4 * Copyright 2012-2015 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.
36 * - RSS hash key and options cannot be modified.
37 * - Hardware counters aren't implemented.
51 #include <arpa/inet.h>
54 #include <sys/ioctl.h>
55 #include <sys/socket.h>
56 #include <netinet/in.h>
58 #include <linux/ethtool.h>
59 #include <linux/sockios.h>
63 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
65 #pragma GCC diagnostic ignored "-pedantic"
67 #include <infiniband/verbs.h>
69 #pragma GCC diagnostic error "-pedantic"
72 /* DPDK headers don't like -pedantic. */
74 #pragma GCC diagnostic ignored "-pedantic"
76 #include <rte_config.h>
77 #include <rte_ether.h>
78 #include <rte_ethdev.h>
81 #include <rte_errno.h>
82 #include <rte_mempool.h>
83 #include <rte_prefetch.h>
84 #include <rte_malloc.h>
85 #include <rte_spinlock.h>
86 #include <rte_atomic.h>
87 #include <rte_version.h>
89 #include <rte_alarm.h>
91 #pragma GCC diagnostic error "-pedantic"
94 /* Generated configuration header. */
95 #include "mlx4_autoconf.h"
100 /* Runtime logging through RTE_LOG() is enabled when not in debugging mode.
101 * Intermediate LOG_*() macros add the required end-of-line characters. */
103 #define INFO(...) DEBUG(__VA_ARGS__)
104 #define WARN(...) DEBUG(__VA_ARGS__)
105 #define ERROR(...) DEBUG(__VA_ARGS__)
107 #define LOG__(level, m, ...) \
108 RTE_LOG(level, PMD, MLX4_DRIVER_NAME ": " m "%c", __VA_ARGS__)
109 #define LOG_(level, ...) LOG__(level, __VA_ARGS__, '\n')
110 #define INFO(...) LOG_(INFO, __VA_ARGS__)
111 #define WARN(...) LOG_(WARNING, __VA_ARGS__)
112 #define ERROR(...) LOG_(ERR, __VA_ARGS__)
115 /* Convenience macros for accessing mbuf fields. */
116 #define NEXT(m) ((m)->next)
117 #define DATA_LEN(m) ((m)->data_len)
118 #define PKT_LEN(m) ((m)->pkt_len)
119 #define DATA_OFF(m) ((m)->data_off)
120 #define SET_DATA_OFF(m, o) ((m)->data_off = (o))
121 #define NB_SEGS(m) ((m)->nb_segs)
122 #define PORT(m) ((m)->port)
124 /* Work Request ID data type (64 bit). */
133 #define WR_ID(o) (((wr_id_t *)&(o))->data)
135 /* Compile-time check. */
136 static inline void wr_id_t_check(void)
138 wr_id_t check[1 + (2 * -!(sizeof(wr_id_t) == sizeof(uint64_t)))];
144 /* Transpose flags. Useful to convert IBV to DPDK flags. */
145 #define TRANSPOSE(val, from, to) \
146 (((from) >= (to)) ? \
147 (((val) & (from)) / ((from) / (to))) : \
148 (((val) & (from)) * ((to) / (from))))
150 struct mlx4_rxq_stats {
151 unsigned int idx; /**< Mapping index. */
152 #ifdef MLX4_PMD_SOFT_COUNTERS
153 uint64_t ipackets; /**< Total of successfully received packets. */
154 uint64_t ibytes; /**< Total of successfully received bytes. */
156 uint64_t idropped; /**< Total of packets dropped when RX ring full. */
157 uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
160 struct mlx4_txq_stats {
161 unsigned int idx; /**< Mapping index. */
162 #ifdef MLX4_PMD_SOFT_COUNTERS
163 uint64_t opackets; /**< Total of successfully sent packets. */
164 uint64_t obytes; /**< Total of successfully sent bytes. */
166 uint64_t odropped; /**< Total of packets not sent when TX ring full. */
169 /* RX element (scattered packets). */
171 struct ibv_recv_wr wr; /* Work Request. */
172 struct ibv_sge sges[MLX4_PMD_SGE_WR_N]; /* Scatter/Gather Elements. */
173 struct rte_mbuf *bufs[MLX4_PMD_SGE_WR_N]; /* SGEs buffers. */
178 struct ibv_recv_wr wr; /* Work Request. */
179 struct ibv_sge sge; /* Scatter/Gather Element. */
180 /* mbuf pointer is derived from WR_ID(wr.wr_id).offset. */
183 /* RX queue descriptor. */
185 struct priv *priv; /* Back pointer to private data. */
186 struct rte_mempool *mp; /* Memory Pool for allocations. */
187 struct ibv_mr *mr; /* Memory Region (for mp). */
188 struct ibv_cq *cq; /* Completion Queue. */
189 struct ibv_qp *qp; /* Queue Pair. */
190 struct ibv_exp_qp_burst_family *if_qp; /* QP burst interface. */
191 struct ibv_exp_cq_family *if_cq; /* CQ interface. */
193 * Each VLAN ID requires a separate flow steering rule.
195 BITFIELD_DECLARE(mac_configured, uint32_t, MLX4_MAX_MAC_ADDRESSES);
196 struct ibv_flow *mac_flow[MLX4_MAX_MAC_ADDRESSES][MLX4_MAX_VLAN_IDS];
197 struct ibv_flow *promisc_flow; /* Promiscuous flow. */
198 struct ibv_flow *allmulti_flow; /* Multicast flow. */
199 unsigned int port_id; /* Port ID for incoming packets. */
200 unsigned int elts_n; /* (*elts)[] length. */
201 unsigned int elts_head; /* Current index in (*elts)[]. */
203 struct rxq_elt_sp (*sp)[]; /* Scattered RX elements. */
204 struct rxq_elt (*no_sp)[]; /* RX elements. */
206 unsigned int sp:1; /* Use scattered RX elements. */
207 unsigned int csum:1; /* Enable checksum offloading. */
208 unsigned int csum_l2tun:1; /* Same for L2 tunnels. */
209 uint32_t mb_len; /* Length of a mp-issued mbuf. */
210 struct mlx4_rxq_stats stats; /* RX queue counters. */
211 unsigned int socket; /* CPU socket ID for allocations. */
212 struct ibv_exp_res_domain *rd; /* Resource Domain. */
217 struct rte_mbuf *buf;
220 /* Linear buffer type. It is used when transmitting buffers with too many
221 * segments that do not fit the hardware queue (see max_send_sge).
222 * Extra segments are copied (linearized) in such buffers, replacing the
223 * last SGE during TX.
224 * The size is arbitrary but large enough to hold a jumbo frame with
225 * 8 segments considering mbuf.buf_len is about 2048 bytes. */
226 typedef uint8_t linear_t[16384];
228 /* TX queue descriptor. */
230 struct priv *priv; /* Back pointer to private data. */
232 struct rte_mempool *mp; /* Cached Memory Pool. */
233 struct ibv_mr *mr; /* Memory Region (for mp). */
234 uint32_t lkey; /* mr->lkey */
235 } mp2mr[MLX4_PMD_TX_MP_CACHE]; /* MP to MR translation table. */
236 struct ibv_cq *cq; /* Completion Queue. */
237 struct ibv_qp *qp; /* Queue Pair. */
238 struct ibv_exp_qp_burst_family *if_qp; /* QP burst interface. */
239 struct ibv_exp_cq_family *if_cq; /* CQ interface. */
240 #if MLX4_PMD_MAX_INLINE > 0
241 uint32_t max_inline; /* Max inline send size <= MLX4_PMD_MAX_INLINE. */
243 unsigned int elts_n; /* (*elts)[] length. */
244 struct txq_elt (*elts)[]; /* TX elements. */
245 unsigned int elts_head; /* Current index in (*elts)[]. */
246 unsigned int elts_tail; /* First element awaiting completion. */
247 unsigned int elts_comp; /* Number of completion requests. */
248 unsigned int elts_comp_cd; /* Countdown for next completion request. */
249 unsigned int elts_comp_cd_init; /* Initial value for countdown. */
250 struct mlx4_txq_stats stats; /* TX queue counters. */
251 linear_t (*elts_linear)[]; /* Linearized buffers. */
252 struct ibv_mr *mr_linear; /* Memory Region for linearized buffers. */
253 unsigned int socket; /* CPU socket ID for allocations. */
254 struct ibv_exp_res_domain *rd; /* Resource Domain. */
258 struct rte_eth_dev *dev; /* Ethernet device. */
259 struct ibv_context *ctx; /* Verbs context. */
260 struct ibv_device_attr device_attr; /* Device properties. */
261 struct ibv_pd *pd; /* Protection Domain. */
263 * MAC addresses array and configuration bit-field.
264 * An extra entry that cannot be modified by the DPDK is reserved
265 * for broadcast frames (destination MAC address ff:ff:ff:ff:ff:ff).
267 struct ether_addr mac[MLX4_MAX_MAC_ADDRESSES];
268 BITFIELD_DECLARE(mac_configured, uint32_t, MLX4_MAX_MAC_ADDRESSES);
271 unsigned int enabled:1; /* If enabled. */
272 unsigned int id:12; /* VLAN ID (0-4095). */
273 } vlan_filter[MLX4_MAX_VLAN_IDS]; /* VLAN filters table. */
274 /* Device properties. */
275 uint16_t mtu; /* Configured MTU. */
276 uint8_t port; /* Physical port number. */
277 unsigned int started:1; /* Device started, flows enabled. */
278 unsigned int promisc:1; /* Device in promiscuous mode. */
279 unsigned int allmulti:1; /* Device receives all multicast packets. */
280 unsigned int hw_qpg:1; /* QP groups are supported. */
281 unsigned int hw_tss:1; /* TSS is supported. */
282 unsigned int hw_rss:1; /* RSS is supported. */
283 unsigned int hw_csum:1; /* Checksum offload is supported. */
284 unsigned int hw_csum_l2tun:1; /* Same for L2 tunnels. */
285 unsigned int rss:1; /* RSS is enabled. */
286 unsigned int vf:1; /* This is a VF device. */
287 unsigned int pending_alarm:1; /* An alarm is pending. */
289 unsigned int inl_recv_size; /* Inline recv size */
291 unsigned int max_rss_tbl_sz; /* Maximum number of RSS queues. */
293 struct rxq rxq_parent; /* Parent queue when RSS is enabled. */
294 unsigned int rxqs_n; /* RX queues array size. */
295 unsigned int txqs_n; /* TX queues array size. */
296 struct rxq *(*rxqs)[]; /* RX queues. */
297 struct txq *(*txqs)[]; /* TX queues. */
298 struct rte_intr_handle intr_handle; /* Interrupt handler. */
299 rte_spinlock_t lock; /* Lock for control functions. */
303 * Lock private structure to protect it from concurrent access in the
307 * Pointer to private structure.
310 priv_lock(struct priv *priv)
312 rte_spinlock_lock(&priv->lock);
316 * Unlock private structure.
319 * Pointer to private structure.
322 priv_unlock(struct priv *priv)
324 rte_spinlock_unlock(&priv->lock);
327 /* Allocate a buffer on the stack and fill it with a printf format string. */
328 #define MKSTR(name, ...) \
329 char name[snprintf(NULL, 0, __VA_ARGS__) + 1]; \
331 snprintf(name, sizeof(name), __VA_ARGS__)
334 * Get interface name from private structure.
337 * Pointer to private structure.
339 * Interface name output buffer.
342 * 0 on success, -1 on failure and errno is set.
345 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
349 unsigned int dev_type = 0;
350 unsigned int dev_port_prev = ~0u;
351 char match[IF_NAMESIZE] = "";
354 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
360 while ((dent = readdir(dir)) != NULL) {
361 char *name = dent->d_name;
363 unsigned int dev_port;
366 if ((name[0] == '.') &&
367 ((name[1] == '\0') ||
368 ((name[1] == '.') && (name[2] == '\0'))))
371 MKSTR(path, "%s/device/net/%s/%s",
372 priv->ctx->device->ibdev_path, name,
373 (dev_type ? "dev_id" : "dev_port"));
375 file = fopen(path, "rb");
380 * Switch to dev_id when dev_port does not exist as
381 * is the case with Linux kernel versions < 3.15.
392 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
397 * Switch to dev_id when dev_port returns the same value for
398 * all ports. May happen when using a MOFED release older than
399 * 3.0 with a Linux kernel >= 3.15.
401 if (dev_port == dev_port_prev)
403 dev_port_prev = dev_port;
404 if (dev_port == (priv->port - 1u))
405 snprintf(match, sizeof(match), "%s", name);
408 if (match[0] == '\0')
410 strncpy(*ifname, match, sizeof(*ifname));
415 * Read from sysfs entry.
418 * Pointer to private structure.
420 * Entry name relative to sysfs path.
422 * Data output buffer.
427 * 0 on success, -1 on failure and errno is set.
430 priv_sysfs_read(const struct priv *priv, const char *entry,
431 char *buf, size_t size)
433 char ifname[IF_NAMESIZE];
438 if (priv_get_ifname(priv, &ifname))
441 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
444 file = fopen(path, "rb");
447 ret = fread(buf, 1, size, file);
449 if (((size_t)ret < size) && (ferror(file)))
459 * Write to sysfs entry.
462 * Pointer to private structure.
464 * Entry name relative to sysfs path.
471 * 0 on success, -1 on failure and errno is set.
474 priv_sysfs_write(const struct priv *priv, const char *entry,
475 char *buf, size_t size)
477 char ifname[IF_NAMESIZE];
482 if (priv_get_ifname(priv, &ifname))
485 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
488 file = fopen(path, "wb");
491 ret = fwrite(buf, 1, size, file);
493 if (((size_t)ret < size) || (ferror(file)))
503 * Get unsigned long sysfs property.
506 * Pointer to private structure.
508 * Entry name relative to sysfs path.
510 * Value output buffer.
513 * 0 on success, -1 on failure and errno is set.
516 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
519 unsigned long value_ret;
522 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
524 DEBUG("cannot read %s value from sysfs: %s",
525 name, strerror(errno));
528 value_str[ret] = '\0';
530 value_ret = strtoul(value_str, NULL, 0);
532 DEBUG("invalid %s value `%s': %s", name, value_str,
541 * Set unsigned long sysfs property.
544 * Pointer to private structure.
546 * Entry name relative to sysfs path.
551 * 0 on success, -1 on failure and errno is set.
554 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
557 MKSTR(value_str, "%lu", value);
559 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
561 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
562 name, value_str, value, strerror(errno));
569 * Perform ifreq ioctl() on associated Ethernet device.
572 * Pointer to private structure.
574 * Request number to pass to ioctl().
576 * Interface request structure output buffer.
579 * 0 on success, -1 on failure and errno is set.
582 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
584 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
589 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
590 ret = ioctl(sock, req, ifr);
599 * Pointer to private structure.
601 * MTU value output buffer.
604 * 0 on success, -1 on failure and errno is set.
607 priv_get_mtu(struct priv *priv, uint16_t *mtu)
609 unsigned long ulong_mtu;
611 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
621 * Pointer to private structure.
626 * 0 on success, -1 on failure and errno is set.
629 priv_set_mtu(struct priv *priv, uint16_t mtu)
631 return priv_set_sysfs_ulong(priv, "mtu", mtu);
638 * Pointer to private structure.
640 * Bitmask for flags that must remain untouched.
642 * Bitmask for flags to modify.
645 * 0 on success, -1 on failure and errno is set.
648 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
652 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
656 return priv_set_sysfs_ulong(priv, "flags", tmp);
659 /* Device configuration. */
662 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
663 unsigned int socket, const struct rte_eth_rxconf *conf,
664 struct rte_mempool *mp);
667 rxq_cleanup(struct rxq *rxq);
670 * Ethernet device configuration.
672 * Prepare the driver for a given number of TX and RX queues.
673 * Allocate parent RSS queue when several RX queues are requested.
676 * Pointer to Ethernet device structure.
679 * 0 on success, errno value on failure.
