void *uar;
uint32_t pdn; /* Protection Domain number. */
uint8_t min_block_size;
+ uint8_t sq_ts_format; /* Whether SQ supports timestamp formats. */
/* Minimum huffman block size supported by the device. */
struct ibv_pd *pd;
struct rte_compressdev_config dev_config;
LIST_HEAD(xform_list, mlx5_compress_xform) xform_list;
rte_spinlock_t xform_sl;
+ struct mlx5_mr_share_cache mr_scache; /* Global shared MR cache. */
+ volatile uint64_t *uar_addr;
+#ifndef RTE_ARCH_64
+ rte_spinlock_t uar32_sl;
+#endif /* RTE_ARCH_64 */
};
struct mlx5_compress_qp {
uint16_t entries_n;
uint16_t pi;
uint16_t ci;
- volatile uint64_t *uar_addr;
+ struct mlx5_mr_ctrl mr_ctrl;
int socket_id;
struct mlx5_devx_cq cq;
struct mlx5_devx_sq sq;
struct mlx5_pmd_mr opaque_mr;
struct rte_comp_op **ops;
struct mlx5_compress_priv *priv;
+ struct rte_compressdev_stats stats;
};
TAILQ_HEAD(mlx5_compress_privs, mlx5_compress_priv) mlx5_compress_priv_list =
int mlx5_compress_logtype;
-const struct rte_compressdev_capabilities mlx5_caps[RTE_COMP_ALGO_LIST_END];
-
+static const struct rte_compressdev_capabilities mlx5_caps[] = {
+ {
+ .algo = RTE_COMP_ALGO_NULL,
+ .comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+ RTE_COMP_FF_CRC32_CHECKSUM |
+ RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+ RTE_COMP_FF_SHAREABLE_PRIV_XFORM,
+ },
+ {
+ .algo = RTE_COMP_ALGO_DEFLATE,
+ .comp_feature_flags = RTE_COMP_FF_ADLER32_CHECKSUM |
+ RTE_COMP_FF_CRC32_CHECKSUM |
+ RTE_COMP_FF_CRC32_ADLER32_CHECKSUM |
+ RTE_COMP_FF_SHAREABLE_PRIV_XFORM |
+ RTE_COMP_FF_HUFFMAN_FIXED |
+ RTE_COMP_FF_HUFFMAN_DYNAMIC,
+ .window_size = {.min = 10, .max = 15, .increment = 1},
+ },
+ {
+ .algo = RTE_COMP_ALGO_LIST_END,
+ }
+};
static void
mlx5_compress_dev_info_get(struct rte_compressdev *dev,
if (opaq != NULL)
rte_free(opaq);
}
+ mlx5_mr_btree_free(&qp->mr_ctrl.cache_bh);
rte_free(qp);
dev->data->queue_pairs[qp_id] = NULL;
return 0;
rte_errno = ENOMEM;
goto err;
}
+ if (mlx5_mr_btree_init(&qp->mr_ctrl.cache_bh, MLX5_MR_BTREE_CACHE_N,
+ priv->dev_config.socket_id)) {
+ DRV_LOG(ERR, "Cannot allocate MR Btree for qp %u.",
+ (uint32_t)qp_id);
+ rte_errno = ENOMEM;
+ goto err;
+ }
qp->entries_n = 1 << log_ops_n;
qp->socket_id = socket_id;
qp->qp_id = qp_id;
qp->priv = priv;
qp->ops = (struct rte_comp_op **)RTE_ALIGN((uintptr_t)(qp + 1),
RTE_CACHE_LINE_SIZE);
- qp->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
- MLX5_ASSERT(qp->uar_addr);
if (mlx5_common_verbs_reg_mr(priv->pd, opaq_buf, qp->entries_n *
sizeof(struct mlx5_gga_compress_opaque),
&qp->opaque_mr) != 0) {
goto err;
}
sq_attr.cqn = qp->cq.cq->id;
+ sq_attr.ts_format = mlx5_ts_format_conv(priv->sq_ts_format);
ret = mlx5_devx_sq_create(priv->ctx, &qp->sq, log_ops_n, &sq_attr,
socket_id);
if (ret != 0) {
DRV_LOG(ERR, "Can't change SQ state to ready.");
goto err;
}
- DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u\n",
+ DRV_LOG(INFO, "QP %u: SQN=0x%X CQN=0x%X entries num = %u",
(uint32_t)qp_id, qp->sq.sq->id, qp->cq.cq->id, qp->entries_n);
return 0;
err:
return -ENOTSUP;
}
+static void
+mlx5_compress_dev_stop(struct rte_compressdev *dev)
+{
+ RTE_SET_USED(dev);
+}
+
+static int
+mlx5_compress_dev_start(struct rte_compressdev *dev)
+{
+ RTE_SET_USED(dev);
+ return 0;
+}
+
+static void
