#include <stdint.h>
#include <unistd.h>
#include <sys/mman.h>
+#include <inttypes.h>
/* Verbs header. */
/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
unsigned int i;
for (i = 0; (i != elts_n); ++i)
- (*txq_ctrl->txq.elts)[i] = NULL;
+ txq_ctrl->txq.elts[i] = NULL;
DRV_LOG(DEBUG, "port %u Tx queue %u allocated and configured %u WRs",
PORT_ID(txq_ctrl->priv), txq_ctrl->txq.idx, elts_n);
txq_ctrl->txq.elts_head = 0;
const uint16_t elts_m = elts_n - 1;
uint16_t elts_head = txq_ctrl->txq.elts_head;
uint16_t elts_tail = txq_ctrl->txq.elts_tail;
- struct rte_mbuf *(*elts)[elts_n] = txq_ctrl->txq.elts;
+ struct rte_mbuf *(*elts)[elts_n] = &txq_ctrl->txq.elts;
DRV_LOG(DEBUG, "port %u Tx queue %u freeing WRs",
PORT_ID(txq_ctrl->priv), txq_ctrl->txq.idx);
return -rte_errno;
}
-/**
- * Check if the burst function is using eMPW.
- *
- * @param tx_pkt_burst
- * Tx burst function pointer.
- *
- * @return
- * 1 if the burst function is using eMPW, 0 otherwise.
- */
-static int
-is_empw_burst_func(eth_tx_burst_t tx_pkt_burst)
-{
- if (tx_pkt_burst == mlx5_tx_burst_raw_vec ||
- tx_pkt_burst == mlx5_tx_burst_vec ||
- tx_pkt_burst == mlx5_tx_burst_empw)
- return 1;
- return 0;
-}
-
/**
* Create the Tx queue Verbs object.
*
struct mlx5_txq_ctrl *txq_ctrl =
container_of(txq_data, struct mlx5_txq_ctrl, txq);
struct mlx5_txq_ibv tmpl;
- struct mlx5_txq_ibv *txq_ibv;
+ struct mlx5_txq_ibv *txq_ibv = NULL;
union {
struct ibv_qp_init_attr_ex init;
struct ibv_cq_init_attr_ex cq;
struct mlx5dv_cq cq_info;
struct mlx5dv_obj obj;
const int desc = 1 << txq_data->elts_n;
- eth_tx_burst_t tx_pkt_burst = mlx5_select_tx_function(dev);
int ret = 0;
assert(txq_data);
attr.cq = (struct ibv_cq_init_attr_ex){
.comp_mask = 0,
};
- cqe_n = ((desc / MLX5_TX_COMP_THRESH) - 1) ?
- ((desc / MLX5_TX_COMP_THRESH) - 1) : 1;
- if (is_empw_burst_func(tx_pkt_burst))
- cqe_n += MLX5_TX_COMP_THRESH_INLINE_DIV;
+ cqe_n = desc / MLX5_TX_COMP_THRESH +
+ 1 + MLX5_TX_COMP_THRESH_INLINE_DIV;
tmpl.cq = mlx5_glue->create_cq(priv->sh->ctx, cqe_n, NULL, NULL, 0);
if (tmpl.cq == NULL) {
DRV_LOG(ERR, "port %u Tx queue %u CQ creation failure",
.pd = priv->sh->pd,
.comp_mask = IBV_QP_INIT_ATTR_PD,
};
- if (txq_data->max_inline)
+ if (txq_data->inlen_send)
attr.init.cap.max_inline_data = txq_ctrl->max_inline_data;
if (txq_data->tso_en) {
attr.init.max_tso_header = txq_ctrl->max_tso_header;
goto error;
}
txq_data->cqe_n = log2above(cq_info.cqe_cnt);
+ txq_data->cqe_s = 1 << txq_data->cqe_n;
+ txq_data->cqe_m = txq_data->cqe_s - 1;
txq_data->qp_num_8s = tmpl.qp->qp_num << 8;
txq_data->wqes = qp.sq.buf;
txq_data->wqe_n = log2above(qp.sq.wqe_cnt);
+ txq_data->wqe_s = 1 << txq_data->wqe_n;
+ txq_data->wqe_m = txq_data->wqe_s - 1;
+ txq_data->wqes_end = txq_data->wqes + txq_data->wqe_s;
txq_data->qp_db = &qp.dbrec[MLX5_SND_DBR];
txq_data->cq_db = cq_info.dbrec;