682 dev_configure(struct rte_eth_dev *dev)
684 struct priv *priv = dev->data->dev_private;
685 unsigned int rxqs_n = dev->data->nb_rx_queues;
686 unsigned int txqs_n = dev->data->nb_tx_queues;
690 priv->rxqs = (void *)dev->data->rx_queues;
691 priv->txqs = (void *)dev->data->tx_queues;
692 if (txqs_n != priv->txqs_n) {
693 INFO("%p: TX queues number update: %u -> %u",
694 (void *)dev, priv->txqs_n, txqs_n);
695 priv->txqs_n = txqs_n;
697 if (rxqs_n == priv->rxqs_n)
699 INFO("%p: RX queues number update: %u -> %u",
700 (void *)dev, priv->rxqs_n, rxqs_n);
701 /* If RSS is enabled, disable it first. */
705 /* Only if there are no remaining child RX queues. */
706 for (i = 0; (i != priv->rxqs_n); ++i)
707 if ((*priv->rxqs)[i] != NULL)
709 rxq_cleanup(&priv->rxq_parent);
714 /* Nothing else to do. */
715 priv->rxqs_n = rxqs_n;
718 /* Allocate a new RSS parent queue if supported by hardware. */
720 ERROR("%p: only a single RX queue can be configured when"
721 " hardware doesn't support RSS",
725 /* Fail if hardware doesn't support that many RSS queues. */
726 if (rxqs_n >= priv->max_rss_tbl_sz) {
727 ERROR("%p: only %u RX queues can be configured for RSS",
728 (void *)dev, priv->max_rss_tbl_sz);
733 priv->rxqs_n = rxqs_n;
734 ret = rxq_setup(dev, &priv->rxq_parent, 0, 0, NULL, NULL);
737 /* Failure, rollback. */
745 * DPDK callback for Ethernet device configuration.
748 * Pointer to Ethernet device structure.
751 * 0 on success, negative errno value on failure.
754 mlx4_dev_configure(struct rte_eth_dev *dev)
756 struct priv *priv = dev->data->dev_private;
760 ret = dev_configure(dev);
766 /* TX queues handling. */
769 * Allocate TX queue elements.
772 * Pointer to TX queue structure.
774 * Number of elements to allocate.
777 * 0 on success, errno value on failure.
780 txq_alloc_elts(struct txq *txq, unsigned int elts_n)
783 struct txq_elt (*elts)[elts_n] =
784 rte_calloc_socket("TXQ", 1, sizeof(*elts), 0, txq->socket);
785 linear_t (*elts_linear)[elts_n] =
786 rte_calloc_socket("TXQ", 1, sizeof(*elts_linear), 0,
788 struct ibv_mr *mr_linear = NULL;
791 if ((elts == NULL) || (elts_linear == NULL)) {
792 ERROR("%p: can't allocate packets array", (void *)txq);
797 ibv_reg_mr(txq->priv->pd, elts_linear, sizeof(*elts_linear),
798 (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
799 if (mr_linear == NULL) {
800 ERROR("%p: unable to configure MR, ibv_reg_mr() failed",
805 for (i = 0; (i != elts_n); ++i) {
806 struct txq_elt *elt = &(*elts)[i];
810 DEBUG("%p: allocated and configured %u WRs", (void *)txq, elts_n);
811 txq->elts_n = elts_n;
816 /* Request send completion every MLX4_PMD_TX_PER_COMP_REQ packets or
817 * at least 4 times per ring. */
818 txq->elts_comp_cd_init =
819 ((MLX4_PMD_TX_PER_COMP_REQ < (elts_n / 4)) ?
820 MLX4_PMD_TX_PER_COMP_REQ : (elts_n / 4));
821 txq->elts_comp_cd = txq->elts_comp_cd_init;
822 txq->elts_linear = elts_linear;
823 txq->mr_linear = mr_linear;
827 if (mr_linear != NULL)
828 claim_zero(ibv_dereg_mr(mr_linear));
830 rte_free(elts_linear);
833 DEBUG("%p: failed, freed everything", (void *)txq);
839 * Free TX queue elements.
842 * Pointer to TX queue structure.
845 txq_free_elts(struct txq *txq)
848 unsigned int elts_n = txq->elts_n;
849 struct txq_elt (*elts)[elts_n] = txq->elts;
850 linear_t (*elts_linear)[elts_n] = txq->elts_linear;
851 struct ibv_mr *mr_linear = txq->mr_linear;
853 DEBUG("%p: freeing WRs", (void *)txq);
856 txq->elts_linear = NULL;
857 txq->mr_linear = NULL;
858 if (mr_linear != NULL)
859 claim_zero(ibv_dereg_mr(mr_linear));
861 rte_free(elts_linear);
864 for (i = 0; (i != elemof(*elts)); ++i) {
865 struct txq_elt *elt = &(*elts)[i];
867 if (elt->buf == NULL)
869 rte_pktmbuf_free(elt->buf);
876 * Clean up a TX queue.
878 * Destroy objects, free allocated memory and reset the structure for reuse.
881 * Pointer to TX queue structure.
884 txq_cleanup(struct txq *txq)
886 struct ibv_exp_release_intf_params params;
889 DEBUG("cleaning up %p", (void *)txq);
891 if (txq->if_qp != NULL) {
892 assert(txq->priv != NULL);
893 assert(txq->priv->ctx != NULL);
894 assert(txq->qp != NULL);
895 params = (struct ibv_exp_release_intf_params){
898 claim_zero(ibv_exp_release_intf(txq->priv->ctx,
902 if (txq->if_cq != NULL) {
903 assert(txq->priv != NULL);
904 assert(txq->priv->ctx != NULL);
905 assert(txq->cq != NULL);
906 params = (struct ibv_exp_release_intf_params){
909 claim_zero(ibv_exp_release_intf(txq->priv->ctx,
914 claim_zero(ibv_destroy_qp(txq->qp));
916 claim_zero(ibv_destroy_cq(txq->cq));
917 if (txq->rd != NULL) {
918 struct ibv_exp_destroy_res_domain_attr attr = {
922 assert(txq->priv != NULL);
923 assert(txq->priv->ctx != NULL);
924 claim_zero(ibv_exp_destroy_res_domain(txq->priv->ctx,
928 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
929 if (txq->mp2mr[i].mp == NULL)
931 assert(txq->mp2mr[i].mr != NULL);
932 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
934 memset(txq, 0, sizeof(*txq));
938 * Manage TX completions.
940 * When sending a burst, mlx4_tx_burst() posts several WRs.
941 * To improve performance, a completion event is only required once every
942 * MLX4_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
943 * for other WRs, but this information would not be used anyway.
946 * Pointer to TX queue structure.
949 * 0 on success, -1 on failure.
952 txq_complete(struct txq *txq)
954 unsigned int elts_comp = txq->elts_comp;
955 unsigned int elts_tail = txq->elts_tail;
956 const unsigned int elts_n = txq->elts_n;
959 if (unlikely(elts_comp == 0))
962 DEBUG("%p: processing %u work requests completions",
963 (void *)txq, elts_comp);
965 wcs_n = txq->if_cq->poll_cnt(txq->cq, elts_comp);
966 if (unlikely(wcs_n == 0))
968 if (unlikely(wcs_n < 0)) {
969 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
974 assert(elts_comp <= txq->elts_comp);
976 * Assume WC status is successful as nothing can be done about it
979 elts_tail += wcs_n * txq->elts_comp_cd_init;
980 if (elts_tail >= elts_n)
982 txq->elts_tail = elts_tail;
983 txq->elts_comp = elts_comp;
988 * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
989 * the cloned mbuf is allocated is returned instead.
995 * Memory pool where data is located for given mbuf.
997 static struct rte_mempool *
998 txq_mb2mp(struct rte_mbuf *buf)
1000 if (unlikely(RTE_MBUF_INDIRECT(buf)))
1001 return rte_mbuf_from_indirect(buf)->pool;
1006 * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
1007 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
1008 * remove an entry first.
1011 * Pointer to TX queue structure.
1013 * Memory Pool for which a Memory Region lkey must be returned.
1016 * mr->lkey on success, (uint32_t)-1 on failure.
1019 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
1024 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
1025 if (unlikely(txq->mp2mr[i].mp == NULL)) {
1026 /* Unknown MP, add a new MR for it. */
1029 if (txq->mp2mr[i].mp == mp) {
1030 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
1031 assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
1032 return txq->mp2mr[i].lkey;
1035 /* Add a new entry, register MR first. */
1036 DEBUG("%p: discovered new memory pool %p", (void *)txq, (void *)mp);
1037 mr = ibv_reg_mr(txq->priv->pd,
1038 (void *)mp->elt_va_start,
1039 (mp->elt_va_end - mp->elt_va_start),
1040 (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
1041 if (unlikely(mr == NULL)) {
1042 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
1044 return (uint32_t)-1;
1046 if (unlikely(i == elemof(txq->mp2mr))) {
1047 /* Table is full, remove oldest entry. */
1048 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
1051 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
1052 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
1053 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
1055 /* Store the new entry. */
1056 txq->mp2mr[i].mp = mp;
1057 txq->mp2mr[i].mr = mr;
1058 txq->mp2mr[i].lkey = mr->lkey;
1059 DEBUG("%p: new MR lkey for MP %p: 0x%08" PRIu32,
1060 (void *)txq, (void *)mp, txq->mp2mr[i].lkey);
1061 return txq->mp2mr[i].lkey;
1064 #if MLX4_PMD_SGE_WR_N > 1
1067 * Copy scattered mbuf contents to a single linear buffer.
1069 * @param[out] linear
1070 * Linear output buffer.
1072 * Scattered input buffer.
1075 * Number of bytes copied to the output buffer or 0 if not large enough.
1078 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
1080 unsigned int size = 0;
1081 unsigned int offset;
1084 unsigned int len = DATA_LEN(buf);
1088 if (unlikely(size > sizeof(*linear)))
1090 memcpy(&(*linear)[offset],
1091 rte_pktmbuf_mtod(buf, uint8_t *),
1094 } while (buf != NULL);
1099 * Handle scattered buffers for mlx4_tx_burst().
1102 * TX queue structure.
1104 * Number of segments in buf.
1106 * TX queue element to fill.
1108 * Buffer to process.
1110 * Index of the linear buffer to use if necessary (normally txq->elts_head).
1112 * Array filled with SGEs on success.
1115 * A structure containing the processed packet size in bytes and the
1116 * number of SGEs. Both fields are set to (unsigned int)-1 in case of
1119 static struct tx_burst_sg_ret {
1120 unsigned int length;
1123 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
1124 struct rte_mbuf *buf, unsigned int elts_head,
1125 struct ibv_sge (*sges)[MLX4_PMD_SGE_WR_N])
1127 unsigned int sent_size = 0;
1131 /* When there are too many segments, extra segments are
1132 * linearized in the last SGE. */
1133 if (unlikely(segs > elemof(*sges))) {
1134 segs = (elemof(*sges) - 1);
1137 /* Update element. */
1139 /* Register segments as SGEs. */
1140 for (j = 0; (j != segs); ++j) {
1141 struct ibv_sge *sge = &(*sges)[j];
1144 /* Retrieve Memory Region key for this memory pool. */
1145 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
1146 if (unlikely(lkey == (uint32_t)-1)) {
1147 /* MR does not exist. */
1148 DEBUG("%p: unable to get MP <-> MR association",
1150 /* Clean up TX element. */
1155 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
1157 rte_prefetch0((volatile void *)
1158 (uintptr_t)sge->addr);
1159 sge->length = DATA_LEN(buf);
1161 sent_size += sge->length;
1164 /* If buf is not NULL here and is not going to be linearized,
1165 * nb_segs is not valid. */
1167 assert((buf == NULL) || (linearize));
1168 /* Linearize extra segments. */
1170 struct ibv_sge *sge = &(*sges)[segs];
1171 linear_t *linear = &(*txq->elts_linear)[elts_head];
1172 unsigned int size = linearize_mbuf(linear, buf);
1174 assert(segs == (elemof(*sges) - 1));
1176 /* Invalid packet. */
1177 DEBUG("%p: packet too large to be linearized.",
1179 /* Clean up TX element. */
1183 /* If MLX4_PMD_SGE_WR_N is 1, free mbuf immediately. */
1184 if (elemof(*sges) == 1) {
1186 struct rte_mbuf *next = NEXT(buf);
1188 rte_pktmbuf_free_seg(buf);
1190 } while (buf != NULL);
1194 sge->addr = (uintptr_t)&(*linear)[0];
1196 sge->lkey = txq->mr_linear->lkey;
1198 /* Include last segment. */
1201 return (struct tx_burst_sg_ret){
1202 .length = sent_size,
1206 return (struct tx_burst_sg_ret){
1212 #endif /* MLX4_PMD_SGE_WR_N > 1 */
1215 * DPDK callback for TX.
1218 * Generic pointer to TX queue structure.
1220 * Packets to transmit.
1222 * Number of packets in array.
1225 * Number of packets successfully transmitted (<= pkts_n).
1228 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1230 struct txq *txq = (struct txq *)dpdk_txq;
1231 unsigned int elts_head = txq->elts_head;
1232 const unsigned int elts_tail = txq->elts_tail;
1233 const unsigned int elts_n = txq->elts_n;
1234 unsigned int elts_comp_cd = txq->elts_comp_cd;
1235 unsigned int elts_comp = 0;
1240 assert(elts_comp_cd != 0);
1242 max = (elts_n - (elts_head - elts_tail));
1246 assert(max <= elts_n);
1247 /* Always leave one free entry in the ring. */
1253 for (i = 0; (i != max); ++i) {
1254 struct rte_mbuf *buf = pkts[i];
1255 unsigned int elts_head_next =
1256 (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
1257 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
1258 struct txq_elt *elt = &(*txq->elts)[elts_head];
1259 unsigned int segs = NB_SEGS(buf);
1260 #ifdef MLX4_PMD_SOFT_COUNTERS
1261 unsigned int sent_size = 0;
1263 uint32_t send_flags = 0;
1265 /* Clean up old buffer. */
1266 if (likely(elt->buf != NULL)) {
1267 struct rte_mbuf *tmp = elt->buf;
1269 /* Faster than rte_pktmbuf_free(). */
1271 struct rte_mbuf *next = NEXT(tmp);
1273 rte_pktmbuf_free_seg(tmp);
1275 } while (tmp != NULL);
1277 /* Request TX completion. */
1278 if (unlikely(--elts_comp_cd == 0)) {
1279 elts_comp_cd = txq->elts_comp_cd_init;
1281 send_flags |= IBV_EXP_QP_BURST_SIGNALED;
1283 /* Should we enable HW CKSUM offload */
1285 (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
1286 send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
1287 /* HW does not support checksum offloads at arbitrary
1288 * offsets but automatically recognizes the packet
1289 * type. For inner L3/L4 checksums, only VXLAN (UDP)
1290 * tunnels are currently supported. */
1291 if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
1292 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
1294 if (likely(segs == 1)) {
1299 /* Retrieve buffer information. */
1300 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1301 length = DATA_LEN(buf);
1302 /* Retrieve Memory Region key for this memory pool. */
1303 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
1304 if (unlikely(lkey == (uint32_t)-1)) {
1305 /* MR does not exist. */
1306 DEBUG("%p: unable to get MP <-> MR"
1307 " association", (void *)txq);
1308 /* Clean up TX element. */
1312 /* Update element. */
1315 rte_prefetch0((volatile void *)
1317 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1318 /* Put packet into send queue. */
1319 #if MLX4_PMD_MAX_INLINE > 0
1320 if (length <= txq->max_inline)
1321 err = txq->if_qp->send_pending_inline
1328 err = txq->if_qp->send_pending
1336 #ifdef MLX4_PMD_SOFT_COUNTERS
1337 sent_size += length;
1340 #if MLX4_PMD_SGE_WR_N > 1
1341 struct ibv_sge sges[MLX4_PMD_SGE_WR_N];
1342 struct tx_burst_sg_ret ret;
1344 ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
1346 if (ret.length == (unsigned int)-1)
1348 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1349 /* Put SG list into send queue. */
1350 err = txq->if_qp->send_pending_sg_list
1357 #ifdef MLX4_PMD_SOFT_COUNTERS
1358 sent_size += ret.length;
1360 #else /* MLX4_PMD_SGE_WR_N > 1 */
1361 DEBUG("%p: TX scattered buffers support not"
1362 " compiled in", (void *)txq);
1364 #endif /* MLX4_PMD_SGE_WR_N > 1 */
1366 elts_head = elts_head_next;
1367 #ifdef MLX4_PMD_SOFT_COUNTERS
1368 /* Increment sent bytes counter. */
1369 txq->stats.obytes += sent_size;
1373 /* Take a shortcut if nothing must be sent. */
1374 if (unlikely(i == 0))
1376 #ifdef MLX4_PMD_SOFT_COUNTERS
1377 /* Increment sent packets counter. */
1378 txq->stats.opackets += i;
1380 /* Ring QP doorbell. */
1381 err = txq->if_qp->send_flush(txq->qp);
1382 if (unlikely(err)) {
1383 /* A nonzero value is not supposed to be returned.
