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>
62 #include <rte_spinlock.h>
64 #include <rte_alarm.h>
65 #include <rte_memory.h>
67 #include <rte_kvargs.h>
68 #include <rte_interrupts.h>
69 #include <rte_branch_prediction.h>
71 /* Generated configuration header. */
72 #include "mlx4_autoconf.h"
76 #include "mlx4_flow.h"
78 /* Convenience macros for accessing mbuf fields. */
79 #define NEXT(m) ((m)->next)
80 #define DATA_LEN(m) ((m)->data_len)
81 #define PKT_LEN(m) ((m)->pkt_len)
82 #define DATA_OFF(m) ((m)->data_off)
83 #define SET_DATA_OFF(m, o) ((m)->data_off = (o))
84 #define NB_SEGS(m) ((m)->nb_segs)
85 #define PORT(m) ((m)->port)
87 /** Configuration structure for device arguments. */
90 uint32_t present; /**< Bit-field for existing ports. */
91 uint32_t enabled; /**< Bit-field for user-enabled ports. */
95 /* Available parameters list. */
96 const char *pmd_mlx4_init_params[] = {
102 mlx4_rx_intr_enable(struct rte_eth_dev *dev, uint16_t idx);
105 mlx4_rx_intr_disable(struct rte_eth_dev *dev, uint16_t idx);
108 priv_rx_intr_vec_enable(struct priv *priv);
111 priv_rx_intr_vec_disable(struct priv *priv);
114 * Lock private structure to protect it from concurrent access in the
118 * Pointer to private structure.
120 void priv_lock(struct priv *priv)
122 rte_spinlock_lock(&priv->lock);
126 * Unlock private structure.
129 * Pointer to private structure.
131 void priv_unlock(struct priv *priv)
133 rte_spinlock_unlock(&priv->lock);
136 /* Allocate a buffer on the stack and fill it with a printf format string. */
137 #define MKSTR(name, ...) \
138 char name[snprintf(NULL, 0, __VA_ARGS__) + 1]; \
140 snprintf(name, sizeof(name), __VA_ARGS__)
143 * Get interface name from private structure.
146 * Pointer to private structure.
148 * Interface name output buffer.
151 * 0 on success, -1 on failure and errno is set.
154 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
158 unsigned int dev_type = 0;
159 unsigned int dev_port_prev = ~0u;
160 char match[IF_NAMESIZE] = "";
163 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
169 while ((dent = readdir(dir)) != NULL) {
170 char *name = dent->d_name;
172 unsigned int dev_port;
175 if ((name[0] == '.') &&
176 ((name[1] == '\0') ||
177 ((name[1] == '.') && (name[2] == '\0'))))
180 MKSTR(path, "%s/device/net/%s/%s",
181 priv->ctx->device->ibdev_path, name,
182 (dev_type ? "dev_id" : "dev_port"));
184 file = fopen(path, "rb");
189 * Switch to dev_id when dev_port does not exist as
190 * is the case with Linux kernel versions < 3.15.
201 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
206 * Switch to dev_id when dev_port returns the same value for
207 * all ports. May happen when using a MOFED release older than
208 * 3.0 with a Linux kernel >= 3.15.
210 if (dev_port == dev_port_prev)
212 dev_port_prev = dev_port;
213 if (dev_port == (priv->port - 1u))
214 snprintf(match, sizeof(match), "%s", name);
217 if (match[0] == '\0')
219 strncpy(*ifname, match, sizeof(*ifname));
224 * Read from sysfs entry.
227 * Pointer to private structure.
229 * Entry name relative to sysfs path.
231 * Data output buffer.
236 * 0 on success, -1 on failure and errno is set.
239 priv_sysfs_read(const struct priv *priv, const char *entry,
240 char *buf, size_t size)
242 char ifname[IF_NAMESIZE];
247 if (priv_get_ifname(priv, &ifname))
250 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
253 file = fopen(path, "rb");
256 ret = fread(buf, 1, size, file);
258 if (((size_t)ret < size) && (ferror(file)))
268 * Write to sysfs entry.
271 * Pointer to private structure.
273 * Entry name relative to sysfs path.
280 * 0 on success, -1 on failure and errno is set.
283 priv_sysfs_write(const struct priv *priv, const char *entry,
284 char *buf, size_t size)
286 char ifname[IF_NAMESIZE];
291 if (priv_get_ifname(priv, &ifname))
294 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
297 file = fopen(path, "wb");
300 ret = fwrite(buf, 1, size, file);
302 if (((size_t)ret < size) || (ferror(file)))
312 * Get unsigned long sysfs property.
315 * Pointer to private structure.
317 * Entry name relative to sysfs path.
319 * Value output buffer.
322 * 0 on success, -1 on failure and errno is set.
325 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
328 unsigned long value_ret;
331 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
333 DEBUG("cannot read %s value from sysfs: %s",
334 name, strerror(errno));
337 value_str[ret] = '\0';
339 value_ret = strtoul(value_str, NULL, 0);
341 DEBUG("invalid %s value `%s': %s", name, value_str,
350 * Set unsigned long sysfs property.
353 * Pointer to private structure.
355 * Entry name relative to sysfs path.
360 * 0 on success, -1 on failure and errno is set.
363 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
366 MKSTR(value_str, "%lu", value);
368 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
370 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
371 name, value_str, value, strerror(errno));
378 * Perform ifreq ioctl() on associated Ethernet device.
381 * Pointer to private structure.
383 * Request number to pass to ioctl().
385 * Interface request structure output buffer.
388 * 0 on success, -1 on failure and errno is set.
391 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
393 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
398 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
399 ret = ioctl(sock, req, ifr);
408 * Pointer to private structure.
410 * MTU value output buffer.
413 * 0 on success, -1 on failure and errno is set.
416 priv_get_mtu(struct priv *priv, uint16_t *mtu)
418 unsigned long ulong_mtu;
420 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
430 * Pointer to private structure.
435 * 0 on success, -1 on failure and errno is set.
438 priv_set_mtu(struct priv *priv, uint16_t mtu)
442 if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
443 priv_get_mtu(priv, &new_mtu))
455 * Pointer to private structure.
457 * Bitmask for flags that must remain untouched.
459 * Bitmask for flags to modify.
462 * 0 on success, -1 on failure and errno is set.
465 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
469 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
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 * 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, errno value on failure.
508 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;
530 * DPDK callback for Ethernet device configuration.
533 * Pointer to Ethernet device structure.
536 * 0 on success, negative errno value on failure.
539 mlx4_dev_configure(struct rte_eth_dev *dev)
541 struct priv *priv = dev->data->dev_private;
545 ret = dev_configure(dev);
551 static uint16_t mlx4_tx_burst(void *, struct rte_mbuf **, uint16_t);
552 static uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
554 /* TX queues handling. */
557 * Allocate TX queue elements.
560 * Pointer to TX queue structure.
562 * Number of elements to allocate.
565 * 0 on success, errno value on failure.
568 txq_alloc_elts(struct txq *txq, unsigned int elts_n)
571 struct txq_elt (*elts)[elts_n] =
572 rte_calloc_socket("TXQ", 1, sizeof(*elts), 0, txq->socket);
576 ERROR("%p: can't allocate packets array", (void *)txq);
580 for (i = 0; (i != elts_n); ++i) {
581 struct txq_elt *elt = &(*elts)[i];
585 DEBUG("%p: allocated and configured %u WRs", (void *)txq, elts_n);
586 txq->elts_n = elts_n;
591 /* Request send completion every MLX4_PMD_TX_PER_COMP_REQ packets or
592 * at least 4 times per ring. */
593 txq->elts_comp_cd_init =
594 ((MLX4_PMD_TX_PER_COMP_REQ < (elts_n / 4)) ?
595 MLX4_PMD_TX_PER_COMP_REQ : (elts_n / 4));
596 txq->elts_comp_cd = txq->elts_comp_cd_init;
602 DEBUG("%p: failed, freed everything", (void *)txq);
608 * Free TX queue elements.
611 * Pointer to TX queue structure.
614 txq_free_elts(struct txq *txq)
616 unsigned int elts_n = txq->elts_n;
617 unsigned int elts_head = txq->elts_head;
618 unsigned int elts_tail = txq->elts_tail;
619 struct txq_elt (*elts)[elts_n] = txq->elts;
621 DEBUG("%p: freeing WRs", (void *)txq);
626 txq->elts_comp_cd = 0;
627 txq->elts_comp_cd_init = 0;
631 while (elts_tail != elts_head) {
632 struct txq_elt *elt = &(*elts)[elts_tail];
634 assert(elt->buf != NULL);
635 rte_pktmbuf_free(elt->buf);
638 memset(elt, 0x77, sizeof(*elt));
640 if (++elts_tail == elts_n)
648 * Clean up a TX queue.
650 * Destroy objects, free allocated memory and reset the structure for reuse.
653 * Pointer to TX queue structure.
656 txq_cleanup(struct txq *txq)
660 DEBUG("cleaning up %p", (void *)txq);
663 claim_zero(ibv_destroy_qp(txq->qp));
665 claim_zero(ibv_destroy_cq(txq->cq));
666 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
667 if (txq->mp2mr[i].mp == NULL)
669 assert(txq->mp2mr[i].mr != NULL);
670 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
672 memset(txq, 0, sizeof(*txq));
676 * Manage TX completions.
