1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2015-2022 Intel Corporation
5 #include <rte_common.h>
6 #include <rte_cycles.h>
8 #include <rte_malloc.h>
9 #include <rte_memzone.h>
11 #include <rte_bus_pci.h>
12 #include <rte_atomic.h>
13 #include <rte_prefetch.h>
16 #include "qat_device.h"
22 #define QAT_CQ_MAX_DEQ_RETRIES 10
24 #define ADF_MAX_DESC 4096
25 #define ADF_MIN_DESC 128
27 struct qat_qp_hw_spec_funcs*
28 qat_qp_hw_spec[QAT_N_GENS];
30 static int qat_qp_check_queue_alignment(uint64_t phys_addr,
31 uint32_t queue_size_bytes);
32 static void qat_queue_delete(struct qat_queue *queue);
33 static int qat_queue_create(struct qat_pci_device *qat_dev,
34 struct qat_queue *queue, struct qat_qp_config *, uint8_t dir);
35 static int adf_verify_queue_size(uint32_t msg_size, uint32_t msg_num,
36 uint32_t *queue_size_for_csr);
37 static int adf_configure_queues(struct qat_qp *queue,
38 enum qat_device_gen qat_dev_gen);
39 static int adf_queue_arb_enable(struct qat_pci_device *qat_dev,
40 struct qat_queue *txq, void *base_addr, rte_spinlock_t *lock);
41 static int adf_queue_arb_disable(enum qat_device_gen qat_dev_gen,
42 struct qat_queue *txq, void *base_addr, rte_spinlock_t *lock);
43 static int qat_qp_build_ring_base(struct qat_pci_device *qat_dev,
44 void *io_addr, struct qat_queue *queue);
45 static const struct rte_memzone *queue_dma_zone_reserve(const char *queue_name,
46 uint32_t queue_size, int socket_id);
47 static int qat_qp_csr_setup(struct qat_pci_device *qat_dev, void *io_addr,
51 qat_qp_setup(struct qat_pci_device *qat_dev,
52 struct qat_qp **qp_addr,
53 uint16_t queue_pair_id,
54 struct qat_qp_config *qat_qp_conf)
56 struct qat_qp *qp = NULL;
57 struct rte_pci_device *pci_dev =
58 qat_pci_devs[qat_dev->qat_dev_id].pci_dev;
59 char op_cookie_pool_name[RTE_RING_NAMESIZE];
60 struct qat_dev_hw_spec_funcs *ops_hw =
61 qat_dev_hw_spec[qat_dev->qat_dev_gen];
65 QAT_LOG(DEBUG, "Setup qp %u on qat pci device %d gen %d",
66 queue_pair_id, qat_dev->qat_dev_id, qat_dev->qat_dev_gen);
68 if ((qat_qp_conf->nb_descriptors > ADF_MAX_DESC) ||
69 (qat_qp_conf->nb_descriptors < ADF_MIN_DESC)) {
70 QAT_LOG(ERR, "Can't create qp for %u descriptors",
71 qat_qp_conf->nb_descriptors);
75 if (ops_hw->qat_dev_get_transport_bar == NULL) {
77 "QAT Internal Error: qat_dev_get_transport_bar not set for gen %d",
78 qat_dev->qat_dev_gen);
82 io_addr = ops_hw->qat_dev_get_transport_bar(pci_dev)->addr;
83 if (io_addr == NULL) {
84 QAT_LOG(ERR, "Could not find VF config space "
85 "(UIO driver attached?).");
89 /* Allocate the queue pair data structure. */
90 qp = rte_zmalloc_socket("qat PMD qp metadata",
91 sizeof(*qp), RTE_CACHE_LINE_SIZE,
92 qat_qp_conf->socket_id);
94 QAT_LOG(ERR, "Failed to alloc mem for qp struct");
97 qp->nb_descriptors = qat_qp_conf->nb_descriptors;
98 qp->op_cookies = rte_zmalloc_socket("qat PMD op cookie pointer",
