Add zsda compressdev dequeue datapath.

Signed-off-by: Hanxiao Li <li.hanxiao@zte.com.cn>
---
 drivers/common/zsda/zsda_qp.c         |  55 +++++++++
 drivers/common/zsda/zsda_qp.h         |   1 +
 drivers/common/zsda/zsda_qp_common.h  |   4 +
 drivers/compress/zsda/zsda_comp.c     | 155 ++++++++++++++++++++++++++
 drivers/compress/zsda/zsda_comp.h     |   9 ++
 drivers/compress/zsda/zsda_comp_pmd.c |  14 ++-
 6 files changed, 235 insertions(+), 3 deletions(-)

diff --git a/drivers/common/zsda/zsda_qp.c b/drivers/common/zsda/zsda_qp.c
index c2a7d9b28b..fd45558868 100644
--- a/drivers/common/zsda/zsda_qp.c
+++ b/drivers/common/zsda/zsda_qp.c
@@ -878,3 +878,58 @@ zsda_enqueue_op_burst(struct zsda_qp *qp, void **ops, const uint16_t nb_ops)
     return nb_send;
 }
 
+static void
+zsda_dequeue(struct qp_srv *srv, void **ops, const uint16_t nb_ops, uint16_t *nb)
+{
+    uint16_t head;
+    struct zsda_cqe *cqe;
+    struct zsda_queue *queue = &srv->rx_q;
+    struct zsda_op_cookie *cookie;
+    head = queue->head;
+
+    while (*nb < nb_ops) {
+        cqe = (struct zsda_cqe *)(
+            (uint8_t *)queue->base_addr + head * queue->msg_size);
+
+        if (!CQE_VALID(cqe->err1))
+            break;
+        cookie = srv->op_cookies[cqe->sid];
+
+        ops[*nb] = cookie->op;
+        if (srv->rx_cb(cookie, cqe) == ZSDA_SUCCESS)
+            srv->stats.dequeued_count++;
+        else {
+            ZSDA_LOG(ERR,
+                 "ERR! Cqe, opcode 0x%x, sid 0x%x, "
+                 "tx_real_length 0x%x, err0 0x%x, err1 0x%x",
+                 cqe->op_code, cqe->sid, cqe->tx_real_length,
+                 cqe->err0, cqe->err1);
+            srv->stats.dequeue_err_count++;
+        }
+        (*nb)++;
+        cookie->used = false;
+
+        head = zsda_modulo_16(head + 1, queue->modulo_mask);
+        queue->head = head;
+        WRITE_CSR_CQ_HEAD(queue->io_addr, queue->hw_queue_number, head);
+        memset(cqe, 0x0, sizeof(struct zsda_cqe));
+    }
+}
+
+uint16_t
+zsda_dequeue_op_burst(struct zsda_qp *qp, void **ops, const uint16_t nb_ops)
+{
+    uint16_t nb = 0;
+    uint32_t type = 0;
+    struct qp_srv *srv;
+
+    for (type = 0; type < ZSDA_SERVICE_INVALID; type++) {
+        if (!qp->srv[type].used)
+            continue;
+        srv = &qp->srv[type];
+        zsda_dequeue(srv, ops, nb_ops, &nb);
+        if (nb >= nb_ops)
+            return nb_ops;
+    }
+    return nb;
+}
diff --git a/drivers/common/zsda/zsda_qp.h b/drivers/common/zsda/zsda_qp.h
index c68c8ed380..9595a91208 100644
--- a/drivers/common/zsda/zsda_qp.h
+++ b/drivers/common/zsda/zsda_qp.h
@@ -177,5 +177,6 @@ int zsda_common_setup_qp(uint32_t dev_id, struct zsda_qp **qp_addr,
         const uint16_t queue_pair_id, const struct zsda_qp_config *conf);
 
 uint16_t zsda_enqueue_op_burst(struct zsda_qp *qp, void **ops, const uint16_t nb_ops);
+uint16_t zsda_dequeue_op_burst(struct zsda_qp *qp, void **ops, const uint16_t nb_ops);
 