678 * When sending a burst, mlx4_tx_burst() posts several WRs.
679 * To improve performance, a completion event is only required once every
680 * MLX4_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
681 * for other WRs, but this information would not be used anyway.
684 * Pointer to TX queue structure.
687 * 0 on success, -1 on failure.
690 txq_complete(struct txq *txq)
692 unsigned int elts_comp = txq->elts_comp;
693 unsigned int elts_tail = txq->elts_tail;
694 const unsigned int elts_n = txq->elts_n;
695 struct ibv_wc wcs[elts_comp];
698 if (unlikely(elts_comp == 0))
700 wcs_n = ibv_poll_cq(txq->cq, elts_comp, wcs);
701 if (unlikely(wcs_n == 0))
703 if (unlikely(wcs_n < 0)) {
704 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
709 assert(elts_comp <= txq->elts_comp);
711 * Assume WC status is successful as nothing can be done about it
714 elts_tail += wcs_n * txq->elts_comp_cd_init;
715 if (elts_tail >= elts_n)
717 txq->elts_tail = elts_tail;
718 txq->elts_comp = elts_comp;
722 struct mlx4_check_mempool_data {
728 /* Called by mlx4_check_mempool() when iterating the memory chunks. */
729 static void mlx4_check_mempool_cb(struct rte_mempool *mp,
730 void *opaque, struct rte_mempool_memhdr *memhdr,
733 struct mlx4_check_mempool_data *data = opaque;
738 /* It already failed, skip the next chunks. */
741 /* It is the first chunk. */
742 if (data->start == NULL && data->end == NULL) {
743 data->start = memhdr->addr;
744 data->end = data->start + memhdr->len;
747 if (data->end == memhdr->addr) {
748 data->end += memhdr->len;
751 if (data->start == (char *)memhdr->addr + memhdr->len) {
752 data->start -= memhdr->len;
755 /* Error, mempool is not virtually contigous. */
760 * Check if a mempool can be used: it must be virtually contiguous.
763 * Pointer to memory pool.
765 * Pointer to the start address of the mempool virtual memory area
767 * Pointer to the end address of the mempool virtual memory area
770 * 0 on success (mempool is virtually contiguous), -1 on error.
772 static int mlx4_check_mempool(struct rte_mempool *mp, uintptr_t *start,
775 struct mlx4_check_mempool_data data;
777 memset(&data, 0, sizeof(data));
778 rte_mempool_mem_iter(mp, mlx4_check_mempool_cb, &data);
779 *start = (uintptr_t)data.start;
780 *end = (uintptr_t)data.end;
785 /* For best performance, this function should not be inlined. */
786 static struct ibv_mr *mlx4_mp2mr(struct ibv_pd *, struct rte_mempool *)
790 * Register mempool as a memory region.
793 * Pointer to protection domain.
795 * Pointer to memory pool.
798 * Memory region pointer, NULL in case of error.
800 static struct ibv_mr *
801 mlx4_mp2mr(struct ibv_pd *pd, struct rte_mempool *mp)
803 const struct rte_memseg *ms = rte_eal_get_physmem_layout();
808 if (mlx4_check_mempool(mp, &start, &end) != 0) {
809 ERROR("mempool %p: not virtually contiguous",
814 DEBUG("mempool %p area start=%p end=%p size=%zu",
815 (void *)mp, (void *)start, (void *)end,
816 (size_t)(end - start));
817 /* Round start and end to page boundary if found in memory segments. */
818 for (i = 0; (i < RTE_MAX_MEMSEG) && (ms[i].addr != NULL); ++i) {
819 uintptr_t addr = (uintptr_t)ms[i].addr;
820 size_t len = ms[i].len;
821 unsigned int align = ms[i].hugepage_sz;
823 if ((start > addr) && (start < addr + len))
824 start = RTE_ALIGN_FLOOR(start, align);
825 if ((end > addr) && (end < addr + len))
826 end = RTE_ALIGN_CEIL(end, align);
828 DEBUG("mempool %p using start=%p end=%p size=%zu for MR",
829 (void *)mp, (void *)start, (void *)end,
830 (size_t)(end - start));
831 return ibv_reg_mr(pd,
834 IBV_ACCESS_LOCAL_WRITE);
838 * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
839 * the cloned mbuf is allocated is returned instead.
845 * Memory pool where data is located for given mbuf.
847 static struct rte_mempool *
848 txq_mb2mp(struct rte_mbuf *buf)
850 if (unlikely(RTE_MBUF_INDIRECT(buf)))
851 return rte_mbuf_from_indirect(buf)->pool;
856 * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
857 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
858 * remove an entry first.
861 * Pointer to TX queue structure.
863 * Memory Pool for which a Memory Region lkey must be returned.
866 * mr->lkey on success, (uint32_t)-1 on failure.
869 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
874 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
875 if (unlikely(txq->mp2mr[i].mp == NULL)) {
876 /* Unknown MP, add a new MR for it. */
879 if (txq->mp2mr[i].mp == mp) {
880 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
881 assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
882 return txq->mp2mr[i].lkey;
885 /* Add a new entry, register MR first. */
886 DEBUG("%p: discovered new memory pool \"%s\" (%p)",
887 (void *)txq, mp->name, (void *)mp);
888 mr = mlx4_mp2mr(txq->priv->pd, mp);
889 if (unlikely(mr == NULL)) {
890 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
894 if (unlikely(i == elemof(txq->mp2mr))) {
895 /* Table is full, remove oldest entry. */
896 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
899 claim_zero(ibv_dereg_mr(txq->mp2mr[0].mr));
900 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
901 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
903 /* Store the new entry. */
904 txq->mp2mr[i].mp = mp;
905 txq->mp2mr[i].mr = mr;
906 txq->mp2mr[i].lkey = mr->lkey;
907 DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
908 (void *)txq, mp->name, (void *)mp, txq->mp2mr[i].lkey);
909 return txq->mp2mr[i].lkey;
912 struct txq_mp2mr_mbuf_check_data {
917 * Callback function for rte_mempool_obj_iter() to check whether a given
918 * mempool object looks like a mbuf.
921 * The mempool pointer
923 * Context data (struct txq_mp2mr_mbuf_check_data). Contains the
928 * Object index, unused.
931 txq_mp2mr_mbuf_check(struct rte_mempool *mp, void *arg, void *obj,
932 uint32_t index __rte_unused)
934 struct txq_mp2mr_mbuf_check_data *data = arg;
935 struct rte_mbuf *buf = obj;
937 /* Check whether mbuf structure fits element size and whether mempool
938 * pointer is valid. */
939 if (sizeof(*buf) > mp->elt_size || buf->pool != mp)
944 * Iterator function for rte_mempool_walk() to register existing mempools and
945 * fill the MP to MR cache of a TX queue.
948 * Memory Pool to register.
950 * Pointer to TX queue structure.
953 txq_mp2mr_iter(struct rte_mempool *mp, void *arg)
955 struct txq *txq = arg;
956 struct txq_mp2mr_mbuf_check_data data = {
960 /* Register mempool only if the first element looks like a mbuf. */
961 if (rte_mempool_obj_iter(mp, txq_mp2mr_mbuf_check, &data) == 0 ||
968 * DPDK callback for TX.
971 * Generic pointer to TX queue structure.
973 * Packets to transmit.
975 * Number of packets in array.
978 * Number of packets successfully transmitted (<= pkts_n).