99 qat_qp_conf->nb_descriptors * sizeof(*qp->op_cookies),
100 RTE_CACHE_LINE_SIZE, qat_qp_conf->socket_id);
101 if (qp->op_cookies == NULL) {
102 QAT_LOG(ERR, "Failed to alloc mem for cookie");
107 qp->mmap_bar_addr = io_addr;
108 qp->enqueued = qp->dequeued = 0;
110 if (qat_queue_create(qat_dev, &(qp->tx_q), qat_qp_conf,
111 ADF_RING_DIR_TX) != 0) {
112 QAT_LOG(ERR, "Tx queue create failed "
113 "queue_pair_id=%u", queue_pair_id);
117 qp->max_inflights = ADF_MAX_INFLIGHTS(qp->tx_q.queue_size,
118 ADF_BYTES_TO_MSG_SIZE(qp->tx_q.msg_size));
120 if (qp->max_inflights < 2) {
121 QAT_LOG(ERR, "Invalid num inflights");
122 qat_queue_delete(&(qp->tx_q));
126 if (qat_queue_create(qat_dev, &(qp->rx_q), qat_qp_conf,
127 ADF_RING_DIR_RX) != 0) {
128 QAT_LOG(ERR, "Rx queue create failed "
129 "queue_pair_id=%hu", queue_pair_id);
130 qat_queue_delete(&(qp->tx_q));
134 snprintf(op_cookie_pool_name, RTE_RING_NAMESIZE,
135 "%s%d_cookies_%s_qp%hu",
136 pci_dev->driver->driver.name, qat_dev->qat_dev_id,
137 qat_qp_conf->service_str, queue_pair_id);
139 QAT_LOG(DEBUG, "cookiepool: %s", op_cookie_pool_name);
140 qp->op_cookie_pool = rte_mempool_lookup(op_cookie_pool_name);
141 if (qp->op_cookie_pool == NULL)
142 qp->op_cookie_pool = rte_mempool_create(op_cookie_pool_name,
144 qat_qp_conf->cookie_size, 64, 0,
145 NULL, NULL, NULL, NULL,
146 pci_dev->device.numa_node,
148 if (!qp->op_cookie_pool) {
149 QAT_LOG(ERR, "QAT PMD Cannot create"
151 qat_queue_delete(&(qp->tx_q));
152 qat_queue_delete(&(qp->rx_q));
156 for (i = 0; i < qp->nb_descriptors; i++) {
157 if (rte_mempool_get(qp->op_cookie_pool, &qp->op_cookies[i])) {
158 QAT_LOG(ERR, "QAT PMD Cannot get op_cookie");
161 memset(qp->op_cookies[i], 0, qat_qp_conf->cookie_size);
164 qp->qat_dev_gen = qat_dev->qat_dev_gen;
165 qp->service_type = qat_qp_conf->hw->service_type;
166 qp->qat_dev = qat_dev;
168 QAT_LOG(DEBUG, "QP setup complete: id: %d, cookiepool: %s",
169 queue_pair_id, op_cookie_pool_name);
171 qat_qp_csr_setup(qat_dev, io_addr, qp);
178 rte_mempool_free(qp->op_cookie_pool);
180 rte_free(qp->op_cookies);
189 qat_queue_create(struct qat_pci_device *qat_dev, struct qat_queue *queue,
190 struct qat_qp_config *qp_conf, uint8_t dir)
192 const struct rte_memzone *qp_mz;
193 struct rte_pci_device *pci_dev =
194 qat_pci_devs[qat_dev->qat_dev_id].pci_dev;
196 uint16_t desc_size = (dir == ADF_RING_DIR_TX ?
197 qp_conf->hw->tx_msg_size : qp_conf->hw->rx_msg_size);
198 uint32_t queue_size_bytes = (qp_conf->nb_descriptors)*(desc_size);
200 queue->hw_bundle_number = qp_conf->hw->hw_bundle_num;
201 queue->hw_queue_number = (dir == ADF_RING_DIR_TX ?
202 qp_conf->hw->tx_ring_num : qp_conf->hw->rx_ring_num);
204 if (desc_size > ADF_MSG_SIZE_TO_BYTES(ADF_MAX_MSG_SIZE)) {
205 QAT_LOG(ERR, "Invalid descriptor size %d", desc_size);
210 * Allocate a memzone for the queue - create a unique name.
