess_type == RTE_CRYPTO_OP_WITH_SESSION)
-			return otx_cpt_enq_single_asym(inst, op, pqueue);
+			return otx_cpt_enq_single_asym(inst, op);
 	}
 
 	/* Should not reach here */
@@ -655,30 +642,33 @@ otx_cpt_pkt_enqueue(void *qptr, struct rte_crypto_op **ops, uint16_t nb_ops,
 		    const uint8_t op_type)
 {
 	struct cpt_instance *instance = (struct cpt_instance *)qptr;
-	uint16_t count;
+	uint16_t count, free_slots;
 	void *req;
 	struct cpt_vf *cptvf = (struct cpt_vf *)instance;
 	struct pending_queue *pqueue = &cptvf->pqueue;
 
-	count = DEFAULT_CMD_QLEN - pqueue->pending_count;
-	if (nb_ops > count)
-		nb_ops = count;
+	free_slots = pending_queue_free_slots(pqueue, DEFAULT_CMD_QLEN,
+				DEFAULT_CMD_QRSVD_SLOTS);
+	if (nb_ops > free_slots)
+		nb_ops = free_slots;
 
 	count = 0;
 	while (likely(count < nb_ops)) {
 
 		/* Enqueue single op */
-		req = otx_cpt_enq_single(instance, ops[count], pqueue, op_type);
+		req = otx_cpt_enq_single(instance, ops[count], op_type);
 
 		if (unlikely(req == NULL))
 			break;
 
-		pqueue->req_queue[pqueue->enq_tail] = (uintptr_t)req;
-		MOD_INC(pqueue->enq_tail, DEFAULT_CMD_QLEN);
-		pqueue->pending_count += 1;
+		pending_queue_push(pqueue, req, count, DEFAULT_CMD_QLEN);
 		count++;
 	}
-	otx_cpt_ring_dbell(instance, count);
+
+	if (likely(count)) {
+		pending_queue_commit(pqueue, count, DEFAULT_CMD_QLEN);
+		otx_cpt_ring_dbell(instance, count);
+	}
 	return count;
 }
 
@@ -756,8 +746,7 @@ otx_crypto_adapter_enqueue(void *port, struct rte_crypto_op *op)
 
 	op_type = op->type == RTE_CRYPTO_OP_TYPE_SYMMETRIC ? OP_TYPE_SYM :
 							     OP_TYPE_ASYM;
-	req = otx_cpt_enq_single(instance, op,
-				 &((struct cpt_vf *)instance)->pqueue, op_type);
+	req = otx_cpt_enq_single(instance, op, op_type);
 	if (unlikely(req == NULL))
 		return 0;
 
@@ -971,17 +960,16 @@ otx_cpt_pkt_dequeue(void *qptr, struct rte_crypto_op **ops, uint16_t nb_ops,
 	int nb_completed;
 	struct pending_queue *pqueue = &cptvf->pqueue;
 
-	pcount = pqueue->pending_count;
+	pcount = pending_queue_level(pqueue, DEFAULT_CMD_QLEN);
+
+	/* Ensure pcount isn't read before data lands */
+	rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
+
 	count = (nb_ops > pcount) ? pcount : nb_ops;
 
 	for (i = 0; i < count; i++) {
-		user_req = (struct cpt_request_info *)
-				pqueue->req_queue[pqueue->deq_head];
-
-		if (likely((i+1) < count)) {
-			rte_prefetch_non_temporal(
-				(void *)pqueue->req_queue[i+1]);
-		}
+		pending_queue_peek(pqueue, (void **) &user_req,
+			DEFAULT_CMD_QLEN, i + 1 < count);
 
 		ret = check_nb_command_id(user_req, instance);
 
@@ -997,8 +985,7 @@ otx_cpt_pkt_dequeue(void *qptr, struct rte_crypto_op **ops, uint16_t nb_ops,
 		CPT_LOG_DP_DEBUG("Request %p Op %p completed with code %d",
 				 user_req, user_req->op, ret);
 
