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);
423 * Pointer to private structure.
428 * 0 on success, negative errno value otherwise and rte_errno is set.
431 priv_set_mtu(struct priv *priv, uint16_t mtu)
434 int ret = priv_set_sysfs_ulong(priv, "mtu", mtu);
438 ret = priv_get_mtu(priv, &new_mtu);
451 * Pointer to private structure.
453 * Bitmask for flags that must remain untouched.
455 * Bitmask for flags to modify.
458 * 0 on success, negative errno value otherwise and rte_errno is set.
461 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
463 unsigned long tmp = 0;
464 int ret = priv_get_sysfs_ulong(priv, "flags", &tmp);
469 tmp |= (flags & (~keep));
470 return priv_set_sysfs_ulong(priv, "flags", tmp);
473 /* Device configuration. */
476 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
477 unsigned int socket, const struct rte_eth_txconf *conf);
480 txq_cleanup(struct txq *txq);
483 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
484 unsigned int socket, const struct rte_eth_rxconf *conf,
485 struct rte_mempool *mp);
488 rxq_cleanup(struct rxq *rxq);
491 priv_mac_addr_del(struct priv *priv);
494 * Ethernet device configuration.
496 * Prepare the driver for a given number of TX and RX queues.
499 * Pointer to Ethernet device structure.
502 * 0 on success, negative errno value otherwise and rte_errno is set.
505 dev_configure(struct rte_eth_dev *dev)
507 struct priv *priv = dev->data->dev_private;
508 unsigned int rxqs_n = dev->data->nb_rx_queues;
509 unsigned int txqs_n = dev->data->nb_tx_queues;
511 priv->rxqs = (void *)dev->data->rx_queues;
512 priv->txqs = (void *)dev->data->tx_queues;
513 if (txqs_n != priv->txqs_n) {
514 INFO("%p: TX queues number update: %u -> %u",
515 (void *)dev, priv->txqs_n, txqs_n);
516 priv->txqs_n = txqs_n;
518 if (rxqs_n != priv->rxqs_n) {
519 INFO("%p: Rx queues number update: %u -> %u",
520 (void *)dev, priv->rxqs_n, rxqs_n);
521 priv->rxqs_n = rxqs_n;
527 * DPDK callback for Ethernet device configuration.
530 * Pointer to Ethernet device structure.
533 * 0 on success, negative errno value otherwise and rte_errno is set.
536 mlx4_dev_configure(struct rte_eth_dev *dev)
538 return dev_configure(dev);
541 static uint16_t mlx4_tx_burst(void *, struct rte_mbuf **, uint16_t);
542 static uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
544 /* TX queues handling. */
547 * Allocate TX queue elements.
550 * Pointer to TX queue structure.
552 * Number of elements to allocate.
555 * 0 on success, negative errno value otherwise and rte_errno is set.
558 txq_alloc_elts(struct txq *txq, unsigned int elts_n)
561 struct txq_elt (*elts)[elts_n] =
562 rte_calloc_socket("TXQ", 1, sizeof(*elts), 0, txq->socket);
566 ERROR("%p: can't allocate packets array", (void *)txq);
570 for (i = 0; (i != elts_n); ++i) {
571 struct txq_elt *elt = &(*elts)[i];
575 DEBUG("%p: allocated and configured %u WRs", (void *)txq, elts_n);
576 txq->elts_n = elts_n;
582 * Request send completion every MLX4_PMD_TX_PER_COMP_REQ packets or
583 * at least 4 times per ring.
585 txq->elts_comp_cd_init =
586 ((MLX4_PMD_TX_PER_COMP_REQ < (elts_n / 4)) ?
587 MLX4_PMD_TX_PER_COMP_REQ : (elts_n / 4));
588 txq->elts_comp_cd = txq->elts_comp_cd_init;
593 DEBUG("%p: failed, freed everything", (void *)txq);
600 * Free TX queue elements.
603 * Pointer to TX queue structure.
606 txq_free_elts(struct txq *txq)
608 unsigned int elts_n = txq->elts_n;
609 unsigned int elts_head = txq->elts_head;
610 unsigned int elts_tail = txq->elts_tail;
611 struct txq_elt (*elts)[elts_n] = txq->elts;
613 DEBUG("%p: freeing WRs", (void *)txq);
618 txq->elts_comp_cd = 0;
619 txq->elts_comp_cd_init = 0;
623 while (elts_tail != elts_head) {
624 struct txq_elt *elt = &(*elts)[elts_tail];
626 assert(elt->buf != NULL);
627 rte_pktmbuf_free(elt->buf);
630 memset(elt, 0x77, sizeof(*elt));
632 if (++elts_tail == elts_n)
639 * Clean up a TX queue.
641 * Destroy objects, free allocated memory and reset the structure for reuse.
644 * Pointer to TX queue structure.
647 txq_cleanup(struct txq *txq)
651 DEBUG("cleaning up %p", (void *)txq);
654 claim_zero(ibv_destroy_qp(txq->qp));
656 claim_zero(ibv_destroy_cq(txq->cq));
657 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
658 if (txq->mp2mr[i].mp == NULL)
660 assert(txq->mp2mr[i].mr != NULL);
661 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
663 memset(txq, 0, sizeof(*txq));
667 * Manage TX completions.
669 * When sending a burst, mlx4_tx_burst() posts several WRs.
670 * To improve performance, a completion event is only required once every
671 * MLX4_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
672 * for other WRs, but this information would not be used anyway.
675 * Pointer to TX queue structure.
678 * 0 on success, -1 on failure.
681 txq_complete(struct txq *txq)
683 unsigned int elts_comp = txq->elts_comp;
684 unsigned int elts_tail = txq->elts_tail;
685 const unsigned int elts_n = txq->elts_n;
686 struct ibv_wc wcs[elts_comp];
689 if (unlikely(elts_comp == 0))
691 wcs_n = ibv_poll_cq(txq->cq, elts_comp, wcs);
692 if (unlikely(wcs_n == 0))
694 if (unlikely(wcs_n < 0)) {
695 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
700 assert(elts_comp <= txq->elts_comp);
702 * Assume WC status is successful as nothing can be done about it
705 elts_tail += wcs_n * txq->elts_comp_cd_init;
706 if (elts_tail >= elts_n)
708 txq->elts_tail = elts_tail;
709 txq->elts_comp = elts_comp;
713 struct mlx4_check_mempool_data {
719 /* Called by mlx4_check_mempool() when iterating the memory chunks. */
720 static void mlx4_check_mempool_cb(struct rte_mempool *mp,
721 void *opaque, struct rte_mempool_memhdr *memhdr,
724 struct mlx4_check_mempool_data *data = opaque;
728 /* It already failed, skip the next chunks. */
731 /* It is the first chunk. */
732 if (data->start == NULL && data->end == NULL) {
733 data->start = memhdr->addr;
734 data->end = data->start + memhdr->len;
737 if (data->end == memhdr->addr) {
738 data->end += memhdr->len;
741 if (data->start == (char *)memhdr->addr + memhdr->len) {
742 data->start -= memhdr->len;
745 /* Error, mempool is not virtually contigous. */
750 * Check if a mempool can be used: it must be virtually contiguous.
753 * Pointer to memory pool.
755 * Pointer to the start address of the mempool virtual memory area
757 * Pointer to the end address of the mempool virtual memory area
760 * 0 on success (mempool is virtually contiguous), -1 on error.
762 static int mlx4_check_mempool(struct rte_mempool *mp, uintptr_t *start,
765 struct mlx4_check_mempool_data data;
767 memset(&data, 0, sizeof(data));
768 rte_mempool_mem_iter(mp, mlx4_check_mempool_cb, &data);
769 *start = (uintptr_t)data.start;
770 *end = (uintptr_t)data.end;
774 /* For best performance, this function should not be inlined. */
775 static struct ibv_mr *mlx4_mp2mr(struct ibv_pd *, struct rte_mempool *)
779 * Register mempool as a memory region.
782 * Pointer to protection domain.
784 * Pointer to memory pool.
787 * Memory region pointer, NULL in case of error and rte_errno is set.
789 static struct ibv_mr *
790 mlx4_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp)
792 const struct rte_memseg *ms = rte_eal_get_physmem_layout();
798 if (mlx4_check_mempool(mp, &start, &end) != 0) {
800 ERROR("mempool %p: not virtually contiguous",
804 DEBUG("mempool %p area start=%p end=%p size=%zu",
805 (void *)mp, (void *)start, (void *)end,
806 (size_t)(end - start));
807 /* Round start and end to page boundary if found in memory segments. */
808 for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) {
809 uintptr_t addr = (uintptr_t)ms[i].addr;
810 size_t len = ms[i].len;
811 unsigned int align = ms[i].hugepage_sz;
813 if ((start > addr) && (start < addr + len))
814 start = RTE_ALIGN_FLOOR(start, align);
815 if ((end > addr) && (end < addr + len))
816 end = RTE_ALIGN_CEIL(end, align);
818 DEBUG("mempool %p using start=%p end=%p size=%zu for MR",
819 (void *)mp, (void *)start, (void *)end,
820 (size_t)(end - start));
824 IBV_ACCESS_LOCAL_WRITE);
826 rte_errno = errno ? errno : EINVAL;
831 * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
832 * the cloned mbuf is allocated is returned instead.