- txq_data->cqes =
- (volatile struct mlx5_cqe (*)[])
- (uintptr_t)cq_info.buf;
+ txq_data->cqes = (volatile struct mlx5_cqe *)cq_info.buf;
txq_data->cq_ci = 0;
#ifndef NDEBUG
txq_data->cq_pi = 0;
#endif
txq_data->wqe_ci = 0;
txq_data->wqe_pi = 0;
+ txq_data->wqe_comp = 0;
+ txq_data->wqe_thres = txq_data->wqe_s / MLX5_TX_COMP_THRESH_INLINE_DIV;
txq_ibv->qp = tmpl.qp;
txq_ibv->cq = tmpl.cq;
rte_atomic32_inc(&txq_ibv->refcnt);
txq_uar_init(txq_ctrl);
if (qp.comp_mask & MLX5DV_QP_MASK_UAR_MMAP_OFFSET) {
txq_ctrl->uar_mmap_offset = qp.uar_mmap_offset;
- DRV_LOG(DEBUG, "port %u: uar_mmap_offset 0x%lx",
+ DRV_LOG(DEBUG, "port %u: uar_mmap_offset 0x%"PRIx64,
dev->data->port_id, txq_ctrl->uar_mmap_offset);
} else {
DRV_LOG(ERR,
claim_zero(mlx5_glue->destroy_cq(tmpl.cq));
if (tmpl.qp)
claim_zero(mlx5_glue->destroy_qp(tmpl.qp));
+ if (txq_ibv)
+ rte_free(txq_ibv);
priv->verbs_alloc_ctx.type = MLX5_VERBS_ALLOC_TYPE_NONE;
rte_errno = ret; /* Restore rte_errno. */
return NULL;
return 1;
}
-/**
- * Return true if a single reference exists on the object.
- *
- * @param txq_ibv
- * Verbs Tx queue object.
- */
-int
-mlx5_txq_ibv_releasable(struct mlx5_txq_ibv *txq_ibv)
-{
- assert(txq_ibv);
- return (rte_atomic32_read(&txq_ibv->refcnt) == 1);
-}
-
/**
* Verify the Verbs Tx queue list is empty
*
return ret;
}
+/**
+ * Calculate the total number of WQEBB for Tx queue.
+ *
+ * Simplified version of calc_sq_size() in rdma-core.
+ *
+ * @param txq_ctrl
+ * Pointer to Tx queue control structure.
+ *
+ * @return
+ * The number of WQEBB.
+ */
+static int
+txq_calc_wqebb_cnt(struct mlx5_txq_ctrl *txq_ctrl)
+{
+ unsigned int wqe_size;
+ const unsigned int desc = 1 << txq_ctrl->txq.elts_n;
+
+ wqe_size = MLX5_WQE_CSEG_SIZE +
+ MLX5_WQE_ESEG_SIZE +
+ MLX5_WSEG_SIZE -
+ MLX5_ESEG_MIN_INLINE_SIZE +
+ txq_ctrl->max_inline_data;
+ return rte_align32pow2(wqe_size * desc) / MLX5_WQE_SIZE;
+}
+
/**
* Set Tx queue parameters from device configuration.
*
{
struct mlx5_priv *priv = txq_ctrl->priv;
struct mlx5_dev_config *config = &priv->config;
- const unsigned int max_tso_inline =
- ((MLX5_MAX_TSO_HEADER + (RTE_CACHE_LINE_SIZE - 1)) /
- RTE_CACHE_LINE_SIZE);
- unsigned int txq_inline;
- unsigned int txqs_inline;
- unsigned int inline_max_packet_sz;
- eth_tx_burst_t tx_pkt_burst =
- mlx5_select_tx_function(ETH_DEV(priv));
- int is_empw_func = is_empw_burst_func(tx_pkt_burst);
- int tso = !!(txq_ctrl->txq.offloads & (DEV_TX_OFFLOAD_TCP_TSO |
- DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
- DEV_TX_OFFLOAD_GRE_TNL_TSO |
- DEV_TX_OFFLOAD_IP_TNL_TSO |
- DEV_TX_OFFLOAD_UDP_TNL_TSO));
-
- txq_inline = (config->txq_inline == MLX5_ARG_UNSET) ?
- 0 : config->txq_inline;
- txqs_inline = (config->txqs_inline == MLX5_ARG_UNSET) ?
- 0 : config->txqs_inline;
- inline_max_packet_sz =
- (config->inline_max_packet_sz == MLX5_ARG_UNSET) ?