981 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
983 struct txq *txq = (struct txq *)dpdk_txq;
984 struct ibv_send_wr *wr_head = NULL;
985 struct ibv_send_wr **wr_next = &wr_head;
986 struct ibv_send_wr *wr_bad = NULL;
987 unsigned int elts_head = txq->elts_head;
988 const unsigned int elts_n = txq->elts_n;
989 unsigned int elts_comp_cd = txq->elts_comp_cd;
990 unsigned int elts_comp = 0;
995 assert(elts_comp_cd != 0);
997 max = (elts_n - (elts_head - txq->elts_tail));
1001 assert(max <= elts_n);
1002 /* Always leave one free entry in the ring. */
1008 for (i = 0; (i != max); ++i) {
1009 struct rte_mbuf *buf = pkts[i];
1010 unsigned int elts_head_next =
1011 (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
1012 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
1013 struct txq_elt *elt = &(*txq->elts)[elts_head];
1014 struct ibv_send_wr *wr = &elt->wr;
1015 unsigned int segs = NB_SEGS(buf);
1016 unsigned int sent_size = 0;
1017 uint32_t send_flags = 0;
1019 /* Clean up old buffer. */
1020 if (likely(elt->buf != NULL)) {
1021 struct rte_mbuf *tmp = elt->buf;
1025 memset(elt, 0x66, sizeof(*elt));
1027 /* Faster than rte_pktmbuf_free(). */
1029 struct rte_mbuf *next = NEXT(tmp);
1031 rte_pktmbuf_free_seg(tmp);
1033 } while (tmp != NULL);
1035 /* Request TX completion. */
1036 if (unlikely(--elts_comp_cd == 0)) {
1037 elts_comp_cd = txq->elts_comp_cd_init;
1039 send_flags |= IBV_SEND_SIGNALED;
1041 if (likely(segs == 1)) {
1042 struct ibv_sge *sge = &elt->sge;
1047 /* Retrieve buffer information. */
1048 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1049 length = DATA_LEN(buf);
1050 /* Retrieve Memory Region key for this memory pool. */
1051 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
1052 if (unlikely(lkey == (uint32_t)-1)) {
1053 /* MR does not exist. */
1054 DEBUG("%p: unable to get MP <-> MR"
1055 " association", (void *)txq);
1056 /* Clean up TX element. */
1060 /* Update element. */
1063 rte_prefetch0((volatile void *)
1065 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1067 sge->length = length;
1069 sent_size += length;
1074 if (sent_size <= txq->max_inline)
1075 send_flags |= IBV_SEND_INLINE;
1076 elts_head = elts_head_next;
1077 /* Increment sent bytes counter. */
1078 txq->stats.obytes += sent_size;
1080 wr->sg_list = &elt->sge;
1082 wr->opcode = IBV_WR_SEND;
1083 wr->send_flags = send_flags;
1085 wr_next = &wr->next;
1088 /* Take a shortcut if nothing must be sent. */
1089 if (unlikely(i == 0))
1091 /* Increment sent packets counter. */
1092 txq->stats.opackets += i;
1093 /* Ring QP doorbell. */
1096 err = ibv_post_send(txq->qp, wr_head, &wr_bad);
1097 if (unlikely(err)) {
1098 uint64_t obytes = 0;
1099 uint64_t opackets = 0;
1101 /* Rewind bad WRs. */
1102 while (wr_bad != NULL) {
1105 /* Force completion request if one was lost. */
1106 if (wr_bad->send_flags & IBV_SEND_SIGNALED) {
1111 for (j = 0; j < wr_bad->num_sge; ++j)
1112 obytes += wr_bad->sg_list[j].length;
1113 elts_head = (elts_head ? elts_head : elts_n) - 1;
1114 wr_bad = wr_bad->next;
1116 txq->stats.opackets -= opackets;
1117 txq->stats.obytes -= obytes;
1119 DEBUG("%p: ibv_post_send() failed, %" PRIu64 " packets"
1120 " (%" PRIu64 " bytes) rejected: %s",
1124 (err <= -1) ? "Internal error" : strerror(err));
1126 txq->elts_head = elts_head;
1127 txq->elts_comp += elts_comp;
1128 txq->elts_comp_cd = elts_comp_cd;
1133 * Configure a TX queue.
1136 * Pointer to Ethernet device structure.
1138 * Pointer to TX queue structure.
1140 * Number of descriptors to configure in queue.
1142 * NUMA socket on which memory must be allocated.
1144 * Thresholds parameters.
1147 * 0 on success, errno value on failure.
1150 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1151 unsigned int socket, const struct rte_eth_txconf *conf)
1153 struct priv *priv = dev->data->dev_private;
1159 struct ibv_qp_init_attr init;
1160 struct ibv_qp_attr mod;
1164 (void)conf; /* Thresholds configuration (ignored). */
1168 ERROR("%p: invalid number of Tx descriptors", (void *)dev);
1171 /* MRs will be registered in mp2mr[] later. */
1172 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
1173 if (tmpl.cq == NULL) {
1175 ERROR("%p: CQ creation failure: %s",
1176 (void *)dev, strerror(ret));
1179 DEBUG("priv->device_attr.max_qp_wr is %d",
1180 priv->device_attr.max_qp_wr);
1181 DEBUG("priv->device_attr.max_sge is %d",
1182 priv->device_attr.max_sge);
1183 attr.init = (struct ibv_qp_init_attr){
1184 /* CQ to be associated with the send queue. */
1186 /* CQ to be associated with the receive queue. */
1189 /* Max number of outstanding WRs. */
1190 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1191 priv->device_attr.max_qp_wr :
1193 /* Max number of scatter/gather elements in a WR. */
1195 .max_inline_data = MLX4_PMD_MAX_INLINE,
1197 .qp_type = IBV_QPT_RAW_PACKET,
1198 /* 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 ret = (errno ? errno : EINVAL);
1205 ERROR("%p: QP creation failure: %s",
1206 (void *)dev, strerror(ret));
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);
1219 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1220 (void *)dev, strerror(ret));
1223 ret = txq_alloc_elts(&tmpl, desc);
1225 ERROR("%p: TXQ allocation failed: %s",
1226 (void *)dev, strerror(ret));
1229 attr.mod = (struct ibv_qp_attr){
1230 .qp_state = IBV_QPS_RTR
1232 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1234 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1235 (void *)dev, strerror(ret));
1238 attr.mod.qp_state = IBV_QPS_RTS;
1239 ret = ibv_modify_qp(tmpl.qp, &attr.mod, IBV_QP_STATE);
1241 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1242 (void *)dev, strerror(ret));
1245 /* Clean up txq in case we're reinitializing it. */
1246 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1249 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1250 /* Pre-register known mempools. */
1251 rte_mempool_walk(txq_mp2mr_iter, txq);
1261 * DPDK callback to configure a TX queue.
1264 * Pointer to Ethernet device structure.
1268 * Number of descriptors to configure in queue.
1270 * NUMA socket on which memory must be allocated.
1272 * Thresholds parameters.
1275 * 0 on success, negative errno value on failure.
1278 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1279 unsigned int socket, const struct rte_eth_txconf *conf)
1281 struct priv *priv = dev->data->dev_private;
1282 struct txq *txq = (*priv->txqs)[idx];
1286 DEBUG("%p: configuring queue %u for %u descriptors",
1287 (void *)dev, idx, desc);
1288 if (idx >= priv->txqs_n) {
1289 ERROR("%p: queue index out of range (%u >= %u)",
1290 (void *)dev, idx, priv->txqs_n);
1295 DEBUG("%p: reusing already allocated queue index %u (%p)",
1296 (void *)dev, idx, (void *)txq);
1297 if (priv->started) {
1301 (*priv->txqs)[idx] = NULL;
1304 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1306 ERROR("%p: unable to allocate queue index %u",
1312 ret = txq_setup(dev, txq, desc, socket, conf);
1316 txq->stats.idx = idx;
1317 DEBUG("%p: adding TX queue %p to list",
1318 (void *)dev, (void *)txq);
1319 (*priv->txqs)[idx] = txq;
1320 /* Update send callback. */
1321 dev->tx_pkt_burst = mlx4_tx_burst;
1328 * DPDK callback to release a TX queue.
1331 * Generic TX queue pointer.
1334 mlx4_tx_queue_release(void *dpdk_txq)
1336 struct txq *txq = (struct txq *)dpdk_txq;
1344 for (i = 0; (i != priv->txqs_n); ++i)
1345 if ((*priv->txqs)[i] == txq) {
1346 DEBUG("%p: removing TX queue %p from list",
1347 (void *)priv->dev, (void *)txq);
1348 (*priv->txqs)[i] = NULL;
1356 /* RX queues handling. */
1359 * Allocate RX queue elements.
1362 * Pointer to RX queue structure.
1364 * Number of elements to allocate.
1367 * 0 on success, errno value on failure.
1370 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n)
1373 struct rxq_elt (*elts)[elts_n] =
1374 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1379 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);
1391 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;
1425 for (i = 0; (i != elemof(*elts)); ++i)
1426 rte_pktmbuf_free_seg((*elts)[i].buf);
1429 DEBUG("%p: failed, freed everything", (void *)rxq);
1435 * Free RX queue elements.
1438 * Pointer to RX queue structure.
1441 rxq_free_elts(struct rxq *rxq)
1444 unsigned int elts_n = rxq->elts_n;
1445 struct rxq_elt (*elts)[elts_n] = rxq->elts;
1447 DEBUG("%p: freeing WRs", (void *)rxq);
1452 for (i = 0; (i != elemof(*elts)); ++i)
1453 rte_pktmbuf_free_seg((*elts)[i].buf);
1458 * Unregister a MAC address.
1461 * Pointer to private structure.
1464 priv_mac_addr_del(struct priv *priv)
1467 uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
1470 if (!priv->mac_flow)
1472 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x",
1474 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
1475 claim_zero(ibv_destroy_flow(priv->mac_flow));
1476 priv->mac_flow = NULL;
1480 * Register a MAC address.
1482 * The MAC address is registered in queue 0.
1485 * Pointer to private structure.
1488 * 0 on success, errno value on failure.
1491 priv_mac_addr_add(struct priv *priv)
1493 uint8_t (*mac)[ETHER_ADDR_LEN] = &priv->mac.addr_bytes;
1495 struct ibv_flow *flow;
1497 /* If device isn't started, this is all we need to do. */
1502 if (*priv->rxqs && (*priv->rxqs)[0])
1503 rxq = (*priv->rxqs)[0];
1507 /* Allocate flow specification on the stack. */
1508 struct __attribute__((packed)) {
1509 struct ibv_flow_attr attr;
1510 struct ibv_flow_spec_eth spec;
1512 struct ibv_flow_attr *attr = &data.attr;
1513 struct ibv_flow_spec_eth *spec = &data.spec;
1516 priv_mac_addr_del(priv);
1518 * No padding must be inserted by the compiler between attr and spec.
1519 * This layout is expected by libibverbs.
