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 * - Multiple RX VLAN filters can be configured, but only the first one
38 * - RSS hash key and options cannot be modified.
39 * - Hardware counters aren't implemented.
53 #include <arpa/inet.h>
56 #include <sys/ioctl.h>
57 #include <sys/socket.h>
58 #include <netinet/in.h>
60 #include <linux/ethtool.h>
61 #include <linux/sockios.h>
64 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
66 #pragma GCC diagnostic ignored "-pedantic"
68 #include <infiniband/verbs.h>
70 #pragma GCC diagnostic error "-pedantic"
73 /* DPDK headers don't like -pedantic. */
75 #pragma GCC diagnostic ignored "-pedantic"
77 #include <rte_config.h>
78 #include <rte_ether.h>
79 #include <rte_ethdev.h>
82 #include <rte_errno.h>
83 #include <rte_mempool.h>
84 #include <rte_prefetch.h>
85 #include <rte_malloc.h>
86 #include <rte_spinlock.h>
87 #include <rte_atomic.h>
88 #include <rte_version.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 /* If raw send operations are available, use them since they are faster. */
145 #ifdef SEND_RAW_WR_SUPPORT
146 typedef struct ibv_send_wr_raw mlx4_send_wr_t;
147 #define mlx4_post_send ibv_post_send_raw
149 typedef struct ibv_send_wr mlx4_send_wr_t;
150 #define mlx4_post_send ibv_post_send
153 struct mlx4_rxq_stats {
154 unsigned int idx; /**< Mapping index. */
155 #ifdef MLX4_PMD_SOFT_COUNTERS
156 uint64_t ipackets; /**< Total of successfully received packets. */
157 uint64_t ibytes; /**< Total of successfully received bytes. */
159 uint64_t idropped; /**< Total of packets dropped when RX ring full. */
160 uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
163 struct mlx4_txq_stats {
164 unsigned int idx; /**< Mapping index. */
165 #ifdef MLX4_PMD_SOFT_COUNTERS
166 uint64_t opackets; /**< Total of successfully sent packets. */
167 uint64_t obytes; /**< Total of successfully sent bytes. */
169 uint64_t odropped; /**< Total of packets not sent when TX ring full. */
172 /* RX element (scattered packets). */
174 struct ibv_recv_wr wr; /* Work Request. */
175 struct ibv_sge sges[MLX4_PMD_SGE_WR_N]; /* Scatter/Gather Elements. */
176 struct rte_mbuf *bufs[MLX4_PMD_SGE_WR_N]; /* SGEs buffers. */
181 struct ibv_recv_wr wr; /* Work Request. */
182 struct ibv_sge sge; /* Scatter/Gather Element. */
183 /* mbuf pointer is derived from WR_ID(wr.wr_id).offset. */
186 /* RX queue descriptor. */
188 struct priv *priv; /* Back pointer to private data. */
189 struct rte_mempool *mp; /* Memory Pool for allocations. */
190 struct ibv_mr *mr; /* Memory Region (for mp). */
191 struct ibv_cq *cq; /* Completion Queue. */
192 struct ibv_qp *qp; /* Queue Pair. */
194 * There is exactly one flow configured per MAC address. Each flow
195 * may contain several specifications, one per configured VLAN ID.
197 BITFIELD_DECLARE(mac_configured, uint32_t, MLX4_MAX_MAC_ADDRESSES);
198 struct ibv_exp_flow *mac_flow[MLX4_MAX_MAC_ADDRESSES];
199 struct ibv_exp_flow *promisc_flow; /* Promiscuous flow. */
200 struct ibv_exp_flow *allmulti_flow; /* Multicast flow. */
201 unsigned int port_id; /* Port ID for incoming packets. */
202 unsigned int elts_n; /* (*elts)[] length. */
204 struct rxq_elt_sp (*sp)[]; /* Scattered RX elements. */
205 struct rxq_elt (*no_sp)[]; /* RX elements. */
207 unsigned int sp:1; /* Use scattered RX elements. */
208 uint32_t mb_len; /* Length of a mp-issued mbuf. */
209 struct mlx4_rxq_stats stats; /* RX queue counters. */
210 unsigned int socket; /* CPU socket ID for allocations. */
215 mlx4_send_wr_t wr; /* Work Request. */
216 struct ibv_sge sges[MLX4_PMD_SGE_WR_N]; /* Scatter/Gather Elements. */
217 /* mbuf pointer is derived from WR_ID(wr.wr_id).offset. */
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 #if MLX4_PMD_MAX_INLINE > 0
239 uint32_t max_inline; /* Max inline send size <= MLX4_PMD_MAX_INLINE. */
241 unsigned int elts_n; /* (*elts)[] length. */
242 struct txq_elt (*elts)[]; /* TX elements. */
243 unsigned int elts_head; /* Current index in (*elts)[]. */
244 unsigned int elts_tail; /* First element awaiting completion. */
245 unsigned int elts_comp; /* Number of completion requests. */
246 struct mlx4_txq_stats stats; /* TX queue counters. */
247 linear_t (*elts_linear)[]; /* Linearized buffers. */
248 struct ibv_mr *mr_linear; /* Memory Region for linearized buffers. */
249 unsigned int socket; /* CPU socket ID for allocations. */
253 struct rte_eth_dev *dev; /* Ethernet device. */
254 struct ibv_context *ctx; /* Verbs context. */
255 struct ibv_device_attr device_attr; /* Device properties. */
256 struct ibv_port_attr port_attr; /* Physical port properties. */
257 struct ibv_pd *pd; /* Protection Domain. */
259 * MAC addresses array and configuration bit-field.
260 * An extra entry that cannot be modified by the DPDK is reserved
261 * for broadcast frames (destination MAC address ff:ff:ff:ff:ff:ff).
