1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2015-2018 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"
21 #include "adf_transport_access_macros.h"
23 #define QAT_CQ_MAX_DEQ_RETRIES 10
25 #define ADF_MAX_DESC 4096
26 #define ADF_MIN_DESC 128
28 #define ADF_ARB_REG_SLOT 0x1000
29 #define ADF_ARB_RINGSRVARBEN_OFFSET 0x19C
31 #define WRITE_CSR_ARB_RINGSRVARBEN(csr_addr, index, value) \
32 ADF_CSR_WR(csr_addr, ADF_ARB_RINGSRVARBEN_OFFSET + \
33 (ADF_ARB_REG_SLOT * index), value)
36 const struct qat_qp_hw_data qat_gen1_qps
[QAT_MAX_SERVICES
]
37 [ADF_MAX_QPS_ON_ANY_SERVICE
] = {
38 /* queue pairs which provide an asymmetric crypto service */
39 [QAT_SERVICE_ASYMMETRIC
] = {
41 .service_type
= QAT_SERVICE_ASYMMETRIC
,
49 .service_type
= QAT_SERVICE_ASYMMETRIC
,
57 /* queue pairs which provide a symmetric crypto service */
58 [QAT_SERVICE_SYMMETRIC
] = {
60 .service_type
= QAT_SERVICE_SYMMETRIC
,
68 .service_type
= QAT_SERVICE_SYMMETRIC
,
76 /* queue pairs which provide a compression service */
77 [QAT_SERVICE_COMPRESSION
] = {
79 .service_type
= QAT_SERVICE_COMPRESSION
,
86 .service_type
= QAT_SERVICE_COMPRESSION
,
97 const struct qat_qp_hw_data qat_gen3_qps
[QAT_MAX_SERVICES
]
98 [ADF_MAX_QPS_ON_ANY_SERVICE
] = {
99 /* queue pairs which provide an asymmetric crypto service */
100 [QAT_SERVICE_ASYMMETRIC
] = {
102 .service_type
= QAT_SERVICE_ASYMMETRIC
,
110 /* queue pairs which provide a symmetric crypto service */
111 [QAT_SERVICE_SYMMETRIC
] = {
113 .service_type
= QAT_SERVICE_SYMMETRIC
,
121 /* queue pairs which provide a compression service */
122 [QAT_SERVICE_COMPRESSION
] = {
124 .service_type
= QAT_SERVICE_COMPRESSION
,
134 static int qat_qp_check_queue_alignment(uint64_t phys_addr
,
135 uint32_t queue_size_bytes
);
136 static void qat_queue_delete(struct qat_queue
*queue
);
137 static int qat_queue_create(struct qat_pci_device
*qat_dev
,
138 struct qat_queue
*queue
, struct qat_qp_config
*, uint8_t dir
);
139 static int adf_verify_queue_size(uint32_t msg_size
, uint32_t msg_num
,
140 uint32_t *queue_size_for_csr
);
141 static void adf_configure_queues(struct qat_qp
*queue
);
142 static void adf_queue_arb_enable(struct qat_queue
*txq
, void *base_addr
,
143 rte_spinlock_t
*lock
);
144 static void adf_queue_arb_disable(struct qat_queue
*txq
, void *base_addr
,
145 rte_spinlock_t
*lock
);
148 int qat_qps_per_service(const struct qat_qp_hw_data
*qp_hw_data
,
149 enum qat_service_type service
)
153 for (i
= 0, count
= 0; i
< ADF_MAX_QPS_ON_ANY_SERVICE
; i
++)
154 if (qp_hw_data
[i
].service_type
== service
)
159 static const struct rte_memzone
*
160 queue_dma_zone_reserve(const char *queue_name
, uint32_t queue_size
,
163 const struct rte_memzone
*mz
;
165 mz
= rte_memzone_lookup(queue_name
);
167 if (((size_t)queue_size
<= mz
->len
) &&
168 ((socket_id
== SOCKET_ID_ANY
) ||
169 (socket_id
== mz
->socket_id
))) {
170 QAT_LOG(DEBUG
, "re-use memzone already "
171 "allocated for %s", queue_name