-		MOD_INC(pqueue->deq_head, DEFAULT_CMD_QLEN);
-		pqueue->pending_count -= 1;
+		pending_queue_pop(pqueue, DEFAULT_CMD_QLEN);
 	}
 
 	nb_completed = i;
diff --git a/drivers/crypto/octeontx2/otx2_cryptodev_hw_access.h b/drivers/crypto/octeontx2/otx2_cryptodev_hw_access.h
index f9981ea..90a338e 100644
--- a/drivers/crypto/octeontx2/otx2_cryptodev_hw_access.h
+++ b/drivers/crypto/octeontx2/otx2_cryptodev_hw_access.h
@@ -17,10 +17,10 @@
 #include "otx2_dev.h"
 #include "otx2_cryptodev_qp.h"
 
-/* CPT instruction queue length */
-#define OTX2_CPT_IQ_LEN			8200
-
-#define OTX2_CPT_DEFAULT_CMD_QLEN	OTX2_CPT_IQ_LEN
+/* CPT instruction queue length.
+ * Use queue size as power of 2 for aiding in pending queue calculations.
+ */
+#define OTX2_CPT_DEFAULT_CMD_QLEN	8192
 
 /* Mask which selects all engine groups */
 #define OTX2_CPT_ENG_GRPS_MASK		0xFF
diff --git a/drivers/crypto/octeontx2/otx2_cryptodev_ops.c b/drivers/crypto/octeontx2/otx2_cryptodev_ops.c
index 952d135..7134fa7 100644
--- a/drivers/crypto/octeontx2/otx2_cryptodev_ops.c
+++ b/drivers/crypto/octeontx2/otx2_cryptodev_ops.c
@@ -49,6 +49,7 @@ otx2_cpt_metabuf_mempool_create(const struct rte_cryptodev *dev,
 {
 	char mempool_name[RTE_MEMPOOL_NAMESIZE];
 	struct cpt_qp_meta_info *meta_info;
+	int lcore_cnt = rte_lcore_count();
 	int ret, max_mlen, mb_pool_sz;
 	struct rte_mempool *pool;
 	int asym_mlen = 0;
@@ -87,7 +88,13 @@ otx2_cpt_metabuf_mempool_create(const struct rte_cryptodev *dev,
 	snprintf(mempool_name, RTE_MEMPOOL_NAMESIZE, "otx2_cpt_mb_%u:%u",
 		 dev->data->dev_id, qp_id);
 
-	mb_pool_sz = RTE_MAX(nb_elements, (METABUF_POOL_CACHE_SIZE * rte_lcore_count()));
+	mb_pool_sz = nb_elements;
+
+	/* For poll mode, core that enqueues and core that dequeues can be
+	 * different. For event mode, all cores are allowed to use same crypto
+	 * queue pair.
+	 */
+	mb_pool_sz += (RTE_MAX(2, lcore_cnt) * METABUF_POOL_CACHE_SIZE);
 
 	pool = rte_mempool_create_empty(mempool_name, mb_pool_sz, max_mlen,
 					METABUF_POOL_CACHE_SIZE, 0,
@@ -187,7 +194,13 @@ otx2_cpt_qp_create(const struct rte_cryptodev *dev, uint16_t qp_id,
 		return NULL;
 	}
 
-	iq_len = OTX2_CPT_IQ_LEN;
+	/*
+	 * Pending queue updates make assumption that queue size is a power
+	 * of 2.
+	 */
+	RTE_BUILD_BUG_ON(!RTE_IS_POWER_OF_2(OTX2_CPT_DEFAULT_CMD_QLEN));
+
+	iq_len = OTX2_CPT_DEFAULT_CMD_QLEN;
 
 	/*
 	 * Queue size must be a multiple of 40 and effective queue size to
@@ -196,7 +209,7 @@ otx2_cpt_qp_create(const struct rte_cryptodev *dev, uint16_t qp_id,
 	size_div40 = (iq_len + 40 - 1) / 40 + 1;
 
 	/* For pending queue */
-	len = iq_len * sizeof(uintptr_t);
+	len = iq_len * RTE_ALIGN(sizeof(qp->pend_q.rid_queue[0]), 8);
 
 	/* Space for instruction group memory */
 	len += size_div40 * 16;
@@ -205,7 +218,7 @@ otx2_cpt_qp_create(const struct rte_cryptodev *dev, uint16_t qp_id,
 	len = RTE_ALIGN(len, pg_sz);
 
