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;
1199 return (struct tx_burst_sg_ret){
1200 .length = sent_size,
1204 return (struct tx_burst_sg_ret){
1210 #endif /* MLX4_PMD_SGE_WR_N > 1 */
1213 * DPDK callback for TX.
1216 * Generic pointer to TX queue structure.
1218 * Packets to transmit.
1220 * Number of packets in array.
1223 * Number of packets successfully transmitted (<= pkts_n).
1226 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1228 struct txq *txq = (struct txq *)dpdk_txq;
1229 unsigned int elts_head = txq->elts_head;
1230 const unsigned int elts_tail = txq->elts_tail;
1231 const unsigned int elts_n = txq->elts_n;
1232 unsigned int elts_comp_cd = txq->elts_comp_cd;
1233 unsigned int elts_comp = 0;
1238 assert(elts_comp_cd != 0);
1240 max = (elts_n - (elts_head - elts_tail));
1244 assert(max <= elts_n);
1245 /* Always leave one free entry in the ring. */
1251 for (i = 0; (i != max); ++i) {
1252 struct rte_mbuf *buf = pkts[i];
1253 unsigned int elts_head_next =
1254 (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
1255 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
1256 struct txq_elt *elt = &(*txq->elts)[elts_head];
1257 unsigned int segs = NB_SEGS(buf);
1258 #ifdef MLX4_PMD_SOFT_COUNTERS
1259 unsigned int sent_size = 0;
1261 uint32_t send_flags = 0;
1263 /* Clean up old buffer. */
1264 if (likely(elt->buf != NULL)) {
1265 struct rte_mbuf *tmp = elt->buf;
1267 /* Faster than rte_pktmbuf_free(). */
1269 struct rte_mbuf *next = NEXT(tmp);
1271 rte_pktmbuf_free_seg(tmp);
1273 } while (tmp != NULL);
1275 /* Request TX completion. */
1276 if (unlikely(--elts_comp_cd == 0)) {
1277 elts_comp_cd = txq->elts_comp_cd_init;
1279 send_flags |= IBV_EXP_QP_BURST_SIGNALED;
1281 /* Should we enable HW CKSUM offload */
1283 (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
1284 send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
1285 /* HW does not support checksum offloads at arbitrary
1286 * offsets but automatically recognizes the packet
1287 * type. For inner L3/L4 checksums, only VXLAN (UDP)
1288 * tunnels are currently supported. */
1289 if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
1290 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
1292 if (likely(segs == 1)) {
1297 /* Retrieve buffer information. */
1298 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1299 length = DATA_LEN(buf);
1300 /* Retrieve Memory Region key for this memory pool. */
1301 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
1302 if (unlikely(lkey == (uint32_t)-1)) {
1303 /* MR does not exist. */
1304 DEBUG("%p: unable to get MP <-> MR"
1305 " association", (void *)txq);
1306 /* Clean up TX element. */
1310 /* Update element. */
1313 rte_prefetch0((volatile void *)
1315 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1316 /* Put packet into send queue. */
1317 #if MLX4_PMD_MAX_INLINE > 0
1318 if (length <= txq->max_inline)
1319 err = txq->if_qp->send_pending_inline
1326 err = txq->if_qp->send_pending
1334 #ifdef MLX4_PMD_SOFT_COUNTERS
1335 sent_size += length;
1338 #if MLX4_PMD_SGE_WR_N > 1
1339 struct ibv_sge sges[MLX4_PMD_SGE_WR_N];
1340 struct tx_burst_sg_ret ret;
1342 ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
1344 if (ret.length == (unsigned int)-1)
1346 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1347 /* Put SG list into send queue. */
1348 err = txq->if_qp->send_pending_sg_list
1355 #ifdef MLX4_PMD_SOFT_COUNTERS
1356 sent_size += ret.length;
1358 #else /* MLX4_PMD_SGE_WR_N > 1 */
1359 DEBUG("%p: TX scattered buffers support not"
1360 " compiled in", (void *)txq);
1362 #endif /* MLX4_PMD_SGE_WR_N > 1 */
1364 elts_head = elts_head_next;
1365 #ifdef MLX4_PMD_SOFT_COUNTERS
1366 /* Increment sent bytes counter. */
1367 txq->stats.obytes += sent_size;
1371 /* Take a shortcut if nothing must be sent. */
1372 if (unlikely(i == 0))
1374 #ifdef MLX4_PMD_SOFT_COUNTERS
1375 /* Increment sent packets counter. */
1376 txq->stats.opackets += i;
1378 /* Ring QP doorbell. */
1379 err = txq->if_qp->send_flush(txq->qp);
1380 if (unlikely(err)) {
1381 /* A nonzero value is not supposed to be returned.
1382 * Nothing can be done about it. */
1383 DEBUG("%p: send_flush() failed with error %d",
1386 txq->elts_head = elts_head;
1387 txq->elts_comp += elts_comp;
1388 txq->elts_comp_cd = elts_comp_cd;
1393 * Configure a TX queue.
1396 * Pointer to Ethernet device structure.
1398 * Pointer to TX queue structure.
1400 * Number of descriptors to configure in queue.
1402 * NUMA socket on which memory must be allocated.
1404 * Thresholds parameters.
1407 * 0 on success, errno value on failure.
1410 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1411 unsigned int socket, const struct rte_eth_txconf *conf)
1413 struct priv *priv = dev->data->dev_private;
1419 struct ibv_exp_query_intf_params params;
1420 struct ibv_exp_qp_init_attr init;
1421 struct ibv_exp_res_domain_init_attr rd;
1422 struct ibv_exp_cq_init_attr cq;
1423 struct ibv_exp_qp_attr mod;
1425 enum ibv_exp_query_intf_status status;
1428 (void)conf; /* Thresholds configuration (ignored). */
1429 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
1430 ERROR("%p: invalid number of TX descriptors (must be a"
1431 " multiple of %d)", (void *)dev, MLX4_PMD_SGE_WR_N);
1434 desc /= MLX4_PMD_SGE_WR_N;
1435 /* MRs will be registered in mp2mr[] later. */
1436 attr.rd = (struct ibv_exp_res_domain_init_attr){
1437 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
1438 IBV_EXP_RES_DOMAIN_MSG_MODEL),
1439 .thread_model = IBV_EXP_THREAD_SINGLE,
1440 .msg_model = IBV_EXP_MSG_HIGH_BW,
1442 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
1443 if (tmpl.rd == NULL) {
1445 ERROR("%p: RD creation failure: %s",
1446 (void *)dev, strerror(ret));
1449 attr.cq = (struct ibv_exp_cq_init_attr){
1450 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
1451 .res_domain = tmpl.rd,
1453 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
1454 if (tmpl.cq == NULL) {
1456 ERROR("%p: CQ creation failure: %s",
1457 (void *)dev, strerror(ret));
1460 DEBUG("priv->device_attr.max_qp_wr is %d",
1461 priv->device_attr.max_qp_wr);
1462 DEBUG("priv->device_attr.max_sge is %d",
1463 priv->device_attr.max_sge);
1464 attr.init = (struct ibv_exp_qp_init_attr){
1465 /* CQ to be associated with the send queue. */
1467 /* CQ to be associated with the receive queue. */
1470 /* Max number of outstanding WRs. */
1471 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1472 priv->device_attr.max_qp_wr :
1474 /* Max number of scatter/gather elements in a WR. */
1475 .max_send_sge = ((priv->device_attr.max_sge <
1476 MLX4_PMD_SGE_WR_N) ?
1477 priv->device_attr.max_sge :
1479 #if MLX4_PMD_MAX_INLINE > 0
1480 .max_inline_data = MLX4_PMD_MAX_INLINE,
1483 .qp_type = IBV_QPT_RAW_PACKET,
1484 /* Do *NOT* enable this, completions events are managed per
1488 .res_domain = tmpl.rd,
1489 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
1490 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
1492 tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
1493 if (tmpl.qp == NULL) {
1494 ret = (errno ? errno : EINVAL);
1495 ERROR("%p: QP creation failure: %s",
1496 (void *)dev, strerror(ret));
1499 #if MLX4_PMD_MAX_INLINE > 0
1500 /* ibv_create_qp() updates this value. */
1501 tmpl.max_inline = attr.init.cap.max_inline_data;
1503 attr.mod = (struct ibv_exp_qp_attr){
1504 /* Move the QP to this state. */
1505 .qp_state = IBV_QPS_INIT,
1506 /* Primary port number. */
1507 .port_num = priv->port
1509 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
1510 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
1512 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1513 (void *)dev, strerror(ret));
1516 ret = txq_alloc_elts(&tmpl, desc);
1518 ERROR("%p: TXQ allocation failed: %s",
1519 (void *)dev, strerror(ret));
1522 attr.mod = (struct ibv_exp_qp_attr){
1523 .qp_state = IBV_QPS_RTR
1525 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1527 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1528 (void *)dev, strerror(ret));
1531 attr.mod.qp_state = IBV_QPS_RTS;
1532 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1534 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1535 (void *)dev, strerror(ret));
1538 attr.params = (struct ibv_exp_query_intf_params){
1539 .intf_scope = IBV_EXP_INTF_GLOBAL,
1540 .intf = IBV_EXP_INTF_CQ,
1543 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
1544 if (tmpl.if_cq == NULL) {
1545 ERROR("%p: CQ interface family query failed with status %d",
1546 (void *)dev, status);
1549 attr.params = (struct ibv_exp_query_intf_params){
1550 .intf_scope = IBV_EXP_INTF_GLOBAL,
1551 .intf = IBV_EXP_INTF_QP_BURST,
1553 #ifdef HAVE_EXP_QP_BURST_CREATE_DISABLE_ETH_LOOPBACK
1554 /* MC loopback must be disabled when not using a VF. */
1557 IBV_EXP_QP_BURST_CREATE_DISABLE_ETH_LOOPBACK :
1561 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
1562 if (tmpl.if_qp == NULL) {
1563 ERROR("%p: QP interface family query failed with status %d",
1564 (void *)dev, status);
1567 /* Clean up txq in case we're reinitializing it. */
1568 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1571 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1581 * DPDK callback to configure a TX queue.
1584 * Pointer to Ethernet device structure.
1588 * Number of descriptors to configure in queue.
1590 * NUMA socket on which memory must be allocated.
1592 * Thresholds parameters.
1595 * 0 on success, negative errno value on failure.
1598 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1599 unsigned int socket, const struct rte_eth_txconf *conf)
1601 struct priv *priv = dev->data->dev_private;
1602 struct txq *txq = (*priv->txqs)[idx];
1606 DEBUG("%p: configuring queue %u for %u descriptors",
1607 (void *)dev, idx, desc);
1608 if (idx >= priv->txqs_n) {
1609 ERROR("%p: queue index out of range (%u >= %u)",
1610 (void *)dev, idx, priv->txqs_n);
1615 DEBUG("%p: reusing already allocated queue index %u (%p)",
1616 (void *)dev, idx, (void *)txq);
1617 if (priv->started) {
1621 (*priv->txqs)[idx] = NULL;
1624 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1626 ERROR("%p: unable to allocate queue index %u",
1632 ret = txq_setup(dev, txq, desc, socket, conf);
1636 txq->stats.idx = idx;
1637 DEBUG("%p: adding TX queue %p to list",
1638 (void *)dev, (void *)txq);
1639 (*priv->txqs)[idx] = txq;
1640 /* Update send callback. */
1641 dev->tx_pkt_burst = mlx4_tx_burst;
1648 * DPDK callback to release a TX queue.
1651 * Generic TX queue pointer.
1654 mlx4_tx_queue_release(void *dpdk_txq)
1656 struct txq *txq = (struct txq *)dpdk_txq;
1664 for (i = 0; (i != priv->txqs_n); ++i)
1665 if ((*priv->txqs)[i] == txq) {
1666 DEBUG("%p: removing TX queue %p from list",
1667 (void *)priv->dev, (void *)txq);
1668 (*priv->txqs)[i] = NULL;
1676 /* RX queues handling. */
1679 * Allocate RX queue elements with scattered packets support.
1682 * Pointer to RX queue structure.
1684 * Number of elements to allocate.
1686 * If not NULL, fetch buffers from this array instead of allocating them
1687 * with rte_pktmbuf_alloc().
1690 * 0 on success, errno value on failure.