 #endif /* _ZSDA_QP_H_ */
diff --git a/drivers/common/zsda/zsda_qp_common.h b/drivers/common/zsda/zsda_qp_common.h
index e4b00d8307..834a89e910 100644
--- a/drivers/common/zsda/zsda_qp_common.h
+++ b/drivers/common/zsda/zsda_qp_common.h
@@ -47,6 +47,10 @@ enum zsda_service_type {
 #define ZSDA_OPC_DECOMP_ZLIB    0x19 /* Decomp inflate-Zlib */
 #define ZSDA_OPC_INVALID        0xff
 
+#define CQE_VALID(value) (value & 0x8000)
+#define CQE_ERR0(value) (value & 0xFFFF)
+#define CQE_ERR1(value) (value & 0x7FFF)
+
 enum wqe_element_type {
     WQE_ELM_TYPE_PHYS_ADDR = 1,
     WQE_ELM_TYPE_LIST,
diff --git a/drivers/compress/zsda/zsda_comp.c b/drivers/compress/zsda/zsda_comp.c
index c00e0c2a4b..c22d690e5e 100644
--- a/drivers/compress/zsda/zsda_comp.c
+++ b/drivers/compress/zsda/zsda_comp.c
@@ -10,6 +10,83 @@
 #define GZIP_TRAILER_SIZE 8
 #define CHECKSUM_SIZE 4
 
+#define POLYNOMIAL 0xEDB88320
+static uint32_t crc32_table[8][256];
+static int table_config;
+
+static void
+build_crc32_table(void)
+{
+    for (uint32_t i = 0; i < 256; i++) {
+        uint32_t crc = i;
+        for (uint32_t j = 0; j < 8; j++)
+            crc = (crc >> 1) ^ ((crc & 1) ? POLYNOMIAL : 0);
+        crc32_table[0][i] = crc;
+    }
+
+    for (int i = 1; i < 8; i++) {
+        for (uint32_t j = 0; j < 256; j++)
+            crc32_table[i][j] = (crc32_table[i-1][j] >> 8) ^
+                    crc32_table[0][crc32_table[i-1][j] & 0xFF];
+    }
+    table_config = 1;
+}
+
+static uint32_t
+zsda_crc32(const uint8_t *data, size_t length)
+{
+    uint32_t crc = 0xFFFFFFFF;
+
+    if (!table_config)
+        build_crc32_table();
+
+    while (length >= 8) {
+        crc ^= *(const uint32_t *)data;
+        crc = crc32_table[7][crc & 0xFF] ^
+              crc32_table[6][(crc >> 8) & 0xFF] ^
+              crc32_table[5][(crc >> 16) & 0xFF] ^
+              crc32_table[4][(crc >> 24) & 0xFF] ^
+              crc32_table[3][data[4]] ^
+              crc32_table[2][data[5]] ^
+              crc32_table[1][data[6]] ^
+              crc32_table[0][data[7]];
+
+        data += 8;
+        length -= 8;
+    }
+
+    for (size_t i = 0; i < length; i++)
+        crc = (crc >> 8) ^ crc32_table[0][(crc ^ data[i]) & 0xFF];
+
+    return crc ^ 0xFFFFFFFF;
+}
+
+#define MOD_ADLER 65521
+#define NMAX 5552
+static uint32_t
+zsda_adler32(const uint8_t *buf, uint32_t len)
+{
+    uint32_t s1 = 1;
+    uint32_t s2 = 0;
+
+    while (len > 0) {
+        uint32_t k = (len < NMAX) ? len : NMAX;
+        len -= k;
+
+        for (uint32_t i = 0; i < k; i++) {
+            s1 += buf[i];
+            s2 += s1;
+        }
+
+        s1 %= MOD_ADLER;
+        s2 %= MOD_ADLER;
+
+        buf += k;
+    }
+
+    return (s2 << 16) | s1;
+}
+
 int
 zsda_comp_match(const void *op_in)
 {
@@ -231,3 +308,81 @@ zsda_build_decomp_request(void *op_in, const struct zsda_queue *queue,
 
