4 * Copyright 2012 6WIND S.A.
5 * Copyright 2012 Mellanox
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of 6WIND S.A. nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
46 #include <sys/ioctl.h>
47 #include <sys/socket.h>
48 #include <netinet/in.h>
49 #include <linux/ethtool.h>
50 #include <linux/sockios.h>
53 #include <rte_ether.h>
54 #include <rte_ethdev.h>
55 #include <rte_ethdev_pci.h>
58 #include <rte_errno.h>
59 #include <rte_mempool.h>
60 #include <rte_prefetch.h>
61 #include <rte_malloc.h>
63 #include <rte_alarm.h>
64 #include <rte_memory.h>
66 #include <rte_kvargs.h>
67 #include <rte_interrupts.h>
68 #include <rte_branch_prediction.h>
70 /* Generated configuration header. */
71 #include "mlx4_autoconf.h"
75 #include "mlx4_flow.h"
77 /* Convenience macros for accessing mbuf fields. */
78 #define NEXT(m) ((m)->next)
79 #define DATA_LEN(m) ((m)->data_len)
80 #define PKT_LEN(m) ((m)->pkt_len)
81 #define DATA_OFF(m) ((m)->data_off)
82 #define SET_DATA_OFF(m, o) ((m)->data_off = (o))
83 #define NB_SEGS(m) ((m)->nb_segs)
84 #define PORT(m) ((m)->port)
86 /** Configuration structure for device arguments. */
89 uint32_t present; /**< Bit-field for existing ports. */
90 uint32_t enabled; /**< Bit-field for user-enabled ports. */
94 /* Available parameters list. */
95 const char *pmd_mlx4_init_params[] = {
101 mlx4_rx_intr_enable(struct rte_eth_dev *dev, uint16_t idx);
104 mlx4_rx_intr_disable(struct rte_eth_dev *dev, uint16_t idx);
107 priv_rx_intr_vec_enable(struct priv *priv);
110 priv_rx_intr_vec_disable(struct priv *priv);
112 /* Allocate a buffer on the stack and fill it with a printf format string. */
113 #define MKSTR(name, ...) \
114 char name[snprintf(NULL, 0, __VA_ARGS__) + 1]; \
116 snprintf(name, sizeof(name), __VA_ARGS__)
119 * Get interface name from private structure.
122 * Pointer to private structure.
124 * Interface name output buffer.
127 * 0 on success, negative errno value otherwise and rte_errno is set.
130 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
134 unsigned int dev_type = 0;
135 unsigned int dev_port_prev = ~0u;
136 char match[IF_NAMESIZE] = "";
139 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
147 while ((dent = readdir(dir)) != NULL) {
148 char *name = dent->d_name;
150 unsigned int dev_port;
153 if ((name[0] == '.') &&
154 ((name[1] == '\0') ||
155 ((name[1] == '.') && (name[2] == '\0'))))
158 MKSTR(path, "%s/device/net/%s/%s",
159 priv->ctx->device->ibdev_path, name,
160 (dev_type ? "dev_id" : "dev_port"));
162 file = fopen(path, "rb");
167 * Switch to dev_id when dev_port does not exist as
168 * is the case with Linux kernel versions < 3.15.
179 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
184 * Switch to dev_id when dev_port returns the same value for
185 * all ports. May happen when using a MOFED release older than
186 * 3.0 with a Linux kernel >= 3.15.
188 if (dev_port == dev_port_prev)
190 dev_port_prev = dev_port;
191 if (dev_port == (priv->port - 1u))
192 snprintf(match, sizeof(match), "%s", name);
195 if (match[0] == '\0') {
199 strncpy(*ifname, match, sizeof(*ifname));
204 * Read from sysfs entry.
207 * Pointer to private structure.
209 * Entry name relative to sysfs path.
211 * Data output buffer.
216 * Number of bytes read on success, negative errno value otherwise and
220 priv_sysfs_read(const struct priv *priv, const char *entry,
221 char *buf, size_t size)
223 char ifname[IF_NAMESIZE];
227 ret = priv_get_ifname(priv, &ifname);
231 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
234 file = fopen(path, "rb");
239 ret = fread(buf, 1, size, file);
240 if ((size_t)ret < size && ferror(file)) {
251 * Write to sysfs entry.
254 * Pointer to private structure.
256 * Entry name relative to sysfs path.
263 * Number of bytes written on success, negative errno value otherwise and
267 priv_sysfs_write(const struct priv *priv, const char *entry,
268 char *buf, size_t size)
270 char ifname[IF_NAMESIZE];
274 ret = priv_get_ifname(priv, &ifname);
278 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
281 file = fopen(path, "wb");
286 ret = fwrite(buf, 1, size, file);
287 if ((size_t)ret < size || ferror(file)) {
298 * Get unsigned long sysfs property.
301 * Pointer to private structure.
303 * Entry name relative to sysfs path.
305 * Value output buffer.
308 * 0 on success, negative errno value otherwise and rte_errno is set.
311 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
314 unsigned long value_ret;
317 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
319 DEBUG("cannot read %s value from sysfs: %s",
320 name, strerror(rte_errno));
323 value_str[ret] = '\0';
325 value_ret = strtoul(value_str, NULL, 0);
328 DEBUG("invalid %s value `%s': %s", name, value_str,
329 strerror(rte_errno));
337 * Set unsigned long sysfs property.
340 * Pointer to private structure.
342 * Entry name relative to sysfs path.
347 * 0 on success, negative errno value otherwise and rte_errno is set.
350 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
353 MKSTR(value_str, "%lu", value);
355 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
357 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
358 name, value_str, value, strerror(rte_errno));
365 * Perform ifreq ioctl() on associated Ethernet device.
368 * Pointer to private structure.
370 * Request number to pass to ioctl().
372 * Interface request structure output buffer.
375 * 0 on success, negative errno value otherwise and rte_errno is set.
378 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
380 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
387 ret = priv_get_ifname(priv, &ifr->ifr_name);
388 if (!ret && ioctl(sock, req, ifr) == -1) {
400 * Pointer to private structure.
402 * MTU value output buffer.
405 * 0 on success, negative errno value otherwise and rte_errno is set.
408 priv_get_mtu(struct priv *priv, uint16_t *mtu)
410 unsigned long ulong_mtu = 0;
411 int ret = priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu);
420 * DPDK callback to change the MTU.
423 * Pointer to Ethernet device structure.
428 * 0 on success, negative errno value otherwise and rte_errno is set.
431 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
433 struct priv *priv = dev->data->dev_private;
435 int ret = priv_set_sysfs_ulong(priv, "mtu", mtu);
439 ret = priv_get_mtu(priv, &new_mtu);
442 if (new_mtu == mtu) {
454 * Pointer to private structure.
456 * Bitmask for flags that must remain untouched.
458 * Bitmask for flags to modify.
461 * 0 on success, negative errno value otherwise and rte_errno is set.
464 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
466 unsigned long tmp = 0;
467 int ret = priv_get_sysfs_ulong(priv, "flags", &tmp);
472 tmp |= (flags & (~keep));
473 return priv_set_sysfs_ulong(priv, "flags", tmp);
476 /* Device configuration. */
479 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
480 unsigned int socket, const struct rte_eth_txconf *conf);
483 txq_cleanup(struct txq *txq);
486 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
487 unsigned int socket, const struct rte_eth_rxconf *conf,
488 struct rte_mempool *mp);
491 rxq_cleanup(struct rxq *rxq);
494 priv_mac_addr_del(struct priv *priv);
497 * DPDK callback for Ethernet device configuration.
499 * Prepare the driver for a given number of TX and RX queues.
502 * Pointer to Ethernet device structure.
505 * 0 on success, negative errno value otherwise and rte_errno is set.
508 mlx4_dev_configure(struct rte_eth_dev *dev)
510 struct priv *priv = dev->data->dev_private;
511 unsigned int rxqs_n = dev->data->nb_rx_queues;
512 unsigned int txqs_n = dev->data->nb_tx_queues;
514 priv->rxqs = (void *)dev->data->rx_queues;
515 priv->txqs = (void *)dev->data->tx_queues;
516 if (txqs_n != priv->txqs_n) {
517 INFO("%p: TX queues number update: %u -> %u",
518 (void *)dev, priv->txqs_n, txqs_n);
519 priv->txqs_n = txqs_n;
521 if (rxqs_n != priv->rxqs_n) {
522 INFO("%p: Rx queues number update: %u -> %u",
523 (void *)dev, priv->rxqs_n, rxqs_n);
524 priv->rxqs_n = rxqs_n;
529 static uint16_t mlx4_tx_burst(void *, struct rte_mbuf **, uint16_t);
530 static uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
532 /* TX queues handling. */
535 * Allocate TX queue elements.
538 * Pointer to TX queue structure.
540 * Number of elements to allocate.
543 * 0 on success, negative errno value otherwise and rte_errno is set.
546 txq_alloc_elts(struct txq *txq, unsigned int elts_n)
549 struct txq_elt (*elts)[elts_n] =
550 rte_calloc_socket("TXQ", 1, sizeof(*elts), 0, txq->socket);
554 ERROR("%p: can't allocate packets array", (void *)txq);
558 for (i = 0; (i != elts_n); ++i) {
559 struct txq_elt *elt = &(*elts)[i];
563 DEBUG("%p: allocated and configured %u WRs", (void *)txq, elts_n);
564 txq->elts_n = elts_n;
570 * Request send completion every MLX4_PMD_TX_PER_COMP_REQ packets or
571 * at least 4 times per ring.
573 txq->elts_comp_cd_init =
574 ((MLX4_PMD_TX_PER_COMP_REQ < (elts_n / 4)) ?
575 MLX4_PMD_TX_PER_COMP_REQ : (elts_n / 4));
576 txq->elts_comp_cd = txq->elts_comp_cd_init;
581 DEBUG("%p: failed, freed everything", (void *)txq);
588 * Free TX queue elements.
