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 Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
989 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
990 * remove an entry first.
993 * Pointer to TX queue structure.
995 * Memory Pool for which a Memory Region lkey must be returned.
998 * mr->lkey on success, (uint32_t)-1 on failure.
1001 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
1006 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
1007 if (unlikely(txq->mp2mr[i].mp == NULL)) {
1008 /* Unknown MP, add a new MR for it. */
1011 if (txq->mp2mr[i].mp == mp) {
1012 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
1013 assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
1014 return txq->mp2mr[i].lkey;
1017 /* Add a new entry, register MR first. */
1018 DEBUG("%p: discovered new memory pool %p", (void *)txq, (void *)mp);
1019 mr = ibv_reg_mr(txq->priv->pd,
1020 (void *)mp->elt_va_start,
1021 (mp->elt_va_end - mp->elt_va_start),
1022 (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
1023 if (unlikely(mr == NULL)) {
1024 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
1026 return (uint32_t)-1;
1028 if (unlikely(i == elemof(txq->mp2mr))) {
1029 /* Table is full, remove oldest entry. */
1030 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
1033 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
1034 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
1035 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
1037 /* Store the new entry. */
1038 txq->mp2mr[i].mp = mp;
1039 txq->mp2mr[i].mr = mr;
1040 txq->mp2mr[i].lkey = mr->lkey;
1041 DEBUG("%p: new MR lkey for MP %p: 0x%08" PRIu32,
1042 (void *)txq, (void *)mp, txq->mp2mr[i].lkey);
1043 return txq->mp2mr[i].lkey;
1046 #if MLX4_PMD_SGE_WR_N > 1
1049 * Copy scattered mbuf contents to a single linear buffer.
1051 * @param[out] linear
1052 * Linear output buffer.
1054 * Scattered input buffer.
1057 * Number of bytes copied to the output buffer or 0 if not large enough.
1060 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
1062 unsigned int size = 0;
1063 unsigned int offset;
1066 unsigned int len = DATA_LEN(buf);
1070 if (unlikely(size > sizeof(*linear)))
1072 memcpy(&(*linear)[offset],
1073 rte_pktmbuf_mtod(buf, uint8_t *),
1076 } while (buf != NULL);
1081 * Handle scattered buffers for mlx4_tx_burst().
1084 * TX queue structure.
1086 * Number of segments in buf.
1088 * TX queue element to fill.
1090 * Buffer to process.
1092 * Index of the linear buffer to use if necessary (normally txq->elts_head).
1094 * Array filled with SGEs on success.
1097 * A structure containing the processed packet size in bytes and the
1098 * number of SGEs. Both fields are set to (unsigned int)-1 in case of
1101 static struct tx_burst_sg_ret {
1102 unsigned int length;
1105 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
1106 struct rte_mbuf *buf, unsigned int elts_head,
1107 struct ibv_sge (*sges)[MLX4_PMD_SGE_WR_N])
1109 unsigned int sent_size = 0;
1113 /* When there are too many segments, extra segments are
1114 * linearized in the last SGE. */
1115 if (unlikely(segs > elemof(*sges))) {
1116 segs = (elemof(*sges) - 1);
1119 /* Update element. */
1121 /* Register segments as SGEs. */
1122 for (j = 0; (j != segs); ++j) {
1123 struct ibv_sge *sge = &(*sges)[j];
1126 /* Retrieve Memory Region key for this memory pool. */
1127 lkey = txq_mp2mr(txq, buf->pool);
1128 if (unlikely(lkey == (uint32_t)-1)) {
1129 /* MR does not exist. */
1130 DEBUG("%p: unable to get MP <-> MR association",
1132 /* Clean up TX element. */
1137 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
1139 rte_prefetch0((volatile void *)
1140 (uintptr_t)sge->addr);
1141 sge->length = DATA_LEN(buf);
1143 sent_size += sge->length;
1146 /* If buf is not NULL here and is not going to be linearized,
1147 * nb_segs is not valid. */
1149 assert((buf == NULL) || (linearize));
1150 /* Linearize extra segments. */
1152 struct ibv_sge *sge = &(*sges)[segs];
1153 linear_t *linear = &(*txq->elts_linear)[elts_head];
1154 unsigned int size = linearize_mbuf(linear, buf);
1156 assert(segs == (elemof(*sges) - 1));
1158 /* Invalid packet. */
1159 DEBUG("%p: packet too large to be linearized.",
1161 /* Clean up TX element. */
1165 /* If MLX4_PMD_SGE_WR_N is 1, free mbuf immediately. */
1166 if (elemof(*sges) == 1) {
1168 struct rte_mbuf *next = NEXT(buf);
1170 rte_pktmbuf_free_seg(buf);
1172 } while (buf != NULL);
1176 sge->addr = (uintptr_t)&(*linear)[0];
1178 sge->lkey = txq->mr_linear->lkey;
1181 return (struct tx_burst_sg_ret){
1182 .length = sent_size,
1186 return (struct tx_burst_sg_ret){
1192 #endif /* MLX4_PMD_SGE_WR_N > 1 */
1195 * DPDK callback for TX.
1198 * Generic pointer to TX queue structure.
1200 * Packets to transmit.
1202 * Number of packets in array.
1205 * Number of packets successfully transmitted (<= pkts_n).
1208 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1210 struct txq *txq = (struct txq *)dpdk_txq;
1211 unsigned int elts_head = txq->elts_head;
1212 const unsigned int elts_tail = txq->elts_tail;
1213 const unsigned int elts_n = txq->elts_n;
1214 unsigned int elts_comp_cd = txq->elts_comp_cd;
1215 unsigned int elts_comp = 0;
1220 assert(elts_comp_cd != 0);
1222 max = (elts_n - (elts_head - elts_tail));
1226 assert(max <= elts_n);
1227 /* Always leave one free entry in the ring. */
1233 for (i = 0; (i != max); ++i) {
1234 struct rte_mbuf *buf = pkts[i];
1235 unsigned int elts_head_next =
1236 (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
1237 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
1238 struct txq_elt *elt = &(*txq->elts)[elts_head];
1239 unsigned int segs = NB_SEGS(buf);
1240 #ifdef MLX4_PMD_SOFT_COUNTERS
1241 unsigned int sent_size = 0;
1243 uint32_t send_flags = 0;
1245 /* Clean up old buffer. */
1246 if (likely(elt->buf != NULL)) {
1247 struct rte_mbuf *tmp = elt->buf;
1249 /* Faster than rte_pktmbuf_free(). */
1251 struct rte_mbuf *next = NEXT(tmp);
1253 rte_pktmbuf_free_seg(tmp);
1255 } while (tmp != NULL);
1257 /* Request TX completion. */
1258 if (unlikely(--elts_comp_cd == 0)) {
1259 elts_comp_cd = txq->elts_comp_cd_init;
1261 send_flags |= IBV_EXP_QP_BURST_SIGNALED;
1263 /* Should we enable HW CKSUM offload */
1265 (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
1266 send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
1267 /* HW does not support checksum offloads at arbitrary
1268 * offsets but automatically recognizes the packet
1269 * type. For inner L3/L4 checksums, only VXLAN (UDP)
1270 * tunnels are currently supported. */
1271 if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
1272 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
1274 if (likely(segs == 1)) {
1279 /* Retrieve buffer information. */
1280 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1281 length = DATA_LEN(buf);
1282 /* Retrieve Memory Region key for this memory pool. */
1283 lkey = txq_mp2mr(txq, buf->pool);
1284 if (unlikely(lkey == (uint32_t)-1)) {
1285 /* MR does not exist. */
1286 DEBUG("%p: unable to get MP <-> MR"
1287 " association", (void *)txq);
1288 /* Clean up TX element. */
1292 /* Update element. */
1295 rte_prefetch0((volatile void *)
1297 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1298 /* Put packet into send queue. */
1299 #if MLX4_PMD_MAX_INLINE > 0
1300 if (length <= txq->max_inline)
1301 err = txq->if_qp->send_pending_inline
1308 err = txq->if_qp->send_pending
1316 #ifdef MLX4_PMD_SOFT_COUNTERS
1317 sent_size += length;
1320 #if MLX4_PMD_SGE_WR_N > 1
1321 struct ibv_sge sges[MLX4_PMD_SGE_WR_N];
1322 struct tx_burst_sg_ret ret;
1324 ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
1326 if (ret.length == (unsigned int)-1)
1328 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
1329 /* Put SG list into send queue. */
1330 err = txq->if_qp->send_pending_sg_list
1337 #ifdef MLX4_PMD_SOFT_COUNTERS
1338 sent_size += ret.length;
1340 #else /* MLX4_PMD_SGE_WR_N > 1 */
1341 DEBUG("%p: TX scattered buffers support not"
1342 " compiled in", (void *)txq);
1344 #endif /* MLX4_PMD_SGE_WR_N > 1 */
1346 elts_head = elts_head_next;
1347 #ifdef MLX4_PMD_SOFT_COUNTERS
1348 /* Increment sent bytes counter. */
1349 txq->stats.obytes += sent_size;
1353 /* Take a shortcut if nothing must be sent. */
1354 if (unlikely(i == 0))
1356 #ifdef MLX4_PMD_SOFT_COUNTERS
1357 /* Increment sent packets counter. */
1358 txq->stats.opackets += i;
1360 /* Ring QP doorbell. */
1361 err = txq->if_qp->send_flush(txq->qp);
1362 if (unlikely(err)) {
1363 /* A nonzero value is not supposed to be returned.
1364 * Nothing can be done about it. */
1365 DEBUG("%p: send_flush() failed with error %d",
1368 txq->elts_head = elts_head;
1369 txq->elts_comp += elts_comp;
1370 txq->elts_comp_cd = elts_comp_cd;
1375 * Configure a TX queue.
1378 * Pointer to Ethernet device structure.
1380 * Pointer to TX queue structure.
1382 * Number of descriptors to configure in queue.
1384 * NUMA socket on which memory must be allocated.
1386 * Thresholds parameters.
1389 * 0 on success, errno value on failure.
1392 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1393 unsigned int socket, const struct rte_eth_txconf *conf)
1395 struct priv *priv = dev->data->dev_private;
1401 struct ibv_exp_query_intf_params params;
1402 struct ibv_exp_qp_init_attr init;
1403 struct ibv_exp_res_domain_init_attr rd;
1404 struct ibv_exp_cq_init_attr cq;
1405 struct ibv_exp_qp_attr mod;
1407 enum ibv_exp_query_intf_status status;
1410 (void)conf; /* Thresholds configuration (ignored). */
1411 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
1412 ERROR("%p: invalid number of TX descriptors (must be a"
1413 " multiple of %d)", (void *)dev, MLX4_PMD_SGE_WR_N);
1416 desc /= MLX4_PMD_SGE_WR_N;
1417 /* MRs will be registered in mp2mr[] later. */
1418 attr.rd = (struct ibv_exp_res_domain_init_attr){
1419 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
1420 IBV_EXP_RES_DOMAIN_MSG_MODEL),
1421 .thread_model = IBV_EXP_THREAD_SINGLE,
1422 .msg_model = IBV_EXP_MSG_HIGH_BW,
1424 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
1425 if (tmpl.rd == NULL) {
1427 ERROR("%p: RD creation failure: %s",
1428 (void *)dev, strerror(ret));
1431 attr.cq = (struct ibv_exp_cq_init_attr){
1432 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
1433 .res_domain = tmpl.rd,
1435 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
1436 if (tmpl.cq == NULL) {
1438 ERROR("%p: CQ creation failure: %s",
1439 (void *)dev, strerror(ret));
1442 DEBUG("priv->device_attr.max_qp_wr is %d",
1443 priv->device_attr.max_qp_wr);
1444 DEBUG("priv->device_attr.max_sge is %d",
1445 priv->device_attr.max_sge);
1446 attr.init = (struct ibv_exp_qp_init_attr){
1447 /* CQ to be associated with the send queue. */
1449 /* CQ to be associated with the receive queue. */
1452 /* Max number of outstanding WRs. */
1453 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1454 priv->device_attr.max_qp_wr :
1456 /* Max number of scatter/gather elements in a WR. */
1457 .max_send_sge = ((priv->device_attr.max_sge <
1458 MLX4_PMD_SGE_WR_N) ?
1459 priv->device_attr.max_sge :
1461 #if MLX4_PMD_MAX_INLINE > 0
1462 .max_inline_data = MLX4_PMD_MAX_INLINE,
1465 .qp_type = IBV_QPT_RAW_PACKET,
1466 /* Do *NOT* enable this, completions events are managed per
1470 .res_domain = tmpl.rd,
1471 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
1472 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
1474 tmpl.qp = ibv_exp_create_qp(priv->ctx, &attr.init);
1475 if (tmpl.qp == NULL) {
1476 ret = (errno ? errno : EINVAL);
1477 ERROR("%p: QP creation failure: %s",
1478 (void *)dev, strerror(ret));
1481 #if MLX4_PMD_MAX_INLINE > 0
1482 /* ibv_create_qp() updates this value. */
1483 tmpl.max_inline = attr.init.cap.max_inline_data;
1485 attr.mod = (struct ibv_exp_qp_attr){
1486 /* Move the QP to this state. */
1487 .qp_state = IBV_QPS_INIT,
1488 /* Primary port number. */
1489 .port_num = priv->port
1491 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
1492 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
1494 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1495 (void *)dev, strerror(ret));
1498 ret = txq_alloc_elts(&tmpl, desc);
1500 ERROR("%p: TXQ allocation failed: %s",
1501 (void *)dev, strerror(ret));
1504 attr.mod = (struct ibv_exp_qp_attr){
1505 .qp_state = IBV_QPS_RTR
1507 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1509 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1510 (void *)dev, strerror(ret));
1513 attr.mod.qp_state = IBV_QPS_RTS;
1514 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1516 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1517 (void *)dev, strerror(ret));
1520 attr.params = (struct ibv_exp_query_intf_params){
1521 .intf_scope = IBV_EXP_INTF_GLOBAL,
1522 .intf = IBV_EXP_INTF_CQ,
1525 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
1526 if (tmpl.if_cq == NULL) {
1527 ERROR("%p: CQ interface family query failed with status %d",
1528 (void *)dev, status);
1531 attr.params = (struct ibv_exp_query_intf_params){
1532 .intf_scope = IBV_EXP_INTF_GLOBAL,
1533 .intf = IBV_EXP_INTF_QP_BURST,
1535 #ifdef HAVE_EXP_QP_BURST_CREATE_DISABLE_ETH_LOOPBACK
1536 /* MC loopback must be disabled when not using a VF. */
1539 IBV_EXP_QP_BURST_CREATE_DISABLE_ETH_LOOPBACK :
1543 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
1544 if (tmpl.if_qp == NULL) {
1545 ERROR("%p: QP interface family query failed with status %d",
1546 (void *)dev, status);
1549 /* Clean up txq in case we're reinitializing it. */
1550 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1553 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1563 * DPDK callback to configure a TX queue.