);
175 QAT_LOG(ERR
, "Incompatible memzone already "
176 "allocated %s, size %u, socket %d. "
177 "Requested size %u, socket %u",
178 queue_name
, (uint32_t)mz
->len
,
179 mz
->socket_id
, queue_size
, socket_id
);
183 QAT_LOG(DEBUG
, "Allocate memzone for %s, size %u on socket %u",
184 queue_name
, queue_size
, socket_id
);
185 return rte_memzone_reserve_aligned(queue_name
, queue_size
,
186 socket_id
, RTE_MEMZONE_IOVA_CONTIG
, queue_size
);
189 int qat_qp_setup(struct qat_pci_device
*qat_dev
,
190 struct qat_qp
**qp_addr
,
191 uint16_t queue_pair_id
,
192 struct qat_qp_config
*qat_qp_conf
)
196 struct rte_pci_device
*pci_dev
= qat_dev
->pci_dev
;
197 char op_cookie_pool_name
[RTE_RING_NAMESIZE
];
200 QAT_LOG(DEBUG
, "Setup qp %u on qat pci device %d gen %d",
201 queue_pair_id
, qat_dev
->qat_dev_id
, qat_dev
->qat_dev_gen
);
203 if ((qat_qp_conf
->nb_descriptors
> ADF_MAX_DESC
) ||
204 (qat_qp_conf
->nb_descriptors
< ADF_MIN_DESC
)) {
205 QAT_LOG(ERR
, "Can't create qp for %u descriptors",
206 qat_qp_conf
->nb_descriptors
);
210 if (pci_dev
->mem_resource
[0].addr
== NULL
) {
211 QAT_LOG(ERR
, "Could not find VF config space "
212 "(UIO driver attached?).");
216 /* Allocate the queue pair data structure. */
217 qp
= rte_zmalloc_socket("qat PMD qp metadata",
218 sizeof(*qp
), RTE_CACHE_LINE_SIZE
,
219 qat_qp_conf
->socket_id
);
221 QAT_LOG(ERR
, "Failed to alloc mem for qp struct");
224 qp
->nb_descriptors
= qat_qp_conf
->nb_descriptors
;
225 qp
->op_cookies
= rte_zmalloc_socket("qat PMD op cookie pointer",
226 qat_qp_conf
->nb_descriptors
* sizeof(*qp
->op_cookies
),
227 RTE_CACHE_LINE_SIZE
, qat_qp_conf
->socket_id
);
228 if (qp
->op_cookies
== NULL
) {
229 QAT_LOG(ERR
, "Failed to alloc mem for cookie");
234 qp
->mmap_bar_addr
= pci_dev
->mem_resource
[0].addr
;
235 qp
->enqueued
= qp
->dequeued
= 0;
237 if (qat_queue_create(qat_dev
, &(qp
->tx_q
), qat_qp_conf
,
238 ADF_RING_DIR_TX
) != 0) {
239 QAT_LOG(ERR
, "Tx queue create failed "
240 "queue_pair_id=%u", queue_pair_id
);
244 qp
->max_inflights
= ADF_MAX_INFLIGHTS(qp
->tx_q
.queue_size
,
245 ADF_BYTES_TO_MSG_SIZE(qp
->tx_q
.msg_size
));
247 if (qp
->max_inflights
< 2) {
248 QAT_LOG(ERR
, "Invalid num inflights");
249 qat_queue_delete(&(qp
->tx_q
));
253 if (qat_queue_create(qat_dev
, &(qp
->rx_q
), qat_qp_conf
,
254 ADF_RING_DIR_RX
) != 0) {
255 QAT_LOG(ERR
, "Rx queue create failed "
256 "queue_pair_id=%hu", queue_pair_id
);
257 qat_queue_delete(&(qp
->tx_q
));
261 adf_configure_queues(qp
);
262 adf_queue_arb_enable(&qp
->tx_q
, qp
->mmap_bar_addr
,
263 &qat_dev
->arb_csr_lock
);
265 snprintf(op_cookie_pool_name
, RTE_RING_NAMESIZE
,
266 "%s%d_cookies_%s_qp%hu",
267 pci_dev
->driver
->driver
.name
, qat_dev
->qat_dev_id
,
268 qat_qp_conf
->service_str
, queue_pair_id
);
270 QAT_LOG(DEBUG
, "cookiepool: %s", op_cookie_pool_name
);
271 qp
->op_cookie_pool
= rte_mempool_lookup(op_cookie_pool_name
);
272 if (qp
->op_cookie_pool
== NULL
)
273 qp