- 0 : config->inline_max_packet_sz;
- if (is_empw_func) {
- if (config->txq_inline == MLX5_ARG_UNSET)
- txq_inline = MLX5_WQE_SIZE_MAX - MLX5_WQE_SIZE;
- if (config->txqs_inline == MLX5_ARG_UNSET)
- txqs_inline = MLX5_EMPW_MIN_TXQS;
- if (config->inline_max_packet_sz == MLX5_ARG_UNSET)
- inline_max_packet_sz = MLX5_EMPW_MAX_INLINE_LEN;
- txq_ctrl->txq.mpw_hdr_dseg = config->mpw_hdr_dseg;
- txq_ctrl->txq.inline_max_packet_sz = inline_max_packet_sz;
- }
- if (txq_inline && priv->txqs_n >= txqs_inline) {
- unsigned int ds_cnt;
-
- txq_ctrl->txq.max_inline =
- ((txq_inline + (RTE_CACHE_LINE_SIZE - 1)) /
- RTE_CACHE_LINE_SIZE);
- if (is_empw_func) {
- /* To minimize the size of data set, avoid requesting
- * too large WQ.
+ unsigned int inlen_send; /* Inline data for ordinary SEND.*/
+ unsigned int inlen_empw; /* Inline data for enhanced MPW. */
+ unsigned int inlen_mode; /* Minimal required Inline data. */
+ unsigned int txqs_inline; /* Min Tx queues to enable inline. */
+ uint64_t dev_txoff = priv->dev_data->dev_conf.txmode.offloads;
+ bool tso = txq_ctrl->txq.offloads & (DEV_TX_OFFLOAD_TCP_TSO |
+ DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
+ DEV_TX_OFFLOAD_GRE_TNL_TSO |
+ DEV_TX_OFFLOAD_IP_TNL_TSO |
+ DEV_TX_OFFLOAD_UDP_TNL_TSO);
+ bool vlan_inline;
+ unsigned int temp;
+
+ if (config->txqs_inline == MLX5_ARG_UNSET)
+ txqs_inline =
+#if defined(RTE_ARCH_ARM64)
+ (priv->sh->pci_dev->id.device_id ==
+ PCI_DEVICE_ID_MELLANOX_CONNECTX5BF) ?
+ MLX5_INLINE_MAX_TXQS_BLUEFIELD :
+#endif
+ MLX5_INLINE_MAX_TXQS;
+ else
+ txqs_inline = (unsigned int)config->txqs_inline;
+ inlen_send = (config->txq_inline_max == MLX5_ARG_UNSET) ?
+ MLX5_SEND_DEF_INLINE_LEN :
+ (unsigned int)config->txq_inline_max;
+ inlen_empw = (config->txq_inline_mpw == MLX5_ARG_UNSET) ?
+ MLX5_EMPW_DEF_INLINE_LEN :
+ (unsigned int)config->txq_inline_mpw;
+ inlen_mode = (config->txq_inline_min == MLX5_ARG_UNSET) ?
+ 0 : (unsigned int)config->txq_inline_min;
+ if (config->mps != MLX5_MPW_ENHANCED)
+ inlen_empw = 0;
+ /*
+ * If there is requested minimal amount of data to inline
+ * we MUST enable inlining. This is a case for ConnectX-4
+ * which usually requires L2 inlined for correct operating
+ * and ConnectX-4LX which requires L2-L4 inlined to
+ * support E-Switch Flows.
+ */
+ if (inlen_mode) {
+ if (inlen_mode <= MLX5_ESEG_MIN_INLINE_SIZE) {
+ /*
+ * Optimize minimal inlining for single
+ * segment packets to fill one WQEBB
+ * without gaps.
*/
- txq_ctrl->max_inline_data =
- ((RTE_MIN(txq_inline,
- inline_max_packet_sz) +
- (RTE_CACHE_LINE_SIZE - 1)) /
- RTE_CACHE_LINE_SIZE) * RTE_CACHE_LINE_SIZE;
+ temp = MLX5_ESEG_MIN_INLINE_SIZE;
} else {
- txq_ctrl->max_inline_data =
- txq_ctrl->txq.max_inline * RTE_CACHE_LINE_SIZE;
+ temp = inlen_mode - MLX5_ESEG_MIN_INLINE_SIZE;
+ temp = RTE_ALIGN(temp, MLX5_WSEG_SIZE) +
+ MLX5_ESEG_MIN_INLINE_SIZE;
+ temp = RTE_MIN(temp, MLX5_SEND_MAX_INLINE_LEN);
+ }
+ if (temp != inlen_mode) {
+ DRV_LOG(INFO,
+ "port %u minimal required inline setting"
+ " aligned from %u to %u",
+ PORT_ID(priv), inlen_mode, temp);
+ inlen_mode = temp;
}
+ }
+ /*
+ * If port is configured to support VLAN insertion and device
+ * does not support this feature by HW (for NICs before ConnectX-5
+ * or in case of wqe_vlan_insert flag is not set) we must enable
+ * data inline on all queues because it is supported by single
+ * tx_burst routine.