1693 rxq_alloc_elts_sp(struct rxq *rxq, unsigned int elts_n,
1694 struct rte_mbuf **pool)
1697 struct rxq_elt_sp (*elts)[elts_n] =
1698 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1703 ERROR("%p: can't allocate packets array", (void *)rxq);
1707 /* For each WR (packet). */
1708 for (i = 0; (i != elts_n); ++i) {
1710 struct rxq_elt_sp *elt = &(*elts)[i];
1711 struct ibv_recv_wr *wr = &elt->wr;
1712 struct ibv_sge (*sges)[(elemof(elt->sges))] = &elt->sges;
1714 /* These two arrays must have the same size. */
1715 assert(elemof(elt->sges) == elemof(elt->bufs));
1718 wr->next = &(*elts)[(i + 1)].wr;
1719 wr->sg_list = &(*sges)[0];
1720 wr->num_sge = elemof(*sges);
1721 /* For each SGE (segment). */
1722 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1723 struct ibv_sge *sge = &(*sges)[j];
1724 struct rte_mbuf *buf;
1728 assert(buf != NULL);
1729 rte_pktmbuf_reset(buf);
1731 buf = rte_pktmbuf_alloc(rxq->mp);
1733 assert(pool == NULL);
1734 ERROR("%p: empty mbuf pool", (void *)rxq);
1739 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1740 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1741 /* Buffer is supposed to be empty. */
1742 assert(rte_pktmbuf_data_len(buf) == 0);
1743 assert(rte_pktmbuf_pkt_len(buf) == 0);
1744 /* sge->addr must be able to store a pointer. */
1745 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1747 /* The first SGE keeps its headroom. */
1748 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
1749 sge->length = (buf->buf_len -
1750 RTE_PKTMBUF_HEADROOM);
1752 /* Subsequent SGEs lose theirs. */
1753 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1754 SET_DATA_OFF(buf, 0);
1755 sge->addr = (uintptr_t)buf->buf_addr;
1756 sge->length = buf->buf_len;
1758 sge->lkey = rxq->mr->lkey;
1759 /* Redundant check for tailroom. */
1760 assert(sge->length == rte_pktmbuf_tailroom(buf));
1763 /* The last WR pointer must be NULL. */
1764 (*elts)[(i - 1)].wr.next = NULL;
1765 DEBUG("%p: allocated and configured %u WRs (%zu segments)",
1766 (void *)rxq, elts_n, (elts_n * elemof((*elts)[0].sges)));
1767 rxq->elts_n = elts_n;
1769 rxq->elts.sp = elts;
1774 assert(pool == NULL);
1775 for (i = 0; (i != elemof(*elts)); ++i) {
1777 struct rxq_elt_sp *elt = &(*elts)[i];
1779 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1780 struct rte_mbuf *buf = elt->bufs[j];
1783 rte_pktmbuf_free_seg(buf);
1788 DEBUG("%p: failed, freed everything", (void *)rxq);
1794 * Free RX queue elements with scattered packets support.
1797 * Pointer to RX queue structure.
1800 rxq_free_elts_sp(struct rxq *rxq)
1803 unsigned int elts_n = rxq->elts_n;
1804 struct rxq_elt_sp (*elts)[elts_n] = rxq->elts.sp;
1806 DEBUG("%p: freeing WRs", (void *)rxq);
1808 rxq->elts.sp = NULL;
1811 for (i = 0; (i != elemof(*elts)); ++i) {
1813 struct rxq_elt_sp *elt = &(*elts)[i];
1815 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1816 struct rte_mbuf *buf = elt->bufs[j];
1819 rte_pktmbuf_free_seg(buf);
1826 * Allocate RX queue elements.
1829 * Pointer to RX queue structure.
1831 * Number of elements to allocate.
1833 * If not NULL, fetch buffers from this array instead of allocating them
1834 * with rte_pktmbuf_alloc().
1837 * 0 on success, errno value on failure.
1840 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n, struct rte_mbuf **pool)
1843 struct rxq_elt (*elts)[elts_n] =
1844 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1849 ERROR("%p: can't allocate packets array", (void *)rxq);
1853 /* For each WR (packet). */
1854 for (i = 0; (i != elts_n); ++i) {
1855 struct rxq_elt *elt = &(*elts)[i];
1856 struct ibv_recv_wr *wr = &elt->wr;
1857 struct ibv_sge *sge = &(*elts)[i].sge;
1858 struct rte_mbuf *buf;
1862 assert(buf != NULL);
1863 rte_pktmbuf_reset(buf);
1865 buf = rte_pktmbuf_alloc(rxq->mp);
1867 assert(pool == NULL);
1868 ERROR("%p: empty mbuf pool", (void *)rxq);
1872 /* Configure WR. Work request ID contains its own index in
1873 * the elts array and the offset between SGE buffer header and
1875 WR_ID(wr->wr_id).id = i;
1876 WR_ID(wr->wr_id).offset =
1877 (((uintptr_t)buf->buf_addr + RTE_PKTMBUF_HEADROOM) -
1879 wr->next = &(*elts)[(i + 1)].wr;
1882 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1883 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1884 /* Buffer is supposed to be empty. */
1885 assert(rte_pktmbuf_data_len(buf) == 0);
1886 assert(rte_pktmbuf_pkt_len(buf) == 0);
1887 /* sge->addr must be able to store a pointer. */
1888 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1889 /* SGE keeps its headroom. */
1890 sge->addr = (uintptr_t)
1891 ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
1892 sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
1893 sge->lkey = rxq->mr->lkey;
1894 /* Redundant check for tailroom. */
1895 assert(sge->length == rte_pktmbuf_tailroom(buf));
1896 /* Make sure elts index and SGE mbuf pointer can be deduced
1898 if ((WR_ID(wr->wr_id).id != i) ||
1899 ((void *)((uintptr_t)sge->addr -
1900 WR_ID(wr->wr_id).offset) != buf)) {
1901 ERROR("%p: cannot store index and offset in WR ID",
1904 rte_pktmbuf_free(buf);
1909 /* The last WR pointer must be NULL. */
1910 (*elts)[(i - 1)].wr.next = NULL;
1911 DEBUG("%p: allocated and configured %u single-segment WRs",
1912 (void *)rxq, elts_n);
1913 rxq->elts_n = elts_n;
1915 rxq->elts.no_sp = elts;
1920 assert(pool == NULL);
1921 for (i = 0; (i != elemof(*elts)); ++i) {
1922 struct rxq_elt *elt = &(*elts)[i];
1923 struct rte_mbuf *buf;
1925 if (elt->sge.addr == 0)
1927 assert(WR_ID(elt->wr.wr_id).id == i);
1928 buf = (void *)((uintptr_t)elt->sge.addr -
1929 WR_ID(elt->wr.wr_id).offset);
1930 rte_pktmbuf_free_seg(buf);
1934 DEBUG("%p: failed, freed everything", (void *)rxq);
1940 * Free RX queue elements.
1943 * Pointer to RX queue structure.
1946 rxq_free_elts(struct rxq *rxq)
1949 unsigned int elts_n = rxq->elts_n;
1950 struct rxq_elt (*elts)[elts_n] = rxq->elts.no_sp;
1952 DEBUG("%p: freeing WRs", (void *)rxq);
1954 rxq->elts.no_sp = NULL;
1957 for (i = 0; (i != elemof(*elts)); ++i) {
1958 struct rxq_elt *elt = &(*elts)[i];
1959 struct rte_mbuf *buf;
1961 if (elt->sge.addr == 0)
1963 assert(WR_ID(elt->wr.wr_id).id == i);
1964 buf = (void *)((uintptr_t)elt->sge.addr -
1965 WR_ID(elt->wr.wr_id).offset);
1966 rte_pktmbuf_free_seg(buf);
1972 * Delete flow steering rule.
1975 * Pointer to RX queue structure.
1977 * MAC address index.
1982 rxq_del_flow(struct rxq *rxq, unsigned int mac_index, unsigned int vlan_index)
1985 struct priv *priv = rxq->priv;
1986 const uint8_t (*mac)[ETHER_ADDR_LEN] =
1987 (const uint8_t (*)[ETHER_ADDR_LEN])
1988 priv->mac[mac_index].addr_bytes;
1990 assert(rxq->mac_flow[mac_index][vlan_index] != NULL);
1991 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
1992 " (VLAN ID %" PRIu16 ")",
1994 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
1995 mac_index, priv->vlan_filter[vlan_index].id);
1996 claim_zero(ibv_destroy_flow(rxq->mac_flow[mac_index][vlan_index]));
1997 rxq->mac_flow[mac_index][vlan_index] = NULL;
2001 * Unregister a MAC address from a RX queue.
2004 * Pointer to RX queue structure.
2006 * MAC address index.
2009 rxq_mac_addr_del(struct rxq *rxq, unsigned int mac_index)
2011 struct priv *priv = rxq->priv;
2013 unsigned int vlans = 0;
2015 assert(mac_index < elemof(priv->mac));
2016 if (!BITFIELD_ISSET(rxq->mac_configured, mac_index))
2018 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
2019 if (!priv->vlan_filter[i].enabled)
2021 rxq_del_flow(rxq, mac_index, i);
2025 rxq_del_flow(rxq, mac_index, 0);
2027 BITFIELD_RESET(rxq->mac_configured, mac_index);
2031 * Unregister all MAC addresses from a RX queue.
2034 * Pointer to RX queue structure.
2037 rxq_mac_addrs_del(struct rxq *rxq)
2039 struct priv *priv = rxq->priv;
2042 for (i = 0; (i != elemof(priv->mac)); ++i)
2043 rxq_mac_addr_del(rxq, i);
2046 static int rxq_promiscuous_enable(struct rxq *);
2047 static void rxq_promiscuous_disable(struct rxq *);
2050 * Add single flow steering rule.
2053 * Pointer to RX queue structure.
2055 * MAC address index to register.
2057 * VLAN index. Use -1 for a flow without VLAN.
2060 * 0 on success, errno value on failure.
2063 rxq_add_flow(struct rxq *rxq, unsigned int mac_index, unsigned int vlan_index)
2065 struct ibv_flow *flow;
2066 struct priv *priv = rxq->priv;
2067 const uint8_t (*mac)[ETHER_ADDR_LEN] =
2068 (const uint8_t (*)[ETHER_ADDR_LEN])
2069 priv->mac[mac_index].addr_bytes;
2071 /* Allocate flow specification on the stack. */
2072 struct __attribute__((packed)) {
2073 struct ibv_flow_attr attr;
2074 struct ibv_flow_spec_eth spec;
2076 struct ibv_flow_attr *attr = &data.attr;
2077 struct ibv_flow_spec_eth *spec = &data.spec;
2079 assert(mac_index < elemof(priv->mac));
2080 assert((vlan_index < elemof(priv->vlan_filter)) || (vlan_index == -1u));
2082 * No padding must be inserted by the compiler between attr and spec.
2083 * This layout is expected by libibverbs.
2085 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
2086 *attr = (struct ibv_flow_attr){
2087 .type = IBV_FLOW_ATTR_NORMAL,
2092 *spec = (struct ibv_flow_spec_eth){
2093 .type = IBV_FLOW_SPEC_ETH,
2094 .size = sizeof(*spec),
2097 (*mac)[0], (*mac)[1], (*mac)[2],
2098 (*mac)[3], (*mac)[4], (*mac)[5]
2100 .vlan_tag = ((vlan_index != -1u) ?
2101 htons(priv->vlan_filter[vlan_index].id) :
2105 .dst_mac = "\xff\xff\xff\xff\xff\xff",
2106 .vlan_tag = ((vlan_index != -1u) ? htons(0xfff) : 0),
2109 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
2110 " (VLAN %s %" PRIu16 ")",
2112 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
2114 ((vlan_index != -1u) ? "ID" : "index"),
2115 ((vlan_index != -1u) ? priv->vlan_filter[vlan_index].id : -1u));
2116 /* Create related flow. */
2118 flow = ibv_create_flow(rxq->qp, attr);
2120 /* It's not clear whether errno is always set in this case. */
2121 ERROR("%p: flow configuration failed, errno=%d: %s",
2123 (errno ? strerror(errno) : "Unknown error"));
2128 if (vlan_index == -1u)
2130 assert(rxq->mac_flow[mac_index][vlan_index] == NULL);
2131 rxq->mac_flow[mac_index][vlan_index] = flow;
2136 * Register a MAC address in a RX queue.
2139 * Pointer to RX queue structure.
2141 * MAC address index to register.
2144 * 0 on success, errno value on failure.
2147 rxq_mac_addr_add(struct rxq *rxq, unsigned int mac_index)
2149 struct priv *priv = rxq->priv;
2151 unsigned int vlans = 0;
2154 assert(mac_index < elemof(priv->mac));
2155 if (BITFIELD_ISSET(rxq->mac_configured, mac_index))
2156 rxq_mac_addr_del(rxq, mac_index);
2157 /* Fill VLAN specifications. */
2158 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
2159 if (!priv->vlan_filter[i].enabled)
2161 /* Create related flow. */
2162 ret = rxq_add_flow(rxq, mac_index, i);
2167 /* Failure, rollback. */
2169 if (priv->vlan_filter[--i].enabled)
2170 rxq_del_flow(rxq, mac_index, i);
2174 /* In case there is no VLAN filter. */
2176 ret = rxq_add_flow(rxq, mac_index, -1);
2180 BITFIELD_SET(rxq->mac_configured, mac_index);
2185 * Register all MAC addresses in a RX queue.