1521 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
1522 *attr = (struct ibv_flow_attr){
1523 .type = IBV_FLOW_ATTR_NORMAL,
1529 *spec = (struct ibv_flow_spec_eth){
1530 .type = IBV_FLOW_SPEC_ETH,
1531 .size = sizeof(*spec),
1534 (*mac)[0], (*mac)[1], (*mac)[2],
1535 (*mac)[3], (*mac)[4], (*mac)[5]
1539 .dst_mac = "\xff\xff\xff\xff\xff\xff",
1542 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x",
1544 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5]);
1545 /* Create related flow. */
1547 flow = ibv_create_flow(rxq->qp, attr);
1549 /* It's not clear whether errno is always set in this case. */
1550 ERROR("%p: flow configuration failed, errno=%d: %s",
1552 (errno ? strerror(errno) : "Unknown error"));
1557 assert(priv->mac_flow == NULL);
1558 priv->mac_flow = flow;
1563 * Clean up a RX queue.
1565 * Destroy objects, free allocated memory and reset the structure for reuse.
1568 * Pointer to RX queue structure.
1571 rxq_cleanup(struct rxq *rxq)
1573 DEBUG("cleaning up %p", (void *)rxq);
1575 if (rxq->qp != NULL)
1576 claim_zero(ibv_destroy_qp(rxq->qp));
1577 if (rxq->cq != NULL)
1578 claim_zero(ibv_destroy_cq(rxq->cq));
1579 if (rxq->channel != NULL)
1580 claim_zero(ibv_destroy_comp_channel(rxq->channel));
1581 if (rxq->mr != NULL)
1582 claim_zero(ibv_dereg_mr(rxq->mr));
1583 memset(rxq, 0, sizeof(*rxq));
1587 * DPDK callback for RX.
1589 * The following function doesn't manage scattered packets.
1592 * Generic pointer to RX queue structure.
1594 * Array to store received packets.
1596 * Maximum number of packets in array.
1599 * Number of packets successfully received (<= pkts_n).
1602 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1604 struct rxq *rxq = (struct rxq *)dpdk_rxq;
1605 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts;
1606 const unsigned int elts_n = rxq->elts_n;
1607 unsigned int elts_head = rxq->elts_head;
1608 struct ibv_wc wcs[pkts_n];
1609 struct ibv_recv_wr *wr_head = NULL;
1610 struct ibv_recv_wr **wr_next = &wr_head;
1611 struct ibv_recv_wr *wr_bad = NULL;
1613 unsigned int pkts_ret = 0;
1616 ret = ibv_poll_cq(rxq->cq, pkts_n, wcs);
1617 if (unlikely(ret == 0))
1619 if (unlikely(ret < 0)) {
1620 DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
1624 assert(ret <= (int)pkts_n);
1625 /* For each work completion. */
1626 for (i = 0; i != (unsigned int)ret; ++i) {
1627 struct ibv_wc *wc = &wcs[i];
1628 struct rxq_elt *elt = &(*elts)[elts_head];
1629 struct ibv_recv_wr *wr = &elt->wr;
1630 uint32_t len = wc->byte_len;
1631 struct rte_mbuf *seg = elt->buf;
1632 struct rte_mbuf *rep;
1634 /* Sanity checks. */
1635 assert(wr->sg_list == &elt->sge);
1636 assert(wr->num_sge == 1);
1637 assert(elts_head < rxq->elts_n);
1638 assert(rxq->elts_head < rxq->elts_n);
1640 * Fetch initial bytes of packet descriptor into a
1641 * cacheline while allocating rep.
1643 rte_mbuf_prefetch_part1(seg);
1644 rte_mbuf_prefetch_part2(seg);
1645 /* Link completed WRs together for repost. */
1647 wr_next = &wr->next;
1648 if (unlikely(wc->status != IBV_WC_SUCCESS)) {
1649 /* Whatever, just repost the offending WR. */
1650 DEBUG("rxq=%p: bad work completion status (%d): %s",
1651 (void *)rxq, wc->status,
1652 ibv_wc_status_str(wc->status));
1653 /* Increment dropped packets counter. */
1654 ++rxq->stats.idropped;
1657 rep = rte_mbuf_raw_alloc(rxq->mp);
1658 if (unlikely(rep == NULL)) {
1660 * Unable to allocate a replacement mbuf,
1663 DEBUG("rxq=%p: can't allocate a new mbuf",
1665 /* Increase out of memory counters. */
1666 ++rxq->stats.rx_nombuf;
1667 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
1671 /* Reconfigure sge to use rep instead of seg. */
1672 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
1673 assert(elt->sge.lkey == rxq->mr->lkey);
1676 /* Update seg information. */
1677 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
1679 PORT(seg) = rxq->port_id;
1682 DATA_LEN(seg) = len;
1683 seg->packet_type = 0;
1686 /* Return packet. */
1689 /* Increase bytes counter. */
1690 rxq->stats.ibytes += len;
1692 if (++elts_head >= elts_n)
1696 if (unlikely(i == 0))
1701 ret = ibv_post_recv(rxq->qp, wr_head, &wr_bad);
1702 if (unlikely(ret)) {
1703 /* Inability to repost WRs is fatal. */
1704 DEBUG("%p: recv_burst(): failed (ret=%d)",
1709 rxq->elts_head = elts_head;
1710 /* Increase packets counter. */
1711 rxq->stats.ipackets += pkts_ret;
1716 * Allocate a Queue Pair.
1717 * Optionally setup inline receive if supported.
1720 * Pointer to private structure.
1722 * Completion queue to associate with QP.
1724 * Number of descriptors in QP (hint only).
1727 * QP pointer or NULL in case of error.
1729 static struct ibv_qp *
1730 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
1732 struct ibv_qp_init_attr attr = {
1733 /* CQ to be associated with the send queue. */
1735 /* CQ to be associated with the receive queue. */
1738 /* Max number of outstanding WRs. */
1739 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
1740 priv->device_attr.max_qp_wr :
1742 /* Max number of scatter/gather elements in a WR. */
1745 .qp_type = IBV_QPT_RAW_PACKET,
1748 return ibv_create_qp(priv->pd, &attr);
1752 * Configure a RX queue.
1755 * Pointer to Ethernet device structure.
1757 * Pointer to RX queue structure.
1759 * Number of descriptors to configure in queue.
1761 * NUMA socket on which memory must be allocated.
1763 * Thresholds parameters.
1765 * Memory pool for buffer allocations.
1768 * 0 on success, errno value on failure.
1771 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
1772 unsigned int socket, const struct rte_eth_rxconf *conf,
1773 struct rte_mempool *mp)
1775 struct priv *priv = dev->data->dev_private;
1781 struct ibv_qp_attr mod;
1782 struct ibv_recv_wr *bad_wr;
1783 unsigned int mb_len;
1786 (void)conf; /* Thresholds configuration (ignored). */
1787 mb_len = rte_pktmbuf_data_room_size(mp);
1789 ERROR("%p: invalid number of Rx descriptors", (void *)dev);
1792 /* Enable scattered packets support for this queue if necessary. */
1793 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
1794 if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
1795 (mb_len - RTE_PKTMBUF_HEADROOM)) {
1797 } else if (dev->data->dev_conf.rxmode.enable_scatter) {
1798 WARN("%p: scattered mode has been requested but is"
1799 " not supported, this may lead to packet loss",
1802 WARN("%p: the requested maximum Rx packet size (%u) is"
1803 " larger than a single mbuf (%u) and scattered"
1804 " mode has not been requested",
1806 dev->data->dev_conf.rxmode.max_rx_pkt_len,
1807 mb_len - RTE_PKTMBUF_HEADROOM);
1809 /* Use the entire RX mempool as the memory region. */
1810 tmpl.mr = mlx4_mp2mr(priv->pd, mp);
1811 if (tmpl.mr == NULL) {
1813 ERROR("%p: MR creation failure: %s",
1814 (void *)dev, strerror(ret));
1817 if (dev->data->dev_conf.intr_conf.rxq) {
1818 tmpl.channel = ibv_create_comp_channel(priv->ctx);
1819 if (tmpl.channel == NULL) {
1821 ERROR("%p: Rx interrupt completion channel creation"
1823 (void *)dev, strerror(ret));
1827 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, tmpl.channel, 0);
1828 if (tmpl.cq == NULL) {
1830 ERROR("%p: CQ creation failure: %s",
1831 (void *)dev, strerror(ret));
1834 DEBUG("priv->device_attr.max_qp_wr is %d",
1835 priv->device_attr.max_qp_wr);
1836 DEBUG("priv->device_attr.max_sge is %d",
1837 priv->device_attr.max_sge);
1838 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc);
1839 if (tmpl.qp == NULL) {
1840 ret = (errno ? errno : EINVAL);
1841 ERROR("%p: QP creation failure: %s",
1842 (void *)dev, strerror(ret));
1845 mod = (struct ibv_qp_attr){
1846 /* Move the QP to this state. */
1847 .qp_state = IBV_QPS_INIT,
1848 /* Primary port number. */
1849 .port_num = priv->port
1851 ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE | IBV_QP_PORT);
1853 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1854 (void *)dev, strerror(ret));
1857 ret = rxq_alloc_elts(&tmpl, desc);
1859 ERROR("%p: RXQ allocation failed: %s",
1860 (void *)dev, strerror(ret));
1863 ret = ibv_post_recv(tmpl.qp, &(*tmpl.elts)[0].wr, &bad_wr);
1865 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
1871 mod = (struct ibv_qp_attr){
1872 .qp_state = IBV_QPS_RTR
1874 ret = ibv_modify_qp(tmpl.qp, &mod, IBV_QP_STATE);
1876 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1877 (void *)dev, strerror(ret));
1881 tmpl.port_id = dev->data->port_id;
1882 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
1883 /* Clean up rxq in case we're reinitializing it. */
1884 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
1887 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
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 on failure.