+mlx5_compress_stats_get(struct rte_compressdev *dev,
+ struct rte_compressdev_stats *stats)
+{
+ int qp_id;
+
+ for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+ struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+ stats->enqueued_count += qp->stats.enqueued_count;
+ stats->dequeued_count += qp->stats.dequeued_count;
+ stats->enqueue_err_count += qp->stats.enqueue_err_count;
+ stats->dequeue_err_count += qp->stats.dequeue_err_count;
+ }
+}
+
+static void
+mlx5_compress_stats_reset(struct rte_compressdev *dev)
+{
+ int qp_id;
+
+ for (qp_id = 0; qp_id < dev->data->nb_queue_pairs; qp_id++) {
+ struct mlx5_compress_qp *qp = dev->data->queue_pairs[qp_id];
+
+ memset(&qp->stats, 0, sizeof(qp->stats));
+ }
+}
+
static struct rte_compressdev_ops mlx5_compress_ops = {
.dev_configure = mlx5_compress_dev_configure,
- .dev_start = NULL,
- .dev_stop = NULL,
+ .dev_start = mlx5_compress_dev_start,
+ .dev_stop = mlx5_compress_dev_stop,
.dev_close = mlx5_compress_dev_close,
.dev_infos_get = mlx5_compress_dev_info_get,
- .stats_get = NULL,
- .stats_reset = NULL,
+ .stats_get = mlx5_compress_stats_get,
+ .stats_reset = mlx5_compress_stats_reset,
.queue_pair_setup = mlx5_compress_qp_setup,
.queue_pair_release = mlx5_compress_qp_release,
.private_xform_create = mlx5_compress_xform_create,
.stream_free = NULL,
};
+static __rte_always_inline uint32_t
+mlx5_compress_dseg_set(struct mlx5_compress_qp *qp,
+ volatile struct mlx5_wqe_dseg *restrict dseg,
+ struct rte_mbuf *restrict mbuf,
+ uint32_t offset, uint32_t len)
+{
+ uintptr_t addr = rte_pktmbuf_mtod_offset(mbuf, uintptr_t, offset);
+
+ dseg->bcount = rte_cpu_to_be_32(len);
+ dseg->lkey = mlx5_mr_addr2mr_bh(qp->priv->pd, 0, &qp->priv->mr_scache,
+ &qp->mr_ctrl, addr,
+ !!(mbuf->ol_flags & EXT_ATTACHED_MBUF));
+ dseg->pbuf = rte_cpu_to_be_64(addr);
+ return dseg->lkey;
+}
+
+/*
+ * Provide safe 64bit store operation to mlx5 UAR region for both 32bit and
+ * 64bit architectures.
+ */
+static __rte_always_inline void
+mlx5_compress_uar_write(uint64_t val, struct mlx5_compress_priv *priv)
+{
+#ifdef RTE_ARCH_64
+ *priv->uar_addr = val;
+#else /* !RTE_ARCH_64 */
+ rte_spinlock_lock(&priv->uar32_sl);
+ *(volatile uint32_t *)priv->uar_addr = val;
+ rte_io_wmb();
+ *((volatile uint32_t *)priv->uar_addr + 1) = val >> 32;
+ rte_spinlock_unlock(&priv->uar32_sl);
+#endif
+}
+
+static uint16_t
+mlx5_compress_enqueue_burst(void *queue_pair, struct rte_comp_op **ops,
+ uint16_t nb_ops)
+{
+ struct mlx5_compress_qp *qp = queue_pair;
+ volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+ qp->sq.wqes, *wqe;
+ struct mlx5_compress_xform *xform;
+ struct rte_comp_op *op;
+ uint16_t mask = qp->entries_n - 1;
+ uint16_t remain = qp->entries_n - (qp->pi - qp->ci);
+ uint16_t idx;
+ bool invalid;
+
+ if (remain < nb_ops)
+ nb_ops = remain;
+ else
+ remain = nb_ops;
+ if (unlikely(remain == 0))
+ return 0;
+ do {
+ idx = qp->pi & mask;
+ wqe = &wqes[idx];
+ rte_prefetch0(&wqes[(qp->pi + 1) & mask]);
+ op = *ops++;
+ xform = op->private_xform;
+ /*
+ * Check operation arguments and error cases:
+ * - Operation type must be state-less.
+ * - Compress operation flush flag must be FULL or FINAL.
+ * - Source and destination buffers must be mapped internally.