212 snprintf(queue->memz_name, sizeof(queue->memz_name),
214 pci_dev->driver->driver.name, qat_dev->qat_dev_id,
215 qp_conf->service_str, "qp_mem",
216 queue->hw_bundle_number, queue->hw_queue_number);
217 qp_mz = queue_dma_zone_reserve(queue->memz_name, queue_size_bytes,
218 pci_dev->device.numa_node);
220 QAT_LOG(ERR, "Failed to allocate ring memzone");
224 queue->base_addr = (char *)qp_mz->addr;
225 queue->base_phys_addr = qp_mz->iova;
226 if (qat_qp_check_queue_alignment(queue->base_phys_addr,
228 QAT_LOG(ERR, "Invalid alignment on queue create "
230 queue->base_phys_addr);
232 goto queue_create_err;
235 if (adf_verify_queue_size(desc_size, qp_conf->nb_descriptors,
236 &(queue->queue_size)) != 0) {
237 QAT_LOG(ERR, "Invalid num inflights");
239 goto queue_create_err;
242 queue->modulo_mask = (1 << ADF_RING_SIZE_MODULO(queue->queue_size)) - 1;
245 queue->msg_size = desc_size;
247 /* For fast calculation of cookie index, relies on msg_size being 2^n */
248 queue->trailz = __builtin_ctz(desc_size);
251 * Write an unused pattern to the queue memory.
253 memset(queue->base_addr, 0x7F, queue_size_bytes);
255 QAT_LOG(DEBUG, "RING: Name:%s, size in CSR: %u, in bytes %u,"
256 " nb msgs %u, msg_size %u, modulo mask %u",
258 queue->queue_size, queue_size_bytes,
259 qp_conf->nb_descriptors, desc_size,
265 rte_memzone_free(qp_mz);
269 static const struct rte_memzone *
270 queue_dma_zone_reserve(const char *queue_name, uint32_t queue_size,
273 const struct rte_memzone *mz;
275 mz = rte_memzone_lookup(queue_name);
277 if (((size_t)queue_size <= mz->len) &&
278 ((socket_id == SOCKET_ID_ANY) ||
279 (socket_id == mz->socket_id))) {
280 QAT_LOG(DEBUG, "re-use memzone already "
281 "allocated for %s", queue_name);
285 QAT_LOG(ERR, "Incompatible memzone already "
286 "allocated %s, size %u, socket %d. "
287 "Requested size %u, socket %u",
288 queue_name, (uint32_t)mz->len,
289 mz->socket_id, queue_size, socket_id);
293 QAT_LOG(DEBUG, "Allocate memzone for %s, size %u on socket %u",
294 queue_name, queue_size, socket_id);
295 return rte_memzone_reserve_aligned(queue_name, queue_size,
296 socket_id, RTE_MEMZONE_IOVA_CONTIG, queue_size);
300 qat_qp_release(enum qat_device_gen qat_dev_gen, struct qat_qp **qp_addr)
303 struct qat_qp *qp = *qp_addr;
307 QAT_LOG(DEBUG, "qp already freed");
311 QAT_LOG(DEBUG, "Free qp on qat_pci device %d",
312 qp->qat_dev->qat_dev_id);
314 /* Don't free memory if there are still responses to be processed */
315 if ((qp->enqueued - qp->dequeued) == 0) {
316 qat_queue_delete(&(qp->tx_q));
317 qat_queue_delete(&(qp->rx_q));
322 ret = adf_queue_arb_disable(qat_dev_gen, &(qp->tx_q),
323 qp->mmap_bar_addr, &qp->qat_dev->arb_csr_lock);
327 for (i = 0; i < qp->nb_descriptors; i++)
328 rte_mempool_put(qp->op_cookie_pool, qp->op_cookies[i]);
330 rte_mempool_free(qp->op_cookie_pool);
332 rte_free(qp->op_cookies);
340 qat_queue_delete(struct qat_queue *queue)
342 const struct rte_memzone *mz;
346 QAT_LOG(DEBUG, "Invalid queue");
349 QAT_LOG(DEBUG, "Free ring %d, memzone: %s",
350 queue->hw_queue_number, queue->memz_name);
352 mz = rte_memzone_lookup(queue->memz_name);
354 /* Write an unused pattern to the queue memory. */
355 memset(queue->base_addr, 0x7F, queue->queue_size);
356 status = rte_memzone_free(mz);
358 QAT_LOG(ERR, "Error %d on freeing queue %s",
359 status, queue->memz_name);
361 QAT_LOG(DEBUG, "queue %s doesn't exist",
366 static int __rte_unused
367 adf_queue_arb_enable(struct qat_pci_device *qat_dev, struct qat_queue *txq,
