#include <errno.h>
#include <string.h>
#include <stdint.h>
+#include <unistd.h>
+#include <sys/mman.h>
/* Verbs header. */
/* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
#ifdef PEDANTIC
-#pragma GCC diagnostic ignored "-pedantic"
+#pragma GCC diagnostic ignored "-Wpedantic"
#endif
#include <infiniband/verbs.h>
#ifdef PEDANTIC
-#pragma GCC diagnostic error "-pedantic"
+#pragma GCC diagnostic error "-Wpedantic"
#endif
-/* DPDK headers don't like -pedantic. */
-#ifdef PEDANTIC
-#pragma GCC diagnostic ignored "-pedantic"
-#endif
#include <rte_mbuf.h>
#include <rte_malloc.h>
#include <rte_ethdev.h>
#include <rte_common.h>
-#ifdef PEDANTIC
-#pragma GCC diagnostic error "-pedantic"
-#endif
#include "mlx5_utils.h"
#include "mlx5_defs.h"
#include "mlx5.h"
#include "mlx5_rxtx.h"
#include "mlx5_autoconf.h"
-#include "mlx5_defs.h"
/**
* Allocate TX queue elements.
for (i = 0; (i != elts_n); ++i)
(*txq_ctrl->txq.elts)[i] = NULL;
- for (i = 0; (i != txq_ctrl->txq.wqe_n); ++i) {
- volatile union mlx5_wqe *wqe = &(*txq_ctrl->txq.wqes)[i];
+ for (i = 0; (i != (1u << txq_ctrl->txq.wqe_n)); ++i) {
+ volatile struct mlx5_wqe64 *wqe =
+ (volatile struct mlx5_wqe64 *)
+ txq_ctrl->txq.wqes + i;
memset((void *)(uintptr_t)wqe, 0x0, sizeof(*wqe));
}
static void
txq_free_elts(struct txq_ctrl *txq_ctrl)
{
- unsigned int elts_n = txq_ctrl->txq.elts_n;
- unsigned int elts_head = txq_ctrl->txq.elts_head;
- unsigned int elts_tail = txq_ctrl->txq.elts_tail;
+ const uint16_t elts_n = 1 << txq_ctrl->txq.elts_n;
+ 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;
DEBUG("%p: freeing WRs", (void *)txq_ctrl);
txq_ctrl->txq.elts_comp = 0;
while (elts_tail != elts_head) {
- struct rte_mbuf *elt = (*elts)[elts_tail];
+ struct rte_mbuf *elt = (*elts)[elts_tail & elts_m];
assert(elt != NULL);
- rte_pktmbuf_free(elt);
+ rte_pktmbuf_free_seg(elt);
#ifndef NDEBUG
/* Poisoning. */
- memset(&(*elts)[elts_tail],
+ memset(&(*elts)[elts_tail & elts_m],
0x77,
- sizeof((*elts)[elts_tail]));
+ sizeof((*elts)[elts_tail & elts_m]));
#endif
- if (++elts_tail == elts_n)
- elts_tail = 0;
+ ++elts_tail;
}
}
void
txq_cleanup(struct txq_ctrl *txq_ctrl)
{
- struct ibv_exp_release_intf_params params;
size_t i;
DEBUG("cleaning up %p", (void *)txq_ctrl);
txq_free_elts(txq_ctrl);
- if (txq_ctrl->if_qp != NULL) {
- assert(txq_ctrl->priv != NULL);
- assert(txq_ctrl->priv->ctx != NULL);
- assert(txq_ctrl->qp != NULL);
- params = (struct ibv_exp_release_intf_params){
- .comp_mask = 0,
- };
- claim_zero(ibv_exp_release_intf(txq_ctrl->priv->ctx,
- txq_ctrl->if_qp,
- ¶ms));
- }
- if (txq_ctrl->if_cq != NULL) {
- assert(txq_ctrl->priv != NULL);
- assert(txq_ctrl->priv->ctx != NULL);
- assert(txq_ctrl->cq != NULL);
- params = (struct ibv_exp_release_intf_params){
- .comp_mask = 0,
- };
- claim_zero(ibv_exp_release_intf(txq_ctrl->priv->ctx,
- txq_ctrl->if_cq,
- ¶ms));
- }
if (txq_ctrl->qp != NULL)
claim_zero(ibv_destroy_qp(txq_ctrl->qp));
if (txq_ctrl->cq != NULL)
claim_zero(ibv_destroy_cq(txq_ctrl->cq));
- if (txq_ctrl->rd != NULL) {
- struct ibv_exp_destroy_res_domain_attr attr = {