1384 * Nothing can be done about it. */
1385 DEBUG("%p: send_flush() failed with error %d",
1388 txq->elts_head = elts_head;
1389 txq->elts_comp += elts_comp;
1390 txq->elts_comp_cd = elts_comp_cd;
1395 * Configure a TX queue.
1398 * Pointer to Ethernet device structure.
1400 * Pointer to TX queue structure.
1402 * Number of descriptors to configure in queue.
1404 * NUMA socket on which memory must be allocated.
1406 * Thresholds parameters.
1409 * 0 on success, errno value on failure.
1412 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1413 unsigned int socket, const struct rte_eth_txconf *conf)
1415 struct priv *priv = dev->data->dev_private;
1421 struct ibv_exp_query_intf_params params;
1422 struct ibv_exp_qp_init_attr init;
1423 struct ibv_exp_res_domain_init_attr rd;
1424 struct ibv_exp_cq_init_attr cq;
1425 struct ibv_exp_qp_attr mod;
1427 enum ibv_exp_query_intf_status status;
1430 (void)conf; /* Thresholds configuration (ignored). */
1431 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
1432 ERROR("%p: invalid number of TX descriptors (must be a"
1433 " multiple of %d)", (void *)dev, MLX4_PMD_SGE_WR_N);
1436 desc /= MLX4_PMD_SGE_WR_N;
1437 /* MRs will be registered in mp2mr[] later. */
1438 attr.rd = (struct ibv_exp_res_domain_init_attr){
1439 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
1440 IBV_EXP_RES_DOMAIN_MSG_MODEL),
1441 .thread_model = IBV_EXP_THREAD_SINGLE,
1442 .msg_model = IBV_EXP_MSG_HIGH_BW,
1444 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
1445 if (tmpl.rd == NULL) {
1447 ERROR("%p: RD creation failure: %s",
1448 (void *)dev, strerror(ret));
1451 attr.cq = (struct ibv_exp_cq_init_attr){
1452 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
1453 .res_domain = tmpl.rd,
1455 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
1456 if (tmpl.cq == NULL) {
1458 ERROR("%p: CQ creation failure: %s",
1459 (void *)dev, strerror(ret));
1462 DEBUG("priv->device_attr.max_qp_wr is %d",
1463 priv->device_attr.max_qp_wr);
1464 DEBUG("priv->device_attr.max_sge is %d",
1465 priv->device_attr.max_sge);
1466 attr.init = (struct ibv_exp_qp_init_attr){
1467 /* CQ to be associated with the send queue. */
1469 /* CQ to be associated with the receive queue. */
1472 /* Max number of outstanding WRs. */
1473 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1474 priv->device_attr.max_qp_wr :
1476 /* Max number of scatter/gather elements in a WR. */
1477 .max_send_sge = ((priv->device_attr.max_sge <
1478 MLX4_PMD_SGE_WR_N) ?
1479 priv->device_attr.max_sge :
1481 #if MLX4_PMD_MAX_INLINE > 0
1482 .max_inline_data = MLX4_PMD_MAX_INLINE,
1485 .qp_type = IBV_QPT_RAW_PACKET,
1486 /* Do *NOT* enable this, completions events are managed per
1490 .res_domain = tmpl.rd,
1491 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
1492 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
1494 tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
1495 if (tmpl.qp == NULL) {
1496 ret = (errno ? errno : EINVAL);
1497 ERROR("%p: QP creation failure: %s",
1498 (void *)dev, strerror(ret));
1501 #if MLX4_PMD_MAX_INLINE > 0
1502 /* ibv_create_qp() updates this value. */
1503 tmpl.max_inline = attr.init.cap.max_inline_data;
1505 attr.mod = (struct ibv_exp_qp_attr){
1506 /* Move the QP to this state. */
1507 .qp_state = IBV_QPS_INIT,
1508 /* Primary port number. */
1509 .port_num = priv->port
1511 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
1512 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
1514 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1515 (void *)dev, strerror(ret));
1518 ret = txq_alloc_elts(&tmpl, desc);
1520 ERROR("%p: TXQ allocation failed: %s",
1521 (void *)dev, strerror(ret));
1524 attr.mod = (struct ibv_exp_qp_attr){
1525 .qp_state = IBV_QPS_RTR
1527 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1529 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1530 (void *)dev, strerror(ret));
1533 attr.mod.qp_state = IBV_QPS_RTS;
1534 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1536 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1537 (void *)dev, strerror(ret));
1540 attr.params = (struct ibv_exp_query_intf_params){
1541 .intf_scope = IBV_EXP_INTF_GLOBAL,
1542 .intf = IBV_EXP_INTF_CQ,
1545 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
1546 if (tmpl.if_cq == NULL) {
1547 ERROR("%p: CQ interface family query failed with status %d",
1548 (void *)dev, status);
1551 attr.params = (struct ibv_exp_query_intf_params){
1552 .intf_scope = IBV_EXP_INTF_GLOBAL,
1553 .intf = IBV_EXP_INTF_QP_BURST,
1555 #ifdef HAVE_EXP_QP_BURST_CREATE_DISABLE_ETH_LOOPBACK
1556 /* MC loopback must be disabled when not using a VF. */
1559 IBV_EXP_QP_BURST_CREATE_DISABLE_ETH_LOOPBACK :
1563 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
1564 if (tmpl.if_qp == NULL) {
1565 ERROR("%p: QP interface family query failed with status %d",
1566 (void *)dev, status);
1569 /* Clean up txq in case we're reinitializing it. */
1570 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1573 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1583 * DPDK callback to configure a TX queue.
1586 * Pointer to Ethernet device structure.
1590 * Number of descriptors to configure in queue.
1592 * NUMA socket on which memory must be allocated.
1594 * Thresholds parameters.
1597 * 0 on success, negative errno value on failure.
1600 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1601 unsigned int socket, const struct rte_eth_txconf *conf)
1603 struct priv *priv = dev->data->dev_private;
1604 struct txq *txq = (*priv->txqs)[idx];
1608 DEBUG("%p: configuring queue %u for %u descriptors",
1609 (void *)dev, idx, desc);
1610 if (idx >= priv->txqs_n) {
1611 ERROR("%p: queue index out of range (%u >= %u)",
1612 (void *)dev, idx, priv->txqs_n);
1617 DEBUG("%p: reusing already allocated queue index %u (%p)",
1618 (void *)dev, idx, (void *)txq);
1619 if (priv->started) {
1623 (*priv->txqs)[idx] = NULL;
1626 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1628 ERROR("%p: unable to allocate queue index %u",
1634 ret = txq_setup(dev, txq, desc, socket, conf);
1638 txq->stats.idx = idx;
1639 DEBUG("%p: adding TX queue %p to list",
1640 (void *)dev, (void *)txq);
1641 (*priv->txqs)[idx] = txq;
1642 /* Update send callback. */
1643 dev->tx_pkt_burst = mlx4_tx_burst;
1650 * DPDK callback to release a TX queue.
1653 * Generic TX queue pointer.
1656 mlx4_tx_queue_release(void *dpdk_txq)
1658 struct txq *txq = (struct txq *)dpdk_txq;
1666 for (i = 0; (i != priv->txqs_n); ++i)
1667 if ((*priv->txqs)[i] == txq) {
1668 DEBUG("%p: removing TX queue %p from list",
1669 (void *)priv->dev, (void *)txq);
1670 (*priv->txqs)[i] = NULL;
1678 /* RX queues handling. */
1681 * Allocate RX queue elements with scattered packets support.
1684 * Pointer to RX queue structure.
1686 * Number of elements to allocate.
1688 * If not NULL, fetch buffers from this array instead of allocating them
1689 * with rte_pktmbuf_alloc().
1692 * 0 on success, errno value on failure.
1695 rxq_alloc_elts_sp(struct rxq *rxq, unsigned int elts_n,
1696 struct rte_mbuf **pool)
1699 struct rxq_elt_sp (*elts)[elts_n] =
1700 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1705 ERROR("%p: can't allocate packets array", (void *)rxq);
1709 /* For each WR (packet). */
1710 for (i = 0; (i != elts_n); ++i) {
1712 struct rxq_elt_sp *elt = &(*elts)[i];
1713 struct ibv_recv_wr *wr = &elt->wr;
1714 struct ibv_sge (*sges)[(elemof(elt->sges))] = &elt->sges;
1716 /* These two arrays must have the same size. */
1717 assert(elemof(elt->sges) == elemof(elt->bufs));
1720 wr->next = &(*elts)[(i + 1)].wr;
1721 wr->sg_list = &(*sges)[0];
1722 wr->num_sge = elemof(*sges);
1723 /* For each SGE (segment). */
1724 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1725 struct ibv_sge *sge = &(*sges)[j];
1726 struct rte_mbuf *buf;
1730 assert(buf != NULL);
1731 rte_pktmbuf_reset(buf);
1733 buf = rte_pktmbuf_alloc(rxq->mp);
1735 assert(pool == NULL);
1736 ERROR("%p: empty mbuf pool", (void *)rxq);
1741 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1742 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1743 /* Buffer is supposed to be empty. */
1744 assert(rte_pktmbuf_data_len(buf) == 0);
1745 assert(rte_pktmbuf_pkt_len(buf) == 0);
1746 /* sge->addr must be able to store a pointer. */
1747 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1749 /* The first SGE keeps its headroom. */
1750 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
1751 sge->length = (buf->buf_len -
1752 RTE_PKTMBUF_HEADROOM);
1754 /* Subsequent SGEs lose theirs. */
1755 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1756 SET_DATA_OFF(buf, 0);
1757 sge->addr = (uintptr_t)buf->buf_addr;
1758 sge->length = buf->buf_len;
1760 sge->lkey = rxq->mr->lkey;
1761 /* Redundant check for tailroom. */
1762 assert(sge->length == rte_pktmbuf_tailroom(buf));
1765 /* The last WR pointer must be NULL. */
1766 (*elts)[(i - 1)].wr.next = NULL;
1767 DEBUG("%p: allocated and configured %u WRs (%zu segments)",
1768 (void *)rxq, elts_n, (elts_n * elemof((*elts)[0].sges)));
1769 rxq->elts_n = elts_n;
1771 rxq->elts.sp = elts;
1776 assert(pool == NULL);
1777 for (i = 0; (i != elemof(*elts)); ++i) {
1779 struct rxq_elt_sp *elt = &(*elts)[i];
1781 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1782 struct rte_mbuf *buf = elt->bufs[j];
1785 rte_pktmbuf_free_seg(buf);
1790 DEBUG("%p: failed, freed everything", (void *)rxq);
1796 * Free RX queue elements with scattered packets support.
1799 * Pointer to RX queue structure.
1802 rxq_free_elts_sp(struct rxq *rxq)
1805 unsigned int elts_n = rxq->elts_n;
1806 struct rxq_elt_sp (*elts)[elts_n] = rxq->elts.sp;
1808 DEBUG("%p: freeing WRs", (void *)rxq);
1810 rxq->elts.sp = NULL;
1813 for (i = 0; (i != elemof(*elts)); ++i) {
1815 struct rxq_elt_sp *elt = &(*elts)[i];
1817 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1818 struct rte_mbuf *buf = elt->bufs[j];
1821 rte_pktmbuf_free_seg(buf);
1828 * Allocate RX queue elements.
1831 * Pointer to RX queue structure.
1833 * Number of elements to allocate.
1835 * If not NULL, fetch buffers from this array instead of allocating them
1836 * with rte_pktmbuf_alloc().
1839 * 0 on success, errno value on failure.
1842 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n, struct rte_mbuf **pool)
1845 struct rxq_elt (*elts)[elts_n] =
1846 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1851 ERROR("%p: can't allocate packets array", (void *)rxq);
1855 /* For each WR (packet). */
1856 for (i = 0; (i != elts_n); ++i) {
1857 struct rxq_elt *elt = &(*elts)[i];
1858 struct ibv_recv_wr *wr = &elt->wr;
1859 struct ibv_sge *sge = &(*elts)[i].sge;
1860 struct rte_mbuf *buf;
1864 assert(buf != NULL);
1865 rte_pktmbuf_reset(buf);
1867 buf = rte_pktmbuf_alloc(rxq->mp);
1869 assert(pool == NULL);
1870 ERROR("%p: empty mbuf pool", (void *)rxq);
1874 /* Configure WR. Work request ID contains its own index in
1875 * the elts array and the offset between SGE buffer header and
1877 WR_ID(wr->wr_id).id = i;
1878 WR_ID(wr->wr_id).offset =
1879 (((uintptr_t)buf->buf_addr + RTE_PKTMBUF_HEADROOM) -
1881 wr->next = &(*elts)[(i + 1)].wr;
1884 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1885 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1886 /* Buffer is supposed to be empty. */
1887 assert(rte_pktmbuf_data_len(buf) == 0);
1888 assert(rte_pktmbuf_pkt_len(buf) == 0);
1889 /* sge->addr must be able to store a pointer. */
1890 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1891 /* SGE keeps its headroom. */
1892 sge->addr = (uintptr_t)
1893 ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
1894 sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
1895 sge->lkey = rxq->mr->lkey;
1896 /* Redundant check for tailroom. */
1897 assert(sge->length == rte_pktmbuf_tailroom(buf));
1898 /* Make sure elts index and SGE mbuf pointer can be deduced
1900 if ((WR_ID(wr->wr_id).id != i) ||
1901 ((void *)((uintptr_t)sge->addr -
1902 WR_ID(wr->wr_id).offset) != buf)) {
1903 ERROR("%p: cannot store index and offset in WR ID",
1906 rte_pktmbuf_free(buf);
1911 /* The last WR pointer must be NULL. */
1912 (*elts)[(i - 1)].wr.next = NULL;
1913 DEBUG("%p: allocated and configured %u single-segment WRs",
1914 (void *)rxq, elts_n);
1915 rxq->elts_n = elts_n;
1917 rxq->elts.no_sp = elts;
1922 assert(pool == NULL);
1923 for (i = 0; (i != elemof(*elts)); ++i) {
1924 struct rxq_elt *elt = &(*elts)[i];
1925 struct rte_mbuf *buf;
1927 if (elt->sge.addr == 0)
1929 assert(WR_ID(elt->wr.wr_id).id == i);
1930 buf = (void *)((uintptr_t)elt->sge.addr -
1931 WR_ID(elt->wr.wr_id).offset);
1932 rte_pktmbuf_free_seg(buf);
1936 DEBUG("%p: failed, freed everything", (void *)rxq);
1942 * Free RX queue elements.
1945 * Pointer to RX queue structure.
1948 rxq_free_elts(struct rxq *rxq)
1951 unsigned int elts_n = rxq->elts_n;
1952 struct rxq_elt (*elts)[elts_n] = rxq->elts.no_sp;
1954 DEBUG("%p: freeing WRs", (void *)rxq);
1956 rxq->elts.no_sp = NULL;
1959 for (i = 0; (i != elemof(*elts)); ++i) {
1960 struct rxq_elt *elt = &(*elts)[i];
1961 struct rte_mbuf *buf;
1963 if (elt->sge.addr == 0)
1965 assert(WR_ID(elt->wr.wr_id).id == i);
1966 buf = (void *)((uintptr_t)elt->sge.addr -
1967 WR_ID(elt->wr.wr_id).offset);
1968 rte_pktmbuf_free_seg(buf);
1974 * Delete flow steering rule.
1977 * Pointer to RX queue structure.
1979 * MAC address index.