838 * Memory pool where data is located for given mbuf.
840 static struct rte_mempool *
841 txq_mb2mp(struct rte_mbuf *buf)
843 if (unlikely(RTE_MBUF_INDIRECT(buf)))
844 return rte_mbuf_from_indirect(buf)->pool;
849 * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
850 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
851 * remove an entry first.
854 * Pointer to TX queue structure.
856 * Memory Pool for which a Memory Region lkey must be returned.
859 * mr->lkey on success, (uint32_t)-1 on failure.
862 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
867 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
868 if (unlikely(txq->mp2mr[i].mp == NULL)) {
869 /* Unknown MP, add a new MR for it. */
872 if (txq->mp2mr[i].mp == mp) {
873 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
874 assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
875 return txq->mp2mr[i].lkey;
878 /* Add a new entry, register MR first. */
879 DEBUG("%p: discovered new memory pool \"%s\" (%p)",
880 (void *)txq, mp->name, (void *)mp);
881 mr = mlx4_mp2mr(txq->priv->pd, mp);
882 if (unlikely(mr == NULL)) {
883 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
887 if (unlikely(i == elemof(txq->mp2mr))) {
888 /* Table is full, remove oldest entry. */
889 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
892 claim_zero(ibv_dereg_mr(txq->mp2mr[0].mr));
893 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
894 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
896 /* Store the new entry. */
897 txq->mp2mr[i].mp = mp;
898 txq->mp2mr[i].mr = mr;
899 txq->mp2mr[i].lkey = mr->lkey;
900 DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
901 (void *)txq, mp->name, (void *)mp, txq->mp2mr[i].lkey);
902 return txq->mp2mr[i].lkey;
905 struct txq_mp2mr_mbuf_check_data {
910 * Callback function for rte_mempool_obj_iter() to check whether a given
911 * mempool object looks like a mbuf.
914 * The mempool pointer
916 * Context data (struct txq_mp2mr_mbuf_check_data). Contains the
921 * Object index, unused.
924 txq_mp2mr_mbuf_check(struct rte_mempool *mp, void *arg, void *obj,
925 uint32_t index __rte_unused)
927 struct txq_mp2mr_mbuf_check_data *data = arg;
928 struct rte_mbuf *buf = obj;
931 * Check whether mbuf structure fits element size and whether mempool
934 if (sizeof(*buf) > mp->elt_size || buf->pool != mp)
939 * Iterator function for rte_mempool_walk() to register existing mempools and
940 * fill the MP to MR cache of a TX queue.
943 * Memory Pool to register.
945 * Pointer to TX queue structure.
948 txq_mp2mr_iter(struct rte_mempool *mp, void *arg)
950 struct txq *txq = arg;
951 struct txq_mp2mr_mbuf_check_data data = {
955 /* Register mempool only if the first element looks like a mbuf. */
956 if (rte_mempool_obj_iter(mp, txq_mp2mr_mbuf_check, &data) == 0 ||
963 * DPDK callback for TX.
966 * Generic pointer to TX queue structure.
968 * Packets to transmit.
970 * Number of packets in array.
973 * Number of packets successfully transmitted (<= pkts_n).
976 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
978 struct txq *txq = (struct txq *)dpdk_txq;
979 struct ibv_send_wr *wr_head = NULL;
980 struct ibv_send_wr **wr_next = &wr_head;
981 struct ibv_send_wr *wr_bad = NULL;
982 unsigned int elts_head = txq->elts_head;
983 const unsigned int elts_n = txq->elts_n;
984 unsigned int elts_comp_cd = txq->elts_comp_cd;
985 unsigned int elts_comp = 0;
990 assert(elts_comp_cd != 0);
992 max = (elts_n - (elts_head - txq->elts_tail));
996 assert(max <= elts_n);
997 /* Always leave one free entry in the ring. */
1003 for (i = 0; (i != max); ++i) {
1004 struct rte_mbuf *buf = pkts[i];
1005 unsigned int elts_head_next =
1006 (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
1007 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
1008 struct txq_elt *elt = &(*txq->elts)[elts_head];
1009 struct ibv_send_wr *wr = &elt->wr;
1010 unsigned int segs = NB_SEGS(buf);
1011 unsigned int sent_size = 0;
1012 uint32_t send_flags = 0;
1014 /* Clean up old buffer. */
1015 if (likely(elt->buf != NULL)) {
1016 struct rte_mbuf *tmp = elt->buf;
1020 memset(elt, 0x66, sizeof(*elt));
1022 /* Faster than rte_pktmbuf_free(). */
1024 struct rte_mbuf *next = NEXT(tmp);
1026 rte_pktmbuf_free_seg(tmp);
1028 } while (tmp != NULL);
1030 /* Request TX completion. */
1031 if (unlikely(--elts_comp_cd == 0)) {
1032 elts_comp_cd = txq->elts_comp_cd_init;
1034 send_flags |= IBV_SEND_SIGNALED;
1036 if (likely(segs == 1)) {
1037 struct ibv_sge *sge = &elt->sge;
1042 /* Retrieve buffer information. */
1043 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1044 length = DATA_LEN(buf);
1045 /* Retrieve Memory Region key for this memory pool. */
1046 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
1047 if (unlikely(lkey == (uint32_t)-1)) {
1048 /* MR does not exist. */
1049 DEBUG("%p: unable to get MP <-> MR"
1050 " association", (void *)txq);
1051 /* Clean up TX element. */
1055 /* Update element. */
1058 rte_prefetch0((volatile void *)
1060 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1062 sge->length = length;
1064 sent_size += length;
1069 if (sent_size <= txq->max_inline)
1070 send_flags |= IBV_SEND_INLINE;
1071 elts_head = elts_head_next;
1072 /* Increment sent bytes counter. */
1073 txq->stats.obytes += sent_size;
1075 wr->sg_list = &elt->sge;
1077 wr->opcode = IBV_WR_SEND;
1078 wr->send_flags = send_flags;
1080 wr_next = &wr->next;
1083 /* Take a shortcut if nothing must be sent. */
1084 if (unlikely(i == 0))
1086 /* Increment sent packets counter. */
1087 txq->stats.opackets += i;
1088 /* Ring QP doorbell. */
1091 err = ibv_post_send(txq->qp, wr_head, &wr_bad);
1092 if (unlikely(err)) {
1093 uint64_t obytes = 0;
1094 uint64_t opackets = 0;
1096 /* Rewind bad WRs. */
1097 while (wr_bad != NULL) {
1100 /* Force completion request if one was lost. */
1101 if (wr_bad->send_flags & IBV_SEND_SIGNALED) {
1106 for (j = 0; j < wr_bad->num_sge; ++j)
1107 obytes += wr_bad->sg_list[j].length;
1108 elts_head = (elts_head ? elts_head : elts_n) - 1;
1109 wr_bad = wr_bad->next;
1111 txq->stats.opackets -= opackets;
1112 txq->stats.obytes -= obytes;
1114 DEBUG("%p: ibv_post_send() failed, %" PRIu64 " packets"
1115 " (%" PRIu64 " bytes) rejected: %s",
1119 (err <= -1) ? "Internal error" : strerror(err));
1121 txq->elts_head = elts_head;
1122 txq->elts_comp += elts_comp;
1123 txq->elts_comp_cd = elts_comp_cd;
1128 * Configure a TX queue.
1131 * Pointer to Ethernet device structure.
1133 * Pointer to TX queue structure.
1135 * Number of descriptors to configure in queue.
1137 * NUMA socket on which memory must be allocated.
1139 * Thresholds parameters.
1142 * 0 on success, negative errno value otherwise and rte_errno is set.
1145 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1146 unsigned int socket, const struct rte_eth_txconf *conf)
1148 struct priv *priv = dev->data->dev_private;
1154 struct ibv_qp_init_attr init;
1155 struct ibv_qp_attr mod;
1159 (void)conf; /* Thresholds configuration (ignored). */
1166 ERROR("%p: invalid number of Tx descriptors", (void *)dev);
1169 /* MRs will be registered in mp2mr[] later. */
1170 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
1171 if (tmpl.cq == NULL) {
1173 ERROR("%p: CQ creation failure: %s",
1174 (void *)dev, strerror(rte_errno));
1177 DEBUG("priv->device_attr.max_qp_wr is %d",
1178 priv->device_attr.max_qp_wr);
1179 DEBUG("priv->device_attr.max_sge is %d",
1180 priv->device_attr.max_sge);
1181 attr.init = (struct ibv_qp_init_attr){
1182 /* CQ to be associated with the send queue. */
1184 /* CQ to be associated with the receive queue. */
1187 /* Max number of outstanding WRs. */
1188 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1189 priv->device_attr.max_qp_wr :
1191 /* Max number of scatter/gather elements in a WR. */
1193 .max_inline_data = MLX4_PMD_MAX_INLINE,
1195 .qp_type = IBV_QPT_RAW_PACKET,
1197 * Do *NOT* enable this, completions events are managed per
1202 tmpl.qp = ibv_create_qp(priv->pd, &attr.init);
1203 if (tmpl.qp == NULL) {
1204 rte_errno = errno ? errno : EINVAL;
1205 ERROR("%p: QP creation failure: %s",
1206 (void *)dev, strerror(rte_errno));
1209 /* ibv_create_qp() updates this value. */
1210 tmpl.max_inline = attr.init.cap.max_inline_data;
1211 attr.mod = (struct ibv_qp_attr){
1212 /* Move the QP to this state. */
1213 .qp_state = IBV_QPS_INIT,
1214 /* Primary port number. */
1215 .port_num = priv->port
1217 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE | IBV_QP_PORT);
1220 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1221 (void *)dev, strerror(rte_errno));
1224 ret = txq_alloc_elts(&tmpl, desc);
1227 ERROR("%p: TXQ allocation failed: %s",
1228 (void *)dev, strerror(rte_errno));
1231 attr.mod = (struct ibv_qp_attr){
1232 .qp_state = IBV_QPS_RTR
1234 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1237 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1238 (void *)dev, strerror(rte_errno));
1241 attr.mod.qp_state = IBV_QPS_RTS;
1242 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1245 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1246 (void *)dev, strerror(rte_errno));
1249 /* Clean up txq in case we're reinitializing it. */
1250 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1253 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1254 /* Pre-register known mempools. */
1255 rte_mempool_walk(txq_mp2mr_iter, txq);
1261 assert(rte_errno > 0);
1266 * DPDK callback to configure a TX queue.