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];
1925 DEBUG("%p: configuring queue %u for %u descriptors",
1926 (void *)dev, idx, desc);
1927 if (idx >= priv->rxqs_n) {
1928 ERROR("%p: queue index out of range (%u >= %u)",
1929 (void *)dev, idx, priv->rxqs_n);
1934 DEBUG("%p: reusing already allocated queue index %u (%p)",
1935 (void *)dev, idx, (void *)rxq);
1936 if (priv->started) {
1940 (*priv->rxqs)[idx] = NULL;
1942 priv_mac_addr_del(priv);
1945 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
1947 ERROR("%p: unable to allocate queue index %u",
1953 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
1957 rxq->stats.idx = idx;
1958 DEBUG("%p: adding RX queue %p to list",
1959 (void *)dev, (void *)rxq);
1960 (*priv->rxqs)[idx] = rxq;
1961 /* Update receive callback. */
1962 dev->rx_pkt_burst = mlx4_rx_burst;
1969 * DPDK callback to release a RX queue.
1972 * Generic RX queue pointer.
1975 mlx4_rx_queue_release(void *dpdk_rxq)
1977 struct rxq *rxq = (struct rxq *)dpdk_rxq;
1985 for (i = 0; (i != priv->rxqs_n); ++i)
1986 if ((*priv->rxqs)[i] == rxq) {
1987 DEBUG("%p: removing RX queue %p from list",
1988 (void *)priv->dev, (void *)rxq);
1989 (*priv->rxqs)[i] = NULL;
1991 priv_mac_addr_del(priv);
2000 priv_dev_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
2003 priv_dev_removal_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
2006 priv_dev_link_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
2009 * DPDK callback to start the device.
2011 * Simulate device start by attaching all configured flows.
2014 * Pointer to Ethernet device structure.
2017 * 0 on success, negative errno value on failure.
2020 mlx4_dev_start(struct rte_eth_dev *dev)
2022 struct priv *priv = dev->data->dev_private;
2026 if (priv->started) {
2030 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
2032 ret = priv_mac_addr_add(priv);
2035 ret = priv_dev_link_interrupt_handler_install(priv, dev);
2037 ERROR("%p: LSC handler install failed",
2041 ret = priv_dev_removal_interrupt_handler_install(priv, dev);
2043 ERROR("%p: RMV handler install failed",
2047 ret = priv_rx_intr_vec_enable(priv);
2049 ERROR("%p: Rx interrupt vector creation failed",
2053 ret = mlx4_priv_flow_start(priv);
2055 ERROR("%p: flow start failed: %s",
2056 (void *)dev, strerror(ret));
2063 priv_mac_addr_del(priv);
2070 * DPDK callback to stop the device.
2072 * Simulate device stop by detaching all configured flows.
2075 * Pointer to Ethernet device structure.
2078 mlx4_dev_stop(struct rte_eth_dev *dev)
2080 struct priv *priv = dev->data->dev_private;
2083 if (!priv->started) {
2087 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
2089 mlx4_priv_flow_stop(priv);
2090 priv_mac_addr_del(priv);
2095 * Dummy DPDK callback for TX.
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 TX queue structure.
2103 * Packets to transmit.
2105 * Number of packets in array.
2108 * Number of packets successfully transmitted (<= pkts_n).
2111 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
2120 * Dummy DPDK callback for RX.
2122 * This function is used to temporarily replace the real callback during
2123 * unsafe control operations on the queue, or in case of error.
2126 * Generic pointer to RX queue structure.
2128 * Array to store received packets.
2130 * Maximum number of packets in array.
2133 * Number of packets successfully received (<= pkts_n).
2136 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2145 priv_dev_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
2148 priv_dev_removal_interrupt_handler_uninstall(struct priv *,
2149 struct rte_eth_dev *);
2152 priv_dev_link_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
2155 * DPDK callback to close the device.
2157 * Destroy all queues and objects, free memory.
2160 * Pointer to Ethernet device structure.
2163 mlx4_dev_close(struct rte_eth_dev *dev)
2165 struct priv *priv = dev->data->dev_private;
2172 DEBUG("%p: closing device \"%s\"",
2174 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
2175 priv_mac_addr_del(priv);
2176 /* Prevent crashes when queues are still in use. This is unfortunately
2177 * still required for DPDK 1.3 because some programs (such as testpmd)
2178 * never release them before closing the device. */
2179 dev->rx_pkt_burst = removed_rx_burst;
2180 dev->tx_pkt_burst = removed_tx_burst;
2181 if (priv->rxqs != NULL) {
2182 /* XXX race condition if mlx4_rx_burst() is still running. */
2184 for (i = 0; (i != priv->rxqs_n); ++i) {
2185 tmp = (*priv->rxqs)[i];
2188 (*priv->rxqs)[i] = NULL;
2195 if (priv->txqs != NULL) {
2196 /* XXX race condition if mlx4_tx_burst() is still running. */
2198 for (i = 0; (i != priv->txqs_n); ++i) {
2199 tmp = (*priv->txqs)[i];
2202 (*priv->txqs)[i] = NULL;
2209 if (priv->pd != NULL) {
2210 assert(priv->ctx != NULL);
2211 claim_zero(ibv_dealloc_pd(priv->pd));
2212 claim_zero(ibv_close_device(priv->ctx));
2214 assert(priv->ctx == NULL);
2215 priv_dev_removal_interrupt_handler_uninstall(priv, dev);
2216 priv_dev_link_interrupt_handler_uninstall(priv, dev);
2217 priv_rx_intr_vec_disable(priv);
2219 memset(priv, 0, sizeof(*priv));
2223 * Change the link state (UP / DOWN).
2226 * Pointer to Ethernet device private data.
2228 * Nonzero for link up, otherwise link down.
2231 * 0 on success, errno value on failure.
2234 priv_set_link(struct priv *priv, int up)
2236 struct rte_eth_dev *dev = priv->dev;
2240 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
2243 dev->rx_pkt_burst = mlx4_rx_burst;
2245 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
2248 dev->rx_pkt_burst = removed_rx_burst;
2249 dev->tx_pkt_burst = removed_tx_burst;
2255 * DPDK callback to bring the link DOWN.
2258 * Pointer to Ethernet device structure.
2261 * 0 on success, errno value on failure.
2264 mlx4_set_link_down(struct rte_eth_dev *dev)
2266 struct priv *priv = dev->data->dev_private;
2270 err = priv_set_link(priv, 0);
2276 * DPDK callback to bring the link UP.
2279 * Pointer to Ethernet device structure.
2282 * 0 on success, errno value on failure.
2285 mlx4_set_link_up(struct rte_eth_dev *dev)
2287 struct priv *priv = dev->data->dev_private;
2291 err = priv_set_link(priv, 1);
2296 * DPDK callback to get information about the device.
2299 * Pointer to Ethernet device structure.
2301 * Info structure output buffer.
2304 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
2306 struct priv *priv = dev->data->dev_private;
2308 char ifname[IF_NAMESIZE];
2310 info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
2315 /* FIXME: we should ask the device for these values. */
2316 info->min_rx_bufsize = 32;
2317 info->max_rx_pktlen = 65536;
2319 * Since we need one CQ per QP, the limit is the minimum number
2320 * between the two values.
2322 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
2323 priv->device_attr.max_qp : priv->device_attr.max_cq);
2324 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
2327 info->max_rx_queues = max;
2328 info->max_tx_queues = max;
2329 /* Last array entry is reserved for broadcast. */
2330 info->max_mac_addrs = 1;
2331 info->rx_offload_capa = 0;
2332 info->tx_offload_capa = 0;
2333 if (priv_get_ifname(priv, &ifname) == 0)
2334 info->if_index = if_nametoindex(ifname);
2337 ETH_LINK_SPEED_10G |
2338 ETH_LINK_SPEED_20G |
2339 ETH_LINK_SPEED_40G |
2345 * DPDK callback to get device statistics.
2348 * Pointer to Ethernet device structure.
2350 * Stats structure output buffer.
2353 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
2355 struct priv *priv = dev->data->dev_private;
2356 struct rte_eth_stats tmp = {0};
2363 /* Add software counters. */
2364 for (i = 0; (i != priv->rxqs_n); ++i) {
2365 struct rxq *rxq = (*priv->rxqs)[i];
2369 idx = rxq->stats.idx;
2370 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
2371 tmp.q_ipackets[idx] += rxq->stats.ipackets;
2372 tmp.q_ibytes[idx] += rxq->stats.ibytes;
2373 tmp.q_errors[idx] += (rxq->stats.idropped +
2374 rxq->stats.rx_nombuf);
2376 tmp.ipackets += rxq->stats.ipackets;
2377 tmp.ibytes += rxq->stats.ibytes;
2378 tmp.ierrors += rxq->stats.idropped;
2379 tmp.rx_nombuf += rxq->stats.rx_nombuf;
2381 for (i = 0; (i != priv->txqs_n); ++i) {
2382 struct txq *txq = (*priv->txqs)[i];
2386 idx = txq->stats.idx;
2387 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
2388 tmp.q_opackets[idx] += txq->stats.opackets;
2389 tmp.q_obytes[idx] += txq->stats.obytes;
2390 tmp.q_errors[idx] += txq->stats.odropped;
2392 tmp.opackets += txq->stats.opackets;
2393 tmp.obytes += txq->stats.obytes;
2394 tmp.oerrors += txq->stats.odropped;
2401 * DPDK callback to clear device statistics.