1566 * Pointer to Ethernet device structure.
1570 * Number of descriptors to configure in queue.
1572 * NUMA socket on which memory must be allocated.
1574 * Thresholds parameters.
1577 * 0 on success, negative errno value on failure.
1580 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1581 unsigned int socket, const struct rte_eth_txconf *conf)
1583 struct priv *priv = dev->data->dev_private;
1584 struct txq *txq = (*priv->txqs)[idx];
1588 DEBUG("%p: configuring queue %u for %u descriptors",
1589 (void *)dev, idx, desc);
1590 if (idx >= priv->txqs_n) {
1591 ERROR("%p: queue index out of range (%u >= %u)",
1592 (void *)dev, idx, priv->txqs_n);
1597 DEBUG("%p: reusing already allocated queue index %u (%p)",
1598 (void *)dev, idx, (void *)txq);
1599 if (priv->started) {
1603 (*priv->txqs)[idx] = NULL;
1606 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1608 ERROR("%p: unable to allocate queue index %u",
1614 ret = txq_setup(dev, txq, desc, socket, conf);
1618 txq->stats.idx = idx;
1619 DEBUG("%p: adding TX queue %p to list",
1620 (void *)dev, (void *)txq);
1621 (*priv->txqs)[idx] = txq;
1622 /* Update send callback. */
1623 dev->tx_pkt_burst = mlx4_tx_burst;
1630 * DPDK callback to release a TX queue.
1633 * Generic TX queue pointer.
1636 mlx4_tx_queue_release(void *dpdk_txq)
1638 struct txq *txq = (struct txq *)dpdk_txq;
1646 for (i = 0; (i != priv->txqs_n); ++i)
1647 if ((*priv->txqs)[i] == txq) {
1648 DEBUG("%p: removing TX queue %p from list",
1649 (void *)priv->dev, (void *)txq);
1650 (*priv->txqs)[i] = NULL;
1658 /* RX queues handling. */
1661 * Allocate RX queue elements with scattered packets support.
1664 * Pointer to RX queue structure.
1666 * Number of elements to allocate.
1668 * If not NULL, fetch buffers from this array instead of allocating them
1669 * with rte_pktmbuf_alloc().
1672 * 0 on success, errno value on failure.
1675 rxq_alloc_elts_sp(struct rxq *rxq, unsigned int elts_n,
1676 struct rte_mbuf **pool)
1679 struct rxq_elt_sp (*elts)[elts_n] =
1680 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1685 ERROR("%p: can't allocate packets array", (void *)rxq);
1689 /* For each WR (packet). */
1690 for (i = 0; (i != elts_n); ++i) {
1692 struct rxq_elt_sp *elt = &(*elts)[i];
1693 struct ibv_recv_wr *wr = &elt->wr;
1694 struct ibv_sge (*sges)[(elemof(elt->sges))] = &elt->sges;
1696 /* These two arrays must have the same size. */
1697 assert(elemof(elt->sges) == elemof(elt->bufs));
1700 wr->next = &(*elts)[(i + 1)].wr;
1701 wr->sg_list = &(*sges)[0];
1702 wr->num_sge = elemof(*sges);
1703 /* For each SGE (segment). */
1704 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1705 struct ibv_sge *sge = &(*sges)[j];
1706 struct rte_mbuf *buf;
1710 assert(buf != NULL);
1711 rte_pktmbuf_reset(buf);
1713 buf = rte_pktmbuf_alloc(rxq->mp);
1715 assert(pool == NULL);
1716 ERROR("%p: empty mbuf pool", (void *)rxq);
1721 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1722 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1723 /* Buffer is supposed to be empty. */
1724 assert(rte_pktmbuf_data_len(buf) == 0);
1725 assert(rte_pktmbuf_pkt_len(buf) == 0);
1726 /* sge->addr must be able to store a pointer. */
1727 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1729 /* The first SGE keeps its headroom. */
1730 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
1731 sge->length = (buf->buf_len -
1732 RTE_PKTMBUF_HEADROOM);
1734 /* Subsequent SGEs lose theirs. */
1735 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1736 SET_DATA_OFF(buf, 0);
1737 sge->addr = (uintptr_t)buf->buf_addr;
1738 sge->length = buf->buf_len;
1740 sge->lkey = rxq->mr->lkey;
1741 /* Redundant check for tailroom. */
1742 assert(sge->length == rte_pktmbuf_tailroom(buf));
1745 /* The last WR pointer must be NULL. */
1746 (*elts)[(i - 1)].wr.next = NULL;
1747 DEBUG("%p: allocated and configured %u WRs (%zu segments)",
1748 (void *)rxq, elts_n, (elts_n * elemof((*elts)[0].sges)));
1749 rxq->elts_n = elts_n;
1751 rxq->elts.sp = elts;
1756 assert(pool == NULL);
1757 for (i = 0; (i != elemof(*elts)); ++i) {
1759 struct rxq_elt_sp *elt = &(*elts)[i];
1761 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1762 struct rte_mbuf *buf = elt->bufs[j];
1765 rte_pktmbuf_free_seg(buf);
1770 DEBUG("%p: failed, freed everything", (void *)rxq);
1776 * Free RX queue elements with scattered packets support.
1779 * Pointer to RX queue structure.
1782 rxq_free_elts_sp(struct rxq *rxq)
1785 unsigned int elts_n = rxq->elts_n;
1786 struct rxq_elt_sp (*elts)[elts_n] = rxq->elts.sp;
1788 DEBUG("%p: freeing WRs", (void *)rxq);
1790 rxq->elts.sp = NULL;
1793 for (i = 0; (i != elemof(*elts)); ++i) {
1795 struct rxq_elt_sp *elt = &(*elts)[i];
1797 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1798 struct rte_mbuf *buf = elt->bufs[j];
1801 rte_pktmbuf_free_seg(buf);
1808 * Allocate RX queue elements.
1811 * Pointer to RX queue structure.
1813 * Number of elements to allocate.
1815 * If not NULL, fetch buffers from this array instead of allocating them
1816 * with rte_pktmbuf_alloc().
1819 * 0 on success, errno value on failure.
1822 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n, struct rte_mbuf **pool)
1825 struct rxq_elt (*elts)[elts_n] =
1826 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1831 ERROR("%p: can't allocate packets array", (void *)rxq);
1835 /* For each WR (packet). */
1836 for (i = 0; (i != elts_n); ++i) {
1837 struct rxq_elt *elt = &(*elts)[i];
1838 struct ibv_recv_wr *wr = &elt->wr;
1839 struct ibv_sge *sge = &(*elts)[i].sge;
1840 struct rte_mbuf *buf;
1844 assert(buf != NULL);
1845 rte_pktmbuf_reset(buf);
1847 buf = rte_pktmbuf_alloc(rxq->mp);
1849 assert(pool == NULL);
1850 ERROR("%p: empty mbuf pool", (void *)rxq);
1854 /* Configure WR. Work request ID contains its own index in
1855 * the elts array and the offset between SGE buffer header and
1857 WR_ID(wr->wr_id).id = i;
1858 WR_ID(wr->wr_id).offset =
1859 (((uintptr_t)buf->buf_addr + RTE_PKTMBUF_HEADROOM) -
1861 wr->next = &(*elts)[(i + 1)].wr;
1864 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1865 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1866 /* Buffer is supposed to be empty. */
1867 assert(rte_pktmbuf_data_len(buf) == 0);
1868 assert(rte_pktmbuf_pkt_len(buf) == 0);
1869 /* sge->addr must be able to store a pointer. */
1870 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1871 /* SGE keeps its headroom. */
1872 sge->addr = (uintptr_t)
1873 ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
1874 sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
1875 sge->lkey = rxq->mr->lkey;
1876 /* Redundant check for tailroom. */
1877 assert(sge->length == rte_pktmbuf_tailroom(buf));
1878 /* Make sure elts index and SGE mbuf pointer can be deduced
1880 if ((WR_ID(wr->wr_id).id != i) ||
1881 ((void *)((uintptr_t)sge->addr -
1882 WR_ID(wr->wr_id).offset) != buf)) {
1883 ERROR("%p: cannot store index and offset in WR ID",
1886 rte_pktmbuf_free(buf);
1891 /* The last WR pointer must be NULL. */
1892 (*elts)[(i - 1)].wr.next = NULL;
1893 DEBUG("%p: allocated and configured %u single-segment WRs",
1894 (void *)rxq, elts_n);
1895 rxq->elts_n = elts_n;
1897 rxq->elts.no_sp = elts;
1902 assert(pool == NULL);
1903 for (i = 0; (i != elemof(*elts)); ++i) {
1904 struct rxq_elt *elt = &(*elts)[i];
1905 struct rte_mbuf *buf;
1907 if (elt->sge.addr == 0)
1909 assert(WR_ID(elt->wr.wr_id).id == i);
1910 buf = (void *)((uintptr_t)elt->sge.addr -
1911 WR_ID(elt->wr.wr_id).offset);
1912 rte_pktmbuf_free_seg(buf);
1916 DEBUG("%p: failed, freed everything", (void *)rxq);
1922 * Free RX queue elements.
1925 * Pointer to RX queue structure.
1928 rxq_free_elts(struct rxq *rxq)
1931 unsigned int elts_n = rxq->elts_n;
1932 struct rxq_elt (*elts)[elts_n] = rxq->elts.no_sp;
1934 DEBUG("%p: freeing WRs", (void *)rxq);
1936 rxq->elts.no_sp = NULL;
1939 for (i = 0; (i != elemof(*elts)); ++i) {
1940 struct rxq_elt *elt = &(*elts)[i];
1941 struct rte_mbuf *buf;
1943 if (elt->sge.addr == 0)
1945 assert(WR_ID(elt->wr.wr_id).id == i);
1946 buf = (void *)((uintptr_t)elt->sge.addr -
1947 WR_ID(elt->wr.wr_id).offset);
1948 rte_pktmbuf_free_seg(buf);
1954 * Delete flow steering rule.
1957 * Pointer to RX queue structure.
1959 * MAC address index.
1964 rxq_del_flow(struct rxq *rxq, unsigned int mac_index, unsigned int vlan_index)
1967 struct priv *priv = rxq->priv;
1968 const uint8_t (*mac)[ETHER_ADDR_LEN] =
1969 (const uint8_t (*)[ETHER_ADDR_LEN])
1970 priv->mac[mac_index].addr_bytes;
1972 assert(rxq->mac_flow[mac_index][vlan_index] != NULL);
1973 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
1974 " (VLAN ID %" PRIu16 ")",
1976 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
1977 mac_index, priv->vlan_filter[vlan_index].id);
1978 claim_zero(ibv_destroy_flow(rxq->mac_flow[mac_index][vlan_index]));
1979 rxq->mac_flow[mac_index][vlan_index] = NULL;
1983 * Unregister a MAC address from a RX queue.
1986 * Pointer to RX queue structure.
1988 * MAC address index.
1991 rxq_mac_addr_del(struct rxq *rxq, unsigned int mac_index)
1993 struct priv *priv = rxq->priv;
1995 unsigned int vlans = 0;
1997 assert(mac_index < elemof(priv->mac));
1998 if (!BITFIELD_ISSET(rxq->mac_configured, mac_index))
2000 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
2001 if (!priv->vlan_filter[i].enabled)
2003 rxq_del_flow(rxq, mac_index, i);
2007 rxq_del_flow(rxq, mac_index, 0);
2009 BITFIELD_RESET(rxq->mac_configured, mac_index);
2013 * Unregister all MAC addresses from a RX queue.
2016 * Pointer to RX queue structure.
2019 rxq_mac_addrs_del(struct rxq *rxq)
2021 struct priv *priv = rxq->priv;
2024 for (i = 0; (i != elemof(priv->mac)); ++i)
2025 rxq_mac_addr_del(rxq, i);
2028 static int rxq_promiscuous_enable(struct rxq *);
2029 static void rxq_promiscuous_disable(struct rxq *);
2032 * Add single flow steering rule.