1269 * Pointer to Ethernet device structure.
1273 * Number of descriptors to configure in queue.
1275 * NUMA socket on which memory must be allocated.
1277 * Thresholds parameters.
1280 * 0 on success, negative errno value otherwise and rte_errno is set.
1283 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1284 unsigned int socket, const struct rte_eth_txconf *conf)
1286 struct priv *priv = dev->data->dev_private;
1287 struct txq *txq = (*priv->txqs)[idx];
1290 DEBUG("%p: configuring queue %u for %u descriptors",
1291 (void *)dev, idx, desc);
1292 if (idx >= priv->txqs_n) {
1293 rte_errno = EOVERFLOW;
1294 ERROR("%p: queue index out of range (%u >= %u)",
1295 (void *)dev, idx, priv->txqs_n);
1299 DEBUG("%p: reusing already allocated queue index %u (%p)",
1300 (void *)dev, idx, (void *)txq);
1301 if (priv->started) {
1305 (*priv->txqs)[idx] = NULL;
1308 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1311 ERROR("%p: unable to allocate queue index %u",
1316 ret = txq_setup(dev, txq, desc, socket, conf);
1320 txq->stats.idx = idx;
1321 DEBUG("%p: adding TX queue %p to list",
1322 (void *)dev, (void *)txq);
1323 (*priv->txqs)[idx] = txq;
1324 /* Update send callback. */
1325 dev->tx_pkt_burst = mlx4_tx_burst;
1331 * DPDK callback to release a TX queue.
1334 * Generic TX queue pointer.
1337 mlx4_tx_queue_release(void *dpdk_txq)
1339 struct txq *txq = (struct txq *)dpdk_txq;
1346 for (i = 0; (i != priv->txqs_n); ++i)
1347 if ((*priv->txqs)[i] == txq) {
1348 DEBUG("%p: removing TX queue %p from list",
1349 (void *)priv->dev, (void *)txq);
1350 (*priv->txqs)[i] = NULL;
1357 /* RX queues handling. */
1360 * Allocate RX queue elements.
1363 * Pointer to RX queue structure.
1365 * Number of elements to allocate.
1368 * 0 on success, negative errno value otherwise and rte_errno is set.
1371 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n)
1374 struct rxq_elt (*elts)[elts_n] =
1375 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1380 ERROR("%p: can't allocate packets array", (void *)rxq);
1383 /* For each WR (packet). */
1384 for (i = 0; (i != elts_n); ++i) {
1385 struct rxq_elt *elt = &(*elts)[i];
1386 struct ibv_recv_wr *wr = &elt->wr;
1387 struct ibv_sge *sge = &(*elts)[i].sge;
1388 struct rte_mbuf *buf = rte_pktmbuf_alloc(rxq->mp);
1392 ERROR("%p: empty mbuf pool", (void *)rxq);
1396 wr->next = &(*elts)[(i + 1)].wr;
1399 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1400 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1401 /* Buffer is supposed to be empty. */
1402 assert(rte_pktmbuf_data_len(buf) == 0);
1403 assert(rte_pktmbuf_pkt_len(buf) == 0);
1404 /* sge->addr must be able to store a pointer. */
1405 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1406 /* SGE keeps its headroom. */
1407 sge->addr = (uintptr_t)
1408 ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
1409 sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
1410 sge->lkey = rxq->mr->lkey;
1411 /* Redundant check for tailroom. */
1412 assert(sge->length == rte_pktmbuf_tailroom(buf));
1414 /* The last WR pointer must be NULL. */
1415 (*elts)[(i - 1)].wr.next = NULL;
1416 DEBUG("%p: allocated and configured %u single-segment WRs",
1417 (void *)rxq, elts_n);
1418 rxq->elts_n = elts_n;
1424 for (i = 0; (i != elemof(*elts)); ++i)
1425 rte_pktmbuf_free_seg((*elts)[i].buf);
1428 DEBUG("%p: failed, freed everything", (void *)rxq);
1429 assert(rte_errno > 0);
1434 * Free RX queue elements.
1437 * Pointer to RX queue structure.
1440 rxq_free_elts(struct rxq *rxq)
1443 unsigned int elts_n = rxq->elts_n;
1444 struct rxq_elt (*elts)[elts_n] = rxq->elts;
1446 DEBUG("%p: freeing WRs", (void *)rxq);
1451 for (i = 0; (i != elemof(*elts)); ++i)
1452 rte_pktmbuf_free_seg((*elts)[i].buf);
1457 * Unregister a MAC address.
1460 * Pointer to private structure.
1463 priv_mac_addr_del(struct priv *priv)
1466 uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
1469 if (!priv->mac_flow)
1471 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x",
1473 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
1474 claim_zero(ibv_destroy_flow(priv->mac_flow));
1475 priv->mac_flow = NULL;
1479 * Register a MAC address.
1481 * The MAC address is registered in queue 0.
1484 * Pointer to private structure.
1487 * 0 on success, negative errno value otherwise and rte_errno is set.
1490 priv_mac_addr_add(struct priv *priv)
1492 uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
1494 struct ibv_flow *flow;
1496 /* If device isn't started, this is all we need to do. */
1501 if (*priv->rxqs && (*priv->rxqs)[0])
1502 rxq = (*priv->rxqs)[0];
1506 /* Allocate flow specification on the stack. */
1507 struct __attribute__((packed)) {
1508 struct ibv_flow_attr attr;
1509 struct ibv_flow_spec_eth spec;
1511 struct ibv_flow_attr *attr = &data.attr;
1512 struct ibv_flow_spec_eth *spec = &data.spec;
1515 priv_mac_addr_del(priv);
1517 * No padding must be inserted by the compiler between attr and spec.
1518 * This layout is expected by libibverbs.
1520 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
1521 *attr = (struct ibv_flow_attr){
1522 .type = IBV_FLOW_ATTR_NORMAL,
1528 *spec = (struct ibv_flow_spec_eth){
1529 .type = IBV_FLOW_SPEC_ETH,
1530 .size = sizeof(*spec),
1533 (*mac)[0], (*mac)[1], (*mac)[2],
1534 (*mac)[3], (*mac)[4], (*mac)[5]
1538 .dst_mac = "\xff\xff\xff\xff\xff\xff",
1541 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x",
1543 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
1544 /* Create related flow. */
1545 flow = ibv_create_flow(rxq->qp, attr);
1547 rte_errno = errno ? errno : EINVAL;
1548 ERROR("%p: flow configuration failed, errno=%d: %s",
1549 (void *)rxq, rte_errno, strerror(errno));
1552 assert(priv->mac_flow == NULL);
1553 priv->mac_flow = flow;
1558 * Clean up a RX queue.
1560 * Destroy objects, free allocated memory and reset the structure for reuse.
1563 * Pointer to RX queue structure.
1566 rxq_cleanup(struct rxq *rxq)
1568 DEBUG("cleaning up %p", (void *)rxq);
1570 if (rxq->qp != NULL)
1571 claim_zero(ibv_destroy_qp(rxq->qp));
1572 if (rxq->cq != NULL)
1573 claim_zero(ibv_destroy_cq(rxq->cq));
1574 if (rxq->channel != NULL)
1575 claim_zero(ibv_destroy_comp_channel(rxq->channel));
1576 if (rxq->mr != NULL)
1577 claim_zero(ibv_dereg_mr(rxq->mr));
1578 memset(rxq, 0, sizeof(*rxq));
1582 * DPDK callback for RX.
1584 * The following function doesn't manage scattered packets.
1587 * Generic pointer to RX queue structure.
1589 * Array to store received packets.
1591 * Maximum number of packets in array.
1594 * Number of packets successfully received (<= pkts_n).