591 * Pointer to TX queue structure.
594 txq_free_elts(struct txq *txq)
596 unsigned int elts_n = txq->elts_n;
597 unsigned int elts_head = txq->elts_head;
598 unsigned int elts_tail = txq->elts_tail;
599 struct txq_elt (*elts)[elts_n] = txq->elts;
601 DEBUG("%p: freeing WRs", (void *)txq);
606 txq->elts_comp_cd = 0;
607 txq->elts_comp_cd_init = 0;
611 while (elts_tail != elts_head) {
612 struct txq_elt *elt = &(*elts)[elts_tail];
614 assert(elt->buf != NULL);
615 rte_pktmbuf_free(elt->buf);
618 memset(elt, 0x77, sizeof(*elt));
620 if (++elts_tail == elts_n)
627 * Clean up a TX queue.
629 * Destroy objects, free allocated memory and reset the structure for reuse.
632 * Pointer to TX queue structure.
635 txq_cleanup(struct txq *txq)
639 DEBUG("cleaning up %p", (void *)txq);
642 claim_zero(ibv_destroy_qp(txq->qp));
644 claim_zero(ibv_destroy_cq(txq->cq));
645 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
646 if (txq->mp2mr[i].mp == NULL)
648 assert(txq->mp2mr[i].mr != NULL);
649 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
651 memset(txq, 0, sizeof(*txq));
655 * Manage TX completions.
657 * When sending a burst, mlx4_tx_burst() posts several WRs.
658 * To improve performance, a completion event is only required once every
659 * MLX4_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
660 * for other WRs, but this information would not be used anyway.
663 * Pointer to TX queue structure.
666 * 0 on success, -1 on failure.
669 txq_complete(struct txq *txq)
671 unsigned int elts_comp = txq->elts_comp;
672 unsigned int elts_tail = txq->elts_tail;
673 const unsigned int elts_n = txq->elts_n;
674 struct ibv_wc wcs[elts_comp];
677 if (unlikely(elts_comp == 0))
679 wcs_n = ibv_poll_cq(txq->cq, elts_comp, wcs);
680 if (unlikely(wcs_n == 0))
682 if (unlikely(wcs_n < 0)) {
683 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
688 assert(elts_comp <= txq->elts_comp);
690 * Assume WC status is successful as nothing can be done about it
693 elts_tail += wcs_n * txq->elts_comp_cd_init;
694 if (elts_tail >= elts_n)
696 txq->elts_tail = elts_tail;
697 txq->elts_comp = elts_comp;
701 struct mlx4_check_mempool_data {
707 /* Called by mlx4_check_mempool() when iterating the memory chunks. */
708 static void mlx4_check_mempool_cb(struct rte_mempool *mp,
709 void *opaque, struct rte_mempool_memhdr *memhdr,
712 struct mlx4_check_mempool_data *data = opaque;
716 /* It already failed, skip the next chunks. */
719 /* It is the first chunk. */
720 if (data->start == NULL && data->end == NULL) {
721 data->start = memhdr->addr;
722 data->end = data->start + memhdr->len;
725 if (data->end == memhdr->addr) {
726 data->end += memhdr->len;
729 if (data->start == (char *)memhdr->addr + memhdr->len) {
730 data->start -= memhdr->len;
733 /* Error, mempool is not virtually contigous. */
738 * Check if a mempool can be used: it must be virtually contiguous.
741 * Pointer to memory pool.
743 * Pointer to the start address of the mempool virtual memory area
745 * Pointer to the end address of the mempool virtual memory area
748 * 0 on success (mempool is virtually contiguous), -1 on error.
750 static int mlx4_check_mempool(struct rte_mempool *mp, uintptr_t *start,
753 struct mlx4_check_mempool_data data;
755 memset(&data, 0, sizeof(data));
756 rte_mempool_mem_iter(mp, mlx4_check_mempool_cb, &data);
757 *start = (uintptr_t)data.start;
758 *end = (uintptr_t)data.end;
762 /* For best performance, this function should not be inlined. */
763 static struct ibv_mr *mlx4_mp2mr(struct ibv_pd *, struct rte_mempool *)
767 * Register mempool as a memory region.
770 * Pointer to protection domain.
772 * Pointer to memory pool.
775 * Memory region pointer, NULL in case of error and rte_errno is set.
777 static struct ibv_mr *
778 mlx4_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp)
780 const struct rte_memseg *ms = rte_eal_get_physmem_layout();
786 if (mlx4_check_mempool(mp, &start, &end) != 0) {
788 ERROR("mempool %p: not virtually contiguous",
792 DEBUG("mempool %p area start=%p end=%p size=%zu",
793 (void *)mp, (void *)start, (void *)end,
794 (size_t)(end - start));
795 /* Round start and end to page boundary if found in memory segments. */
796 for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) {
797 uintptr_t addr = (uintptr_t)ms[i].addr;
798 size_t len = ms[i].len;
799 unsigned int align = ms[i].hugepage_sz;
801 if ((start > addr) && (start < addr + len))
802 start = RTE_ALIGN_FLOOR(start, align);
803 if ((end > addr) && (end < addr + len))
804 end = RTE_ALIGN_CEIL(end, align);
806 DEBUG("mempool %p using start=%p end=%p size=%zu for MR",
807 (void *)mp, (void *)start, (void *)end,
808 (size_t)(end - start));
812 IBV_ACCESS_LOCAL_WRITE);
814 rte_errno = errno ? errno : EINVAL;
819 * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
820 * the cloned mbuf is allocated is returned instead.
826 * Memory pool where data is located for given mbuf.
828 static struct rte_mempool *
829 txq_mb2mp(struct rte_mbuf *buf)
831 if (unlikely(RTE_MBUF_INDIRECT(buf)))
832 return rte_mbuf_from_indirect(buf)->pool;
837 * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
838 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
839 * remove an entry first.
842 * Pointer to TX queue structure.
844 * Memory Pool for which a Memory Region lkey must be returned.
847 * mr->lkey on success, (uint32_t)-1 on failure.
850 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
855 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
856 if (unlikely(txq->mp2mr[i].mp == NULL)) {
857 /* Unknown MP, add a new MR for it. */
860 if (txq->mp2mr[i].mp == mp) {
861 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
862 assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
863 return txq->mp2mr[i].lkey;
866 /* Add a new entry, register MR first. */
867 DEBUG("%p: discovered new memory pool \"%s\" (%p)",
868 (void *)txq, mp->name, (void *)mp);
869 mr = mlx4_mp2mr(txq->priv->pd, mp);
870 if (unlikely(mr == NULL)) {
871 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
875 if (unlikely(i == elemof(txq->mp2mr))) {
876 /* Table is full, remove oldest entry. */
877 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
880 claim_zero(ibv_dereg_mr(txq->mp2mr[0].mr));
881 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
882 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
884 /* Store the new entry. */
885 txq->mp2mr[i].mp = mp;
886 txq->mp2mr[i].mr = mr;
887 txq->mp2mr[i].lkey = mr->lkey;
888 DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
889 (void *)txq, mp->name, (void *)mp, txq->mp2mr[i].lkey);
890 return txq->mp2mr[i].lkey;
893 struct txq_mp2mr_mbuf_check_data {
898 * Callback function for rte_mempool_obj_iter() to check whether a given
899 * mempool object looks like a mbuf.
902 * The mempool pointer
904 * Context data (struct txq_mp2mr_mbuf_check_data). Contains the
909 * Object index, unused.
912 txq_mp2mr_mbuf_check(struct rte_mempool *mp, void *arg, void *obj,
913 uint32_t index __rte_unused)
915 struct txq_mp2mr_mbuf_check_data *data = arg;
916 struct rte_mbuf *buf = obj;
919 * Check whether mbuf structure fits element size and whether mempool
922 if (sizeof(*buf) > mp->elt_size || buf->pool != mp)
927 * Iterator function for rte_mempool_walk() to register existing mempools and
928 * fill the MP to MR cache of a TX queue.
931 * Memory Pool to register.
933 * Pointer to TX queue structure.
936 txq_mp2mr_iter(struct rte_mempool *mp, void *arg)
938 struct txq *txq = arg;
939 struct txq_mp2mr_mbuf_check_data data = {
943 /* Register mempool only if the first element looks like a mbuf. */
944 if (rte_mempool_obj_iter(mp, txq_mp2mr_mbuf_check, &data) == 0 ||
951 * DPDK callback for TX.
954 * Generic pointer to TX queue structure.
956 * Packets to transmit.
958 * Number of packets in array.
961 * Number of packets successfully transmitted (<= pkts_n).