 	/* For instruction queues */
-	len += OTX2_CPT_IQ_LEN * sizeof(union cpt_inst_s);
+	len += OTX2_CPT_DEFAULT_CMD_QLEN * sizeof(union cpt_inst_s);
 
 	/* Wastage after instruction queues */
 	len = RTE_ALIGN(len, pg_sz);
@@ -233,12 +246,11 @@ otx2_cpt_qp_create(const struct rte_cryptodev *dev, uint16_t qp_id,
 	}
 
 	/* Initialize pending queue */
-	qp->pend_q.req_queue = (uintptr_t *)va;
-	qp->pend_q.enq_tail = 0;
-	qp->pend_q.deq_head = 0;
-	qp->pend_q.pending_count = 0;
+	qp->pend_q.rid_queue = (void **)va;
+	qp->pend_q.tail = 0;
+	qp->pend_q.head = 0;
 
-	used_len = iq_len * sizeof(uintptr_t);
+	used_len = iq_len * RTE_ALIGN(sizeof(qp->pend_q.rid_queue[0]), 8);
 	used_len += size_div40 * 16;
 	used_len = RTE_ALIGN(used_len, pg_sz);
 	iova += used_len;
@@ -514,7 +526,8 @@ otx2_cpt_enqueue_req(const struct otx2_cpt_qp *qp,
 		     struct pending_queue *pend_q,
 		     struct cpt_request_info *req,
 		     struct rte_crypto_op *op,
-		     uint64_t cpt_inst_w7)
+		     uint64_t cpt_inst_w7,
+		     unsigned int burst_index)
 {
 	void *lmtline = qp->lmtline;
 	union cpt_inst_s inst;
@@ -523,9 +536,6 @@ otx2_cpt_enqueue_req(const struct otx2_cpt_qp *qp,
 	if (qp->ca_enable)
 		return otx2_ca_enqueue_req(qp, req, lmtline, op, cpt_inst_w7);
 
-	if (unlikely(pend_q->pending_count >= OTX2_CPT_DEFAULT_CMD_QLEN))
-		return -EAGAIN;
-
 	inst.u[0] = 0;
 	inst.s9x.res_addr = req->comp_baddr;
 	inst.u[2] = 0;
@@ -553,11 +563,7 @@ otx2_cpt_enqueue_req(const struct otx2_cpt_qp *qp,
 		lmt_status = otx2_lmt_submit(qp->lf_nq_reg);
 	} while (lmt_status == 0);
 
-	pend_q->req_queue[pend_q->enq_tail] = (uintptr_t)req;
-
-	/* We will use soft queue length here to limit requests */
-	MOD_INC(pend_q->enq_tail, OTX2_CPT_DEFAULT_CMD_QLEN);
-	pend_q->pending_count += 1;
+	pending_queue_push(pend_q, req, burst_index, OTX2_CPT_DEFAULT_CMD_QLEN);
 
 	return 0;
 }
@@ -565,7 +571,8 @@ otx2_cpt_enqueue_req(const struct otx2_cpt_qp *qp,
 static __rte_always_inline int32_t __rte_hot
 otx2_cpt_enqueue_asym(struct otx2_cpt_qp *qp,
 		      struct rte_crypto_op *op,
-		      struct pending_queue *pend_q)
+		      struct pending_queue *pend_q,
+		      unsigned int burst_index)
 {
 	struct cpt_qp_meta_info *minfo = &qp->meta_info;
 	struct rte_crypto_asym_op *asym_op = op->asym;
@@ -626,8 +633,7 @@ otx2_cpt_enqueue_asym(struct otx2_cpt_qp *qp,
 	}
 
 	ret = otx2_cpt_enqueue_req(qp, pend_q, params.req, op,
-				   sess->cpt_inst_w7);
-
+				   sess->cpt_inst_w7, burst_index);
 	if (unlikely(ret)) {
 		CPT_LOG_DP_ERR("Could not enqueue crypto req");
 		goto req_fail;
@@ -643,7 +649,7 @@ otx2_cpt_enqueue_asym(struct otx2_cpt_qp *qp,
 
 static __rte_always_inline int __rte_hot
 otx2_cpt_enqueue_sym(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
-		     struct pending_queue *pend_q)
+		     struct pending_queue *pend_q, unsigned int burst_index)
 {
 	struct rte_crypto_sym_op *sym_op = op->sym;
 	struct cpt_request_info *req;
@@ -670,8 +676,8 @@ otx2_cpt_enqueue_sym(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
 		return ret;
 	}
 
