- char *start;
- char *end;
-};
-
-/* Called by mlx4_check_mempool() when iterating the memory chunks. */
-static void mlx4_check_mempool_cb(struct rte_mempool *mp,
- void *opaque, struct rte_mempool_memhdr *memhdr,
- unsigned mem_idx)
-{
- struct mlx4_check_mempool_data *data = opaque;
-
- (void)mp;
- (void)mem_idx;
- /* It already failed, skip the next chunks. */
- if (data->ret != 0)
- return;
- /* It is the first chunk. */
- if (data->start == NULL && data->end == NULL) {
- data->start = memhdr->addr;
- data->end = data->start + memhdr->len;
- return;
- }
- if (data->end == memhdr->addr) {
- data->end += memhdr->len;
- return;
- }
- if (data->start == (char *)memhdr->addr + memhdr->len) {
- data->start -= memhdr->len;
- return;
- }
- /* Error, mempool is not virtually contigous. */
- data->ret = -1;
-}
-
-/**
- * Check if a mempool can be used: it must be virtually contiguous.
- *
- * @param[in] mp
- * Pointer to memory pool.
- * @param[out] start
- * Pointer to the start address of the mempool virtual memory area
- * @param[out] end
- * Pointer to the end address of the mempool virtual memory area
- *
- * @return
- * 0 on success (mempool is virtually contiguous), -1 on error.
- */
-static int mlx4_check_mempool(struct rte_mempool *mp, uintptr_t *start,
- uintptr_t *end)
-{
- struct mlx4_check_mempool_data data;
-
- memset(&data, 0, sizeof(data));
- rte_mempool_mem_iter(mp, mlx4_check_mempool_cb, &data);
- *start = (uintptr_t)data.start;
- *end = (uintptr_t)data.end;
- return data.ret;
-}
-
-/**
- * Register mempool as a memory region.
- *
- * @param pd
- * Pointer to protection domain.
- * @param mp
- * Pointer to memory pool.
- *
- * @return
- * Memory region pointer, NULL in case of error and rte_errno is set.
- */
-struct ibv_mr *
-mlx4_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp)
-{
- const struct rte_memseg *ms = rte_eal_get_physmem_layout();
- uintptr_t start;
- uintptr_t end;
- unsigned int i;
- struct ibv_mr *mr;
-
- if (mlx4_check_mempool(mp, &start, &end) != 0) {
- rte_errno = EINVAL;
- ERROR("mempool %p: not virtually contiguous",
- (void *)mp);
- return NULL;
- }
- DEBUG("mempool %p area start=%p end=%p size=%zu",
- (void *)mp, (void *)start, (void *)end,
- (size_t)(end - start));
- /* Round start and end to page boundary if found in memory segments. */
- for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) {
- uintptr_t addr = (uintptr_t)ms[i].addr;
- size_t len = ms[i].len;
- unsigned int align = ms[i].hugepage_sz;
-
- if ((start > addr) && (start < addr + len))
- start = RTE_ALIGN_FLOOR(start, align);
- if ((end > addr) && (end < addr + len))
- end = RTE_ALIGN_CEIL(end, align);
- }
- DEBUG("mempool %p using start=%p end=%p size=%zu for MR",
- (void *)mp, (void *)start, (void *)end,
- (size_t)(end - start));
- mr = ibv_reg_mr(pd,
- (void *)start,
- end - start,
- IBV_ACCESS_LOCAL_WRITE);
- if (!mr)
- rte_errno = errno ? errno : EINVAL;
- return mr;
-}
-
-/* RX queues handling. */
-
-/**
- * Allocate RX queue elements.
- *
- * @param rxq
- * Pointer to RX queue structure.
- * @param elts_n
- * Number of elements to allocate.
- *
- * @return
- * 0 on success, negative errno value otherwise and rte_errno is set.
- */
-static int
-rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n)
-{
- unsigned int i;
- struct rxq_elt (*elts)[elts_n] =
- rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
- rxq->socket);
-
- if (elts == NULL) {
- rte_errno = ENOMEM;
- ERROR("%p: can't allocate packets array", (void *)rxq);
- goto error;
- }
- /* For each WR (packet). */
- for (i = 0; (i != elts_n); ++i) {
- struct rxq_elt *elt = &(*elts)[i];
- struct ibv_recv_wr *wr = &elt->wr;
- struct ibv_sge *sge = &(*elts)[i].sge;
- struct rte_mbuf *buf = rte_pktmbuf_alloc(rxq->mp);
-
- if (buf == NULL) {
- rte_errno = ENOMEM;
- ERROR("%p: empty mbuf pool", (void *)rxq);
- goto error;
- }
- elt->buf = buf;
- wr->next = &(*elts)[(i + 1)].wr;
- wr->sg_list = sge;
- wr->num_sge = 1;
- /* Headroom is reserved by rte_pktmbuf_alloc(). */
- assert(buf->data_off == RTE_PKTMBUF_HEADROOM);
- /* Buffer is supposed to be empty. */
- assert(rte_pktmbuf_data_len(buf) == 0);
- assert(rte_pktmbuf_pkt_len(buf) == 0);
- /* sge->addr must be able to store a pointer. */
- assert(sizeof(sge->addr) >= sizeof(uintptr_t));
- /* SGE keeps its headroom. */
- sge->addr = (uintptr_t)
- ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
- sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
- sge->lkey = rxq->mr->lkey;
- /* Redundant check for tailroom. */
- assert(sge->length == rte_pktmbuf_tailroom(buf));
- }
- /* The last WR pointer must be NULL. */
- (*elts)[(i - 1)].wr.next = NULL;
- DEBUG("%p: allocated and configured %u single-segment WRs",
- (void *)rxq, elts_n);
- rxq->elts_n = elts_n;
- rxq->elts_head = 0;
- rxq->elts = elts;
- return 0;
-error:
- if (elts != NULL) {
- for (i = 0; (i != RTE_DIM(*elts)); ++i)
- rte_pktmbuf_free_seg((*elts)[i].buf);
- rte_free(elts);
- }
- DEBUG("%p: failed, freed everything", (void *)rxq);
- assert(rte_errno > 0);
- return -rte_errno;
-}
-
-/**
- * Free RX queue elements.