1984 rxq_del_flow(struct rxq *rxq, unsigned int mac_index, unsigned int vlan_index)
1987 struct priv *priv = rxq->priv;
1988 const uint8_t (*mac)[ETHER_ADDR_LEN] =
1989 (const uint8_t (*)[ETHER_ADDR_LEN])
1990 priv->mac[mac_index].addr_bytes;
1992 assert(rxq->mac_flow[mac_index][vlan_index] != NULL);
1993 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
1994 " (VLAN ID %" PRIu16 ")",
1996 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
1997 mac_index, priv->vlan_filter[vlan_index].id);
1998 claim_zero(ibv_destroy_flow(rxq->mac_flow[mac_index][vlan_index]));
1999 rxq->mac_flow[mac_index][vlan_index] = NULL;
2003 * Unregister a MAC address from a RX queue.
2006 * Pointer to RX queue structure.
2008 * MAC address index.
2011 rxq_mac_addr_del(struct rxq *rxq, unsigned int mac_index)
2013 struct priv *priv = rxq->priv;
2015 unsigned int vlans = 0;
2017 assert(mac_index < elemof(priv->mac));
2018 if (!BITFIELD_ISSET(rxq->mac_configured, mac_index))
2020 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
2021 if (!priv->vlan_filter[i].enabled)
2023 rxq_del_flow(rxq, mac_index, i);
2027 rxq_del_flow(rxq, mac_index, 0);
2029 BITFIELD_RESET(rxq->mac_configured, mac_index);
2033 * Unregister all MAC addresses from a RX queue.
2036 * Pointer to RX queue structure.
2039 rxq_mac_addrs_del(struct rxq *rxq)
2041 struct priv *priv = rxq->priv;
2044 for (i = 0; (i != elemof(priv->mac)); ++i)
2045 rxq_mac_addr_del(rxq, i);
2048 static int rxq_promiscuous_enable(struct rxq *);
2049 static void rxq_promiscuous_disable(struct rxq *);
2052 * Add single flow steering rule.
2055 * Pointer to RX queue structure.
2057 * MAC address index to register.
2059 * VLAN index. Use -1 for a flow without VLAN.
2062 * 0 on success, errno value on failure.
2065 rxq_add_flow(struct rxq *rxq, unsigned int mac_index, unsigned int vlan_index)
2067 struct ibv_flow *flow;
2068 struct priv *priv = rxq->priv;
2069 const uint8_t (*mac)[ETHER_ADDR_LEN] =
2070 (const uint8_t (*)[ETHER_ADDR_LEN])
2071 priv->mac[mac_index].addr_bytes;
2073 /* Allocate flow specification on the stack. */
2074 struct __attribute__((packed)) {
2075 struct ibv_flow_attr attr;
2076 struct ibv_flow_spec_eth spec;
2078 struct ibv_flow_attr *attr = &data.attr;
2079 struct ibv_flow_spec_eth *spec = &data.spec;
2081 assert(mac_index < elemof(priv->mac));
2082 assert((vlan_index < elemof(priv->vlan_filter)) || (vlan_index == -1u));
2084 * No padding must be inserted by the compiler between attr and spec.
2085 * This layout is expected by libibverbs.
2087 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
2088 *attr = (struct ibv_flow_attr){
2089 .type = IBV_FLOW_ATTR_NORMAL,
2094 *spec = (struct ibv_flow_spec_eth){
2095 .type = IBV_FLOW_SPEC_ETH,
2096 .size = sizeof(*spec),
2099 (*mac)[0], (*mac)[1], (*mac)[2],
2100 (*mac)[3], (*mac)[4], (*mac)[5]
2102 .vlan_tag = ((vlan_index != -1u) ?
2103 htons(priv->vlan_filter[vlan_index].id) :
2107 .dst_mac = "\xff\xff\xff\xff\xff\xff",
2108 .vlan_tag = ((vlan_index != -1u) ? htons(0xfff) : 0),
2111 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
2112 " (VLAN %s %" PRIu16 ")",
2114 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
2116 ((vlan_index != -1u) ? "ID" : "index"),
2117 ((vlan_index != -1u) ? priv->vlan_filter[vlan_index].id : -1u));
2118 /* Create related flow. */
2120 flow = ibv_create_flow(rxq->qp, attr);
2122 /* It's not clear whether errno is always set in this case. */
2123 ERROR("%p: flow configuration failed, errno=%d: %s",
2125 (errno ? strerror(errno) : "Unknown error"));
2130 if (vlan_index == -1u)
2132 assert(rxq->mac_flow[mac_index][vlan_index] == NULL);
2133 rxq->mac_flow[mac_index][vlan_index] = flow;
2138 * Register a MAC address in a RX queue.
2141 * Pointer to RX queue structure.
2143 * MAC address index to register.
2146 * 0 on success, errno value on failure.
2149 rxq_mac_addr_add(struct rxq *rxq, unsigned int mac_index)
2151 struct priv *priv = rxq->priv;
2153 unsigned int vlans = 0;
2156 assert(mac_index < elemof(priv->mac));
2157 if (BITFIELD_ISSET(rxq->mac_configured, mac_index))
2158 rxq_mac_addr_del(rxq, mac_index);
2159 /* Fill VLAN specifications. */
2160 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
2161 if (!priv->vlan_filter[i].enabled)
2163 /* Create related flow. */
2164 ret = rxq_add_flow(rxq, mac_index, i);
2169 /* Failure, rollback. */
2171 if (priv->vlan_filter[--i].enabled)
2172 rxq_del_flow(rxq, mac_index, i);
2176 /* In case there is no VLAN filter. */
2178 ret = rxq_add_flow(rxq, mac_index, -1);
2182 BITFIELD_SET(rxq->mac_configured, mac_index);
2187 * Register all MAC addresses in a RX queue.
2190 * Pointer to RX queue structure.
2193 * 0 on success, errno value on failure.
2196 rxq_mac_addrs_add(struct rxq *rxq)
2198 struct priv *priv = rxq->priv;
2202 for (i = 0; (i != elemof(priv->mac)); ++i) {
2203 if (!BITFIELD_ISSET(priv->mac_configured, i))
2205 ret = rxq_mac_addr_add(rxq, i);
2208 /* Failure, rollback. */
2210 rxq_mac_addr_del(rxq, --i);
2218 * Unregister a MAC address.
2220 * In RSS mode, the MAC address is unregistered from the parent queue,
2221 * otherwise it is unregistered from each queue directly.
2224 * Pointer to private structure.
2226 * MAC address index.
2229 priv_mac_addr_del(struct priv *priv, unsigned int mac_index)
2233 assert(mac_index < elemof(priv->mac));
2234 if (!BITFIELD_ISSET(priv->mac_configured, mac_index))
2237 rxq_mac_addr_del(&priv->rxq_parent, mac_index);
2240 for (i = 0; (i != priv->dev->data->nb_rx_queues); ++i)
2241 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2243 BITFIELD_RESET(priv->mac_configured, mac_index);
2247 * Register a MAC address.
2249 * In RSS mode, the MAC address is registered in the parent queue,
2250 * otherwise it is registered in each queue directly.
2253 * Pointer to private structure.
2255 * MAC address index to use.
2257 * MAC address to register.
2260 * 0 on success, errno value on failure.
2263 priv_mac_addr_add(struct priv *priv, unsigned int mac_index,
2264 const uint8_t (*mac)[ETHER_ADDR_LEN])
2269 assert(mac_index < elemof(priv->mac));
2270 /* First, make sure this address isn't already configured. */
2271 for (i = 0; (i != elemof(priv->mac)); ++i) {
2272 /* Skip this index, it's going to be reconfigured. */
2275 if (!BITFIELD_ISSET(priv->mac_configured, i))
2277 if (memcmp(priv->mac[i].addr_bytes, *mac, sizeof(*mac)))
2279 /* Address already configured elsewhere, return with error. */
2282 if (BITFIELD_ISSET(priv->mac_configured, mac_index))
2283 priv_mac_addr_del(priv, mac_index);
2284 priv->mac[mac_index] = (struct ether_addr){
2286 (*mac)[0], (*mac)[1], (*mac)[2],
2287 (*mac)[3], (*mac)[4], (*mac)[5]
2290 /* If device isn't started, this is all we need to do. */
2291 if (!priv->started) {
2293 /* Verify that all queues have this index disabled. */
2294 for (i = 0; (i != priv->rxqs_n); ++i) {
2295 if ((*priv->rxqs)[i] == NULL)
2297 assert(!BITFIELD_ISSET
2298 ((*priv->rxqs)[i]->mac_configured, mac_index));
2304 ret = rxq_mac_addr_add(&priv->rxq_parent, mac_index);
2309 for (i = 0; (i != priv->rxqs_n); ++i) {
2310 if ((*priv->rxqs)[i] == NULL)
2312 ret = rxq_mac_addr_add((*priv->rxqs)[i], mac_index);
2315 /* Failure, rollback. */
2317 if ((*priv->rxqs)[(--i)] != NULL)
2318 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2322 BITFIELD_SET(priv->mac_configured, mac_index);
2327 * Enable allmulti mode in a RX queue.
2330 * Pointer to RX queue structure.
2333 * 0 on success, errno value on failure.
2336 rxq_allmulticast_enable(struct rxq *rxq)
2338 struct ibv_flow *flow;
2339 struct ibv_flow_attr attr = {
2340 .type = IBV_FLOW_ATTR_MC_DEFAULT,
2342 .port = rxq->priv->port,
2346 DEBUG("%p: enabling allmulticast mode", (void *)rxq);
2347 if (rxq->allmulti_flow != NULL)
2350 flow = ibv_create_flow(rxq->qp, &attr);
2352 /* It's not clear whether errno is always set in this case. */
2353 ERROR("%p: flow configuration failed, errno=%d: %s",
2355 (errno ? strerror(errno) : "Unknown error"));
2360 rxq->allmulti_flow = flow;
2361 DEBUG("%p: allmulticast mode enabled", (void *)rxq);
2366 * Disable allmulti mode in a RX queue.
2369 * Pointer to RX queue structure.
2372 rxq_allmulticast_disable(struct rxq *rxq)
2374 DEBUG("%p: disabling allmulticast mode", (void *)rxq);
2375 if (rxq->allmulti_flow == NULL)
2377 claim_zero(ibv_destroy_flow(rxq->allmulti_flow));
2378 rxq->allmulti_flow = NULL;
2379 DEBUG("%p: allmulticast mode disabled", (void *)rxq);
2383 * Enable promiscuous mode in a RX queue.
2386 * Pointer to RX queue structure.
2389 * 0 on success, errno value on failure.
2392 rxq_promiscuous_enable(struct rxq *rxq)
2394 struct ibv_flow *flow;
2395 struct ibv_flow_attr attr = {
2396 .type = IBV_FLOW_ATTR_ALL_DEFAULT,
2398 .port = rxq->priv->port,
2404 DEBUG("%p: enabling promiscuous mode", (void *)rxq);
2405 if (rxq->promisc_flow != NULL)
2408 flow = ibv_create_flow(rxq->qp, &attr);
2410 /* It's not clear whether errno is always set in this case. */
2411 ERROR("%p: flow configuration failed, errno=%d: %s",
2413 (errno ? strerror(errno) : "Unknown error"));
2418 rxq->promisc_flow = flow;
2419 DEBUG("%p: promiscuous mode enabled", (void *)rxq);
2424 * Disable promiscuous mode in a RX queue.
2427 * Pointer to RX queue structure.
2430 rxq_promiscuous_disable(struct rxq *rxq)
2434 DEBUG("%p: disabling promiscuous mode", (void *)rxq);
2435 if (rxq->promisc_flow == NULL)
2437 claim_zero(ibv_destroy_flow(rxq->promisc_flow));
2438 rxq->promisc_flow = NULL;
2439 DEBUG("%p: promiscuous mode disabled", (void *)rxq);
2443 * Clean up a RX queue.
2445 * Destroy objects, free allocated memory and reset the structure for reuse.
2448 * Pointer to RX queue structure.
2451 rxq_cleanup(struct rxq *rxq)
2453 struct ibv_exp_release_intf_params params;
2455 DEBUG("cleaning up %p", (void *)rxq);
2457 rxq_free_elts_sp(rxq);
2460 if (rxq->if_qp != NULL) {
2461 assert(rxq->priv != NULL);
2462 assert(rxq->priv->ctx != NULL);
2463 assert(rxq->qp != NULL);
2464 params = (struct ibv_exp_release_intf_params){
2467 claim_zero(ibv_exp_release_intf(rxq->priv->ctx,
2471 if (rxq->if_cq != NULL) {
2472 assert(rxq->priv != NULL);
2473 assert(rxq->priv->ctx != NULL);
2474 assert(rxq->cq != NULL);
2475 params = (struct ibv_exp_release_intf_params){
2478 claim_zero(ibv_exp_release_intf(rxq->priv->ctx,
2482 if (rxq->qp != NULL) {
2483 rxq_promiscuous_disable(rxq);
2484 rxq_allmulticast_disable(rxq);
2485 rxq_mac_addrs_del(rxq);
2486 claim_zero(ibv_destroy_qp(rxq->qp));
2488 if (rxq->cq != NULL)
2489 claim_zero(ibv_destroy_cq(rxq->cq));
2490 if (rxq->rd != NULL) {
2491 struct ibv_exp_destroy_res_domain_attr attr = {
2495 assert(rxq->priv != NULL);
2496 assert(rxq->priv->ctx != NULL);
2497 claim_zero(ibv_exp_destroy_res_domain(rxq->priv->ctx,
2501 if (rxq->mr != NULL)
2502 claim_zero(ibv_dereg_mr(rxq->mr));
2503 memset(rxq, 0, sizeof(*rxq));
2507 * Translate RX completion flags to packet type.
2510 * RX completion flags returned by poll_length_flags().
2513 * Packet type for struct rte_mbuf.
2515 static inline uint32_t
2516 rxq_cq_to_pkt_type(uint32_t flags)
2520 if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
2523 IBV_EXP_CQ_RX_OUTER_IPV4_PACKET, RTE_PTYPE_L3_IPV4) |
2525 IBV_EXP_CQ_RX_OUTER_IPV6_PACKET, RTE_PTYPE_L3_IPV6) |
2527 IBV_EXP_CQ_RX_IPV4_PACKET, RTE_PTYPE_INNER_L3_IPV4) |
2529 IBV_EXP_CQ_RX_IPV6_PACKET, RTE_PTYPE_INNER_L3_IPV6);
2533 IBV_EXP_CQ_RX_IPV4_PACKET, RTE_PTYPE_L3_IPV4) |
2535 IBV_EXP_CQ_RX_IPV6_PACKET, RTE_PTYPE_L3_IPV6);
2540 * Translate RX completion flags to offload flags.
2543 * Pointer to RX queue structure.
2545 * RX completion flags returned by poll_length_flags().
2548 * Offload flags (ol_flags) for struct rte_mbuf.
2550 static inline uint32_t
2551 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
2553 uint32_t ol_flags = 0;
2558 IBV_EXP_CQ_RX_IP_CSUM_OK,
2559 PKT_RX_IP_CKSUM_BAD) |
2561 IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
2562 PKT_RX_L4_CKSUM_BAD);
2564 * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
2565 * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
2568 if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
2571 IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
2572 PKT_RX_IP_CKSUM_BAD) |
2574 IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
2575 PKT_RX_L4_CKSUM_BAD);
2580 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
2583 * DPDK callback for RX with scattered packets support.
2586 * Generic pointer to RX queue structure.
2588 * Array to store received packets.
2590 * Maximum number of packets in array.
2593 * Number of packets successfully received (<= pkts_n).