964 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
966 struct txq *txq = (struct txq *)dpdk_txq;
967 struct ibv_send_wr *wr_head = NULL;
968 struct ibv_send_wr **wr_next = &wr_head;
969 struct ibv_send_wr *wr_bad = NULL;
970 unsigned int elts_head = txq->elts_head;
971 const unsigned int elts_n = txq->elts_n;
972 unsigned int elts_comp_cd = txq->elts_comp_cd;
973 unsigned int elts_comp = 0;
978 assert(elts_comp_cd != 0);
980 max = (elts_n - (elts_head - txq->elts_tail));
984 assert(max <= elts_n);
985 /* Always leave one free entry in the ring. */
991 for (i = 0; (i != max); ++i) {
992 struct rte_mbuf *buf = pkts[i];
993 unsigned int elts_head_next =
994 (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
995 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
996 struct txq_elt *elt = &(*txq->elts)[elts_head];
997 struct ibv_send_wr *wr = &elt->wr;
998 unsigned int segs = NB_SEGS(buf);
999 unsigned int sent_size = 0;
1000 uint32_t send_flags = 0;
1002 /* Clean up old buffer. */
1003 if (likely(elt->buf != NULL)) {
1004 struct rte_mbuf *tmp = elt->buf;
1008 memset(elt, 0x66, sizeof(*elt));
1010 /* Faster than rte_pktmbuf_free(). */
1012 struct rte_mbuf *next = NEXT(tmp);
1014 rte_pktmbuf_free_seg(tmp);
1016 } while (tmp != NULL);
1018 /* Request TX completion. */
1019 if (unlikely(--elts_comp_cd == 0)) {
1020 elts_comp_cd = txq->elts_comp_cd_init;
1022 send_flags |= IBV_SEND_SIGNALED;
1024 if (likely(segs == 1)) {
1025 struct ibv_sge *sge = &elt->sge;
1030 /* Retrieve buffer information. */
1031 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1032 length = DATA_LEN(buf);
1033 /* Retrieve Memory Region key for this memory pool. */
1034 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
1035 if (unlikely(lkey == (uint32_t)-1)) {
1036 /* MR does not exist. */
1037 DEBUG("%p: unable to get MP <-> MR"
1038 " association", (void *)txq);
1039 /* Clean up TX element. */
1043 /* Update element. */
1046 rte_prefetch0((volatile void *)
1048 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1050 sge->length = length;
1052 sent_size += length;
1057 if (sent_size <= txq->max_inline)
1058 send_flags |= IBV_SEND_INLINE;
1059 elts_head = elts_head_next;
1060 /* Increment sent bytes counter. */
1061 txq->stats.obytes += sent_size;
1063 wr->sg_list = &elt->sge;
1065 wr->opcode = IBV_WR_SEND;
1066 wr->send_flags = send_flags;
1068 wr_next = &wr->next;
1071 /* Take a shortcut if nothing must be sent. */
1072 if (unlikely(i == 0))
1074 /* Increment sent packets counter. */
1075 txq->stats.opackets += i;
1076 /* Ring QP doorbell. */
1079 err = ibv_post_send(txq->qp, wr_head, &wr_bad);
1080 if (unlikely(err)) {
1081 uint64_t obytes = 0;
1082 uint64_t opackets = 0;
1084 /* Rewind bad WRs. */
1085 while (wr_bad != NULL) {
1088 /* Force completion request if one was lost. */
1089 if (wr_bad->send_flags & IBV_SEND_SIGNALED) {
1094 for (j = 0; j < wr_bad->num_sge; ++j)
1095 obytes += wr_bad->sg_list[j].length;
1096 elts_head = (elts_head ? elts_head : elts_n) - 1;
1097 wr_bad = wr_bad->next;
1099 txq->stats.opackets -= opackets;
1100 txq->stats.obytes -= obytes;
1102 DEBUG("%p: ibv_post_send() failed, %" PRIu64 " packets"
1103 " (%" PRIu64 " bytes) rejected: %s",
1107 (err <= -1) ? "Internal error" : strerror(err));
1109 txq->elts_head = elts_head;
1110 txq->elts_comp += elts_comp;
1111 txq->elts_comp_cd = elts_comp_cd;
1116 * Configure a TX queue.
1119 * Pointer to Ethernet device structure.
1121 * Pointer to TX queue structure.
1123 * Number of descriptors to configure in queue.
1125 * NUMA socket on which memory must be allocated.
1127 * Thresholds parameters.
1130 * 0 on success, negative errno value otherwise and rte_errno is set.
1133 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1134 unsigned int socket, const struct rte_eth_txconf *conf)
1136 struct priv *priv = dev->data->dev_private;
1142 struct ibv_qp_init_attr init;
1143 struct ibv_qp_attr mod;
1147 (void)conf; /* Thresholds configuration (ignored). */
1154 ERROR("%p: invalid number of Tx descriptors", (void *)dev);
1157 /* MRs will be registered in mp2mr[] later. */
1158 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
1159 if (tmpl.cq == NULL) {
1161 ERROR("%p: CQ creation failure: %s",
1162 (void *)dev, strerror(rte_errno));
1165 DEBUG("priv->device_attr.max_qp_wr is %d",
1166 priv->device_attr.max_qp_wr);
1167 DEBUG("priv->device_attr.max_sge is %d",
1168 priv->device_attr.max_sge);
1169 attr.init = (struct ibv_qp_init_attr){
1170 /* CQ to be associated with the send queue. */
1172 /* CQ to be associated with the receive queue. */
1175 /* Max number of outstanding WRs. */
1176 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1177 priv->device_attr.max_qp_wr :
1179 /* Max number of scatter/gather elements in a WR. */
1181 .max_inline_data = MLX4_PMD_MAX_INLINE,
1183 .qp_type = IBV_QPT_RAW_PACKET,
1185 * Do *NOT* enable this, completions events are managed per
1190 tmpl.qp = ibv_create_qp(priv->pd, &attr.init);
1191 if (tmpl.qp == NULL) {
1192 rte_errno = errno ? errno : EINVAL;
1193 ERROR("%p: QP creation failure: %s",
1194 (void *)dev, strerror(rte_errno));
1197 /* ibv_create_qp() updates this value. */
1198 tmpl.max_inline = attr.init.cap.max_inline_data;
1199 attr.mod = (struct ibv_qp_attr){
1200 /* Move the QP to this state. */
1201 .qp_state = IBV_QPS_INIT,
1202 /* Primary port number. */
1203 .port_num = priv->port
1205 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE | IBV_QP_PORT);
1208 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1209 (void *)dev, strerror(rte_errno));
1212 ret = txq_alloc_elts(&tmpl, desc);
1215 ERROR("%p: TXQ allocation failed: %s",
1216 (void *)dev, strerror(rte_errno));
1219 attr.mod = (struct ibv_qp_attr){
1220 .qp_state = IBV_QPS_RTR
1222 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1225 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1226 (void *)dev, strerror(rte_errno));
1229 attr.mod.qp_state = IBV_QPS_RTS;
1230 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1233 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1234 (void *)dev, strerror(rte_errno));
1237 /* Clean up txq in case we're reinitializing it. */
1238 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1241 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1242 /* Pre-register known mempools. */
1243 rte_mempool_walk(txq_mp2mr_iter, txq);
1249 assert(rte_errno > 0);
1254 * DPDK callback to configure a TX queue.
1257 * Pointer to Ethernet device structure.
1261 * Number of descriptors to configure in queue.
1263 * NUMA socket on which memory must be allocated.
1265 * Thresholds parameters.
1268 * 0 on success, negative errno value otherwise and rte_errno is set.
1271 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1272 unsigned int socket, const struct rte_eth_txconf *conf)
1274 struct priv *priv = dev->data->dev_private;
1275 struct txq *txq = (*priv->txqs)[idx];
1278 DEBUG("%p: configuring queue %u for %u descriptors",
1279 (void *)dev, idx, desc);
1280 if (idx >= priv->txqs_n) {
1281 rte_errno = EOVERFLOW;
1282 ERROR("%p: queue index out of range (%u >= %u)",
1283 (void *)dev, idx, priv->txqs_n);
1287 DEBUG("%p: reusing already allocated queue index %u (%p)",
1288 (void *)dev, idx, (void *)txq);
1289 if (priv->started) {
1293 (*priv->txqs)[idx] = NULL;
1296 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1299 ERROR("%p: unable to allocate queue index %u",
1304 ret = txq_setup(dev, txq, desc, socket, conf);
1308 txq->stats.idx = idx;
1309 DEBUG("%p: adding TX queue %p to list",
1310 (void *)dev, (void *)txq);
1311 (*priv->txqs)[idx] = txq;
1312 /* Update send callback. */
1313 dev->tx_pkt_burst = mlx4_tx_burst;
1319 * DPDK callback to release a TX queue.
1322 * Generic TX queue pointer.
1325 mlx4_tx_queue_release(void *dpdk_txq)
1327 struct txq *txq = (struct txq *)dpdk_txq;
1334 for (i = 0; (i != priv->txqs_n); ++i)
1335 if ((*priv->txqs)[i] == txq) {
1336 DEBUG("%p: removing TX queue %p from list",
1337 (void *)priv->dev, (void *)txq);
1338 (*priv->txqs)[i] = NULL;
1345 /* RX queues handling. */
1348 * Allocate RX queue elements.
1351 * Pointer to RX queue structure.
1353 * Number of elements to allocate.
1356 * 0 on success, negative errno value otherwise and rte_errno is set.
1359 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n)
1362 struct rxq_elt (*elts)[elts_n] =
1363 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1368 ERROR("%p: can't allocate packets array", (void *)rxq);
1371 /* For each WR (packet). */
1372 for (i = 0; (i != elts_n); ++i) {
1373 struct rxq_elt *elt = &(*elts)[i];
1374 struct ibv_recv_wr *wr = &elt->wr;
1375 struct ibv_sge *sge = &(*elts)[i].sge;
1376 struct rte_mbuf *buf = rte_pktmbuf_alloc(rxq->mp);
1380 ERROR("%p: empty mbuf pool", (void *)rxq);
1384 wr->next = &(*elts)[(i + 1)].wr;
1387 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1388 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1389 /* Buffer is supposed to be empty. */
1390 assert(rte_pktmbuf_data_len(buf) == 0);
1391 assert(rte_pktmbuf_pkt_len(buf) == 0);
1392 /* sge->addr must be able to store a pointer. */
1393 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1394 /* SGE keeps its headroom. */
1395 sge->addr = (uintptr_t)
1396 ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
1397 sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
1398 sge->lkey = rxq->mr->lkey;
1399 /* Redundant check for tailroom. */
1400 assert(sge->length == rte_pktmbuf_tailroom(buf));
1402 /* The last WR pointer must be NULL. */
1403 (*elts)[(i - 1)].wr.next = NULL;
1404 DEBUG("%p: allocated and configured %u single-segment WRs",
1405 (void *)rxq, elts_n);
1406 rxq->elts_n = elts_n;
1412 for (i = 0; (i != elemof(*elts)); ++i)
1413 rte_pktmbuf_free_seg((*elts)[i].buf);
1416 DEBUG("%p: failed, freed everything", (void *)rxq);
1417 assert(rte_errno > 0);
1422 * Free RX queue elements.