1597 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1599 struct rxq *rxq = (struct rxq *)dpdk_rxq;
1600 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts;
1601 const unsigned int elts_n = rxq->elts_n;
1602 unsigned int elts_head = rxq->elts_head;
1603 struct ibv_wc wcs[pkts_n];
1604 struct ibv_recv_wr *wr_head = NULL;
1605 struct ibv_recv_wr **wr_next = &wr_head;
1606 struct ibv_recv_wr *wr_bad = NULL;
1608 unsigned int pkts_ret = 0;
1611 ret = ibv_poll_cq(rxq->cq, pkts_n, wcs);
1612 if (unlikely(ret == 0))
1614 if (unlikely(ret < 0)) {
1615 DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
1619 assert(ret <= (int)pkts_n);
1620 /* For each work completion. */
1621 for (i = 0; i != (unsigned int)ret; ++i) {
1622 struct ibv_wc *wc = &wcs[i];
1623 struct rxq_elt *elt = &(*elts)[elts_head];
1624 struct ibv_recv_wr *wr = &elt->wr;
1625 uint32_t len = wc->byte_len;
1626 struct rte_mbuf *seg = elt->buf;
1627 struct rte_mbuf *rep;
1629 /* Sanity checks. */
1630 assert(wr->sg_list == &elt->sge);
1631 assert(wr->num_sge == 1);
1632 assert(elts_head < rxq->elts_n);
1633 assert(rxq->elts_head < rxq->elts_n);
1635 * Fetch initial bytes of packet descriptor into a
1636 * cacheline while allocating rep.
1638 rte_mbuf_prefetch_part1(seg);
1639 rte_mbuf_prefetch_part2(seg);
1640 /* Link completed WRs together for repost. */
1642 wr_next = &wr->next;
1643 if (unlikely(wc->status != IBV_WC_SUCCESS)) {
1644 /* Whatever, just repost the offending WR. */
1645 DEBUG("rxq=%p: bad work completion status (%d): %s",
1646 (void *)rxq, wc->status,
1647 ibv_wc_status_str(wc->status));
1648 /* Increment dropped packets counter. */
1649 ++rxq->stats.idropped;
1652 rep = rte_mbuf_raw_alloc(rxq->mp);
1653 if (unlikely(rep == NULL)) {
1655 * Unable to allocate a replacement mbuf,
1658 DEBUG("rxq=%p: can't allocate a new mbuf",
1660 /* Increase out of memory counters. */
1661 ++rxq->stats.rx_nombuf;
1662 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
1665 /* Reconfigure sge to use rep instead of seg. */
1666 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
1667 assert(elt->sge.lkey == rxq->mr->lkey);
1669 /* Update seg information. */
1670 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
1672 PORT(seg) = rxq->port_id;
1675 DATA_LEN(seg) = len;
1676 seg->packet_type = 0;
1678 /* Return packet. */
1681 /* Increase bytes counter. */
1682 rxq->stats.ibytes += len;
1684 if (++elts_head >= elts_n)
1688 if (unlikely(i == 0))
1693 ret = ibv_post_recv(rxq->qp, wr_head, &wr_bad);
1694 if (unlikely(ret)) {
1695 /* Inability to repost WRs is fatal. */
1696 DEBUG("%p: recv_burst(): failed (ret=%d)",
1701 rxq->elts_head = elts_head;
1702 /* Increase packets counter. */
1703 rxq->stats.ipackets += pkts_ret;
1708 * Allocate a Queue Pair.
1709 * Optionally setup inline receive if supported.
1712 * Pointer to private structure.
1714 * Completion queue to associate with QP.
1716 * Number of descriptors in QP (hint only).
1719 * QP pointer or NULL in case of error and rte_errno is set.
1721 static struct ibv_qp *
1722 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
1725 struct ibv_qp_init_attr attr = {
1726 /* CQ to be associated with the send queue. */
1728 /* CQ to be associated with the receive queue. */
1731 /* Max number of outstanding WRs. */
1732 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
1733 priv->device_attr.max_qp_wr :
1735 /* Max number of scatter/gather elements in a WR. */
1738 .qp_type = IBV_QPT_RAW_PACKET,
1741 qp = ibv_create_qp(priv->pd, &attr);
1743 rte_errno = errno ? errno : EINVAL;
1748 * Configure a RX queue.
1751 * Pointer to Ethernet device structure.
1753 * Pointer to RX queue structure.
1755 * Number of descriptors to configure in queue.
1757 * NUMA socket on which memory must be allocated.
1759 * Thresholds parameters.
1761 * Memory pool for buffer allocations.
1764 * 0 on success, negative errno value otherwise and rte_errno is set.
1767 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
1768 unsigned int socket, const struct rte_eth_rxconf *conf,
1769 struct rte_mempool *mp)
1771 struct priv *priv = dev->data->dev_private;
1777 struct ibv_qp_attr mod;
1778 struct ibv_recv_wr *bad_wr;
1779 unsigned int mb_len;
1782 (void)conf; /* Thresholds configuration (ignored). */
1783 mb_len = rte_pktmbuf_data_room_size(mp);
1786 ERROR("%p: invalid number of Rx descriptors", (void *)dev);
1789 /* Enable scattered packets support for this queue if necessary. */
1790 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
1791 if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
1792 (mb_len - RTE_PKTMBUF_HEADROOM)) {
1794 } else if (dev->data->dev_conf.rxmode.enable_scatter) {
1795 WARN("%p: scattered mode has been requested but is"
1796 " not supported, this may lead to packet loss",
1799 WARN("%p: the requested maximum Rx packet size (%u) is"
1800 " larger than a single mbuf (%u) and scattered"
1801 " mode has not been requested",
1803 dev->data->dev_conf.rxmode.max_rx_pkt_len,
1804 mb_len - RTE_PKTMBUF_HEADROOM);
1806 /* Use the entire RX mempool as the memory region. */
1807 tmpl.mr = mlx4_mp2mr(priv->pd, mp);
1808 if (tmpl.mr == NULL) {
1810 ERROR("%p: MR creation failure: %s",
1811 (void *)dev, strerror(rte_errno));
1814 if (dev->data->dev_conf.intr_conf.rxq) {
1815 tmpl.channel = ibv_create_comp_channel(priv->ctx);
1816 if (tmpl.channel == NULL) {
1818 ERROR("%p: Rx interrupt completion channel creation"
1820 (void *)dev, strerror(rte_errno));
1824 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, tmpl.channel, 0);
1825 if (tmpl.cq == NULL) {
1827 ERROR("%p: CQ creation failure: %s",
1828 (void *)dev, strerror(rte_errno));
1831 DEBUG("priv->device_attr.max_qp_wr is %d",
1832 priv->device_attr.max_qp_wr);
1833 DEBUG("priv->device_attr.max_sge is %d",
1834 priv->device_attr.max_sge);
1835 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc);
1836 if (tmpl.qp == NULL) {
1837 ERROR("%p: QP creation failure: %s",
1838 (void *)dev, strerror(rte_errno));
1841 mod = (struct ibv_qp_attr){
1842 /* Move the QP to this state. */
1843 .qp_state = IBV_QPS_INIT,
1844 /* Primary port number. */
1845 .port_num = priv->port
1847 ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE | IBV_QP_PORT);
1850 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1851 (void *)dev, strerror(rte_errno));
1854 ret = rxq_alloc_elts(&tmpl, desc);
1856 ERROR("%p: RXQ allocation failed: %s",
1857 (void *)dev, strerror(rte_errno));
1860 ret = ibv_post_recv(tmpl.qp, &(*tmpl.elts)[0].wr, &bad_wr);
1863 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
1866 strerror(rte_errno));
1869 mod = (struct ibv_qp_attr){
1870 .qp_state = IBV_QPS_RTR
1872 ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE);
1875 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1876 (void *)dev, strerror(rte_errno));
1880 tmpl.port_id = dev->data->port_id;
1881 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
1882 /* Clean up rxq in case we're reinitializing it. */
1883 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
1886 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
1892 assert(rte_errno > 0);
1897 * DPDK callback to configure a RX queue.
1900 * Pointer to Ethernet device structure.
1904 * Number of descriptors to configure in queue.
1906 * NUMA socket on which memory must be allocated.
1908 * Thresholds parameters.
1910 * Memory pool for buffer allocations.
1913 * 0 on success, negative errno value otherwise and rte_errno is set.
1916 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1917 unsigned int socket, const struct rte_eth_rxconf *conf,
1918 struct rte_mempool *mp)
1920 struct priv *priv = dev->data->dev_private;
1921 struct rxq *rxq = (*priv->rxqs)[idx];
1924 DEBUG("%p: configuring queue %u for %u descriptors",
1925 (void *)dev, idx, desc);
1926 if (idx >= priv->rxqs_n) {
1927 rte_errno = EOVERFLOW;
1928 ERROR("%p: queue index out of range (%u >= %u)",
1929 (void *)dev, idx, priv->rxqs_n);
1933 DEBUG("%p: reusing already allocated queue index %u (%p)",
1934 (void *)dev, idx, (void *)rxq);
1935 if (priv->started) {
1939 (*priv->rxqs)[idx] = NULL;
1941 priv_mac_addr_del(priv);
1944 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
1947 ERROR("%p: unable to allocate queue index %u",
1952 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
1956 rxq->stats.idx = idx;
1957 DEBUG("%p: adding RX queue %p to list",
1958 (void *)dev, (void *)rxq);
1959 (*priv->rxqs)[idx] = rxq;
1960 /* Update receive callback. */
1961 dev->rx_pkt_burst = mlx4_rx_burst;
1967 * DPDK callback to release a RX queue.