2596 mlx4_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2598 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2599 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
2600 const unsigned int elts_n = rxq->elts_n;
2601 unsigned int elts_head = rxq->elts_head;
2602 struct ibv_recv_wr head;
2603 struct ibv_recv_wr **next = &head.next;
2604 struct ibv_recv_wr *bad_wr;
2606 unsigned int pkts_ret = 0;
2609 if (unlikely(!rxq->sp))
2610 return mlx4_rx_burst(dpdk_rxq, pkts, pkts_n);
2611 if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
2613 for (i = 0; (i != pkts_n); ++i) {
2614 struct rxq_elt_sp *elt = &(*elts)[elts_head];
2615 struct ibv_recv_wr *wr = &elt->wr;
2616 uint64_t wr_id = wr->wr_id;
2618 unsigned int pkt_buf_len;
2619 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
2620 struct rte_mbuf **pkt_buf_next = &pkt_buf;
2621 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
2625 /* Sanity checks. */
2629 assert(wr_id < rxq->elts_n);
2630 assert(wr->sg_list == elt->sges);
2631 assert(wr->num_sge == elemof(elt->sges));
2632 assert(elts_head < rxq->elts_n);
2633 assert(rxq->elts_head < rxq->elts_n);
2634 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
2636 if (unlikely(ret < 0)) {
2640 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
2642 /* ibv_poll_cq() must be used in case of failure. */
2643 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
2644 if (unlikely(wcs_n == 0))
2646 if (unlikely(wcs_n < 0)) {
2647 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
2648 (void *)rxq, wcs_n);
2652 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
2653 /* Whatever, just repost the offending WR. */
2654 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
2655 " completion status (%d): %s",
2656 (void *)rxq, wc.wr_id, wc.status,
2657 ibv_wc_status_str(wc.status));
2658 #ifdef MLX4_PMD_SOFT_COUNTERS
2659 /* Increment dropped packets counter. */
2660 ++rxq->stats.idropped;
2662 /* Link completed WRs together for repost. */
2673 /* Link completed WRs together for repost. */
2677 * Replace spent segments with new ones, concatenate and
2678 * return them as pkt_buf.
2681 struct ibv_sge *sge = &elt->sges[j];
2682 struct rte_mbuf *seg = elt->bufs[j];
2683 struct rte_mbuf *rep;
2684 unsigned int seg_tailroom;
2687 * Fetch initial bytes of packet descriptor into a
2688 * cacheline while allocating rep.
2691 rep = __rte_mbuf_raw_alloc(rxq->mp);
2692 if (unlikely(rep == NULL)) {
2694 * Unable to allocate a replacement mbuf,
2697 DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
2698 " can't allocate a new mbuf",
2699 (void *)rxq, wr_id);
2700 if (pkt_buf != NULL) {
2701 *pkt_buf_next = NULL;
2702 rte_pktmbuf_free(pkt_buf);
2704 /* Increase out of memory counters. */
2705 ++rxq->stats.rx_nombuf;
2706 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2710 /* Poison user-modifiable fields in rep. */
2711 NEXT(rep) = (void *)((uintptr_t)-1);
2712 SET_DATA_OFF(rep, 0xdead);
2713 DATA_LEN(rep) = 0xd00d;
2714 PKT_LEN(rep) = 0xdeadd00d;
2715 NB_SEGS(rep) = 0x2a;
2719 assert(rep->buf_len == seg->buf_len);
2720 assert(rep->buf_len == rxq->mb_len);
2721 /* Reconfigure sge to use rep instead of seg. */
2722 assert(sge->lkey == rxq->mr->lkey);
2723 sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
2726 /* Update pkt_buf if it's the first segment, or link
2727 * seg to the previous one and update pkt_buf_next. */
2728 *pkt_buf_next = seg;
2729 pkt_buf_next = &NEXT(seg);
2730 /* Update seg information. */
2731 seg_tailroom = (seg->buf_len - seg_headroom);
2732 assert(sge->length == seg_tailroom);
2733 SET_DATA_OFF(seg, seg_headroom);
2734 if (likely(len <= seg_tailroom)) {
2736 DATA_LEN(seg) = len;
2739 assert(rte_pktmbuf_headroom(seg) ==
2741 assert(rte_pktmbuf_tailroom(seg) ==
2742 (seg_tailroom - len));
2745 DATA_LEN(seg) = seg_tailroom;
2746 PKT_LEN(seg) = seg_tailroom;
2748 assert(rte_pktmbuf_headroom(seg) == seg_headroom);
2749 assert(rte_pktmbuf_tailroom(seg) == 0);
2750 /* Fix len and clear headroom for next segments. */
2751 len -= seg_tailroom;
2754 /* Update head and tail segments. */
2755 *pkt_buf_next = NULL;
2756 assert(pkt_buf != NULL);
2758 NB_SEGS(pkt_buf) = j;
2759 PORT(pkt_buf) = rxq->port_id;
2760 PKT_LEN(pkt_buf) = pkt_buf_len;
2761 pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
2762 pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
2764 /* Return packet. */
2765 *(pkts++) = pkt_buf;
2767 #ifdef MLX4_PMD_SOFT_COUNTERS
2768 /* Increase bytes counter. */
2769 rxq->stats.ibytes += pkt_buf_len;
2772 if (++elts_head >= elts_n)
2776 if (unlikely(i == 0))
2781 DEBUG("%p: reposting %d WRs", (void *)rxq, i);
2783 ret = ibv_post_recv(rxq->qp, head.next, &bad_wr);
2784 if (unlikely(ret)) {
2785 /* Inability to repost WRs is fatal. */
2786 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
2792 rxq->elts_head = elts_head;
2793 #ifdef MLX4_PMD_SOFT_COUNTERS
2794 /* Increase packets counter. */
2795 rxq->stats.ipackets += pkts_ret;
2801 * DPDK callback for RX.
2803 * The following function is the same as mlx4_rx_burst_sp(), except it doesn't
2804 * manage scattered packets. Improves performance when MRU is lower than the
2805 * size of the first segment.
2808 * Generic pointer to RX queue structure.
2810 * Array to store received packets.
2812 * Maximum number of packets in array.
2815 * Number of packets successfully received (<= pkts_n).
2818 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2820 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2821 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
2822 const unsigned int elts_n = rxq->elts_n;
2823 unsigned int elts_head = rxq->elts_head;
2824 struct ibv_sge sges[pkts_n];
2826 unsigned int pkts_ret = 0;
2829 if (unlikely(rxq->sp))
2830 return mlx4_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
2831 for (i = 0; (i != pkts_n); ++i) {
2832 struct rxq_elt *elt = &(*elts)[elts_head];
2833 struct ibv_recv_wr *wr = &elt->wr;
2834 uint64_t wr_id = wr->wr_id;
2836 struct rte_mbuf *seg = (void *)((uintptr_t)elt->sge.addr -
2837 WR_ID(wr_id).offset);
2838 struct rte_mbuf *rep;
2841 /* Sanity checks. */
2842 assert(WR_ID(wr_id).id < rxq->elts_n);
2843 assert(wr->sg_list == &elt->sge);
2844 assert(wr->num_sge == 1);
2845 assert(elts_head < rxq->elts_n);
2846 assert(rxq->elts_head < rxq->elts_n);
2848 * Fetch initial bytes of packet descriptor into a
2849 * cacheline while allocating rep.
2852 rte_prefetch0(&seg->cacheline1);
2853 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
2855 if (unlikely(ret < 0)) {
2859 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
2861 /* ibv_poll_cq() must be used in case of failure. */
2862 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
2863 if (unlikely(wcs_n == 0))
2865 if (unlikely(wcs_n < 0)) {
2866 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
2867 (void *)rxq, wcs_n);
2871 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
2872 /* Whatever, just repost the offending WR. */
2873 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
2874 " completion status (%d): %s",
2875 (void *)rxq, wc.wr_id, wc.status,
2876 ibv_wc_status_str(wc.status));
2877 #ifdef MLX4_PMD_SOFT_COUNTERS
2878 /* Increment dropped packets counter. */
2879 ++rxq->stats.idropped;
2881 /* Add SGE to array for repost. */
2890 rep = __rte_mbuf_raw_alloc(rxq->mp);
2891 if (unlikely(rep == NULL)) {
2893 * Unable to allocate a replacement mbuf,
2896 DEBUG("rxq=%p, wr_id=%" PRIu32 ":"
2897 " can't allocate a new mbuf",
2898 (void *)rxq, WR_ID(wr_id).id);
2899 /* Increase out of memory counters. */
2900 ++rxq->stats.rx_nombuf;
2901 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2905 /* Reconfigure sge to use rep instead of seg. */
2906 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
2907 assert(elt->sge.lkey == rxq->mr->lkey);
2908 WR_ID(wr->wr_id).offset =
2909 (((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
2911 assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
2913 /* Add SGE to array for repost. */
2916 /* Update seg information. */
2917 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
2919 PORT(seg) = rxq->port_id;
2922 DATA_LEN(seg) = len;
2923 seg->packet_type = rxq_cq_to_pkt_type(flags);
2924 seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
2926 /* Return packet. */
2929 #ifdef MLX4_PMD_SOFT_COUNTERS
2930 /* Increase bytes counter. */
2931 rxq->stats.ibytes += len;
2934 if (++elts_head >= elts_n)
2938 if (unlikely(i == 0))
2942 DEBUG("%p: reposting %u WRs", (void *)rxq, i);
2944 ret = rxq->if_qp->recv_burst(rxq->qp, sges, i);
2945 if (unlikely(ret)) {
2946 /* Inability to repost WRs is fatal. */
2947 DEBUG("%p: recv_burst(): failed (ret=%d)",
2952 rxq->elts_head = elts_head;
2953 #ifdef MLX4_PMD_SOFT_COUNTERS
2954 /* Increase packets counter. */
2955 rxq->stats.ipackets += pkts_ret;
2961 * Allocate a Queue Pair.
2962 * Optionally setup inline receive if supported.
2965 * Pointer to private structure.
2967 * Completion queue to associate with QP.
2969 * Number of descriptors in QP (hint only).
2972 * QP pointer or NULL in case of error.
2974 static struct ibv_qp *
2975 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
2976 struct ibv_exp_res_domain *rd)
2978 struct ibv_exp_qp_init_attr attr = {
2979 /* CQ to be associated with the send queue. */
2981 /* CQ to be associated with the receive queue. */
2984 /* Max number of outstanding WRs. */
2985 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2986 priv->device_attr.max_qp_wr :
2988 /* Max number of scatter/gather elements in a WR. */
2989 .max_recv_sge = ((priv->device_attr.max_sge <
2990 MLX4_PMD_SGE_WR_N) ?
2991 priv->device_attr.max_sge :
2994 .qp_type = IBV_QPT_RAW_PACKET,
2995 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
2996 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
3002 attr.max_inl_recv = priv->inl_recv_size;
3003 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
3005 return ibv_exp_create_qp(priv->ctx, &attr);
3011 * Allocate a RSS Queue Pair.
3012 * Optionally setup inline receive if supported.
3015 * Pointer to private structure.
3017 * Completion queue to associate with QP.
3019 * Number of descriptors in QP (hint only).
3021 * If nonzero, create a parent QP, otherwise a child.
3024 * QP pointer or NULL in case of error.
3026 static struct ibv_qp *
3027 rxq_setup_qp_rss(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
3028 int parent, struct ibv_exp_res_domain *rd)
3030 struct ibv_exp_qp_init_attr attr = {
3031 /* CQ to be associated with the send queue. */
3033 /* CQ to be associated with the receive queue. */
3036 /* Max number of outstanding WRs. */
3037 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
3038 priv->device_attr.max_qp_wr :
3040 /* Max number of scatter/gather elements in a WR. */
3041 .max_recv_sge = ((priv->device_attr.max_sge <
3042 MLX4_PMD_SGE_WR_N) ?
3043 priv->device_attr.max_sge :
3046 .qp_type = IBV_QPT_RAW_PACKET,
3047 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
3048 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN |
3049 IBV_EXP_QP_INIT_ATTR_QPG),
3055 attr.max_inl_recv = priv->inl_recv_size,
3056 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
3059 attr.qpg.qpg_type = IBV_EXP_QPG_PARENT;
3060 /* TSS isn't necessary. */
3061 attr.qpg.parent_attrib.tss_child_count = 0;
3062 attr.qpg.parent_attrib.rss_child_count = priv->rxqs_n;
3063 DEBUG("initializing parent RSS queue");
3065 attr.qpg.qpg_type = IBV_EXP_QPG_CHILD_RX;
3066 attr.qpg.qpg_parent = priv->rxq_parent.qp;
3067 DEBUG("initializing child RSS queue");
3069 return ibv_exp_create_qp(priv->ctx, &attr);
3072 #endif /* RSS_SUPPORT */
3075 * Reconfigure a RX queue with new parameters.
3077 * rxq_rehash() does not allocate mbufs, which, if not done from the right
3078 * thread (such as a control thread), may corrupt the pool.
3079 * In case of failure, the queue is left untouched.
3082 * Pointer to Ethernet device structure.
3087 * 0 on success, errno value on failure.
3090 rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
3092 struct priv *priv = rxq->priv;
3093 struct rxq tmpl = *rxq;
3094 unsigned int mbuf_n;
3095 unsigned int desc_n;
3096 struct rte_mbuf **pool;
3098 struct ibv_exp_qp_attr mod;
3099 struct ibv_recv_wr *bad_wr;
3101 int parent = (rxq == &priv->rxq_parent);
3104 ERROR("%p: cannot rehash parent queue %p",
3105 (void *)dev, (void *)rxq);
3108 DEBUG("%p: rehashing queue %p", (void *)dev, (void *)rxq);
3109 /* Number of descriptors and mbufs currently allocated. */
3110 desc_n = (tmpl.elts_n * (tmpl.sp ? MLX4_PMD_SGE_WR_N : 1));
3112 /* Toggle RX checksum offload if hardware supports it. */
3113 if (priv->hw_csum) {
3114 tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3115 rxq->csum = tmpl.csum;
3117 if (priv->hw_csum_l2tun) {
3118 tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3119 rxq->csum_l2tun = tmpl.csum_l2tun;
3121 /* Enable scattered packets support for this queue if necessary. */
3122 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3123 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3124 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3126 desc_n /= MLX4_PMD_SGE_WR_N;
3129 DEBUG("%p: %s scattered packets support (%u WRs)",
3130 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc_n);
3131 /* If scatter mode is the same as before, nothing to do. */
3132 if (tmpl.sp == rxq->sp) {
3133 DEBUG("%p: nothing to do", (void *)dev);
3136 /* Remove attached flows if RSS is disabled (no parent queue). */
3138 rxq_allmulticast_disable(&tmpl);
3139 rxq_promiscuous_disable(&tmpl);
3140 rxq_mac_addrs_del(&tmpl);
3141 /* Update original queue in case of failure. */
3142 rxq->allmulti_flow = tmpl.allmulti_flow;
3143 rxq->promisc_flow = tmpl.promisc_flow;
3144 memcpy(rxq->mac_configured, tmpl.mac_configured,
3145 sizeof(rxq->mac_configured));
3146 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
3148 /* From now on, any failure will render the queue unusable.
3149 * Reinitialize QP. */
3150 mod = (struct ibv_exp_qp_attr){ .qp_state = IBV_QPS_RESET };
3151 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3153 ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
3157 err = ibv_resize_cq(tmpl.cq, desc_n);
3159 ERROR("%p: cannot resize CQ: %s", (void *)dev, strerror(err));
3163 mod = (struct ibv_exp_qp_attr){
3164 /* Move the QP to this state. */
3165 .qp_state = IBV_QPS_INIT,
3166 /* Primary port number. */
3167 .port_num = priv->port
3169 err = ibv_exp_modify_qp(tmpl.qp, &mod,
3172 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3173 #endif /* RSS_SUPPORT */
3176 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3177 (void *)dev, strerror(err));
3181 /* Reconfigure flows. Do not care for errors. */
3183 rxq_mac_addrs_add(&tmpl);
3185 rxq_promiscuous_enable(&tmpl);
3187 rxq_allmulticast_enable(&tmpl);
3188 /* Update original queue in case of failure. */
3189 rxq->allmulti_flow = tmpl.allmulti_flow;
3190 rxq->promisc_flow = tmpl.promisc_flow;
3191 memcpy(rxq->mac_configured, tmpl.mac_configured,
3192 sizeof(rxq->mac_configured));
3193 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
3195 /* Allocate pool. */
3196 pool = rte_malloc(__func__, (mbuf_n * sizeof(*pool)), 0);
3198 ERROR("%p: cannot allocate memory", (void *)dev);
3201 /* Snatch mbufs from original queue. */
3204 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
3206 for (i = 0; (i != elemof(*elts)); ++i) {
3207 struct rxq_elt_sp *elt = &(*elts)[i];
3210 for (j = 0; (j != elemof(elt->bufs)); ++j) {
3211 assert(elt->bufs[j] != NULL);
3212 pool[k++] = elt->bufs[j];
3216 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
3218 for (i = 0; (i != elemof(*elts)); ++i) {
3219 struct rxq_elt *elt = &(*elts)[i];
3220 struct rte_mbuf *buf = (void *)
3221 ((uintptr_t)elt->sge.addr -
3222 WR_ID(elt->wr.wr_id).offset);
3224 assert(WR_ID(elt->wr.wr_id).id == i);
3228 assert(k == mbuf_n);
3230 tmpl.elts.sp = NULL;
3231 assert((void *)&tmpl.elts.sp == (void *)&tmpl.elts.no_sp);
3233 rxq_alloc_elts_sp(&tmpl, desc_n, pool) :
3234 rxq_alloc_elts(&tmpl, desc_n, pool));
3236 ERROR("%p: cannot reallocate WRs, aborting", (void *)dev);
3241 assert(tmpl.elts_n == desc_n);
3242 assert(tmpl.elts.sp != NULL);
3244 /* Clean up original data. */
3246 rte_free(rxq->elts.sp);
3247 rxq->elts.sp = NULL;
3249 err = ibv_post_recv(tmpl.qp,
3251 &(*tmpl.elts.sp)[0].wr :
3252 &(*tmpl.elts.no_sp)[0].wr),
3255 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3261 mod = (struct ibv_exp_qp_attr){
3262 .qp_state = IBV_QPS_RTR
3264 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3266 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3267 (void *)dev, strerror(err));
3275 * Configure a RX queue.