263 struct ether_addr mac[MLX4_MAX_MAC_ADDRESSES];
264 BITFIELD_DECLARE(mac_configured, uint32_t, MLX4_MAX_MAC_ADDRESSES);
267 unsigned int enabled:1; /* If enabled. */
268 unsigned int id:12; /* VLAN ID (0-4095). */
269 } vlan_filter[MLX4_MAX_VLAN_IDS]; /* VLAN filters table. */
270 /* Device properties. */
271 uint16_t mtu; /* Configured MTU. */
272 uint8_t port; /* Physical port number. */
273 unsigned int started:1; /* Device started, flows enabled. */
274 unsigned int promisc:1; /* Device in promiscuous mode. */
275 unsigned int promisc_ok:1; /* Promiscuous flow is supported. */
276 unsigned int allmulti:1; /* Device receives all multicast packets. */
277 unsigned int hw_qpg:1; /* QP groups are supported. */
278 unsigned int hw_tss:1; /* TSS is supported. */
279 unsigned int hw_rss:1; /* RSS is supported. */
280 unsigned int rss:1; /* RSS is enabled. */
281 unsigned int vf:1; /* This is a VF device. */
283 unsigned int inl_recv_size; /* Inline recv size */
285 unsigned int max_rss_tbl_sz; /* Maximum number of RSS queues. */
287 struct rxq rxq_parent; /* Parent queue when RSS is enabled. */
288 unsigned int rxqs_n; /* RX queues array size. */
289 unsigned int txqs_n; /* TX queues array size. */
290 struct rxq *(*rxqs)[]; /* RX queues. */
291 struct txq *(*txqs)[]; /* TX queues. */
292 rte_spinlock_t lock; /* Lock for control functions. */
296 * Lock private structure to protect it from concurrent access in the
300 * Pointer to private structure.
303 priv_lock(struct priv *priv)
305 rte_spinlock_lock(&priv->lock);
309 * Unlock private structure.
312 * Pointer to private structure.
315 priv_unlock(struct priv *priv)
317 rte_spinlock_unlock(&priv->lock);
320 /* Allocate a buffer on the stack and fill it with a printf format string. */
321 #define MKSTR(name, ...) \
322 char name[snprintf(NULL, 0, __VA_ARGS__) + 1]; \
324 snprintf(name, sizeof(name), __VA_ARGS__)
327 * Get interface name from private structure.
330 * Pointer to private structure.
332 * Interface name output buffer.
335 * 0 on success, -1 on failure and errno is set.
338 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
345 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
351 while ((dent = readdir(dir)) != NULL) {
352 char *name = dent->d_name;
357 if ((name[0] == '.') &&
358 ((name[1] == '\0') ||
359 ((name[1] == '.') && (name[2] == '\0'))))
362 MKSTR(path, "%s/device/net/%s/dev_id",
363 priv->ctx->device->ibdev_path, name);
365 file = fopen(path, "rb");
368 r = fscanf(file, "%x", &dev_id);
370 if ((r == 1) && (dev_id == (priv->port - 1u))) {
371 snprintf(*ifname, sizeof(*ifname), "%s", name);
381 * Read from sysfs entry.
384 * Pointer to private structure.
386 * Entry name relative to sysfs path.
388 * Data output buffer.
393 * 0 on success, -1 on failure and errno is set.
396 priv_sysfs_read(const struct priv *priv, const char *entry,
397 char *buf, size_t size)
399 char ifname[IF_NAMESIZE];
404 if (priv_get_ifname(priv, &ifname))
407 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
410 file = fopen(path, "rb");
413 ret = fread(buf, 1, size, file);
415 if (((size_t)ret < size) && (ferror(file)))
425 * Write to sysfs entry.
428 * Pointer to private structure.
430 * Entry name relative to sysfs path.
437 * 0 on success, -1 on failure and errno is set.
440 priv_sysfs_write(const struct priv *priv, const char *entry,
441 char *buf, size_t size)
443 char ifname[IF_NAMESIZE];
448 if (priv_get_ifname(priv, &ifname))
451 MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
454 file = fopen(path, "wb");
457 ret = fwrite(buf, 1, size, file);
459 if (((size_t)ret < size) || (ferror(file)))
469 * Get unsigned long sysfs property.
472 * Pointer to private structure.
474 * Entry name relative to sysfs path.
476 * Value output buffer.
479 * 0 on success, -1 on failure and errno is set.
482 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
485 unsigned long value_ret;
488 ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
490 DEBUG("cannot read %s value from sysfs: %s",
491 name, strerror(errno));
494 value_str[ret] = '\0';
496 value_ret = strtoul(value_str, NULL, 0);
498 DEBUG("invalid %s value `%s': %s", name, value_str,
507 * Set unsigned long sysfs property.
510 * Pointer to private structure.
512 * Entry name relative to sysfs path.
517 * 0 on success, -1 on failure and errno is set.
520 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
523 MKSTR(value_str, "%lu", value);
525 ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
527 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
528 name, value_str, value, strerror(errno));
535 * Perform ifreq ioctl() on associated Ethernet device.
538 * Pointer to private structure.
540 * Request number to pass to ioctl().
542 * Interface request structure output buffer.
545 * 0 on success, -1 on failure and errno is set.
548 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
550 int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
555 if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
556 ret = ioctl(sock, req, ifr);
565 * Pointer to private structure.
567 * MTU value output buffer.
570 * 0 on success, -1 on failure and errno is set.
573 priv_get_mtu(struct priv *priv, uint16_t *mtu)
575 unsigned long ulong_mtu;
577 if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
587 * Pointer to private structure.
592 * 0 on success, -1 on failure and errno is set.
595 priv_set_mtu(struct priv *priv, uint16_t mtu)
597 return priv_set_sysfs_ulong(priv, "mtu", mtu);
604 * Pointer to private structure.
606 * Bitmask for flags that must remain untouched.
608 * Bitmask for flags to modify.
611 * 0 on success, -1 on failure and errno is set.
614 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
618 if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
622 return priv_set_sysfs_ulong(priv, "flags", tmp);
625 /* Device configuration. */
628 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
629 unsigned int socket, const struct rte_eth_rxconf *conf,
630 struct rte_mempool *mp);
633 rxq_cleanup(struct rxq *rxq);
636 * Ethernet device configuration.
638 * Prepare the driver for a given number of TX and RX queues.
639 * Allocate parent RSS queue when several RX queues are requested.
642 * Pointer to Ethernet device structure.
645 * 0 on success, errno value on failure.