- *
- * @param rxq
- * Pointer to RX queue structure.
- */
-static void
-rxq_free_elts(struct rxq *rxq)
-{
- unsigned int i;
- unsigned int elts_n = rxq->elts_n;
- struct rxq_elt (*elts)[elts_n] = rxq->elts;
-
- DEBUG("%p: freeing WRs", (void *)rxq);
- rxq->elts_n = 0;
- rxq->elts = NULL;
- if (elts == NULL)
- return;
- for (i = 0; (i != RTE_DIM(*elts)); ++i)
- rte_pktmbuf_free_seg((*elts)[i].buf);
- rte_free(elts);
-}
-
-/**
- * Unregister a MAC address.
- *
- * @param priv
- * Pointer to private structure.
- */
-static void
-priv_mac_addr_del(struct priv *priv)
-{
-#ifndef NDEBUG
- uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
-#endif
-
- if (!priv->mac_flow)
- return;
- DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x",
- (void *)priv,
- (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
- claim_zero(ibv_destroy_flow(priv->mac_flow));
- priv->mac_flow = NULL;
-}
-
-/**
- * Register a MAC address.
- *
- * The MAC address is registered in queue 0.
- *
- * @param priv
- * Pointer to private structure.
- *
- * @return
- * 0 on success, negative errno value otherwise and rte_errno is set.
- */
-static int
-priv_mac_addr_add(struct priv *priv)
-{
- uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
- struct rxq *rxq;
- struct ibv_flow *flow;
-
- /* If device isn't started, this is all we need to do. */
- if (!priv->started)
- return 0;
- if (priv->isolated)
- return 0;
- if (*priv->rxqs && (*priv->rxqs)[0])
- rxq = (*priv->rxqs)[0];
- else
- return 0;
-
- /* Allocate flow specification on the stack. */
- struct __attribute__((packed)) {
- struct ibv_flow_attr attr;
- struct ibv_flow_spec_eth spec;
- } data;
- struct ibv_flow_attr *attr = &data.attr;
- struct ibv_flow_spec_eth *spec = &data.spec;
-
- if (priv->mac_flow)
- priv_mac_addr_del(priv);
- /*
- * No padding must be inserted by the compiler between attr and spec.
- * This layout is expected by libibverbs.
- */
- assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
- *attr = (struct ibv_flow_attr){
- .type = IBV_FLOW_ATTR_NORMAL,
- .priority = 3,
- .num_of_specs = 1,
- .port = priv->port,
- .flags = 0
- };
- *spec = (struct ibv_flow_spec_eth){
- .type = IBV_FLOW_SPEC_ETH,
- .size = sizeof(*spec),
- .val = {
- .dst_mac = {
- (*mac)[0], (*mac)[1], (*mac)[2],
- (*mac)[3], (*mac)[4], (*mac)[5]
- },
- },
- .mask = {
- .dst_mac = "\xff\xff\xff\xff\xff\xff",
- }
- };
- DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x",
- (void *)priv,
- (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
- /* Create related flow. */
- flow = ibv_create_flow(rxq->qp, attr);
- if (flow == NULL) {
- rte_errno = errno ? errno : EINVAL;
- ERROR("%p: flow configuration failed, errno=%d: %s",
- (void *)rxq, rte_errno, strerror(errno));
- return -rte_errno;
- }
- assert(priv->mac_flow == NULL);
- priv->mac_flow = flow;
- return 0;
-}
-
-/**
- * Clean up a RX queue.
- *
- * Destroy objects, free allocated memory and reset the structure for reuse.
- *
- * @param rxq
- * Pointer to RX queue structure.
- */
-static void
-rxq_cleanup(struct rxq *rxq)
-{
- DEBUG("cleaning up %p", (void *)rxq);
- rxq_free_elts(rxq);
- if (rxq->qp != NULL)
- claim_zero(ibv_destroy_qp(rxq->qp));
- if (rxq->cq != NULL)
- claim_zero(ibv_destroy_cq(rxq->cq));
- if (rxq->channel != NULL)
- claim_zero(ibv_destroy_comp_channel(rxq->channel));
- if (rxq->mr != NULL)
- claim_zero(ibv_dereg_mr(rxq->mr));
- memset(rxq, 0, sizeof(*rxq));
-}
-
-/**
- * Allocate a Queue Pair.
- * Optionally setup inline receive if supported.
- *
- * @param priv
- * Pointer to private structure.
- * @param cq
- * Completion queue to associate with QP.
- * @param desc
- * Number of descriptors in QP (hint only).
- *
- * @return
- * QP pointer or NULL in case of error and rte_errno is set.
- */
-static struct ibv_qp *
-rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
-{
- struct ibv_qp *qp;
- struct ibv_qp_init_attr attr = {
- /* CQ to be associated with the send queue. */
- .send_cq = cq,
- /* CQ to be associated with the receive queue. */
- .recv_cq = cq,
- .cap = {
- /* Max number of outstanding WRs. */
- .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
- priv->device_attr.max_qp_wr :
- desc),
- /* Max number of scatter/gather elements in a WR. */
- .max_recv_sge = 1,
- },
- .qp_type = IBV_QPT_RAW_PACKET,
- };