3278 * Pointer to Ethernet device structure.
3280 * Pointer to RX queue structure.
3282 * Number of descriptors to configure in queue.
3284 * NUMA socket on which memory must be allocated.
3286 * Thresholds parameters.
3288 * Memory pool for buffer allocations.
3291 * 0 on success, errno value on failure.
3294 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
3295 unsigned int socket, const struct rte_eth_rxconf *conf,
3296 struct rte_mempool *mp)
3298 struct priv *priv = dev->data->dev_private;
3304 struct ibv_exp_qp_attr mod;
3306 struct ibv_exp_query_intf_params params;
3307 struct ibv_exp_cq_init_attr cq;
3308 struct ibv_exp_res_domain_init_attr rd;
3310 enum ibv_exp_query_intf_status status;
3311 struct ibv_recv_wr *bad_wr;
3312 struct rte_mbuf *buf;
3314 int parent = (rxq == &priv->rxq_parent);
3316 (void)conf; /* Thresholds configuration (ignored). */
3318 * If this is a parent queue, hardware must support RSS and
3319 * RSS must be enabled.
3321 assert((!parent) || ((priv->hw_rss) && (priv->rss)));
3323 /* Even if unused, ibv_create_cq() requires at least one
3328 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
3329 ERROR("%p: invalid number of RX descriptors (must be a"
3330 " multiple of %d)", (void *)dev, MLX4_PMD_SGE_WR_N);
3333 /* Get mbuf length. */
3334 buf = rte_pktmbuf_alloc(mp);
3336 ERROR("%p: unable to allocate mbuf", (void *)dev);
3339 tmpl.mb_len = buf->buf_len;
3340 assert((rte_pktmbuf_headroom(buf) +
3341 rte_pktmbuf_tailroom(buf)) == tmpl.mb_len);
3342 assert(rte_pktmbuf_headroom(buf) == RTE_PKTMBUF_HEADROOM);
3343 rte_pktmbuf_free(buf);
3344 /* Toggle RX checksum offload if hardware supports it. */
3346 tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3347 if (priv->hw_csum_l2tun)
3348 tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3349 /* Enable scattered packets support for this queue if necessary. */
3350 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3351 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3352 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3354 desc /= MLX4_PMD_SGE_WR_N;
3356 DEBUG("%p: %s scattered packets support (%u WRs)",
3357 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc);
3358 /* Use the entire RX mempool as the memory region. */
3359 tmpl.mr = ibv_reg_mr(priv->pd,
3360 (void *)mp->elt_va_start,
3361 (mp->elt_va_end - mp->elt_va_start),
3362 (IBV_ACCESS_LOCAL_WRITE |
3363 IBV_ACCESS_REMOTE_WRITE));
3364 if (tmpl.mr == NULL) {
3366 ERROR("%p: MR creation failure: %s",
3367 (void *)dev, strerror(ret));
3371 attr.rd = (struct ibv_exp_res_domain_init_attr){
3372 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
3373 IBV_EXP_RES_DOMAIN_MSG_MODEL),
3374 .thread_model = IBV_EXP_THREAD_SINGLE,
3375 .msg_model = IBV_EXP_MSG_HIGH_BW,
3377 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
3378 if (tmpl.rd == NULL) {
3380 ERROR("%p: RD creation failure: %s",
3381 (void *)dev, strerror(ret));
3384 attr.cq = (struct ibv_exp_cq_init_attr){
3385 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
3386 .res_domain = tmpl.rd,
3388 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
3389 if (tmpl.cq == NULL) {
3391 ERROR("%p: CQ creation failure: %s",
3392 (void *)dev, strerror(ret));
3395 DEBUG("priv->device_attr.max_qp_wr is %d",
3396 priv->device_attr.max_qp_wr);
3397 DEBUG("priv->device_attr.max_sge is %d",
3398 priv->device_attr.max_sge);
3401 tmpl.qp = rxq_setup_qp_rss(priv, tmpl.cq, desc, parent,
3404 #endif /* RSS_SUPPORT */
3405 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc, tmpl.rd);
3406 if (tmpl.qp == NULL) {
3407 ret = (errno ? errno : EINVAL);
3408 ERROR("%p: QP creation failure: %s",
3409 (void *)dev, strerror(ret));
3412 mod = (struct ibv_exp_qp_attr){
3413 /* Move the QP to this state. */
3414 .qp_state = IBV_QPS_INIT,
3415 /* Primary port number. */
3416 .port_num = priv->port
3418 ret = ibv_exp_modify_qp(tmpl.qp, &mod,
3421 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3422 #endif /* RSS_SUPPORT */
3425 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3426 (void *)dev, strerror(ret));
3429 if ((parent) || (!priv->rss)) {
3430 /* Configure MAC and broadcast addresses. */
3431 ret = rxq_mac_addrs_add(&tmpl);
3433 ERROR("%p: QP flow attachment failed: %s",
3434 (void *)dev, strerror(ret));
3438 /* Allocate descriptors for RX queues, except for the RSS parent. */
3442 ret = rxq_alloc_elts_sp(&tmpl, desc, NULL);
3444 ret = rxq_alloc_elts(&tmpl, desc, NULL);
3446 ERROR("%p: RXQ allocation failed: %s",
3447 (void *)dev, strerror(ret));
3450 ret = ibv_post_recv(tmpl.qp,
3452 &(*tmpl.elts.sp)[0].wr :
3453 &(*tmpl.elts.no_sp)[0].wr),
3456 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3463 mod = (struct ibv_exp_qp_attr){
3464 .qp_state = IBV_QPS_RTR
3466 ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3468 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3469 (void *)dev, strerror(ret));
3473 tmpl.port_id = dev->data->port_id;
3474 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
3475 attr.params = (struct ibv_exp_query_intf_params){
3476 .intf_scope = IBV_EXP_INTF_GLOBAL,
3477 .intf = IBV_EXP_INTF_CQ,
3480 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
3481 if (tmpl.if_cq == NULL) {
3482 ERROR("%p: CQ interface family query failed with status %d",
3483 (void *)dev, status);
3486 attr.params = (struct ibv_exp_query_intf_params){
3487 .intf_scope = IBV_EXP_INTF_GLOBAL,
3488 .intf = IBV_EXP_INTF_QP_BURST,
3491 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
3492 if (tmpl.if_qp == NULL) {
3493 ERROR("%p: QP interface family query failed with status %d",
3494 (void *)dev, status);
3497 /* Clean up rxq in case we're reinitializing it. */
3498 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
3501 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
3511 * DPDK callback to configure a RX queue.
3514 * Pointer to Ethernet device structure.
3518 * Number of descriptors to configure in queue.
3520 * NUMA socket on which memory must be allocated.
3522 * Thresholds parameters.
3524 * Memory pool for buffer allocations.
3527 * 0 on success, negative errno value on failure.
3530 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
3531 unsigned int socket, const struct rte_eth_rxconf *conf,
3532 struct rte_mempool *mp)
3534 struct priv *priv = dev->data->dev_private;
3535 struct rxq *rxq = (*priv->rxqs)[idx];
3539 DEBUG("%p: configuring queue %u for %u descriptors",
3540 (void *)dev, idx, desc);
3541 if (idx >= priv->rxqs_n) {
3542 ERROR("%p: queue index out of range (%u >= %u)",
3543 (void *)dev, idx, priv->rxqs_n);
3548 DEBUG("%p: reusing already allocated queue index %u (%p)",
3549 (void *)dev, idx, (void *)rxq);
3550 if (priv->started) {
3554 (*priv->rxqs)[idx] = NULL;
3557 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
3559 ERROR("%p: unable to allocate queue index %u",
3565 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
3569 rxq->stats.idx = idx;
3570 DEBUG("%p: adding RX queue %p to list",
3571 (void *)dev, (void *)rxq);
3572 (*priv->rxqs)[idx] = rxq;
3573 /* Update receive callback. */
3575 dev->rx_pkt_burst = mlx4_rx_burst_sp;
3577 dev->rx_pkt_burst = mlx4_rx_burst;
3584 * DPDK callback to release a RX queue.
3587 * Generic RX queue pointer.
3590 mlx4_rx_queue_release(void *dpdk_rxq)
3592 struct rxq *rxq = (struct rxq *)dpdk_rxq;
3600 assert(rxq != &priv->rxq_parent);
3601 for (i = 0; (i != priv->rxqs_n); ++i)
3602 if ((*priv->rxqs)[i] == rxq) {
3603 DEBUG("%p: removing RX queue %p from list",
3604 (void *)priv->dev, (void *)rxq);
3605 (*priv->rxqs)[i] = NULL;
3614 priv_dev_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
3617 * DPDK callback to start the device.
3619 * Simulate device start by attaching all configured flows.
3622 * Pointer to Ethernet device structure.
3625 * 0 on success, negative errno value on failure.
3628 mlx4_dev_start(struct rte_eth_dev *dev)
3630 struct priv *priv = dev->data->dev_private;
3636 if (priv->started) {
3640 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
3643 rxq = &priv->rxq_parent;
3646 rxq = (*priv->rxqs)[0];
3649 /* Iterate only once when RSS is enabled. */
3653 /* Ignore nonexistent RX queues. */
3656 ret = rxq_mac_addrs_add(rxq);
3657 if (!ret && priv->promisc)
3658 ret = rxq_promiscuous_enable(rxq);
3659 if (!ret && priv->allmulti)
3660 ret = rxq_allmulticast_enable(rxq);
3663 WARN("%p: QP flow attachment failed: %s",
3664 (void *)dev, strerror(ret));
3667 rxq = (*priv->rxqs)[--i];
3669 rxq_allmulticast_disable(rxq);
3670 rxq_promiscuous_disable(rxq);
3671 rxq_mac_addrs_del(rxq);
3677 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3678 priv_dev_interrupt_handler_install(priv, dev);
3684 * DPDK callback to stop the device.
3686 * Simulate device stop by detaching all configured flows.
3689 * Pointer to Ethernet device structure.
3692 mlx4_dev_stop(struct rte_eth_dev *dev)
3694 struct priv *priv = dev->data->dev_private;
3700 if (!priv->started) {
3704 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
3707 rxq = &priv->rxq_parent;
3710 rxq = (*priv->rxqs)[0];
3713 /* Iterate only once when RSS is enabled. */
3715 /* Ignore nonexistent RX queues. */
3718 rxq_allmulticast_disable(rxq);
3719 rxq_promiscuous_disable(rxq);
3720 rxq_mac_addrs_del(rxq);
3721 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3726 * Dummy DPDK callback for TX.
3728 * This function is used to temporarily replace the real callback during
3729 * unsafe control operations on the queue, or in case of error.
3732 * Generic pointer to TX queue structure.
3734 * Packets to transmit.
3736 * Number of packets in array.
3739 * Number of packets successfully transmitted (<= pkts_n).
3742 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
3751 * Dummy DPDK callback for RX.
3753 * This function is used to temporarily replace the real callback during
3754 * unsafe control operations on the queue, or in case of error.
3757 * Generic pointer to RX queue structure.
3759 * Array to store received packets.
3761 * Maximum number of packets in array.
3764 * Number of packets successfully received (<= pkts_n).
3767 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
3776 priv_dev_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
3779 * DPDK callback to close the device.
3781 * Destroy all queues and objects, free memory.
3784 * Pointer to Ethernet device structure.
3787 mlx4_dev_close(struct rte_eth_dev *dev)
3789 struct priv *priv = dev->data->dev_private;
3794 DEBUG("%p: closing device \"%s\"",
3796 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
3797 /* Prevent crashes when queues are still in use. This is unfortunately
3798 * still required for DPDK 1.3 because some programs (such as testpmd)
3799 * never release them before closing the device. */
3800 dev->rx_pkt_burst = removed_rx_burst;
3801 dev->tx_pkt_burst = removed_tx_burst;
3802 if (priv->rxqs != NULL) {
3803 /* XXX race condition if mlx4_rx_burst() is still running. */
3805 for (i = 0; (i != priv->rxqs_n); ++i) {
3806 tmp = (*priv->rxqs)[i];
3809 (*priv->rxqs)[i] = NULL;
3816 if (priv->txqs != NULL) {
3817 /* XXX race condition if mlx4_tx_burst() is still running. */
3819 for (i = 0; (i != priv->txqs_n); ++i) {
3820 tmp = (*priv->txqs)[i];
3823 (*priv->txqs)[i] = NULL;
3831 rxq_cleanup(&priv->rxq_parent);
3832 if (priv->pd != NULL) {
3833 assert(priv->ctx != NULL);
3834 claim_zero(ibv_dealloc_pd(priv->pd));
3835 claim_zero(ibv_close_device(priv->ctx));
3837 assert(priv->ctx == NULL);
3838 priv_dev_interrupt_handler_uninstall(priv, dev);
3840 memset(priv, 0, sizeof(*priv));
3844 * DPDK callback to get information about the device.
3847 * Pointer to Ethernet device structure.
3849 * Info structure output buffer.
3852 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
3854 struct priv *priv = dev->data->dev_private;
3856 char ifname[IF_NAMESIZE];
3859 /* FIXME: we should ask the device for these values. */
3860 info->min_rx_bufsize = 32;
3861 info->max_rx_pktlen = 65536;
3863 * Since we need one CQ per QP, the limit is the minimum number
3864 * between the two values.
3866 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
3867 priv->device_attr.max_qp : priv->device_attr.max_cq);
3868 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
3871 info->max_rx_queues = max;
3872 info->max_tx_queues = max;
3873 /* Last array entry is reserved for broadcast. */
3874 info->max_mac_addrs = (elemof(priv->mac) - 1);
3875 info->rx_offload_capa =
3877 (DEV_RX_OFFLOAD_IPV4_CKSUM |
3878 DEV_RX_OFFLOAD_UDP_CKSUM |
3879 DEV_RX_OFFLOAD_TCP_CKSUM) :
3881 info->tx_offload_capa =
3883 (DEV_TX_OFFLOAD_IPV4_CKSUM |
3884 DEV_TX_OFFLOAD_UDP_CKSUM |
3885 DEV_TX_OFFLOAD_TCP_CKSUM) :
3887 if (priv_get_ifname(priv, &ifname) == 0)
3888 info->if_index = if_nametoindex(ifname);
3893 * DPDK callback to get device statistics.
3896 * Pointer to Ethernet device structure.
3898 * Stats structure output buffer.