648 dev_configure(struct rte_eth_dev *dev)
650 struct priv *priv = dev->data->dev_private;
651 unsigned int rxqs_n = dev->data->nb_rx_queues;
652 unsigned int txqs_n = dev->data->nb_tx_queues;
656 priv->rxqs = (void *)dev->data->rx_queues;
657 priv->txqs = (void *)dev->data->tx_queues;
658 if (txqs_n != priv->txqs_n) {
659 INFO("%p: TX queues number update: %u -> %u",
660 (void *)dev, priv->txqs_n, txqs_n);
661 priv->txqs_n = txqs_n;
663 if (rxqs_n == priv->rxqs_n)
665 INFO("%p: RX queues number update: %u -> %u",
666 (void *)dev, priv->rxqs_n, rxqs_n);
667 /* If RSS is enabled, disable it first. */
671 /* Only if there are no remaining child RX queues. */
672 for (i = 0; (i != priv->rxqs_n); ++i)
673 if ((*priv->rxqs)[i] != NULL)
675 rxq_cleanup(&priv->rxq_parent);
680 /* Nothing else to do. */
681 priv->rxqs_n = rxqs_n;
684 /* Allocate a new RSS parent queue if supported by hardware. */
686 ERROR("%p: only a single RX queue can be configured when"
687 " hardware doesn't support RSS",
691 /* Fail if hardware doesn't support that many RSS queues. */
692 if (rxqs_n >= priv->max_rss_tbl_sz) {
693 ERROR("%p: only %u RX queues can be configured for RSS",
694 (void *)dev, priv->max_rss_tbl_sz);
699 priv->rxqs_n = rxqs_n;
700 ret = rxq_setup(dev, &priv->rxq_parent, 0, 0, NULL, NULL);
703 /* Failure, rollback. */
711 * DPDK callback for Ethernet device configuration.
714 * Pointer to Ethernet device structure.
717 * 0 on success, negative errno value on failure.
720 mlx4_dev_configure(struct rte_eth_dev *dev)
722 struct priv *priv = dev->data->dev_private;
726 ret = dev_configure(dev);
732 /* TX queues handling. */
735 * Allocate TX queue elements.
738 * Pointer to TX queue structure.
740 * Number of elements to allocate.
743 * 0 on success, errno value on failure.
746 txq_alloc_elts(struct txq *txq, unsigned int elts_n)
749 struct txq_elt (*elts)[elts_n] =
750 rte_calloc_socket("TXQ", 1, sizeof(*elts), 0, txq->socket);
751 linear_t (*elts_linear)[elts_n] =
752 rte_calloc_socket("TXQ", 1, sizeof(*elts_linear), 0,
754 struct ibv_mr *mr_linear = NULL;
757 if ((elts == NULL) || (elts_linear == NULL)) {
758 ERROR("%p: can't allocate packets array", (void *)txq);
763 ibv_reg_mr(txq->priv->pd, elts_linear, sizeof(*elts_linear),
764 (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
765 if (mr_linear == NULL) {
766 ERROR("%p: unable to configure MR, ibv_reg_mr() failed",
771 for (i = 0; (i != elts_n); ++i) {
772 struct txq_elt *elt = &(*elts)[i];
773 mlx4_send_wr_t *wr = &elt->wr;
776 WR_ID(wr->wr_id).id = i;
777 WR_ID(wr->wr_id).offset = 0;
778 wr->sg_list = &elt->sges[0];
779 wr->opcode = IBV_WR_SEND;
780 /* Other fields are updated during TX. */
782 DEBUG("%p: allocated and configured %u WRs", (void *)txq, elts_n);
783 txq->elts_n = elts_n;
788 txq->elts_linear = elts_linear;
789 txq->mr_linear = mr_linear;
793 if (mr_linear != NULL)
794 claim_zero(ibv_dereg_mr(mr_linear));
796 rte_free(elts_linear);
799 DEBUG("%p: failed, freed everything", (void *)txq);
805 * Free TX queue elements.
808 * Pointer to TX queue structure.
811 txq_free_elts(struct txq *txq)
814 unsigned int elts_n = txq->elts_n;
815 struct txq_elt (*elts)[elts_n] = txq->elts;
816 linear_t (*elts_linear)[elts_n] = txq->elts_linear;
817 struct ibv_mr *mr_linear = txq->mr_linear;
819 DEBUG("%p: freeing WRs", (void *)txq);
822 txq->elts_linear = NULL;
823 txq->mr_linear = NULL;
824 if (mr_linear != NULL)
825 claim_zero(ibv_dereg_mr(mr_linear));
827 rte_free(elts_linear);
830 for (i = 0; (i != elemof(*elts)); ++i) {
831 struct txq_elt *elt = &(*elts)[i];
833 if (WR_ID(elt->wr.wr_id).offset == 0)
835 rte_pktmbuf_free((void *)((uintptr_t)elt->sges[0].addr -
836 WR_ID(elt->wr.wr_id).offset));
843 * Clean up a TX queue.
845 * Destroy objects, free allocated memory and reset the structure for reuse.
848 * Pointer to TX queue structure.
851 txq_cleanup(struct txq *txq)
855 DEBUG("cleaning up %p", (void *)txq);
858 claim_zero(ibv_destroy_qp(txq->qp));
860 claim_zero(ibv_destroy_cq(txq->cq));
861 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
862 if (txq->mp2mr[i].mp == NULL)
864 assert(txq->mp2mr[i].mr != NULL);
865 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
867 memset(txq, 0, sizeof(*txq));
871 * Manage TX completions.
873 * When sending a burst, mlx4_tx_burst() posts several WRs.
874 * To improve performance, a completion event is only required for the last of
875 * them. Doing so discards completion information for other WRs, but this
876 * information would not be used anyway.
879 * Pointer to TX queue structure.
882 * 0 on success, -1 on failure.
885 txq_complete(struct txq *txq)
887 unsigned int elts_comp = txq->elts_comp;
888 unsigned int elts_tail;
889 const unsigned int elts_n = txq->elts_n;
890 struct ibv_wc wcs[elts_comp];
893 if (unlikely(elts_comp == 0))
896 DEBUG("%p: processing %u work requests completions",
897 (void *)txq, elts_comp);
899 wcs_n = ibv_poll_cq(txq->cq, elts_comp, wcs);
900 if (unlikely(wcs_n == 0))
902 if (unlikely(wcs_n < 0)) {
903 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
908 assert(elts_comp <= txq->elts_comp);
910 * Work Completion ID contains the associated element index in
911 * (*txq->elts)[]. Since WCs are returned in order, we only need to
912 * look at the last WC to clear older Work Requests.
914 * Assume WC status is successful as nothing can be done about it
917 elts_tail = WR_ID(wcs[wcs_n - 1].wr_id).id;
918 /* Consume the last WC. */
919 if (++elts_tail >= elts_n)
921 txq->elts_tail = elts_tail;
922 txq->elts_comp = elts_comp;
927 * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
928 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
929 * remove an entry first.