2035 * Pointer to RX queue structure.
2037 * MAC address index to register.
2039 * VLAN index. Use -1 for a flow without VLAN.
2042 * 0 on success, errno value on failure.
2045 rxq_add_flow(struct rxq *rxq, unsigned int mac_index, unsigned int vlan_index)
2047 struct ibv_flow *flow;
2048 struct priv *priv = rxq->priv;
2049 const uint8_t (*mac)[ETHER_ADDR_LEN] =
2050 (const uint8_t (*)[ETHER_ADDR_LEN])
2051 priv->mac[mac_index].addr_bytes;
2053 /* Allocate flow specification on the stack. */
2054 struct __attribute__((packed)) {
2055 struct ibv_flow_attr attr;
2056 struct ibv_flow_spec_eth spec;
2058 struct ibv_flow_attr *attr = &data.attr;
2059 struct ibv_flow_spec_eth *spec = &data.spec;
2061 assert(mac_index < elemof(priv->mac));
2062 assert((vlan_index < elemof(priv->vlan_filter)) || (vlan_index == -1u));
2064 * No padding must be inserted by the compiler between attr and spec.
2065 * This layout is expected by libibverbs.
2067 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
2068 *attr = (struct ibv_flow_attr){
2069 .type = IBV_FLOW_ATTR_NORMAL,
2074 *spec = (struct ibv_flow_spec_eth){
2075 .type = IBV_FLOW_SPEC_ETH,
2076 .size = sizeof(*spec),
2079 (*mac)[0], (*mac)[1], (*mac)[2],
2080 (*mac)[3], (*mac)[4], (*mac)[5]
2082 .vlan_tag = ((vlan_index != -1u) ?
2083 htons(priv->vlan_filter[vlan_index].id) :
2087 .dst_mac = "\xff\xff\xff\xff\xff\xff",
2088 .vlan_tag = ((vlan_index != -1u) ? htons(0xfff) : 0),
2091 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
2092 " (VLAN %s %" PRIu16 ")",
2094 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
2096 ((vlan_index != -1u) ? "ID" : "index"),
2097 ((vlan_index != -1u) ? priv->vlan_filter[vlan_index].id : -1u));
2098 /* Create related flow. */
2100 flow = ibv_create_flow(rxq->qp, attr);
2102 /* It's not clear whether errno is always set in this case. */
2103 ERROR("%p: flow configuration failed, errno=%d: %s",
2105 (errno ? strerror(errno) : "Unknown error"));
2110 if (vlan_index == -1u)
2112 assert(rxq->mac_flow[mac_index][vlan_index] == NULL);
2113 rxq->mac_flow[mac_index][vlan_index] = flow;
2118 * Register a MAC address in a RX queue.
2121 * Pointer to RX queue structure.
2123 * MAC address index to register.
2126 * 0 on success, errno value on failure.
2129 rxq_mac_addr_add(struct rxq *rxq, unsigned int mac_index)
2131 struct priv *priv = rxq->priv;
2133 unsigned int vlans = 0;
2136 assert(mac_index < elemof(priv->mac));
2137 if (BITFIELD_ISSET(rxq->mac_configured, mac_index))
2138 rxq_mac_addr_del(rxq, mac_index);
2139 /* Fill VLAN specifications. */
2140 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
2141 if (!priv->vlan_filter[i].enabled)
2143 /* Create related flow. */
2144 ret = rxq_add_flow(rxq, mac_index, i);
2149 /* Failure, rollback. */
2151 if (priv->vlan_filter[--i].enabled)
2152 rxq_del_flow(rxq, mac_index, i);
2156 /* In case there is no VLAN filter. */
2158 ret = rxq_add_flow(rxq, mac_index, -1);
2162 BITFIELD_SET(rxq->mac_configured, mac_index);
2167 * Register all MAC addresses in a RX queue.
2170 * Pointer to RX queue structure.
2173 * 0 on success, errno value on failure.
2176 rxq_mac_addrs_add(struct rxq *rxq)
2178 struct priv *priv = rxq->priv;
2182 for (i = 0; (i != elemof(priv->mac)); ++i) {
2183 if (!BITFIELD_ISSET(priv->mac_configured, i))
2185 ret = rxq_mac_addr_add(rxq, i);
2188 /* Failure, rollback. */
2190 rxq_mac_addr_del(rxq, --i);
2198 * Unregister a MAC address.
2200 * In RSS mode, the MAC address is unregistered from the parent queue,
2201 * otherwise it is unregistered from each queue directly.
2204 * Pointer to private structure.
2206 * MAC address index.
2209 priv_mac_addr_del(struct priv *priv, unsigned int mac_index)
2213 assert(mac_index < elemof(priv->mac));
2214 if (!BITFIELD_ISSET(priv->mac_configured, mac_index))
2217 rxq_mac_addr_del(&priv->rxq_parent, mac_index);
2220 for (i = 0; (i != priv->dev->data->nb_rx_queues); ++i)
2221 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2223 BITFIELD_RESET(priv->mac_configured, mac_index);
2227 * Register a MAC address.
2229 * In RSS mode, the MAC address is registered in the parent queue,
2230 * otherwise it is registered in each queue directly.
2233 * Pointer to private structure.
2235 * MAC address index to use.
2237 * MAC address to register.
2240 * 0 on success, errno value on failure.
2243 priv_mac_addr_add(struct priv *priv, unsigned int mac_index,
2244 const uint8_t (*mac)[ETHER_ADDR_LEN])
2249 assert(mac_index < elemof(priv->mac));
2250 /* First, make sure this address isn't already configured. */
2251 for (i = 0; (i != elemof(priv->mac)); ++i) {
2252 /* Skip this index, it's going to be reconfigured. */
2255 if (!BITFIELD_ISSET(priv->mac_configured, i))
2257 if (memcmp(priv->mac[i].addr_bytes, *mac, sizeof(*mac)))
2259 /* Address already configured elsewhere, return with error. */
2262 if (BITFIELD_ISSET(priv->mac_configured, mac_index))
2263 priv_mac_addr_del(priv, mac_index);
2264 priv->mac[mac_index] = (struct ether_addr){
2266 (*mac)[0], (*mac)[1], (*mac)[2],
2267 (*mac)[3], (*mac)[4], (*mac)[5]
2270 /* If device isn't started, this is all we need to do. */
2271 if (!priv->started) {
2273 /* Verify that all queues have this index disabled. */
2274 for (i = 0; (i != priv->rxqs_n); ++i) {
2275 if ((*priv->rxqs)[i] == NULL)
2277 assert(!BITFIELD_ISSET
2278 ((*priv->rxqs)[i]->mac_configured, mac_index));
2284 ret = rxq_mac_addr_add(&priv->rxq_parent, mac_index);
2289 for (i = 0; (i != priv->rxqs_n); ++i) {
2290 if ((*priv->rxqs)[i] == NULL)
2292 ret = rxq_mac_addr_add((*priv->rxqs)[i], mac_index);
2295 /* Failure, rollback. */
2297 if ((*priv->rxqs)[(--i)] != NULL)
2298 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2302 BITFIELD_SET(priv->mac_configured, mac_index);
2307 * Enable allmulti mode in a RX queue.
2310 * Pointer to RX queue structure.
2313 * 0 on success, errno value on failure.
2316 rxq_allmulticast_enable(struct rxq *rxq)
2318 struct ibv_flow *flow;
2319 struct ibv_flow_attr attr = {
2320 .type = IBV_FLOW_ATTR_MC_DEFAULT,
2322 .port = rxq->priv->port,
2326 DEBUG("%p: enabling allmulticast mode", (void *)rxq);
2327 if (rxq->allmulti_flow != NULL)
2330 flow = ibv_create_flow(rxq->qp, &attr);
2332 /* It's not clear whether errno is always set in this case. */
2333 ERROR("%p: flow configuration failed, errno=%d: %s",
2335 (errno ? strerror(errno) : "Unknown error"));
2340 rxq->allmulti_flow = flow;
2341 DEBUG("%p: allmulticast mode enabled", (void *)rxq);
2346 * Disable allmulti mode in a RX queue.
2349 * Pointer to RX queue structure.
2352 rxq_allmulticast_disable(struct rxq *rxq)
2354 DEBUG("%p: disabling allmulticast mode", (void *)rxq);
2355 if (rxq->allmulti_flow == NULL)
2357 claim_zero(ibv_destroy_flow(rxq->allmulti_flow));
2358 rxq->allmulti_flow = NULL;
2359 DEBUG("%p: allmulticast mode disabled", (void *)rxq);
2363 * Enable promiscuous mode in a RX queue.
2366 * Pointer to RX queue structure.
2369 * 0 on success, errno value on failure.
2372 rxq_promiscuous_enable(struct rxq *rxq)
2374 struct ibv_flow *flow;
2375 struct ibv_flow_attr attr = {
2376 .type = IBV_FLOW_ATTR_ALL_DEFAULT,
2378 .port = rxq->priv->port,
2384 DEBUG("%p: enabling promiscuous mode", (void *)rxq);
2385 if (rxq->promisc_flow != NULL)
2388 flow = ibv_create_flow(rxq->qp, &attr);
2390 /* It's not clear whether errno is always set in this case. */
2391 ERROR("%p: flow configuration failed, errno=%d: %s",
2393 (errno ? strerror(errno) : "Unknown error"));
2398 rxq->promisc_flow = flow;
2399 DEBUG("%p: promiscuous mode enabled", (void *)rxq);
2404 * Disable promiscuous mode in a RX queue.
2407 * Pointer to RX queue structure.
2410 rxq_promiscuous_disable(struct rxq *rxq)
2414 DEBUG("%p: disabling promiscuous mode", (void *)rxq);
2415 if (rxq->promisc_flow == NULL)
2417 claim_zero(ibv_destroy_flow(rxq->promisc_flow));
2418 rxq->promisc_flow = NULL;
2419 DEBUG("%p: promiscuous mode disabled", (void *)rxq);
2423 * Clean up a RX queue.
2425 * Destroy objects, free allocated memory and reset the structure for reuse.
2428 * Pointer to RX queue structure.
2431 rxq_cleanup(struct rxq *rxq)
2433 struct ibv_exp_release_intf_params params;
2435 DEBUG("cleaning up %p", (void *)rxq);
2437 rxq_free_elts_sp(rxq);
2440 if (rxq->if_qp != NULL) {
2441 assert(rxq->priv != NULL);
2442 assert(rxq->priv->ctx != NULL);
2443 assert(rxq->qp != NULL);
2444 params = (struct ibv_exp_release_intf_params){
2447 claim_zero(ibv_exp_release_intf(rxq->priv->ctx,
2451 if (rxq->if_cq != NULL) {
2452 assert(rxq->priv != NULL);
2453 assert(rxq->priv->ctx != NULL);
2454 assert(rxq->cq != NULL);
2455 params = (struct ibv_exp_release_intf_params){
2458 claim_zero(ibv_exp_release_intf(rxq->priv->ctx,
2462 if (rxq->qp != NULL) {
2463 rxq_promiscuous_disable(rxq);
2464 rxq_allmulticast_disable(rxq);
2465 rxq_mac_addrs_del(rxq);
2466 claim_zero(ibv_destroy_qp(rxq->qp));
2468 if (rxq->cq != NULL)
2469 claim_zero(ibv_destroy_cq(rxq->cq));
2470 if (rxq->rd != NULL) {
2471 struct ibv_exp_destroy_res_domain_attr attr = {
2475 assert(rxq->priv != NULL);
2476 assert(rxq->priv->ctx != NULL);
2477 claim_zero(ibv_exp_destroy_res_domain(rxq->priv->ctx,
2481 if (rxq->mr != NULL)
2482 claim_zero(ibv_dereg_mr(rxq->mr));
2483 memset(rxq, 0, sizeof(*rxq));
2487 * Translate RX completion flags to packet type.
2490 * RX completion flags returned by poll_length_flags().
2493 * Packet type for struct rte_mbuf.
2495 static inline uint32_t
2496 rxq_cq_to_pkt_type(uint32_t flags)
2500 if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
2503 IBV_EXP_CQ_RX_OUTER_IPV4_PACKET, RTE_PTYPE_L3_IPV4) |
2505 IBV_EXP_CQ_RX_OUTER_IPV6_PACKET, RTE_PTYPE_L3_IPV6) |
2507 IBV_EXP_CQ_RX_IPV4_PACKET, RTE_PTYPE_INNER_L3_IPV4) |
2509 IBV_EXP_CQ_RX_IPV6_PACKET, RTE_PTYPE_INNER_L3_IPV6);
2513 IBV_EXP_CQ_RX_IPV4_PACKET, RTE_PTYPE_L3_IPV4) |
2515 IBV_EXP_CQ_RX_IPV6_PACKET, RTE_PTYPE_L3_IPV6);
2520 * Translate RX completion flags to offload flags.
2523 * Pointer to RX queue structure.
2525 * RX completion flags returned by poll_length_flags().
2528 * Offload flags (ol_flags) for struct rte_mbuf.
2530 static inline uint32_t
2531 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
2533 uint32_t ol_flags = 0;
2538 IBV_EXP_CQ_RX_IP_CSUM_OK,
2539 PKT_RX_IP_CKSUM_BAD) |
2541 IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
2542 PKT_RX_L4_CKSUM_BAD);
2544 * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
2545 * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
2548 if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
2551 IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
2552 PKT_RX_IP_CKSUM_BAD) |
2554 IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
2555 PKT_RX_L4_CKSUM_BAD);
2560 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
2563 * DPDK callback for RX with scattered packets support.