     return ret;
 }
+
+static uint32_t
+zsda_read_chksum(uint8_t *data_addr, uint8_t op_code, uint32_t produced)
+{
+    uint8_t *chk_addr;
+    uint32_t chksum = 0;
+    int i = 0;
+
+    if (op_code == ZSDA_OPC_COMP_ZLIB) {
+        chk_addr = data_addr + produced - ZLIB_TRAILER_SIZE;
+        for (i = 0; i < CHECKSUM_SIZE; i++) {
+            chksum = chksum << 8;
+            chksum |= (*(chk_addr + i));
+        }
+    } else if (op_code == ZSDA_OPC_COMP_GZIP) {
+        chk_addr = data_addr + produced - GZIP_TRAILER_SIZE;
+        for (i = 0; i < CHECKSUM_SIZE; i++)
+            chksum |= (*(chk_addr + i) << (i * 8));
+    }
+
+    return chksum;
+}
+
+int
+zsda_comp_callback(void *cookie_in, struct zsda_cqe *cqe)
+{
+    struct zsda_op_cookie *tmp_cookie = cookie_in;
+    struct rte_comp_op *tmp_op = tmp_cookie->op;
+    uint8_t *data_addr =
+        (uint8_t *)tmp_op->m_dst->buf_addr + tmp_op->m_dst->data_off;
+    uint32_t chksum = 0;
+    uint16_t head_len;
+    uint16_t tail_len;
+
+    if (tmp_cookie->decomp_no_tail && CQE_ERR0_RIGHT(cqe->err0))
+        cqe->err0 = 0x0000;
+
+    if (!(CQE_ERR0(cqe->err0) || CQE_ERR1(cqe->err1)))
+        tmp_op->status = RTE_COMP_OP_STATUS_SUCCESS;
+    else {
+        tmp_op->status = RTE_COMP_OP_STATUS_ERROR;
+        return ZSDA_FAILED;
+    }
+
+    /* handle chksum */
+    tmp_op->produced = cqe->tx_real_length;
+    if (cqe->op_code == ZSDA_OPC_COMP_ZLIB) {
+        head_len = ZLIB_HEADER_SIZE;
+        tail_len = ZLIB_TRAILER_SIZE;
+        chksum = zsda_read_chksum(data_addr, cqe->op_code,
+                          tmp_op->produced - head_len);
+    }
+    if (cqe->op_code == ZSDA_OPC_COMP_GZIP) {
+        head_len = GZIP_HEADER_SIZE;
+        tail_len = GZIP_TRAILER_SIZE;
+        chksum = zsda_read_chksum(data_addr, cqe->op_code,
+                          tmp_op->produced - head_len);
+    } else if (cqe->op_code == ZSDA_OPC_DECOMP_ZLIB) {
+        head_len = ZLIB_HEADER_SIZE;
+        tail_len = ZLIB_TRAILER_SIZE;
+        chksum = zsda_adler32(data_addr, tmp_op->produced);
+    } else if (cqe->op_code == ZSDA_OPC_DECOMP_GZIP) {
+        head_len = GZIP_HEADER_SIZE;
+        tail_len = GZIP_TRAILER_SIZE;
+        chksum = zsda_crc32(data_addr, tmp_op->produced);
+    }
+    tmp_op->output_chksum = chksum;
+
+    if (cqe->op_code == ZSDA_OPC_COMP_ZLIB ||
+        cqe->op_code == ZSDA_OPC_COMP_GZIP) {
+        /* remove tail data*/
+        rte_pktmbuf_trim(tmp_op->m_dst, GZIP_TRAILER_SIZE);
+        /* remove head and tail length */
+        tmp_op->produced = tmp_op->produced - (head_len + tail_len);
+    }
+
+    return ZSDA_SUCCESS;
+}
diff --git a/drivers/compress/zsda/zsda_comp.h b/drivers/compress/zsda/zsda_comp.h
index e7dea7e5a4..bb2feecb31 100644
--- a/drivers/compress/zsda/zsda_comp.h
+++ b/drivers/compress/zsda/zsda_comp.h
@@ -24,6 +24,14 @@ struct zsda_wqe_comp {
     struct compress_cfg cfg;
 } __rte_packed;
 