1970 * Generic RX queue pointer.
1973 mlx4_rx_queue_release(void *dpdk_rxq)
1975 struct rxq *rxq = (struct rxq *)dpdk_rxq;
1982 for (i = 0; (i != priv->rxqs_n); ++i)
1983 if ((*priv->rxqs)[i] == rxq) {
1984 DEBUG("%p: removing RX queue %p from list",
1985 (void *)priv->dev, (void *)rxq);
1986 (*priv->rxqs)[i] = NULL;
1988 priv_mac_addr_del(priv);
1996 priv_dev_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
1999 priv_dev_removal_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
2002 priv_dev_link_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
2005 * DPDK callback to start the device.
2007 * Simulate device start by attaching all configured flows.
2010 * Pointer to Ethernet device structure.
2013 * 0 on success, negative errno value otherwise and rte_errno is set.
2016 mlx4_dev_start(struct rte_eth_dev *dev)
2018 struct priv *priv = dev->data->dev_private;
2023 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
2025 ret = priv_mac_addr_add(priv);
2028 ret = priv_dev_link_interrupt_handler_install(priv, dev);
2030 ERROR("%p: LSC handler install failed",
2034 ret = priv_dev_removal_interrupt_handler_install(priv, dev);
2036 ERROR("%p: RMV handler install failed",
2040 ret = priv_rx_intr_vec_enable(priv);
2042 ERROR("%p: Rx interrupt vector creation failed",
2046 ret = mlx4_priv_flow_start(priv);
2048 ERROR("%p: flow start failed: %s",
2049 (void *)dev, strerror(ret));
2055 priv_mac_addr_del(priv);
2061 * DPDK callback to stop the device.
2063 * Simulate device stop by detaching all configured flows.
2066 * Pointer to Ethernet device structure.
2069 mlx4_dev_stop(struct rte_eth_dev *dev)
2071 struct priv *priv = dev->data->dev_private;
2075 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
2077 mlx4_priv_flow_stop(priv);
2078 priv_mac_addr_del(priv);
2082 * Dummy DPDK callback for TX.
2084 * This function is used to temporarily replace the real callback during
2085 * unsafe control operations on the queue, or in case of error.
2088 * Generic pointer to TX queue structure.
2090 * Packets to transmit.
2092 * Number of packets in array.
2095 * Number of packets successfully transmitted (<= pkts_n).
2098 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
2107 * Dummy DPDK callback for RX.
2109 * This function is used to temporarily replace the real callback during
2110 * unsafe control operations on the queue, or in case of error.
2113 * Generic pointer to RX queue structure.
2115 * Array to store received packets.
2117 * Maximum number of packets in array.
2120 * Number of packets successfully received (<= pkts_n).
2123 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2132 priv_dev_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
2135 priv_dev_removal_interrupt_handler_uninstall(struct priv *,
2136 struct rte_eth_dev *);
2139 priv_dev_link_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
2142 * DPDK callback to close the device.
2144 * Destroy all queues and objects, free memory.
2147 * Pointer to Ethernet device structure.
2150 mlx4_dev_close(struct rte_eth_dev *dev)
2152 struct priv *priv = dev->data->dev_private;
2158 DEBUG("%p: closing device \"%s\"",
2160 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
2161 priv_mac_addr_del(priv);
2163 * Prevent crashes when queues are still in use. This is unfortunately
2164 * still required for DPDK 1.3 because some programs (such as testpmd)
2165 * never release them before closing the device.
2167 dev->rx_pkt_burst = removed_rx_burst;
2168 dev->tx_pkt_burst = removed_tx_burst;
2169 if (priv->rxqs != NULL) {
2170 /* XXX race condition if mlx4_rx_burst() is still running. */
2172 for (i = 0; (i != priv->rxqs_n); ++i) {
2173 tmp = (*priv->rxqs)[i];
2176 (*priv->rxqs)[i] = NULL;
2183 if (priv->txqs != NULL) {
2184 /* XXX race condition if mlx4_tx_burst() is still running. */
2186 for (i = 0; (i != priv->txqs_n); ++i) {
2187 tmp = (*priv->txqs)[i];
2190 (*priv->txqs)[i] = NULL;
2197 if (priv->pd != NULL) {
2198 assert(priv->ctx != NULL);
2199 claim_zero(ibv_dealloc_pd(priv->pd));
2200 claim_zero(ibv_close_device(priv->ctx));
2202 assert(priv->ctx == NULL);
2203 priv_dev_removal_interrupt_handler_uninstall(priv, dev);
2204 priv_dev_link_interrupt_handler_uninstall(priv, dev);
2205 priv_rx_intr_vec_disable(priv);
2206 memset(priv, 0, sizeof(*priv));
2210 * Change the link state (UP / DOWN).
2213 * Pointer to Ethernet device private data.
2215 * Nonzero for link up, otherwise link down.
2218 * 0 on success, negative errno value otherwise and rte_errno is set.
2221 priv_set_link(struct priv *priv, int up)
2223 struct rte_eth_dev *dev = priv->dev;
2227 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
2230 dev->rx_pkt_burst = mlx4_rx_burst;
2232 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
2235 dev->rx_pkt_burst = removed_rx_burst;
2236 dev->tx_pkt_burst = removed_tx_burst;
2242 * DPDK callback to bring the link DOWN.
2245 * Pointer to Ethernet device structure.
2248 * 0 on success, negative errno value otherwise and rte_errno is set.
2251 mlx4_set_link_down(struct rte_eth_dev *dev)
2253 struct priv *priv = dev->data->dev_private;
2255 return priv_set_link(priv, 0);
2259 * DPDK callback to bring the link UP.
2262 * Pointer to Ethernet device structure.
2265 * 0 on success, negative errno value otherwise and rte_errno is set.
2268 mlx4_set_link_up(struct rte_eth_dev *dev)
2270 struct priv *priv = dev->data->dev_private;
2272 return priv_set_link(priv, 1);
2276 * DPDK callback to get information about the device.
2279 * Pointer to Ethernet device structure.
2281 * Info structure output buffer.
2284 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
2286 struct priv *priv = dev->data->dev_private;
2288 char ifname[IF_NAMESIZE];
2290 info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
2293 /* FIXME: we should ask the device for these values. */
2294 info->min_rx_bufsize = 32;
2295 info->max_rx_pktlen = 65536;
2297 * Since we need one CQ per QP, the limit is the minimum number
2298 * between the two values.
2300 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
2301 priv->device_attr.max_qp : priv->device_attr.max_cq);
2302 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
2305 info->max_rx_queues = max;
2306 info->max_tx_queues = max;
2307 /* Last array entry is reserved for broadcast. */
2308 info->max_mac_addrs = 1;
2309 info->rx_offload_capa = 0;
2310 info->tx_offload_capa = 0;
2311 if (priv_get_ifname(priv, &ifname) == 0)
2312 info->if_index = if_nametoindex(ifname);
2315 ETH_LINK_SPEED_10G |
2316 ETH_LINK_SPEED_20G |
2317 ETH_LINK_SPEED_40G |
2322 * DPDK callback to get device statistics.
2325 * Pointer to Ethernet device structure.
2327 * Stats structure output buffer.
2330 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
2332 struct priv *priv = dev->data->dev_private;
2333 struct rte_eth_stats tmp = {0};
2339 /* Add software counters. */
2340 for (i = 0; (i != priv->rxqs_n); ++i) {
2341 struct rxq *rxq = (*priv->rxqs)[i];
2345 idx = rxq->stats.idx;
2346 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
2347 tmp.q_ipackets[idx] += rxq->stats.ipackets;
2348 tmp.q_ibytes[idx] += rxq->stats.ibytes;
2349 tmp.q_errors[idx] += (rxq->stats.idropped +
2350 rxq->stats.rx_nombuf);
2352 tmp.ipackets += rxq->stats.ipackets;
2353 tmp.ibytes += rxq->stats.ibytes;
2354 tmp.ierrors += rxq->stats.idropped;
2355 tmp.rx_nombuf += rxq->stats.rx_nombuf;
2357 for (i = 0; (i != priv->txqs_n); ++i) {
2358 struct txq *txq = (*priv->txqs)[i];
2362 idx = txq->stats.idx;
2363 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
2364 tmp.q_opackets[idx] += txq->stats.opackets;
2365 tmp.q_obytes[idx] += txq->stats.obytes;
2366 tmp.q_errors[idx] += txq->stats.odropped;
2368 tmp.opackets += txq->stats.opackets;
2369 tmp.obytes += txq->stats.obytes;
2370 tmp.oerrors += txq->stats.odropped;
2376 * DPDK callback to clear device statistics.
2379 * Pointer to Ethernet device structure.