368 void *base_addr, rte_spinlock_t *lock)
370 struct qat_qp_hw_spec_funcs *ops =
371 qat_qp_hw_spec[qat_dev->qat_dev_gen];
373 RTE_FUNC_PTR_OR_ERR_RET(ops->qat_qp_adf_arb_enable,
375 ops->qat_qp_adf_arb_enable(txq, base_addr, lock);
380 adf_queue_arb_disable(enum qat_device_gen qat_dev_gen, struct qat_queue *txq,
381 void *base_addr, rte_spinlock_t *lock)
383 struct qat_qp_hw_spec_funcs *ops =
384 qat_qp_hw_spec[qat_dev_gen];
386 RTE_FUNC_PTR_OR_ERR_RET(ops->qat_qp_adf_arb_disable,
388 ops->qat_qp_adf_arb_disable(txq, base_addr, lock);
392 static int __rte_unused
393 qat_qp_build_ring_base(struct qat_pci_device *qat_dev, void *io_addr,
394 struct qat_queue *queue)
396 struct qat_qp_hw_spec_funcs *ops =
397 qat_qp_hw_spec[qat_dev->qat_dev_gen];
399 RTE_FUNC_PTR_OR_ERR_RET(ops->qat_qp_build_ring_base,
401 ops->qat_qp_build_ring_base(io_addr, queue);
406 qat_qps_per_service(struct qat_pci_device *qat_dev,
407 enum qat_service_type service)
409 struct qat_qp_hw_spec_funcs *ops =
410 qat_qp_hw_spec[qat_dev->qat_dev_gen];
412 RTE_FUNC_PTR_OR_ERR_RET(ops->qat_qp_rings_per_service,
414 return ops->qat_qp_rings_per_service(qat_dev, service);
417 const struct qat_qp_hw_data *
418 qat_qp_get_hw_data(struct qat_pci_device *qat_dev,
419 enum qat_service_type service, uint16_t qp_id)
421 struct qat_qp_hw_spec_funcs *ops =
422 qat_qp_hw_spec[qat_dev->qat_dev_gen];
424 RTE_FUNC_PTR_OR_ERR_RET(ops->qat_qp_get_hw_data, NULL);
425 return ops->qat_qp_get_hw_data(qat_dev, service, qp_id);
429 qat_read_qp_config(struct qat_pci_device *qat_dev)
431 struct qat_dev_hw_spec_funcs *ops_hw =
432 qat_dev_hw_spec[qat_dev->qat_dev_gen];
434 RTE_FUNC_PTR_OR_ERR_RET(ops_hw->qat_dev_read_config,
436 return ops_hw->qat_dev_read_config(qat_dev);
439 static int __rte_unused
440 adf_configure_queues(struct qat_qp *qp, enum qat_device_gen qat_dev_gen)
442 struct qat_qp_hw_spec_funcs *ops =
443 qat_qp_hw_spec[qat_dev_gen];
445 RTE_FUNC_PTR_OR_ERR_RET(ops->qat_qp_adf_configure_queues,
447 ops->qat_qp_adf_configure_queues(qp);
452 txq_write_tail(enum qat_device_gen qat_dev_gen,
453 struct qat_qp *qp, struct qat_queue *q)
455 struct qat_qp_hw_spec_funcs *ops =
456 qat_qp_hw_spec[qat_dev_gen];
459 * Pointer check should be done during
462 ops->qat_qp_csr_write_tail(qp, q);
466 qat_qp_csr_write_head(enum qat_device_gen qat_dev_gen, struct qat_qp *qp,
467 struct qat_queue *q, uint32_t new_head)
469 struct qat_qp_hw_spec_funcs *ops =
470 qat_qp_hw_spec[qat_dev_gen];
473 * Pointer check should be done during
476 ops->qat_qp_csr_write_head(qp, q, new_head);
480 qat_qp_csr_setup(struct qat_pci_device *qat_dev,
481 void *io_addr, struct qat_qp *qp)
483 struct qat_qp_hw_spec_funcs *ops =
484 qat_qp_hw_spec[qat_dev->qat_dev_gen];
486 RTE_FUNC_PTR_OR_ERR_RET(ops->qat_qp_csr_setup,
488 ops->qat_qp_csr_setup(qat_dev, io_addr, qp);
494 void rxq_free_desc(enum qat_device_gen qat_dev_gen, struct qat_qp *qp,
497 uint32_t old_head, new_head;
500 old_head = q->csr_head;
502 max_head = qp->nb_descriptors * q->msg_size;
504 /* write out free descriptors */
505 void *cur_desc = (uint8_t *)q->base_addr + old_head;
507 if (new_head < old_head) {
508 memset(cur_desc, ADF_RING_EMPTY_SIG_BYTE, max_head - old_head);
509 memset(q->base_addr, ADF_RING_EMPTY_SIG_BYTE, new_head);
511 memset(cur_desc, ADF_RING_EMPTY_SIG_BYTE, new_head - old_head);
513 q->nb_processed_responses = 0;
514 q->csr_head = new_head;
516 qat_qp_csr_write_head(qat_dev_gen, qp, q, new_head);