->op_cookie_pool
= rte_mempool_create(op_cookie_pool_name
,
275 qat_qp_conf
->cookie_size
, 64, 0,
276 NULL
, NULL
, NULL
, NULL
,
277 qat_dev
->pci_dev
->device
.numa_node
,
279 if (!qp
->op_cookie_pool
) {
280 QAT_LOG(ERR
, "QAT PMD Cannot create"
285 for (i
= 0; i
< qp
->nb_descriptors
; i
++) {
286 if (rte_mempool_get(qp
->op_cookie_pool
, &qp
->op_cookies
[i
])) {
287 QAT_LOG(ERR
, "QAT PMD Cannot get op_cookie");
290 memset(qp
->op_cookies
[i
], 0, qat_qp_conf
->cookie_size
);
293 qp
->qat_dev_gen
= qat_dev
->qat_dev_gen
;
294 qp
->build_request
= qat_qp_conf
->build_request
;
295 qp
->service_type
= qat_qp_conf
->hw
->service_type
;
296 qp
->qat_dev
= qat_dev
;
298 QAT_LOG(DEBUG
, "QP setup complete: id: %d, cookiepool: %s",
299 queue_pair_id
, op_cookie_pool_name
);
305 if (qp
->op_cookie_pool
)
306 rte_mempool_free(qp
->op_cookie_pool
);
307 rte_free(qp
->op_cookies
);
312 int qat_qp_release(struct qat_qp
**qp_addr
)
314 struct qat_qp
*qp
= *qp_addr
;
318 QAT_LOG(DEBUG
, "qp already freed");
322 QAT_LOG(DEBUG
, "Free qp on qat_pci device %d",
323 qp
->qat_dev
->qat_dev_id
);
325 /* Don't free memory if there are still responses to be processed */
326 if ((qp
->enqueued
- qp
->dequeued
) == 0) {
327 qat_queue_delete(&(qp
->tx_q
));
328 qat_queue_delete(&(qp
->rx_q
));
333 adf_queue_arb_disable(&(qp
->tx_q
), qp
->mmap_bar_addr
,
334 &qp
->qat_dev
->arb_csr_lock
);
336 for (i
= 0; i
< qp
->nb_descriptors
; i
++)
337 rte_mempool_put(qp
->op_cookie_pool
, qp
->op_cookies
[i
]);
339 if (qp
->op_cookie_pool
)
340 rte_mempool_free(qp
->op_cookie_pool
);
342 rte_free(qp
->op_cookies
);
349 static void qat_queue_delete(struct qat_queue
*queue
)
351 const struct rte_memzone
*mz
;
355 QAT_LOG(DEBUG
, "Invalid queue");
358 QAT_LOG(DEBUG
, "Free ring %d, memzone: %s",
359 queue
->hw_queue_number
, queue
->memz_name
);
361 mz
= rte_memzone_lookup(queue
->memz_name
);
363 /* Write an unused pattern to the queue memory. */
364 memset(queue
->base_addr
, 0x7F, queue
->queue_size
);
365 status
= rte_memzone_free(mz
);
367 QAT_LOG(ERR
, "Error %d on freeing queue %s",
368 status
, queue
->memz_name
);
370 QAT_LOG(DEBUG
, "queue %s doesn't exist",
376 qat_queue_create(struct qat_pci_device
*qat_dev
, struct qat_queue
*queue
,
377 struct qat_qp_config
*qp_conf
, uint8_t dir
)
381 const struct rte_memzone
*qp_mz
;
382 struct rte_pci_device
*pci_dev
= qat_dev
->pci_dev
;
384 uint16_t desc_size
= (dir
== ADF_RING_DIR_TX
?
385 qp_conf
->hw
->tx_msg_size
: qp_conf
->hw
->rx_msg_size
);
386 uint32_t queue_size_bytes
= (qp_conf
->nb_descriptors
)*(desc_size
);
388 queue
->hw_bundle_number
= qp_conf
->hw
->hw_bundle_num
;
389 queue
->hw_queue_number
= (dir
== ADF_RING_DIR_TX
?
390 qp_conf
->hw
->tx_ring_num
: qp_conf
->hw
->rx_ring_num
);
392 if (desc_size
> ADF_MSG_SIZE_TO_BYTES(ADF_MAX_MSG_SIZE
)) {
393 QAT_LOG(ERR
, "Invalid descriptor size %d", desc_size
);
398 * Allocate a memzone for the queue - create a unique name.