- .comp_mask = 0,
- };
-
- assert(txq_ctrl->priv != NULL);
- assert(txq_ctrl->priv->ctx != NULL);
- claim_zero(ibv_exp_destroy_res_domain(txq_ctrl->priv->ctx,
- txq_ctrl->rd,
- &attr));
- }
for (i = 0; (i != RTE_DIM(txq_ctrl->txq.mp2mr)); ++i) {
- if (txq_ctrl->txq.mp2mr[i].mp == NULL)
+ if (txq_ctrl->txq.mp2mr[i].mr == NULL)
break;
- assert(txq_ctrl->txq.mp2mr[i].mr != NULL);
claim_zero(ibv_dereg_mr(txq_ctrl->txq.mp2mr[i].mr));
}
memset(txq_ctrl, 0, sizeof(*txq_ctrl));
static inline int
txq_setup(struct txq_ctrl *tmpl, struct txq_ctrl *txq_ctrl)
{
- struct mlx5_qp *qp = to_mqp(tmpl->qp);
+ struct mlx5dv_qp qp;
struct ibv_cq *ibcq = tmpl->cq;
- struct mlx5_cq *cq = to_mxxx(cq, cq);
+ struct mlx5dv_cq cq_info;
+ struct mlx5dv_obj obj;
+ int ret = 0;
- if (cq->cqe_sz != RTE_CACHE_LINE_SIZE) {
+ qp.comp_mask = MLX5DV_QP_MASK_UAR_MMAP_OFFSET;
+ obj.cq.in = ibcq;
+ obj.cq.out = &cq_info;
+ obj.qp.in = tmpl->qp;
+ obj.qp.out = &qp;
+ ret = mlx5dv_init_obj(&obj, MLX5DV_OBJ_CQ | MLX5DV_OBJ_QP);
+ if (ret != 0) {
+ return -EINVAL;
+ }
+ if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
ERROR("Wrong MLX5_CQE_SIZE environment variable value: "
"it should be set to %u", RTE_CACHE_LINE_SIZE);
return EINVAL;
}
- tmpl->txq.cqe_n = ibcq->cqe + 1;
- tmpl->txq.qp_num_8s = qp->ctrl_seg.qp_num << 8;
- tmpl->txq.wqes =
- (volatile union mlx5_wqe (*)[])
- (uintptr_t)qp->gen_data.sqstart;
- tmpl->txq.wqe_n = qp->sq.wqe_cnt;
- tmpl->txq.qp_db = &qp->gen_data.db[MLX5_SND_DBR];
- tmpl->txq.bf_reg = qp->gen_data.bf->reg;
- tmpl->txq.bf_offset = qp->gen_data.bf->offset;
- tmpl->txq.bf_buf_size = qp->gen_data.bf->buf_size;
- tmpl->txq.cq_db = cq->dbrec;
+ tmpl->txq.cqe_n = log2above(cq_info.cqe_cnt);
+ tmpl->txq.qp_num_8s = tmpl->qp->qp_num << 8;
+ tmpl->txq.wqes = qp.sq.buf;
+ tmpl->txq.wqe_n = log2above(qp.sq.wqe_cnt);
+ tmpl->txq.qp_db = &qp.dbrec[MLX5_SND_DBR];
+ tmpl->txq.bf_reg = qp.bf.reg;
+ tmpl->txq.cq_db = cq_info.dbrec;
tmpl->txq.cqes =
(volatile struct mlx5_cqe (*)[])
- (uintptr_t)cq->active_buf->buf;
+ (uintptr_t)cq_info.buf;
tmpl->txq.elts =
- (struct rte_mbuf *(*)[tmpl->txq.elts_n])
+ (struct rte_mbuf *(*)[1 << tmpl->txq.elts_n])
((uintptr_t)txq_ctrl + sizeof(*txq_ctrl));
+ if (qp.comp_mask | MLX5DV_QP_MASK_UAR_MMAP_OFFSET) {
+ tmpl->uar_mmap_offset = qp.uar_mmap_offset;
+ } else {
+ ERROR("Failed to retrieve UAR info, invalid libmlx5.so version");
+ return EINVAL;
+ }
+
return 0;
}
.socket = socket,
};
union {
- struct ibv_exp_query_intf_params params;
- struct ibv_exp_qp_init_attr init;
- struct ibv_exp_res_domain_init_attr rd;
- struct ibv_exp_cq_init_attr cq;
- struct ibv_exp_qp_attr mod;
- struct ibv_exp_cq_attr cq_attr;
+ struct ibv_qp_init_attr_ex init;
+ struct ibv_cq_init_attr_ex cq;
+ struct ibv_qp_attr mod;
+ struct ibv_cq_ex cq_attr;
} attr;
- enum ibv_exp_query_intf_status status;
+ unsigned int cqe_n;
+ const unsigned int max_tso_inline = ((MLX5_MAX_TSO_HEADER +
+ (RTE_CACHE_LINE_SIZE - 1)) /
+ RTE_CACHE_LINE_SIZE);
int ret = 0;
if (mlx5_getenv_int("MLX5_ENABLE_CQE_COMPRESSION")) {