1425 * Pointer to RX queue structure.
1428 rxq_free_elts(struct rxq *rxq)
1431 unsigned int elts_n = rxq->elts_n;
1432 struct rxq_elt (*elts)[elts_n] = rxq->elts;
1434 DEBUG("%p: freeing WRs", (void *)rxq);
1439 for (i = 0; (i != elemof(*elts)); ++i)
1440 rte_pktmbuf_free_seg((*elts)[i].buf);
1445 * Unregister a MAC address.
1448 * Pointer to private structure.
1451 priv_mac_addr_del(struct priv *priv)
1454 uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
1457 if (!priv->mac_flow)
1459 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x",
1461 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
1462 claim_zero(ibv_destroy_flow(priv->mac_flow));
1463 priv->mac_flow = NULL;
1467 * Register a MAC address.
1469 * The MAC address is registered in queue 0.
1472 * Pointer to private structure.
1475 * 0 on success, negative errno value otherwise and rte_errno is set.
1478 priv_mac_addr_add(struct priv *priv)
1480 uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
1482 struct ibv_flow *flow;
1484 /* If device isn't started, this is all we need to do. */
1489 if (*priv->rxqs && (*priv->rxqs)[0])
1490 rxq = (*priv->rxqs)[0];
1494 /* Allocate flow specification on the stack. */
1495 struct __attribute__((packed)) {
1496 struct ibv_flow_attr attr;
1497 struct ibv_flow_spec_eth spec;
1499 struct ibv_flow_attr *attr = &data.attr;
1500 struct ibv_flow_spec_eth *spec = &data.spec;
1503 priv_mac_addr_del(priv);
1505 * No padding must be inserted by the compiler between attr and spec.
1506 * This layout is expected by libibverbs.
1508 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
1509 *attr = (struct ibv_flow_attr){
1510 .type = IBV_FLOW_ATTR_NORMAL,
1516 *spec = (struct ibv_flow_spec_eth){
1517 .type = IBV_FLOW_SPEC_ETH,
1518 .size = sizeof(*spec),
1521 (*mac)[0], (*mac)[1], (*mac)[2],
1522 (*mac)[3], (*mac)[4], (*mac)[5]
1526 .dst_mac = "\xff\xff\xff\xff\xff\xff",
1529 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x",
1531 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
1532 /* Create related flow. */
1533 flow = ibv_create_flow(rxq->qp, attr);
1535 rte_errno = errno ? errno : EINVAL;
1536 ERROR("%p: flow configuration failed, errno=%d: %s",
1537 (void *)rxq, rte_errno, strerror(errno));
1540 assert(priv->mac_flow == NULL);
1541 priv->mac_flow = flow;
1546 * Clean up a RX queue.
1548 * Destroy objects, free allocated memory and reset the structure for reuse.
1551 * Pointer to RX queue structure.
1554 rxq_cleanup(struct rxq *rxq)
1556 DEBUG("cleaning up %p", (void *)rxq);
1558 if (rxq->qp != NULL)
1559 claim_zero(ibv_destroy_qp(rxq->qp));
1560 if (rxq->cq != NULL)
1561 claim_zero(ibv_destroy_cq(rxq->cq));
1562 if (rxq->channel != NULL)
1563 claim_zero(ibv_destroy_comp_channel(rxq->channel));
1564 if (rxq->mr != NULL)
1565 claim_zero(ibv_dereg_mr(rxq->mr));
1566 memset(rxq, 0, sizeof(*rxq));
1570 * DPDK callback for RX.
1572 * The following function doesn't manage scattered packets.
1575 * Generic pointer to RX queue structure.
1577 * Array to store received packets.
1579 * Maximum number of packets in array.
1582 * Number of packets successfully received (<= pkts_n).
1585 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1587 struct rxq *rxq = (struct rxq *)dpdk_rxq;
1588 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts;
1589 const unsigned int elts_n = rxq->elts_n;
1590 unsigned int elts_head = rxq->elts_head;
1591 struct ibv_wc wcs[pkts_n];
1592 struct ibv_recv_wr *wr_head = NULL;
1593 struct ibv_recv_wr **wr_next = &wr_head;
1594 struct ibv_recv_wr *wr_bad = NULL;
1596 unsigned int pkts_ret = 0;
1599 ret = ibv_poll_cq(rxq->cq, pkts_n, wcs);
1600 if (unlikely(ret == 0))
1602 if (unlikely(ret < 0)) {
1603 DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
1607 assert(ret <= (int)pkts_n);
1608 /* For each work completion. */
1609 for (i = 0; i != (unsigned int)ret; ++i) {
1610 struct ibv_wc *wc = &wcs[i];
1611 struct rxq_elt *elt = &(*elts)[elts_head];
1612 struct ibv_recv_wr *wr = &elt->wr;
1613 uint32_t len = wc->byte_len;
1614 struct rte_mbuf *seg = elt->buf;
1615 struct rte_mbuf *rep;
1617 /* Sanity checks. */
1618 assert(wr->sg_list == &elt->sge);
1619 assert(wr->num_sge == 1);
1620 assert(elts_head < rxq->elts_n);
1621 assert(rxq->elts_head < rxq->elts_n);
1623 * Fetch initial bytes of packet descriptor into a
1624 * cacheline while allocating rep.
1626 rte_mbuf_prefetch_part1(seg);
1627 rte_mbuf_prefetch_part2(seg);
1628 /* Link completed WRs together for repost. */
1630 wr_next = &wr->next;
1631 if (unlikely(wc->status != IBV_WC_SUCCESS)) {
1632 /* Whatever, just repost the offending WR. */
1633 DEBUG("rxq=%p: bad work completion status (%d): %s",
1634 (void *)rxq, wc->status,
1635 ibv_wc_status_str(wc->status));
1636 /* Increment dropped packets counter. */
1637 ++rxq->stats.idropped;
1640 rep = rte_mbuf_raw_alloc(rxq->mp);
1641 if (unlikely(rep == NULL)) {
1643 * Unable to allocate a replacement mbuf,
1646 DEBUG("rxq=%p: can't allocate a new mbuf",
1648 /* Increase out of memory counters. */
1649 ++rxq->stats.rx_nombuf;
1650 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
1653 /* Reconfigure sge to use rep instead of seg. */
1654 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
1655 assert(elt->sge.lkey == rxq->mr->lkey);
1657 /* Update seg information. */
1658 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
1660 PORT(seg) = rxq->port_id;
1663 DATA_LEN(seg) = len;
1664 seg->packet_type = 0;
1666 /* Return packet. */
1669 /* Increase bytes counter. */
1670 rxq->stats.ibytes += len;
1672 if (++elts_head >= elts_n)
1676 if (unlikely(i == 0))
1681 ret = ibv_post_recv(rxq->qp, wr_head, &wr_bad);
1682 if (unlikely(ret)) {
1683 /* Inability to repost WRs is fatal. */
1684 DEBUG("%p: recv_burst(): failed (ret=%d)",
1689 rxq->elts_head = elts_head;
1690 /* Increase packets counter. */
1691 rxq->stats.ipackets += pkts_ret;
1696 * Allocate a Queue Pair.
1697 * Optionally setup inline receive if supported.
1700 * Pointer to private structure.
1702 * Completion queue to associate with QP.
1704 * Number of descriptors in QP (hint only).
1707 * QP pointer or NULL in case of error and rte_errno is set.
1709 static struct ibv_qp *
1710 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
1713 struct ibv_qp_init_attr attr = {
1714 /* CQ to be associated with the send queue. */
1716 /* CQ to be associated with the receive queue. */
1719 /* Max number of outstanding WRs. */
1720 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
1721 priv->device_attr.max_qp_wr :
1723 /* Max number of scatter/gather elements in a WR. */
1726 .qp_type = IBV_QPT_RAW_PACKET,
1729 qp = ibv_create_qp(priv->pd, &attr);
1731 rte_errno = errno ? errno : EINVAL;
1736 * Configure a RX queue.
1739 * Pointer to Ethernet device structure.
1741 * Pointer to RX queue structure.
1743 * Number of descriptors to configure in queue.
1745 * NUMA socket on which memory must be allocated.
1747 * Thresholds parameters.
1749 * Memory pool for buffer allocations.
1752 * 0 on success, negative errno value otherwise and rte_errno is set.