2382 mlx4_stats_reset(struct rte_eth_dev *dev)
2384 struct priv *priv = dev->data->dev_private;
2390 for (i = 0; (i != priv->rxqs_n); ++i) {
2391 if ((*priv->rxqs)[i] == NULL)
2393 idx = (*priv->rxqs)[i]->stats.idx;
2394 (*priv->rxqs)[i]->stats =
2395 (struct mlx4_rxq_stats){ .idx = idx };
2397 for (i = 0; (i != priv->txqs_n); ++i) {
2398 if ((*priv->txqs)[i] == NULL)
2400 idx = (*priv->txqs)[i]->stats.idx;
2401 (*priv->txqs)[i]->stats =
2402 (struct mlx4_txq_stats){ .idx = idx };
2407 * DPDK callback to retrieve physical link information.
2410 * Pointer to Ethernet device structure.
2411 * @param wait_to_complete
2412 * Wait for request completion (ignored).
2415 * 0 on success, negative errno value otherwise and rte_errno is set.
2418 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
2420 const struct priv *priv = dev->data->dev_private;
2421 struct ethtool_cmd edata = {
2425 struct rte_eth_link dev_link;
2432 (void)wait_to_complete;
2433 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
2434 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(rte_errno));
2437 memset(&dev_link, 0, sizeof(dev_link));
2438 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
2439 (ifr.ifr_flags & IFF_RUNNING));
2440 ifr.ifr_data = (void *)&edata;
2441 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2442 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
2443 strerror(rte_errno));
2446 link_speed = ethtool_cmd_speed(&edata);
2447 if (link_speed == -1)
2448 dev_link.link_speed = 0;
2450 dev_link.link_speed = link_speed;
2451 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
2452 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
2453 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
2454 ETH_LINK_SPEED_FIXED);
2455 dev->data->dev_link = dev_link;
2460 * DPDK callback to change the MTU.
2463 * Pointer to Ethernet device structure.
2468 * 0 on success, negative errno value otherwise and rte_errno is set.
2471 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
2473 struct priv *priv = dev->data->dev_private;
2476 /* Set kernel interface MTU first. */
2477 if (priv_set_mtu(priv, mtu)) {
2479 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
2480 strerror(rte_errno));
2483 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
2491 * DPDK callback to get flow control status.
2494 * Pointer to Ethernet device structure.
2495 * @param[out] fc_conf
2496 * Flow control output buffer.
2499 * 0 on success, negative errno value otherwise and rte_errno is set.
2502 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
2504 struct priv *priv = dev->data->dev_private;
2506 struct ethtool_pauseparam ethpause = {
2507 .cmd = ETHTOOL_GPAUSEPARAM
2511 ifr.ifr_data = (void *)ðpause;
2512 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2514 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
2516 strerror(rte_errno));
2519 fc_conf->autoneg = ethpause.autoneg;
2520 if (ethpause.rx_pause && ethpause.tx_pause)
2521 fc_conf->mode = RTE_FC_FULL;
2522 else if (ethpause.rx_pause)
2523 fc_conf->mode = RTE_FC_RX_PAUSE;
2524 else if (ethpause.tx_pause)
2525 fc_conf->mode = RTE_FC_TX_PAUSE;
2527 fc_conf->mode = RTE_FC_NONE;
2535 * DPDK callback to modify flow control parameters.
2538 * Pointer to Ethernet device structure.
2539 * @param[in] fc_conf
2540 * Flow control parameters.
2543 * 0 on success, negative errno value otherwise and rte_errno is set.
2546 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
2548 struct priv *priv = dev->data->dev_private;
2550 struct ethtool_pauseparam ethpause = {
2551 .cmd = ETHTOOL_SPAUSEPARAM
2555 ifr.ifr_data = (void *)ðpause;
2556 ethpause.autoneg = fc_conf->autoneg;
2557 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
2558 (fc_conf->mode & RTE_FC_RX_PAUSE))
2559 ethpause.rx_pause = 1;
2561 ethpause.rx_pause = 0;
2562 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
2563 (fc_conf->mode & RTE_FC_TX_PAUSE))
2564 ethpause.tx_pause = 1;
2566 ethpause.tx_pause = 0;
2567 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2569 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
2571 strerror(rte_errno));
2580 const struct rte_flow_ops mlx4_flow_ops = {
2581 .validate = mlx4_flow_validate,
2582 .create = mlx4_flow_create,
2583 .destroy = mlx4_flow_destroy,
2584 .flush = mlx4_flow_flush,
2586 .isolate = mlx4_flow_isolate,
2590 * Manage filter operations.
2593 * Pointer to Ethernet device structure.
2594 * @param filter_type
2597 * Operation to perform.
2599 * Pointer to operation-specific structure.
2602 * 0 on success, negative errno value otherwise and rte_errno is set.
2605 mlx4_dev_filter_ctrl(struct rte_eth_dev *dev,
2606 enum rte_filter_type filter_type,
2607 enum rte_filter_op filter_op,
2610 switch (filter_type) {
2611 case RTE_ETH_FILTER_GENERIC:
2612 if (filter_op != RTE_ETH_FILTER_GET)
2614 *(const void **)arg = &mlx4_flow_ops;
2617 ERROR("%p: filter type (%d) not supported",
2618 (void *)dev, filter_type);
2621 rte_errno = ENOTSUP;
2625 static const struct eth_dev_ops mlx4_dev_ops = {
2626 .dev_configure = mlx4_dev_configure,
2627 .dev_start = mlx4_dev_start,
2628 .dev_stop = mlx4_dev_stop,
2629 .dev_set_link_down = mlx4_set_link_down,
2630 .dev_set_link_up = mlx4_set_link_up,
2631 .dev_close = mlx4_dev_close,
2632 .link_update = mlx4_link_update,
2633 .stats_get = mlx4_stats_get,
2634 .stats_reset = mlx4_stats_reset,
2635 .dev_infos_get = mlx4_dev_infos_get,
2636 .rx_queue_setup = mlx4_rx_queue_setup,
2637 .tx_queue_setup = mlx4_tx_queue_setup,
2638 .rx_queue_release = mlx4_rx_queue_release,
2639 .tx_queue_release = mlx4_tx_queue_release,
2640 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
2641 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
2642 .mtu_set = mlx4_dev_set_mtu,
2643 .filter_ctrl = mlx4_dev_filter_ctrl,
2644 .rx_queue_intr_enable = mlx4_rx_intr_enable,
2645 .rx_queue_intr_disable = mlx4_rx_intr_disable,
2649 * Get PCI information from struct ibv_device.
2652 * Pointer to Ethernet device structure.
2653 * @param[out] pci_addr
2654 * PCI bus address output buffer.
2657 * 0 on success, negative errno value otherwise and rte_errno is set.
2660 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
2661 struct rte_pci_addr *pci_addr)
2665 MKSTR(path, "%s/device/uevent", device->ibdev_path);
2667 file = fopen(path, "rb");
2672 while (fgets(line, sizeof(line), file) == line) {
2673 size_t len = strlen(line);
2676 /* Truncate long lines. */
2677 if (len == (sizeof(line) - 1))
2678 while (line[(len - 1)] != '\n') {
2682 line[(len - 1)] = ret;
2684 /* Extract information. */
2687 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
2691 &pci_addr->function) == 4) {
2701 * Get MAC address by querying netdevice.
2704 * struct priv for the requested device.
2706 * MAC address output buffer.
2709 * 0 on success, negative errno value otherwise and rte_errno is set.
2712 priv_get_mac(struct priv *priv, uint8_t (*mac)[ETHER_ADDR_LEN])
2714 struct ifreq request;
2715 int ret = priv_ifreq(priv, SIOCGIFHWADDR, &request);
2719 memcpy(mac, request.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
2724 mlx4_dev_link_status_handler(void *);
2726 mlx4_dev_interrupt_handler(void *);
2729 * Link/device status handler.
2732 * Pointer to private structure.
2734 * Pointer to the rte_eth_dev structure.
2736 * Pointer to event flags holder.
2742 priv_dev_status_handler(struct priv *priv, struct rte_eth_dev *dev,
2745 struct ibv_async_event event;
2746 int port_change = 0;
2747 struct rte_eth_link *link = &dev->data->dev_link;
2751 /* Read all message and acknowledge them. */
2753 if (ibv_get_async_event(priv->ctx, &event))
2755 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
2756 event.event_type == IBV_EVENT_PORT_ERR) &&
2757 (priv->intr_conf.lsc == 1)) {
2760 } else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
2761 priv->intr_conf.rmv == 1) {
2762 *events |= (1 << RTE_ETH_EVENT_INTR_RMV);
2765 DEBUG("event type %d on port %d not handled",
2766 event.event_type, event.element.port_num);
2767 ibv_ack_async_event(&event);
2771 mlx4_link_update(dev, 0);
2772 if (((link->link_speed == 0) && link->link_status) ||
2773 ((link->link_speed != 0) && !link->link_status)) {
2774 if (!priv->pending_alarm) {
2775 /* Inconsistent status, check again later. */
2776 priv->pending_alarm = 1;
2777 rte_eal_alarm_set(MLX4_ALARM_TIMEOUT_US,
2778 mlx4_dev_link_status_handler,
2782 *events |= (1 << RTE_ETH_EVENT_INTR_LSC);
2788 * Handle delayed link status event.
