*/
#include <rte_common.h>
+#include <rte_cycles.h>
#include <rte_dev.h>
#include <rte_malloc.h>
#include <rte_memzone.h>
#include "qat_device.h"
#include "qat_qp.h"
#include "qat_sym.h"
+#include "qat_asym.h"
#include "qat_comp.h"
#include "adf_transport_access_macros.h"
+#define QAT_CQ_MAX_DEQ_RETRIES 10
#define ADF_MAX_DESC 4096
#define ADF_MIN_DESC 128
}
/* Allocate the queue pair data structure. */
- qp = rte_zmalloc("qat PMD qp metadata",
- sizeof(*qp), RTE_CACHE_LINE_SIZE);
+ qp = rte_zmalloc_socket("qat PMD qp metadata",
+ sizeof(*qp), RTE_CACHE_LINE_SIZE,
+ qat_qp_conf->socket_id);
if (qp == NULL) {
QAT_LOG(ERR, "Failed to alloc mem for qp struct");
return -ENOMEM;
}
qp->nb_descriptors = qat_qp_conf->nb_descriptors;
- qp->op_cookies = rte_zmalloc("qat PMD op cookie pointer",
+ qp->op_cookies = rte_zmalloc_socket("qat PMD op cookie pointer",
qat_qp_conf->nb_descriptors * sizeof(*qp->op_cookies),
- RTE_CACHE_LINE_SIZE);
+ RTE_CACHE_LINE_SIZE, qat_qp_conf->socket_id);
if (qp->op_cookies == NULL) {
QAT_LOG(ERR, "Failed to alloc mem for cookie");
rte_free(qp);
}
qp->mmap_bar_addr = pci_dev->mem_resource[0].addr;
- qp->inflights16 = 0;
+ qp->enqueued = qp->dequeued = 0;
if (qat_queue_create(qat_dev, &(qp->tx_q), qat_qp_conf,
ADF_RING_DIR_TX) != 0) {
goto create_err;
}
+ qp->max_inflights = ADF_MAX_INFLIGHTS(qp->tx_q.queue_size,
+ ADF_BYTES_TO_MSG_SIZE(qp->tx_q.msg_size));
+
+ if (qp->max_inflights < 2) {
+ QAT_LOG(ERR, "Invalid num inflights");
+ qat_queue_delete(&(qp->tx_q));
+ goto create_err;
+ }
+
if (qat_queue_create(qat_dev, &(qp->rx_q), qat_qp_conf,
ADF_RING_DIR_RX) != 0) {
QAT_LOG(ERR, "Rx queue create failed "
qp->op_cookie_pool = rte_mempool_create(op_cookie_pool_name,
qp->nb_descriptors,
qat_qp_conf->cookie_size, 64, 0,
- NULL, NULL, NULL, NULL, qat_qp_conf->socket_id,
+ NULL, NULL, NULL, NULL,
+ qat_dev->pci_dev->device.numa_node,
0);
if (!qp->op_cookie_pool) {
QAT_LOG(ERR, "QAT PMD Cannot create"
QAT_LOG(ERR, "QAT PMD Cannot get op_cookie");
goto create_err;
}
+ memset(qp->op_cookies[i], 0, qat_qp_conf->cookie_size);
}
qp->qat_dev_gen = qat_dev->qat_dev_gen;
qp->qat_dev->qat_dev_id);
/* Don't free memory if there are still responses to be processed */
- if (qp->inflights16 == 0) {
+ if ((qp->enqueued - qp->dequeued) == 0) {
qat_queue_delete(&(qp->tx_q));
qat_queue_delete(&(qp->rx_q));
} else {
qp_conf->service_str, "qp_mem",
queue->hw_bundle_number, queue->hw_queue_number);
qp_mz = queue_dma_zone_reserve(queue->memz_name, queue_size_bytes,
- qp_conf->socket_id);
+ qat_dev->pci_dev->device.numa_node);
if (qp_mz == NULL) {
QAT_LOG(ERR, "Failed to allocate ring memzone");
return -ENOMEM;
goto queue_create_err;
}
- queue->max_inflights = ADF_MAX_INFLIGHTS(queue->queue_size,
- ADF_BYTES_TO_MSG_SIZE(desc_size));
queue->modulo_mask = (1 << ADF_RING_SIZE_MODULO(queue->queue_size)) - 1;
-
- if (queue->max_inflights < 2) {
- QAT_LOG(ERR, "Invalid num inflights");
- ret = -EINVAL;
- goto queue_create_err;
- }
queue->head = 0;
queue->tail = 0;
queue->msg_size = desc_size;
+ /* For fast calculation of cookie index, relies on msg_size being 2^n */
+ queue->trailz = __builtin_ctz(desc_size);
+
/*
* Write an unused pattern to the queue memory.