2404 * Pointer to Ethernet device structure.
2407 mlx4_stats_reset(struct rte_eth_dev *dev)
2409 struct priv *priv = dev->data->dev_private;
2416 for (i = 0; (i != priv->rxqs_n); ++i) {
2417 if ((*priv->rxqs)[i] == NULL)
2419 idx = (*priv->rxqs)[i]->stats.idx;
2420 (*priv->rxqs)[i]->stats =
2421 (struct mlx4_rxq_stats){ .idx = idx };
2423 for (i = 0; (i != priv->txqs_n); ++i) {
2424 if ((*priv->txqs)[i] == NULL)
2426 idx = (*priv->txqs)[i]->stats.idx;
2427 (*priv->txqs)[i]->stats =
2428 (struct mlx4_txq_stats){ .idx = idx };
2434 * DPDK callback to retrieve physical link information.
2437 * Pointer to Ethernet device structure.
2438 * @param wait_to_complete
2439 * Wait for request completion (ignored).
2442 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
2444 const struct priv *priv = dev->data->dev_private;
2445 struct ethtool_cmd edata = {
2449 struct rte_eth_link dev_link;
2452 /* priv_lock() is not taken to allow concurrent calls. */
2456 (void)wait_to_complete;
2457 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
2458 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
2461 memset(&dev_link, 0, sizeof(dev_link));
2462 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
2463 (ifr.ifr_flags & IFF_RUNNING));
2464 ifr.ifr_data = (void *)&edata;
2465 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2466 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
2470 link_speed = ethtool_cmd_speed(&edata);
2471 if (link_speed == -1)
2472 dev_link.link_speed = 0;
2474 dev_link.link_speed = link_speed;
2475 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
2476 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
2477 dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
2478 ETH_LINK_SPEED_FIXED);
2479 dev->data->dev_link = dev_link;
2484 * DPDK callback to change the MTU.
2487 * Pointer to Ethernet device structure.
2492 * 0 on success, negative errno value on failure.
2495 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
2497 struct priv *priv = dev->data->dev_private;
2501 /* Set kernel interface MTU first. */
2502 if (priv_set_mtu(priv, mtu)) {
2504 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
2508 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
2517 * DPDK callback to get flow control status.
2520 * Pointer to Ethernet device structure.
2521 * @param[out] fc_conf
2522 * Flow control output buffer.
2525 * 0 on success, negative errno value on failure.
2528 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
2530 struct priv *priv = dev->data->dev_private;
2532 struct ethtool_pauseparam ethpause = {
2533 .cmd = ETHTOOL_GPAUSEPARAM
2537 ifr.ifr_data = (void *)ðpause;
2539 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2541 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
2547 fc_conf->autoneg = ethpause.autoneg;
2548 if (ethpause.rx_pause && ethpause.tx_pause)
2549 fc_conf->mode = RTE_FC_FULL;
2550 else if (ethpause.rx_pause)
2551 fc_conf->mode = RTE_FC_RX_PAUSE;
2552 else if (ethpause.tx_pause)
2553 fc_conf->mode = RTE_FC_TX_PAUSE;
2555 fc_conf->mode = RTE_FC_NONE;
2565 * DPDK callback to modify flow control parameters.
2568 * Pointer to Ethernet device structure.
2569 * @param[in] fc_conf
2570 * Flow control parameters.
2573 * 0 on success, negative errno value on failure.
2576 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
2578 struct priv *priv = dev->data->dev_private;
2580 struct ethtool_pauseparam ethpause = {
2581 .cmd = ETHTOOL_SPAUSEPARAM
2585 ifr.ifr_data = (void *)ðpause;
2586 ethpause.autoneg = fc_conf->autoneg;
2587 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
2588 (fc_conf->mode & RTE_FC_RX_PAUSE))
2589 ethpause.rx_pause = 1;
2591 ethpause.rx_pause = 0;
2593 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
2594 (fc_conf->mode & RTE_FC_TX_PAUSE))
2595 ethpause.tx_pause = 1;
2597 ethpause.tx_pause = 0;
2600 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
2602 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
2615 const struct rte_flow_ops mlx4_flow_ops = {
2616 .validate = mlx4_flow_validate,
2617 .create = mlx4_flow_create,
2618 .destroy = mlx4_flow_destroy,
2619 .flush = mlx4_flow_flush,
2621 .isolate = mlx4_flow_isolate,
2625 * Manage filter operations.
2628 * Pointer to Ethernet device structure.
2629 * @param filter_type
2632 * Operation to perform.
2634 * Pointer to operation-specific structure.
2637 * 0 on success, negative errno value on failure.
2640 mlx4_dev_filter_ctrl(struct rte_eth_dev *dev,
2641 enum rte_filter_type filter_type,
2642 enum rte_filter_op filter_op,
2647 switch (filter_type) {
2648 case RTE_ETH_FILTER_GENERIC:
2649 if (filter_op != RTE_ETH_FILTER_GET)
2651 *(const void **)arg = &mlx4_flow_ops;
2654 ERROR("%p: filter type (%d) not supported",
2655 (void *)dev, filter_type);
2661 static const struct eth_dev_ops mlx4_dev_ops = {
2662 .dev_configure = mlx4_dev_configure,
2663 .dev_start = mlx4_dev_start,
2664 .dev_stop = mlx4_dev_stop,
2665 .dev_set_link_down = mlx4_set_link_down,
2666 .dev_set_link_up = mlx4_set_link_up,
2667 .dev_close = mlx4_dev_close,
2668 .link_update = mlx4_link_update,
2669 .stats_get = mlx4_stats_get,
2670 .stats_reset = mlx4_stats_reset,
2671 .dev_infos_get = mlx4_dev_infos_get,
2672 .rx_queue_setup = mlx4_rx_queue_setup,
2673 .tx_queue_setup = mlx4_tx_queue_setup,
2674 .rx_queue_release = mlx4_rx_queue_release,
2675 .tx_queue_release = mlx4_tx_queue_release,
2676 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
2677 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
2678 .mtu_set = mlx4_dev_set_mtu,
2679 .filter_ctrl = mlx4_dev_filter_ctrl,
2680 .rx_queue_intr_enable = mlx4_rx_intr_enable,
2681 .rx_queue_intr_disable = mlx4_rx_intr_disable,
2685 * Get PCI information from struct ibv_device.
2688 * Pointer to Ethernet device structure.
2689 * @param[out] pci_addr
2690 * PCI bus address output buffer.
2693 * 0 on success, -1 on failure and errno is set.
2696 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
2697 struct rte_pci_addr *pci_addr)
2701 MKSTR(path, "%s/device/uevent", device->ibdev_path);
2703 file = fopen(path, "rb");
2706 while (fgets(line, sizeof(line), file) == line) {
2707 size_t len = strlen(line);
2710 /* Truncate long lines. */
2711 if (len == (sizeof(line) - 1))
2712 while (line[(len - 1)] != '\n') {
2716 line[(len - 1)] = ret;
2718 /* Extract information. */
2721 "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
2725 &pci_addr->function) == 4) {
2735 * Get MAC address by querying netdevice.
2738 * struct priv for the requested device.
2740 * MAC address output buffer.
2743 * 0 on success, -1 on failure and errno is set.
2746 priv_get_mac(struct priv *priv, uint8_t (*mac)[ETHER_ADDR_LEN])
2748 struct ifreq request;
2750 if (priv_ifreq(priv, SIOCGIFHWADDR, &request))
2752 memcpy(mac, request.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
2757 mlx4_dev_link_status_handler(void *);
2759 mlx4_dev_interrupt_handler(void *);
2762 * Link/device status handler.
2765 * Pointer to private structure.
2767 * Pointer to the rte_eth_dev structure.
2769 * Pointer to event flags holder.
2775 priv_dev_status_handler(struct priv *priv, struct rte_eth_dev *dev,
2778 struct ibv_async_event event;
2779 int port_change = 0;
2780 struct rte_eth_link *link = &dev->data->dev_link;
2784 /* Read all message and acknowledge them. */
2786 if (ibv_get_async_event(priv->ctx, &event))
2788 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
2789 event.event_type == IBV_EVENT_PORT_ERR) &&
2790 (priv->intr_conf.lsc == 1)) {
2793 } else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
2794 priv->intr_conf.rmv == 1) {
2795 *events |= (1 << RTE_ETH_EVENT_INTR_RMV);
2798 DEBUG("event type %d on port %d not handled",
2799 event.event_type, event.element.port_num);
2800 ibv_ack_async_event(&event);
2804 mlx4_link_update(dev, 0);
2805 if (((link->link_speed == 0) && link->link_status) ||
2806 ((link->link_speed != 0) && !link->link_status)) {
2807 if (!priv->pending_alarm) {
2808 /* Inconsistent status, check again later. */
2809 priv->pending_alarm = 1;
2810 rte_eal_alarm_set(MLX4_ALARM_TIMEOUT_US,
2811 mlx4_dev_link_status_handler,
2815 *events |= (1 << RTE_ETH_EVENT_INTR_LSC);
2821 * Handle delayed link status event.
2824 * Registered argument.