2566 * Generic pointer to RX queue structure.
2568 * Array to store received packets.
2570 * Maximum number of packets in array.
2573 * Number of packets successfully received (<= pkts_n).
2576 mlx4_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2578 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2579 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
2580 const unsigned int elts_n = rxq->elts_n;
2581 unsigned int elts_head = rxq->elts_head;
2582 struct ibv_recv_wr head;
2583 struct ibv_recv_wr **next = &head.next;
2584 struct ibv_recv_wr *bad_wr;
2586 unsigned int pkts_ret = 0;
2589 if (unlikely(!rxq->sp))
2590 return mlx4_rx_burst(dpdk_rxq, pkts, pkts_n);
2591 if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
2593 for (i = 0; (i != pkts_n); ++i) {
2594 struct rxq_elt_sp *elt = &(*elts)[elts_head];
2595 struct ibv_recv_wr *wr = &elt->wr;
2596 uint64_t wr_id = wr->wr_id;
2598 unsigned int pkt_buf_len;
2599 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
2600 struct rte_mbuf **pkt_buf_next = &pkt_buf;
2601 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
2605 /* Sanity checks. */
2609 assert(wr_id < rxq->elts_n);
2610 assert(wr->sg_list == elt->sges);
2611 assert(wr->num_sge == elemof(elt->sges));
2612 assert(elts_head < rxq->elts_n);
2613 assert(rxq->elts_head < rxq->elts_n);
2614 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
2616 if (unlikely(ret < 0)) {
2620 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
2622 /* ibv_poll_cq() must be used in case of failure. */
2623 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
2624 if (unlikely(wcs_n == 0))
2626 if (unlikely(wcs_n < 0)) {
2627 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
2628 (void *)rxq, wcs_n);
2632 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
2633 /* Whatever, just repost the offending WR. */
2634 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
2635 " completion status (%d): %s",
2636 (void *)rxq, wc.wr_id, wc.status,
2637 ibv_wc_status_str(wc.status));
2638 #ifdef MLX4_PMD_SOFT_COUNTERS
2639 /* Increment dropped packets counter. */
2640 ++rxq->stats.idropped;
2642 /* Link completed WRs together for repost. */
2653 /* Link completed WRs together for repost. */
2657 * Replace spent segments with new ones, concatenate and
2658 * return them as pkt_buf.
2661 struct ibv_sge *sge = &elt->sges[j];
2662 struct rte_mbuf *seg = elt->bufs[j];
2663 struct rte_mbuf *rep;
2664 unsigned int seg_tailroom;
2667 * Fetch initial bytes of packet descriptor into a
2668 * cacheline while allocating rep.
2671 rep = __rte_mbuf_raw_alloc(rxq->mp);
2672 if (unlikely(rep == NULL)) {
2674 * Unable to allocate a replacement mbuf,
2677 DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
2678 " can't allocate a new mbuf",
2679 (void *)rxq, wr_id);
2680 if (pkt_buf != NULL) {
2681 *pkt_buf_next = NULL;
2682 rte_pktmbuf_free(pkt_buf);
2684 /* Increase out of memory counters. */
2685 ++rxq->stats.rx_nombuf;
2686 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2690 /* Poison user-modifiable fields in rep. */
2691 NEXT(rep) = (void *)((uintptr_t)-1);
2692 SET_DATA_OFF(rep, 0xdead);
2693 DATA_LEN(rep) = 0xd00d;
2694 PKT_LEN(rep) = 0xdeadd00d;
2695 NB_SEGS(rep) = 0x2a;
2699 assert(rep->buf_len == seg->buf_len);
2700 assert(rep->buf_len == rxq->mb_len);
2701 /* Reconfigure sge to use rep instead of seg. */
2702 assert(sge->lkey == rxq->mr->lkey);
2703 sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
2706 /* Update pkt_buf if it's the first segment, or link
2707 * seg to the previous one and update pkt_buf_next. */
2708 *pkt_buf_next = seg;
2709 pkt_buf_next = &NEXT(seg);
2710 /* Update seg information. */
2711 seg_tailroom = (seg->buf_len - seg_headroom);
2712 assert(sge->length == seg_tailroom);
2713 SET_DATA_OFF(seg, seg_headroom);
2714 if (likely(len <= seg_tailroom)) {
2716 DATA_LEN(seg) = len;
2719 assert(rte_pktmbuf_headroom(seg) ==
2721 assert(rte_pktmbuf_tailroom(seg) ==
2722 (seg_tailroom - len));
2725 DATA_LEN(seg) = seg_tailroom;
2726 PKT_LEN(seg) = seg_tailroom;
2728 assert(rte_pktmbuf_headroom(seg) == seg_headroom);
2729 assert(rte_pktmbuf_tailroom(seg) == 0);
2730 /* Fix len and clear headroom for next segments. */
2731 len -= seg_tailroom;
2734 /* Update head and tail segments. */
2735 *pkt_buf_next = NULL;
2736 assert(pkt_buf != NULL);
2738 NB_SEGS(pkt_buf) = j;
2739 PORT(pkt_buf) = rxq->port_id;
2740 PKT_LEN(pkt_buf) = pkt_buf_len;
2741 pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
2742 pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
2744 /* Return packet. */
2745 *(pkts++) = pkt_buf;
2747 #ifdef MLX4_PMD_SOFT_COUNTERS
2748 /* Increase bytes counter. */
2749 rxq->stats.ibytes += pkt_buf_len;
2752 if (++elts_head >= elts_n)
2756 if (unlikely(i == 0))
2761 DEBUG("%p: reposting %d WRs", (void *)rxq, i);
2763 ret = ibv_post_recv(rxq->qp, head.next, &bad_wr);
2764 if (unlikely(ret)) {
2765 /* Inability to repost WRs is fatal. */
2766 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
2772 rxq->elts_head = elts_head;
2773 #ifdef MLX4_PMD_SOFT_COUNTERS
2774 /* Increase packets counter. */
2775 rxq->stats.ipackets += pkts_ret;
2781 * DPDK callback for RX.
2783 * The following function is the same as mlx4_rx_burst_sp(), except it doesn't
2784 * manage scattered packets. Improves performance when MRU is lower than the
2785 * size of the first segment.
2788 * Generic pointer to RX queue structure.
2790 * Array to store received packets.
2792 * Maximum number of packets in array.
2795 * Number of packets successfully received (<= pkts_n).
2798 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2800 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2801 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
2802 const unsigned int elts_n = rxq->elts_n;
2803 unsigned int elts_head = rxq->elts_head;
2804 struct ibv_sge sges[pkts_n];
2806 unsigned int pkts_ret = 0;
2809 if (unlikely(rxq->sp))
2810 return mlx4_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
2811 for (i = 0; (i != pkts_n); ++i) {
2812 struct rxq_elt *elt = &(*elts)[elts_head];
2813 struct ibv_recv_wr *wr = &elt->wr;
2814 uint64_t wr_id = wr->wr_id;
2816 struct rte_mbuf *seg = (void *)((uintptr_t)elt->sge.addr -
2817 WR_ID(wr_id).offset);
2818 struct rte_mbuf *rep;
2821 /* Sanity checks. */
2822 assert(WR_ID(wr_id).id < rxq->elts_n);
2823 assert(wr->sg_list == &elt->sge);
2824 assert(wr->num_sge == 1);
2825 assert(elts_head < rxq->elts_n);
2826 assert(rxq->elts_head < rxq->elts_n);
2828 * Fetch initial bytes of packet descriptor into a
2829 * cacheline while allocating rep.
2832 rte_prefetch0(&seg->cacheline1);
2833 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
2835 if (unlikely(ret < 0)) {
2839 DEBUG("rxq=%p, poll_length() failed (ret=%d)",
2841 /* ibv_poll_cq() must be used in case of failure. */
2842 wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
2843 if (unlikely(wcs_n == 0))
2845 if (unlikely(wcs_n < 0)) {
2846 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
2847 (void *)rxq, wcs_n);
2851 if (unlikely(wc.status != IBV_WC_SUCCESS)) {
2852 /* Whatever, just repost the offending WR. */
2853 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
2854 " completion status (%d): %s",
2855 (void *)rxq, wc.wr_id, wc.status,
2856 ibv_wc_status_str(wc.status));
2857 #ifdef MLX4_PMD_SOFT_COUNTERS
2858 /* Increment dropped packets counter. */
2859 ++rxq->stats.idropped;
2861 /* Add SGE to array for repost. */
2870 rep = __rte_mbuf_raw_alloc(rxq->mp);
2871 if (unlikely(rep == NULL)) {
2873 * Unable to allocate a replacement mbuf,
2876 DEBUG("rxq=%p, wr_id=%" PRIu32 ":"
2877 " can't allocate a new mbuf",
2878 (void *)rxq, WR_ID(wr_id).id);
2879 /* Increase out of memory counters. */
2880 ++rxq->stats.rx_nombuf;
2881 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2885 /* Reconfigure sge to use rep instead of seg. */
2886 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
2887 assert(elt->sge.lkey == rxq->mr->lkey);
2888 WR_ID(wr->wr_id).offset =
2889 (((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
2891 assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
2893 /* Add SGE to array for repost. */
2896 /* Update seg information. */
2897 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
2899 PORT(seg) = rxq->port_id;
2902 DATA_LEN(seg) = len;
2903 seg->packet_type = rxq_cq_to_pkt_type(flags);
2904 seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
2906 /* Return packet. */
2909 #ifdef MLX4_PMD_SOFT_COUNTERS
2910 /* Increase bytes counter. */
2911 rxq->stats.ibytes += len;
2914 if (++elts_head >= elts_n)
2918 if (unlikely(i == 0))
2922 DEBUG("%p: reposting %u WRs", (void *)rxq, i);
2924 ret = rxq->if_qp->recv_burst(rxq->qp, sges, i);
2925 if (unlikely(ret)) {
2926 /* Inability to repost WRs is fatal. */
2927 DEBUG("%p: recv_burst(): failed (ret=%d)",
2932 rxq->elts_head = elts_head;
2933 #ifdef MLX4_PMD_SOFT_COUNTERS
2934 /* Increase packets counter. */
2935 rxq->stats.ipackets += pkts_ret;
2941 * Allocate a Queue Pair.
2942 * Optionally setup inline receive if supported.
2945 * Pointer to private structure.
2947 * Completion queue to associate with QP.
2949 * Number of descriptors in QP (hint only).
2952 * QP pointer or NULL in case of error.
2954 static struct ibv_qp *
2955 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
2956 struct ibv_exp_res_domain *rd)
2958 struct ibv_exp_qp_init_attr attr = {
2959 /* CQ to be associated with the send queue. */
2961 /* CQ to be associated with the receive queue. */
2964 /* Max number of outstanding WRs. */
2965 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2966 priv->device_attr.max_qp_wr :
2968 /* Max number of scatter/gather elements in a WR. */
2969 .max_recv_sge = ((priv->device_attr.max_sge <
2970 MLX4_PMD_SGE_WR_N) ?
2971 priv->device_attr.max_sge :
2974 .qp_type = IBV_QPT_RAW_PACKET,
2975 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
2976 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN),
2982 attr.max_inl_recv = priv->inl_recv_size;
2983 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
2985 return ibv_exp_create_qp(priv->ctx, &attr);
2991 * Allocate a RSS Queue Pair.
2992 * Optionally setup inline receive if supported.
2995 * Pointer to private structure.
2997 * Completion queue to associate with QP.
2999 * Number of descriptors in QP (hint only).
3001 * If nonzero, create a parent QP, otherwise a child.
3004 * QP pointer or NULL in case of error.
3006 static struct ibv_qp *
3007 rxq_setup_qp_rss(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
3008 int parent, struct ibv_exp_res_domain *rd)
3010 struct ibv_exp_qp_init_attr attr = {
3011 /* CQ to be associated with the send queue. */
3013 /* CQ to be associated with the receive queue. */
3016 /* Max number of outstanding WRs. */
3017 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
3018 priv->device_attr.max_qp_wr :
3020 /* Max number of scatter/gather elements in a WR. */
3021 .max_recv_sge = ((priv->device_attr.max_sge <
3022 MLX4_PMD_SGE_WR_N) ?
3023 priv->device_attr.max_sge :
3026 .qp_type = IBV_QPT_RAW_PACKET,
3027 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
3028 IBV_EXP_QP_INIT_ATTR_RES_DOMAIN |
3029 IBV_EXP_QP_INIT_ATTR_QPG),
3035 attr.max_inl_recv = priv->inl_recv_size,
3036 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
3039 attr.qpg.qpg_type = IBV_EXP_QPG_PARENT;
3040 /* TSS isn't necessary. */
3041 attr.qpg.parent_attrib.tss_child_count = 0;
3042 attr.qpg.parent_attrib.rss_child_count = priv->rxqs_n;
3043 DEBUG("initializing parent RSS queue");
3045 attr.qpg.qpg_type = IBV_EXP_QPG_CHILD_RX;
3046 attr.qpg.qpg_parent = priv->rxq_parent.qp;
3047 DEBUG("initializing child RSS queue");
3049 return ibv_exp_create_qp(priv->ctx, &attr);
3052 #endif /* RSS_SUPPORT */
3055 * Reconfigure a RX queue with new parameters.
3057 * rxq_rehash() does not allocate mbufs, which, if not done from the right
3058 * thread (such as a control thread), may corrupt the pool.