932 * Pointer to TX queue structure.
934 * Memory Pool for which a Memory Region lkey must be returned.
937 * mr->lkey on success, (uint32_t)-1 on failure.
940 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
945 for (i = 0; (i != elemof(txq->mp2mr)); ++i) {
946 if (unlikely(txq->mp2mr[i].mp == NULL)) {
947 /* Unknown MP, add a new MR for it. */
950 if (txq->mp2mr[i].mp == mp) {
951 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
952 assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
953 return txq->mp2mr[i].lkey;
956 /* Add a new entry, register MR first. */
957 DEBUG("%p: discovered new memory pool %p", (void *)txq, (void *)mp);
958 mr = ibv_reg_mr(txq->priv->pd,
959 (void *)mp->elt_va_start,
960 (mp->elt_va_end - mp->elt_va_start),
961 (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
962 if (unlikely(mr == NULL)) {
963 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
967 if (unlikely(i == elemof(txq->mp2mr))) {
968 /* Table is full, remove oldest entry. */
969 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
972 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
973 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
974 (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
976 /* Store the new entry. */
977 txq->mp2mr[i].mp = mp;
978 txq->mp2mr[i].mr = mr;
979 txq->mp2mr[i].lkey = mr->lkey;
980 DEBUG("%p: new MR lkey for MP %p: 0x%08" PRIu32,
981 (void *)txq, (void *)mp, txq->mp2mr[i].lkey);
982 return txq->mp2mr[i].lkey;
986 * Copy scattered mbuf contents to a single linear buffer.
989 * Linear output buffer.
991 * Scattered input buffer.
994 * Number of bytes copied to the output buffer or 0 if not large enough.
997 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
999 unsigned int size = 0;
1000 unsigned int offset;
1003 unsigned int len = DATA_LEN(buf);
1007 if (unlikely(size > sizeof(*linear)))
1009 memcpy(&(*linear)[offset],
1010 rte_pktmbuf_mtod(buf, uint8_t *),
1013 } while (buf != NULL);
1018 * DPDK callback for TX.
1021 * Generic pointer to TX queue structure.
1023 * Packets to transmit.
1025 * Number of packets in array.
1028 * Number of packets successfully transmitted (<= pkts_n).
1031 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1033 struct txq *txq = (struct txq *)dpdk_txq;
1034 mlx4_send_wr_t head;
1035 mlx4_send_wr_t **wr_next = &head.next;
1036 mlx4_send_wr_t *bad_wr;
1037 unsigned int elts_head = txq->elts_head;
1038 const unsigned int elts_tail = txq->elts_tail;
1039 const unsigned int elts_n = txq->elts_n;
1045 max = (elts_n - (elts_head - elts_tail));
1049 assert(max <= elts_n);
1050 /* Always leave one free entry in the ring. */
1056 for (i = 0; (i != max); ++i) {
1057 struct rte_mbuf *buf = pkts[i];
1058 struct txq_elt *elt = &(*txq->elts)[elts_head];
1059 mlx4_send_wr_t *wr = &elt->wr;
1060 unsigned int segs = NB_SEGS(buf);
1061 #if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
1062 unsigned int sent_size = 0;
1067 /* Clean up old buffer. */
1068 if (likely(WR_ID(wr->wr_id).offset != 0)) {
1069 struct rte_mbuf *tmp = (void *)
1070 ((uintptr_t)elt->sges[0].addr -
1071 WR_ID(wr->wr_id).offset);
1073 /* Faster than rte_pktmbuf_free(). */
1075 struct rte_mbuf *next = NEXT(tmp);
1077 rte_pktmbuf_free_seg(tmp);
1079 } while (tmp != NULL);
1083 WR_ID(wr->wr_id).offset = 0;
1084 for (j = 0; ((int)j < wr->num_sge); ++j) {
1085 elt->sges[j].addr = 0;
1086 elt->sges[j].length = 0;
1087 elt->sges[j].lkey = 0;
1092 /* Sanity checks, most of which are only relevant with
1093 * debugging enabled. */
1094 assert(WR_ID(wr->wr_id).id == elts_head);
1095 assert(WR_ID(wr->wr_id).offset == 0);
1096 assert(wr->next == NULL);
1097 assert(wr->sg_list == &elt->sges[0]);
1098 assert(wr->num_sge == 0);
1099 assert(wr->opcode == IBV_WR_SEND);
1100 /* When there are too many segments, extra segments are
1101 * linearized in the last SGE. */
1102 if (unlikely(segs > elemof(elt->sges))) {
1103 segs = (elemof(elt->sges) - 1);
1106 /* Set WR fields. */
1107 assert(((uintptr_t)rte_pktmbuf_mtod(buf, char *) -
1108 (uintptr_t)buf) <= 0xffff);
1109 WR_ID(wr->wr_id).offset =
1110 ((uintptr_t)rte_pktmbuf_mtod(buf, char *) -
1113 /* Register segments as SGEs. */
1114 for (j = 0; (j != segs); ++j) {
1115 struct ibv_sge *sge = &elt->sges[j];
1118 /* Retrieve Memory Region key for this memory pool. */
1119 lkey = txq_mp2mr(txq, buf->pool);
1120 if (unlikely(lkey == (uint32_t)-1)) {
1121 /* MR does not exist. */
1122 DEBUG("%p: unable to get MP <-> MR"
1123 " association", (void *)txq);
1124 /* Clean up TX element. */
1125 WR_ID(elt->wr.wr_id).offset = 0;
1139 /* Sanity checks, only relevant with debugging
1141 assert(sge->addr == 0);
1142 assert(sge->length == 0);
1143 assert(sge->lkey == 0);
1145 sge->addr = (uintptr_t)rte_pktmbuf_mtod(buf, char *);
1147 rte_prefetch0((volatile void *)
1148 (uintptr_t)sge->addr);
1149 sge->length = DATA_LEN(buf);
1151 #if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
1152 sent_size += sge->length;
1156 /* If buf is not NULL here and is not going to be linearized,
1157 * nb_segs is not valid. */
1159 assert((buf == NULL) || (linearize));
1160 /* Linearize extra segments. */
1162 struct ibv_sge *sge = &elt->sges[segs];
1163 linear_t *linear = &(*txq->elts_linear)[elts_head];
1164 unsigned int size = linearize_mbuf(linear, buf);
1166 assert(segs == (elemof(elt->sges) - 1));
1168 /* Invalid packet. */
1169 DEBUG("%p: packet too large to be linearized.",
1171 /* Clean up TX element. */
1172 WR_ID(elt->wr.wr_id).offset = 0;
1186 /* If MLX4_PMD_SGE_WR_N is 1, free mbuf immediately
1187 * and clear offset from WR ID. */
1188 if (elemof(elt->sges) == 1) {
1190 struct rte_mbuf *next = NEXT(buf);