3901 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
3903 struct priv *priv = dev->data->dev_private;
3904 struct rte_eth_stats tmp = {0};
3909 /* Add software counters. */
3910 for (i = 0; (i != priv->rxqs_n); ++i) {
3911 struct rxq *rxq = (*priv->rxqs)[i];
3915 idx = rxq->stats.idx;
3916 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3917 #ifdef MLX4_PMD_SOFT_COUNTERS
3918 tmp.q_ipackets[idx] += rxq->stats.ipackets;
3919 tmp.q_ibytes[idx] += rxq->stats.ibytes;
3921 tmp.q_errors[idx] += (rxq->stats.idropped +
3922 rxq->stats.rx_nombuf);
3924 #ifdef MLX4_PMD_SOFT_COUNTERS
3925 tmp.ipackets += rxq->stats.ipackets;
3926 tmp.ibytes += rxq->stats.ibytes;
3928 tmp.ierrors += rxq->stats.idropped;
3929 tmp.rx_nombuf += rxq->stats.rx_nombuf;
3931 for (i = 0; (i != priv->txqs_n); ++i) {
3932 struct txq *txq = (*priv->txqs)[i];
3936 idx = txq->stats.idx;
3937 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3938 #ifdef MLX4_PMD_SOFT_COUNTERS
3939 tmp.q_opackets[idx] += txq->stats.opackets;
3940 tmp.q_obytes[idx] += txq->stats.obytes;
3942 tmp.q_errors[idx] += txq->stats.odropped;
3944 #ifdef MLX4_PMD_SOFT_COUNTERS
3945 tmp.opackets += txq->stats.opackets;
3946 tmp.obytes += txq->stats.obytes;
3948 tmp.oerrors += txq->stats.odropped;
3950 #ifndef MLX4_PMD_SOFT_COUNTERS
3951 /* FIXME: retrieve and add hardware counters. */
3958 * DPDK callback to clear device statistics.
3961 * Pointer to Ethernet device structure.
3964 mlx4_stats_reset(struct rte_eth_dev *dev)
3966 struct priv *priv = dev->data->dev_private;
3971 for (i = 0; (i != priv->rxqs_n); ++i) {
3972 if ((*priv->rxqs)[i] == NULL)
3974 idx = (*priv->rxqs)[i]->stats.idx;
3975 (*priv->rxqs)[i]->stats =
3976 (struct mlx4_rxq_stats){ .idx = idx };
3978 for (i = 0; (i != priv->txqs_n); ++i) {
3979 if ((*priv->txqs)[i] == NULL)
3981 idx = (*priv->txqs)[i]->stats.idx;
3982 (*priv->txqs)[i]->stats =
3983 (struct mlx4_txq_stats){ .idx = idx };
3985 #ifndef MLX4_PMD_SOFT_COUNTERS
3986 /* FIXME: reset hardware counters. */
3992 * DPDK callback to remove a MAC address.
3995 * Pointer to Ethernet device structure.
3997 * MAC address index.
4000 mlx4_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
4002 struct priv *priv = dev->data->dev_private;
4005 DEBUG("%p: removing MAC address from index %" PRIu32,
4006 (void *)dev, index);
4007 /* Last array entry is reserved for broadcast. */
4008 if (index >= (elemof(priv->mac) - 1))
4010 priv_mac_addr_del(priv, index);
4016 * DPDK callback to add a MAC address.
4019 * Pointer to Ethernet device structure.
4021 * MAC address to register.
4023 * MAC address index.
4025 * VMDq pool index to associate address with (ignored).
4028 mlx4_mac_addr_add(struct rte_eth_dev *dev, struct ether_addr *mac_addr,
4029 uint32_t index, uint32_t vmdq)
4031 struct priv *priv = dev->data->dev_private;
4035 DEBUG("%p: adding MAC address at index %" PRIu32,
4036 (void *)dev, index);
4037 /* Last array entry is reserved for broadcast. */
4038 if (index >= (elemof(priv->mac) - 1))
4040 priv_mac_addr_add(priv, index,
4041 (const uint8_t (*)[ETHER_ADDR_LEN])
4042 mac_addr->addr_bytes);
4048 * DPDK callback to enable promiscuous mode.
4051 * Pointer to Ethernet device structure.
4054 mlx4_promiscuous_enable(struct rte_eth_dev *dev)
4056 struct priv *priv = dev->data->dev_private;
4061 if (priv->promisc) {
4065 /* If device isn't started, this is all we need to do. */
4069 ret = rxq_promiscuous_enable(&priv->rxq_parent);
4076 for (i = 0; (i != priv->rxqs_n); ++i) {
4077 if ((*priv->rxqs)[i] == NULL)
4079 ret = rxq_promiscuous_enable((*priv->rxqs)[i]);
4082 /* Failure, rollback. */
4084 if ((*priv->rxqs)[--i] != NULL)
4085 rxq_promiscuous_disable((*priv->rxqs)[i]);
4095 * DPDK callback to disable promiscuous mode.
4098 * Pointer to Ethernet device structure.
4101 mlx4_promiscuous_disable(struct rte_eth_dev *dev)
4103 struct priv *priv = dev->data->dev_private;
4107 if (!priv->promisc) {
4112 rxq_promiscuous_disable(&priv->rxq_parent);
4115 for (i = 0; (i != priv->rxqs_n); ++i)
4116 if ((*priv->rxqs)[i] != NULL)
4117 rxq_promiscuous_disable((*priv->rxqs)[i]);
4124 * DPDK callback to enable allmulti mode.
4127 * Pointer to Ethernet device structure.
4130 mlx4_allmulticast_enable(struct rte_eth_dev *dev)
4132 struct priv *priv = dev->data->dev_private;
4137 if (priv->allmulti) {
4141 /* If device isn't started, this is all we need to do. */
4145 ret = rxq_allmulticast_enable(&priv->rxq_parent);
4152 for (i = 0; (i != priv->rxqs_n); ++i) {
4153 if ((*priv->rxqs)[i] == NULL)
4155 ret = rxq_allmulticast_enable((*priv->rxqs)[i]);
4158 /* Failure, rollback. */
4160 if ((*priv->rxqs)[--i] != NULL)
4161 rxq_allmulticast_disable((*priv->rxqs)[i]);
4171 * DPDK callback to disable allmulti mode.
4174 * Pointer to Ethernet device structure.
4177 mlx4_allmulticast_disable(struct rte_eth_dev *dev)
4179 struct priv *priv = dev->data->dev_private;
4183 if (!priv->allmulti) {
4188 rxq_allmulticast_disable(&priv->rxq_parent);
4191 for (i = 0; (i != priv->rxqs_n); ++i)
4192 if ((*priv->rxqs)[i] != NULL)
4193 rxq_allmulticast_disable((*priv->rxqs)[i]);
4200 * DPDK callback to retrieve physical link information (unlocked version).
4203 * Pointer to Ethernet device structure.
4204 * @param wait_to_complete
4205 * Wait for request completion (ignored).
4208 mlx4_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
4210 struct priv *priv = dev->data->dev_private;
4211 struct ethtool_cmd edata = {
4215 struct rte_eth_link dev_link;
4218 (void)wait_to_complete;
4219 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
4220 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
4223 memset(&dev_link, 0, sizeof(dev_link));
4224 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
4225 (ifr.ifr_flags & IFF_RUNNING));
4226 ifr.ifr_data = &edata;
4227 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4228 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
4232 link_speed = ethtool_cmd_speed(&edata);
4233 if (link_speed == -1)
4234 dev_link.link_speed = 0;
4236 dev_link.link_speed = link_speed;
4237 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
4238 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
4239 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
4240 /* Link status changed. */
4241 dev->data->dev_link = dev_link;
4244 /* Link status is still the same. */
4249 * DPDK callback to retrieve physical link information.
4252 * Pointer to Ethernet device structure.
4253 * @param wait_to_complete
4254 * Wait for request completion (ignored).
4257 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
4259 struct priv *priv = dev->data->dev_private;
4263 ret = mlx4_link_update_unlocked(dev, wait_to_complete);
4269 * DPDK callback to change the MTU.
4271 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
4272 * received). Use this as a hint to enable/disable scattered packets support
4273 * and improve performance when not needed.
4274 * Since failure is not an option, reconfiguring queues on the fly is not
4278 * Pointer to Ethernet device structure.
4283 * 0 on success, negative errno value on failure.
4286 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
4288 struct priv *priv = dev->data->dev_private;
4291 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
4295 /* Set kernel interface MTU first. */
4296 if (priv_set_mtu(priv, mtu)) {
4298 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
4302 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
4304 /* Temporarily replace RX handler with a fake one, assuming it has not
4305 * been copied elsewhere. */
4306 dev->rx_pkt_burst = removed_rx_burst;
4307 /* Make sure everyone has left mlx4_rx_burst() and uses
4308 * removed_rx_burst() instead. */
4311 /* Reconfigure each RX queue. */
4312 for (i = 0; (i != priv->rxqs_n); ++i) {
4313 struct rxq *rxq = (*priv->rxqs)[i];
4314 unsigned int max_frame_len;
4319 /* Calculate new maximum frame length according to MTU and
4320 * toggle scattered support (sp) if necessary. */
4321 max_frame_len = (priv->mtu + ETHER_HDR_LEN +
4322 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
4323 sp = (max_frame_len > (rxq->mb_len - RTE_PKTMBUF_HEADROOM));
4324 /* Provide new values to rxq_setup(). */
4325 dev->data->dev_conf.rxmode.jumbo_frame = sp;
4326 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
4327 ret = rxq_rehash(dev, rxq);
4329 /* Force SP RX if that queue requires it and abort. */
4331 rx_func = mlx4_rx_burst_sp;
4334 /* Reenable non-RSS queue attributes. No need to check
4335 * for errors at this stage. */
4337 rxq_mac_addrs_add(rxq);
4339 rxq_promiscuous_enable(rxq);
4341 rxq_allmulticast_enable(rxq);
4343 /* Scattered burst function takes priority. */
4345 rx_func = mlx4_rx_burst_sp;
4347 /* Burst functions can now be called again. */
4349 dev->rx_pkt_burst = rx_func;
4357 * DPDK callback to get flow control status.
4360 * Pointer to Ethernet device structure.
4361 * @param[out] fc_conf
4362 * Flow control output buffer.
4365 * 0 on success, negative errno value on failure.
4368 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4370 struct priv *priv = dev->data->dev_private;
4372 struct ethtool_pauseparam ethpause = {
4373 .cmd = ETHTOOL_GPAUSEPARAM
4377 ifr.ifr_data = ðpause;
4379 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4381 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
4387 fc_conf->autoneg = ethpause.autoneg;
4388 if (ethpause.rx_pause && ethpause.tx_pause)
4389 fc_conf->mode = RTE_FC_FULL;
4390 else if (ethpause.rx_pause)
4391 fc_conf->mode = RTE_FC_RX_PAUSE;
4392 else if (ethpause.tx_pause)
4393 fc_conf->mode = RTE_FC_TX_PAUSE;
4395 fc_conf->mode = RTE_FC_NONE;
4405 * DPDK callback to modify flow control parameters.
4408 * Pointer to Ethernet device structure.
4409 * @param[in] fc_conf
4410 * Flow control parameters.
4413 * 0 on success, negative errno value on failure.
4416 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4418 struct priv *priv = dev->data->dev_private;
4420 struct ethtool_pauseparam ethpause = {
4421 .cmd = ETHTOOL_SPAUSEPARAM
4425 ifr.ifr_data = ðpause;
4426 ethpause.autoneg = fc_conf->autoneg;
4427 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4428 (fc_conf->mode & RTE_FC_RX_PAUSE))
4429 ethpause.rx_pause = 1;
4431 ethpause.rx_pause = 0;
4433 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4434 (fc_conf->mode & RTE_FC_TX_PAUSE))
4435 ethpause.tx_pause = 1;
4437 ethpause.tx_pause = 0;
4440 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4442 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
4456 * Configure a VLAN filter.
4459 * Pointer to Ethernet device structure.
4461 * VLAN ID to filter.
4466 * 0 on success, errno value on failure.
4469 vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4471 struct priv *priv = dev->data->dev_private;
4473 unsigned int j = -1;
4475 DEBUG("%p: %s VLAN filter ID %" PRIu16,
4476 (void *)dev, (on ? "enable" : "disable"), vlan_id);
4477 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
4478 if (!priv->vlan_filter[i].enabled) {
4479 /* Unused index, remember it. */
4483 if (priv->vlan_filter[i].id != vlan_id)
4485 /* This VLAN ID is already known, use its index. */
4489 /* Check if there's room for another VLAN filter. */
4490 if (j == (unsigned int)-1)
4493 * VLAN filters apply to all configured MAC addresses, flow
4494 * specifications must be reconfigured accordingly.
4496 priv->vlan_filter[j].id = vlan_id;
4497 if ((on) && (!priv->vlan_filter[j].enabled)) {
4499 * Filter is disabled, enable it.
4500 * Rehashing flows in all RX queues is necessary.
4503 rxq_mac_addrs_del(&priv->rxq_parent);
4505 for (i = 0; (i != priv->rxqs_n); ++i)
4506 if ((*priv->rxqs)[i] != NULL)
4507 rxq_mac_addrs_del((*priv->rxqs)[i]);
4508 priv->vlan_filter[j].enabled = 1;
4509 if (priv->started) {
4511 rxq_mac_addrs_add(&priv->rxq_parent);
4513 for (i = 0; (i != priv->rxqs_n); ++i) {
4514 if ((*priv->rxqs)[i] == NULL)
4516 rxq_mac_addrs_add((*priv->rxqs)[i]);
4519 } else if ((!on) && (priv->vlan_filter[j].enabled)) {
4521 * Filter is enabled, disable it.
4522 * Rehashing flows in all RX queues is necessary.
4525 rxq_mac_addrs_del(&priv->rxq_parent);
4527 for (i = 0; (i != priv->rxqs_n); ++i)
4528 if ((*priv->rxqs)[i] != NULL)
4529 rxq_mac_addrs_del((*priv->rxqs)[i]);
4530 priv->vlan_filter[j].enabled = 0;
4531 if (priv->started) {
4533 rxq_mac_addrs_add(&priv->rxq_parent);
4535 for (i = 0; (i != priv->rxqs_n); ++i) {
4536 if ((*priv->rxqs)[i] == NULL)
4538 rxq_mac_addrs_add((*priv->rxqs)[i]);
4546 * DPDK callback to configure a VLAN filter.
4549 * Pointer to Ethernet device structure.
4551 * VLAN ID to filter.
4556 * 0 on success, negative errno value on failure.
4559 mlx4_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4561 struct priv *priv = dev->data->dev_private;
4565 ret = vlan_filter_set(dev, vlan_id, on);
4571 static const struct eth_dev_ops mlx4_dev_ops = {
4572 .dev_configure = mlx4_dev_configure,
4573 .dev_start = mlx4_dev_start,
4574 .dev_stop = mlx4_dev_stop,
4575 .dev_close = mlx4_dev_close,
4576 .promiscuous_enable = mlx4_promiscuous_enable,
4577 .promiscuous_disable = mlx4_promiscuous_disable,
4578 .allmulticast_enable = mlx4_allmulticast_enable,
4579 .allmulticast_disable = mlx4_allmulticast_disable,
4580 .link_update = mlx4_link_update,
4581 .stats_get = mlx4_stats_get,
4582 .stats_reset = mlx4_stats_reset,
4583 .queue_stats_mapping_set = NULL,
4584 .dev_infos_get = mlx4_dev_infos_get,
4585 .vlan_filter_set = mlx4_vlan_filter_set,
4586 .vlan_tpid_set = NULL,
4587 .vlan_strip_queue_set = NULL,
4588 .vlan_offload_set = NULL,
4589 .rx_queue_setup = mlx4_rx_queue_setup,
4590 .tx_queue_setup = mlx4_tx_queue_setup,
4591 .rx_queue_release = mlx4_rx_queue_release,
4592 .tx_queue_release = mlx4_tx_queue_release,
4594 .dev_led_off = NULL,
4595 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
4596 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
4597 .priority_flow_ctrl_set = NULL,
4598 .mac_addr_remove = mlx4_mac_addr_remove,
4599 .mac_addr_add = mlx4_mac_addr_add,
4600 .mtu_set = mlx4_dev_set_mtu,
4601 .udp_tunnel_add = NULL,
4602 .udp_tunnel_del = NULL,
4606 * Get PCI information from struct ibv_device.
4609 * Pointer to Ethernet device structure.
4610 * @param[out] pci_addr
4611 * PCI bus address output buffer.
4614 * 0 on success, -1 on failure and errno is set.
4617 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
4618 struct rte_pci_addr *pci_addr)
4622 MKSTR(path, "%s/device/uevent", device->ibdev_path);
4624 file = fopen(path, "rb");
4627 while (fgets(line, sizeof(line), file) == line) {
4628 size_t len = strlen(line);
4631 /* Truncate long lines. */
4632 if (len == (sizeof(line) - 1))
4633 while (line[(len - 1)] != '\n') {
4637 line[(len - 1)] = ret;
4639 /* Extract information. */
4642 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
4646 &pci_addr->function) == 4) {
4656 * Get MAC address by querying netdevice.