-	ret = otx2_cpt_enqueue_req(qp, pend_q, req, op, sess->cpt_inst_w7);
-
+	ret = otx2_cpt_enqueue_req(qp, pend_q, req, op, sess->cpt_inst_w7,
+				    burst_index);
 	if (unlikely(ret)) {
 		/* Free buffer allocated by fill params routines */
 		free_op_meta(mdata, qp->meta_info.pool);
@@ -682,7 +688,8 @@ otx2_cpt_enqueue_sym(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
 
 static __rte_always_inline int __rte_hot
 otx2_cpt_enqueue_sec(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
-		     struct pending_queue *pend_q)
+		     struct pending_queue *pend_q,
+		     const unsigned int burst_index)
 {
 	uint32_t winsz, esn_low = 0, esn_hi = 0, seql = 0, seqh = 0;
 	struct rte_mbuf *m_src = op->sym->m_src;
@@ -739,7 +746,8 @@ otx2_cpt_enqueue_sec(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
 		return ret;
 	}
 
-	ret = otx2_cpt_enqueue_req(qp, pend_q, req, op, sess->cpt_inst_w7);
+	ret = otx2_cpt_enqueue_req(qp, pend_q, req, op, sess->cpt_inst_w7,
+				    burst_index);
 
 	if (winsz && esn) {
 		seq_in_sa = ((uint64_t)esn_hi << 32) | esn_low;
@@ -754,7 +762,8 @@ otx2_cpt_enqueue_sec(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
 
 static __rte_always_inline int __rte_hot
 otx2_cpt_enqueue_sym_sessless(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
-			      struct pending_queue *pend_q)
+			      struct pending_queue *pend_q,
+			      unsigned int burst_index)
 {
 	const int driver_id = otx2_cryptodev_driver_id;
 	struct rte_crypto_sym_op *sym_op = op->sym;
@@ -773,7 +782,7 @@ otx2_cpt_enqueue_sym_sessless(struct otx2_cpt_qp *qp, struct rte_crypto_op *op,
 
 	sym_op->session = sess;
 
-	ret = otx2_cpt_enqueue_sym(qp, op, pend_q);
+	ret = otx2_cpt_enqueue_sym(qp, op, pend_q, burst_index);
 
 	if (unlikely(ret))
 		goto priv_put;
@@ -798,23 +807,26 @@ otx2_cpt_enqueue_burst(void *qptr, struct rte_crypto_op **ops, uint16_t nb_ops)
 
 	pend_q = &qp->pend_q;
 
-	nb_allowed = OTX2_CPT_DEFAULT_CMD_QLEN - pend_q->pending_count;
-	if (nb_ops > nb_allowed)
-		nb_ops = nb_allowed;
+	nb_allowed = pending_queue_free_slots(pend_q,
+				OTX2_CPT_DEFAULT_CMD_QLEN, 0);
+	nb_ops = RTE_MIN(nb_ops, nb_allowed);
 
 	for (count = 0; count < nb_ops; count++) {
 		op = ops[count];
 		if (op->type == RTE_CRYPTO_OP_TYPE_SYMMETRIC) {
 			if (op->sess_type == RTE_CRYPTO_OP_SECURITY_SESSION)
-				ret = otx2_cpt_enqueue_sec(qp, op, pend_q);
+				ret = otx2_cpt_enqueue_sec(qp, op, pend_q,
+							   count);
 			else if (op->sess_type == RTE_CRYPTO_OP_WITH_SESSION)
-				ret = otx2_cpt_enqueue_sym(qp, op, pend_q);
+				ret = otx2_cpt_enqueue_sym(qp, op, pend_q,
+							   count);
 			else
 				ret = otx2_cpt_enqueue_sym_sessless(qp, op,
-								    pend_q);
+						pend_q, count);
 		} else if (op->type == RTE_CRYPTO_OP_TYPE_ASYMMETRIC) {
 			if (op->sess_type == RTE_CRYPTO_OP_WITH_SESSION)
-				ret = otx2_cpt_enqueue_asym(qp, op, pend_q);
+				ret = otx2_cpt_enqueue_asym(qp, op, pend_q,
+								count);
 			else
 				break;
 		} else
@@ -824,6 +836,9 @@ otx2_cpt_enqueue_burst(void *qptr, struct rte_crypto_op **ops, uint16_t nb_ops)
 			break;
 	}
 