1755 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
1756 unsigned int socket, const struct rte_eth_rxconf *conf,
1757 struct rte_mempool *mp)
1759 struct priv *priv = dev->data->dev_private;
1765 struct ibv_qp_attr mod;
1766 struct ibv_recv_wr *bad_wr;
1767 unsigned int mb_len;
1770 (void)conf; /* Thresholds configuration (ignored). */
1771 mb_len = rte_pktmbuf_data_room_size(mp);
1774 ERROR("%p: invalid number of Rx descriptors", (void *)dev);
1777 /* Enable scattered packets support for this queue if necessary. */
1778 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
1779 if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
1780 (mb_len - RTE_PKTMBUF_HEADROOM)) {
1782 } else if (dev->data->dev_conf.rxmode.enable_scatter) {
1783 WARN("%p: scattered mode has been requested but is"
1784 " not supported, this may lead to packet loss",
1787 WARN("%p: the requested maximum Rx packet size (%u) is"
1788 " larger than a single mbuf (%u) and scattered"
1789 " mode has not been requested",
1791 dev->data->dev_conf.rxmode.max_rx_pkt_len,
1792 mb_len - RTE_PKTMBUF_HEADROOM);
1794 /* Use the entire RX mempool as the memory region. */
1795 tmpl.mr = mlx4_mp2mr(priv->pd, mp);
1796 if (tmpl.mr == NULL) {
1798 ERROR("%p: MR creation failure: %s",
1799 (void *)dev, strerror(rte_errno));
1802 if (dev->data->dev_conf.intr_conf.rxq) {
1803 tmpl.channel = ibv_create_comp_channel(priv->ctx);
1804 if (tmpl.channel == NULL) {
1806 ERROR("%p: Rx interrupt completion channel creation"
1808 (void *)dev, strerror(rte_errno));
1812 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, tmpl.channel, 0);
1813 if (tmpl.cq == NULL) {
1815 ERROR("%p: CQ creation failure: %s",
1816 (void *)dev, strerror(rte_errno));
1819 DEBUG("priv->device_attr.max_qp_wr is %d",
1820 priv->device_attr.max_qp_wr);
1821 DEBUG("priv->device_attr.max_sge is %d",
1822 priv->device_attr.max_sge);
1823 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc);
1824 if (tmpl.qp == NULL) {
1825 ERROR("%p: QP creation failure: %s",
1826 (void *)dev, strerror(rte_errno));
1829 mod = (struct ibv_qp_attr){
1830 /* Move the QP to this state. */
1831 .qp_state = IBV_QPS_INIT,
1832 /* Primary port number. */
1833 .port_num = priv->port
1835 ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE | IBV_QP_PORT);
1838 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1839 (void *)dev, strerror(rte_errno));
1842 ret = rxq_alloc_elts(&tmpl, desc);
1844 ERROR("%p: RXQ allocation failed: %s",
1845 (void *)dev, strerror(rte_errno));
1848 ret = ibv_post_recv(tmpl.qp, &(*tmpl.elts)[0].wr, &bad_wr);
1851 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
1854 strerror(rte_errno));
1857 mod = (struct ibv_qp_attr){
1858 .qp_state = IBV_QPS_RTR
1860 ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE);
1863 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1864 (void *)dev, strerror(rte_errno));
1868 tmpl.port_id = dev->data->port_id;
1869 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
1870 /* Clean up rxq in case we're reinitializing it. */
1871 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
1874 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
1880 assert(rte_errno > 0);
1885 * DPDK callback to configure a RX queue.
1888 * Pointer to Ethernet device structure.
1892 * Number of descriptors to configure in queue.
1894 * NUMA socket on which memory must be allocated.
1896 * Thresholds parameters.
1898 * Memory pool for buffer allocations.
1901 * 0 on success, negative errno value otherwise and rte_errno is set.
1904 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1905 unsigned int socket, const struct rte_eth_rxconf *conf,
1906 struct rte_mempool *mp)
1908 struct priv *priv = dev->data->dev_private;
1909 struct rxq *rxq = (*priv->rxqs)[idx];
1912 DEBUG("%p: configuring queue %u for %u descriptors",
1913 (void *)dev, idx, desc);
1914 if (idx >= priv->rxqs_n) {
1915 rte_errno = EOVERFLOW;
1916 ERROR("%p: queue index out of range (%u >= %u)",
1917 (void *)dev, idx, priv->rxqs_n);
1921 DEBUG("%p: reusing already allocated queue index %u (%p)",
1922 (void *)dev, idx, (void *)rxq);
1923 if (priv->started) {
1927 (*priv->rxqs)[idx] = NULL;
1929 priv_mac_addr_del(priv);
1932 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
1935 ERROR("%p: unable to allocate queue index %u",
1940 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
1944 rxq->stats.idx = idx;
1945 DEBUG("%p: adding RX queue %p to list",
1946 (void *)dev, (void *)rxq);
1947 (*priv->rxqs)[idx] = rxq;
1948 /* Update receive callback. */
1949 dev->rx_pkt_burst = mlx4_rx_burst;
1955 * DPDK callback to release a RX queue.
1958 * Generic RX queue pointer.
1961 mlx4_rx_queue_release(void *dpdk_rxq)
1963 struct rxq *rxq = (struct rxq *)dpdk_rxq;
1970 for (i = 0; (i != priv->rxqs_n); ++i)
1971 if ((*priv->rxqs)[i] == rxq) {
1972 DEBUG("%p: removing RX queue %p from list",
1973 (void *)priv->dev, (void *)rxq);
1974 (*priv->rxqs)[i] = NULL;
1976 priv_mac_addr_del(priv);
1984 priv_dev_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
1987 priv_dev_removal_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
1990 priv_dev_link_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
1993 * DPDK callback to start the device.
1995 * Simulate device start by attaching all configured flows.
1998 * Pointer to Ethernet device structure.
2001 * 0 on success, negative errno value otherwise and rte_errno is set.
2004 mlx4_dev_start(struct rte_eth_dev *dev)
2006 struct priv *priv = dev->data->dev_private;
2011 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
2013 ret = priv_mac_addr_add(priv);
2016 ret = priv_dev_link_interrupt_handler_install(priv, dev);
2018 ERROR("%p: LSC handler install failed",
2022 ret = priv_dev_removal_interrupt_handler_install(priv, dev);
2024 ERROR("%p: RMV handler install failed",
2028 ret = priv_rx_intr_vec_enable(priv);
2030 ERROR("%p: Rx interrupt vector creation failed",
2034 ret = mlx4_priv_flow_start(priv);
2036 ERROR("%p: flow start failed: %s",
2037 (void *)dev, strerror(ret));
2043 priv_mac_addr_del(priv);
2049 * DPDK callback to stop the device.
2051 * Simulate device stop by detaching all configured flows.
2054 * Pointer to Ethernet device structure.
2057 mlx4_dev_stop(struct rte_eth_dev *dev)
2059 struct priv *priv = dev->data->dev_private;
2063 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
2065 mlx4_priv_flow_stop(priv);
2066 priv_mac_addr_del(priv);
2070 * Dummy DPDK callback for TX.
2072 * This function is used to temporarily replace the real callback during
2073 * unsafe control operations on the queue, or in case of error.
2076 * Generic pointer to TX queue structure.
2078 * Packets to transmit.
2080 * Number of packets in array.
2083 * Number of packets successfully transmitted (<= pkts_n).
2086 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
2095 * Dummy DPDK callback for RX.
2097 * This function is used to temporarily replace the real callback during
2098 * unsafe control operations on the queue, or in case of error.
2101 * Generic pointer to RX queue structure.
2103 * Array to store received packets.
2105 * Maximum number of packets in array.
2108 * Number of packets successfully received (<= pkts_n).
2111 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2120 priv_dev_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
2123 priv_dev_removal_interrupt_handler_uninstall(struct priv *,
2124 struct rte_eth_dev *);
2127 priv_dev_link_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
2130 * DPDK callback to close the device.
2132 * Destroy all queues and objects, free memory.
2135 * Pointer to Ethernet device structure.
2138 mlx4_dev_close(struct rte_eth_dev *dev)
2140 struct priv *priv = dev->data->dev_private;
2146 DEBUG("%p: closing device \"%s\"",
2148 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
2149 priv_mac_addr_del(priv);
2151 * Prevent crashes when queues are still in use. This is unfortunately
2152 * still required for DPDK 1.3 because some programs (such as testpmd)
2153 * never release them before closing the device.
2155 dev->rx_pkt_burst = removed_rx_burst;
2156 dev->tx_pkt_burst = removed_tx_burst;
2157 if (priv->rxqs != NULL) {
2158 /* XXX race condition if mlx4_rx_burst() is still running. */
2160 for (i = 0; (i != priv->rxqs_n); ++i) {
2161 tmp = (*priv->rxqs)[i];
2164 (*priv->rxqs)[i] = NULL;
2171 if (priv->txqs != NULL) {
2172 /* XXX race condition if mlx4_tx_burst() is still running. */
2174 for (i = 0; (i != priv->txqs_n); ++i) {
2175 tmp = (*priv->txqs)[i];
2178 (*priv->txqs)[i] = NULL;
2185 if (priv->pd != NULL) {
2186 assert(priv->ctx != NULL);
2187 claim_zero(ibv_dealloc_pd(priv->pd));
2188 claim_zero(ibv_close_device(priv->ctx));
2190 assert(priv->ctx == NULL);
2191 priv_dev_removal_interrupt_handler_uninstall(priv, dev);
2192 priv_dev_link_interrupt_handler_uninstall(priv, dev);
2193 priv_rx_intr_vec_disable(priv);
2194 memset(priv, 0, sizeof(*priv));
2198 * Change the link state (UP / DOWN).
2201 * Pointer to Ethernet device private data.
2203 * Nonzero for link up, otherwise link down.
2206 * 0 on success, negative errno value otherwise and rte_errno is set.
2209 priv_set_link(struct priv *priv, int up)
2211 struct rte_eth_dev *dev = priv->dev;
2215 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
2218 dev->rx_pkt_burst = mlx4_rx_burst;
2220 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
2223 dev->rx_pkt_burst = removed_rx_burst;
2224 dev->tx_pkt_burst = removed_tx_burst;
2230 * DPDK callback to bring the link DOWN.
2233 * Pointer to Ethernet device structure.
2236 * 0 on success, negative errno value otherwise and rte_errno is set.
2239 mlx4_set_link_down(struct rte_eth_dev *dev)
2241 struct priv *priv = dev->data->dev_private;
2243 return priv_set_link(priv, 0);
2247 * DPDK callback to bring the link UP.
2250 * Pointer to Ethernet device structure.
2253 * 0 on success, negative errno value otherwise and rte_errno is set.
2256 mlx4_set_link_up(struct rte_eth_dev *dev)
2258 struct priv *priv = dev->data->dev_private;
2260 return priv_set_link(priv, 1);
2264 * DPDK callback to get information about the device.
2267 * Pointer to Ethernet device structure.
2269 * Info structure output buffer.
2272 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
2274 struct priv *priv = dev->data->dev_private;
2276 char ifname[IF_NAMESIZE];
2278 info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
2281 /* FIXME: we should ask the device for these values. */
2282 info->min_rx_bufsize = 32;
2283 info->max_rx_pktlen = 65536;
2285 * Since we need one CQ per QP, the limit is the minimum number
2286 * between the two values.