2188 * Pointer to RX queue structure.
2191 * 0 on success, errno value on failure.
2194 rxq_mac_addrs_add(struct rxq *rxq)
2196 struct priv *priv = rxq->priv;
2200 for (i = 0; (i != elemof(priv->mac)); ++i) {
2201 if (!BITFIELD_ISSET(priv->mac_configured, i))
2203 ret = rxq_mac_addr_add(rxq, i);
2206 /* Failure, rollback. */
2208 rxq_mac_addr_del(rxq, --i);
2216 * Unregister a MAC address.
2218 * In RSS mode, the MAC address is unregistered from the parent queue,
2219 * otherwise it is unregistered from each queue directly.
2222 * Pointer to private structure.
2224 * MAC address index.
2227 priv_mac_addr_del(struct priv *priv, unsigned int mac_index)
2231 assert(mac_index < elemof(priv->mac));
2232 if (!BITFIELD_ISSET(priv->mac_configured, mac_index))
2235 rxq_mac_addr_del(&priv->rxq_parent, mac_index);
2238 for (i = 0; (i != priv->dev->data->nb_rx_queues); ++i)
2239 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2241 BITFIELD_RESET(priv->mac_configured, mac_index);
2245 * Register a MAC address.
2247 * In RSS mode, the MAC address is registered in the parent queue,
2248 * otherwise it is registered in each queue directly.
2251 * Pointer to private structure.
2253 * MAC address index to use.
2255 * MAC address to register.
2258 * 0 on success, errno value on failure.
2261 priv_mac_addr_add(struct priv *priv, unsigned int mac_index,
2262 const uint8_t (*mac)[ETHER_ADDR_LEN])
2267 assert(mac_index < elemof(priv->mac));
2268 /* First, make sure this address isn't already configured. */
2269 for (i = 0; (i != elemof(priv->mac)); ++i) {
2270 /* Skip this index, it's going to be reconfigured. */
2273 if (!BITFIELD_ISSET(priv->mac_configured, i))
2275 if (memcmp(priv->mac[i].addr_bytes, *mac, sizeof(*mac)))
2277 /* Address already configured elsewhere, return with error. */
2280 if (BITFIELD_ISSET(priv->mac_configured, mac_index))
2281 priv_mac_addr_del(priv, mac_index);
2282 priv->mac[mac_index] = (struct ether_addr){
2284 (*mac)[0], (*mac)[1], (*mac)[2],
2285 (*mac)[3], (*mac)[4], (*mac)[5]
2288 /* If device isn't started, this is all we need to do. */
2289 if (!priv->started) {
2291 /* Verify that all queues have this index disabled. */
2292 for (i = 0; (i != priv->rxqs_n); ++i) {
2293 if ((*priv->rxqs)[i] == NULL)
2295 assert(!BITFIELD_ISSET
2296 ((*priv->rxqs)[i]->mac_configured, mac_index));
2302 ret = rxq_mac_addr_add(&priv->rxq_parent, mac_index);
2307 for (i = 0; (i != priv->rxqs_n); ++i) {
2308 if ((*priv->rxqs)[i] == NULL)
2310 ret = rxq_mac_addr_add((*priv->rxqs)[i], mac_index);
2313 /* Failure, rollback. */
2315 if ((*priv->rxqs)[(--i)] != NULL)
2316 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2320 BITFIELD_SET(priv->mac_configured, mac_index);
2325 * Enable allmulti mode in a RX queue.
2328 * Pointer to RX queue structure.
2331 * 0 on success, errno value on failure.
2334 rxq_allmulticast_enable(struct rxq *rxq)
2336 struct ibv_flow *flow;
2337 struct ibv_flow_attr attr = {
2338 .type = IBV_FLOW_ATTR_MC_DEFAULT,
2340 .port = rxq->priv->port,
2344 DEBUG("%p: enabling allmulticast mode", (void *)rxq);
2345 if (rxq->allmulti_flow != NULL)
2348 flow = ibv_create_flow(rxq->qp, &attr);
2350 /* It's not clear whether errno is always set in this case. */
2351 ERROR("%p: flow configuration failed, errno=%d: %s",
2353 (errno ? strerror(errno) : "Unknown error"));
2358 rxq->allmulti_flow = flow;
2359 DEBUG("%p: allmulticast mode enabled", (void *)rxq);
2364 * Disable allmulti mode in a RX queue.
2367 * Pointer to RX queue structure.
2370 rxq_allmulticast_disable(struct rxq *rxq)
2372 DEBUG("%p: disabling allmulticast mode", (void *)rxq);
2373 if (rxq->allmulti_flow == NULL)
2375 claim_zero(ibv_destroy_flow(rxq->allmulti_flow));
2376 rxq->allmulti_flow = NULL;
2377 DEBUG("%p: allmulticast mode disabled", (void *)rxq);
2381 * Enable promiscuous mode in a RX queue.
2384 * Pointer to RX queue structure.
2387 * 0 on success, errno value on failure.
2390 rxq_promiscuous_enable(struct rxq *rxq)
2392 struct ibv_flow *flow;
2393 struct ibv_flow_attr attr = {
2394 .type = IBV_FLOW_ATTR_ALL_DEFAULT,
2396 .port = rxq->priv->port,
2402 DEBUG("%p: enabling promiscuous mode", (void *)rxq);
2403 if (rxq->promisc_flow != NULL)
2406 flow = ibv_create_flow(rxq->qp, &attr);
2408 /* It's not clear whether errno is always set in this case. */
2409 ERROR("%p: flow configuration failed, errno=%d: %s",
2411 (errno ? strerror(errno) : "Unknown error"));
2416 rxq->promisc_flow = flow;
2417 DEBUG("%p: promiscuous mode enabled", (void *)rxq);
2422 * Disable promiscuous mode in a RX queue.
2425 * Pointer to RX queue structure.
2428 rxq_promiscuous_disable(struct rxq *rxq)
2432 DEBUG("%p: disabling promiscuous mode", (void *)rxq);
2433 if (rxq->promisc_flow == NULL)
2435 claim_zero(ibv_destroy_flow(rxq->promisc_flow));
2436 rxq->promisc_flow = NULL;
2437 DEBUG("%p: promiscuous mode disabled", (void *)rxq);
2441 * Clean up a RX queue.
2443 * Destroy objects, free allocated memory and reset the structure for reuse.
2446 * Pointer to RX queue structure.
2449 rxq_cleanup(struct rxq *rxq)
2451 struct ibv_exp_release_intf_params params;
2453 DEBUG("cleaning up %p", (void *)rxq);
2455 rxq_free_elts_sp(rxq);
2458 if (rxq->if_qp != NULL) {
2459 assert(rxq->priv != NULL);
2460 assert(rxq->priv->ctx != NULL);
2461 assert(rxq->qp != NULL);
2462 params = (struct ibv_exp_release_intf_params){
2465 claim_zero(ibv_exp_release_intf(rxq->priv->ctx,
2469 if (rxq->if_cq != NULL) {
2470 assert(rxq->priv != NULL);
2471 assert(rxq->priv->ctx != NULL);
2472 assert(rxq->cq != NULL);
2473 params = (struct ibv_exp_release_intf_params){
2476 claim_zero(ibv_exp_release_intf(rxq->priv->ctx,
2480 if (rxq->qp != NULL) {
2481 rxq_promiscuous_disable(rxq);
2482 rxq_allmulticast_disable(rxq);
2483 rxq_mac_addrs_del(rxq);
2484 claim_zero(ibv_destroy_qp(rxq->qp));
2486 if (rxq->cq != NULL)
2487 claim_zero(ibv_destroy_cq(rxq->cq));
2488 if (rxq->rd != NULL) {
2489 struct ibv_exp_destroy_res_domain_attr attr = {
2493 assert(rxq->priv != NULL);
2494 assert(rxq->priv->ctx != NULL);
2495 claim_zero(ibv_exp_destroy_res_domain(rxq->priv->ctx,
2499 if (rxq->mr != NULL)
2500 claim_zero(ibv_dereg_mr(rxq->mr));
2501 memset(rxq, 0, sizeof(*rxq));
2505 * Translate RX completion flags to packet type.
2508 * RX completion flags returned by poll_length_flags().
2511 * Packet type for struct rte_mbuf.
2513 static inline uint32_t
2514 rxq_cq_to_pkt_type(uint32_t flags)
2518 if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
2521 IBV_EXP_CQ_RX_OUTER_IPV4_PACKET, RTE_PTYPE_L3_IPV4) |
2523 IBV_EXP_CQ_RX_OUTER_IPV6_PACKET, RTE_PTYPE_L3_IPV6) |
2525 IBV_EXP_CQ_RX_IPV4_PACKET, RTE_PTYPE_INNER_L3_IPV4) |
2527 IBV_EXP_CQ_RX_IPV6_PACKET, RTE_PTYPE_INNER_L3_IPV6);
2531 IBV_EXP_CQ_RX_IPV4_PACKET, RTE_PTYPE_L3_IPV4) |
2533 IBV_EXP_CQ_RX_IPV6_PACKET, RTE_PTYPE_L3_IPV6);
2538 * Translate RX completion flags to offload flags.
2541 * Pointer to RX queue structure.
2543 * RX completion flags returned by poll_length_flags().
2546 * Offload flags (ol_flags) for struct rte_mbuf.
2548 static inline uint32_t
2549 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
2551 uint32_t ol_flags = 0;
2556 IBV_EXP_CQ_RX_IP_CSUM_OK,
2557 PKT_RX_IP_CKSUM_BAD) |
2559 IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
2560 PKT_RX_L4_CKSUM_BAD);
2562 * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
2563 * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
2566 if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
2569 IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
2570 PKT_RX_IP_CKSUM_BAD) |
2572 IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
2573 PKT_RX_L4_CKSUM_BAD);
2578 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
2581 * DPDK callback for RX with scattered packets support.
2584 * Generic pointer to RX queue structure.
2586 * Array to store received packets.
2588 * Maximum number of packets in array.
2591 * Number of packets successfully received (<= pkts_n).
2594 mlx4_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2596 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2597 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
2598 const unsigned int elts_n = rxq->elts_n;
2599 unsigned int elts_head = rxq->elts_head;
2600 struct ibv_recv_wr head;
2601 struct ibv_recv_wr **next = &head.next;
2602 struct ibv_recv_wr *bad_wr;
2604 unsigned int pkts_ret = 0;
2607 if (unlikely(!rxq->sp))
2608 return mlx4_rx_burst(dpdk_rxq, pkts, pkts_n);
2609 if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
2611 for (i = 0; (i != pkts_n); ++i) {
2612 struct rxq_elt_sp *elt = &(*elts)[elts_head];
2613 struct ibv_recv_wr *wr = &elt->wr;
2614 uint64_t wr_id = wr->wr_id;
2616 unsigned int pkt_buf_len;
2617 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
2618 struct rte_mbuf **pkt_buf_next = &pkt_buf;
2619 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
2623 /* Sanity checks. */
2627 assert(wr_id < rxq->elts_n);
2628 assert(wr->sg_list == elt->sges);
2629 assert(wr->num_sge == elemof(elt->sges));
2630 assert(elts_head < rxq->elts_n);
2631 assert(rxq->elts_head < rxq->elts_n);
2632 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
2634 if (unlikely(ret < 0)) {
2638 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
2640 /* ibv_poll_cq() must be used in case of failure. */
2641 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
2642 if (unlikely(wcs_n == 0))
2644 if (unlikely(wcs_n < 0)) {
2645 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
2646 (void *)rxq, wcs_n);
2650 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
2651 /* Whatever, just repost the offending WR. */
2652 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
2653 " completion status (%d): %s",
2654 (void *)rxq, wc.wr_id, wc.status,
2655 ibv_wc_status_str(wc.status));
2656 #ifdef MLX4_PMD_SOFT_COUNTERS
2657 /* Increment dropped packets counter. */
2658 ++rxq->stats.idropped;
2660 /* Link completed WRs together for repost. */
2671 /* Link completed WRs together for repost. */
2675 * Replace spent segments with new ones, concatenate and
2676 * return them as pkt_buf.
2679 struct ibv_sge *sge = &elt->sges[j];
2680 struct rte_mbuf *seg = elt->bufs[j];
2681 struct rte_mbuf *rep;
2682 unsigned int seg_tailroom;
2685 * Fetch initial bytes of packet descriptor into a
2686 * cacheline while allocating rep.