520 qat_qp_check_queue_alignment(uint64_t phys_addr, uint32_t queue_size_bytes)
522 if (((queue_size_bytes - 1) & phys_addr) != 0)
528 adf_verify_queue_size(uint32_t msg_size, uint32_t msg_num,
529 uint32_t *p_queue_size_for_csr)
531 uint8_t i = ADF_MIN_RING_SIZE;
533 for (; i <= ADF_MAX_RING_SIZE; i++)
534 if ((msg_size * msg_num) ==
535 (uint32_t)ADF_SIZE_TO_RING_SIZE_IN_BYTES(i)) {
536 *p_queue_size_for_csr = i;
539 QAT_LOG(ERR, "Invalid ring size %d", msg_size * msg_num);
543 static inline uint32_t
544 adf_modulo(uint32_t data, uint32_t modulo_mask)
546 return data & modulo_mask;
550 qat_enqueue_op_burst(void *qp,
551 __rte_unused qat_op_build_request_t op_build_request,
552 void **ops, uint16_t nb_ops)
554 register struct qat_queue *queue;
555 struct qat_qp *tmp_qp = (struct qat_qp *)qp;
556 register uint32_t nb_ops_sent = 0;
557 register int ret = -1;
558 uint16_t nb_ops_possible = nb_ops;
559 register uint8_t *base_addr;
560 register uint32_t tail;
562 if (unlikely(nb_ops == 0))
565 /* read params used a lot in main loop into registers */
566 queue = &(tmp_qp->tx_q);
567 base_addr = (uint8_t *)queue->base_addr;
570 /* Find how many can actually fit on the ring */
572 /* dequeued can only be written by one thread, but it may not
573 * be this thread. As it's 4-byte aligned it will be read
574 * atomically here by any Intel CPU.
575 * enqueued can wrap before dequeued, but cannot
576 * lap it as var size of enq/deq (uint32_t) > var size of
577 * max_inflights (uint16_t). In reality inflights is never
578 * even as big as max uint16_t, as it's <= ADF_MAX_DESC.
579 * On wrapping, the calculation still returns the correct
580 * positive value as all three vars are unsigned.
583 tmp_qp->enqueued - tmp_qp->dequeued;
585 if ((inflights + nb_ops) > tmp_qp->max_inflights) {
586 nb_ops_possible = tmp_qp->max_inflights - inflights;
587 if (nb_ops_possible == 0)
590 /* QAT has plenty of work queued already, so don't waste cycles
591 * enqueueing, wait til the application has gathered a bigger
592 * burst or some completed ops have been dequeued
594 if (tmp_qp->min_enq_burst_threshold && inflights >
595 QAT_QP_MIN_INFL_THRESHOLD && nb_ops_possible <
596 tmp_qp->min_enq_burst_threshold) {
597 tmp_qp->stats.threshold_hit_count++;
603 if (tmp_qp->service_type == QAT_SERVICE_SYMMETRIC)
604 qat_sym_preprocess_requests(ops, nb_ops_possible);
607 while (nb_ops_sent != nb_ops_possible) {
608 if (tmp_qp->service_type == QAT_SERVICE_SYMMETRIC) {
610 ret = qat_sym_build_request(*ops, base_addr + tail,
611 tmp_qp->op_cookies[tail >> queue->trailz],
612 tmp_qp->qat_dev_gen);
614 } else if (tmp_qp->service_type == QAT_SERVICE_COMPRESSION) {
615 ret = qat_comp_build_request(*ops, base_addr + tail,
616 tmp_qp->op_cookies[tail >> queue->trailz],
617 tmp_qp->qat_dev_gen);
618 } else if (tmp_qp->service_type == QAT_SERVICE_ASYMMETRIC) {
619 #ifdef BUILD_QAT_ASYM
620 ret = qat_asym_build_request(*ops, base_addr + tail,
621 tmp_qp->op_cookies[tail >> queue->trailz],
622 NULL, tmp_qp->qat_dev_gen);
626 tmp_qp->stats.enqueue_err_count++;
627 /* This message cannot be enqueued */
628 if (nb_ops_sent == 0)
633 tail = adf_modulo(tail + queue->msg_size, queue->modulo_mask);
639 tmp_qp->enqueued += nb_ops_sent;
640 tmp_qp->stats.enqueued_count += nb_ops_sent;
641 txq_write_tail(tmp_qp->qat_dev_gen, tmp_qp, queue);
645 /* Use this for compression only - but keep consistent with above common
646 * function as much as possible.