400 snprintf(queue
->memz_name
, sizeof(queue
->memz_name
),
402 pci_dev
->driver
->driver
.name
, qat_dev
->qat_dev_id
,
403 qp_conf
->service_str
, "qp_mem",
404 queue
->hw_bundle_number
, queue
->hw_queue_number
);
405 qp_mz
= queue_dma_zone_reserve(queue
->memz_name
, queue_size_bytes
,
406 qat_dev
->pci_dev
->device
.numa_node
);
408 QAT_LOG(ERR
, "Failed to allocate ring memzone");
412 queue
->base_addr
= (char *)qp_mz
->addr
;
413 queue
->base_phys_addr
= qp_mz
->iova
;
414 if (qat_qp_check_queue_alignment(queue
->base_phys_addr
,
416 QAT_LOG(ERR
, "Invalid alignment on queue create "
418 queue
->base_phys_addr
);
420 goto queue_create_err
;
423 if (adf_verify_queue_size(desc_size
, qp_conf
->nb_descriptors
,
424 &(queue
->queue_size
)) != 0) {
425 QAT_LOG(ERR
, "Invalid num inflights");
427 goto queue_create_err
;
430 queue
->modulo_mask
= (1 << ADF_RING_SIZE_MODULO(queue
->queue_size
)) - 1;
433 queue
->msg_size
= desc_size
;
435 /* For fast calculation of cookie index, relies on msg_size being 2^n */
436 queue
->trailz
= __builtin_ctz(desc_size
);
439 * Write an unused pattern to the queue memory.
441 memset(queue
->base_addr
, 0x7F, queue_size_bytes
);
443 queue_base
= BUILD_RING_BASE_ADDR(queue
->base_phys_addr
,
446 io_addr
= pci_dev
->mem_resource
[0].addr
;
448 WRITE_CSR_RING_BASE(io_addr
, queue
->hw_bundle_number
,
449 queue
->hw_queue_number
, queue_base
);
451 QAT_LOG(DEBUG
, "RING: Name:%s, size in CSR: %u, in bytes %u,"
452 " nb msgs %u, msg_size %u, modulo mask %u",
454 queue
->queue_size
, queue_size_bytes
,
455 qp_conf
->nb_descriptors
, desc_size
,
461 rte_memzone_free(qp_mz
);
465 static int qat_qp_check_queue_alignment(uint64_t phys_addr
,
466 uint32_t queue_size_bytes
)
468 if (((queue_size_bytes
- 1) & phys_addr
) != 0)
473 static int adf_verify_queue_size(uint32_t msg_size
, uint32_t msg_num
,
474 uint32_t *p_queue_size_for_csr
)
476 uint8_t i
= ADF_MIN_RING_SIZE
;
478 for (; i
<= ADF_MAX_RING_SIZE
; i
++)
479 if ((msg_size
* msg_num
) ==
480 (uint32_t)ADF_SIZE_TO_RING_SIZE_IN_BYTES(i
)) {
481 *p_queue_size_for_csr
= i
;
484 QAT_LOG(ERR
, "Invalid ring size %d", msg_size
* msg_num
);
488 static void adf_queue_arb_enable(struct qat_queue
*txq
, void *base_addr
,
489 rte_spinlock_t
*lock
)
491 uint32_t arb_csr_offset
= ADF_ARB_RINGSRVARBEN_OFFSET
+
493 txq
->hw_bundle_number
);
496 rte_spinlock_lock(lock
);
497 value
= ADF_CSR_RD(base_addr
, arb_csr_offset
);
498 value
|= (0x01 << txq
->hw_queue_number
);
499 ADF_CSR_WR(base_addr
, arb_csr_offset
, value
);
500 rte_spinlock_unlock(lock
);
503 static void adf_queue_arb_disable(struct qat_queue
*txq
, void *base_addr
,
504 rte_spinlock_t
*lock
)
506 uint32_t arb_csr_offset
= ADF_ARB_RINGSRVARBEN_OFFSET
+
508 txq
->hw_bundle_number
);
511 rte_spinlock_lock(lock
);
512 value
= ADF_CSR_RD(base_addr
, arb_csr_offset
);
513 value
&= ~(0x01 << txq
->hw_queue_number
);
514 ADF_CSR_WR(base_addr
, arb_csr_offset
, value
);
515 rte_spinlock_unlock(lock
);
518 static void adf_configure_queues(struct qat_qp
*qp
)
520 uint32_t queue_config
;
521 struct qat_queue
*queue
= &qp
->tx_q
;
523 queue_config
= BUILD_RING_CONFIG(queue
->queue_size
);
525 WRITE_CSR_RING_CONFIG(qp
->mmap_bar_addr
, queue
->hw_bundle_number
,
526 queue
->hw_queue_number
, queue_config
);
530 BUILD_RESP_RING_CONFIG(queue
->queue_size
,
531 ADF_RING_NEAR_WATERMARK_512
,
532 ADF_RING_NEAR_WATERMARK_0