2689 rep = __rte_mbuf_raw_alloc(rxq->mp);
2690 if (unlikely(rep == NULL)) {
2692 * Unable to allocate a replacement mbuf,
2695 DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
2696 " can't allocate a new mbuf",
2697 (void *)rxq, wr_id);
2698 if (pkt_buf != NULL) {
2699 *pkt_buf_next = NULL;
2700 rte_pktmbuf_free(pkt_buf);
2702 /* Increase out of memory counters. */
2703 ++rxq->stats.rx_nombuf;
2704 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2708 /* Poison user-modifiable fields in rep. */
2709 NEXT(rep) = (void *)((uintptr_t)-1);
2710 SET_DATA_OFF(rep, 0xdead);
2711 DATA_LEN(rep) = 0xd00d;
2712 PKT_LEN(rep) = 0xdeadd00d;
2713 NB_SEGS(rep) = 0x2a;
2717 assert(rep->buf_len == seg->buf_len);
2718 assert(rep->buf_len == rxq->mb_len);
2719 /* Reconfigure sge to use rep instead of seg. */
2720 assert(sge->lkey == rxq->mr->lkey);
2721 sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
2724 /* Update pkt_buf if it's the first segment, or link
2725 * seg to the previous one and update pkt_buf_next. */
2726 *pkt_buf_next = seg;
2727 pkt_buf_next = &NEXT(seg);
2728 /* Update seg information. */
2729 seg_tailroom = (seg->buf_len - seg_headroom);
2730 assert(sge->length == seg_tailroom);
2731 SET_DATA_OFF(seg, seg_headroom);
2732 if (likely(len <= seg_tailroom)) {
2734 DATA_LEN(seg) = len;
2737 assert(rte_pktmbuf_headroom(seg) ==
2739 assert(rte_pktmbuf_tailroom(seg) ==
2740 (seg_tailroom - len));
2743 DATA_LEN(seg) = seg_tailroom;
2744 PKT_LEN(seg) = seg_tailroom;
2746 assert(rte_pktmbuf_headroom(seg) == seg_headroom);
2747 assert(rte_pktmbuf_tailroom(seg) == 0);
2748 /* Fix len and clear headroom for next segments. */
2749 len -= seg_tailroom;
2752 /* Update head and tail segments. */
2753 *pkt_buf_next = NULL;
2754 assert(pkt_buf != NULL);
2756 NB_SEGS(pkt_buf) = j;
2757 PORT(pkt_buf) = rxq->port_id;
2758 PKT_LEN(pkt_buf) = pkt_buf_len;
2759 pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
2760 pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
2762 /* Return packet. */
2763 *(pkts++) = pkt_buf;
2765 #ifdef MLX4_PMD_SOFT_COUNTERS
2766 /* Increase bytes counter. */
2767 rxq->stats.ibytes += pkt_buf_len;
2770 if (++elts_head >= elts_n)
2774 if (unlikely(i == 0))
2779 DEBUG("%p: reposting %d WRs", (void *)rxq, i);
2781 ret = ibv_post_recv(rxq->qp, head.next, &bad_wr);
2782 if (unlikely(ret)) {
2783 /* Inability to repost WRs is fatal. */
2784 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
2790 rxq->elts_head = elts_head;
2791 #ifdef MLX4_PMD_SOFT_COUNTERS
2792 /* Increase packets counter. */
2793 rxq->stats.ipackets += pkts_ret;
2799 * DPDK callback for RX.
2801 * The following function is the same as mlx4_rx_burst_sp(), except it doesn't
2802 * manage scattered packets. Improves performance when MRU is lower than the
2803 * size of the first segment.
2806 * Generic pointer to RX queue structure.
2808 * Array to store received packets.
2810 * Maximum number of packets in array.
2813 * Number of packets successfully received (<= pkts_n).
2816 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2818 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2819 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
2820 const unsigned int elts_n = rxq->elts_n;
2821 unsigned int elts_head = rxq->elts_head;
2822 struct ibv_sge sges[pkts_n];
2824 unsigned int pkts_ret = 0;
2827 if (unlikely(rxq->sp))
2828 return mlx4_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
2829 for (i = 0; (i != pkts_n); ++i) {
2830 struct rxq_elt *elt = &(*elts)[elts_head];
2831 struct ibv_recv_wr *wr = &elt->wr;
2832 uint64_t wr_id = wr->wr_id;
2834 struct rte_mbuf *seg = (void *)((uintptr_t)elt->sge.addr -
2835 WR_ID(wr_id).offset);
2836 struct rte_mbuf *rep;
2839 /* Sanity checks. */
2840 assert(WR_ID(wr_id).id < rxq->elts_n);
2841 assert(wr->sg_list == &elt->sge);
2842 assert(wr->num_sge == 1);
2843 assert(elts_head < rxq->elts_n);
2844 assert(rxq->elts_head < rxq->elts_n);
2846 * Fetch initial bytes of packet descriptor into a
2847 * cacheline while allocating rep.
2850 rte_prefetch0(&seg->cacheline1);
2851 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
2853 if (unlikely(ret < 0)) {
2857 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
2859 /* ibv_poll_cq() must be used in case of failure. */
2860 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
2861 if (unlikely(wcs_n == 0))
2863 if (unlikely(wcs_n < 0)) {
2864 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
2865 (void *)rxq, wcs_n);
2869 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
2870 /* Whatever, just repost the offending WR. */
2871 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
2872 " completion status (%d): %s",
2873 (void *)rxq, wc.wr_id, wc.status,
2874 ibv_wc_status_str(wc.status));
2875 #ifdef MLX4_PMD_SOFT_COUNTERS
2876 /* Increment dropped packets counter. */
2877 ++rxq->stats.idropped;
2879 /* Add SGE to array for repost. */
2888 rep = __rte_mbuf_raw_alloc(rxq->mp);
2889 if (unlikely(rep == NULL)) {
2891 * Unable to allocate a replacement mbuf,
2894 DEBUG("rxq=%p, wr_id=%" PRIu32 ":"
2895 " can't allocate a new mbuf",
2896 (void *)rxq, WR_ID(wr_id).id);
2897 /* Increase out of memory counters. */
2898 ++rxq->stats.rx_nombuf;
2899 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2903 /* Reconfigure sge to use rep instead of seg. */
2904 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
2905 assert(elt->sge.lkey == rxq->mr->lkey);
2906 WR_ID(wr->wr_id).offset =
2907 (((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
2909 assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
2911 /* Add SGE to array for repost. */
2914 /* Update seg information. */
2915 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
2917 PORT(seg) = rxq->port_id;
2920 DATA_LEN(seg) = len;
2921 seg->packet_type = rxq_cq_to_pkt_type(flags);
2922 seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
2924 /* Return packet. */
2927 #ifdef MLX4_PMD_SOFT_COUNTERS
2928 /* Increase bytes counter. */
2929 rxq->stats.ibytes += len;
2932 if (++elts_head >= elts_n)
2936 if (unlikely(i == 0))
2940 DEBUG("%p: reposting %u WRs", (void *)rxq, i);
2942 ret = rxq->if_qp->recv_burst(rxq->qp, sges, i);
2943 if (unlikely(ret)) {
2944 /* Inability to repost WRs is fatal. */
2945 DEBUG("%p: recv_burst(): failed (ret=%d)",
2950 rxq->elts_head = elts_head;
2951 #ifdef MLX4_PMD_SOFT_COUNTERS
2952 /* Increase packets counter. */
2953 rxq->stats.ipackets += pkts_ret;
2959 * Allocate a Queue Pair.
2960 * Optionally setup inline receive if supported.
2963 * Pointer to private structure.
2965 * Completion queue to associate with QP.
2967 * Number of descriptors in QP (hint only).
2970 * QP pointer or NULL in case of error.
2972 static struct ibv_qp *
2973 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
2974 struct ibv_exp_res_domain *rd)
2976 struct ibv_exp_qp_init_attr attr = {
2977 /* CQ to be associated with the send queue. */
2979 /* CQ to be associated with the receive queue. */
2982 /* Max number of outstanding WRs. */
2983 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2984 priv->device_attr.max_qp_wr :
2986 /* Max number of scatter/gather elements in a WR. */
2987 .max_recv_sge = ((priv->device_attr.max_sge <
2988 MLX4_PMD_SGE_WR_N) ?
2989 priv->device_attr.max_sge :
2992 .qp_type = IBV_QPT_RAW_PACKET,
2993 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
2994 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
3000 attr.max_inl_recv = priv->inl_recv_size;
3001 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
3003 return ibv_exp_create_qp(priv->ctx, &attr);
3009 * Allocate a RSS Queue Pair.
3010 * Optionally setup inline receive if supported.
3013 * Pointer to private structure.
3015 * Completion queue to associate with QP.
3017 * Number of descriptors in QP (hint only).
3019 * If nonzero, create a parent QP, otherwise a child.
3022 * QP pointer or NULL in case of error.
3024 static struct ibv_qp *
3025 rxq_setup_qp_rss(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
3026 int parent, struct ibv_exp_res_domain *rd)
3028 struct ibv_exp_qp_init_attr attr = {
3029 /* CQ to be associated with the send queue. */
3031 /* CQ to be associated with the receive queue. */
3034 /* Max number of outstanding WRs. */
3035 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
3036 priv->device_attr.max_qp_wr :
3038 /* Max number of scatter/gather elements in a WR. */
3039 .max_recv_sge = ((priv->device_attr.max_sge <
3040 MLX4_PMD_SGE_WR_N) ?
3041 priv->device_attr.max_sge :
3044 .qp_type = IBV_QPT_RAW_PACKET,
3045 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
3046 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN |
3047 IBV_EXP_QP_INIT_ATTR_QPG),
3053 attr.max_inl_recv = priv->inl_recv_size,
3054 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
3057 attr.qpg.qpg_type = IBV_EXP_QPG_PARENT;
3058 /* TSS isn't necessary. */
3059 attr.qpg.parent_attrib.tss_child_count = 0;
3060 attr.qpg.parent_attrib.rss_child_count = priv->rxqs_n;
3061 DEBUG("initializing parent RSS queue");
3063 attr.qpg.qpg_type = IBV_EXP_QPG_CHILD_RX;
3064 attr.qpg.qpg_parent = priv->rxq_parent.qp;
3065 DEBUG("initializing child RSS queue");
3067 return ibv_exp_create_qp(priv->ctx, &attr);
3070 #endif /* RSS_SUPPORT */
3073 * Reconfigure a RX queue with new parameters.
3075 * rxq_rehash() does not allocate mbufs, which, if not done from the right
3076 * thread (such as a control thread), may corrupt the pool.
3077 * In case of failure, the queue is left untouched.
3080 * Pointer to Ethernet device structure.
3085 * 0 on success, errno value on failure.
3088 rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
3090 struct priv *priv = rxq->priv;
3091 struct rxq tmpl = *rxq;
3092 unsigned int mbuf_n;
3093 unsigned int desc_n;
3094 struct rte_mbuf **pool;
3096 struct ibv_exp_qp_attr mod;
3097 struct ibv_recv_wr *bad_wr;
3099 int parent = (rxq == &priv->rxq_parent);
3102 ERROR("%p: cannot rehash parent queue %p",
3103 (void *)dev, (void *)rxq);
3106 DEBUG("%p: rehashing queue %p", (void *)dev, (void *)rxq);
3107 /* Number of descriptors and mbufs currently allocated. */
3108 desc_n = (tmpl.elts_n * (tmpl.sp ? MLX4_PMD_SGE_WR_N : 1));
3110 /* Toggle RX checksum offload if hardware supports it. */
3111 if (priv->hw_csum) {
3112 tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3113 rxq->csum = tmpl.csum;
3115 if (priv->hw_csum_l2tun) {
3116 tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3117 rxq->csum_l2tun = tmpl.csum_l2tun;
3119 /* Enable scattered packets support for this queue if necessary. */
3120 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3121 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3122 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3124 desc_n /= MLX4_PMD_SGE_WR_N;
3127 DEBUG("%p: %s scattered packets support (%u WRs)",
3128 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc_n);
3129 /* If scatter mode is the same as before, nothing to do. */
3130 if (tmpl.sp == rxq->sp) {
3131 DEBUG("%p: nothing to do", (void *)dev);
3134 /* Remove attached flows if RSS is disabled (no parent queue). */
3136 rxq_allmulticast_disable(&tmpl);
3137 rxq_promiscuous_disable(&tmpl);
3138 rxq_mac_addrs_del(&tmpl);
3139 /* Update original queue in case of failure. */
3140 rxq->allmulti_flow = tmpl.allmulti_flow;
3141 rxq->promisc_flow = tmpl.promisc_flow;
3142 memcpy(rxq->mac_configured, tmpl.mac_configured,
3143 sizeof(rxq->mac_configured));
3144 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
3146 /* From now on, any failure will render the queue unusable.