ERROR("MLX5_ENABLE_CQE_COMPRESSION must never be set");
goto error;
}
- (void)conf; /* Thresholds configuration (ignored). */
+ tmpl.txq.flags = conf->txq_flags;
assert(desc > MLX5_TX_COMP_THRESH);
- tmpl.txq.elts_n = desc;
+ tmpl.txq.elts_n = log2above(desc);
+ if (priv->mps == MLX5_MPW_ENHANCED)
+ tmpl.txq.mpw_hdr_dseg = priv->mpw_hdr_dseg;
/* MRs will be registered in mp2mr[] later. */
- attr.rd = (struct ibv_exp_res_domain_init_attr){
- .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
- IBV_EXP_RES_DOMAIN_MSG_MODEL),
- .thread_model = IBV_EXP_THREAD_SINGLE,
- .msg_model = IBV_EXP_MSG_HIGH_BW,
- };
- tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
- if (tmpl.rd == NULL) {
- ret = ENOMEM;
- ERROR("%p: RD creation failure: %s",
- (void *)dev, strerror(ret));
- goto error;
- }
- attr.cq = (struct ibv_exp_cq_init_attr){
- .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
- .res_domain = tmpl.rd,
+ attr.cq = (struct ibv_cq_init_attr_ex){
+ .comp_mask = 0,
};
- tmpl.cq = ibv_exp_create_cq(priv->ctx,
- (((desc / MLX5_TX_COMP_THRESH) - 1) ?
- ((desc / MLX5_TX_COMP_THRESH) - 1) : 1),
- NULL, NULL, 0, &attr.cq);
+ cqe_n = ((desc / MLX5_TX_COMP_THRESH) - 1) ?
+ ((desc / MLX5_TX_COMP_THRESH) - 1) : 1;
+ if (priv->mps == MLX5_MPW_ENHANCED)
+ cqe_n += MLX5_TX_COMP_THRESH_INLINE_DIV;
+ tmpl.cq = ibv_create_cq(priv->ctx,
+ cqe_n,
+ NULL, NULL, 0);
if (tmpl.cq == NULL) {
ret = ENOMEM;
ERROR("%p: CQ creation failure: %s",
goto error;
}
DEBUG("priv->device_attr.max_qp_wr is %d",
- priv->device_attr.max_qp_wr);
+ priv->device_attr.orig_attr.max_qp_wr);
DEBUG("priv->device_attr.max_sge is %d",
- priv->device_attr.max_sge);
- attr.init = (struct ibv_exp_qp_init_attr){
+ priv->device_attr.orig_attr.max_sge);
+ attr.init = (struct ibv_qp_init_attr_ex){
/* CQ to be associated with the send queue. */
.send_cq = tmpl.cq,
/* CQ to be associated with the receive queue. */
.recv_cq = tmpl.cq,
.cap = {
/* Max number of outstanding WRs. */
- .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
- priv->device_attr.max_qp_wr :
- desc),
+ .max_send_wr =
+ ((priv->device_attr.orig_attr.max_qp_wr < desc) ?
+ priv->device_attr.orig_attr.max_qp_wr :
+ desc),
/*
* Max number of scatter/gather elements in a WR,
* must be 1 to prevent libmlx5 from trying to affect
* TX burst. */
.sq_sig_all = 0,
.pd = priv->pd,
- .res_domain = tmpl.rd,
- .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
- IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
+ .comp_mask = IBV_QP_INIT_ATTR_PD,
};
if (priv->txq_inline && (priv->txqs_n >= priv->txqs_inline)) {
+ unsigned int ds_cnt;
+
tmpl.txq.max_inline =
((priv->txq_inline + (RTE_CACHE_LINE_SIZE - 1)) /
RTE_CACHE_LINE_SIZE);
- attr.init.cap.max_inline_data =
- tmpl.txq.max_inline * RTE_CACHE_LINE_SIZE;
+ tmpl.txq.inline_en = 1;
+ /* TSO and MPS can't be enabled concurrently. */
+ assert(!priv->tso || !priv->mps);
+ if (priv->mps == MLX5_MPW_ENHANCED) {
+ tmpl.txq.inline_max_packet_sz =
+ priv->inline_max_packet_sz;
+ /* To minimize the size of data set, avoid requesting
+ * too large WQ.