+ */
+ invalid = op->op_type != RTE_COMP_OP_STATELESS ||
+ (xform->type == RTE_COMP_COMPRESS &&
+ op->flush_flag < RTE_COMP_FLUSH_FULL);
+ if (unlikely(invalid ||
+ (mlx5_compress_dseg_set(qp, &wqe->gather,
+ op->m_src,
+ op->src.offset,
+ op->src.length) ==
+ UINT32_MAX) ||
+ (mlx5_compress_dseg_set(qp, &wqe->scatter,
+ op->m_dst,
+ op->dst.offset,
+ rte_pktmbuf_pkt_len(op->m_dst) -
+ op->dst.offset) ==
+ UINT32_MAX))) {
+ op->status = invalid ? RTE_COMP_OP_STATUS_INVALID_ARGS :
+ RTE_COMP_OP_STATUS_ERROR;
+ nb_ops -= remain;
+ if (unlikely(nb_ops == 0))
+ return 0;
+ break;
+ }
+ wqe->gga_ctrl1 = xform->gga_ctrl1;
+ wqe->opcode = rte_cpu_to_be_32(xform->opcode + (qp->pi << 8));
+ qp->ops[idx] = op;
+ qp->pi++;
+ } while (--remain);
+ qp->stats.enqueued_count += nb_ops;
+ rte_io_wmb();
+ qp->sq.db_rec[MLX5_SND_DBR] = rte_cpu_to_be_32(qp->pi);
+ rte_wmb();
+ mlx5_compress_uar_write(*(volatile uint64_t *)wqe, qp->priv);
+ rte_wmb();
+ return nb_ops;
+}
+
+static void
+mlx5_compress_dump_err_objs(volatile uint32_t *cqe, volatile uint32_t *wqe,
+ volatile uint32_t *opaq)
+{
+ size_t i;
+
+ DRV_LOG(ERR, "Error cqe:");
+ for (i = 0; i < sizeof(struct mlx5_err_cqe) >> 2; i += 4)
+ DRV_LOG(ERR, "%08X %08X %08X %08X", cqe[i], cqe[i + 1],
+ cqe[i + 2], cqe[i + 3]);
+ DRV_LOG(ERR, "\nError wqe:");
+ for (i = 0; i < sizeof(struct mlx5_gga_wqe) >> 2; i += 4)
+ DRV_LOG(ERR, "%08X %08X %08X %08X", wqe[i], wqe[i + 1],
+ wqe[i + 2], wqe[i + 3]);
+ DRV_LOG(ERR, "\nError opaq:");
+ for (i = 0; i < sizeof(struct mlx5_gga_compress_opaque) >> 2; i += 4)
+ DRV_LOG(ERR, "%08X %08X %08X %08X", opaq[i], opaq[i + 1],
+ opaq[i + 2], opaq[i + 3]);
+}
+
+static void
+mlx5_compress_cqe_err_handle(struct mlx5_compress_qp *qp,
+ struct rte_comp_op *op)
+{
+ const uint32_t idx = qp->ci & (qp->entries_n - 1);
+ volatile struct mlx5_err_cqe *cqe = (volatile struct mlx5_err_cqe *)
+ &qp->cq.cqes[idx];
+ volatile struct mlx5_gga_wqe *wqes = (volatile struct mlx5_gga_wqe *)
+ qp->sq.wqes;
+ volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+
+ op->status = RTE_COMP_OP_STATUS_ERROR;
+ op->consumed = 0;
+ op->produced = 0;
+ op->output_chksum = 0;
+ op->debug_status = rte_be_to_cpu_32(opaq[idx].syndrom) |
+ ((uint64_t)rte_be_to_cpu_32(cqe->syndrome) << 32);
+ mlx5_compress_dump_err_objs((volatile uint32_t *)cqe,
+ (volatile uint32_t *)&wqes[idx],
+ (volatile uint32_t *)&opaq[idx]);
+ qp->stats.dequeue_err_count++;
+}
+
+static uint16_t
+mlx5_compress_dequeue_burst(void *queue_pair, struct rte_comp_op **ops,
+ uint16_t nb_ops)
+{
+ struct mlx5_compress_qp *qp = queue_pair;
+ volatile struct mlx5_compress_xform *restrict xform;
+ volatile struct mlx5_cqe *restrict cqe;
+ volatile struct mlx5_gga_compress_opaque *opaq = qp->opaque_mr.addr;
+ struct rte_comp_op *restrict op;
+ const unsigned int cq_size = qp->entries_n;
+ const unsigned int mask = cq_size - 1;
+ uint32_t idx;
+ uint32_t next_idx = qp->ci & mask;
+ const uint16_t max = RTE_MIN((uint16_t)(qp->pi - qp->ci), nb_ops);
+ uint16_t i = 0;
+ int ret;
+
+ if (unlikely(max == 0))
+ return 0;
+ do {
+ idx = next_idx;
+ next_idx = (qp->ci + 1) & mask;
+ rte_prefetch0(&qp->cq.cqes[next_idx]);
+ rte_prefetch0(qp->ops[next_idx]);
+ op = qp->ops[idx];
+ cqe = &qp->cq.cqes[idx];
+ ret = check_cqe(cqe, cq_size, qp->ci);
+ /*
+ * Be sure owner read is done before any other cookie field or
+ * opaque field.