3147 * Reinitialize QP. */
3148 mod = (struct ibv_exp_qp_attr){ .qp_state = IBV_QPS_RESET };
3149 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3151 ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
3155 err = ibv_resize_cq(tmpl.cq, desc_n);
3157 ERROR("%p: cannot resize CQ: %s", (void *)dev, strerror(err));
3161 mod = (struct ibv_exp_qp_attr){
3162 /* Move the QP to this state. */
3163 .qp_state = IBV_QPS_INIT,
3164 /* Primary port number. */
3165 .port_num = priv->port
3167 err = ibv_exp_modify_qp(tmpl.qp, &mod,
3170 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3171 #endif /* RSS_SUPPORT */
3174 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3175 (void *)dev, strerror(err));
3179 /* Reconfigure flows. Do not care for errors. */
3181 rxq_mac_addrs_add(&tmpl);
3183 rxq_promiscuous_enable(&tmpl);
3185 rxq_allmulticast_enable(&tmpl);
3186 /* Update original queue in case of failure. */
3187 rxq->allmulti_flow = tmpl.allmulti_flow;
3188 rxq->promisc_flow = tmpl.promisc_flow;
3189 memcpy(rxq->mac_configured, tmpl.mac_configured,
3190 sizeof(rxq->mac_configured));
3191 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
3193 /* Allocate pool. */
3194 pool = rte_malloc(__func__, (mbuf_n * sizeof(*pool)), 0);
3196 ERROR("%p: cannot allocate memory", (void *)dev);
3199 /* Snatch mbufs from original queue. */
3202 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
3204 for (i = 0; (i != elemof(*elts)); ++i) {
3205 struct rxq_elt_sp *elt = &(*elts)[i];
3208 for (j = 0; (j != elemof(elt->bufs)); ++j) {
3209 assert(elt->bufs[j] != NULL);
3210 pool[k++] = elt->bufs[j];
3214 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
3216 for (i = 0; (i != elemof(*elts)); ++i) {
3217 struct rxq_elt *elt = &(*elts)[i];
3218 struct rte_mbuf *buf = (void *)
3219 ((uintptr_t)elt->sge.addr -
3220 WR_ID(elt->wr.wr_id).offset);
3222 assert(WR_ID(elt->wr.wr_id).id == i);
3226 assert(k == mbuf_n);
3228 tmpl.elts.sp = NULL;
3229 assert((void *)&tmpl.elts.sp == (void *)&tmpl.elts.no_sp);
3231 rxq_alloc_elts_sp(&tmpl, desc_n, pool) :
3232 rxq_alloc_elts(&tmpl, desc_n, pool));
3234 ERROR("%p: cannot reallocate WRs, aborting", (void *)dev);
3239 assert(tmpl.elts_n == desc_n);
3240 assert(tmpl.elts.sp != NULL);
3242 /* Clean up original data. */
3244 rte_free(rxq->elts.sp);
3245 rxq->elts.sp = NULL;
3247 err = ibv_post_recv(tmpl.qp,
3249 &(*tmpl.elts.sp)[0].wr :
3250 &(*tmpl.elts.no_sp)[0].wr),
3253 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3259 mod = (struct ibv_exp_qp_attr){
3260 .qp_state = IBV_QPS_RTR
3262 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3264 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3265 (void *)dev, strerror(err));
3273 * Configure a RX queue.
3276 * Pointer to Ethernet device structure.
3278 * Pointer to RX queue structure.
3280 * Number of descriptors to configure in queue.
3282 * NUMA socket on which memory must be allocated.
3284 * Thresholds parameters.
3286 * Memory pool for buffer allocations.
3289 * 0 on success, errno value on failure.
3292 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
3293 unsigned int socket, const struct rte_eth_rxconf *conf,
3294 struct rte_mempool *mp)
3296 struct priv *priv = dev->data->dev_private;
3302 struct ibv_exp_qp_attr mod;
3304 struct ibv_exp_query_intf_params params;
3305 struct ibv_exp_cq_init_attr cq;
3306 struct ibv_exp_res_domain_init_attr rd;
3308 enum ibv_exp_query_intf_status status;
3309 struct ibv_recv_wr *bad_wr;
3310 struct rte_mbuf *buf;
3312 int parent = (rxq == &priv->rxq_parent);
3314 (void)conf; /* Thresholds configuration (ignored). */
3316 * If this is a parent queue, hardware must support RSS and
3317 * RSS must be enabled.
3319 assert((!parent) || ((priv->hw_rss) && (priv->rss)));
3321 /* Even if unused, ibv_create_cq() requires at least one
3326 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
3327 ERROR("%p: invalid number of RX descriptors (must be a"
3328 " multiple of %d)", (void *)dev, MLX4_PMD_SGE_WR_N);
3331 /* Get mbuf length. */
3332 buf = rte_pktmbuf_alloc(mp);
3334 ERROR("%p: unable to allocate mbuf", (void *)dev);
3337 tmpl.mb_len = buf->buf_len;
3338 assert((rte_pktmbuf_headroom(buf) +
3339 rte_pktmbuf_tailroom(buf)) == tmpl.mb_len);
3340 assert(rte_pktmbuf_headroom(buf) == RTE_PKTMBUF_HEADROOM);
3341 rte_pktmbuf_free(buf);
3342 /* Toggle RX checksum offload if hardware supports it. */
3344 tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3345 if (priv->hw_csum_l2tun)
3346 tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3347 /* Enable scattered packets support for this queue if necessary. */
3348 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3349 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3350 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3352 desc /= MLX4_PMD_SGE_WR_N;
3354 DEBUG("%p: %s scattered packets support (%u WRs)",
3355 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc);
3356 /* Use the entire RX mempool as the memory region. */
3357 tmpl.mr = ibv_reg_mr(priv->pd,
3358 (void *)mp->elt_va_start,
3359 (mp->elt_va_end - mp->elt_va_start),
3360 (IBV_ACCESS_LOCAL_WRITE |
3361 IBV_ACCESS_REMOTE_WRITE));
3362 if (tmpl.mr == NULL) {
3364 ERROR("%p: MR creation failure: %s",
3365 (void *)dev, strerror(ret));
3369 attr.rd = (struct ibv_exp_res_domain_init_attr){
3370 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
3371 IBV_EXP_RES_DOMAIN_MSG_MODEL),
3372 .thread_model = IBV_EXP_THREAD_SINGLE,
3373 .msg_model = IBV_EXP_MSG_HIGH_BW,
3375 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
3376 if (tmpl.rd == NULL) {
3378 ERROR("%p: RD creation failure: %s",
3379 (void *)dev, strerror(ret));
3382 attr.cq = (struct ibv_exp_cq_init_attr){
3383 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
3384 .res_domain = tmpl.rd,
3386 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
3387 if (tmpl.cq == NULL) {
3389 ERROR("%p: CQ creation failure: %s",
3390 (void *)dev, strerror(ret));
3393 DEBUG("priv->device_attr.max_qp_wr is %d",
3394 priv->device_attr.max_qp_wr);
3395 DEBUG("priv->device_attr.max_sge is %d",
3396 priv->device_attr.max_sge);
3399 tmpl.qp = rxq_setup_qp_rss(priv, tmpl.cq, desc, parent,
3402 #endif /* RSS_SUPPORT */
3403 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc, tmpl.rd);
3404 if (tmpl.qp == NULL) {
3405 ret = (errno ? errno : EINVAL);
3406 ERROR("%p: QP creation failure: %s",
3407 (void *)dev, strerror(ret));
3410 mod = (struct ibv_exp_qp_attr){
3411 /* Move the QP to this state. */
3412 .qp_state = IBV_QPS_INIT,
3413 /* Primary port number. */
3414 .port_num = priv->port
3416 ret = ibv_exp_modify_qp(tmpl.qp, &mod,
3419 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3420 #endif /* RSS_SUPPORT */
3423 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3424 (void *)dev, strerror(ret));
3427 if ((parent) || (!priv->rss)) {
3428 /* Configure MAC and broadcast addresses. */
3429 ret = rxq_mac_addrs_add(&tmpl);
3431 ERROR("%p: QP flow attachment failed: %s",
3432 (void *)dev, strerror(ret));
3436 /* Allocate descriptors for RX queues, except for the RSS parent. */
3440 ret = rxq_alloc_elts_sp(&tmpl, desc, NULL);
3442 ret = rxq_alloc_elts(&tmpl, desc, NULL);
3444 ERROR("%p: RXQ allocation failed: %s",
3445 (void *)dev, strerror(ret));
3448 ret = ibv_post_recv(tmpl.qp,
3450 &(*tmpl.elts.sp)[0].wr :
3451 &(*tmpl.elts.no_sp)[0].wr),
3454 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3461 mod = (struct ibv_exp_qp_attr){
3462 .qp_state = IBV_QPS_RTR
3464 ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3466 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3467 (void *)dev, strerror(ret));
3471 tmpl.port_id = dev->data->port_id;
3472 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
3473 attr.params = (struct ibv_exp_query_intf_params){
3474 .intf_scope = IBV_EXP_INTF_GLOBAL,
3475 .intf = IBV_EXP_INTF_CQ,
3478 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
3479 if (tmpl.if_cq == NULL) {
3480 ERROR("%p: CQ interface family query failed with status %d",
3481 (void *)dev, status);
3484 attr.params = (struct ibv_exp_query_intf_params){
3485 .intf_scope = IBV_EXP_INTF_GLOBAL,
3486 .intf = IBV_EXP_INTF_QP_BURST,
3489 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
3490 if (tmpl.if_qp == NULL) {
3491 ERROR("%p: QP interface family query failed with status %d",
3492 (void *)dev, status);
3495 /* Clean up rxq in case we're reinitializing it. */
3496 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
3499 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
3509 * DPDK callback to configure a RX queue.
3512 * Pointer to Ethernet device structure.
3516 * Number of descriptors to configure in queue.
3518 * NUMA socket on which memory must be allocated.
3520 * Thresholds parameters.
3522 * Memory pool for buffer allocations.
3525 * 0 on success, negative errno value on failure.
3528 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
3529 unsigned int socket, const struct rte_eth_rxconf *conf,
3530 struct rte_mempool *mp)
3532 struct priv *priv = dev->data->dev_private;
3533 struct rxq *rxq = (*priv->rxqs)[idx];
3537 DEBUG("%p: configuring queue %u for %u descriptors",
3538 (void *)dev, idx, desc);
3539 if (idx >= priv->rxqs_n) {
3540 ERROR("%p: queue index out of range (%u >= %u)",
3541 (void *)dev, idx, priv->rxqs_n);
3546 DEBUG("%p: reusing already allocated queue index %u (%p)",
3547 (void *)dev, idx, (void *)rxq);
3548 if (priv->started) {
3552 (*priv->rxqs)[idx] = NULL;
3555 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
3557 ERROR("%p: unable to allocate queue index %u",
3563 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
3567 rxq->stats.idx = idx;
3568 DEBUG("%p: adding RX queue %p to list",
3569 (void *)dev, (void *)rxq);
3570 (*priv->rxqs)[idx] = rxq;
3571 /* Update receive callback. */
3573 dev->rx_pkt_burst = mlx4_rx_burst_sp;
3575 dev->rx_pkt_burst = mlx4_rx_burst;
3582 * DPDK callback to release a RX queue.
3585 * Generic RX queue pointer.
3588 mlx4_rx_queue_release(void *dpdk_rxq)
3590 struct rxq *rxq = (struct rxq *)dpdk_rxq;
3598 assert(rxq != &priv->rxq_parent);
3599 for (i = 0; (i != priv->rxqs_n); ++i)
3600 if ((*priv->rxqs)[i] == rxq) {
3601 DEBUG("%p: removing RX queue %p from list",
3602 (void *)priv->dev, (void *)rxq);
3603 (*priv->rxqs)[i] = NULL;
3612 priv_dev_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
3615 * DPDK callback to start the device.
3617 * Simulate device start by attaching all configured flows.
3620 * Pointer to Ethernet device structure.
3623 * 0 on success, negative errno value on failure.
3626 mlx4_dev_start(struct rte_eth_dev *dev)
3628 struct priv *priv = dev->data->dev_private;
3634 if (priv->started) {
3638 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
3641 rxq = &priv->rxq_parent;
3644 rxq = (*priv->rxqs)[0];
3647 /* Iterate only once when RSS is enabled. */
3651 /* Ignore nonexistent RX queues. */
3654 ret = rxq_mac_addrs_add(rxq);
3655 if (!ret && priv->promisc)
3656 ret = rxq_promiscuous_enable(rxq);
3657 if (!ret && priv->allmulti)
3658 ret = rxq_allmulticast_enable(rxq);
3661 WARN("%p: QP flow attachment failed: %s",
3662 (void *)dev, strerror(ret));
3665 rxq = (*priv->rxqs)[--i];
3667 rxq_allmulticast_disable(rxq);
3668 rxq_promiscuous_disable(rxq);
3669 rxq_mac_addrs_del(rxq);
3675 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3676 priv_dev_interrupt_handler_install(priv, dev);
3682 * DPDK callback to stop the device.