+ */
+ txq_ctrl->txq.vlan_en = config->hw_vlan_insert;
+ vlan_inline = (dev_txoff & DEV_TX_OFFLOAD_VLAN_INSERT) &&
+ !config->hw_vlan_insert;
+ if (vlan_inline)
+ inlen_send = RTE_MAX(inlen_send, MLX5_ESEG_MIN_INLINE_SIZE);
+ /*
+ * If there are few Tx queues it is prioritized
+ * to save CPU cycles and disable data inlining at all.
+ */
+ if ((inlen_send && priv->txqs_n >= txqs_inline) || vlan_inline) {
/*
- * Check if the inline size is too large in a way which
- * can make the WQE DS to overflow.
- * Considering in calculation:
- * WQE CTRL (1 DS)
- * WQE ETH (1 DS)
- * Inline part (N DS)
+ * The data sent with ordinal MLX5_OPCODE_SEND
+ * may be inlined in Ethernet Segment, align the
+ * length accordingly to fit entire WQEBBs.
*/
- ds_cnt = 2 + (txq_ctrl->txq.max_inline / MLX5_WQE_DWORD_SIZE);
- if (ds_cnt > MLX5_DSEG_MAX) {
- unsigned int max_inline = (MLX5_DSEG_MAX - 2) *
- MLX5_WQE_DWORD_SIZE;
-
- max_inline = max_inline - (max_inline %
- RTE_CACHE_LINE_SIZE);
- DRV_LOG(WARNING,
- "port %u txq inline is too large (%d) setting"
- " it to the maximum possible: %d\n",
- PORT_ID(priv), txq_inline, max_inline);
- txq_ctrl->txq.max_inline = max_inline /
- RTE_CACHE_LINE_SIZE;
+ temp = (inlen_send / MLX5_WQE_SIZE) * MLX5_WQE_SIZE +
+ MLX5_ESEG_MIN_INLINE_SIZE + MLX5_WQE_DSEG_SIZE;
+ temp = RTE_MIN(temp, MLX5_WQE_SIZE_MAX +
+ MLX5_ESEG_MIN_INLINE_SIZE -
+ MLX5_WQE_CSEG_SIZE -
+ MLX5_WQE_ESEG_SIZE -
+ MLX5_WQE_DSEG_SIZE * 2);
+ temp = RTE_MIN(temp, MLX5_SEND_MAX_INLINE_LEN);
+ temp = RTE_MAX(temp, inlen_mode);
+ if (temp != inlen_send) {
+ DRV_LOG(INFO,
+ "port %u ordinary send inline setting"
+ " aligned from %u to %u",
+ PORT_ID(priv), inlen_send, temp);
+ inlen_send = temp;
}
+ /*
+ * Not aligned to cache lines, but to WQEs.
+ * First bytes of data (initial alignment)
+ * is going to be copied explicitly at the
+ * beginning of inlining buffer in Ethernet
+ * Segment.
+ */
+ assert(inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
+ assert(inlen_send <= MLX5_WQE_SIZE_MAX +
+ MLX5_ESEG_MIN_INLINE_SIZE -
+ MLX5_WQE_CSEG_SIZE -
+ MLX5_WQE_ESEG_SIZE -
+ MLX5_WQE_DSEG_SIZE * 2);
+ txq_ctrl->txq.inlen_send = inlen_send;
+ txq_ctrl->txq.inlen_mode = inlen_mode;
+ txq_ctrl->txq.inlen_empw = 0;
+ } else {
+ /*
+ * If minimal inlining is requested we must
+ * enable inlining in general, despite the
+ * number of configured queues.
+ */
+ inlen_send = inlen_mode;
+ if (inlen_mode) {
+ /*
+ * Extend space for inline data to allow
+ * optional alignment of data buffer
+ * start address, it may improve PCIe
+ * performance.