);
534 WRITE_CSR_RING_CONFIG(qp
->mmap_bar_addr
, queue
->hw_bundle_number
,
535 queue
->hw_queue_number
, queue_config
);
538 static inline uint32_t adf_modulo(uint32_t data
, uint32_t modulo_mask
)
540 return data
& modulo_mask
;
544 txq_write_tail(struct qat_qp
*qp
, struct qat_queue
*q
) {
545 WRITE_CSR_RING_TAIL(qp
->mmap_bar_addr
, q
->hw_bundle_number
,
546 q
->hw_queue_number
, q
->tail
);
547 q
->csr_tail
= q
->tail
;
551 void rxq_free_desc(struct qat_qp
*qp
, struct qat_queue
*q
)
553 uint32_t old_head
, new_head
;
556 old_head
= q
->csr_head
;
558 max_head
= qp
->nb_descriptors
* q
->msg_size
;
560 /* write out free descriptors */
561 void *cur_desc
= (uint8_t *)q
->base_addr
+ old_head
;
563 if (new_head
< old_head
) {
564 memset(cur_desc
, ADF_RING_EMPTY_SIG_BYTE
, max_head
- old_head
);
565 memset(q
->base_addr
, ADF_RING_EMPTY_SIG_BYTE
, new_head
);
567 memset(cur_desc
, ADF_RING_EMPTY_SIG_BYTE
, new_head
- old_head
);
569 q
->nb_processed_responses
= 0;
570 q
->csr_head
= new_head
;
572 /* write current head to CSR */
573 WRITE_CSR_RING_HEAD(qp
->mmap_bar_addr
, q
->hw_bundle_number
,
574 q
->hw_queue_number
, new_head
);
578 qat_enqueue_op_burst(void *qp
, void **ops
, uint16_t nb_ops
)
580 register struct qat_queue
*queue
;
581 struct qat_qp
*tmp_qp
= (struct qat_qp
*)qp
;
582 register uint32_t nb_ops_sent
= 0;
584 uint16_t nb_ops_possible
= nb_ops
;
585 register uint8_t *base_addr
;
586 register uint32_t tail
;
588 if (unlikely(nb_ops
== 0))
591 /* read params used a lot in main loop into registers */
592 queue
= &(tmp_qp
->tx_q
);
593 base_addr
= (uint8_t *)queue
->base_addr
;
596 /* Find how many can actually fit on the ring */
598 /* dequeued can only be written by one thread, but it may not
599 * be this thread. As it's 4-byte aligned it will be read
600 * atomically here by any Intel CPU.
601 * enqueued can wrap before dequeued, but cannot
602 * lap it as var size of enq/deq (uint32_t) > var size of
603 * max_inflights (uint16_t). In reality inflights is never
604 * even as big as max uint16_t, as it's <= ADF_MAX_DESC.
605 * On wrapping, the calculation still returns the correct
606 * positive value as all three vars are unsigned.
609 tmp_qp
->enqueued
- tmp_qp
->dequeued
;
611 if ((inflights
+ nb_ops
) > tmp_qp
->max_inflights
) {
612 nb_ops_possible
= tmp_qp
->max_inflights
- inflights
;
613 if (nb_ops_possible
== 0)
616 /* QAT has plenty of work queued already, so don't waste cycles
617 * enqueueing, wait til the application has gathered a bigger
618 * burst or some completed ops have been dequeued
620 if (tmp_qp
->min_enq_burst_threshold
&& inflights
>
621 QAT_QP_MIN_INFL_THRESHOLD
&& nb_ops_possible
<
622 tmp_qp
->min_enq_burst_threshold
) {
623 tmp_qp
->stats
.threshold_hit_count
++;
629 while (nb_ops_sent
!= nb_ops_possible
) {
630 ret
= tmp_qp
->build_request(*ops
, base_addr
+ tail
,
631 tmp_qp
->op_cookies
[tail
>> queue
->trailz
],
632 tmp_qp
->qat_dev_gen
);
634 tmp_qp
->stats
.enqueue_err_count
++;
635 /* This message cannot be enqueued */
636 if (nb_ops_sent
== 0)
641 tail
= adf_modulo(tail
+ queue
->msg_size
, queue
->modulo_mask
);
647 tmp_qp
->enqueued
+= nb_ops_sent
;
648 tmp_qp
->stats
.enqueued_count
+= nb_ops_sent
;
649 txq_write_tail(tmp_qp
, queue
);
653 /* Use this for compression only - but keep consistent with above common
654 * function as much as possible.