1192 rte_pktmbuf_free_seg(buf);
1194 } while (buf != NULL);
1195 WR_ID(wr->wr_id).offset = 0;
1197 /* Set WR fields and fill SGE with linear buffer. */
1199 /* Sanity checks, only relevant with debugging
1201 assert(sge->addr == 0);
1202 assert(sge->length == 0);
1203 assert(sge->lkey == 0);
1205 sge->addr = (uintptr_t)&(*linear)[0];
1207 sge->lkey = txq->mr_linear->lkey;
1208 #if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
1212 /* Link WRs together for ibv_post_send(). */
1214 wr_next = &wr->next;
1215 #if MLX4_PMD_MAX_INLINE > 0
1216 if (sent_size <= txq->max_inline)
1217 wr->send_flags = IBV_SEND_INLINE;
1221 if (++elts_head >= elts_n)
1223 #ifdef MLX4_PMD_SOFT_COUNTERS
1224 /* Increment sent bytes counter. */
1225 txq->stats.obytes += sent_size;
1229 /* Take a shortcut if nothing must be sent. */
1230 if (unlikely(i == 0))
1232 #ifdef MLX4_PMD_SOFT_COUNTERS
1233 /* Increment sent packets counter. */
1234 txq->stats.opackets += i;
1237 /* The last WR is the only one asking for a completion event. */
1238 containerof(wr_next, mlx4_send_wr_t, next)->
1239 send_flags |= IBV_SEND_SIGNALED;
1240 err = mlx4_post_send(txq->qp, head.next, &bad_wr);
1241 if (unlikely(err)) {
1242 unsigned int unsent = 0;
1244 /* An error occurred, completion event is lost. Fix counters. */
1245 while (bad_wr != NULL) {
1246 struct txq_elt *elt =
1247 containerof(bad_wr, struct txq_elt, wr);
1248 mlx4_send_wr_t *wr = &elt->wr;
1249 mlx4_send_wr_t *next = wr->next;
1250 #if defined(MLX4_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1254 assert(wr == bad_wr);
1255 /* Clean up TX element without freeing it, caller
1256 * should take care of this. */
1257 WR_ID(elt->wr.wr_id).offset = 0;
1258 #ifdef MLX4_PMD_SOFT_COUNTERS
1259 for (j = 0; ((int)j < wr->num_sge); ++j)
1260 txq->stats.obytes -= wr->sg_list[j].length;
1265 for (j = 0; ((int)j < wr->num_sge); ++j) {
1266 elt->sges[j].addr = 0;
1267 elt->sges[j].length = 0;
1268 elt->sges[j].lkey = 0;
1275 #ifdef MLX4_PMD_SOFT_COUNTERS
1276 txq->stats.opackets -= unsent;
1278 assert(i >= unsent);
1280 /* "Unsend" remaining packets. */
1281 elts_head -= unsent;
1282 if (elts_head >= elts_n)
1283 elts_head += elts_n;
1284 assert(elts_head < elts_n);
1285 DEBUG("%p: mlx4_post_send() failed, %u unprocessed WRs: %s",
1286 (void *)txq, unsent,
1287 ((err <= -1) ? "Internal error" : strerror(err)));
1290 txq->elts_head = elts_head;
1295 * Configure a TX queue.
1298 * Pointer to Ethernet device structure.
1300 * Pointer to TX queue structure.
1302 * Number of descriptors to configure in queue.
1304 * NUMA socket on which memory must be allocated.
1306 * Thresholds parameters.
1309 * 0 on success, errno value on failure.
1312 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1313 unsigned int socket, const struct rte_eth_txconf *conf)
1315 struct priv *priv = dev->data->dev_private;
1321 struct ibv_qp_init_attr init;
1322 struct ibv_exp_qp_attr mod;
1326 (void)conf; /* Thresholds configuration (ignored). */
1327 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
1328 ERROR("%p: invalid number of TX descriptors (must be a"
1329 " multiple of %d)", (void *)dev, desc);
1332 desc /= MLX4_PMD_SGE_WR_N;
1333 /* MRs will be registered in mp2mr[] later. */
1334 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
1335 if (tmpl.cq == NULL) {
1337 ERROR("%p: CQ creation failure: %s",
1338 (void *)dev, strerror(ret));
1341 DEBUG("priv->device_attr.max_qp_wr is %d",
1342 priv->device_attr.max_qp_wr);
1343 DEBUG("priv->device_attr.max_sge is %d",
1344 priv->device_attr.max_sge);
1345 attr.init = (struct ibv_qp_init_attr){
1346 /* CQ to be associated with the send queue. */
1348 /* CQ to be associated with the receive queue. */
1351 /* Max number of outstanding WRs. */
1352 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1353 priv->device_attr.max_qp_wr :
1355 /* Max number of scatter/gather elements in a WR. */
1356 .max_send_sge = ((priv->device_attr.max_sge <
1357 MLX4_PMD_SGE_WR_N) ?
1358 priv->device_attr.max_sge :
1360 #if MLX4_PMD_MAX_INLINE > 0
1361 .max_inline_data = MLX4_PMD_MAX_INLINE,
1364 .qp_type = IBV_QPT_RAW_PACKET,
1365 /* Do *NOT* enable this, completions events are managed per
1369 tmpl.qp = ibv_create_qp(priv->pd, &attr.init);
1370 if (tmpl.qp == NULL) {
1371 ret = (errno ? errno : EINVAL);
1372 ERROR("%p: QP creation failure: %s",
1373 (void *)dev, strerror(ret));
1376 #if MLX4_PMD_MAX_INLINE > 0
1377 /* ibv_create_qp() updates this value. */
1378 tmpl.max_inline = attr.init.cap.max_inline_data;
1380 attr.mod = (struct ibv_exp_qp_attr){
1381 /* Move the QP to this state. */
1382 .qp_state = IBV_QPS_INIT,
1383 /* Primary port number. */
1384 .port_num = priv->port
1386 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
1387 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
1389 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1390 (void *)dev, strerror(ret));
1393 ret = txq_alloc_elts(&tmpl, desc);
1395 ERROR("%p: TXQ allocation failed: %s",
1396 (void *)dev, strerror(ret));
1399 attr.mod = (struct ibv_exp_qp_attr){
1400 .qp_state = IBV_QPS_RTR
1402 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1404 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1405 (void *)dev, strerror(ret));
1408 attr.mod.qp_state = IBV_QPS_RTS;
1409 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1411 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1412 (void *)dev, strerror(ret));
1415 /* Clean up txq in case we're reinitializing it. */
1416 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1419 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1429 * DPDK callback to configure a TX queue.