2791 * Registered argument.
2794 mlx4_dev_link_status_handler(void *arg)
2796 struct rte_eth_dev *dev = arg;
2797 struct priv *priv = dev->data->dev_private;
2801 assert(priv->pending_alarm == 1);
2802 priv->pending_alarm = 0;
2803 ret = priv_dev_status_handler(priv, dev, &events);
2804 if (ret > 0 && events & (1 << RTE_ETH_EVENT_INTR_LSC))
2805 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC,
2810 * Handle interrupts from the NIC.
2812 * @param[in] intr_handle
2813 * Interrupt handler.
2815 * Callback argument.
2818 mlx4_dev_interrupt_handler(void *cb_arg)
2820 struct rte_eth_dev *dev = cb_arg;
2821 struct priv *priv = dev->data->dev_private;
2826 ret = priv_dev_status_handler(priv, dev, &ev);
2828 for (i = RTE_ETH_EVENT_UNKNOWN;
2829 i < RTE_ETH_EVENT_MAX;
2831 if (ev & (1 << i)) {
2833 _rte_eth_dev_callback_process(dev, i, NULL,
2839 WARN("%d event%s not processed", ret,
2840 (ret > 1 ? "s were" : " was"));
2845 * Uninstall interrupt handler.
2848 * Pointer to private structure.
2850 * Pointer to the rte_eth_dev structure.
2853 * 0 on success, negative errno value otherwise and rte_errno is set.
2856 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
2860 if (priv->intr_conf.lsc ||
2861 priv->intr_conf.rmv)
2863 ret = rte_intr_callback_unregister(&priv->intr_handle,
2864 mlx4_dev_interrupt_handler,
2868 ERROR("rte_intr_callback_unregister failed with %d %s",
2869 ret, strerror(rte_errno));
2871 priv->intr_handle.fd = 0;
2872 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
2877 * Install interrupt handler.
2880 * Pointer to private structure.
2882 * Pointer to the rte_eth_dev structure.
2885 * 0 on success, negative errno value otherwise and rte_errno is set.
2888 priv_dev_interrupt_handler_install(struct priv *priv,
2889 struct rte_eth_dev *dev)
2895 * Check whether the interrupt handler has already been installed
2896 * for either type of interrupt.
2898 if (priv->intr_conf.lsc &&
2899 priv->intr_conf.rmv &&
2900 priv->intr_handle.fd)
2902 assert(priv->ctx->async_fd > 0);
2903 flags = fcntl(priv->ctx->async_fd, F_GETFL);
2904 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2906 rte_errno = errno ? errno : EINVAL;
2907 INFO("failed to change file descriptor async event queue");
2908 dev->data->dev_conf.intr_conf.lsc = 0;
2909 dev->data->dev_conf.intr_conf.rmv = 0;
2912 priv->intr_handle.fd = priv->ctx->async_fd;
2913 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
2914 rc = rte_intr_callback_register(&priv->intr_handle,
2915 mlx4_dev_interrupt_handler,
2919 ERROR("rte_intr_callback_register failed "
2920 " (rte_errno: %s)", strerror(rte_errno));
2928 * Uninstall interrupt handler.
2931 * Pointer to private structure.
2933 * Pointer to the rte_eth_dev structure.
2936 * 0 on success, negative errno value otherwise and rte_errno is set.
2939 priv_dev_removal_interrupt_handler_uninstall(struct priv *priv,
2940 struct rte_eth_dev *dev)
2942 if (dev->data->dev_conf.intr_conf.rmv) {
2943 priv->intr_conf.rmv = 0;
2944 return priv_dev_interrupt_handler_uninstall(priv, dev);
2950 * Uninstall interrupt handler.
2953 * Pointer to private structure.
2955 * Pointer to the rte_eth_dev structure.
2958 * 0 on success, negative errno value otherwise and rte_errno is set.
2961 priv_dev_link_interrupt_handler_uninstall(struct priv *priv,
2962 struct rte_eth_dev *dev)
2966 if (dev->data->dev_conf.intr_conf.lsc) {
2967 priv->intr_conf.lsc = 0;
2968 ret = priv_dev_interrupt_handler_uninstall(priv, dev);
2972 if (priv->pending_alarm)
2973 if (rte_eal_alarm_cancel(mlx4_dev_link_status_handler,
2975 ERROR("rte_eal_alarm_cancel failed "
2976 " (rte_errno: %s)", strerror(rte_errno));
2979 priv->pending_alarm = 0;
2984 * Install link interrupt handler.
2987 * Pointer to private structure.
2989 * Pointer to the rte_eth_dev structure.
2992 * 0 on success, negative errno value otherwise and rte_errno is set.
2995 priv_dev_link_interrupt_handler_install(struct priv *priv,
2996 struct rte_eth_dev *dev)
3000 if (dev->data->dev_conf.intr_conf.lsc) {
3001 ret = priv_dev_interrupt_handler_install(priv, dev);
3004 priv->intr_conf.lsc = 1;
3010 * Install removal interrupt handler.
3013 * Pointer to private structure.
3015 * Pointer to the rte_eth_dev structure.
3018 * 0 on success, negative errno value otherwise and rte_errno is set.
3021 priv_dev_removal_interrupt_handler_install(struct priv *priv,
3022 struct rte_eth_dev *dev)
3026 if (dev->data->dev_conf.intr_conf.rmv) {
3027 ret = priv_dev_interrupt_handler_install(priv, dev);
3030 priv->intr_conf.rmv = 1;
3036 * Allocate queue vector and fill epoll fd list for Rx interrupts.
3039 * Pointer to private structure.
3042 * 0 on success, negative errno value otherwise and rte_errno is set.
3045 priv_rx_intr_vec_enable(struct priv *priv)
3048 unsigned int rxqs_n = priv->rxqs_n;
3049 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
3050 unsigned int count = 0;
3051 struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
3053 if (!priv->dev->data->dev_conf.intr_conf.rxq)
3055 priv_rx_intr_vec_disable(priv);
3056 intr_handle->intr_vec = malloc(sizeof(intr_handle->intr_vec[rxqs_n]));
3057 if (intr_handle->intr_vec == NULL) {
3059 ERROR("failed to allocate memory for interrupt vector,"
3060 " Rx interrupts will not be supported");
3063 intr_handle->type = RTE_INTR_HANDLE_EXT;
3064 for (i = 0; i != n; ++i) {
3065 struct rxq *rxq = (*priv->rxqs)[i];
3070 /* Skip queues that cannot request interrupts. */
3071 if (!rxq || !rxq->channel) {
3072 /* Use invalid intr_vec[] index to disable entry. */
3073 intr_handle->intr_vec[i] =
3074 RTE_INTR_VEC_RXTX_OFFSET +
3075 RTE_MAX_RXTX_INTR_VEC_ID;
3078 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
3080 ERROR("too many Rx queues for interrupt vector size"
3081 " (%d), Rx interrupts cannot be enabled",
3082 RTE_MAX_RXTX_INTR_VEC_ID);
3083 priv_rx_intr_vec_disable(priv);
3086 fd = rxq->channel->fd;
3087 flags = fcntl(fd, F_GETFL);
3088 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
3091 ERROR("failed to make Rx interrupt file descriptor"
3092 " %d non-blocking for queue index %d", fd, i);
3093 priv_rx_intr_vec_disable(priv);
3096 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
3097 intr_handle->efds[count] = fd;
3101 priv_rx_intr_vec_disable(priv);
3103 intr_handle->nb_efd = count;
3108 * Clean up Rx interrupts handler.
3111 * Pointer to private structure.
3114 priv_rx_intr_vec_disable(struct priv *priv)
3116 struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
3118 rte_intr_free_epoll_fd(intr_handle);
3119 free(intr_handle->intr_vec);
3120 intr_handle->nb_efd = 0;
3121 intr_handle->intr_vec = NULL;
3125 * DPDK callback for Rx queue interrupt enable.
3128 * Pointer to Ethernet device structure.
3133 * 0 on success, negative errno value otherwise and rte_errno is set.
3136 mlx4_rx_intr_enable(struct rte_eth_dev *dev, uint16_t idx)
3138 struct priv *priv = dev->data->dev_private;
3139 struct rxq *rxq = (*priv->rxqs)[idx];
3142 if (!rxq || !rxq->channel)
3145 ret = ibv_req_notify_cq(rxq->cq, 0);
3148 WARN("unable to arm interrupt on rx queue %d", idx);
3154 * DPDK callback for Rx queue interrupt disable.
3157 * Pointer to Ethernet device structure.
3162 * 0 on success, negative errno value otherwise and rte_errno is set.
3165 mlx4_rx_intr_disable(struct rte_eth_dev *dev, uint16_t idx)
3167 struct priv *priv = dev->data->dev_private;
3168 struct rxq *rxq = (*priv->rxqs)[idx];
3169 struct ibv_cq *ev_cq;
3173 if (!rxq || !rxq->channel) {
3176 ret = ibv_get_cq_event(rxq->cq->channel, &ev_cq, &ev_ctx);
3177 if (ret || ev_cq != rxq->cq)
3182 WARN("unable to disable interrupt on rx queue %d",
3185 ibv_ack_cq_events(rxq->cq, 1);
3191 * Verify and store value for device argument.