649 qat_enqueue_comp_op_burst(void *qp, void **ops, uint16_t nb_ops)
651 register struct qat_queue *queue;
652 struct qat_qp *tmp_qp = (struct qat_qp *)qp;
653 register uint32_t nb_ops_sent = 0;
654 register int nb_desc_to_build;
655 uint16_t nb_ops_possible = nb_ops;
656 register uint8_t *base_addr;
657 register uint32_t tail;
659 int descriptors_built, total_descriptors_built = 0;
660 int nb_remaining_descriptors;
663 if (unlikely(nb_ops == 0))
666 /* read params used a lot in main loop into registers */
667 queue = &(tmp_qp->tx_q);
668 base_addr = (uint8_t *)queue->base_addr;
671 /* Find how many can actually fit on the ring */
673 /* dequeued can only be written by one thread, but it may not
674 * be this thread. As it's 4-byte aligned it will be read
675 * atomically here by any Intel CPU.
676 * enqueued can wrap before dequeued, but cannot
677 * lap it as var size of enq/deq (uint32_t) > var size of
678 * max_inflights (uint16_t). In reality inflights is never
679 * even as big as max uint16_t, as it's <= ADF_MAX_DESC.
680 * On wrapping, the calculation still returns the correct
681 * positive value as all three vars are unsigned.
684 tmp_qp->enqueued - tmp_qp->dequeued;
686 /* Find how many can actually fit on the ring */
687 overflow = (inflights + nb_ops) - tmp_qp->max_inflights;
689 nb_ops_possible = nb_ops - overflow;
690 if (nb_ops_possible == 0)
694 /* QAT has plenty of work queued already, so don't waste cycles
695 * enqueueing, wait til the application has gathered a bigger
696 * burst or some completed ops have been dequeued
698 if (tmp_qp->min_enq_burst_threshold && inflights >
699 QAT_QP_MIN_INFL_THRESHOLD && nb_ops_possible <
700 tmp_qp->min_enq_burst_threshold) {
701 tmp_qp->stats.threshold_hit_count++;
706 /* At this point nb_ops_possible is assuming a 1:1 mapping
707 * between ops and descriptors.
708 * Fewer may be sent if some ops have to be split.
709 * nb_ops_possible is <= burst size.