1432 * Pointer to Ethernet device structure.
1436 * Number of descriptors to configure in queue.
1438 * NUMA socket on which memory must be allocated.
1440 * Thresholds parameters.
1443 * 0 on success, negative errno value on failure.
1446 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1447 unsigned int socket, const struct rte_eth_txconf *conf)
1449 struct priv *priv = dev->data->dev_private;
1450 struct txq *txq = (*priv->txqs)[idx];
1454 DEBUG("%p: configuring queue %u for %u descriptors",
1455 (void *)dev, idx, desc);
1456 if (idx >= priv->txqs_n) {
1457 ERROR("%p: queue index out of range (%u >= %u)",
1458 (void *)dev, idx, priv->txqs_n);
1463 DEBUG("%p: reusing already allocated queue index %u (%p)",
1464 (void *)dev, idx, (void *)txq);
1465 if (priv->started) {
1469 (*priv->txqs)[idx] = NULL;
1472 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1474 ERROR("%p: unable to allocate queue index %u",
1480 ret = txq_setup(dev, txq, desc, socket, conf);
1484 txq->stats.idx = idx;
1485 DEBUG("%p: adding TX queue %p to list",
1486 (void *)dev, (void *)txq);
1487 (*priv->txqs)[idx] = txq;
1488 /* Update send callback. */
1489 dev->tx_pkt_burst = mlx4_tx_burst;
1496 * DPDK callback to release a TX queue.
1499 * Generic TX queue pointer.
1502 mlx4_tx_queue_release(void *dpdk_txq)
1504 struct txq *txq = (struct txq *)dpdk_txq;
1512 for (i = 0; (i != priv->txqs_n); ++i)
1513 if ((*priv->txqs)[i] == txq) {
1514 DEBUG("%p: removing TX queue %p from list",
1515 (void *)priv->dev, (void *)txq);
1516 (*priv->txqs)[i] = NULL;
1524 /* RX queues handling. */
1527 * Allocate RX queue elements with scattered packets support.
1530 * Pointer to RX queue structure.
1532 * Number of elements to allocate.
1534 * If not NULL, fetch buffers from this array instead of allocating them
1535 * with rte_pktmbuf_alloc().
1538 * 0 on success, errno value on failure.
1541 rxq_alloc_elts_sp(struct rxq *rxq, unsigned int elts_n,
1542 struct rte_mbuf **pool)
1545 struct rxq_elt_sp (*elts)[elts_n] =
1546 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1551 ERROR("%p: can't allocate packets array", (void *)rxq);
1555 /* For each WR (packet). */
1556 for (i = 0; (i != elts_n); ++i) {
1558 struct rxq_elt_sp *elt = &(*elts)[i];
1559 struct ibv_recv_wr *wr = &elt->wr;
1560 struct ibv_sge (*sges)[(elemof(elt->sges))] = &elt->sges;
1562 /* These two arrays must have the same size. */
1563 assert(elemof(elt->sges) == elemof(elt->bufs));
1566 wr->next = &(*elts)[(i + 1)].wr;
1567 wr->sg_list = &(*sges)[0];
1568 wr->num_sge = elemof(*sges);
1569 /* For each SGE (segment). */
1570 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1571 struct ibv_sge *sge = &(*sges)[j];
1572 struct rte_mbuf *buf;
1576 assert(buf != NULL);
1577 rte_pktmbuf_reset(buf);
1579 buf = rte_pktmbuf_alloc(rxq->mp);
1581 assert(pool == NULL);
1582 ERROR("%p: empty mbuf pool", (void *)rxq);
1587 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1588 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1589 /* Buffer is supposed to be empty. */
1590 assert(rte_pktmbuf_data_len(buf) == 0);
1591 assert(rte_pktmbuf_pkt_len(buf) == 0);
1592 /* sge->addr must be able to store a pointer. */
1593 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1595 /* The first SGE keeps its headroom. */
1596 sge->addr = (uintptr_t)rte_pktmbuf_mtod(buf,
1598 sge->length = (buf->buf_len -
1599 RTE_PKTMBUF_HEADROOM);
1601 /* Subsequent SGEs lose theirs. */
1602 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1603 SET_DATA_OFF(buf, 0);
1604 sge->addr = (uintptr_t)buf->buf_addr;
1605 sge->length = buf->buf_len;
1607 sge->lkey = rxq->mr->lkey;
1608 /* Redundant check for tailroom. */
1609 assert(sge->length == rte_pktmbuf_tailroom(buf));
1612 /* The last WR pointer must be NULL. */
1613 (*elts)[(i - 1)].wr.next = NULL;
1614 DEBUG("%p: allocated and configured %u WRs (%zu segments)",
1615 (void *)rxq, elts_n, (elts_n * elemof((*elts)[0].sges)));
1616 rxq->elts_n = elts_n;
1617 rxq->elts.sp = elts;
1622 assert(pool == NULL);
1623 for (i = 0; (i != elemof(*elts)); ++i) {
1625 struct rxq_elt_sp *elt = &(*elts)[i];
1627 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1628 struct rte_mbuf *buf = elt->bufs[j];
1631 rte_pktmbuf_free_seg(buf);
1636 DEBUG("%p: failed, freed everything", (void *)rxq);
1642 * Free RX queue elements with scattered packets support.
1645 * Pointer to RX queue structure.