+ */
+ rte_io_rmb();
+ if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
+ if (likely(ret == MLX5_CQE_STATUS_HW_OWN))
+ break;
+ mlx5_compress_cqe_err_handle(qp, op);
+ } else {
+ xform = op->private_xform;
+ op->status = RTE_COMP_OP_STATUS_SUCCESS;
+ op->consumed = op->src.length;
+ op->produced = rte_be_to_cpu_32(cqe->byte_cnt);
+ MLX5_ASSERT(cqe->byte_cnt ==
+ opaq[idx].scattered_length);
+ switch (xform->csum_type) {
+ case RTE_COMP_CHECKSUM_CRC32:
+ op->output_chksum = (uint64_t)rte_be_to_cpu_32
+ (opaq[idx].crc32);
+ break;
+ case RTE_COMP_CHECKSUM_ADLER32:
+ op->output_chksum = (uint64_t)rte_be_to_cpu_32
+ (opaq[idx].adler32) << 32;
+ break;
+ case RTE_COMP_CHECKSUM_CRC32_ADLER32:
+ op->output_chksum = (uint64_t)rte_be_to_cpu_32
+ (opaq[idx].crc32) |
+ ((uint64_t)rte_be_to_cpu_32
+ (opaq[idx].adler32) << 32);
+ break;
+ default:
+ break;
+ }
+ }
+ ops[i++] = op;
+ qp->ci++;
+ } while (i < max);
+ if (likely(i != 0)) {
+ rte_io_wmb();
+ qp->cq.db_rec[0] = rte_cpu_to_be_32(qp->ci);
+ qp->stats.dequeued_count += i;
+ }
+ return i;
+}
+
static struct ibv_device *
mlx5_compress_get_ib_device_match(struct rte_pci_addr *addr)
{
DRV_LOG(ERR, "Failed to allocate UAR.");
return -1;
}
+ priv->uar_addr = mlx5_os_get_devx_uar_reg_addr(priv->uar);
+ MLX5_ASSERT(priv->uar_addr);
+#ifndef RTE_ARCH_64
+ rte_spinlock_init(&priv->uar32_sl);
+#endif /* RTE_ARCH_64 */
return 0;
}
DRV_LOG(INFO,
"Compress device %s was created successfully.", ibv->name);
cdev->dev_ops = &mlx5_compress_ops;
- cdev->dequeue_burst = NULL;
- cdev->enqueue_burst = NULL;
+ cdev->dequeue_burst = mlx5_compress_dequeue_burst;
+ cdev->enqueue_burst = mlx5_compress_enqueue_burst;
cdev->feature_flags = RTE_COMPDEV_FF_HW_ACCELERATED;
priv = cdev->data->dev_private;
priv->ctx = ctx;
priv->pci_dev = pci_dev;
priv->cdev = cdev;
priv->min_block_size = att.compress_min_block_size;
+ priv->sq_ts_format = att.sq_ts_format;
if (mlx5_compress_hw_global_prepare(priv) != 0) {
rte_compressdev_pmd_destroy(priv->cdev);
claim_zero(mlx5_glue->close_device(priv->ctx));
return -1;
}
+ if (mlx5_mr_btree_init(&priv->mr_scache.cache,
+ MLX5_MR_BTREE_CACHE_N * 2, rte_socket_id()) != 0) {
+ DRV_LOG(ERR, "Failed to allocate shared cache MR memory.");
+ mlx5_compress_hw_global_release(priv);
+ rte_compressdev_pmd_destroy(priv->cdev);
+ claim_zero(mlx5_glue->close_device(priv->ctx));
+ rte_errno = ENOMEM;
+ return -rte_errno;
+ }
+ priv->mr_scache.reg_mr_cb = mlx5_common_verbs_reg_mr;
+ priv->mr_scache.dereg_mr_cb = mlx5_common_verbs_dereg_mr;
pthread_mutex_lock(&priv_list_lock);
TAILQ_INSERT_TAIL(&mlx5_compress_priv_list, priv, next);
pthread_mutex_unlock(&priv_list_lock);
TAILQ_REMOVE(&mlx5_compress_priv_list, priv, next);
pthread_mutex_unlock(&priv_list_lock);
if (priv) {
+ mlx5_mr_release_cache(&priv->mr_scache);
mlx5_compress_hw_global_release(priv);
rte_compressdev_pmd_destroy(priv->cdev);
claim_zero(mlx5_glue->close_device(priv->ctx));