3684 * Simulate device stop by detaching all configured flows.
3687 * Pointer to Ethernet device structure.
3690 mlx4_dev_stop(struct rte_eth_dev *dev)
3692 struct priv *priv = dev->data->dev_private;
3698 if (!priv->started) {
3702 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
3705 rxq = &priv->rxq_parent;
3708 rxq = (*priv->rxqs)[0];
3711 /* Iterate only once when RSS is enabled. */
3713 /* Ignore nonexistent RX queues. */
3716 rxq_allmulticast_disable(rxq);
3717 rxq_promiscuous_disable(rxq);
3718 rxq_mac_addrs_del(rxq);
3719 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3724 * Dummy DPDK callback for TX.
3726 * This function is used to temporarily replace the real callback during
3727 * unsafe control operations on the queue, or in case of error.
3730 * Generic pointer to TX queue structure.
3732 * Packets to transmit.
3734 * Number of packets in array.
3737 * Number of packets successfully transmitted (<= pkts_n).
3740 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
3749 * Dummy DPDK callback for RX.
3751 * This function is used to temporarily replace the real callback during
3752 * unsafe control operations on the queue, or in case of error.
3755 * Generic pointer to RX queue structure.
3757 * Array to store received packets.
3759 * Maximum number of packets in array.
3762 * Number of packets successfully received (<= pkts_n).
3765 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
3774 priv_dev_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
3777 * DPDK callback to close the device.
3779 * Destroy all queues and objects, free memory.
3782 * Pointer to Ethernet device structure.
3785 mlx4_dev_close(struct rte_eth_dev *dev)
3787 struct priv *priv = dev->data->dev_private;
3792 DEBUG("%p: closing device \"%s\"",
3794 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
3795 /* Prevent crashes when queues are still in use. This is unfortunately
3796 * still required for DPDK 1.3 because some programs (such as testpmd)
3797 * never release them before closing the device. */
3798 dev->rx_pkt_burst = removed_rx_burst;
3799 dev->tx_pkt_burst = removed_tx_burst;
3800 if (priv->rxqs != NULL) {
3801 /* XXX race condition if mlx4_rx_burst() is still running. */
3803 for (i = 0; (i != priv->rxqs_n); ++i) {
3804 tmp = (*priv->rxqs)[i];
3807 (*priv->rxqs)[i] = NULL;
3814 if (priv->txqs != NULL) {
3815 /* XXX race condition if mlx4_tx_burst() is still running. */
3817 for (i = 0; (i != priv->txqs_n); ++i) {
3818 tmp = (*priv->txqs)[i];
3821 (*priv->txqs)[i] = NULL;
3829 rxq_cleanup(&priv->rxq_parent);
3830 if (priv->pd != NULL) {
3831 assert(priv->ctx != NULL);
3832 claim_zero(ibv_dealloc_pd(priv->pd));
3833 claim_zero(ibv_close_device(priv->ctx));
3835 assert(priv->ctx == NULL);
3836 priv_dev_interrupt_handler_uninstall(priv, dev);
3838 memset(priv, 0, sizeof(*priv));
3842 * DPDK callback to get information about the device.
3845 * Pointer to Ethernet device structure.
3847 * Info structure output buffer.
3850 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
3852 struct priv *priv = dev->data->dev_private;
3854 char ifname[IF_NAMESIZE];
3857 /* FIXME: we should ask the device for these values. */
3858 info->min_rx_bufsize = 32;
3859 info->max_rx_pktlen = 65536;
3861 * Since we need one CQ per QP, the limit is the minimum number
3862 * between the two values.
3864 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
3865 priv->device_attr.max_qp : priv->device_attr.max_cq);
3866 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
3869 info->max_rx_queues = max;
3870 info->max_tx_queues = max;
3871 /* Last array entry is reserved for broadcast. */
3872 info->max_mac_addrs = (elemof(priv->mac) - 1);
3873 info->rx_offload_capa =
3875 (DEV_RX_OFFLOAD_IPV4_CKSUM |
3876 DEV_RX_OFFLOAD_UDP_CKSUM |
3877 DEV_RX_OFFLOAD_TCP_CKSUM) :
3879 info->tx_offload_capa =
3881 (DEV_TX_OFFLOAD_IPV4_CKSUM |
3882 DEV_TX_OFFLOAD_UDP_CKSUM |
3883 DEV_TX_OFFLOAD_TCP_CKSUM) :
3885 if (priv_get_ifname(priv, &ifname) == 0)
3886 info->if_index = if_nametoindex(ifname);
3891 * DPDK callback to get device statistics.
3894 * Pointer to Ethernet device structure.
3896 * Stats structure output buffer.
3899 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
3901 struct priv *priv = dev->data->dev_private;
3902 struct rte_eth_stats tmp = {0};
3907 /* Add software counters. */
3908 for (i = 0; (i != priv->rxqs_n); ++i) {
3909 struct rxq *rxq = (*priv->rxqs)[i];
3913 idx = rxq->stats.idx;
3914 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3915 #ifdef MLX4_PMD_SOFT_COUNTERS
3916 tmp.q_ipackets[idx] += rxq->stats.ipackets;
3917 tmp.q_ibytes[idx] += rxq->stats.ibytes;
3919 tmp.q_errors[idx] += (rxq->stats.idropped +
3920 rxq->stats.rx_nombuf);
3922 #ifdef MLX4_PMD_SOFT_COUNTERS
3923 tmp.ipackets += rxq->stats.ipackets;
3924 tmp.ibytes += rxq->stats.ibytes;
3926 tmp.ierrors += rxq->stats.idropped;
3927 tmp.rx_nombuf += rxq->stats.rx_nombuf;
3929 for (i = 0; (i != priv->txqs_n); ++i) {
3930 struct txq *txq = (*priv->txqs)[i];
3934 idx = txq->stats.idx;
3935 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3936 #ifdef MLX4_PMD_SOFT_COUNTERS
3937 tmp.q_opackets[idx] += txq->stats.opackets;
3938 tmp.q_obytes[idx] += txq->stats.obytes;
3940 tmp.q_errors[idx] += txq->stats.odropped;
3942 #ifdef MLX4_PMD_SOFT_COUNTERS
3943 tmp.opackets += txq->stats.opackets;
3944 tmp.obytes += txq->stats.obytes;
3946 tmp.oerrors += txq->stats.odropped;
3948 #ifndef MLX4_PMD_SOFT_COUNTERS
3949 /* FIXME: retrieve and add hardware counters. */
3956 * DPDK callback to clear device statistics.
3959 * Pointer to Ethernet device structure.
3962 mlx4_stats_reset(struct rte_eth_dev *dev)
3964 struct priv *priv = dev->data->dev_private;
3969 for (i = 0; (i != priv->rxqs_n); ++i) {
3970 if ((*priv->rxqs)[i] == NULL)
3972 idx = (*priv->rxqs)[i]->stats.idx;
3973 (*priv->rxqs)[i]->stats =
3974 (struct mlx4_rxq_stats){ .idx = idx };
3976 for (i = 0; (i != priv->txqs_n); ++i) {
3977 if ((*priv->txqs)[i] == NULL)
3979 idx = (*priv->txqs)[i]->stats.idx;
3980 (*priv->txqs)[i]->stats =
3981 (struct mlx4_txq_stats){ .idx = idx };
3983 #ifndef MLX4_PMD_SOFT_COUNTERS
3984 /* FIXME: reset hardware counters. */
3990 * DPDK callback to remove a MAC address.
3993 * Pointer to Ethernet device structure.
3995 * MAC address index.
3998 mlx4_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
4000 struct priv *priv = dev->data->dev_private;
4003 DEBUG("%p: removing MAC address from index %" PRIu32,
4004 (void *)dev, index);
4005 /* Last array entry is reserved for broadcast. */
4006 if (index >= (elemof(priv->mac) - 1))
4008 priv_mac_addr_del(priv, index);
4014 * DPDK callback to add a MAC address.
4017 * Pointer to Ethernet device structure.
4019 * MAC address to register.
4021 * MAC address index.
4023 * VMDq pool index to associate address with (ignored).
4026 mlx4_mac_addr_add(struct rte_eth_dev *dev, struct ether_addr *mac_addr,
4027 uint32_t index, uint32_t vmdq)
4029 struct priv *priv = dev->data->dev_private;
4033 DEBUG("%p: adding MAC address at index %" PRIu32,
4034 (void *)dev, index);
4035 /* Last array entry is reserved for broadcast. */
4036 if (index >= (elemof(priv->mac) - 1))
4038 priv_mac_addr_add(priv, index,
4039 (const uint8_t (*)[ETHER_ADDR_LEN])
4040 mac_addr->addr_bytes);
4046 * DPDK callback to enable promiscuous mode.
4049 * Pointer to Ethernet device structure.
4052 mlx4_promiscuous_enable(struct rte_eth_dev *dev)
4054 struct priv *priv = dev->data->dev_private;
4059 if (priv->promisc) {
4063 /* If device isn't started, this is all we need to do. */
4067 ret = rxq_promiscuous_enable(&priv->rxq_parent);
4074 for (i = 0; (i != priv->rxqs_n); ++i) {
4075 if ((*priv->rxqs)[i] == NULL)
4077 ret = rxq_promiscuous_enable((*priv->rxqs)[i]);
4080 /* Failure, rollback. */
4082 if ((*priv->rxqs)[--i] != NULL)
4083 rxq_promiscuous_disable((*priv->rxqs)[i]);
4093 * DPDK callback to disable promiscuous mode.
4096 * Pointer to Ethernet device structure.
4099 mlx4_promiscuous_disable(struct rte_eth_dev *dev)
4101 struct priv *priv = dev->data->dev_private;
4105 if (!priv->promisc) {
4110 rxq_promiscuous_disable(&priv->rxq_parent);
4113 for (i = 0; (i != priv->rxqs_n); ++i)
4114 if ((*priv->rxqs)[i] != NULL)
4115 rxq_promiscuous_disable((*priv->rxqs)[i]);
4122 * DPDK callback to enable allmulti mode.
4125 * Pointer to Ethernet device structure.
4128 mlx4_allmulticast_enable(struct rte_eth_dev *dev)
4130 struct priv *priv = dev->data->dev_private;
4135 if (priv->allmulti) {
4139 /* If device isn't started, this is all we need to do. */
4143 ret = rxq_allmulticast_enable(&priv->rxq_parent);
4150 for (i = 0; (i != priv->rxqs_n); ++i) {
4151 if ((*priv->rxqs)[i] == NULL)
4153 ret = rxq_allmulticast_enable((*priv->rxqs)[i]);
4156 /* Failure, rollback. */
4158 if ((*priv->rxqs)[--i] != NULL)
4159 rxq_allmulticast_disable((*priv->rxqs)[i]);
4169 * DPDK callback to disable allmulti mode.
4172 * Pointer to Ethernet device structure.
4175 mlx4_allmulticast_disable(struct rte_eth_dev *dev)
4177 struct priv *priv = dev->data->dev_private;
4181 if (!priv->allmulti) {
4186 rxq_allmulticast_disable(&priv->rxq_parent);
4189 for (i = 0; (i != priv->rxqs_n); ++i)
4190 if ((*priv->rxqs)[i] != NULL)
4191 rxq_allmulticast_disable((*priv->rxqs)[i]);
4198 * DPDK callback to retrieve physical link information (unlocked version).
4201 * Pointer to Ethernet device structure.
4202 * @param wait_to_complete
4203 * Wait for request completion (ignored).
4206 mlx4_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
4208 struct priv *priv = dev->data->dev_private;
4209 struct ethtool_cmd edata = {
4213 struct rte_eth_link dev_link;
4216 (void)wait_to_complete;
4217 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
4218 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
4221 memset(&dev_link, 0, sizeof(dev_link));
4222 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
4223 (ifr.ifr_flags & IFF_RUNNING));
4224 ifr.ifr_data = &edata;
4225 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4226 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
4230 link_speed = ethtool_cmd_speed(&edata);
4231 if (link_speed == -1)
4232 dev_link.link_speed = 0;
4234 dev_link.link_speed = link_speed;
4235 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
4236 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
4237 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
4238 /* Link status changed. */
4239 dev->data->dev_link = dev_link;
4242 /* Link status is still the same. */
4247 * DPDK callback to retrieve physical link information.
4250 * Pointer to Ethernet device structure.
4251 * @param wait_to_complete
4252 * Wait for request completion (ignored).