4659 * struct priv for the requested device.
4661 * MAC address output buffer.
4664 * 0 on success, -1 on failure and errno is set.
4667 priv_get_mac(struct priv *priv, uint8_t (*mac)[ETHER_ADDR_LEN])
4669 struct ifreq request;
4671 if (priv_ifreq(priv, SIOCGIFHWADDR, &request))
4673 memcpy(mac, request.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
4677 /* Support up to 32 adapters. */
4679 struct rte_pci_addr pci_addr; /* associated PCI address */
4680 uint32_t ports; /* physical ports bitfield. */
4684 * Get device index in mlx4_dev[] from PCI bus address.
4686 * @param[in] pci_addr
4687 * PCI bus address to look for.
4690 * mlx4_dev[] index on success, -1 on failure.
4693 mlx4_dev_idx(struct rte_pci_addr *pci_addr)
4698 assert(pci_addr != NULL);
4699 for (i = 0; (i != elemof(mlx4_dev)); ++i) {
4700 if ((mlx4_dev[i].pci_addr.domain == pci_addr->domain) &&
4701 (mlx4_dev[i].pci_addr.bus == pci_addr->bus) &&
4702 (mlx4_dev[i].pci_addr.devid == pci_addr->devid) &&
4703 (mlx4_dev[i].pci_addr.function == pci_addr->function))
4705 if ((mlx4_dev[i].ports == 0) && (ret == -1))
4712 * Retrieve integer value from environment variable.
4715 * Environment variable name.
4718 * Integer value, 0 if the variable is not set.
4721 mlx4_getenv_int(const char *name)
4723 const char *val = getenv(name);
4731 mlx4_dev_link_status_handler(void *);
4733 mlx4_dev_interrupt_handler(struct rte_intr_handle *, void *);
4736 * Link status handler.
4739 * Pointer to private structure.
4741 * Pointer to the rte_eth_dev structure.
4744 * Nonzero if the callback process can be called immediately.
4747 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
4749 struct ibv_async_event event;
4750 int port_change = 0;
4753 /* Read all message and acknowledge them. */
4755 if (ibv_get_async_event(priv->ctx, &event))
4758 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
4759 event.event_type == IBV_EVENT_PORT_ERR)
4762 DEBUG("event type %d on port %d not handled",
4763 event.event_type, event.element.port_num);
4764 ibv_ack_async_event(&event);
4767 if (port_change ^ priv->pending_alarm) {
4768 struct rte_eth_link *link = &dev->data->dev_link;
4770 priv->pending_alarm = 0;
4771 mlx4_link_update_unlocked(dev, 0);
4772 if (((link->link_speed == 0) && link->link_status) ||
4773 ((link->link_speed != 0) && !link->link_status)) {
4774 /* Inconsistent status, check again later. */
4775 priv->pending_alarm = 1;
4776 rte_eal_alarm_set(MLX4_ALARM_TIMEOUT_US,
4777 mlx4_dev_link_status_handler,
4786 * Handle delayed link status event.
4789 * Registered argument.
4792 mlx4_dev_link_status_handler(void *arg)
4794 struct rte_eth_dev *dev = arg;
4795 struct priv *priv = dev->data->dev_private;
4799 assert(priv->pending_alarm == 1);
4800 ret = priv_dev_link_status_handler(priv, dev);
4803 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
4807 * Handle interrupts from the NIC.
4809 * @param[in] intr_handle
4810 * Interrupt handler.
4812 * Callback argument.
4815 mlx4_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
4817 struct rte_eth_dev *dev = cb_arg;
4818 struct priv *priv = dev->data->dev_private;
4823 ret = priv_dev_link_status_handler(priv, dev);
4826 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
4830 * Uninstall interrupt handler.
4833 * Pointer to private structure.
4835 * Pointer to the rte_eth_dev structure.
4838 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
4840 if (!dev->data->dev_conf.intr_conf.lsc)
4842 rte_intr_callback_unregister(&priv->intr_handle,
4843 mlx4_dev_interrupt_handler,
4845 if (priv->pending_alarm)
4846 rte_eal_alarm_cancel(mlx4_dev_link_status_handler, dev);
4847 priv->pending_alarm = 0;
4848 priv->intr_handle.fd = 0;
4849 priv->intr_handle.type = 0;
4853 * Install interrupt handler.
4856 * Pointer to private structure.
4858 * Pointer to the rte_eth_dev structure.
4861 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
4865 if (!dev->data->dev_conf.intr_conf.lsc)
4867 assert(priv->ctx->async_fd > 0);
4868 flags = fcntl(priv->ctx->async_fd, F_GETFL);
4869 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
4871 INFO("failed to change file descriptor async event queue");
4872 dev->data->dev_conf.intr_conf.lsc = 0;
4874 priv->intr_handle.fd = priv->ctx->async_fd;
4875 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
4876 rte_intr_callback_register(&priv->intr_handle,
4877 mlx4_dev_interrupt_handler,
4882 static struct eth_driver mlx4_driver;
4885 * DPDK callback to register a PCI device.
4887 * This function creates an Ethernet device for each port of a given
4890 * @param[in] pci_drv
4891 * PCI driver structure (mlx4_driver).
4892 * @param[in] pci_dev
4893 * PCI device information.
4896 * 0 on success, negative errno value on failure.
4899 mlx4_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
4901 struct ibv_device **list;
4902 struct ibv_device *ibv_dev;
4904 struct ibv_context *attr_ctx = NULL;
4905 struct ibv_device_attr device_attr;
4911 assert(pci_drv == &mlx4_driver.pci_drv);
4912 /* Get mlx4_dev[] index. */
4913 idx = mlx4_dev_idx(&pci_dev->addr);
4915 ERROR("this driver cannot support any more adapters");
4918 DEBUG("using driver device index %d", idx);
4920 /* Save PCI address. */
4921 mlx4_dev[idx].pci_addr = pci_dev->addr;
4922 list = ibv_get_device_list(&i);
4925 if (errno == ENOSYS) {
4926 WARN("cannot list devices, is ib_uverbs loaded?");
4933 * For each listed device, check related sysfs entry against
4934 * the provided PCI ID.
4937 struct rte_pci_addr pci_addr;
4940 DEBUG("checking device \"%s\"", list[i]->name);
4941 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
4943 if ((pci_dev->addr.domain != pci_addr.domain) ||
4944 (pci_dev->addr.bus != pci_addr.bus) ||
4945 (pci_dev->addr.devid != pci_addr.devid) ||
4946 (pci_dev->addr.function != pci_addr.function))
4948 vf = (pci_dev->id.device_id ==
4949 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
4950 INFO("PCI information matches, using device \"%s\" (VF: %s)",
4951 list[i]->name, (vf ? "true" : "false"));
4952 attr_ctx = ibv_open_device(list[i]);
4956 if (attr_ctx == NULL) {
4957 ibv_free_device_list(list);
4960 WARN("cannot access device, is mlx4_ib loaded?");
4963 WARN("cannot use device, are drivers up to date?");
4971 DEBUG("device opened");
4972 if (ibv_query_device(attr_ctx, &device_attr))
4974 INFO("%u port(s) detected", device_attr.phys_port_cnt);
4976 for (i = 0; i < device_attr.phys_port_cnt; i++) {
4977 uint32_t port = i + 1; /* ports are indexed from one */
4978 uint32_t test = (1 << i);
4979 struct ibv_context *ctx = NULL;
4980 struct ibv_port_attr port_attr;
4981 struct ibv_pd *pd = NULL;
4982 struct priv *priv = NULL;
4983 struct rte_eth_dev *eth_dev;
4984 #ifdef HAVE_EXP_QUERY_DEVICE
4985 struct ibv_exp_device_attr exp_device_attr;
4986 #endif /* HAVE_EXP_QUERY_DEVICE */
4987 struct ether_addr mac;
4989 #ifdef HAVE_EXP_QUERY_DEVICE
4990 exp_device_attr.comp_mask = IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS;
4992 exp_device_attr.comp_mask |= IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ;
4993 #endif /* RSS_SUPPORT */
4994 #endif /* HAVE_EXP_QUERY_DEVICE */
4996 DEBUG("using port %u (%08" PRIx32 ")", port, test);
4998 ctx = ibv_open_device(ibv_dev);
5002 /* Check port status. */
5003 err = ibv_query_port(ctx, port, &port_attr);
5005 ERROR("port query failed: %s", strerror(err));
5008 if (port_attr.state != IBV_PORT_ACTIVE)
5009 DEBUG("port %d is not active: \"%s\" (%d)",
5010 port, ibv_port_state_str(port_attr.state),
5013 /* Allocate protection domain. */
5014 pd = ibv_alloc_pd(ctx);
5016 ERROR("PD allocation failure");
5021 mlx4_dev[idx].ports |= test;
5023 /* from rte_ethdev.c */
5024 priv = rte_zmalloc("ethdev private structure",
5026 RTE_CACHE_LINE_SIZE);
5028 ERROR("priv allocation failure");
5034 priv->device_attr = device_attr;
5037 priv->mtu = ETHER_MTU;
5038 #ifdef HAVE_EXP_QUERY_DEVICE
5039 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
5040 ERROR("ibv_exp_query_device() failed");
5044 if ((exp_device_attr.exp_device_cap_flags &
5045 IBV_EXP_DEVICE_QPG) &&
5046 (exp_device_attr.exp_device_cap_flags &
5047 IBV_EXP_DEVICE_UD_RSS) &&
5048 (exp_device_attr.comp_mask &
5049 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ) &&
5050 (exp_device_attr.max_rss_tbl_sz > 0)) {
5053 priv->max_rss_tbl_sz = exp_device_attr.max_rss_tbl_sz;
5057 priv->max_rss_tbl_sz = 0;
5059 priv->hw_tss = !!(exp_device_attr.exp_device_cap_flags &
5060 IBV_EXP_DEVICE_UD_TSS);
5061 DEBUG("device flags: %s%s%s",
5062 (priv->hw_qpg ? "IBV_DEVICE_QPG " : ""),
5063 (priv->hw_tss ? "IBV_DEVICE_TSS " : ""),
5064 (priv->hw_rss ? "IBV_DEVICE_RSS " : ""));
5066 DEBUG("maximum RSS indirection table size: %u",
5067 exp_device_attr.max_rss_tbl_sz);
5068 #endif /* RSS_SUPPORT */
5071 ((exp_device_attr.exp_device_cap_flags &
5072 IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
5073 (exp_device_attr.exp_device_cap_flags &
5074 IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
5075 DEBUG("checksum offloading is %ssupported",
5076 (priv->hw_csum ? "" : "not "));
5078 priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
5079 IBV_EXP_DEVICE_VXLAN_SUPPORT);
5080 DEBUG("L2 tunnel checksum offloads are %ssupported",
5081 (priv->hw_csum_l2tun ? "" : "not "));
5084 priv->inl_recv_size = mlx4_getenv_int("MLX4_INLINE_RECV_SIZE");
5086 if (priv->inl_recv_size) {
5087 exp_device_attr.comp_mask =
5088 IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ;
5089 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
5090 INFO("Couldn't query device for inline-receive"
5092 priv->inl_recv_size = 0;
5094 if ((unsigned)exp_device_attr.inline_recv_sz <
5095 priv->inl_recv_size) {
5096 INFO("Max inline-receive (%d) <"
5097 " requested inline-receive (%u)",
5098 exp_device_attr.inline_recv_sz,
5099 priv->inl_recv_size);
5100 priv->inl_recv_size =
5101 exp_device_attr.inline_recv_sz;
5104 INFO("Set inline receive size to %u",
5105 priv->inl_recv_size);
5107 #endif /* INLINE_RECV */
5108 #endif /* HAVE_EXP_QUERY_DEVICE */
5110 (void)mlx4_getenv_int;
5112 /* Configure the first MAC address by default. */
5113 if (priv_get_mac(priv, &mac.addr_bytes)) {
5114 ERROR("cannot get MAC address, is mlx4_en loaded?"
5115 " (errno: %s)", strerror(errno));
5118 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
5120 mac.addr_bytes[0], mac.addr_bytes[1],
5121 mac.addr_bytes[2], mac.addr_bytes[3],
5122 mac.addr_bytes[4], mac.addr_bytes[5]);
5123 /* Register MAC and broadcast addresses. */
5124 claim_zero(priv_mac_addr_add(priv, 0,
5125 (const uint8_t (*)[ETHER_ADDR_LEN])
5127 claim_zero(priv_mac_addr_add(priv, (elemof(priv->mac) - 1),
5128 &(const uint8_t [ETHER_ADDR_LEN])
5129 { "\xff\xff\xff\xff\xff\xff" }));
5132 char ifname[IF_NAMESIZE];
5134 if (priv_get_ifname(priv, &ifname) == 0)
5135 DEBUG("port %u ifname is \"%s\"",
5136 priv->port, ifname);
5138 DEBUG("port %u ifname is unknown", priv->port);
5141 /* Get actual MTU if possible. */
5142 priv_get_mtu(priv, &priv->mtu);
5143 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
5145 /* from rte_ethdev.c */
5147 char name[RTE_ETH_NAME_MAX_LEN];
5149 snprintf(name, sizeof(name), "%s port %u",
5150 ibv_get_device_name(ibv_dev), port);
5151 eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
5153 if (eth_dev == NULL) {
5154 ERROR("can not allocate rte ethdev");
5159 eth_dev->data->dev_private = priv;
5160 eth_dev->pci_dev = pci_dev;
5162 rte_eth_copy_pci_info(eth_dev, pci_dev);
5164 eth_dev->driver = &mlx4_driver;
5165 eth_dev->data->rx_mbuf_alloc_failed = 0;
5166 eth_dev->data->mtu = ETHER_MTU;
5168 priv->dev = eth_dev;
5169 eth_dev->dev_ops = &mlx4_dev_ops;
5170 eth_dev->data->mac_addrs = priv->mac;
5171 TAILQ_INIT(ð_dev->link_intr_cbs);
5173 /* Bring Ethernet device up. */
5174 DEBUG("forcing Ethernet interface up");
5175 priv_set_flags(priv, ~IFF_UP, IFF_UP);
5181 claim_zero(ibv_dealloc_pd(pd));
5183 claim_zero(ibv_close_device(ctx));
5188 * XXX if something went wrong in the loop above, there is a resource
5189 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
5190 * long as the dpdk does not provide a way to deallocate a ethdev and a
5191 * way to enumerate the registered ethdevs to free the previous ones.
5194 /* no port found, complain */
5195 if (!mlx4_dev[idx].ports) {
5202 claim_zero(ibv_close_device(attr_ctx));
5204 ibv_free_device_list(list);
5209 static const struct rte_pci_id mlx4_pci_id_map[] = {
5211 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5212 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3,
5213 .subsystem_vendor_id = PCI_ANY_ID,
5214 .subsystem_device_id = PCI_ANY_ID
5217 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5218 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO,
5219 .subsystem_vendor_id = PCI_ANY_ID,
5220 .subsystem_device_id = PCI_ANY_ID
5223 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5224 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3VF,
5225 .subsystem_vendor_id = PCI_ANY_ID,
5226 .subsystem_device_id = PCI_ANY_ID
5233 static struct eth_driver mlx4_driver = {
5235 .name = MLX4_DRIVER_NAME,
5236 .id_table = mlx4_pci_id_map,
5237 .devinit = mlx4_pci_devinit,
5238 .drv_flags = RTE_PCI_DRV_INTR_LSC,
5240 .dev_private_size = sizeof(struct priv)
5244 * Driver initialization routine.
5247 rte_mlx4_pmd_init(const char *name, const char *args)
5252 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
5253 * huge pages. Calling ibv_fork_init() during init allows
5254 * applications to use fork() safely for purposes other than
5255 * using this PMD, which is not supported in forked processes.
5257 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
5259 rte_eal_pci_register(&mlx4_driver.pci_drv);
5263 static struct rte_driver rte_mlx4_driver = {
5265 .name = MLX4_DRIVER_NAME,
5266 .init = rte_mlx4_pmd_init,
5269 PMD_REGISTER_DRIVER(rte_mlx4_driver)