2827 mlx4_dev_link_status_handler(void *arg)
2829 struct rte_eth_dev *dev = arg;
2830 struct priv *priv = dev->data->dev_private;
2835 assert(priv->pending_alarm == 1);
2836 priv->pending_alarm = 0;
2837 ret = priv_dev_status_handler(priv, dev, &events);
2839 if (ret > 0 && events & (1 << RTE_ETH_EVENT_INTR_LSC))
2840 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL,
2845 * Handle interrupts from the NIC.
2847 * @param[in] intr_handle
2848 * Interrupt handler.
2850 * Callback argument.
2853 mlx4_dev_interrupt_handler(void *cb_arg)
2855 struct rte_eth_dev *dev = cb_arg;
2856 struct priv *priv = dev->data->dev_private;
2862 ret = priv_dev_status_handler(priv, dev, &ev);
2865 for (i = RTE_ETH_EVENT_UNKNOWN;
2866 i < RTE_ETH_EVENT_MAX;
2868 if (ev & (1 << i)) {
2870 _rte_eth_dev_callback_process(dev, i, NULL,
2876 WARN("%d event%s not processed", ret,
2877 (ret > 1 ? "s were" : " was"));
2882 * Uninstall interrupt handler.
2885 * Pointer to private structure.
2887 * Pointer to the rte_eth_dev structure.
2889 * 0 on success, negative errno value on failure.
2892 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
2896 if (priv->intr_conf.lsc ||
2897 priv->intr_conf.rmv)
2899 ret = rte_intr_callback_unregister(&priv->intr_handle,
2900 mlx4_dev_interrupt_handler,
2903 ERROR("rte_intr_callback_unregister failed with %d"
2905 (errno ? " (errno: " : ""),
2906 (errno ? strerror(errno) : ""),
2907 (errno ? ")" : ""));
2909 priv->intr_handle.fd = 0;
2910 priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
2915 * Install interrupt handler.
2918 * Pointer to private structure.
2920 * Pointer to the rte_eth_dev structure.
2922 * 0 on success, negative errno value on failure.
2925 priv_dev_interrupt_handler_install(struct priv *priv,
2926 struct rte_eth_dev *dev)
2931 /* Check whether the interrupt handler has already been installed
2932 * for either type of interrupt
2934 if (priv->intr_conf.lsc &&
2935 priv->intr_conf.rmv &&
2936 priv->intr_handle.fd)
2938 assert(priv->ctx->async_fd > 0);
2939 flags = fcntl(priv->ctx->async_fd, F_GETFL);
2940 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
2942 INFO("failed to change file descriptor async event queue");
2943 dev->data->dev_conf.intr_conf.lsc = 0;
2944 dev->data->dev_conf.intr_conf.rmv = 0;
2947 priv->intr_handle.fd = priv->ctx->async_fd;
2948 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
2949 rc = rte_intr_callback_register(&priv->intr_handle,
2950 mlx4_dev_interrupt_handler,
2953 ERROR("rte_intr_callback_register failed "
2954 " (errno: %s)", strerror(errno));
2962 * Uninstall interrupt handler.
2965 * Pointer to private structure.
2967 * Pointer to the rte_eth_dev structure.
2969 * 0 on success, negative value on error.
2972 priv_dev_removal_interrupt_handler_uninstall(struct priv *priv,
2973 struct rte_eth_dev *dev)
2975 if (dev->data->dev_conf.intr_conf.rmv) {
2976 priv->intr_conf.rmv = 0;
2977 return priv_dev_interrupt_handler_uninstall(priv, dev);
2983 * Uninstall interrupt handler.
2986 * Pointer to private structure.
2988 * Pointer to the rte_eth_dev structure.
2990 * 0 on success, negative value on error,
2993 priv_dev_link_interrupt_handler_uninstall(struct priv *priv,
2994 struct rte_eth_dev *dev)
2998 if (dev->data->dev_conf.intr_conf.lsc) {
2999 priv->intr_conf.lsc = 0;
3000 ret = priv_dev_interrupt_handler_uninstall(priv, dev);
3004 if (priv->pending_alarm)
3005 if (rte_eal_alarm_cancel(mlx4_dev_link_status_handler,
3007 ERROR("rte_eal_alarm_cancel failed "
3008 " (errno: %s)", strerror(rte_errno));
3011 priv->pending_alarm = 0;
3016 * Install link interrupt handler.
3019 * Pointer to private structure.
3021 * Pointer to the rte_eth_dev structure.
3023 * 0 on success, negative value on error.
3026 priv_dev_link_interrupt_handler_install(struct priv *priv,
3027 struct rte_eth_dev *dev)
3031 if (dev->data->dev_conf.intr_conf.lsc) {
3032 ret = priv_dev_interrupt_handler_install(priv, dev);
3035 priv->intr_conf.lsc = 1;
3041 * Install removal interrupt handler.
3044 * Pointer to private structure.
3046 * Pointer to the rte_eth_dev structure.
3048 * 0 on success, negative value on error.
3051 priv_dev_removal_interrupt_handler_install(struct priv *priv,
3052 struct rte_eth_dev *dev)
3056 if (dev->data->dev_conf.intr_conf.rmv) {
3057 ret = priv_dev_interrupt_handler_install(priv, dev);
3060 priv->intr_conf.rmv = 1;
3066 * Allocate queue vector and fill epoll fd list for Rx interrupts.
3069 * Pointer to private structure.
3072 * 0 on success, negative on failure.
3075 priv_rx_intr_vec_enable(struct priv *priv)
3078 unsigned int rxqs_n = priv->rxqs_n;
3079 unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
3080 unsigned int count = 0;
3081 struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
3083 if (!priv->dev->data->dev_conf.intr_conf.rxq)
3085 priv_rx_intr_vec_disable(priv);
3086 intr_handle->intr_vec = malloc(sizeof(intr_handle->intr_vec[rxqs_n]));
3087 if (intr_handle->intr_vec == NULL) {
3088 ERROR("failed to allocate memory for interrupt vector,"
3089 " Rx interrupts will not be supported");
3092 intr_handle->type = RTE_INTR_HANDLE_EXT;
3093 for (i = 0; i != n; ++i) {
3094 struct rxq *rxq = (*priv->rxqs)[i];
3099 /* Skip queues that cannot request interrupts. */
3100 if (!rxq || !rxq->channel) {
3101 /* Use invalid intr_vec[] index to disable entry. */
3102 intr_handle->intr_vec[i] =
3103 RTE_INTR_VEC_RXTX_OFFSET +
3104 RTE_MAX_RXTX_INTR_VEC_ID;
3107 if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
3108 ERROR("too many Rx queues for interrupt vector size"
3109 " (%d), Rx interrupts cannot be enabled",
3110 RTE_MAX_RXTX_INTR_VEC_ID);
3111 priv_rx_intr_vec_disable(priv);
3114 fd = rxq->channel->fd;
3115 flags = fcntl(fd, F_GETFL);
3116 rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
3118 ERROR("failed to make Rx interrupt file descriptor"
3119 " %d non-blocking for queue index %d", fd, i);
3120 priv_rx_intr_vec_disable(priv);
3123 intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
3124 intr_handle->efds[count] = fd;
3128 priv_rx_intr_vec_disable(priv);
3130 intr_handle->nb_efd = count;
3135 * Clean up Rx interrupts handler.
3138 * Pointer to private structure.
3141 priv_rx_intr_vec_disable(struct priv *priv)
3143 struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
3145 rte_intr_free_epoll_fd(intr_handle);
3146 free(intr_handle->intr_vec);
3147 intr_handle->nb_efd = 0;
3148 intr_handle->intr_vec = NULL;
3152 * DPDK callback for Rx queue interrupt enable.
3155 * Pointer to Ethernet device structure.
3160 * 0 on success, negative on failure.
3163 mlx4_rx_intr_enable(struct rte_eth_dev *dev, uint16_t idx)
3165 struct priv *priv = dev->data->dev_private;
3166 struct rxq *rxq = (*priv->rxqs)[idx];
3169 if (!rxq || !rxq->channel)
3172 ret = ibv_req_notify_cq(rxq->cq, 0);
3174 WARN("unable to arm interrupt on rx queue %d", idx);
3179 * DPDK callback for Rx queue interrupt disable.
3182 * Pointer to Ethernet device structure.
3187 * 0 on success, negative on failure.
3190 mlx4_rx_intr_disable(struct rte_eth_dev *dev, uint16_t idx)
3192 struct priv *priv = dev->data->dev_private;
3193 struct rxq *rxq = (*priv->rxqs)[idx];
3194 struct ibv_cq *ev_cq;
3198 if (!rxq || !rxq->channel) {
3201 ret = ibv_get_cq_event(rxq->cq->channel, &ev_cq, &ev_ctx);
3202 if (ret || ev_cq != rxq->cq)
3206 WARN("unable to disable interrupt on rx queue %d",
3209 ibv_ack_cq_events(rxq->cq, 1);
3214 * Verify and store value for device argument.
3217 * Key argument to verify.
3219 * Value associated with key.
3220 * @param[in, out] conf
3221 * Shared configuration data.
3224 * 0 on success, negative errno value on failure.