657 qat_enqueue_comp_op_burst(void *qp
, void **ops
, uint16_t nb_ops
)
659 register struct qat_queue
*queue
;
660 struct qat_qp
*tmp_qp
= (struct qat_qp
*)qp
;
661 register uint32_t nb_ops_sent
= 0;
662 register int nb_desc_to_build
;
663 uint16_t nb_ops_possible
= nb_ops
;
664 register uint8_t *base_addr
;
665 register uint32_t tail
;
667 int descriptors_built
, total_descriptors_built
= 0;
668 int nb_remaining_descriptors
;
671 if (unlikely(nb_ops
== 0))
674 /* read params used a lot in main loop into registers */
675 queue
= &(tmp_qp
->tx_q
);
676 base_addr
= (uint8_t *)queue
->base_addr
;
679 /* Find how many can actually fit on the ring */
681 /* dequeued can only be written by one thread, but it may not
682 * be this thread. As it's 4-byte aligned it will be read
683 * atomically here by any Intel CPU.
684 * enqueued can wrap before dequeued, but cannot
685 * lap it as var size of enq/deq (uint32_t) > var size of
686 * max_inflights (uint16_t). In reality inflights is never
687 * even as big as max uint16_t, as it's <= ADF_MAX_DESC.
688 * On wrapping, the calculation still returns the correct
689 * positive value as all three vars are unsigned.
692 tmp_qp
->enqueued
- tmp_qp
->dequeued
;
694 /* Find how many can actually fit on the ring */
695 overflow
= (inflights
+ nb_ops
) - tmp_qp
->max_inflights
;
697 nb_ops_possible
= nb_ops
- overflow
;
698 if (nb_ops_possible
== 0)
702 /* QAT has plenty of work queued already, so don't waste cycles
703 * enqueueing, wait til the application has gathered a bigger
704 * burst or some completed ops have been dequeued
706 if (tmp_qp
->min_enq_burst_threshold
&& inflights
>
707 QAT_QP_MIN_INFL_THRESHOLD
&& nb_ops_possible
<
708 tmp_qp
->min_enq_burst_threshold
) {
709 tmp_qp
->stats
.threshold_hit_count
++;
714 /* At this point nb_ops_possible is assuming a 1:1 mapping
715 * between ops and descriptors.
716 * Fewer may be sent if some ops have to be split.
717 * nb_ops_possible is <= burst size.
718 * Find out how many spaces are actually available on the qp in case
721 nb_remaining_descriptors
= nb_ops_possible
722 + ((overflow
>= 0) ? 0 : overflow
* (-1));
723 QAT_DP_LOG(DEBUG
, "Nb ops requested %d, nb descriptors remaining %d",
724 nb_ops
, nb_remaining_descriptors
);
726 while (nb_ops_sent
!= nb_ops_possible
&&
727 nb_remaining_descriptors
> 0) {
728 struct qat_comp_op_cookie
*cookie
=
729 tmp_qp
->op_cookies
[tail
>> queue
->trailz
];
731 descriptors_built
= 0;
733 QAT_DP_LOG(DEBUG
, "--- data length: %u",
734 ((struct rte_comp_op
*)*ops
)->src
.length
);
736 nb_desc_to_build
= qat_comp_build_request(*ops
,
737 base_addr
+ tail
, cookie
, tmp_qp
->qat_dev_gen
);
738 QAT_DP_LOG(DEBUG
, "%d descriptors built, %d remaining, "
739 "%d ops sent, %d descriptors needed",
740 total_descriptors_built
, nb_remaining_descriptors
,
741 nb_ops_sent
, nb_desc_to_build
);
743 if (unlikely(nb_desc_to_build
< 0)) {
744 /* this message cannot be enqueued */
745 tmp_qp
->stats
.enqueue_err_count
++;
746 if (nb_ops_sent
== 0)
749 } else if (unlikely(nb_desc_to_build
> 1)) {
750 /* this op is too big and must be split - get more
751 * descriptors and retry
754 QAT_DP_LOG(DEBUG
, "Build %d descriptors for this op",
757 nb_remaining_descriptors
-= nb_desc_to_build
;