*/
queue->hw_queue_number, queue_base);
QAT_LOG(DEBUG, "RING: Name:%s, size in CSR: %u, in bytes %u,"
- " nb msgs %u, msg_size %u, max_inflights %u modulo mask %u",
+ " nb msgs %u, msg_size %u, modulo mask %u",
queue->memz_name,
queue->queue_size, queue_size_bytes,
qp_conf->nb_descriptors, desc_size,
- queue->max_inflights, queue->modulo_mask);
+ queue->modulo_mask);
return 0;
txq_write_tail(struct qat_qp *qp, struct qat_queue *q) {
WRITE_CSR_RING_TAIL(qp->mmap_bar_addr, q->hw_bundle_number,
q->hw_queue_number, q->tail);
- q->nb_pending_requests = 0;
q->csr_tail = q->tail;
}
uint16_t nb_ops_possible = nb_ops;
register uint8_t *base_addr;
register uint32_t tail;
- int overflow;
if (unlikely(nb_ops == 0))
return 0;
tail = queue->tail;
/* Find how many can actually fit on the ring */
- tmp_qp->inflights16 += nb_ops;
- overflow = tmp_qp->inflights16 - queue->max_inflights;
- if (overflow > 0) {
- tmp_qp->inflights16 -= overflow;
- nb_ops_possible = nb_ops - overflow;
- if (nb_ops_possible == 0)
+ {
+ /* dequeued can only be written by one thread, but it may not
+ * be this thread. As it's 4-byte aligned it will be read
+ * atomically here by any Intel CPU.
+ * enqueued can wrap before dequeued, but cannot
+ * lap it as var size of enq/deq (uint32_t) > var size of
+ * max_inflights (uint16_t). In reality inflights is never
+ * even as big as max uint16_t, as it's <= ADF_MAX_DESC.
+ * On wrapping, the calculation still returns the correct
+ * positive value as all three vars are unsigned.
+ */
+ uint32_t inflights =
+ tmp_qp->enqueued - tmp_qp->dequeued;
+
+ if ((inflights + nb_ops) > tmp_qp->max_inflights) {
+ nb_ops_possible = tmp_qp->max_inflights - inflights;
+ if (nb_ops_possible == 0)
+ return 0;
+ }
+ /* QAT has plenty of work queued already, so don't waste cycles
+ * enqueueing, wait til the application has gathered a bigger
+ * burst or some completed ops have been dequeued
+ */
+ if (tmp_qp->min_enq_burst_threshold && inflights >
+ QAT_QP_MIN_INFL_THRESHOLD && nb_ops_possible <
+ tmp_qp->min_enq_burst_threshold) {
+ tmp_qp->stats.threshold_hit_count++;
return 0;
+ }
}
+
while (nb_ops_sent != nb_ops_possible) {
ret = tmp_qp->build_request(*ops, base_addr + tail,
- tmp_qp->op_cookies[tail / queue->msg_size],
+ tmp_qp->op_cookies[tail >> queue->trailz],
tmp_qp->qat_dev_gen);
if (ret != 0) {
tmp_qp->stats.enqueue_err_count++;
- /*
- * This message cannot be enqueued,
- * decrease number of ops that wasn't sent
- */
- tmp_qp->inflights16 -= nb_ops_possible - nb_ops_sent;
+ /* This message cannot be enqueued */
if (nb_ops_sent == 0)
return 0;
goto kick_tail;
}
kick_tail:
queue->tail = tail;
+ tmp_qp->enqueued += nb_ops_sent;
tmp_qp->stats.enqueued_count += nb_ops_sent;
- queue->nb_pending_requests += nb_ops_sent;
- if (tmp_qp->inflights16 < QAT_CSR_TAIL_FORCE_WRITE_THRESH ||
- queue->nb_pending_requests > QAT_CSR_TAIL_WRITE_THRESH) {
- txq_write_tail(tmp_qp, queue);
+ txq_write_tail(tmp_qp, queue);
+ return nb_ops_sent;
+}
+
+/* Use this for compression only - but keep consistent with above common
+ * function as much as possible.