710 * Find out how many spaces are actually available on the qp in case
713 nb_remaining_descriptors = nb_ops_possible
714 + ((overflow >= 0) ? 0 : overflow * (-1));
715 QAT_DP_LOG(DEBUG, "Nb ops requested %d, nb descriptors remaining %d",
716 nb_ops, nb_remaining_descriptors);
718 while (nb_ops_sent != nb_ops_possible &&
719 nb_remaining_descriptors > 0) {
720 struct qat_comp_op_cookie *cookie =
721 tmp_qp->op_cookies[tail >> queue->trailz];
723 descriptors_built = 0;
725 QAT_DP_LOG(DEBUG, "--- data length: %u",
726 ((struct rte_comp_op *)*ops)->src.length);
728 nb_desc_to_build = qat_comp_build_request(*ops,
729 base_addr + tail, cookie, tmp_qp->qat_dev_gen);
730 QAT_DP_LOG(DEBUG, "%d descriptors built, %d remaining, "
731 "%d ops sent, %d descriptors needed",
732 total_descriptors_built, nb_remaining_descriptors,
733 nb_ops_sent, nb_desc_to_build);
735 if (unlikely(nb_desc_to_build < 0)) {
736 /* this message cannot be enqueued */
737 tmp_qp->stats.enqueue_err_count++;
738 if (nb_ops_sent == 0)
741 } else if (unlikely(nb_desc_to_build > 1)) {
742 /* this op is too big and must be split - get more
743 * descriptors and retry
746 QAT_DP_LOG(DEBUG, "Build %d descriptors for this op",
749 nb_remaining_descriptors -= nb_desc_to_build;
750 if (nb_remaining_descriptors >= 0) {
751 /* There are enough remaining descriptors
754 int ret2 = qat_comp_build_multiple_requests(
758 if (unlikely(ret2 < 1)) {
760 "Failed to build (%d) descriptors, status %d",
761 nb_desc_to_build, ret2);
763 qat_comp_free_split_op_memzones(cookie,
764 nb_desc_to_build - 1);
766 tmp_qp->stats.enqueue_err_count++;
768 /* This message cannot be enqueued */
769 if (nb_ops_sent == 0)
773 descriptors_built = ret2;
774 total_descriptors_built +=
776 nb_remaining_descriptors -=
779 "Multiple descriptors (%d) built ok",
783 QAT_DP_LOG(ERR, "For the current op, number of requested descriptors (%d) "
784 "exceeds number of available descriptors (%d)",
786 nb_remaining_descriptors +
789 qat_comp_free_split_op_memzones(cookie,
790 nb_desc_to_build - 1);
792 /* Not enough extra descriptors */
793 if (nb_ops_sent == 0)
798 descriptors_built = 1;
799 total_descriptors_built++;
800 nb_remaining_descriptors--;
801 QAT_DP_LOG(DEBUG, "Single descriptor built ok");
804 tail = adf_modulo(tail + (queue->msg_size * descriptors_built),
812 tmp_qp->enqueued += total_descriptors_built;
813 tmp_qp->stats.enqueued_count += nb_ops_sent;
814 txq_write_tail(tmp_qp->qat_dev_gen, tmp_qp, queue);
819 qat_dequeue_op_burst(void *qp, void **ops,
820 __rte_unused qat_op_dequeue_t qat_dequeue_process_response,
823 struct qat_queue *rx_queue;
824 struct qat_qp *tmp_qp = (struct qat_qp *)qp;
826 uint32_t op_resp_counter = 0, fw_resp_counter = 0;
830 rx_queue = &(tmp_qp->rx_q);
831 head = rx_queue->head;
832 resp_msg = (uint8_t *)rx_queue->base_addr + rx_queue->head;
834 while (*(uint32_t *)resp_msg != ADF_RING_EMPTY_SIG &&
835 op_resp_counter != nb_ops) {
839 if (tmp_qp->service_type == QAT_SERVICE_SYMMETRIC)
840 qat_sym_process_response(ops, resp_msg,
841 tmp_qp->op_cookies[head >> rx_queue->trailz],
843 else if (tmp_qp->service_type == QAT_SERVICE_COMPRESSION)
844 nb_fw_responses = qat_comp_process_response(
846 tmp_qp->op_cookies[head >> rx_queue->trailz],
847 &tmp_qp->stats.dequeue_err_count);
848 #ifdef BUILD_QAT_ASYM
849 else if (tmp_qp->service_type == QAT_SERVICE_ASYMMETRIC)
850 qat_asym_process_response(ops, resp_msg,
851 tmp_qp->op_cookies[head >> rx_queue->trailz],
855 head = adf_modulo(head + rx_queue->msg_size,
856 rx_queue->modulo_mask);
858 resp_msg = (uint8_t *)rx_queue->base_addr + head;
860 if (nb_fw_responses) {
861 /* only move on to next op if one was ready to return
868 /* A compression op may be broken up into multiple fw requests.
869 * Only count fw responses as complete once ALL the responses
870 * associated with an op have been processed, as the cookie
871 * data from the first response must be available until
872 * finished with all firmware responses.