+	if (unlikely(!qp->ca_enable))
+		pending_queue_commit(pend_q, count, OTX2_CPT_DEFAULT_CMD_QLEN);
+
 	return count;
 }
 
@@ -1059,14 +1074,16 @@ otx2_cpt_dequeue_burst(void *qptr, struct rte_crypto_op **ops, uint16_t nb_ops)
 
 	pend_q = &qp->pend_q;
 
-	nb_pending = pend_q->pending_count;
+	nb_pending = pending_queue_level(pend_q, OTX2_CPT_DEFAULT_CMD_QLEN);
+
+	/* Ensure pcount isn't read before data lands */
+	rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
 
-	if (nb_ops > nb_pending)
-		nb_ops = nb_pending;
+	nb_ops = RTE_MIN(nb_ops, nb_pending);
 
 	for (i = 0; i < nb_ops; i++) {
-		req = (struct cpt_request_info *)
-				pend_q->req_queue[pend_q->deq_head];
+		pending_queue_peek(pend_q, (void **)&req,
+			OTX2_CPT_DEFAULT_CMD_QLEN, 0);
 
 		cc[i] = otx2_cpt_compcode_get(req);
 
@@ -1075,8 +1092,7 @@ otx2_cpt_dequeue_burst(void *qptr, struct rte_crypto_op **ops, uint16_t nb_ops)
 
 		ops[i] = req->op;
 
-		MOD_INC(pend_q->deq_head, OTX2_CPT_DEFAULT_CMD_QLEN);
-		pend_q->pending_count -= 1;
+		pending_queue_pop(pend_q, OTX2_CPT_DEFAULT_CMD_QLEN);
 	}
 
 	nb_completed = i;
-- 
2.7.4


^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [dpdk-dev] [PATCH] common/cpt: rework pending queue
  2021-09-24 11:20 [dpdk-dev] [PATCH] common/cpt: rework pending queue Anoob Joseph
@ 2021-09-24 17:45 ` Akhil Goyal
  0 siblings, 0 replies; 2+ messages in thread
From: Akhil Goyal @ 2021-09-24 17:45 UTC (permalink / raw)
  To: Anoob Joseph
  Cc: David George, Jerin Jacob Kollanukkaran, Ankur Dwivedi,
	Tejasree Kondoj, dev, Anoob Joseph

> From: David George <david.george@sophos.com>
> 
> Replace pending queue with one that allows concurrent single producer and
> single consumer. This relaxes the restriction of only allowing a single
> lcore to operate on a given queue pair.
> 
> Signed-off-by: David George <david.george@sophos.com>
> Signed-off-by: Anoob Joseph <anoobj@marvell.com>
> ---
>  doc/guides/cryptodevs/octeontx.rst                 |   6 --
>  doc/guides/cryptodevs/octeontx2.rst                |   6 --
>  drivers/common/cpt/cpt_common.h                    |  73 +++++++++++++--
>  drivers/crypto/octeontx/otx_cryptodev_hw_access.c  |  19 ++--
>  drivers/crypto/octeontx/otx_cryptodev_hw_access.h  |  14 ++-
>  drivers/crypto/octeontx/otx_cryptodev_ops.c        |  75 +++++++---------
>  .../crypto/octeontx2/otx2_cryptodev_hw_access.h    |   8 +-
>  drivers/crypto/octeontx2/otx2_cryptodev_ops.c      | 100 ++++++++++++------
> ---
Applied to dpdk-next-crypto

Thanks.

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2021-09-24 17:45 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-09-24 11:20 [dpdk-dev] [PATCH] common/cpt: rework pending queue Anoob Joseph
2021-09-24 17:45 ` Akhil Goyal

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).