+ */
+ attr.init.cap.max_inline_data =
+ ((RTE_MIN(priv->txq_inline,
+ priv->inline_max_packet_sz) +
+ (RTE_CACHE_LINE_SIZE - 1)) /
+ RTE_CACHE_LINE_SIZE) * RTE_CACHE_LINE_SIZE;
+ } else if (priv->tso) {
+ int inline_diff = tmpl.txq.max_inline - max_tso_inline;
+
+ /*
+ * Adjust inline value as Verbs aggregates
+ * tso_inline and txq_inline fields.
+ */
+ attr.init.cap.max_inline_data = inline_diff > 0 ?
+ inline_diff *
+ RTE_CACHE_LINE_SIZE :
+ 0;
+ } else {
+ attr.init.cap.max_inline_data =
+ tmpl.txq.max_inline * RTE_CACHE_LINE_SIZE;
+ }
+ /*
+ * 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)
+ */
+ ds_cnt = 2 +
+ (attr.init.cap.max_inline_data / 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);
+ WARN("txq inline is too large (%d) setting it to "
+ "the maximum possible: %d\n",
+ priv->txq_inline, max_inline);
+ tmpl.txq.max_inline = max_inline / RTE_CACHE_LINE_SIZE;
+ attr.init.cap.max_inline_data = max_inline;
+ }
+ }
+ if (priv->tso) {
+ attr.init.max_tso_header =
+ max_tso_inline * RTE_CACHE_LINE_SIZE;
+ attr.init.comp_mask |= IBV_QP_INIT_ATTR_MAX_TSO_HEADER;
+ tmpl.txq.max_inline = RTE_MAX(tmpl.txq.max_inline,
+ max_tso_inline);
+ tmpl.txq.tso_en = 1;
}
- tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
+ if (priv->tunnel_en)
+ tmpl.txq.tunnel_en = 1;
+ tmpl.qp = ibv_create_qp_ex(priv->ctx, &attr.init);
if (tmpl.qp == NULL) {
ret = (errno ? errno : EINVAL);
ERROR("%p: QP creation failure: %s",
attr.init.cap.max_send_wr,
attr.init.cap.max_send_sge,
attr.init.cap.max_inline_data);
- attr.mod = (struct ibv_exp_qp_attr){
+ attr.mod = (struct ibv_qp_attr){
/* Move the QP to this state. */
.qp_state = IBV_QPS_INIT,
/* Primary port number. */
.port_num = priv->port
};
- ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
- (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
+ ret = ibv_modify_qp(tmpl.qp, &attr.mod,
+ (IBV_QP_STATE | IBV_QP_PORT));
if (ret) {
ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
(void *)dev, strerror(ret));
goto error;
}
txq_alloc_elts(&tmpl, desc);
- attr.mod = (struct ibv_exp_qp_attr){
+ attr.mod = (struct ibv_qp_attr){
.qp_state = IBV_QPS_RTR
};
- ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
+ ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
if (ret) {
ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
(void *)dev, strerror(ret));
goto error;
}
attr.mod.qp_state = IBV_QPS_RTS;
- ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
+ ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
if (ret) {
ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
(void *)dev, strerror(ret));
goto error;
}
- attr.params = (struct ibv_exp_query_intf_params){
- .intf_scope = IBV_EXP_INTF_GLOBAL,
- .intf = IBV_EXP_INTF_CQ,
- .obj = tmpl.cq,
- };
- tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
- if (tmpl.if_cq == NULL) {
- ret = EINVAL;
- ERROR("%p: CQ interface family query failed with status %d",
- (void *)dev, status);
- goto error;
- }
- attr.params = (struct ibv_exp_query_intf_params){
- .intf_scope = IBV_EXP_INTF_GLOBAL,
- .intf = IBV_EXP_INTF_QP_BURST,
- .intf_version = 1,
- .obj = tmpl.qp,
- /* Enable multi-packet send if supported. */
- .family_flags =
- ((priv->mps && !priv->sriov) ?