3059 * In case of failure, the queue is left untouched.
3062 * Pointer to Ethernet device structure.
3067 * 0 on success, errno value on failure.
3070 rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
3072 struct priv *priv = rxq->priv;
3073 struct rxq tmpl = *rxq;
3074 unsigned int mbuf_n;
3075 unsigned int desc_n;
3076 struct rte_mbuf **pool;
3078 struct ibv_exp_qp_attr mod;
3079 struct ibv_recv_wr *bad_wr;
3081 int parent = (rxq == &priv->rxq_parent);
3084 ERROR("%p: cannot rehash parent queue %p",
3085 (void *)dev, (void *)rxq);
3088 DEBUG("%p: rehashing queue %p", (void *)dev, (void *)rxq);
3089 /* Number of descriptors and mbufs currently allocated. */
3090 desc_n = (tmpl.elts_n * (tmpl.sp ? MLX4_PMD_SGE_WR_N : 1));
3092 /* Toggle RX checksum offload if hardware supports it. */
3093 if (priv->hw_csum) {
3094 tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3095 rxq->csum = tmpl.csum;
3097 if (priv->hw_csum_l2tun) {
3098 tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3099 rxq->csum_l2tun = tmpl.csum_l2tun;
3101 /* Enable scattered packets support for this queue if necessary. */
3102 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3103 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3104 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3106 desc_n /= MLX4_PMD_SGE_WR_N;
3109 DEBUG("%p: %s scattered packets support (%u WRs)",
3110 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc_n);
3111 /* If scatter mode is the same as before, nothing to do. */
3112 if (tmpl.sp == rxq->sp) {
3113 DEBUG("%p: nothing to do", (void *)dev);
3116 /* Remove attached flows if RSS is disabled (no parent queue). */
3118 rxq_allmulticast_disable(&tmpl);
3119 rxq_promiscuous_disable(&tmpl);
3120 rxq_mac_addrs_del(&tmpl);
3121 /* Update original queue in case of failure. */
3122 rxq->allmulti_flow = tmpl.allmulti_flow;
3123 rxq->promisc_flow = tmpl.promisc_flow;
3124 memcpy(rxq->mac_configured, tmpl.mac_configured,
3125 sizeof(rxq->mac_configured));
3126 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
3128 /* From now on, any failure will render the queue unusable.
3129 * Reinitialize QP. */
3130 mod = (struct ibv_exp_qp_attr){ .qp_state = IBV_QPS_RESET };
3131 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3133 ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
3137 err = ibv_resize_cq(tmpl.cq, desc_n);
3139 ERROR("%p: cannot resize CQ: %s", (void *)dev, strerror(err));
3143 mod = (struct ibv_exp_qp_attr){
3144 /* Move the QP to this state. */
3145 .qp_state = IBV_QPS_INIT,
3146 /* Primary port number. */
3147 .port_num = priv->port
3149 err = ibv_exp_modify_qp(tmpl.qp, &mod,
3152 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3153 #endif /* RSS_SUPPORT */
3156 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3157 (void *)dev, strerror(err));
3161 /* Reconfigure flows. Do not care for errors. */
3163 rxq_mac_addrs_add(&tmpl);
3165 rxq_promiscuous_enable(&tmpl);
3167 rxq_allmulticast_enable(&tmpl);
3168 /* Update original queue in case of failure. */
3169 rxq->allmulti_flow = tmpl.allmulti_flow;
3170 rxq->promisc_flow = tmpl.promisc_flow;
3171 memcpy(rxq->mac_configured, tmpl.mac_configured,
3172 sizeof(rxq->mac_configured));
3173 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
3175 /* Allocate pool. */
3176 pool = rte_malloc(__func__, (mbuf_n * sizeof(*pool)), 0);
3178 ERROR("%p: cannot allocate memory", (void *)dev);
3181 /* Snatch mbufs from original queue. */
3184 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
3186 for (i = 0; (i != elemof(*elts)); ++i) {
3187 struct rxq_elt_sp *elt = &(*elts)[i];
3190 for (j = 0; (j != elemof(elt->bufs)); ++j) {
3191 assert(elt->bufs[j] != NULL);
3192 pool[k++] = elt->bufs[j];
3196 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
3198 for (i = 0; (i != elemof(*elts)); ++i) {
3199 struct rxq_elt *elt = &(*elts)[i];
3200 struct rte_mbuf *buf = (void *)
3201 ((uintptr_t)elt->sge.addr -
3202 WR_ID(elt->wr.wr_id).offset);
3204 assert(WR_ID(elt->wr.wr_id).id == i);
3208 assert(k == mbuf_n);
3210 tmpl.elts.sp = NULL;
3211 assert((void *)&tmpl.elts.sp == (void *)&tmpl.elts.no_sp);
3213 rxq_alloc_elts_sp(&tmpl, desc_n, pool) :
3214 rxq_alloc_elts(&tmpl, desc_n, pool));
3216 ERROR("%p: cannot reallocate WRs, aborting", (void *)dev);
3221 assert(tmpl.elts_n == desc_n);
3222 assert(tmpl.elts.sp != NULL);
3224 /* Clean up original data. */
3226 rte_free(rxq->elts.sp);
3227 rxq->elts.sp = NULL;
3229 err = ibv_post_recv(tmpl.qp,
3231 &(*tmpl.elts.sp)[0].wr :
3232 &(*tmpl.elts.no_sp)[0].wr),
3235 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3241 mod = (struct ibv_exp_qp_attr){
3242 .qp_state = IBV_QPS_RTR
3244 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3246 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3247 (void *)dev, strerror(err));
3255 * Configure a RX queue.
3258 * Pointer to Ethernet device structure.
3260 * Pointer to RX queue structure.
3262 * Number of descriptors to configure in queue.
3264 * NUMA socket on which memory must be allocated.
3266 * Thresholds parameters.
3268 * Memory pool for buffer allocations.
3271 * 0 on success, errno value on failure.
3274 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
3275 unsigned int socket, const struct rte_eth_rxconf *conf,
3276 struct rte_mempool *mp)
3278 struct priv *priv = dev->data->dev_private;
3284 struct ibv_exp_qp_attr mod;
3286 struct ibv_exp_query_intf_params params;
3287 struct ibv_exp_cq_init_attr cq;
3288 struct ibv_exp_res_domain_init_attr rd;
3290 enum ibv_exp_query_intf_status status;
3291 struct ibv_recv_wr *bad_wr;
3292 struct rte_mbuf *buf;
3294 int parent = (rxq == &priv->rxq_parent);
3296 (void)conf; /* Thresholds configuration (ignored). */
3298 * If this is a parent queue, hardware must support RSS and
3299 * RSS must be enabled.
3301 assert((!parent) || ((priv->hw_rss) && (priv->rss)));
3303 /* Even if unused, ibv_create_cq() requires at least one
3308 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
3309 ERROR("%p: invalid number of RX descriptors (must be a"
3310 " multiple of %d)", (void *)dev, MLX4_PMD_SGE_WR_N);
3313 /* Get mbuf length. */
3314 buf = rte_pktmbuf_alloc(mp);
3316 ERROR("%p: unable to allocate mbuf", (void *)dev);
3319 tmpl.mb_len = buf->buf_len;
3320 assert((rte_pktmbuf_headroom(buf) +
3321 rte_pktmbuf_tailroom(buf)) == tmpl.mb_len);
3322 assert(rte_pktmbuf_headroom(buf) == RTE_PKTMBUF_HEADROOM);
3323 rte_pktmbuf_free(buf);
3324 /* Toggle RX checksum offload if hardware supports it. */
3326 tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3327 if (priv->hw_csum_l2tun)
3328 tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
3329 /* Enable scattered packets support for this queue if necessary. */
3330 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3331 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3332 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3334 desc /= MLX4_PMD_SGE_WR_N;
3336 DEBUG("%p: %s scattered packets support (%u WRs)",
3337 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc);
3338 /* Use the entire RX mempool as the memory region. */
3339 tmpl.mr = ibv_reg_mr(priv->pd,
3340 (void *)mp->elt_va_start,
3341 (mp->elt_va_end - mp->elt_va_start),
3342 (IBV_ACCESS_LOCAL_WRITE |
3343 IBV_ACCESS_REMOTE_WRITE));
3344 if (tmpl.mr == NULL) {
3346 ERROR("%p: MR creation failure: %s",
3347 (void *)dev, strerror(ret));
3351 attr.rd = (struct ibv_exp_res_domain_init_attr){
3352 .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
3353 IBV_EXP_RES_DOMAIN_MSG_MODEL),
3354 .thread_model = IBV_EXP_THREAD_SINGLE,
3355 .msg_model = IBV_EXP_MSG_HIGH_BW,
3357 tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
3358 if (tmpl.rd == NULL) {
3360 ERROR("%p: RD creation failure: %s",
3361 (void *)dev, strerror(ret));
3364 attr.cq = (struct ibv_exp_cq_init_attr){
3365 .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
3366 .res_domain = tmpl.rd,
3368 tmpl.cq = ibv_exp_create_cq(priv->ctx, desc, NULL, NULL, 0, &attr.cq);
3369 if (tmpl.cq == NULL) {
3371 ERROR("%p: CQ creation failure: %s",
3372 (void *)dev, strerror(ret));
3375 DEBUG("priv->device_attr.max_qp_wr is %d",
3376 priv->device_attr.max_qp_wr);
3377 DEBUG("priv->device_attr.max_sge is %d",
3378 priv->device_attr.max_sge);
3381 tmpl.qp = rxq_setup_qp_rss(priv, tmpl.cq, desc, parent,
3384 #endif /* RSS_SUPPORT */
3385 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc, tmpl.rd);
3386 if (tmpl.qp == NULL) {
3387 ret = (errno ? errno : EINVAL);
3388 ERROR("%p: QP creation failure: %s",
3389 (void *)dev, strerror(ret));
3392 mod = (struct ibv_exp_qp_attr){
3393 /* Move the QP to this state. */
3394 .qp_state = IBV_QPS_INIT,
3395 /* Primary port number. */
3396 .port_num = priv->port
3398 ret = ibv_exp_modify_qp(tmpl.qp, &mod,
3401 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3402 #endif /* RSS_SUPPORT */
3405 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3406 (void *)dev, strerror(ret));
3409 if ((parent) || (!priv->rss)) {
3410 /* Configure MAC and broadcast addresses. */
3411 ret = rxq_mac_addrs_add(&tmpl);
3413 ERROR("%p: QP flow attachment failed: %s",
3414 (void *)dev, strerror(ret));
3418 /* Allocate descriptors for RX queues, except for the RSS parent. */
3422 ret = rxq_alloc_elts_sp(&tmpl, desc, NULL);
3424 ret = rxq_alloc_elts(&tmpl, desc, NULL);
3426 ERROR("%p: RXQ allocation failed: %s",
3427 (void *)dev, strerror(ret));
3430 ret = ibv_post_recv(tmpl.qp,
3432 &(*tmpl.elts.sp)[0].wr :
3433 &(*tmpl.elts.no_sp)[0].wr),
3436 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3443 mod = (struct ibv_exp_qp_attr){
3444 .qp_state = IBV_QPS_RTR
3446 ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3448 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3449 (void *)dev, strerror(ret));
3453 tmpl.port_id = dev->data->port_id;
3454 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
3455 attr.params = (struct ibv_exp_query_intf_params){
3456 .intf_scope = IBV_EXP_INTF_GLOBAL,
3457 .intf = IBV_EXP_INTF_CQ,
3460 tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
3461 if (tmpl.if_cq == NULL) {
3462 ERROR("%p: CQ interface family query failed with status %d",
3463 (void *)dev, status);
3466 attr.params = (struct ibv_exp_query_intf_params){
3467 .intf_scope = IBV_EXP_INTF_GLOBAL,
3468 .intf = IBV_EXP_INTF_QP_BURST,
3471 tmpl.if_qp = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
3472 if (tmpl.if_qp == NULL) {
3473 ERROR("%p: QP interface family query failed with status %d",
3474 (void *)dev, status);
3477 /* Clean up rxq in case we're reinitializing it. */
3478 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
3481 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
3491 * DPDK callback to configure a RX queue.
3494 * Pointer to Ethernet device structure.
3498 * Number of descriptors to configure in queue.
3500 * NUMA socket on which memory must be allocated.
3502 * Thresholds parameters.
3504 * Memory pool for buffer allocations.
3507 * 0 on success, negative errno value on failure.
3510 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
3511 unsigned int socket, const struct rte_eth_rxconf *conf,
3512 struct rte_mempool *mp)
3514 struct priv *priv = dev->data->dev_private;
3515 struct rxq *rxq = (*priv->rxqs)[idx];
3519 DEBUG("%p: configuring queue %u for %u descriptors",
3520 (void *)dev, idx, desc);
3521 if (idx >= priv->rxqs_n) {
3522 ERROR("%p: queue index out of range (%u >= %u)",
3523 (void *)dev, idx, priv->rxqs_n);
3528 DEBUG("%p: reusing already allocated queue index %u (%p)",
3529 (void *)dev, idx, (void *)rxq);
3530 if (priv->started) {
3534 (*priv->rxqs)[idx] = NULL;
3537 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
3539 ERROR("%p: unable to allocate queue index %u",
3545 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
3549 rxq->stats.idx = idx;
3550 DEBUG("%p: adding RX queue %p to list",
3551 (void *)dev, (void *)rxq);
3552 (*priv->rxqs)[idx] = rxq;
3553 /* Update receive callback. */
3555 dev->rx_pkt_burst = mlx4_rx_burst_sp;
3557 dev->rx_pkt_burst = mlx4_rx_burst;
3564 * DPDK callback to release a RX queue.