758 if (nb_remaining_descriptors
>= 0) {
759 /* There are enough remaining descriptors
762 int ret2
= qat_comp_build_multiple_requests(
766 if (unlikely(ret2
< 1)) {
768 "Failed to build (%d) descriptors, status %d",
769 nb_desc_to_build
, ret2
);
771 qat_comp_free_split_op_memzones(cookie
,
772 nb_desc_to_build
- 1);
774 tmp_qp
->stats
.enqueue_err_count
++;
776 /* This message cannot be enqueued */
777 if (nb_ops_sent
== 0)
781 descriptors_built
= ret2
;
782 total_descriptors_built
+=
784 nb_remaining_descriptors
-=
787 "Multiple descriptors (%d) built ok",
791 QAT_DP_LOG(ERR
, "For the current op, number of requested descriptors (%d) "
792 "exceeds number of available descriptors (%d)",
794 nb_remaining_descriptors
+
797 qat_comp_free_split_op_memzones(cookie
,
798 nb_desc_to_build
- 1);
800 /* Not enough extra descriptors */
801 if (nb_ops_sent
== 0)
806 descriptors_built
= 1;
807 total_descriptors_built
++;
808 nb_remaining_descriptors
--;
809 QAT_DP_LOG(DEBUG
, "Single descriptor built ok");
812 tail
= adf_modulo(tail
+ (queue
->msg_size
* descriptors_built
),
820 tmp_qp
->enqueued
+= total_descriptors_built
;
821 tmp_qp
->stats
.enqueued_count
+= nb_ops_sent
;
822 txq_write_tail(tmp_qp
, queue
);
827 qat_dequeue_op_burst(void *qp
, void **ops
, uint16_t nb_ops
)
829 struct qat_queue
*rx_queue
;
830 struct qat_qp
*tmp_qp
= (struct qat_qp
*)qp
;
832 uint32_t op_resp_counter
= 0, fw_resp_counter
= 0;
836 rx_queue
= &(tmp_qp
->rx_q
);
837 head
= rx_queue
->head
;
838 resp_msg
= (uint8_t *)rx_queue
->base_addr
+ rx_queue
->head
;
840 while (*(uint32_t *)resp_msg
!= ADF_RING_EMPTY_SIG
&&
841 op_resp_counter
!= nb_ops
) {
845 if (tmp_qp
->service_type
== QAT_SERVICE_SYMMETRIC
)
846 qat_sym_process_response(ops
, resp_msg
);
847 else if (tmp_qp
->service_type
== QAT_SERVICE_COMPRESSION
)
848 nb_fw_responses
= qat_comp_process_response(
850 tmp_qp
->op_cookies
[head
>> rx_queue
->trailz
],
851 &tmp_qp
->stats
.dequeue_err_count
);
852 #ifdef BUILD_QAT_ASYM
853 else if (tmp_qp
->service_type
== QAT_SERVICE_ASYMMETRIC
)
854 qat_asym_process_response(ops
, resp_msg
,
855 tmp_qp
->op_cookies
[head
>> rx_queue
->trailz
]);
858 head
= adf_modulo(head
+ rx_queue
->msg_size
,
859 rx_queue
->modulo_mask
);
861 resp_msg
= (uint8_t *)rx_queue
->base_addr
+ head
;
863 if (nb_fw_responses
) {
864 /* only move on to next op if one was ready to return
871 /* A compression op may be broken up into multiple fw requests.
872 * Only count fw responses as complete once ALL the responses
873 * associated with an op have been processed, as the cookie
874 * data from the first response must be available until
875 * finished with all firmware responses.
877 fw_resp_counter
+= nb_fw_responses
;
879 rx_queue
->nb_processed_responses
++;
882 tmp_qp
->dequeued
+= fw_resp_counter
;
883 tmp_qp
->stats
.dequeued_count
+= op_resp_counter
;
885 rx_queue
->head
= head
;
886 if (rx_queue
->nb_processed_responses
> QAT_CSR_HEAD_WRITE_THRESH
)
887 rxq_free_desc(tmp_qp
, rx_queue
);
889 QAT_DP_LOG(DEBUG
, "Dequeue burst return: %u, QAT responses: %u",
890 op_resp_counter
, fw_resp_counter
);
892 return op_resp_counter
;
895 /* This is almost same as dequeue_op_burst, without the atomic, without stats
896 * and without the op. Dequeues one response.