1648 rxq_free_elts_sp(struct rxq *rxq)
1651 unsigned int elts_n = rxq->elts_n;
1652 struct rxq_elt_sp (*elts)[elts_n] = rxq->elts.sp;
1654 DEBUG("%p: freeing WRs", (void *)rxq);
1656 rxq->elts.sp = NULL;
1659 for (i = 0; (i != elemof(*elts)); ++i) {
1661 struct rxq_elt_sp *elt = &(*elts)[i];
1663 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1664 struct rte_mbuf *buf = elt->bufs[j];
1667 rte_pktmbuf_free_seg(buf);
1674 * Allocate RX queue elements.
1677 * Pointer to RX queue structure.
1679 * Number of elements to allocate.
1681 * If not NULL, fetch buffers from this array instead of allocating them
1682 * with rte_pktmbuf_alloc().
1685 * 0 on success, errno value on failure.
1688 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n, struct rte_mbuf **pool)
1691 struct rxq_elt (*elts)[elts_n] =
1692 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1697 ERROR("%p: can't allocate packets array", (void *)rxq);
1701 /* For each WR (packet). */
1702 for (i = 0; (i != elts_n); ++i) {
1703 struct rxq_elt *elt = &(*elts)[i];
1704 struct ibv_recv_wr *wr = &elt->wr;
1705 struct ibv_sge *sge = &(*elts)[i].sge;
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);
1720 /* Configure WR. Work request ID contains its own index in
1721 * the elts array and the offset between SGE buffer header and
1723 WR_ID(wr->wr_id).id = i;
1724 WR_ID(wr->wr_id).offset =
1725 (((uintptr_t)buf->buf_addr + RTE_PKTMBUF_HEADROOM) -
1727 wr->next = &(*elts)[(i + 1)].wr;
1730 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1731 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1732 /* Buffer is supposed to be empty. */
1733 assert(rte_pktmbuf_data_len(buf) == 0);
1734 assert(rte_pktmbuf_pkt_len(buf) == 0);
1735 /* sge->addr must be able to store a pointer. */
1736 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1737 /* SGE keeps its headroom. */
1738 sge->addr = (uintptr_t)
1739 ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
1740 sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
1741 sge->lkey = rxq->mr->lkey;
1742 /* Redundant check for tailroom. */
1743 assert(sge->length == rte_pktmbuf_tailroom(buf));
1744 /* Make sure elts index and SGE mbuf pointer can be deduced
1746 if ((WR_ID(wr->wr_id).id != i) ||
1747 ((void *)((uintptr_t)sge->addr -
1748 WR_ID(wr->wr_id).offset) != buf)) {
1749 ERROR("%p: cannot store index and offset in WR ID",
1752 rte_pktmbuf_free(buf);
1757 /* The last WR pointer must be NULL. */
1758 (*elts)[(i - 1)].wr.next = NULL;
1759 DEBUG("%p: allocated and configured %u single-segment WRs",
1760 (void *)rxq, elts_n);
1761 rxq->elts_n = elts_n;
1762 rxq->elts.no_sp = elts;
1767 assert(pool == NULL);
1768 for (i = 0; (i != elemof(*elts)); ++i) {
1769 struct rxq_elt *elt = &(*elts)[i];
1770 struct rte_mbuf *buf;
1772 if (elt->sge.addr == 0)
1774 assert(WR_ID(elt->wr.wr_id).id == i);
1775 buf = (void *)((uintptr_t)elt->sge.addr -
1776 WR_ID(elt->wr.wr_id).offset);
1777 rte_pktmbuf_free_seg(buf);
1781 DEBUG("%p: failed, freed everything", (void *)rxq);
1787 * Free RX queue elements.
1790 * Pointer to RX queue structure.
1793 rxq_free_elts(struct rxq *rxq)
1796 unsigned int elts_n = rxq->elts_n;
1797 struct rxq_elt (*elts)[elts_n] = rxq->elts.no_sp;
1799 DEBUG("%p: freeing WRs", (void *)rxq);
1801 rxq->elts.no_sp = NULL;
1804 for (i = 0; (i != elemof(*elts)); ++i) {
1805 struct rxq_elt *elt = &(*elts)[i];
1806 struct rte_mbuf *buf;
1808 if (elt->sge.addr == 0)
1810 assert(WR_ID(elt->wr.wr_id).id == i);
1811 buf = (void *)((uintptr_t)elt->sge.addr -
1812 WR_ID(elt->wr.wr_id).offset);
1813 rte_pktmbuf_free_seg(buf);
1819 * Unregister a MAC address from a RX queue.
1822 * Pointer to RX queue structure.
1824 * MAC address index.
1827 rxq_mac_addr_del(struct rxq *rxq, unsigned int mac_index)
1830 struct priv *priv = rxq->priv;
1831 const uint8_t (*mac)[ETHER_ADDR_LEN] =
1832 (const uint8_t (*)[ETHER_ADDR_LEN])
1833 priv->mac[mac_index].addr_bytes;
1836 assert(mac_index < elemof(priv->mac));
1837 if (!BITFIELD_ISSET(rxq->mac_configured, mac_index)) {
1838 assert(rxq->mac_flow[mac_index] == NULL);
1841 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x"
1844 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
1846 assert(rxq->mac_flow[mac_index] != NULL);
1847 claim_zero(ibv_exp_destroy_flow(rxq->mac_flow[mac_index]));
1848 rxq->mac_flow[mac_index] = NULL;
1849 BITFIELD_RESET(rxq->mac_configured, mac_index);
1853 * Unregister all MAC addresses from a RX queue.
1856 * Pointer to RX queue structure.
1859 rxq_mac_addrs_del(struct rxq *rxq)
1861 struct priv *priv = rxq->priv;
1864 for (i = 0; (i != elemof(priv->mac)); ++i)
1865 rxq_mac_addr_del(rxq, i);
1868 static int rxq_promiscuous_enable(struct rxq *);
1869 static void rxq_promiscuous_disable(struct rxq *);
1872 * Register a MAC address in a RX queue.
1875 * Pointer to RX queue structure.
1877 * MAC address index to register.
1880 * 0 on success, errno value on failure.