2288 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
2289 priv->device_attr.max_qp : priv->device_attr.max_cq);
2290 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
2293 info->max_rx_queues = max;
2294 info->max_tx_queues = max;
2295 /* Last array entry is reserved for broadcast. */
2296 info->max_mac_addrs = 1;
2297 info->rx_offload_capa = 0;
2298 info->tx_offload_capa = 0;
2299 if (priv_get_ifname(priv, &ifname) == 0)
2300 info->if_index = if_nametoindex(ifname);
2303 ETH_LINK_SPEED_10G |
2304 ETH_LINK_SPEED_20G |
2305 ETH_LINK_SPEED_40G |
2310 * DPDK callback to get device statistics.
2313 * Pointer to Ethernet device structure.
2315 * Stats structure output buffer.
2318 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
2320 struct priv *priv = dev->data->dev_private;
2321 struct rte_eth_stats tmp = {0};
2327 /* Add software counters. */
2328 for (i = 0; (i != priv->rxqs_n); ++i) {
2329 struct rxq *rxq = (*priv->rxqs)[i];
2333 idx = rxq->stats.idx;
2334 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
2335 tmp.q_ipackets[idx] += rxq->stats.ipackets;
2336 tmp.q_ibytes[idx] += rxq->stats.ibytes;
2337 tmp.q_errors[idx] += (rxq->stats.idropped +
2338 rxq->stats.rx_nombuf);
2340 tmp.ipackets += rxq->stats.ipackets;
2341 tmp.ibytes += rxq->stats.ibytes;
2342 tmp.ierrors += rxq->stats.idropped;
2343 tmp.rx_nombuf += rxq->stats.rx_nombuf;
2345 for (i = 0; (i != priv->txqs_n); ++i) {
2346 struct txq *txq = (*priv->txqs)[i];
2350 idx = txq->stats.idx;
2351 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
2352 tmp.q_opackets[idx] += txq->stats.opackets;
2353 tmp.q_obytes[idx] += txq->stats.obytes;
2354 tmp.q_errors[idx] += txq->stats.odropped;
2356 tmp.opackets += txq->stats.opackets;
2357 tmp.obytes += txq->stats.obytes;
2358 tmp.oerrors += txq->stats.odropped;
2364 * DPDK callback to clear device statistics.
2367 * Pointer to Ethernet device structure.
2370 mlx4_stats_reset(struct rte_eth_dev *dev)
2372 struct priv *priv = dev->data->dev_private;
2378 for (i = 0; (i != priv->rxqs_n); ++i) {
2379 if ((*priv->rxqs)[i] == NULL)
2381 idx = (*priv->rxqs)[i]->stats.idx;
2382 (*priv->rxqs)[i]->stats =
2383 (struct mlx4_rxq_stats){ .idx = idx };
2385 for (i = 0; (i != priv->txqs_n); ++i) {
2386 if ((*priv->txqs)[i] == NULL)
2388 idx = (*priv->txqs)[i]->stats.idx;
2389 (*priv->txqs)[i]->stats =
2390 (struct mlx4_txq_stats){ .idx = idx };
2395 * DPDK callback to retrieve physical link information.
2398 * Pointer to Ethernet device structure.
2399 * @param wait_to_complete
2400 * Wait for request completion (ignored).
2403 * 0 on success, negative errno value otherwise and rte_errno is set.
2406 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
2408 const struct priv *priv = dev->data->dev_private;
2409 struct ethtool_cmd edata = {
2413 struct rte_eth_link dev_link;
2420 (void)wait_to_complete;
2421 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
2422 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(rte_errno));
2425 memset(&dev_link, 0, sizeof(dev_link));
2426 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
2427 (ifr.ifr_flags & IFF_RUNNING));
2428 ifr.ifr_data = (void *)&edata;
2429 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2430 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
2431 strerror(rte_errno));
2434 link_speed = ethtool_cmd_speed(&edata);
2435 if (link_speed == -1)
2436 dev_link.link_speed = 0;
2438 dev_link.link_speed = link_speed;
2439 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
2440 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
2441 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
2442 ETH_LINK_SPEED_FIXED);
2443 dev->data->dev_link = dev_link;
2448 * DPDK callback to get flow control status.
2451 * Pointer to Ethernet device structure.
2452 * @param[out] fc_conf
2453 * Flow control output buffer.
2456 * 0 on success, negative errno value otherwise and rte_errno is set.
2459 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
2461 struct priv *priv = dev->data->dev_private;
2463 struct ethtool_pauseparam ethpause = {
2464 .cmd = ETHTOOL_GPAUSEPARAM
2468 ifr.ifr_data = (void *)ðpause;
2469 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2471 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
2473 strerror(rte_errno));
2476 fc_conf->autoneg = ethpause.autoneg;
2477 if (ethpause.rx_pause && ethpause.tx_pause)
2478 fc_conf->mode = RTE_FC_FULL;
2479 else if (ethpause.rx_pause)
2480 fc_conf->mode = RTE_FC_RX_PAUSE;
2481 else if (ethpause.tx_pause)
2482 fc_conf->mode = RTE_FC_TX_PAUSE;
2484 fc_conf->mode = RTE_FC_NONE;
2492 * DPDK callback to modify flow control parameters.
2495 * Pointer to Ethernet device structure.
2496 * @param[in] fc_conf
2497 * Flow control parameters.
2500 * 0 on success, negative errno value otherwise and rte_errno is set.
2503 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
2505 struct priv *priv = dev->data->dev_private;
2507 struct ethtool_pauseparam ethpause = {
2508 .cmd = ETHTOOL_SPAUSEPARAM
2512 ifr.ifr_data = (void *)ðpause;
2513 ethpause.autoneg = fc_conf->autoneg;
2514 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
2515 (fc_conf->mode & RTE_FC_RX_PAUSE))
2516 ethpause.rx_pause = 1;
2518 ethpause.rx_pause = 0;
2519 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
2520 (fc_conf->mode & RTE_FC_TX_PAUSE))
2521 ethpause.tx_pause = 1;
2523 ethpause.tx_pause = 0;
2524 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2526 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
2528 strerror(rte_errno));
2537 const struct rte_flow_ops mlx4_flow_ops = {
2538 .validate = mlx4_flow_validate,
2539 .create = mlx4_flow_create,
2540 .destroy = mlx4_flow_destroy,
2541 .flush = mlx4_flow_flush,
2543 .isolate = mlx4_flow_isolate,
2547 * Manage filter operations.
2550 * Pointer to Ethernet device structure.
2551 * @param filter_type
2554 * Operation to perform.
2556 * Pointer to operation-specific structure.
2559 * 0 on success, negative errno value otherwise and rte_errno is set.
2562 mlx4_dev_filter_ctrl(struct rte_eth_dev *dev,
2563 enum rte_filter_type filter_type,
2564 enum rte_filter_op filter_op,
2567 switch (filter_type) {
2568 case RTE_ETH_FILTER_GENERIC:
2569 if (filter_op != RTE_ETH_FILTER_GET)
2571 *(const void **)arg = &mlx4_flow_ops;
2574 ERROR("%p: filter type (%d) not supported",
2575 (void *)dev, filter_type);
2578 rte_errno = ENOTSUP;
2582 static const struct eth_dev_ops mlx4_dev_ops = {
2583 .dev_configure = mlx4_dev_configure,
2584 .dev_start = mlx4_dev_start,
2585 .dev_stop = mlx4_dev_stop,
2586 .dev_set_link_down = mlx4_set_link_down,
2587 .dev_set_link_up = mlx4_set_link_up,
2588 .dev_close = mlx4_dev_close,
2589 .link_update = mlx4_link_update,
2590 .stats_get = mlx4_stats_get,
2591 .stats_reset = mlx4_stats_reset,
2592 .dev_infos_get = mlx4_dev_infos_get,
2593 .rx_queue_setup = mlx4_rx_queue_setup,
2594 .tx_queue_setup = mlx4_tx_queue_setup,
2595 .rx_queue_release = mlx4_rx_queue_release,
2596 .tx_queue_release = mlx4_tx_queue_release,
2597 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
2598 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
2599 .mtu_set = mlx4_dev_set_mtu,
2600 .filter_ctrl = mlx4_dev_filter_ctrl,
2601 .rx_queue_intr_enable = mlx4_rx_intr_enable,
2602 .rx_queue_intr_disable = mlx4_rx_intr_disable,
2606 * Get PCI information from struct ibv_device.
2609 * Pointer to Ethernet device structure.
2610 * @param[out] pci_addr
2611 * PCI bus address output buffer.
2614 * 0 on success, negative errno value otherwise and rte_errno is set.
2617 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
2618 struct rte_pci_addr *pci_addr)
2622 MKSTR(path, "%s/device/uevent", device->ibdev_path);
2624 file = fopen(path, "rb");
2629 while (fgets(line, sizeof(line), file) == line) {
2630 size_t len = strlen(line);
2633 /* Truncate long lines. */
2634 if (len == (sizeof(line) - 1))
2635 while (line[(len - 1)] != '\n') {
2639 line[(len - 1)] = ret;
2641 /* Extract information. */
2644 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
2648 &pci_addr->function) == 4) {
2658 * Get MAC address by querying netdevice.
2661 * struct priv for the requested device.
2663 * MAC address output buffer.
2666 * 0 on success, negative errno value otherwise and rte_errno is set.
2669 priv_get_mac(struct priv *priv, uint8_t (*mac)[ETHER_ADDR_LEN])
2671 struct ifreq request;
2672 int ret = priv_ifreq(priv, SIOCGIFHWADDR, &request);
2676 memcpy(mac, request.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
2681 mlx4_dev_link_status_handler(void *);
2683 mlx4_dev_interrupt_handler(void *);
2686 * Link/device status handler.