+ */
+ inlen_send = RTE_MIN(inlen_send + MLX5_WQE_SIZE,
+ MLX5_SEND_MAX_INLINE_LEN);
+ }
+ txq_ctrl->txq.inlen_send = inlen_send;
+ txq_ctrl->txq.inlen_mode = inlen_mode;
+ txq_ctrl->txq.inlen_empw = 0;
+ inlen_send = 0;
+ inlen_empw = 0;
}
+ if (inlen_send && inlen_empw && priv->txqs_n >= txqs_inline) {
+ /*
+ * The data sent with MLX5_OPCODE_ENHANCED_MPSW
+ * may be inlined in Data Segment, align the
+ * length accordingly to fit entire WQEBBs.
+ */
+ temp = (inlen_empw + MLX5_WQE_SIZE - 1) / MLX5_WQE_SIZE;
+ temp = temp * MLX5_WQE_SIZE +
+ MLX5_DSEG_MIN_INLINE_SIZE - MLX5_WQE_DSEG_SIZE;
+ temp = RTE_MIN(temp, MLX5_WQE_SIZE_MAX +
+ MLX5_DSEG_MIN_INLINE_SIZE -
+ MLX5_WQE_CSEG_SIZE -
+ MLX5_WQE_ESEG_SIZE -
+ MLX5_WQE_DSEG_SIZE);
+ temp = RTE_MIN(temp, MLX5_EMPW_MAX_INLINE_LEN);
+ if (temp != inlen_empw) {
+ DRV_LOG(INFO,
+ "port %u enhanced empw inline setting"
+ " aligned from %u to %u",
+ PORT_ID(priv), inlen_empw, temp);
+ inlen_empw = temp;
+ }
+ assert(inlen_empw >= MLX5_ESEG_MIN_INLINE_SIZE);
+ assert(inlen_empw <= MLX5_WQE_SIZE_MAX +
+ MLX5_DSEG_MIN_INLINE_SIZE -
+ MLX5_WQE_CSEG_SIZE -
+ MLX5_WQE_ESEG_SIZE -
+ MLX5_WQE_DSEG_SIZE);
+ txq_ctrl->txq.inlen_empw = inlen_empw;
+ }
+ txq_ctrl->max_inline_data = RTE_MAX(inlen_send, inlen_empw);
if (tso) {
- txq_ctrl->max_tso_header = max_tso_inline * RTE_CACHE_LINE_SIZE;
- txq_ctrl->txq.max_inline = RTE_MAX(txq_ctrl->txq.max_inline,
- max_tso_inline);
+ txq_ctrl->max_tso_header = MLX5_MAX_TSO_HEADER;
+ txq_ctrl->max_inline_data = RTE_MAX(txq_ctrl->max_inline_data,
+ MLX5_MAX_TSO_HEADER);
txq_ctrl->txq.tso_en = 1;
}
txq_ctrl->txq.tunnel_en = config->tunnel_en | config->swp;
tmpl->priv = priv;
tmpl->socket = socket;
tmpl->txq.elts_n = log2above(desc);
+ tmpl->txq.elts_s = desc;
+ tmpl->txq.elts_m = desc - 1;
tmpl->txq.port_id = dev->data->port_id;
tmpl->txq.idx = idx;
txq_set_params(tmpl);
- DRV_LOG(DEBUG, "port %u device_attr.max_qp_wr is %d",
- dev->data->port_id, priv->sh->device_attr.orig_attr.max_qp_wr);
- DRV_LOG(DEBUG, "port %u device_attr.max_sge is %d",
- dev->data->port_id, priv->sh->device_attr.orig_attr.max_sge);
- tmpl->txq.elts =
- (struct rte_mbuf *(*)[1 << tmpl->txq.elts_n])(tmpl + 1);
+ if (txq_calc_wqebb_cnt(tmpl) >
+ priv->sh->device_attr.orig_attr.max_qp_wr) {
+ DRV_LOG(ERR,
+ "port %u Tx WQEBB count (%d) exceeds the limit (%d),"
+ " try smaller queue size",
+ dev->data->port_id, txq_calc_wqebb_cnt(tmpl),
+ priv->sh->device_attr.orig_attr.max_qp_wr);
+ rte_errno = ENOMEM;
+ goto error;
+ }
rte_atomic32_inc(&tmpl->refcnt);
LIST_INSERT_HEAD(&priv->txqsctrl, tmpl, next);
return tmpl;
git.droids-corp.org / dpdk.git / blobdiff