3567 * Generic RX queue pointer.
3570 mlx4_rx_queue_release(void *dpdk_rxq)
3572 struct rxq *rxq = (struct rxq *)dpdk_rxq;
3580 assert(rxq != &priv->rxq_parent);
3581 for (i = 0; (i != priv->rxqs_n); ++i)
3582 if ((*priv->rxqs)[i] == rxq) {
3583 DEBUG("%p: removing RX queue %p from list",
3584 (void *)priv->dev, (void *)rxq);
3585 (*priv->rxqs)[i] = NULL;
3594 priv_dev_interrupt_handler_install(struct priv *, struct rte_eth_dev *);
3597 * DPDK callback to start the device.
3599 * Simulate device start by attaching all configured flows.
3602 * Pointer to Ethernet device structure.
3605 * 0 on success, negative errno value on failure.
3608 mlx4_dev_start(struct rte_eth_dev *dev)
3610 struct priv *priv = dev->data->dev_private;
3616 if (priv->started) {
3620 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
3623 rxq = &priv->rxq_parent;
3626 rxq = (*priv->rxqs)[0];
3629 /* Iterate only once when RSS is enabled. */
3633 /* Ignore nonexistent RX queues. */
3636 ret = rxq_mac_addrs_add(rxq);
3637 if (!ret && priv->promisc)
3638 ret = rxq_promiscuous_enable(rxq);
3639 if (!ret && priv->allmulti)
3640 ret = rxq_allmulticast_enable(rxq);
3643 WARN("%p: QP flow attachment failed: %s",
3644 (void *)dev, strerror(ret));
3647 rxq = (*priv->rxqs)[--i];
3649 rxq_allmulticast_disable(rxq);
3650 rxq_promiscuous_disable(rxq);
3651 rxq_mac_addrs_del(rxq);
3657 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3658 priv_dev_interrupt_handler_install(priv, dev);
3664 * DPDK callback to stop the device.
3666 * Simulate device stop by detaching all configured flows.
3669 * Pointer to Ethernet device structure.
3672 mlx4_dev_stop(struct rte_eth_dev *dev)
3674 struct priv *priv = dev->data->dev_private;
3680 if (!priv->started) {
3684 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
3687 rxq = &priv->rxq_parent;
3690 rxq = (*priv->rxqs)[0];
3693 /* Iterate only once when RSS is enabled. */
3695 /* Ignore nonexistent RX queues. */
3698 rxq_allmulticast_disable(rxq);
3699 rxq_promiscuous_disable(rxq);
3700 rxq_mac_addrs_del(rxq);
3701 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3706 * Dummy DPDK callback for TX.
3708 * This function is used to temporarily replace the real callback during
3709 * unsafe control operations on the queue, or in case of error.
3712 * Generic pointer to TX queue structure.
3714 * Packets to transmit.
3716 * Number of packets in array.
3719 * Number of packets successfully transmitted (<= pkts_n).
3722 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
3731 * Dummy DPDK callback for RX.
3733 * This function is used to temporarily replace the real callback during
3734 * unsafe control operations on the queue, or in case of error.
3737 * Generic pointer to RX queue structure.
3739 * Array to store received packets.
3741 * Maximum number of packets in array.
3744 * Number of packets successfully received (<= pkts_n).
3747 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
3756 priv_dev_interrupt_handler_uninstall(struct priv *, struct rte_eth_dev *);
3759 * DPDK callback to close the device.
3761 * Destroy all queues and objects, free memory.
3764 * Pointer to Ethernet device structure.
3767 mlx4_dev_close(struct rte_eth_dev *dev)
3769 struct priv *priv = dev->data->dev_private;
3774 DEBUG("%p: closing device \"%s\"",
3776 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
3777 /* Prevent crashes when queues are still in use. This is unfortunately
3778 * still required for DPDK 1.3 because some programs (such as testpmd)
3779 * never release them before closing the device. */
3780 dev->rx_pkt_burst = removed_rx_burst;
3781 dev->tx_pkt_burst = removed_tx_burst;
3782 if (priv->rxqs != NULL) {
3783 /* XXX race condition if mlx4_rx_burst() is still running. */
3785 for (i = 0; (i != priv->rxqs_n); ++i) {
3786 tmp = (*priv->rxqs)[i];
3789 (*priv->rxqs)[i] = NULL;
3796 if (priv->txqs != NULL) {
3797 /* XXX race condition if mlx4_tx_burst() is still running. */
3799 for (i = 0; (i != priv->txqs_n); ++i) {
3800 tmp = (*priv->txqs)[i];
3803 (*priv->txqs)[i] = NULL;
3811 rxq_cleanup(&priv->rxq_parent);
3812 if (priv->pd != NULL) {
3813 assert(priv->ctx != NULL);
3814 claim_zero(ibv_dealloc_pd(priv->pd));
3815 claim_zero(ibv_close_device(priv->ctx));
3817 assert(priv->ctx == NULL);
3818 priv_dev_interrupt_handler_uninstall(priv, dev);
3820 memset(priv, 0, sizeof(*priv));
3824 * DPDK callback to get information about the device.
3827 * Pointer to Ethernet device structure.
3829 * Info structure output buffer.
3832 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
3834 struct priv *priv = dev->data->dev_private;
3836 char ifname[IF_NAMESIZE];
3839 /* FIXME: we should ask the device for these values. */
3840 info->min_rx_bufsize = 32;
3841 info->max_rx_pktlen = 65536;
3843 * Since we need one CQ per QP, the limit is the minimum number
3844 * between the two values.
3846 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
3847 priv->device_attr.max_qp : priv->device_attr.max_cq);
3848 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
3851 info->max_rx_queues = max;
3852 info->max_tx_queues = max;
3853 /* Last array entry is reserved for broadcast. */
3854 info->max_mac_addrs = (elemof(priv->mac) - 1);
3855 info->rx_offload_capa =
3857 (DEV_RX_OFFLOAD_IPV4_CKSUM |
3858 DEV_RX_OFFLOAD_UDP_CKSUM |
3859 DEV_RX_OFFLOAD_TCP_CKSUM) :
3861 info->tx_offload_capa =
3863 (DEV_TX_OFFLOAD_IPV4_CKSUM |
3864 DEV_TX_OFFLOAD_UDP_CKSUM |
3865 DEV_TX_OFFLOAD_TCP_CKSUM) :
3867 if (priv_get_ifname(priv, &ifname) == 0)
3868 info->if_index = if_nametoindex(ifname);
3873 * DPDK callback to get device statistics.
3876 * Pointer to Ethernet device structure.
3878 * Stats structure output buffer.
3881 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
3883 struct priv *priv = dev->data->dev_private;
3884 struct rte_eth_stats tmp = {0};
3889 /* Add software counters. */
3890 for (i = 0; (i != priv->rxqs_n); ++i) {
3891 struct rxq *rxq = (*priv->rxqs)[i];
3895 idx = rxq->stats.idx;
3896 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3897 #ifdef MLX4_PMD_SOFT_COUNTERS
3898 tmp.q_ipackets[idx] += rxq->stats.ipackets;
3899 tmp.q_ibytes[idx] += rxq->stats.ibytes;
3901 tmp.q_errors[idx] += (rxq->stats.idropped +
3902 rxq->stats.rx_nombuf);
3904 #ifdef MLX4_PMD_SOFT_COUNTERS
3905 tmp.ipackets += rxq->stats.ipackets;
3906 tmp.ibytes += rxq->stats.ibytes;
3908 tmp.ierrors += rxq->stats.idropped;
3909 tmp.rx_nombuf += rxq->stats.rx_nombuf;
3911 for (i = 0; (i != priv->txqs_n); ++i) {
3912 struct txq *txq = (*priv->txqs)[i];
3916 idx = txq->stats.idx;
3917 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3918 #ifdef MLX4_PMD_SOFT_COUNTERS
3919 tmp.q_opackets[idx] += txq->stats.opackets;
3920 tmp.q_obytes[idx] += txq->stats.obytes;
3922 tmp.q_errors[idx] += txq->stats.odropped;
3924 #ifdef MLX4_PMD_SOFT_COUNTERS
3925 tmp.opackets += txq->stats.opackets;
3926 tmp.obytes += txq->stats.obytes;
3928 tmp.oerrors += txq->stats.odropped;
3930 #ifndef MLX4_PMD_SOFT_COUNTERS
3931 /* FIXME: retrieve and add hardware counters. */
3938 * DPDK callback to clear device statistics.
3941 * Pointer to Ethernet device structure.
3944 mlx4_stats_reset(struct rte_eth_dev *dev)
3946 struct priv *priv = dev->data->dev_private;
3951 for (i = 0; (i != priv->rxqs_n); ++i) {
3952 if ((*priv->rxqs)[i] == NULL)
3954 idx = (*priv->rxqs)[i]->stats.idx;
3955 (*priv->rxqs)[i]->stats =
3956 (struct mlx4_rxq_stats){ .idx = idx };
3958 for (i = 0; (i != priv->txqs_n); ++i) {
3959 if ((*priv->txqs)[i] == NULL)
3961 idx = (*priv->rxqs)[i]->stats.idx;
3962 (*priv->txqs)[i]->stats =
3963 (struct mlx4_txq_stats){ .idx = idx };
3965 #ifndef MLX4_PMD_SOFT_COUNTERS
3966 /* FIXME: reset hardware counters. */
3972 * DPDK callback to remove a MAC address.
3975 * Pointer to Ethernet device structure.
3977 * MAC address index.
3980 mlx4_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
3982 struct priv *priv = dev->data->dev_private;
3985 DEBUG("%p: removing MAC address from index %" PRIu32,
3986 (void *)dev, index);
3987 /* Last array entry is reserved for broadcast. */
3988 if (index >= (elemof(priv->mac) - 1))
3990 priv_mac_addr_del(priv, index);
3996 * DPDK callback to add a MAC address.
3999 * Pointer to Ethernet device structure.
4001 * MAC address to register.
4003 * MAC address index.
4005 * VMDq pool index to associate address with (ignored).
4008 mlx4_mac_addr_add(struct rte_eth_dev *dev, struct ether_addr *mac_addr,
4009 uint32_t index, uint32_t vmdq)
4011 struct priv *priv = dev->data->dev_private;
4015 DEBUG("%p: adding MAC address at index %" PRIu32,
4016 (void *)dev, index);
4017 /* Last array entry is reserved for broadcast. */
4018 if (index >= (elemof(priv->mac) - 1))
4020 priv_mac_addr_add(priv, index,
4021 (const uint8_t (*)[ETHER_ADDR_LEN])
4022 mac_addr->addr_bytes);
4028 * DPDK callback to enable promiscuous mode.
4031 * Pointer to Ethernet device structure.
4034 mlx4_promiscuous_enable(struct rte_eth_dev *dev)
4036 struct priv *priv = dev->data->dev_private;
4041 if (priv->promisc) {
4045 /* If device isn't started, this is all we need to do. */
4049 ret = rxq_promiscuous_enable(&priv->rxq_parent);
4056 for (i = 0; (i != priv->rxqs_n); ++i) {
4057 if ((*priv->rxqs)[i] == NULL)
4059 ret = rxq_promiscuous_enable((*priv->rxqs)[i]);
4062 /* Failure, rollback. */
4064 if ((*priv->rxqs)[--i] != NULL)
4065 rxq_promiscuous_disable((*priv->rxqs)[i]);
4075 * DPDK callback to disable promiscuous mode.
4078 * Pointer to Ethernet device structure.
4081 mlx4_promiscuous_disable(struct rte_eth_dev *dev)
4083 struct priv *priv = dev->data->dev_private;
4087 if (!priv->promisc) {
4092 rxq_promiscuous_disable(&priv->rxq_parent);
4095 for (i = 0; (i != priv->rxqs_n); ++i)
4096 if ((*priv->rxqs)[i] != NULL)
4097 rxq_promiscuous_disable((*priv->rxqs)[i]);
4104 * DPDK callback to enable allmulti mode.
4107 * Pointer to Ethernet device structure.
4110 mlx4_allmulticast_enable(struct rte_eth_dev *dev)
4112 struct priv *priv = dev->data->dev_private;
4117 if (priv->allmulti) {
4121 /* If device isn't started, this is all we need to do. */
4125 ret = rxq_allmulticast_enable(&priv->rxq_parent);
4132 for (i = 0; (i != priv->rxqs_n); ++i) {
4133 if ((*priv->rxqs)[i] == NULL)
4135 ret = rxq_allmulticast_enable((*priv->rxqs)[i]);
4138 /* Failure, rollback. */
4140 if ((*priv->rxqs)[--i] != NULL)
4141 rxq_allmulticast_disable((*priv->rxqs)[i]);
4151 * DPDK callback to disable allmulti mode.
4154 * Pointer to Ethernet device structure.
4157 mlx4_allmulticast_disable(struct rte_eth_dev *dev)
4159 struct priv *priv = dev->data->dev_private;
4163 if (!priv->allmulti) {
4168 rxq_allmulticast_disable(&priv->rxq_parent);
4171 for (i = 0; (i != priv->rxqs_n); ++i)
4172 if ((*priv->rxqs)[i] != NULL)
4173 rxq_allmulticast_disable((*priv->rxqs)[i]);
4180 * DPDK callback to retrieve physical link information (unlocked version).