4255 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
4257 struct priv *priv = dev->data->dev_private;
4261 ret = mlx4_link_update_unlocked(dev, wait_to_complete);
4267 * DPDK callback to change the MTU.
4269 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
4270 * received). Use this as a hint to enable/disable scattered packets support
4271 * and improve performance when not needed.
4272 * Since failure is not an option, reconfiguring queues on the fly is not
4276 * Pointer to Ethernet device structure.
4281 * 0 on success, negative errno value on failure.
4284 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
4286 struct priv *priv = dev->data->dev_private;
4289 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
4293 /* Set kernel interface MTU first. */
4294 if (priv_set_mtu(priv, mtu)) {
4296 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
4300 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
4302 /* Temporarily replace RX handler with a fake one, assuming it has not
4303 * been copied elsewhere. */
4304 dev->rx_pkt_burst = removed_rx_burst;
4305 /* Make sure everyone has left mlx4_rx_burst() and uses
4306 * removed_rx_burst() instead. */
4309 /* Reconfigure each RX queue. */
4310 for (i = 0; (i != priv->rxqs_n); ++i) {
4311 struct rxq *rxq = (*priv->rxqs)[i];
4312 unsigned int max_frame_len;
4317 /* Calculate new maximum frame length according to MTU and
4318 * toggle scattered support (sp) if necessary. */
4319 max_frame_len = (priv->mtu + ETHER_HDR_LEN +
4320 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
4321 sp = (max_frame_len > (rxq->mb_len - RTE_PKTMBUF_HEADROOM));
4322 /* Provide new values to rxq_setup(). */
4323 dev->data->dev_conf.rxmode.jumbo_frame = sp;
4324 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
4325 ret = rxq_rehash(dev, rxq);
4327 /* Force SP RX if that queue requires it and abort. */
4329 rx_func = mlx4_rx_burst_sp;
4332 /* Reenable non-RSS queue attributes. No need to check
4333 * for errors at this stage. */
4335 rxq_mac_addrs_add(rxq);
4337 rxq_promiscuous_enable(rxq);
4339 rxq_allmulticast_enable(rxq);
4341 /* Scattered burst function takes priority. */
4343 rx_func = mlx4_rx_burst_sp;
4345 /* Burst functions can now be called again. */
4347 dev->rx_pkt_burst = rx_func;
4355 * DPDK callback to get flow control status.
4358 * Pointer to Ethernet device structure.
4359 * @param[out] fc_conf
4360 * Flow control output buffer.
4363 * 0 on success, negative errno value on failure.
4366 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4368 struct priv *priv = dev->data->dev_private;
4370 struct ethtool_pauseparam ethpause = {
4371 .cmd = ETHTOOL_GPAUSEPARAM
4375 ifr.ifr_data = ðpause;
4377 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4379 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
4385 fc_conf->autoneg = ethpause.autoneg;
4386 if (ethpause.rx_pause && ethpause.tx_pause)
4387 fc_conf->mode = RTE_FC_FULL;
4388 else if (ethpause.rx_pause)
4389 fc_conf->mode = RTE_FC_RX_PAUSE;
4390 else if (ethpause.tx_pause)
4391 fc_conf->mode = RTE_FC_TX_PAUSE;
4393 fc_conf->mode = RTE_FC_NONE;
4403 * DPDK callback to modify flow control parameters.
4406 * Pointer to Ethernet device structure.
4407 * @param[in] fc_conf
4408 * Flow control parameters.
4411 * 0 on success, negative errno value on failure.
4414 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4416 struct priv *priv = dev->data->dev_private;
4418 struct ethtool_pauseparam ethpause = {
4419 .cmd = ETHTOOL_SPAUSEPARAM
4423 ifr.ifr_data = ðpause;
4424 ethpause.autoneg = fc_conf->autoneg;
4425 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4426 (fc_conf->mode & RTE_FC_RX_PAUSE))
4427 ethpause.rx_pause = 1;
4429 ethpause.rx_pause = 0;
4431 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4432 (fc_conf->mode & RTE_FC_TX_PAUSE))
4433 ethpause.tx_pause = 1;
4435 ethpause.tx_pause = 0;
4438 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4440 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
4454 * Configure a VLAN filter.
4457 * Pointer to Ethernet device structure.
4459 * VLAN ID to filter.
4464 * 0 on success, errno value on failure.
4467 vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4469 struct priv *priv = dev->data->dev_private;
4471 unsigned int j = -1;
4473 DEBUG("%p: %s VLAN filter ID %" PRIu16,
4474 (void *)dev, (on ? "enable" : "disable"), vlan_id);
4475 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
4476 if (!priv->vlan_filter[i].enabled) {
4477 /* Unused index, remember it. */
4481 if (priv->vlan_filter[i].id != vlan_id)
4483 /* This VLAN ID is already known, use its index. */
4487 /* Check if there's room for another VLAN filter. */
4488 if (j == (unsigned int)-1)
4491 * VLAN filters apply to all configured MAC addresses, flow
4492 * specifications must be reconfigured accordingly.
4494 priv->vlan_filter[j].id = vlan_id;
4495 if ((on) && (!priv->vlan_filter[j].enabled)) {
4497 * Filter is disabled, enable it.
4498 * Rehashing flows in all RX queues is necessary.
4501 rxq_mac_addrs_del(&priv->rxq_parent);
4503 for (i = 0; (i != priv->rxqs_n); ++i)
4504 if ((*priv->rxqs)[i] != NULL)
4505 rxq_mac_addrs_del((*priv->rxqs)[i]);
4506 priv->vlan_filter[j].enabled = 1;
4507 if (priv->started) {
4509 rxq_mac_addrs_add(&priv->rxq_parent);
4511 for (i = 0; (i != priv->rxqs_n); ++i) {
4512 if ((*priv->rxqs)[i] == NULL)
4514 rxq_mac_addrs_add((*priv->rxqs)[i]);
4517 } else if ((!on) && (priv->vlan_filter[j].enabled)) {
4519 * Filter is enabled, disable it.
4520 * Rehashing flows in all RX queues is necessary.
4523 rxq_mac_addrs_del(&priv->rxq_parent);
4525 for (i = 0; (i != priv->rxqs_n); ++i)
4526 if ((*priv->rxqs)[i] != NULL)
4527 rxq_mac_addrs_del((*priv->rxqs)[i]);
4528 priv->vlan_filter[j].enabled = 0;
4529 if (priv->started) {
4531 rxq_mac_addrs_add(&priv->rxq_parent);
4533 for (i = 0; (i != priv->rxqs_n); ++i) {
4534 if ((*priv->rxqs)[i] == NULL)
4536 rxq_mac_addrs_add((*priv->rxqs)[i]);
4544 * DPDK callback to configure a VLAN filter.
4547 * Pointer to Ethernet device structure.
4549 * VLAN ID to filter.
4554 * 0 on success, negative errno value on failure.
4557 mlx4_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4559 struct priv *priv = dev->data->dev_private;
4563 ret = vlan_filter_set(dev, vlan_id, on);
4569 static const struct eth_dev_ops mlx4_dev_ops = {
4570 .dev_configure = mlx4_dev_configure,
4571 .dev_start = mlx4_dev_start,
4572 .dev_stop = mlx4_dev_stop,
4573 .dev_close = mlx4_dev_close,
4574 .promiscuous_enable = mlx4_promiscuous_enable,
4575 .promiscuous_disable = mlx4_promiscuous_disable,
4576 .allmulticast_enable = mlx4_allmulticast_enable,
4577 .allmulticast_disable = mlx4_allmulticast_disable,
4578 .link_update = mlx4_link_update,
4579 .stats_get = mlx4_stats_get,
4580 .stats_reset = mlx4_stats_reset,
4581 .queue_stats_mapping_set = NULL,
4582 .dev_infos_get = mlx4_dev_infos_get,
4583 .vlan_filter_set = mlx4_vlan_filter_set,
4584 .vlan_tpid_set = NULL,
4585 .vlan_strip_queue_set = NULL,
4586 .vlan_offload_set = NULL,
4587 .rx_queue_setup = mlx4_rx_queue_setup,
4588 .tx_queue_setup = mlx4_tx_queue_setup,
4589 .rx_queue_release = mlx4_rx_queue_release,
4590 .tx_queue_release = mlx4_tx_queue_release,
4592 .dev_led_off = NULL,
4593 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
4594 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
4595 .priority_flow_ctrl_set = NULL,
4596 .mac_addr_remove = mlx4_mac_addr_remove,
4597 .mac_addr_add = mlx4_mac_addr_add,
4598 .mtu_set = mlx4_dev_set_mtu,
4599 .udp_tunnel_add = NULL,
4600 .udp_tunnel_del = NULL,
4604 * Get PCI information from struct ibv_device.
4607 * Pointer to Ethernet device structure.
4608 * @param[out] pci_addr
4609 * PCI bus address output buffer.
4612 * 0 on success, -1 on failure and errno is set.
4615 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
4616 struct rte_pci_addr *pci_addr)
4620 MKSTR(path, "%s/device/uevent", device->ibdev_path);
4622 file = fopen(path, "rb");
4625 while (fgets(line, sizeof(line), file) == line) {
4626 size_t len = strlen(line);
4629 /* Truncate long lines. */
4630 if (len == (sizeof(line) - 1))
4631 while (line[(len - 1)] != '\n') {
4635 line[(len - 1)] = ret;
4637 /* Extract information. */
4640 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
4644 &pci_addr->function) == 4) {
4654 * Get MAC address by querying netdevice.
4657 * struct priv for the requested device.
4659 * MAC address output buffer.
4662 * 0 on success, -1 on failure and errno is set.
4665 priv_get_mac(struct priv *priv, uint8_t (*mac)[ETHER_ADDR_LEN])
4667 struct ifreq request;
4669 if (priv_ifreq(priv, SIOCGIFHWADDR, &request))
4671 memcpy(mac, request.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
4675 /* Support up to 32 adapters. */
4677 struct rte_pci_addr pci_addr; /* associated PCI address */
4678 uint32_t ports; /* physical ports bitfield. */
4682 * Get device index in mlx4_dev[] from PCI bus address.
4684 * @param[in] pci_addr
4685 * PCI bus address to look for.
4688 * mlx4_dev[] index on success, -1 on failure.
4691 mlx4_dev_idx(struct rte_pci_addr *pci_addr)
4696 assert(pci_addr != NULL);
4697 for (i = 0; (i != elemof(mlx4_dev)); ++i) {
4698 if ((mlx4_dev[i].pci_addr.domain == pci_addr->domain) &&
4699 (mlx4_dev[i].pci_addr.bus == pci_addr->bus) &&
4700 (mlx4_dev[i].pci_addr.devid == pci_addr->devid) &&
4701 (mlx4_dev[i].pci_addr.function == pci_addr->function))
4703 if ((mlx4_dev[i].ports == 0) && (ret == -1))
4710 * Retrieve integer value from environment variable.
4713 * Environment variable name.
4716 * Integer value, 0 if the variable is not set.
4719 mlx4_getenv_int(const char *name)
4721 const char *val = getenv(name);
4729 mlx4_dev_link_status_handler(void *);
4731 mlx4_dev_interrupt_handler(struct rte_intr_handle *, void *);
4734 * Link status handler.
4737 * Pointer to private structure.
4739 * Pointer to the rte_eth_dev structure.
4742 * Nonzero if the callback process can be called immediately.
4745 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
4747 struct ibv_async_event event;
4748 int port_change = 0;
4751 /* Read all message and acknowledge them. */
4753 if (ibv_get_async_event(priv->ctx, &event))
4756 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
4757 event.event_type == IBV_EVENT_PORT_ERR)
4760 DEBUG("event type %d on port %d not handled",
4761 event.event_type, event.element.port_num);
4762 ibv_ack_async_event(&event);
4765 if (port_change ^ priv->pending_alarm) {
4766 struct rte_eth_link *link = &dev->data->dev_link;
4768 priv->pending_alarm = 0;
4769 mlx4_link_update_unlocked(dev, 0);
4770 if (((link->link_speed == 0) && link->link_status) ||
4771 ((link->link_speed != 0) && !link->link_status)) {
4772 /* Inconsistent status, check again later. */
4773 priv->pending_alarm = 1;
4774 rte_eal_alarm_set(MLX4_ALARM_TIMEOUT_US,
4775 mlx4_dev_link_status_handler,
4784 * Handle delayed link status event.
4787 * Registered argument.
4790 mlx4_dev_link_status_handler(void *arg)
4792 struct rte_eth_dev *dev = arg;
4793 struct priv *priv = dev->data->dev_private;
4797 assert(priv->pending_alarm == 1);
4798 ret = priv_dev_link_status_handler(priv, dev);
4801 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
4805 * Handle interrupts from the NIC.
4807 * @param[in] intr_handle
4808 * Interrupt handler.
4810 * Callback argument.