3194 * Key argument to verify.
3196 * Value associated with key.
3197 * @param[in, out] conf
3198 * Shared configuration data.
3201 * 0 on success, negative errno value otherwise and rte_errno is set.
3204 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
3209 tmp = strtoul(val, NULL, 0);
3212 WARN("%s: \"%s\" is not a valid integer", key, val);
3215 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
3216 uint32_t ports = rte_log2_u32(conf->ports.present);
3219 ERROR("port index %lu outside range [0,%" PRIu32 ")",
3223 if (!(conf->ports.present & (1 << tmp))) {
3225 ERROR("invalid port index %lu", tmp);
3228 conf->ports.enabled |= 1 << tmp;
3231 WARN("%s: unknown parameter", key);
3238 * Parse device parameters.
3241 * Device arguments structure.
3244 * 0 on success, negative errno value otherwise and rte_errno is set.
3247 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
3249 struct rte_kvargs *kvlist;
3250 unsigned int arg_count;
3254 if (devargs == NULL)
3256 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
3257 if (kvlist == NULL) {
3259 ERROR("failed to parse kvargs");
3262 /* Process parameters. */
3263 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
3264 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
3265 while (arg_count-- > 0) {
3266 ret = rte_kvargs_process(kvlist,
3267 MLX4_PMD_PORT_KVARG,
3268 (int (*)(const char *,
3278 rte_kvargs_free(kvlist);
3282 static struct rte_pci_driver mlx4_driver;
3285 * DPDK callback to register a PCI device.
3287 * This function creates an Ethernet device for each port of a given
3290 * @param[in] pci_drv
3291 * PCI driver structure (mlx4_driver).
3292 * @param[in] pci_dev
3293 * PCI device information.
3296 * 0 on success, negative errno value otherwise and rte_errno is set.
3299 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
3301 struct ibv_device **list;
3302 struct ibv_device *ibv_dev;
3304 struct ibv_context *attr_ctx = NULL;
3305 struct ibv_device_attr device_attr;
3306 struct mlx4_conf conf = {
3313 assert(pci_drv == &mlx4_driver);
3314 list = ibv_get_device_list(&i);
3318 if (rte_errno == ENOSYS)
3319 ERROR("cannot list devices, is ib_uverbs loaded?");
3324 * For each listed device, check related sysfs entry against
3325 * the provided PCI ID.
3328 struct rte_pci_addr pci_addr;
3331 DEBUG("checking device \"%s\"", list[i]->name);
3332 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
3334 if ((pci_dev->addr.domain != pci_addr.domain) ||
3335 (pci_dev->addr.bus != pci_addr.bus) ||
3336 (pci_dev->addr.devid != pci_addr.devid) ||
3337 (pci_dev->addr.function != pci_addr.function))
3339 vf = (pci_dev->id.device_id ==
3340 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
3341 INFO("PCI information matches, using device \"%s\" (VF: %s)",
3342 list[i]->name, (vf ? "true" : "false"));
3343 attr_ctx = ibv_open_device(list[i]);
3347 if (attr_ctx == NULL) {
3348 ibv_free_device_list(list);
3352 ERROR("cannot access device, is mlx4_ib loaded?");
3356 ERROR("cannot use device, are drivers up to date?");
3364 DEBUG("device opened");
3365 if (ibv_query_device(attr_ctx, &device_attr)) {
3369 INFO("%u port(s) detected", device_attr.phys_port_cnt);
3370 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
3371 if (mlx4_args(pci_dev->device.devargs, &conf)) {
3372 ERROR("failed to process device arguments");
3376 /* Use all ports when none are defined */
3377 if (!conf.ports.enabled)
3378 conf.ports.enabled = conf.ports.present;
3379 for (i = 0; i < device_attr.phys_port_cnt; i++) {
3380 uint32_t port = i + 1; /* ports are indexed from one */
3381 struct ibv_context *ctx = NULL;
3382 struct ibv_port_attr port_attr;
3383 struct ibv_pd *pd = NULL;
3384 struct priv *priv = NULL;
3385 struct rte_eth_dev *eth_dev = NULL;
3386 struct ether_addr mac;
3388 /* If port is not enabled, skip. */
3389 if (!(conf.ports.enabled & (1 << i)))
3391 DEBUG("using port %u", port);
3392 ctx = ibv_open_device(ibv_dev);
3397 /* Check port status. */
3398 err = ibv_query_port(ctx, port, &port_attr);
3401 ERROR("port query failed: %s", strerror(rte_errno));
3404 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
3405 rte_errno = ENOTSUP;
3406 ERROR("port %d is not configured in Ethernet mode",
3410 if (port_attr.state != IBV_PORT_ACTIVE)
3411 DEBUG("port %d is not active: \"%s\" (%d)",
3412 port, ibv_port_state_str(port_attr.state),
3414 /* Allocate protection domain. */
3415 pd = ibv_alloc_pd(ctx);
3418 ERROR("PD allocation failure");
3421 /* from rte_ethdev.c */
3422 priv = rte_zmalloc("ethdev private structure",
3424 RTE_CACHE_LINE_SIZE);
3427 ERROR("priv allocation failure");
3431 priv->device_attr = device_attr;
3434 priv->mtu = ETHER_MTU;
3436 /* Configure the first MAC address by default. */
3437 if (priv_get_mac(priv, &mac.addr_bytes)) {
3438 ERROR("cannot get MAC address, is mlx4_en loaded?"
3439 " (rte_errno: %s)", strerror(rte_errno));
3442 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
3444 mac.addr_bytes[0], mac.addr_bytes[1],
3445 mac.addr_bytes[2], mac.addr_bytes[3],
3446 mac.addr_bytes[4], mac.addr_bytes[5]);
3447 /* Register MAC address. */
3449 if (priv_mac_addr_add(priv))
3453 char ifname[IF_NAMESIZE];
3455 if (priv_get_ifname(priv, &ifname) == 0)
3456 DEBUG("port %u ifname is \"%s\"",
3457 priv->port, ifname);
3459 DEBUG("port %u ifname is unknown", priv->port);
3462 /* Get actual MTU if possible. */
3463 priv_get_mtu(priv, &priv->mtu);
3464 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
3465 /* from rte_ethdev.c */
3467 char name[RTE_ETH_NAME_MAX_LEN];
3469 snprintf(name, sizeof(name), "%s port %u",
3470 ibv_get_device_name(ibv_dev), port);
3471 eth_dev = rte_eth_dev_allocate(name);
3473 if (eth_dev == NULL) {
3474 ERROR("can not allocate rte ethdev");
3478 eth_dev->data->dev_private = priv;
3479 eth_dev->data->mac_addrs = &priv->mac;
3480 eth_dev->device = &pci_dev->device;
3481 rte_eth_copy_pci_info(eth_dev, pci_dev);
3482 eth_dev->device->driver = &mlx4_driver.driver;
3484 * Copy and override interrupt handle to prevent it from
3485 * being shared between all ethdev instances of a given PCI
3486 * device. This is required to properly handle Rx interrupts
3489 priv->intr_handle_dev = *eth_dev->intr_handle;
3490 eth_dev->intr_handle = &priv->intr_handle_dev;
3491 priv->dev = eth_dev;
3492 eth_dev->dev_ops = &mlx4_dev_ops;
3493 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
3494 /* Bring Ethernet device up. */
3495 DEBUG("forcing Ethernet interface up");
3496 priv_set_flags(priv, ~IFF_UP, IFF_UP);
3497 /* Update link status once if waiting for LSC. */
3498 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
3499 mlx4_link_update(eth_dev, 0);
3504 claim_zero(ibv_dealloc_pd(pd));
3506 claim_zero(ibv_close_device(ctx));
3508 rte_eth_dev_release_port(eth_dev);
3511 if (i == device_attr.phys_port_cnt)
3514 * XXX if something went wrong in the loop above, there is a resource
3515 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
3516 * long as the dpdk does not provide a way to deallocate a ethdev and a
3517 * way to enumerate the registered ethdevs to free the previous ones.
3521 claim_zero(ibv_close_device(attr_ctx));
3523 ibv_free_device_list(list);
3524 assert(rte_errno >= 0);
3528 static const struct rte_pci_id mlx4_pci_id_map[] = {
3530 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3531 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
3534 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3535 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
3538 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3539 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
3546 static struct rte_pci_driver mlx4_driver = {
3548 .name = MLX4_DRIVER_NAME
3550 .id_table = mlx4_pci_id_map,
3551 .probe = mlx4_pci_probe,
3552 .drv_flags = RTE_PCI_DRV_INTR_LSC |
3553 RTE_PCI_DRV_INTR_RMV,
3557 * Driver initialization routine.
3559 RTE_INIT(rte_mlx4_pmd_init);
3561 rte_mlx4_pmd_init(void)
3564 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3565 * huge pages. Calling ibv_fork_init() during init allows
3566 * applications to use fork() safely for purposes other than
3567 * using this PMD, which is not supported in forked processes.
3569 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3571 rte_pci_register(&mlx4_driver);
3574 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
3575 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
3576 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
3577 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");