4183 * Pointer to Ethernet device structure.
4184 * @param wait_to_complete
4185 * Wait for request completion (ignored).
4188 mlx4_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
4190 struct priv *priv = dev->data->dev_private;
4191 struct ethtool_cmd edata = {
4195 struct rte_eth_link dev_link;
4198 (void)wait_to_complete;
4199 if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
4200 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
4203 memset(&dev_link, 0, sizeof(dev_link));
4204 dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
4205 (ifr.ifr_flags & IFF_RUNNING));
4206 ifr.ifr_data = &edata;
4207 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4208 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
4212 link_speed = ethtool_cmd_speed(&edata);
4213 if (link_speed == -1)
4214 dev_link.link_speed = 0;
4216 dev_link.link_speed = link_speed;
4217 dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
4218 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
4219 if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
4220 /* Link status changed. */
4221 dev->data->dev_link = dev_link;
4224 /* Link status is still the same. */
4229 * DPDK callback to retrieve physical link information.
4232 * Pointer to Ethernet device structure.
4233 * @param wait_to_complete
4234 * Wait for request completion (ignored).
4237 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
4239 struct priv *priv = dev->data->dev_private;
4243 ret = mlx4_link_update_unlocked(dev, wait_to_complete);
4249 * DPDK callback to change the MTU.
4251 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
4252 * received). Use this as a hint to enable/disable scattered packets support
4253 * and improve performance when not needed.
4254 * Since failure is not an option, reconfiguring queues on the fly is not
4258 * Pointer to Ethernet device structure.
4263 * 0 on success, negative errno value on failure.
4266 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
4268 struct priv *priv = dev->data->dev_private;
4271 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
4275 /* Set kernel interface MTU first. */
4276 if (priv_set_mtu(priv, mtu)) {
4278 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
4282 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
4284 /* Temporarily replace RX handler with a fake one, assuming it has not
4285 * been copied elsewhere. */
4286 dev->rx_pkt_burst = removed_rx_burst;
4287 /* Make sure everyone has left mlx4_rx_burst() and uses
4288 * removed_rx_burst() instead. */
4291 /* Reconfigure each RX queue. */
4292 for (i = 0; (i != priv->rxqs_n); ++i) {
4293 struct rxq *rxq = (*priv->rxqs)[i];
4294 unsigned int max_frame_len;
4299 /* Calculate new maximum frame length according to MTU and
4300 * toggle scattered support (sp) if necessary. */
4301 max_frame_len = (priv->mtu + ETHER_HDR_LEN +
4302 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
4303 sp = (max_frame_len > (rxq->mb_len - RTE_PKTMBUF_HEADROOM));
4304 /* Provide new values to rxq_setup(). */
4305 dev->data->dev_conf.rxmode.jumbo_frame = sp;
4306 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
4307 ret = rxq_rehash(dev, rxq);
4309 /* Force SP RX if that queue requires it and abort. */
4311 rx_func = mlx4_rx_burst_sp;
4314 /* Reenable non-RSS queue attributes. No need to check
4315 * for errors at this stage. */
4317 rxq_mac_addrs_add(rxq);
4319 rxq_promiscuous_enable(rxq);
4321 rxq_allmulticast_enable(rxq);
4323 /* Scattered burst function takes priority. */
4325 rx_func = mlx4_rx_burst_sp;
4327 /* Burst functions can now be called again. */
4329 dev->rx_pkt_burst = rx_func;
4337 * DPDK callback to get flow control status.
4340 * Pointer to Ethernet device structure.
4341 * @param[out] fc_conf
4342 * Flow control output buffer.
4345 * 0 on success, negative errno value on failure.
4348 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4350 struct priv *priv = dev->data->dev_private;
4352 struct ethtool_pauseparam ethpause = {
4353 .cmd = ETHTOOL_GPAUSEPARAM
4357 ifr.ifr_data = ðpause;
4359 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4361 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
4367 fc_conf->autoneg = ethpause.autoneg;
4368 if (ethpause.rx_pause && ethpause.tx_pause)
4369 fc_conf->mode = RTE_FC_FULL;
4370 else if (ethpause.rx_pause)
4371 fc_conf->mode = RTE_FC_RX_PAUSE;
4372 else if (ethpause.tx_pause)
4373 fc_conf->mode = RTE_FC_TX_PAUSE;
4375 fc_conf->mode = RTE_FC_NONE;
4385 * DPDK callback to modify flow control parameters.
4388 * Pointer to Ethernet device structure.
4389 * @param[in] fc_conf
4390 * Flow control parameters.
4393 * 0 on success, negative errno value on failure.
4396 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4398 struct priv *priv = dev->data->dev_private;
4400 struct ethtool_pauseparam ethpause = {
4401 .cmd = ETHTOOL_SPAUSEPARAM
4405 ifr.ifr_data = ðpause;
4406 ethpause.autoneg = fc_conf->autoneg;
4407 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4408 (fc_conf->mode & RTE_FC_RX_PAUSE))
4409 ethpause.rx_pause = 1;
4411 ethpause.rx_pause = 0;
4413 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4414 (fc_conf->mode & RTE_FC_TX_PAUSE))
4415 ethpause.tx_pause = 1;
4417 ethpause.tx_pause = 0;
4420 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4422 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
4436 * Configure a VLAN filter.
4439 * Pointer to Ethernet device structure.
4441 * VLAN ID to filter.
4446 * 0 on success, errno value on failure.
4449 vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4451 struct priv *priv = dev->data->dev_private;
4453 unsigned int j = -1;
4455 DEBUG("%p: %s VLAN filter ID %" PRIu16,
4456 (void *)dev, (on ? "enable" : "disable"), vlan_id);
4457 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
4458 if (!priv->vlan_filter[i].enabled) {
4459 /* Unused index, remember it. */
4463 if (priv->vlan_filter[i].id != vlan_id)
4465 /* This VLAN ID is already known, use its index. */
4469 /* Check if there's room for another VLAN filter. */
4470 if (j == (unsigned int)-1)
4473 * VLAN filters apply to all configured MAC addresses, flow
4474 * specifications must be reconfigured accordingly.
4476 priv->vlan_filter[j].id = vlan_id;
4477 if ((on) && (!priv->vlan_filter[j].enabled)) {
4479 * Filter is disabled, enable it.
4480 * Rehashing flows in all RX queues is necessary.
4483 rxq_mac_addrs_del(&priv->rxq_parent);
4485 for (i = 0; (i != priv->rxqs_n); ++i)
4486 if ((*priv->rxqs)[i] != NULL)
4487 rxq_mac_addrs_del((*priv->rxqs)[i]);
4488 priv->vlan_filter[j].enabled = 1;
4489 if (priv->started) {
4491 rxq_mac_addrs_add(&priv->rxq_parent);
4493 for (i = 0; (i != priv->rxqs_n); ++i) {
4494 if ((*priv->rxqs)[i] == NULL)
4496 rxq_mac_addrs_add((*priv->rxqs)[i]);
4499 } else if ((!on) && (priv->vlan_filter[j].enabled)) {
4501 * Filter is enabled, disable it.
4502 * Rehashing flows in all RX queues is necessary.
4505 rxq_mac_addrs_del(&priv->rxq_parent);
4507 for (i = 0; (i != priv->rxqs_n); ++i)
4508 if ((*priv->rxqs)[i] != NULL)
4509 rxq_mac_addrs_del((*priv->rxqs)[i]);
4510 priv->vlan_filter[j].enabled = 0;
4511 if (priv->started) {
4513 rxq_mac_addrs_add(&priv->rxq_parent);
4515 for (i = 0; (i != priv->rxqs_n); ++i) {
4516 if ((*priv->rxqs)[i] == NULL)
4518 rxq_mac_addrs_add((*priv->rxqs)[i]);
4526 * DPDK callback to configure a VLAN filter.
4529 * Pointer to Ethernet device structure.
4531 * VLAN ID to filter.
4536 * 0 on success, negative errno value on failure.
4539 mlx4_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4541 struct priv *priv = dev->data->dev_private;
4545 ret = vlan_filter_set(dev, vlan_id, on);
4551 static const struct eth_dev_ops mlx4_dev_ops = {
4552 .dev_configure = mlx4_dev_configure,
4553 .dev_start = mlx4_dev_start,
4554 .dev_stop = mlx4_dev_stop,
4555 .dev_close = mlx4_dev_close,
4556 .promiscuous_enable = mlx4_promiscuous_enable,
4557 .promiscuous_disable = mlx4_promiscuous_disable,
4558 .allmulticast_enable = mlx4_allmulticast_enable,
4559 .allmulticast_disable = mlx4_allmulticast_disable,
4560 .link_update = mlx4_link_update,
4561 .stats_get = mlx4_stats_get,
4562 .stats_reset = mlx4_stats_reset,
4563 .queue_stats_mapping_set = NULL,
4564 .dev_infos_get = mlx4_dev_infos_get,
4565 .vlan_filter_set = mlx4_vlan_filter_set,
4566 .vlan_tpid_set = NULL,
4567 .vlan_strip_queue_set = NULL,
4568 .vlan_offload_set = NULL,
4569 .rx_queue_setup = mlx4_rx_queue_setup,
4570 .tx_queue_setup = mlx4_tx_queue_setup,
4571 .rx_queue_release = mlx4_rx_queue_release,
4572 .tx_queue_release = mlx4_tx_queue_release,
4574 .dev_led_off = NULL,
4575 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
4576 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
4577 .priority_flow_ctrl_set = NULL,
4578 .mac_addr_remove = mlx4_mac_addr_remove,
4579 .mac_addr_add = mlx4_mac_addr_add,
4580 .mtu_set = mlx4_dev_set_mtu,
4581 .udp_tunnel_add = NULL,
4582 .udp_tunnel_del = NULL,
4586 * Get PCI information from struct ibv_device.
4589 * Pointer to Ethernet device structure.
4590 * @param[out] pci_addr
4591 * PCI bus address output buffer.
4594 * 0 on success, -1 on failure and errno is set.
4597 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
4598 struct rte_pci_addr *pci_addr)
4602 MKSTR(path, "%s/device/uevent", device->ibdev_path);
4604 file = fopen(path, "rb");
4607 while (fgets(line, sizeof(line), file) == line) {
4608 size_t len = strlen(line);
4611 /* Truncate long lines. */
4612 if (len == (sizeof(line) - 1))
4613 while (line[(len - 1)] != '\n') {
4617 line[(len - 1)] = ret;
4619 /* Extract information. */
4622 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
4626 &pci_addr->function) == 4) {
4636 * Get MAC address by querying netdevice.
4639 * struct priv for the requested device.
4641 * MAC address output buffer.
4644 * 0 on success, -1 on failure and errno is set.
4647 priv_get_mac(struct priv *priv, uint8_t (*mac)[ETHER_ADDR_LEN])
4649 struct ifreq request;
4651 if (priv_ifreq(priv, SIOCGIFHWADDR, &request))
4653 memcpy(mac, request.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
4657 /* Support up to 32 adapters. */
4659 struct rte_pci_addr pci_addr; /* associated PCI address */
4660 uint32_t ports; /* physical ports bitfield. */
4664 * Get device index in mlx4_dev[] from PCI bus address.
4666 * @param[in] pci_addr
4667 * PCI bus address to look for.
4670 * mlx4_dev[] index on success, -1 on failure.
4673 mlx4_dev_idx(struct rte_pci_addr *pci_addr)
4678 assert(pci_addr != NULL);
4679 for (i = 0; (i != elemof(mlx4_dev)); ++i) {
4680 if ((mlx4_dev[i].pci_addr.domain == pci_addr->domain) &&
4681 (mlx4_dev[i].pci_addr.bus == pci_addr->bus) &&
4682 (mlx4_dev[i].pci_addr.devid == pci_addr->devid) &&
4683 (mlx4_dev[i].pci_addr.function == pci_addr->function))
4685 if ((mlx4_dev[i].ports == 0) && (ret == -1))
4692 * Retrieve integer value from environment variable.
4695 * Environment variable name.
4698 * Integer value, 0 if the variable is not set.
4701 mlx4_getenv_int(const char *name)
4703 const char *val = getenv(name);
4711 mlx4_dev_link_status_handler(void *);
4713 mlx4_dev_interrupt_handler(struct rte_intr_handle *, void *);
4716 * Link status handler.
4719 * Pointer to private structure.
4721 * Pointer to the rte_eth_dev structure.
4724 * Nonzero if the callback process can be called immediately.
4727 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
4729 struct ibv_async_event event;
4730 int port_change = 0;
4733 /* Read all message and acknowledge them. */
4735 if (ibv_get_async_event(priv->ctx, &event))
4738 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
4739 event.event_type == IBV_EVENT_PORT_ERR)
4742 DEBUG("event type %d on port %d not handled",
4743 event.event_type, event.element.port_num);
4744 ibv_ack_async_event(&event);
4747 if (port_change ^ priv->pending_alarm) {
4748 struct rte_eth_link *link = &dev->data->dev_link;
4750 priv->pending_alarm = 0;
4751 mlx4_link_update_unlocked(dev, 0);
4752 if (((link->link_speed == 0) && link->link_status) ||
4753 ((link->link_speed != 0) && !link->link_status)) {
4754 /* Inconsistent status, check again later. */
4755 priv->pending_alarm = 1;
4756 rte_eal_alarm_set(MLX4_ALARM_TIMEOUT_US,
4757 mlx4_dev_link_status_handler,
4766 * Handle delayed link status event.