+/* For situations where err0 are reported but the results are correct */
+#define DECOMP_RIGHT_ERR0_0 0xC710
+#define DECOMP_RIGHT_ERR0_1 0xC727
+#define DECOMP_RIGHT_ERR0_2 0xC729
+#define CQE_ERR0_RIGHT(value)                                                  \
+    (value == DECOMP_RIGHT_ERR0_0 || value == DECOMP_RIGHT_ERR0_1 ||       \
+     value == DECOMP_RIGHT_ERR0_2)
+
 int zsda_comp_match(const void *op_in);
 int zsda_decomp_match(const void *op_in);
 
@@ -32,5 +40,6 @@ int zsda_build_comp_request(void *op_in, const struct zsda_queue *queue,
 
 int zsda_build_decomp_request(void *op_in, const struct zsda_queue *queue,
              void **op_cookies, const uint16_t new_tail);
+int zsda_comp_callback(void *cookie_in, struct zsda_cqe *cqe);
 
 #endif /* _ZSDA_COMP_H_ */
diff --git a/drivers/compress/zsda/zsda_comp_pmd.c b/drivers/compress/zsda/zsda_comp_pmd.c
index 9cb2169d68..ea3de2f505 100644
--- a/drivers/compress/zsda/zsda_comp_pmd.c
+++ b/drivers/compress/zsda/zsda_comp_pmd.c
@@ -229,7 +229,7 @@ zsda_setup_comp_queue(struct zsda_pci_device *zsda_pci_dev, const uint16_t qp_id
     conf.service_str = "comp";
 
     ret = zsda_common_setup_qp(zsda_pci_dev->zsda_dev_id, &qp, qp_id, &conf);
-    qp->srv[type].rx_cb = NULL;
+    qp->srv[type].rx_cb = zsda_comp_callback;
     qp->srv[type].tx_cb = zsda_build_comp_request;
     qp->srv[type].match = zsda_comp_match;
 
@@ -254,7 +254,7 @@ zsda_setup_decomp_queue(struct zsda_pci_device *zsda_pci_dev, const uint16_t qp_
     conf.service_str = "decomp";
 
     ret = zsda_common_setup_qp(zsda_pci_dev->zsda_dev_id, &qp, qp_id, &conf);
-    qp->srv[type].rx_cb = NULL;
+    qp->srv[type].rx_cb = zsda_comp_callback;
     qp->srv[type].tx_cb = zsda_build_decomp_request;
     qp->srv[type].match = zsda_decomp_match;
 
@@ -348,6 +348,14 @@ zsda_comp_pmd_enqueue_op_burst(void *qp, struct rte_comp_op **ops,
                      nb_ops);
 }
 
+static uint16_t
+zsda_comp_pmd_dequeue_op_burst(void *qp, struct rte_comp_op **ops,
+                   uint16_t nb_ops)
+{
+    return zsda_dequeue_op_burst((struct zsda_qp *)qp, (void **)ops,
+                     nb_ops);
+}
+
 int
 zsda_comp_dev_create(struct zsda_pci_device *zsda_pci_dev)
 {
@@ -386,7 +394,7 @@ zsda_comp_dev_create(struct zsda_pci_device *zsda_pci_dev)
     compressdev->dev_ops = &compress_zsda_ops;
 
     compressdev->enqueue_burst = zsda_comp_pmd_enqueue_op_burst;
-    compressdev->dequeue_burst = NULL;
+    compressdev->dequeue_burst = zsda_comp_pmd_dequeue_op_burst;
 
     compressdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
 
-- 
2.27.0