+ */
+uint16_t
+qat_enqueue_comp_op_burst(void *qp, void **ops, uint16_t nb_ops)
+{
+ register struct qat_queue *queue;
+ struct qat_qp *tmp_qp = (struct qat_qp *)qp;
+ register uint32_t nb_ops_sent = 0;
+ register int nb_desc_to_build;
+ uint16_t nb_ops_possible = nb_ops;
+ register uint8_t *base_addr;
+ register uint32_t tail;
+
+ int descriptors_built, total_descriptors_built = 0;
+ int nb_remaining_descriptors;
+ int overflow = 0;
+
+ if (unlikely(nb_ops == 0))
+ return 0;
+
+ /* read params used a lot in main loop into registers */
+ queue = &(tmp_qp->tx_q);
+ base_addr = (uint8_t *)queue->base_addr;
+ tail = queue->tail;
+
+ /* Find how many can actually fit on the ring */
+ {
+ /* dequeued can only be written by one thread, but it may not
+ * be this thread. As it's 4-byte aligned it will be read
+ * atomically here by any Intel CPU.
+ * enqueued can wrap before dequeued, but cannot
+ * lap it as var size of enq/deq (uint32_t) > var size of
+ * max_inflights (uint16_t). In reality inflights is never
+ * even as big as max uint16_t, as it's <= ADF_MAX_DESC.
+ * On wrapping, the calculation still returns the correct
+ * positive value as all three vars are unsigned.
+ */
+ uint32_t inflights =
+ tmp_qp->enqueued - tmp_qp->dequeued;
+
+ /* Find how many can actually fit on the ring */
+ overflow = (inflights + nb_ops) - tmp_qp->max_inflights;
+ if (overflow > 0) {
+ nb_ops_possible = nb_ops - overflow;
+ if (nb_ops_possible == 0)
+ return 0;
+ }
+
+ /* QAT has plenty of work queued already, so don't waste cycles
+ * enqueueing, wait til the application has gathered a bigger
+ * burst or some completed ops have been dequeued
+ */
+ if (tmp_qp->min_enq_burst_threshold && inflights >
+ QAT_QP_MIN_INFL_THRESHOLD && nb_ops_possible <
+ tmp_qp->min_enq_burst_threshold) {
+ tmp_qp->stats.threshold_hit_count++;
+ return 0;
+ }
+ }
+
+ /* At this point nb_ops_possible is assuming a 1:1 mapping
+ * between ops and descriptors.
+ * Fewer may be sent if some ops have to be split.
+ * nb_ops_possible is <= burst size.
+ * Find out how many spaces are actually available on the qp in case
+ * more are needed.