2689 * Pointer to private structure.
2691 * Pointer to the rte_eth_dev structure.
2693 * Pointer to event flags holder.
2699 priv_dev_status_handler(struct priv *priv, struct rte_eth_dev *dev,
2702 struct ibv_async_event event;
2703 int port_change = 0;
2704 struct rte_eth_link *link = &dev->data->dev_link;
2708 /* Read all message and acknowledge them. */
2710 if (ibv_get_async_event(priv->ctx, &event))
2712 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
2713 event.event_type == IBV_EVENT_PORT_ERR) &&
2714 (priv->intr_conf.lsc == 1)) {
2717 } else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
2718 priv->intr_conf.rmv == 1) {
2719 *events |= (1 << RTE_ETH_EVENT_INTR_RMV);
2722 DEBUG("event type %d on port %d not handled",
2723 event.event_type, event.element.port_num);
2724 ibv_ack_async_event(&event);
2728 mlx4_link_update(dev, 0);
2729 if (((link->link_speed == 0) && link->link_status) ||
2730 ((link->link_speed != 0) && !link->link_status)) {
2731 if (!priv->pending_alarm) {
2732 /* Inconsistent status, check again later. */
2733 priv->pending_alarm = 1;
2734 rte_eal_alarm_set(MLX4_ALARM_TIMEOUT_US,
2735 mlx4_dev_link_status_handler,
2739 *events |= (1 << RTE_ETH_EVENT_INTR_LSC);
2745 * Handle delayed link status event.
2748 * Registered argument.
2751 mlx4_dev_link_status_handler(void *arg)
2753 struct rte_eth_dev *dev = arg;
2754 struct priv *priv = dev->data->dev_private;
2758 assert(priv->pending_alarm == 1);
2759 priv->pending_alarm = 0;
2760 ret = priv_dev_status_handler(priv, dev, &events);
2761 if (ret > 0 && events & (1 << RTE_ETH_EVENT_INTR_LSC))
2762 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC,
2767 * Handle interrupts from the NIC.
2769 * @param[in] intr_handle
2770 * Interrupt handler.
2772 * Callback argument.
2775 mlx4_dev_interrupt_handler(void *cb_arg)
2777 struct rte_eth_dev *dev = cb_arg;
2778 struct priv *priv = dev->data->dev_private;
2783 ret = priv_dev_status_handler(priv, dev, &ev);
2785 for (i = RTE_ETH_EVENT_UNKNOWN;
2786 i < RTE_ETH_EVENT_MAX;
2788 if (ev & (1 << i)) {
2790 _rte_eth_dev_callback_process(dev, i, NULL,
2796 WARN("%d event%s not processed", ret,
2797 (ret > 1 ? "s were" : " was"));
2802 * Uninstall interrupt handler.
2805 * Pointer to private structure.
2807 * Pointer to the rte_eth_dev structure.
2810 * 0 on success, negative errno value otherwise and rte_errno is set.
2813 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
2817 if (priv->intr_conf.lsc ||
2818 priv->intr_conf.rmv)
2820 ret = rte_intr_callback_unregister(&priv->intr_handle,
2821 mlx4_dev_interrupt_handler,
2825 ERROR("rte_intr_callback_unregister failed with %d %s",
2826 ret, strerror(rte_errno));
2828 priv->intr_handle.fd = 0;
2829 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
2834 * Install interrupt handler.
2837 * Pointer to private structure.
2839 * Pointer to the rte_eth_dev structure.
2842 * 0 on success, negative errno value otherwise and rte_errno is set.
2845 priv_dev_interrupt_handler_install(struct priv *priv,
2846 struct rte_eth_dev *dev)
2852 * Check whether the interrupt handler has already been installed
2853 * for either type of interrupt.
2855 if (priv->intr_conf.lsc &&
2856 priv->intr_conf.rmv &&
2857 priv->intr_handle.fd)
2859 assert(priv->ctx->async_fd > 0);
2860 flags = fcntl(priv->ctx->async_fd, F_GETFL);
2861 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2863 rte_errno = errno ? errno : EINVAL;
2864 INFO("failed to change file descriptor async event queue");
2865 dev->data->dev_conf.intr_conf.lsc = 0;
2866 dev->data->dev_conf.intr_conf.rmv = 0;
2869 priv->intr_handle.fd = priv->ctx->async_fd;
2870 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
2871 rc = rte_intr_callback_register(&priv->intr_handle,
2872 mlx4_dev_interrupt_handler,
2876 ERROR("rte_intr_callback_register failed "
2877 " (rte_errno: %s)", strerror(rte_errno));
2885 * Uninstall interrupt handler.
2888 * Pointer to private structure.
2890 * Pointer to the rte_eth_dev structure.
2893 * 0 on success, negative errno value otherwise and rte_errno is set.
2896 priv_dev_removal_interrupt_handler_uninstall(struct priv *priv,
2897 struct rte_eth_dev *dev)
2899 if (dev->data->dev_conf.intr_conf.rmv) {
2900 priv->intr_conf.rmv = 0;
2901 return priv_dev_interrupt_handler_uninstall(priv, dev);
2907 * Uninstall interrupt handler.
2910 * Pointer to private structure.
2912 * Pointer to the rte_eth_dev structure.
2915 * 0 on success, negative errno value otherwise and rte_errno is set.
2918 priv_dev_link_interrupt_handler_uninstall(struct priv *priv,
2919 struct rte_eth_dev *dev)
2923 if (dev->data->dev_conf.intr_conf.lsc) {
2924 priv->intr_conf.lsc = 0;
2925 ret = priv_dev_interrupt_handler_uninstall(priv, dev);
2929 if (priv->pending_alarm)
2930 if (rte_eal_alarm_cancel(mlx4_dev_link_status_handler,
2932 ERROR("rte_eal_alarm_cancel failed "
2933 " (rte_errno: %s)", strerror(rte_errno));
2936 priv->pending_alarm = 0;
2941 * Install link interrupt handler.
2944 * Pointer to private structure.
2946 * Pointer to the rte_eth_dev structure.
2949 * 0 on success, negative errno value otherwise and rte_errno is set.
2952 priv_dev_link_interrupt_handler_install(struct priv *priv,
2953 struct rte_eth_dev *dev)
2957 if (dev->data->dev_conf.intr_conf.lsc) {
2958 ret = priv_dev_interrupt_handler_install(priv, dev);
2961 priv->intr_conf.lsc = 1;
2967 * Install removal interrupt handler.
2970 * Pointer to private structure.
2972 * Pointer to the rte_eth_dev structure.
2975 * 0 on success, negative errno value otherwise and rte_errno is set.
2978 priv_dev_removal_interrupt_handler_install(struct priv *priv,
2979 struct rte_eth_dev *dev)
2983 if (dev->data->dev_conf.intr_conf.rmv) {
2984 ret = priv_dev_interrupt_handler_install(priv, dev);
2987 priv->intr_conf.rmv = 1;
2993 * Allocate queue vector and fill epoll fd list for Rx interrupts.
2996 * Pointer to private structure.
2999 * 0 on success, negative errno value otherwise and rte_errno is set.
3002 priv_rx_intr_vec_enable(struct priv *priv)
3005 unsigned int rxqs_n = priv->rxqs_n;
3006 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
3007 unsigned int count = 0;
3008 struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
3010 if (!priv->dev->data->dev_conf.intr_conf.rxq)
3012 priv_rx_intr_vec_disable(priv);
3013 intr_handle->intr_vec = malloc(sizeof(intr_handle->intr_vec[rxqs_n]));
3014 if (intr_handle->intr_vec == NULL) {
3016 ERROR("failed to allocate memory for interrupt vector,"
3017 " Rx interrupts will not be supported");
3020 intr_handle->type = RTE_INTR_HANDLE_EXT;
3021 for (i = 0; i != n; ++i) {
3022 struct rxq *rxq = (*priv->rxqs)[i];
3027 /* Skip queues that cannot request interrupts. */
3028 if (!rxq || !rxq->channel) {
3029 /* Use invalid intr_vec[] index to disable entry. */
3030 intr_handle->intr_vec[i] =
3031 RTE_INTR_VEC_RXTX_OFFSET +
3032 RTE_MAX_RXTX_INTR_VEC_ID;
3035 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
3037 ERROR("too many Rx queues for interrupt vector size"
3038 " (%d), Rx interrupts cannot be enabled",
3039 RTE_MAX_RXTX_INTR_VEC_ID);
3040 priv_rx_intr_vec_disable(priv);
3043 fd = rxq->channel->fd;
3044 flags = fcntl(fd, F_GETFL);
3045 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
3048 ERROR("failed to make Rx interrupt file descriptor"
3049 " %d non-blocking for queue index %d", fd, i);
3050 priv_rx_intr_vec_disable(priv);
3053 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
3054 intr_handle->efds[count] = fd;
3058 priv_rx_intr_vec_disable(priv);
3060 intr_handle->nb_efd = count;
3065 * Clean up Rx interrupts handler.
3068 * Pointer to private structure.
3071 priv_rx_intr_vec_disable(struct priv *priv)
3073 struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
3075 rte_intr_free_epoll_fd(intr_handle);
3076 free(intr_handle->intr_vec);
3077 intr_handle->nb_efd = 0;
3078 intr_handle->intr_vec = NULL;
3082 * DPDK callback for Rx queue interrupt enable.
3085 * Pointer to Ethernet device structure.
3090 * 0 on success, negative errno value otherwise and rte_errno is set.
3093 mlx4_rx_intr_enable(struct rte_eth_dev *dev, uint16_t idx)
3095 struct priv *priv = dev->data->dev_private;
3096 struct rxq *rxq = (*priv->rxqs)[idx];
3099 if (!rxq || !rxq->channel)
3102 ret = ibv_req_notify_cq(rxq->cq, 0);
3105 WARN("unable to arm interrupt on rx queue %d", idx);
3111 * DPDK callback for Rx queue interrupt disable.