3227 mlx4_arg_parse(const char *key, const char *val, struct mlx4_conf *conf)
3232 tmp = strtoul(val, NULL, 0);
3234 WARN("%s: \"%s\" is not a valid integer", key, val);
3237 if (strcmp(MLX4_PMD_PORT_KVARG, key) == 0) {
3238 uint32_t ports = rte_log2_u32(conf->ports.present);
3241 ERROR("port index %lu outside range [0,%" PRIu32 ")",
3245 if (!(conf->ports.present & (1 << tmp))) {
3246 ERROR("invalid port index %lu", tmp);
3249 conf->ports.enabled |= 1 << tmp;
3251 WARN("%s: unknown parameter", key);
3258 * Parse device parameters.
3261 * Device arguments structure.
3264 * 0 on success, negative errno value on failure.
3267 mlx4_args(struct rte_devargs *devargs, struct mlx4_conf *conf)
3269 struct rte_kvargs *kvlist;
3270 unsigned int arg_count;
3274 if (devargs == NULL)
3276 kvlist = rte_kvargs_parse(devargs->args, pmd_mlx4_init_params);
3277 if (kvlist == NULL) {
3278 ERROR("failed to parse kvargs");
3281 /* Process parameters. */
3282 for (i = 0; pmd_mlx4_init_params[i]; ++i) {
3283 arg_count = rte_kvargs_count(kvlist, MLX4_PMD_PORT_KVARG);
3284 while (arg_count-- > 0) {
3285 ret = rte_kvargs_process(kvlist,
3286 MLX4_PMD_PORT_KVARG,
3287 (int (*)(const char *,
3297 rte_kvargs_free(kvlist);
3301 static struct rte_pci_driver mlx4_driver;
3304 * DPDK callback to register a PCI device.
3306 * This function creates an Ethernet device for each port of a given
3309 * @param[in] pci_drv
3310 * PCI driver structure (mlx4_driver).
3311 * @param[in] pci_dev
3312 * PCI device information.
3315 * 0 on success, negative errno value on failure.
3318 mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
3320 struct ibv_device **list;
3321 struct ibv_device *ibv_dev;
3323 struct ibv_context *attr_ctx = NULL;
3324 struct ibv_device_attr device_attr;
3325 struct mlx4_conf conf = {
3332 assert(pci_drv == &mlx4_driver);
3334 list = ibv_get_device_list(&i);
3337 if (errno == ENOSYS)
3338 ERROR("cannot list devices, is ib_uverbs loaded?");
3343 * For each listed device, check related sysfs entry against
3344 * the provided PCI ID.
3347 struct rte_pci_addr pci_addr;
3350 DEBUG("checking device \"%s\"", list[i]->name);
3351 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
3353 if ((pci_dev->addr.domain != pci_addr.domain) ||
3354 (pci_dev->addr.bus != pci_addr.bus) ||
3355 (pci_dev->addr.devid != pci_addr.devid) ||
3356 (pci_dev->addr.function != pci_addr.function))
3358 vf = (pci_dev->id.device_id ==
3359 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
3360 INFO("PCI information matches, using device \"%s\" (VF: %s)",
3361 list[i]->name, (vf ? "true" : "false"));
3362 attr_ctx = ibv_open_device(list[i]);
3366 if (attr_ctx == NULL) {
3367 ibv_free_device_list(list);
3370 ERROR("cannot access device, is mlx4_ib loaded?");
3373 ERROR("cannot use device, are drivers up to date?");
3381 DEBUG("device opened");
3382 if (ibv_query_device(attr_ctx, &device_attr)) {
3386 INFO("%u port(s) detected", device_attr.phys_port_cnt);
3388 conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
3389 if (mlx4_args(pci_dev->device.devargs, &conf)) {
3390 ERROR("failed to process device arguments");
3394 /* Use all ports when none are defined */
3395 if (!conf.ports.enabled)
3396 conf.ports.enabled = conf.ports.present;
3397 for (i = 0; i < device_attr.phys_port_cnt; i++) {
3398 uint32_t port = i + 1; /* ports are indexed from one */
3399 struct ibv_context *ctx = NULL;
3400 struct ibv_port_attr port_attr;
3401 struct ibv_pd *pd = NULL;
3402 struct priv *priv = NULL;
3403 struct rte_eth_dev *eth_dev = NULL;
3404 struct ether_addr mac;
3406 /* If port is not enabled, skip. */
3407 if (!(conf.ports.enabled & (1 << i)))
3410 DEBUG("using port %u", port);
3412 ctx = ibv_open_device(ibv_dev);
3418 /* Check port status. */
3419 err = ibv_query_port(ctx, port, &port_attr);
3421 ERROR("port query failed: %s", strerror(err));
3426 if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
3427 ERROR("port %d is not configured in Ethernet mode",
3433 if (port_attr.state != IBV_PORT_ACTIVE)
3434 DEBUG("port %d is not active: \"%s\" (%d)",
3435 port, ibv_port_state_str(port_attr.state),
3438 /* Allocate protection domain. */
3439 pd = ibv_alloc_pd(ctx);
3441 ERROR("PD allocation failure");
3446 /* from rte_ethdev.c */
3447 priv = rte_zmalloc("ethdev private structure",
3449 RTE_CACHE_LINE_SIZE);
3451 ERROR("priv allocation failure");
3457 priv->device_attr = device_attr;
3460 priv->mtu = ETHER_MTU;
3463 /* Configure the first MAC address by default. */
3464 if (priv_get_mac(priv, &mac.addr_bytes)) {
3465 ERROR("cannot get MAC address, is mlx4_en loaded?"
3466 " (errno: %s)", strerror(errno));
3470 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
3472 mac.addr_bytes[0], mac.addr_bytes[1],
3473 mac.addr_bytes[2], mac.addr_bytes[3],
3474 mac.addr_bytes[4], mac.addr_bytes[5]);
3475 /* Register MAC address. */
3477 if (priv_mac_addr_add(priv))
3481 char ifname[IF_NAMESIZE];
3483 if (priv_get_ifname(priv, &ifname) == 0)
3484 DEBUG("port %u ifname is \"%s\"",
3485 priv->port, ifname);
3487 DEBUG("port %u ifname is unknown", priv->port);
3490 /* Get actual MTU if possible. */
3491 priv_get_mtu(priv, &priv->mtu);
3492 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
3494 /* from rte_ethdev.c */
3496 char name[RTE_ETH_NAME_MAX_LEN];
3498 snprintf(name, sizeof(name), "%s port %u",
3499 ibv_get_device_name(ibv_dev), port);
3500 eth_dev = rte_eth_dev_allocate(name);
3502 if (eth_dev == NULL) {
3503 ERROR("can not allocate rte ethdev");
3508 eth_dev->data->dev_private = priv;
3509 eth_dev->data->mac_addrs = &priv->mac;
3510 eth_dev->device = &pci_dev->device;
3512 rte_eth_copy_pci_info(eth_dev, pci_dev);
3514 eth_dev->device->driver = &mlx4_driver.driver;
3517 * Copy and override interrupt handle to prevent it from
3518 * being shared between all ethdev instances of a given PCI
3519 * device. This is required to properly handle Rx interrupts
3522 priv->intr_handle_dev = *eth_dev->intr_handle;
3523 eth_dev->intr_handle = &priv->intr_handle_dev;
3525 priv->dev = eth_dev;
3526 eth_dev->dev_ops = &mlx4_dev_ops;
3527 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
3529 /* Bring Ethernet device up. */
3530 DEBUG("forcing Ethernet interface up");
3531 priv_set_flags(priv, ~IFF_UP, IFF_UP);
3532 /* Update link status once if waiting for LSC. */
3533 if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
3534 mlx4_link_update(eth_dev, 0);
3540 claim_zero(ibv_dealloc_pd(pd));
3542 claim_zero(ibv_close_device(ctx));
3544 rte_eth_dev_release_port(eth_dev);
3547 if (i == device_attr.phys_port_cnt)
3551 * XXX if something went wrong in the loop above, there is a resource
3552 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
3553 * long as the dpdk does not provide a way to deallocate a ethdev and a
3554 * way to enumerate the registered ethdevs to free the previous ones.
3559 claim_zero(ibv_close_device(attr_ctx));
3561 ibv_free_device_list(list);
3566 static const struct rte_pci_id mlx4_pci_id_map[] = {
3568 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3569 PCI_DEVICE_ID_MELLANOX_CONNECTX3)
3572 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3573 PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO)
3576 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
3577 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF)
3584 static struct rte_pci_driver mlx4_driver = {
3586 .name = MLX4_DRIVER_NAME
3588 .id_table = mlx4_pci_id_map,
3589 .probe = mlx4_pci_probe,
3590 .drv_flags = RTE_PCI_DRV_INTR_LSC |
3591 RTE_PCI_DRV_INTR_RMV,
3595 * Driver initialization routine.
3597 RTE_INIT(rte_mlx4_pmd_init);
3599 rte_mlx4_pmd_init(void)
3602 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
3603 * huge pages. Calling ibv_fork_init() during init allows
3604 * applications to use fork() safely for purposes other than
3605 * using this PMD, which is not supported in forked processes.
3607 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
3609 rte_pci_register(&mlx4_driver);
3612 RTE_PMD_EXPORT_NAME(net_mlx4, __COUNTER__);
3613 RTE_PMD_REGISTER_PCI_TABLE(net_mlx4, mlx4_pci_id_map);
3614 RTE_PMD_REGISTER_KMOD_DEP(net_mlx4,
3615 "* ib_uverbs & mlx4_en & mlx4_core & mlx4_ib");