+ */
+ nb_remaining_descriptors = nb_ops_possible
+ + ((overflow >= 0) ? 0 : overflow * (-1));
+ QAT_DP_LOG(DEBUG, "Nb ops requested %d, nb descriptors remaining %d",
+ nb_ops, nb_remaining_descriptors);
+
+ while (nb_ops_sent != nb_ops_possible &&
+ nb_remaining_descriptors > 0) {
+ struct qat_comp_op_cookie *cookie =
+ tmp_qp->op_cookies[tail >> queue->trailz];
+
+ descriptors_built = 0;
+
+ QAT_DP_LOG(DEBUG, "--- data length: %u",
+ ((struct rte_comp_op *)*ops)->src.length);
+
+ nb_desc_to_build = qat_comp_build_request(*ops,
+ base_addr + tail, cookie, tmp_qp->qat_dev_gen);
+ QAT_DP_LOG(DEBUG, "%d descriptors built, %d remaining, "
+ "%d ops sent, %d descriptors needed",
+ total_descriptors_built, nb_remaining_descriptors,
+ nb_ops_sent, nb_desc_to_build);
+
+ if (unlikely(nb_desc_to_build < 0)) {
+ /* this message cannot be enqueued */
+ tmp_qp->stats.enqueue_err_count++;
+ if (nb_ops_sent == 0)
+ return 0;
+ goto kick_tail;
+ } else if (unlikely(nb_desc_to_build > 1)) {
+ /* this op is too big and must be split - get more
+ * descriptors and retry
+ */
+
+ QAT_DP_LOG(DEBUG, "Build %d descriptors for this op",
+ nb_desc_to_build);
+
+ nb_remaining_descriptors -= nb_desc_to_build;
+ if (nb_remaining_descriptors >= 0) {
+ /* There are enough remaining descriptors
+ * so retry
+ */
+ int ret2 = qat_comp_build_multiple_requests(
+ *ops, tmp_qp, tail,
+ nb_desc_to_build);
+
+ if (unlikely(ret2 < 1)) {
+ QAT_DP_LOG(DEBUG,
+ "Failed to build (%d) descriptors, status %d",
+ nb_desc_to_build, ret2);
+
+ qat_comp_free_split_op_memzones(cookie,
+ nb_desc_to_build - 1);
+
+ tmp_qp->stats.enqueue_err_count++;
+
+ /* This message cannot be enqueued */
+ if (nb_ops_sent == 0)
+ return 0;
+ goto kick_tail;
+ } else {
+ descriptors_built = ret2;
+ total_descriptors_built +=
+ descriptors_built;
+ nb_remaining_descriptors -=
+ descriptors_built;
+ QAT_DP_LOG(DEBUG,
+ "Multiple descriptors (%d) built ok",
+ descriptors_built);
+ }
+ } else {
+ QAT_DP_LOG(ERR, "For the current op, number of requested descriptors (%d) "
+ "exceeds number of available descriptors (%d)",
+ nb_desc_to_build,
+ nb_remaining_descriptors +
+ nb_desc_to_build);
+
+ qat_comp_free_split_op_memzones(cookie,
+ nb_desc_to_build - 1);
+
+ /* Not enough extra descriptors */
+ if (nb_ops_sent == 0)
+ return 0;
+ goto kick_tail;
+ }
+ } else {
+ descriptors_built = 1;
+ total_descriptors_built++;
+ nb_remaining_descriptors--;
+ QAT_DP_LOG(DEBUG, "Single descriptor built ok");
+ }
+
+ tail = adf_modulo(tail + (queue->msg_size * descriptors_built),
+ queue->modulo_mask);
+ ops++;
+ nb_ops_sent++;
}
+
+kick_tail:
+ queue->tail = tail;
+ tmp_qp->enqueued += total_descriptors_built;
+ tmp_qp->stats.enqueued_count += total_descriptors_built;
+ txq_write_tail(tmp_qp, queue);
return nb_ops_sent;
}
uint16_t
qat_dequeue_op_burst(void *qp, void **ops, uint16_t nb_ops)
{
- struct qat_queue *rx_queue, *tx_queue;
+ struct qat_queue *rx_queue;
struct qat_qp *tmp_qp = (struct qat_qp *)qp;