4769 * Registered argument.
4772 mlx4_dev_link_status_handler(void *arg)
4774 struct rte_eth_dev *dev = arg;
4775 struct priv *priv = dev->data->dev_private;
4779 assert(priv->pending_alarm == 1);
4780 ret = priv_dev_link_status_handler(priv, dev);
4783 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
4787 * Handle interrupts from the NIC.
4789 * @param[in] intr_handle
4790 * Interrupt handler.
4792 * Callback argument.
4795 mlx4_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
4797 struct rte_eth_dev *dev = cb_arg;
4798 struct priv *priv = dev->data->dev_private;
4803 ret = priv_dev_link_status_handler(priv, dev);
4806 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
4810 * Uninstall interrupt handler.
4813 * Pointer to private structure.
4815 * Pointer to the rte_eth_dev structure.
4818 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
4820 if (!dev->data->dev_conf.intr_conf.lsc)
4822 rte_intr_callback_unregister(&priv->intr_handle,
4823 mlx4_dev_interrupt_handler,
4825 if (priv->pending_alarm)
4826 rte_eal_alarm_cancel(mlx4_dev_link_status_handler, dev);
4827 priv->pending_alarm = 0;
4828 priv->intr_handle.fd = 0;
4829 priv->intr_handle.type = 0;
4833 * Install interrupt handler.
4836 * Pointer to private structure.
4838 * Pointer to the rte_eth_dev structure.
4841 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
4845 if (!dev->data->dev_conf.intr_conf.lsc)
4847 assert(priv->ctx->async_fd > 0);
4848 flags = fcntl(priv->ctx->async_fd, F_GETFL);
4849 rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
4851 INFO("failed to change file descriptor async event queue");
4852 dev->data->dev_conf.intr_conf.lsc = 0;
4854 priv->intr_handle.fd = priv->ctx->async_fd;
4855 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
4856 rte_intr_callback_register(&priv->intr_handle,
4857 mlx4_dev_interrupt_handler,
4862 static struct eth_driver mlx4_driver;
4865 * DPDK callback to register a PCI device.
4867 * This function creates an Ethernet device for each port of a given
4870 * @param[in] pci_drv
4871 * PCI driver structure (mlx4_driver).
4872 * @param[in] pci_dev
4873 * PCI device information.
4876 * 0 on success, negative errno value on failure.
4879 mlx4_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
4881 struct ibv_device **list;
4882 struct ibv_device *ibv_dev;
4884 struct ibv_context *attr_ctx = NULL;
4885 struct ibv_device_attr device_attr;
4891 assert(pci_drv == &mlx4_driver.pci_drv);
4892 /* Get mlx4_dev[] index. */
4893 idx = mlx4_dev_idx(&pci_dev->addr);
4895 ERROR("this driver cannot support any more adapters");
4898 DEBUG("using driver device index %d", idx);
4900 /* Save PCI address. */
4901 mlx4_dev[idx].pci_addr = pci_dev->addr;
4902 list = ibv_get_device_list(&i);
4905 if (errno == ENOSYS) {
4906 WARN("cannot list devices, is ib_uverbs loaded?");
4913 * For each listed device, check related sysfs entry against
4914 * the provided PCI ID.
4917 struct rte_pci_addr pci_addr;
4920 DEBUG("checking device \"%s\"", list[i]->name);
4921 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
4923 if ((pci_dev->addr.domain != pci_addr.domain) ||
4924 (pci_dev->addr.bus != pci_addr.bus) ||
4925 (pci_dev->addr.devid != pci_addr.devid) ||
4926 (pci_dev->addr.function != pci_addr.function))
4928 vf = (pci_dev->id.device_id ==
4929 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
4930 INFO("PCI information matches, using device \"%s\" (VF: %s)",
4931 list[i]->name, (vf ? "true" : "false"));
4932 attr_ctx = ibv_open_device(list[i]);
4936 if (attr_ctx == NULL) {
4937 ibv_free_device_list(list);
4940 WARN("cannot access device, is mlx4_ib loaded?");
4943 WARN("cannot use device, are drivers up to date?");
4951 DEBUG("device opened");
4952 if (ibv_query_device(attr_ctx, &device_attr))
4954 INFO("%u port(s) detected", device_attr.phys_port_cnt);
4956 for (i = 0; i < device_attr.phys_port_cnt; i++) {
4957 uint32_t port = i + 1; /* ports are indexed from one */
4958 uint32_t test = (1 << i);
4959 struct ibv_context *ctx = NULL;
4960 struct ibv_port_attr port_attr;
4961 struct ibv_pd *pd = NULL;
4962 struct priv *priv = NULL;
4963 struct rte_eth_dev *eth_dev;
4964 #ifdef HAVE_EXP_QUERY_DEVICE
4965 struct ibv_exp_device_attr exp_device_attr;
4966 #endif /* HAVE_EXP_QUERY_DEVICE */
4967 struct ether_addr mac;
4969 #ifdef HAVE_EXP_QUERY_DEVICE
4970 exp_device_attr.comp_mask = IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS;
4972 exp_device_attr.comp_mask |= IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ;
4973 #endif /* RSS_SUPPORT */
4974 #endif /* HAVE_EXP_QUERY_DEVICE */
4976 DEBUG("using port %u (%08" PRIx32 ")", port, test);
4978 ctx = ibv_open_device(ibv_dev);
4982 /* Check port status. */
4983 err = ibv_query_port(ctx, port, &port_attr);
4985 ERROR("port query failed: %s", strerror(err));
4988 if (port_attr.state != IBV_PORT_ACTIVE)
4989 DEBUG("port %d is not active: \"%s\" (%d)",
4990 port, ibv_port_state_str(port_attr.state),
4993 /* Allocate protection domain. */
4994 pd = ibv_alloc_pd(ctx);
4996 ERROR("PD allocation failure");
5001 mlx4_dev[idx].ports |= test;
5003 /* from rte_ethdev.c */
5004 priv = rte_zmalloc("ethdev private structure",
5006 RTE_CACHE_LINE_SIZE);
5008 ERROR("priv allocation failure");
5014 priv->device_attr = device_attr;
5017 priv->mtu = ETHER_MTU;
5018 #ifdef HAVE_EXP_QUERY_DEVICE
5019 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
5020 ERROR("ibv_exp_query_device() failed");
5024 if ((exp_device_attr.exp_device_cap_flags &
5025 IBV_EXP_DEVICE_QPG) &&
5026 (exp_device_attr.exp_device_cap_flags &
5027 IBV_EXP_DEVICE_UD_RSS) &&
5028 (exp_device_attr.comp_mask &
5029 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ) &&
5030 (exp_device_attr.max_rss_tbl_sz > 0)) {
5033 priv->max_rss_tbl_sz = exp_device_attr.max_rss_tbl_sz;
5037 priv->max_rss_tbl_sz = 0;
5039 priv->hw_tss = !!(exp_device_attr.exp_device_cap_flags &
5040 IBV_EXP_DEVICE_UD_TSS);
5041 DEBUG("device flags: %s%s%s",
5042 (priv->hw_qpg ? "IBV_DEVICE_QPG " : ""),
5043 (priv->hw_tss ? "IBV_DEVICE_TSS " : ""),
5044 (priv->hw_rss ? "IBV_DEVICE_RSS " : ""));
5046 DEBUG("maximum RSS indirection table size: %u",
5047 exp_device_attr.max_rss_tbl_sz);
5048 #endif /* RSS_SUPPORT */
5051 ((exp_device_attr.exp_device_cap_flags &
5052 IBV_EXP_DEVICE_RX_CSUM_TCP_UDP_PKT) &&
5053 (exp_device_attr.exp_device_cap_flags &
5054 IBV_EXP_DEVICE_RX_CSUM_IP_PKT));
5055 DEBUG("checksum offloading is %ssupported",
5056 (priv->hw_csum ? "" : "not "));
5058 priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
5059 IBV_EXP_DEVICE_VXLAN_SUPPORT);
5060 DEBUG("L2 tunnel checksum offloads are %ssupported",
5061 (priv->hw_csum_l2tun ? "" : "not "));
5064 priv->inl_recv_size = mlx4_getenv_int("MLX4_INLINE_RECV_SIZE");
5066 if (priv->inl_recv_size) {
5067 exp_device_attr.comp_mask =
5068 IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ;
5069 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
5070 INFO("Couldn't query device for inline-receive"
5072 priv->inl_recv_size = 0;
5074 if ((unsigned)exp_device_attr.inline_recv_sz <
5075 priv->inl_recv_size) {
5076 INFO("Max inline-receive (%d) <"
5077 " requested inline-receive (%u)",
5078 exp_device_attr.inline_recv_sz,
5079 priv->inl_recv_size);
5080 priv->inl_recv_size =
5081 exp_device_attr.inline_recv_sz;
5084 INFO("Set inline receive size to %u",
5085 priv->inl_recv_size);
5087 #endif /* INLINE_RECV */
5088 #endif /* HAVE_EXP_QUERY_DEVICE */
5090 (void)mlx4_getenv_int;
5092 /* Configure the first MAC address by default. */
5093 if (priv_get_mac(priv, &mac.addr_bytes)) {
5094 ERROR("cannot get MAC address, is mlx4_en loaded?"
5095 " (errno: %s)", strerror(errno));
5098 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
5100 mac.addr_bytes[0], mac.addr_bytes[1],
5101 mac.addr_bytes[2], mac.addr_bytes[3],
5102 mac.addr_bytes[4], mac.addr_bytes[5]);
5103 /* Register MAC and broadcast addresses. */
5104 claim_zero(priv_mac_addr_add(priv, 0,
5105 (const uint8_t (*)[ETHER_ADDR_LEN])
5107 claim_zero(priv_mac_addr_add(priv, (elemof(priv->mac) - 1),
5108 &(const uint8_t [ETHER_ADDR_LEN])
5109 { "\xff\xff\xff\xff\xff\xff" }));
5112 char ifname[IF_NAMESIZE];
5114 if (priv_get_ifname(priv, &ifname) == 0)
5115 DEBUG("port %u ifname is \"%s\"",
5116 priv->port, ifname);
5118 DEBUG("port %u ifname is unknown", priv->port);
5121 /* Get actual MTU if possible. */
5122 priv_get_mtu(priv, &priv->mtu);
5123 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
5125 /* from rte_ethdev.c */
5127 char name[RTE_ETH_NAME_MAX_LEN];
5129 snprintf(name, sizeof(name), "%s port %u",
5130 ibv_get_device_name(ibv_dev), port);
5131 eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
5133 if (eth_dev == NULL) {
5134 ERROR("can not allocate rte ethdev");
5139 eth_dev->data->dev_private = priv;
5140 eth_dev->pci_dev = pci_dev;
5142 rte_eth_copy_pci_info(eth_dev, pci_dev);
5144 eth_dev->driver = &mlx4_driver;
5145 eth_dev->data->rx_mbuf_alloc_failed = 0;
5146 eth_dev->data->mtu = ETHER_MTU;
5148 priv->dev = eth_dev;
5149 eth_dev->dev_ops = &mlx4_dev_ops;
5150 eth_dev->data->mac_addrs = priv->mac;
5151 TAILQ_INIT(ð_dev->link_intr_cbs);
5153 /* Bring Ethernet device up. */
5154 DEBUG("forcing Ethernet interface up");
5155 priv_set_flags(priv, ~IFF_UP, IFF_UP);
5161 claim_zero(ibv_dealloc_pd(pd));
5163 claim_zero(ibv_close_device(ctx));
5168 * XXX if something went wrong in the loop above, there is a resource
5169 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
5170 * long as the dpdk does not provide a way to deallocate a ethdev and a
5171 * way to enumerate the registered ethdevs to free the previous ones.
5174 /* no port found, complain */
5175 if (!mlx4_dev[idx].ports) {
5182 claim_zero(ibv_close_device(attr_ctx));
5184 ibv_free_device_list(list);
5189 static const struct rte_pci_id mlx4_pci_id_map[] = {
5191 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5192 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3,
5193 .subsystem_vendor_id = PCI_ANY_ID,
5194 .subsystem_device_id = PCI_ANY_ID
5197 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5198 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO,
5199 .subsystem_vendor_id = PCI_ANY_ID,
5200 .subsystem_device_id = PCI_ANY_ID
5203 .vendor_id = PCI_VENDOR_ID_MELLANOX,
5204 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3VF,
5205 .subsystem_vendor_id = PCI_ANY_ID,
5206 .subsystem_device_id = PCI_ANY_ID
5213 static struct eth_driver mlx4_driver = {
5215 .name = MLX4_DRIVER_NAME,
5216 .id_table = mlx4_pci_id_map,
5217 .devinit = mlx4_pci_devinit,
5218 .drv_flags = RTE_PCI_DRV_INTR_LSC,
5220 .dev_private_size = sizeof(struct priv)
5224 * Driver initialization routine.
5227 rte_mlx4_pmd_init(const char *name, const char *args)
5232 * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
5233 * huge pages. Calling ibv_fork_init() during init allows
5234 * applications to use fork() safely for purposes other than
5235 * using this PMD, which is not supported in forked processes.
5237 setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
5239 rte_eal_pci_register(&mlx4_driver.pci_drv);
5243 static struct rte_driver rte_mlx4_driver = {
5245 .name = MLX4_DRIVER_NAME,
5246 .init = rte_mlx4_pmd_init,
5249 PMD_REGISTER_DRIVER(rte_mlx4_driver)