899 qat_cq_dequeue_response(struct qat_qp
*qp
, void *out_data
)
903 struct qat_queue
*queue
= &(qp
->rx_q
);
904 struct icp_qat_fw_comn_resp
*resp_msg
= (struct icp_qat_fw_comn_resp
*)
905 ((uint8_t *)queue
->base_addr
+ queue
->head
);
907 while (retries
++ < QAT_CQ_MAX_DEQ_RETRIES
&&
908 *(uint32_t *)resp_msg
== ADF_RING_EMPTY_SIG
) {
909 /* loop waiting for response until we reach the timeout */
913 if (*(uint32_t *)resp_msg
!= ADF_RING_EMPTY_SIG
) {
914 /* response received */
917 /* check status flag */
918 if (ICP_QAT_FW_COMN_RESP_CRYPTO_STAT_GET(
919 resp_msg
->comn_hdr
.comn_status
) ==
920 ICP_QAT_FW_COMN_STATUS_FLAG_OK
) {
922 memcpy(out_data
, resp_msg
, queue
->msg_size
);
924 memset(out_data
, 0, queue
->msg_size
);
927 queue
->head
= adf_modulo(queue
->head
+ queue
->msg_size
,
929 rxq_free_desc(qp
, queue
);
935 /* Sends a NULL message and extracts QAT fw version from the response.
936 * Used to determine detailed capabilities based on the fw version number.
937 * This assumes that there are no inflight messages, i.e. assumes there's space
938 * on the qp, one message is sent and only one response collected.
939 * Returns fw version number or 0 for unknown version or a negative error code.
942 qat_cq_get_fw_version(struct qat_qp
*qp
)
944 struct qat_queue
*queue
= &(qp
->tx_q
);
945 uint8_t *base_addr
= (uint8_t *)queue
->base_addr
;
946 struct icp_qat_fw_comn_req null_msg
;
947 struct icp_qat_fw_comn_resp response
;
949 /* prepare the NULL request */
950 memset(&null_msg
, 0, sizeof(null_msg
));
951 null_msg
.comn_hdr
.hdr_flags
=
952 ICP_QAT_FW_COMN_HDR_FLAGS_BUILD(ICP_QAT_FW_COMN_REQ_FLAG_SET
);
953 null_msg
.comn_hdr
.service_type
= ICP_QAT_FW_COMN_REQ_NULL
;
954 null_msg
.comn_hdr
.service_cmd_id
= ICP_QAT_FW_NULL_REQ_SERV_ID
;
956 #if RTE_LOG_DP_LEVEL >= RTE_LOG_DEBUG
957 QAT_DP_HEXDUMP_LOG(DEBUG
, "NULL request", &null_msg
, sizeof(null_msg
));
960 /* send the NULL request */
961 memcpy(base_addr
+ queue
->tail
, &null_msg
, sizeof(null_msg
));
962 queue
->tail
= adf_modulo(queue
->tail
+ queue
->msg_size
,
964 txq_write_tail(qp
, queue
);
966 /* receive a response */
967 if (qat_cq_dequeue_response(qp
, &response
)) {
969 #if RTE_LOG_DP_LEVEL >= RTE_LOG_DEBUG
970 QAT_DP_HEXDUMP_LOG(DEBUG
, "NULL response:", &response
,
973 /* if LW0 bit 24 is set - then the fw version was returned */
974 if (QAT_FIELD_GET(response
.comn_hdr
.hdr_flags
,
975 ICP_QAT_FW_COMN_NULL_VERSION_FLAG_BITPOS
,
976 ICP_QAT_FW_COMN_NULL_VERSION_FLAG_MASK
))
977 return response
.resrvd
[0]; /* return LW4 */
979 return 0; /* not set - we don't know fw version */
982 QAT_LOG(ERR
, "No response received");
987 qat_comp_process_response(void **op __rte_unused
, uint8_t *resp __rte_unused
,
988 void *op_cookie __rte_unused
,
989 uint64_t *dequeue_err_count __rte_unused
)