uint32_t head;
- uint32_t resp_counter = 0;
+ uint32_t op_resp_counter = 0, fw_resp_counter = 0;
uint8_t *resp_msg;
- uint8_t hdr_flags;
+ int nb_fw_responses = 0;
rx_queue = &(tmp_qp->rx_q);
- tx_queue = &(tmp_qp->tx_q);
head = rx_queue->head;
resp_msg = (uint8_t *)rx_queue->base_addr + rx_queue->head;
- hdr_flags = ((struct icp_qat_fw_comn_resp_hdr *)resp_msg)->hdr_flags;
while (*(uint32_t *)resp_msg != ADF_RING_EMPTY_SIG &&
- resp_counter != nb_ops) {
+ op_resp_counter != nb_ops) {
- if (unlikely(!ICP_QAT_FW_COMN_VALID_FLAG_GET(hdr_flags))) {
- /* Fatal firmware error */
- QAT_LOG(ERR, "QAT Firmware returned invalid response");
- return 0;
- }
-
- if (tmp_qp->service_type == QAT_SERVICE_SYMMETRIC)
+ nb_fw_responses = 0;
+ if (tmp_qp->service_type == QAT_SERVICE_SYMMETRIC) {
qat_sym_process_response(ops, resp_msg);
- else if (tmp_qp->service_type == QAT_SERVICE_COMPRESSION)
- qat_comp_process_response(ops, resp_msg);
+ nb_fw_responses = 1;
+ } else if (tmp_qp->service_type == QAT_SERVICE_COMPRESSION)
+
+ nb_fw_responses = qat_comp_process_response(
+ ops, resp_msg,
+ tmp_qp->op_cookies[head >> rx_queue->trailz],
+ &tmp_qp->stats.dequeue_err_count);
+
+ else if (tmp_qp->service_type == QAT_SERVICE_ASYMMETRIC) {
+#ifdef BUILD_QAT_ASYM
+ qat_asym_process_response(ops, resp_msg,
+ tmp_qp->op_cookies[head >> rx_queue->trailz]);
+ nb_fw_responses = 1;
+#endif
+ }
head = adf_modulo(head + rx_queue->msg_size,
rx_queue->modulo_mask);
resp_msg = (uint8_t *)rx_queue->base_addr + head;
- ops++;
- resp_counter++;
+
+ if (ops != NULL && nb_fw_responses) {
+ /* only move on to next op if one was ready to return
+ * to API
+ */
+ ops++;
+ op_resp_counter++;
+ }
+
+ /* A compression op may be broken up into multiple fw requests.
+ * Only count fw responses as complete once ALL the responses
+ * associated with an op have been processed, as the cookie
+ * data from the first response must be available until
+ * finished with all firmware responses.
+ */
+ fw_resp_counter += nb_fw_responses;
}
- if (resp_counter > 0) {
+
+ if (fw_resp_counter > 0) {
rx_queue->head = head;
- tmp_qp->stats.dequeued_count += resp_counter;
- rx_queue->nb_processed_responses += resp_counter;
- tmp_qp->inflights16 -= resp_counter;
+ tmp_qp->dequeued += fw_resp_counter;
+ tmp_qp->stats.dequeued_count += fw_resp_counter;
+ rx_queue->nb_processed_responses += fw_resp_counter;
if (rx_queue->nb_processed_responses >
- QAT_CSR_HEAD_WRITE_THRESH)
+ QAT_CSR_HEAD_WRITE_THRESH)
rxq_free_desc(tmp_qp, rx_queue);
}
- /* also check if tail needs to be advanced */
- if (tmp_qp->inflights16 <= QAT_CSR_TAIL_FORCE_WRITE_THRESH &&
- tx_queue->tail != tx_queue->csr_tail) {
- txq_write_tail(tmp_qp, tx_queue);
+ QAT_DP_LOG(DEBUG, "Dequeue burst return: %u, QAT responses: %u",
+ op_resp_counter, fw_resp_counter);
+
+ return op_resp_counter;
+}
+
+/* This is almost same as dequeue_op_burst, without the atomic, without stats
+ * and without the op. Dequeues one response.