3114 * Pointer to Ethernet device structure.
3119 * 0 on success, negative errno value otherwise and rte_errno is set.
3122 mlx4_rx_intr_disable(struct rte_eth_dev *dev, uint16_t idx)
3124 struct priv *priv = dev->data->dev_private;
3125 struct rxq *rxq = (*priv->rxqs)[idx];
3126 struct ibv_cq *ev_cq;
3130 if (!rxq || !rxq->channel) {
3133 ret = ibv_get_cq_event(rxq->cq->channel, &ev_cq, &ev_ctx);
3134 if (ret || ev_cq != rxq->cq)
3139 WARN("unable to disable interrupt on rx queue %d",
3142 ibv_ack_cq_events(rxq->cq, 1);
3148 * Verify and store value for device argument.
3151 * Key argument to verify.
3153 * Value associated with key.
3154 * @param[in, out] conf
3155 * Shared configuration data.
3158 * 0 on success, negative errno value otherwise and rte_errno is set.
3161 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
3166 tmp = strtoul(val, NULL, 0);
3169 WARN("%s: \"%s\" is not a valid integer", key, val);
3172 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
3173 uint32_t ports = rte_log2_u32(conf->ports.present);
3176 ERROR("port index %lu outside range [0,%" PRIu32 ")",
3180 if (!(conf->ports.present & (1 << tmp))) {
3182 ERROR("invalid port index %lu", tmp);
3185 conf->ports.enabled |= 1 << tmp;
3188 WARN("%s: unknown parameter", key);
3195 * Parse device parameters.
3198 * Device arguments structure.
3201 * 0 on success, negative errno value otherwise and rte_errno is set.
3204 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
3206 struct rte_kvargs *kvlist;
3207 unsigned int arg_count;
3211 if (devargs == NULL)
3213 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
3214 if (kvlist == NULL) {
3216 ERROR("failed to parse kvargs");
3219 /* Process parameters. */
3220 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
3221 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
3222 while (arg_count-- > 0) {
3223 ret = rte_kvargs_process(kvlist,
3224 MLX4_PMD_PORT_KVARG,
3225 (int (*)(const char *,
3235 rte_kvargs_free(kvlist);
3239 static struct rte_pci_driver mlx4_driver;
3242 * DPDK callback to register a PCI device.
3244 * This function creates an Ethernet device for each port of a given
3247 * @param[in] pci_drv
3248 * PCI driver structure (mlx4_driver).
3249 * @param[in] pci_dev
3250 * PCI device information.
3253 * 0 on success, negative errno value otherwise and rte_errno is set.
3256 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
3258 struct ibv_device **list;
3259 struct ibv_device *ibv_dev;
3261 struct ibv_context *attr_ctx = NULL;
3262 struct ibv_device_attr device_attr;
3263 struct mlx4_conf conf = {
3270 assert(pci_drv == &mlx4_driver);
3271 list = ibv_get_device_list(&i);
3275 if (rte_errno == ENOSYS)
3276 ERROR("cannot list devices, is ib_uverbs loaded?");
3281 * For each listed device, check related sysfs entry against
3282 * the provided PCI ID.
3285 struct rte_pci_addr pci_addr;
3288 DEBUG("checking device \"%s\"", list[i]->name);
3289 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
3291 if ((pci_dev->addr.domain != pci_addr.domain) ||
3292 (pci_dev->addr.bus != pci_addr.bus) ||
3293 (pci_dev->addr.devid != pci_addr.devid) ||
3294 (pci_dev->addr.function != pci_addr.function))
3296 vf = (pci_dev->id.device_id ==
3297 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
3298 INFO("PCI information matches, using device \"%s\" (VF: %s)",
3299 list[i]->name, (vf ? "true" : "false"));
3300 attr_ctx = ibv_open_device(list[i]);
3304 if (attr_ctx == NULL) {
3305 ibv_free_device_list(list);
3309 ERROR("cannot access device, is mlx4_ib loaded?");
3313 ERROR("cannot use device, are drivers up to date?");
3321 DEBUG("device opened");
3322 if (ibv_query_device(attr_ctx, &device_attr)) {
3326 INFO("%u port(s) detected", device_attr.phys_port_cnt);
3327 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
3328 if (mlx4_args(pci_dev->device.devargs, &conf)) {
3329 ERROR("failed to process device arguments");
3333 /* Use all ports when none are defined */
3334 if (!conf.ports.enabled)
3335 conf.ports.enabled = conf.ports.present;
3336 for (i = 0; i < device_attr.phys_port_cnt; i++) {
3337 uint32_t port = i + 1; /* ports are indexed from one */
3338 struct ibv_context *ctx = NULL;
3339 struct ibv_port_attr port_attr;
3340 struct ibv_pd *pd = NULL;
3341 struct priv *priv = NULL;
3342 struct rte_eth_dev *eth_dev = NULL;
3343 struct ether_addr mac;
3345 /* If port is not enabled, skip. */
3346 if (!(conf.ports.enabled & (1 << i)))
3348 DEBUG("using port %u", port);
3349 ctx = ibv_open_device(ibv_dev);
3354 /* Check port status. */
3355 err = ibv_query_port(ctx, port, &port_attr);
3358 ERROR("port query failed: %s", strerror(rte_errno));
3361 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
3362 rte_errno = ENOTSUP;
3363 ERROR("port %d is not configured in Ethernet mode",
3367 if (port_attr.state != IBV_PORT_ACTIVE)
3368 DEBUG("port %d is not active: \"%s\" (%d)",
3369 port, ibv_port_state_str(port_attr.state),
3371 /* Allocate protection domain. */
3372 pd = ibv_alloc_pd(ctx);
3375 ERROR("PD allocation failure");
3378 /* from rte_ethdev.c */
3379 priv = rte_zmalloc("ethdev private structure",
3381 RTE_CACHE_LINE_SIZE);
3384 ERROR("priv allocation failure");
3388 priv->device_attr = device_attr;
3391 priv->mtu = ETHER_MTU;
3393 /* Configure the first MAC address by default. */
3394 if (priv_get_mac(priv, &mac.addr_bytes)) {
3395 ERROR("cannot get MAC address, is mlx4_en loaded?"
3396 " (rte_errno: %s)", strerror(rte_errno));
3399 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
3401 mac.addr_bytes[0], mac.addr_bytes[1],
3402 mac.addr_bytes[2], mac.addr_bytes[3],
3403 mac.addr_bytes[4], mac.addr_bytes[5]);
3404 /* Register MAC address. */
3406 if (priv_mac_addr_add(priv))
3410 char ifname[IF_NAMESIZE];
3412 if (priv_get_ifname(priv, &ifname) == 0)
3413 DEBUG("port %u ifname is \"%s\"",
3414 priv->port, ifname);
3416 DEBUG("port %u ifname is unknown", priv->port);
3419 /* Get actual MTU if possible. */
3420 priv_get_mtu(priv, &priv->mtu);
3421 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
3422 /* from rte_ethdev.c */
3424 char name[RTE_ETH_NAME_MAX_LEN];
3426 snprintf(name, sizeof(name), "%s port %u",
3427 ibv_get_device_name(ibv_dev), port);
3428 eth_dev = rte_eth_dev_allocate(name);
3430 if (eth_dev == NULL) {
3431 ERROR("can not allocate rte ethdev");
3435 eth_dev->data->dev_private = priv;
3436 eth_dev->data->mac_addrs = &priv->mac;
3437 eth_dev->device = &pci_dev->device;
3438 rte_eth_copy_pci_info(eth_dev, pci_dev);
3439 eth_dev->device->driver = &mlx4_driver.driver;
3441 * Copy and override interrupt handle to prevent it from
3442 * being shared between all ethdev instances of a given PCI
3443 * device. This is required to properly handle Rx interrupts
3446 priv->intr_handle_dev = *eth_dev->intr_handle;
3447 eth_dev->intr_handle = &priv->intr_handle_dev;
3448 priv->dev = eth_dev;
3449 eth_dev->dev_ops = &mlx4_dev_ops;
3450 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
3451 /* Bring Ethernet device up. */
3452 DEBUG("forcing Ethernet interface up");
3453 priv_set_flags(priv, ~IFF_UP, IFF_UP);
3454 /* Update link status once if waiting for LSC. */
3455 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
3456 mlx4_link_update(eth_dev, 0);
3461 claim_zero(ibv_dealloc_pd(pd));
3463 claim_zero(ibv_close_device(ctx));
3465 rte_eth_dev_release_port(eth_dev);
3468 if (i == device_attr.phys_port_cnt)
3471 * XXX if something went wrong in the loop above, there is a resource
3472 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
3473 * long as the dpdk does not provide a way to deallocate a ethdev and a
3474 * way to enumerate the registered ethdevs to free the previous ones.
3478 claim_zero(ibv_close_device(attr_ctx));
3480 ibv_free_device_list(list);
3481 assert(rte_errno >= 0);
3485 static const struct rte_pci_id mlx4_pci_id_map[] = {
3487 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3488 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
3491 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3492 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
3495 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3496 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
3503 static struct rte_pci_driver mlx4_driver = {
3505 .name = MLX4_DRIVER_NAME
3507 .id_table = mlx4_pci_id_map,
3508 .probe = mlx4_pci_probe,
3509 .drv_flags = RTE_PCI_DRV_INTR_LSC |
3510 RTE_PCI_DRV_INTR_RMV,
3514 * Driver initialization routine.
3516 RTE_INIT(rte_mlx4_pmd_init);
3518 rte_mlx4_pmd_init(void)
3521 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3522 * huge pages. Calling ibv_fork_init() during init allows
3523 * applications to use fork() safely for purposes other than
3524 * using this PMD, which is not supported in forked processes.
3526 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3528 rte_pci_register(&mlx4_driver);
3531 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
3532 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
3533 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
3534 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");