- IBV_EXP_QP_BURST_CREATE_ENABLE_MULTI_PACKET_SEND_WR :
- 0),
- };
- tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
- if (tmpl.if_qp == NULL) {
- ret = EINVAL;
- ERROR("%p: QP interface family query failed with status %d",
- (void *)dev, status);
- goto error;
- }
/* Clean up txq in case we're reinitializing it. */
DEBUG("%p: cleaning-up old txq just in case", (void *)txq_ctrl);
txq_cleanup(txq_ctrl);
}
(*priv->txqs)[idx] = NULL;
txq_cleanup(txq_ctrl);
+ /* Resize if txq size is changed. */
+ if (txq_ctrl->txq.elts_n != log2above(desc)) {
+ txq_ctrl = rte_realloc(txq_ctrl,
+ sizeof(*txq_ctrl) +
+ desc * sizeof(struct rte_mbuf *),
+ RTE_CACHE_LINE_SIZE);
+ if (!txq_ctrl) {
+ ERROR("%p: unable to reallocate queue index %u",
+ (void *)dev, idx);
+ priv_unlock(priv);
+ return -ENOMEM;
+ }
+ }
} else {
txq_ctrl =
rte_calloc_socket("TXQ", 1,
DEBUG("%p: adding TX queue %p to list",
(void *)dev, (void *)txq_ctrl);
(*priv->txqs)[idx] = &txq_ctrl->txq;
- /* Update send callback. */
- priv_select_tx_function(priv);
}
priv_unlock(priv);
return -ret;
priv_unlock(priv);
}
+
/**
- * DPDK callback for TX in secondary processes.
- *
- * This function configures all queues from primary process information
- * if necessary before reverting to the normal TX burst callback.
+ * Map locally UAR used in Tx queues for BlueFlame doorbell.
*
- * @param dpdk_txq
- * Generic pointer to TX queue structure.
- * @param[in] pkts
- * Packets to transmit.
- * @param pkts_n
- * Number of packets in array.
+ * @param[in] priv
+ * Pointer to private structure.
+ * @param fd
+ * Verbs file descriptor to map UAR pages.
*
* @return
- * Number of packets successfully transmitted (<= pkts_n).
+ * 0 on success, errno value on failure.
*/
-uint16_t
-mlx5_tx_burst_secondary_setup(void *dpdk_txq, struct rte_mbuf **pkts,
- uint16_t pkts_n)
+int
+priv_tx_uar_remap(struct priv *priv, int fd)
{
- struct txq *txq = dpdk_txq;
- struct txq_ctrl *txq_ctrl = container_of(txq, struct txq_ctrl, txq);
- struct priv *priv = mlx5_secondary_data_setup(txq_ctrl->priv);
- struct priv *primary_priv;
- unsigned int index;
+ unsigned int i, j;
+ uintptr_t pages[priv->txqs_n];
+ unsigned int pages_n = 0;
+ uintptr_t uar_va;
+ void *addr;
+ struct txq *txq;
+ struct txq_ctrl *txq_ctrl;
+ int already_mapped;
+ size_t page_size = sysconf(_SC_PAGESIZE);
- if (priv == NULL)
- return 0;
- primary_priv =
- mlx5_secondary_data[priv->dev->data->port_id].primary_priv;
- /* Look for queue index in both private structures. */
- for (index = 0; index != priv->txqs_n; ++index)
- if (((*primary_priv->txqs)[index] == txq) ||
- ((*priv->txqs)[index] == txq))
- break;
- if (index == priv->txqs_n)
- return 0;
- txq = (*priv->txqs)[index];
- return priv->dev->tx_pkt_burst(txq, pkts, pkts_n);
+ /*
+ * As rdma-core, UARs are mapped in size of OS page size.
+ * Use aligned address to avoid duplicate mmap.
+ * Ref to libmlx5 function: mlx5_init_context()
+ */
+ for (i = 0; i != priv->txqs_n; ++i) {
+ txq = (*priv->txqs)[i];
+ txq_ctrl = container_of(txq, struct txq_ctrl, txq);
+ uar_va = (uintptr_t)txq_ctrl->txq.bf_reg;
+ uar_va = RTE_ALIGN_FLOOR(uar_va, page_size);
+ already_mapped = 0;
+ for (j = 0; j != pages_n; ++j) {
+ if (pages[j] == uar_va) {
+ already_mapped = 1;
+ break;
+ }
+ }
+ if (already_mapped)
+ continue;
+ pages[pages_n++] = uar_va;
+ addr = mmap((void *)uar_va, page_size,
+ PROT_WRITE, MAP_FIXED | MAP_SHARED, fd,
+ txq_ctrl->uar_mmap_offset);
+ if (addr != (void *)uar_va) {
+ ERROR("call to mmap failed on UAR for txq %d\n", i);
+ return -1;
+ }
+ }
+ return 0;
}
git.droids-corp.org / dpdk.git / blobdiff