874 fw_resp_counter += nb_fw_responses;
876 rx_queue->nb_processed_responses++;
879 tmp_qp->dequeued += fw_resp_counter;
880 tmp_qp->stats.dequeued_count += op_resp_counter;
882 rx_queue->head = head;
883 if (rx_queue->nb_processed_responses > QAT_CSR_HEAD_WRITE_THRESH)
884 rxq_free_desc(tmp_qp->qat_dev_gen, tmp_qp, rx_queue);
886 QAT_DP_LOG(DEBUG, "Dequeue burst return: %u, QAT responses: %u",
887 op_resp_counter, fw_resp_counter);
889 return op_resp_counter;
892 /* This is almost same as dequeue_op_burst, without the atomic, without stats
893 * and without the op. Dequeues one response.
896 qat_cq_dequeue_response(struct qat_qp *qp, void *out_data)
900 struct qat_queue *queue = &(qp->rx_q);
901 struct icp_qat_fw_comn_resp *resp_msg = (struct icp_qat_fw_comn_resp *)
902 ((uint8_t *)queue->base_addr + queue->head);
904 while (retries++ < QAT_CQ_MAX_DEQ_RETRIES &&
905 *(uint32_t *)resp_msg == ADF_RING_EMPTY_SIG) {
906 /* loop waiting for response until we reach the timeout */
910 if (*(uint32_t *)resp_msg != ADF_RING_EMPTY_SIG) {
911 /* response received */
914 /* check status flag */
915 if (ICP_QAT_FW_COMN_RESP_CRYPTO_STAT_GET(
916 resp_msg->comn_hdr.comn_status) ==
917 ICP_QAT_FW_COMN_STATUS_FLAG_OK) {
919 memcpy(out_data, resp_msg, queue->msg_size);
921 memset(out_data, 0, queue->msg_size);
924 queue->head = adf_modulo(queue->head + queue->msg_size,
926 rxq_free_desc(qp->qat_dev_gen, qp, queue);
932 /* Sends a NULL message and extracts QAT fw version from the response.
933 * Used to determine detailed capabilities based on the fw version number.
934 * This assumes that there are no inflight messages, i.e. assumes there's space
935 * on the qp, one message is sent and only one response collected.
936 * Returns fw version number or 0 for unknown version or a negative error code.
939 qat_cq_get_fw_version(struct qat_qp *qp)
941 struct qat_queue *queue = &(qp->tx_q);
942 uint8_t *base_addr = (uint8_t *)queue->base_addr;
943 struct icp_qat_fw_comn_req null_msg;
944 struct icp_qat_fw_comn_resp response;
946 /* prepare the NULL request */
947 memset(&null_msg, 0, sizeof(null_msg));
948 null_msg.comn_hdr.hdr_flags =
949 ICP_QAT_FW_COMN_HDR_FLAGS_BUILD(ICP_QAT_FW_COMN_REQ_FLAG_SET);
950 null_msg.comn_hdr.service_type = ICP_QAT_FW_COMN_REQ_NULL;
951 null_msg.comn_hdr.service_cmd_id = ICP_QAT_FW_NULL_REQ_SERV_ID;
953 #if RTE_LOG_DP_LEVEL >= RTE_LOG_DEBUG
954 QAT_DP_HEXDUMP_LOG(DEBUG, "NULL request", &null_msg, sizeof(null_msg));
957 /* send the NULL request */
958 memcpy(base_addr + queue->tail, &null_msg, sizeof(null_msg));
959 queue->tail = adf_modulo(queue->tail + queue->msg_size,
961 txq_write_tail(qp->qat_dev_gen, qp, queue);
963 /* receive a response */
964 if (qat_cq_dequeue_response(qp, &response)) {
966 #if RTE_LOG_DP_LEVEL >= RTE_LOG_DEBUG
967 QAT_DP_HEXDUMP_LOG(DEBUG, "NULL response:", &response,
970 /* if LW0 bit 24 is set - then the fw version was returned */
971 if (QAT_FIELD_GET(response.comn_hdr.hdr_flags,
972 ICP_QAT_FW_COMN_NULL_VERSION_FLAG_BITPOS,
973 ICP_QAT_FW_COMN_NULL_VERSION_FLAG_MASK))
974 return response.resrvd[0]; /* return LW4 */
976 return 0; /* not set - we don't know fw version */
979 QAT_LOG(ERR, "No response received");
984 qat_comp_process_response(void **op __rte_unused, uint8_t *resp __rte_unused,
985 void *op_cookie __rte_unused,
986 uint64_t *dequeue_err_count __rte_unused)