4813 mlx4_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
4815 struct rte_eth_dev *dev = cb_arg;
4816 struct priv *priv = dev->data->dev_private;
4821 ret = priv_dev_link_status_handler(priv, dev);
4824 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
4828 * Uninstall interrupt handler.
4831 * Pointer to private structure.
4833 * Pointer to the rte_eth_dev structure.
4836 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
4838 if (!dev->data->dev_conf.intr_conf.lsc)
4840 rte_intr_callback_unregister(&priv->intr_handle,
4841 mlx4_dev_interrupt_handler,
4843 if (priv->pending_alarm)
4844 rte_eal_alarm_cancel(mlx4_dev_link_status_handler, dev);
4845 priv->pending_alarm = 0;
4846 priv->intr_handle.fd = 0;
4847 priv->intr_handle.type = 0;
4851 * Install interrupt handler.
4854 * Pointer to private structure.
4856 * Pointer to the rte_eth_dev structure.
4859 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
4863 if (!dev->data->dev_conf.intr_conf.lsc)
4865 assert(priv->ctx->async_fd > 0);
4866 flags = fcntl(priv->ctx->async_fd, F_GETFL);
4867 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
4869 INFO("failed to change file descriptor async event queue");
4870 dev->data->dev_conf.intr_conf.lsc = 0;
4872 priv->intr_handle.fd = priv->ctx->async_fd;
4873 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
4874 rte_intr_callback_register(&priv->intr_handle,
4875 mlx4_dev_interrupt_handler,
4880 static struct eth_driver mlx4_driver;
4883 * DPDK callback to register a PCI device.
4885 * This function creates an Ethernet device for each port of a given
4888 * @param[in] pci_drv
4889 * PCI driver structure (mlx4_driver).
4890 * @param[in] pci_dev
4891 * PCI device information.
4894 * 0 on success, negative errno value on failure.
4897 mlx4_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
4899 struct ibv_device **list;
4900 struct ibv_device *ibv_dev;
4902 struct ibv_context *attr_ctx = NULL;
4903 struct ibv_device_attr device_attr;
4909 assert(pci_drv == &mlx4_driver.pci_drv);
4910 /* Get mlx4_dev[] index. */
4911 idx = mlx4_dev_idx(&pci_dev->addr);
4913 ERROR("this driver cannot support any more adapters");
4916 DEBUG("using driver device index %d", idx);
4918 /* Save PCI address. */
4919 mlx4_dev[idx].pci_addr = pci_dev->addr;
4920 list = ibv_get_device_list(&i);
4923 if (errno == ENOSYS) {
4924 WARN("cannot list devices, is ib_uverbs loaded?");
4931 * For each listed device, check related sysfs entry against
4932 * the provided PCI ID.
4935 struct rte_pci_addr pci_addr;
4938 DEBUG("checking device \"%s\"", list[i]->name);
4939 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
4941 if ((pci_dev->addr.domain != pci_addr.domain) ||
4942 (pci_dev->addr.bus != pci_addr.bus) ||
4943 (pci_dev->addr.devid != pci_addr.devid) ||
4944 (pci_dev->addr.function != pci_addr.function))
4946 vf = (pci_dev->id.device_id ==
4947 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
4948 INFO("PCI information matches, using device \"%s\" (VF: %s)",
4949 list[i]->name, (vf ? "true" : "false"));
4950 attr_ctx = ibv_open_device(list[i]);
4954 if (attr_ctx == NULL) {
4955 ibv_free_device_list(list);
4958 WARN("cannot access device, is mlx4_ib loaded?");
4961 WARN("cannot use device, are drivers up to date?");
4969 DEBUG("device opened");
4970 if (ibv_query_device(attr_ctx, &device_attr))
4972 INFO("%u port(s) detected", device_attr.phys_port_cnt);
4974 for (i = 0; i < device_attr.phys_port_cnt; i++) {
4975 uint32_t port = i + 1; /* ports are indexed from one */
4976 uint32_t test = (1 << i);
4977 struct ibv_context *ctx = NULL;
4978 struct ibv_port_attr port_attr;
4979 struct ibv_pd *pd = NULL;
4980 struct priv *priv = NULL;
4981 struct rte_eth_dev *eth_dev;
4982 #ifdef HAVE_EXP_QUERY_DEVICE
4983 struct ibv_exp_device_attr exp_device_attr;
4984 #endif /* HAVE_EXP_QUERY_DEVICE */
4985 struct ether_addr mac;
4987 #ifdef HAVE_EXP_QUERY_DEVICE
4988 exp_device_attr.comp_mask = IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS;
4990 exp_device_attr.comp_mask |= IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ;
4991 #endif /* RSS_SUPPORT */
4992 #endif /* HAVE_EXP_QUERY_DEVICE */
4994 DEBUG("using port %u (%08" PRIx32 ")", port, test);
4996 ctx = ibv_open_device(ibv_dev);
5000 /* Check port status. */
5001 err = ibv_query_port(ctx, port, &port_attr);
5003 ERROR("port query failed: %s", strerror(err));
5006 if (port_attr.state != IBV_PORT_ACTIVE)
5007 DEBUG("port %d is not active: \"%s\" (%d)",
5008 port, ibv_port_state_str(port_attr.state),
5011 /* Allocate protection domain. */
5012 pd = ibv_alloc_pd(ctx);
5014 ERROR("PD allocation failure");
5019 mlx4_dev[idx].ports |= test;
5021 /* from rte_ethdev.c */
5022 priv = rte_zmalloc("ethdev private structure",
5024 RTE_CACHE_LINE_SIZE);
5026 ERROR("priv allocation failure");
5032 priv->device_attr = device_attr;
5035 priv->mtu = ETHER_MTU;
5036 #ifdef HAVE_EXP_QUERY_DEVICE
5037 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
5038 ERROR("ibv_exp_query_device() failed");
5042 if ((exp_device_attr.exp_device_cap_flags &
5043 IBV_EXP_DEVICE_QPG) &&
5044 (exp_device_attr.exp_device_cap_flags &
5045 IBV_EXP_DEVICE_UD_RSS) &&
5046 (exp_device_attr.comp_mask &
5047 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ) &&
5048 (exp_device_attr.max_rss_tbl_sz > 0)) {
5051 priv->max_rss_tbl_sz = exp_device_attr.max_rss_tbl_sz;
5055 priv->max_rss_tbl_sz = 0;
5057 priv->hw_tss = !!(exp_device_attr.exp_device_cap_flags &
5058 IBV_EXP_DEVICE_UD_TSS);
5059 DEBUG("device flags: %s%s%s",
5060 (priv->hw_qpg ? "IBV_DEVICE_QPG " : ""),
5061 (priv->hw_tss ? "IBV_DEVICE_TSS " : ""),
5062 (priv->hw_rss ? "IBV_DEVICE_RSS " : ""));
5064 DEBUG("maximum RSS indirection table size: %u",
5065 exp_device_attr.max_rss_tbl_sz);
5066 #endif /* RSS_SUPPORT */
5069 ((exp_device_attr.exp_device_cap_flags &
5070 IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
5071 (exp_device_attr.exp_device_cap_flags &
5072 IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
5073 DEBUG("checksum offloading is %ssupported",
5074 (priv->hw_csum ? "" : "not "));
5076 priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
5077 IBV_EXP_DEVICE_VXLAN_SUPPORT);
5078 DEBUG("L2 tunnel checksum offloads are %ssupported",
5079 (priv->hw_csum_l2tun ? "" : "not "));
5082 priv->inl_recv_size = mlx4_getenv_int("MLX4_INLINE_RECV_SIZE");
5084 if (priv->inl_recv_size) {
5085 exp_device_attr.comp_mask =
5086 IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ;
5087 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
5088 INFO("Couldn't query device for inline-receive"
5090 priv->inl_recv_size = 0;
5092 if ((unsigned)exp_device_attr.inline_recv_sz <
5093 priv->inl_recv_size) {
5094 INFO("Max inline-receive (%d) <"
5095 " requested inline-receive (%u)",
5096 exp_device_attr.inline_recv_sz,
5097 priv->inl_recv_size);
5098 priv->inl_recv_size =
5099 exp_device_attr.inline_recv_sz;
5102 INFO("Set inline receive size to %u",
5103 priv->inl_recv_size);
5105 #endif /* INLINE_RECV */
5106 #endif /* HAVE_EXP_QUERY_DEVICE */
5108 (void)mlx4_getenv_int;
5110 /* Configure the first MAC address by default. */
5111 if (priv_get_mac(priv, &mac.addr_bytes)) {
5112 ERROR("cannot get MAC address, is mlx4_en loaded?"
5113 " (errno: %s)", strerror(errno));
5116 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
5118 mac.addr_bytes[0], mac.addr_bytes[1],
5119 mac.addr_bytes[2], mac.addr_bytes[3],
5120 mac.addr_bytes[4], mac.addr_bytes[5]);
5121 /* Register MAC and broadcast addresses. */
5122 claim_zero(priv_mac_addr_add(priv, 0,
5123 (const uint8_t (*)[ETHER_ADDR_LEN])
5125 claim_zero(priv_mac_addr_add(priv, (elemof(priv->mac) - 1),
5126 &(const uint8_t [ETHER_ADDR_LEN])
5127 { "\xff\xff\xff\xff\xff\xff" }));
5130 char ifname[IF_NAMESIZE];
5132 if (priv_get_ifname(priv, &ifname) == 0)
5133 DEBUG("port %u ifname is \"%s\"",
5134 priv->port, ifname);
5136 DEBUG("port %u ifname is unknown", priv->port);
5139 /* Get actual MTU if possible. */
5140 priv_get_mtu(priv, &priv->mtu);
5141 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
5143 /* from rte_ethdev.c */
5145 char name[RTE_ETH_NAME_MAX_LEN];
5147 snprintf(name, sizeof(name), "%s port %u",
5148 ibv_get_device_name(ibv_dev), port);
5149 eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
5151 if (eth_dev == NULL) {
5152 ERROR("can not allocate rte ethdev");
5157 eth_dev->data->dev_private = priv;
5158 eth_dev->pci_dev = pci_dev;
5160 rte_eth_copy_pci_info(eth_dev, pci_dev);
5162 eth_dev->driver = &mlx4_driver;
5163 eth_dev->data->rx_mbuf_alloc_failed = 0;
5164 eth_dev->data->mtu = ETHER_MTU;
5166 priv->dev = eth_dev;
5167 eth_dev->dev_ops = &mlx4_dev_ops;
5168 eth_dev->data->mac_addrs = priv->mac;
5169 TAILQ_INIT(ð_dev->link_intr_cbs);
5171 /* Bring Ethernet device up. */
5172 DEBUG("forcing Ethernet interface up");
5173 priv_set_flags(priv, ~IFF_UP, IFF_UP);
5179 claim_zero(ibv_dealloc_pd(pd));
5181 claim_zero(ibv_close_device(ctx));
5186 * XXX if something went wrong in the loop above, there is a resource
5187 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
5188 * long as the dpdk does not provide a way to deallocate a ethdev and a
5189 * way to enumerate the registered ethdevs to free the previous ones.
5192 /* no port found, complain */
5193 if (!mlx4_dev[idx].ports) {
5200 claim_zero(ibv_close_device(attr_ctx));
5202 ibv_free_device_list(list);
5207 static const struct rte_pci_id mlx4_pci_id_map[] = {
5209 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5210 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3,
5211 .subsystem_vendor_id = PCI_ANY_ID,
5212 .subsystem_device_id = PCI_ANY_ID
5215 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5216 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO,
5217 .subsystem_vendor_id = PCI_ANY_ID,
5218 .subsystem_device_id = PCI_ANY_ID
5221 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5222 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3VF,
5223 .subsystem_vendor_id = PCI_ANY_ID,
5224 .subsystem_device_id = PCI_ANY_ID
5231 static struct eth_driver mlx4_driver = {
5233 .name = MLX4_DRIVER_NAME,
5234 .id_table = mlx4_pci_id_map,
5235 .devinit = mlx4_pci_devinit,
5236 .drv_flags = RTE_PCI_DRV_INTR_LSC,
5238 .dev_private_size = sizeof(struct priv)
5242 * Driver initialization routine.
5245 rte_mlx4_pmd_init(const char *name, const char *args)
5250 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
5251 * huge pages. Calling ibv_fork_init() during init allows
5252 * applications to use fork() safely for purposes other than
5253 * using this PMD, which is not supported in forked processes.
5255 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
5257 rte_eal_pci_register(&mlx4_driver.pci_drv);
5261 static struct rte_driver rte_mlx4_driver = {
5263 .name = MLX4_DRIVER_NAME,
5264 .init = rte_mlx4_pmd_init,
5267 PMD_REGISTER_DRIVER(rte_mlx4_driver)