+ */
+static uint8_t
+qat_cq_dequeue_response(struct qat_qp *qp, void *out_data)
+{
+ uint8_t result = 0;
+ uint8_t retries = 0;
+ struct qat_queue *queue = &(qp->rx_q);
+ struct icp_qat_fw_comn_resp *resp_msg = (struct icp_qat_fw_comn_resp *)
+ ((uint8_t *)queue->base_addr + queue->head);
+
+ while (retries++ < QAT_CQ_MAX_DEQ_RETRIES &&
+ *(uint32_t *)resp_msg == ADF_RING_EMPTY_SIG) {
+ /* loop waiting for response until we reach the timeout */
+ rte_delay_ms(20);
+ }
+
+ if (*(uint32_t *)resp_msg != ADF_RING_EMPTY_SIG) {
+ /* response received */
+ result = 1;
+
+ /* check status flag */
+ if (ICP_QAT_FW_COMN_RESP_CRYPTO_STAT_GET(
+ resp_msg->comn_hdr.comn_status) ==
+ ICP_QAT_FW_COMN_STATUS_FLAG_OK) {
+ /* success */
+ memcpy(out_data, resp_msg, queue->msg_size);
+ } else {
+ memset(out_data, 0, queue->msg_size);
+ }
+
+ queue->head = adf_modulo(queue->head + queue->msg_size,
+ queue->modulo_mask);
+ rxq_free_desc(qp, queue);
}
- return resp_counter;
+
+ return result;
+}
+
+/* Sends a NULL message and extracts QAT fw version from the response.
+ * Used to determine detailed capabilities based on the fw version number.
+ * This assumes that there are no inflight messages, i.e. assumes there's space
+ * on the qp, one message is sent and only one response collected.
+ * Returns fw version number or 0 for unknown version or a negative error code.
+ */
+int
+qat_cq_get_fw_version(struct qat_qp *qp)
+{
+ struct qat_queue *queue = &(qp->tx_q);
+ uint8_t *base_addr = (uint8_t *)queue->base_addr;
+ struct icp_qat_fw_comn_req null_msg;
+ struct icp_qat_fw_comn_resp response;
+
+ /* prepare the NULL request */
+ memset(&null_msg, 0, sizeof(null_msg));
+ null_msg.comn_hdr.hdr_flags =
+ ICP_QAT_FW_COMN_HDR_FLAGS_BUILD(ICP_QAT_FW_COMN_REQ_FLAG_SET);
+ null_msg.comn_hdr.service_type = ICP_QAT_FW_COMN_REQ_NULL;
+ null_msg.comn_hdr.service_cmd_id = ICP_QAT_FW_NULL_REQ_SERV_ID;
+
+#if RTE_LOG_DP_LEVEL >= RTE_LOG_DEBUG
+ QAT_DP_HEXDUMP_LOG(DEBUG, "NULL request", &null_msg, sizeof(null_msg));
+#endif
+
+ /* send the NULL request */
+ memcpy(base_addr + queue->tail, &null_msg, sizeof(null_msg));
+ queue->tail = adf_modulo(queue->tail + queue->msg_size,
+ queue->modulo_mask);
+ txq_write_tail(qp, queue);
+
+ /* receive a response */
+ if (qat_cq_dequeue_response(qp, &response)) {
+
+#if RTE_LOG_DP_LEVEL >= RTE_LOG_DEBUG
+ QAT_DP_HEXDUMP_LOG(DEBUG, "NULL response:", &response,
+ sizeof(response));
+#endif
+ /* if LW0 bit 24 is set - then the fw version was returned */
+ if (QAT_FIELD_GET(response.comn_hdr.hdr_flags,
+ ICP_QAT_FW_COMN_NULL_VERSION_FLAG_BITPOS,
+ ICP_QAT_FW_COMN_NULL_VERSION_FLAG_MASK))
+ return response.resrvd[0]; /* return LW4 */
+ else
+ return 0; /* not set - we don't know fw version */
+ }
+
+ QAT_LOG(ERR, "No response received");
+ return -EINVAL;
}
__rte_weak int
-qat_comp_process_response(void **op __rte_unused, uint8_t *resp __rte_unused)
+qat_comp_process_response(void **op __rte_unused, uint8_t *resp __rte_unused,
+ void *op_cookie __rte_unused,
+ uint64_t *dequeue_err_count __rte_unused)
{
return 0;
}