1883 rxq_mac_addr_add(struct rxq *rxq, unsigned int mac_index)
1885 struct priv *priv = rxq->priv;
1886 const uint8_t (*mac)[ETHER_ADDR_LEN] =
1887 (const uint8_t (*)[ETHER_ADDR_LEN])
1888 priv->mac[mac_index].addr_bytes;
1889 unsigned int vlans = 0;
1890 unsigned int specs = 0;
1892 struct ibv_exp_flow *flow;
1894 assert(mac_index < elemof(priv->mac));
1895 if (BITFIELD_ISSET(rxq->mac_configured, mac_index))
1896 rxq_mac_addr_del(rxq, mac_index);
1897 /* Number of configured VLANs. */
1898 for (i = 0; (i != elemof(priv->vlan_filter)); ++i)
1899 if (priv->vlan_filter[i].enabled)
1901 specs = (vlans ? vlans : 1);
1903 /* Allocate flow specification on the stack. */
1904 struct ibv_exp_flow_attr data
1906 (sizeof(struct ibv_exp_flow_spec_eth[specs]) /
1907 sizeof(struct ibv_exp_flow_attr)) +
1908 !!(sizeof(struct ibv_exp_flow_spec_eth[specs]) %
1909 sizeof(struct ibv_exp_flow_attr))];
1910 struct ibv_exp_flow_attr *attr = (void *)&data[0];
1911 struct ibv_exp_flow_spec_eth *spec = (void *)&data[1];
1914 * No padding must be inserted by the compiler between attr and spec.
1915 * This layout is expected by libibverbs.
1917 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
1918 *attr = (struct ibv_exp_flow_attr){
1919 .type = IBV_EXP_FLOW_ATTR_NORMAL,
1920 .num_of_specs = specs,
1924 *spec = (struct ibv_exp_flow_spec_eth){
1925 .type = IBV_EXP_FLOW_SPEC_ETH,
1926 .size = sizeof(*spec),
1929 (*mac)[0], (*mac)[1], (*mac)[2],
1930 (*mac)[3], (*mac)[4], (*mac)[5]
1934 .dst_mac = "\xff\xff\xff\xff\xff\xff",
1935 .vlan_tag = (vlans ? htons(0xfff) : 0)
1938 /* Fill VLAN specifications. */
1939 for (i = 0, j = 0; (i != elemof(priv->vlan_filter)); ++i) {
1940 if (!priv->vlan_filter[i].enabled)
1945 spec[j].val.vlan_tag = htons(priv->vlan_filter[i].id);
1948 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
1949 " (%u VLAN(s) configured)",
1951 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
1954 /* Create related flow. */
1956 flow = ibv_exp_create_flow(rxq->qp, attr);
1960 /* Flow creation failure is not fatal when in DMFS A0 mode.
1961 * Ignore error if promiscuity is already enabled or can be
1963 if (priv->promisc_ok)
1965 if ((rxq->promisc_flow != NULL) ||
1966 (rxq_promiscuous_enable(rxq) == 0)) {
1967 if (rxq->promisc_flow != NULL)
1968 rxq_promiscuous_disable(rxq);
1969 WARN("cannot configure normal flow but promiscuous"
1970 " mode is fine, assuming promiscuous optimization"
1972 " (options mlx4_core log_num_mgm_entry_size=-7)");
1973 priv->promisc_ok = 1;
1977 /* It's not clear whether errno is always set in this case. */
1978 ERROR("%p: flow configuration failed, errno=%d: %s",
1980 (errno ? strerror(errno) : "Unknown error"));
1985 assert(rxq->mac_flow[mac_index] == NULL);
1986 rxq->mac_flow[mac_index] = flow;
1987 BITFIELD_SET(rxq->mac_configured, mac_index);
1992 * Register all MAC addresses in a RX queue.
1995 * Pointer to RX queue structure.
1998 * 0 on success, errno value on failure.
2001 rxq_mac_addrs_add(struct rxq *rxq)
2003 struct priv *priv = rxq->priv;
2007 for (i = 0; (i != elemof(priv->mac)); ++i) {
2008 if (!BITFIELD_ISSET(priv->mac_configured, i))
2010 ret = rxq_mac_addr_add(rxq, i);
2013 /* Failure, rollback. */
2015 rxq_mac_addr_del(rxq, --i);
2023 * Unregister a MAC address.
2025 * In RSS mode, the MAC address is unregistered from the parent queue,
2026 * otherwise it is unregistered from each queue directly.
2029 * Pointer to private structure.
2031 * MAC address index.
2034 priv_mac_addr_del(struct priv *priv, unsigned int mac_index)
2038 assert(mac_index < elemof(priv->mac));
2039 if (!BITFIELD_ISSET(priv->mac_configured, mac_index))
2042 rxq_mac_addr_del(&priv->rxq_parent, mac_index);
2045 for (i = 0; (i != priv->dev->data->nb_rx_queues); ++i)
2046 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2048 BITFIELD_RESET(priv->mac_configured, mac_index);
2052 * Register a MAC address.
2054 * In RSS mode, the MAC address is registered in the parent queue,
2055 * otherwise it is registered in each queue directly.
2058 * Pointer to private structure.
2060 * MAC address index to use.
2062 * MAC address to register.
2065 * 0 on success, errno value on failure.
2068 priv_mac_addr_add(struct priv *priv, unsigned int mac_index,
2069 const uint8_t (*mac)[ETHER_ADDR_LEN])
2074 assert(mac_index < elemof(priv->mac));
2075 /* First, make sure this address isn't already configured. */
2076 for (i = 0; (i != elemof(priv->mac)); ++i) {
2077 /* Skip this index, it's going to be reconfigured. */
2080 if (!BITFIELD_ISSET(priv->mac_configured, i))
2082 if (memcmp(priv->mac[i].addr_bytes, *mac, sizeof(*mac)))
2084 /* Address already configured elsewhere, return with error. */
2087 if (BITFIELD_ISSET(priv->mac_configured, mac_index))
2088 priv_mac_addr_del(priv, mac_index);
2089 priv->mac[mac_index] = (struct ether_addr){
2091 (*mac)[0], (*mac)[1], (*mac)[2],
2092 (*mac)[3], (*mac)[4], (*mac)[5]
2095 /* If device isn't started, this is all we need to do. */
2096 if (!priv->started) {
2098 /* Verify that all queues have this index disabled. */
2099 for (i = 0; (i != priv->rxqs_n); ++i) {
2100 if ((*priv->rxqs)[i] == NULL)
2102 assert(!BITFIELD_ISSET
2103 ((*priv->rxqs)[i]->mac_configured, mac_index));
2109 ret = rxq_mac_addr_add(&priv->rxq_parent, mac_index);
2114 for (i = 0; (i != priv->rxqs_n); ++i) {
2115 if ((*priv->rxqs)[i] == NULL)
2117 ret = rxq_mac_addr_add((*priv->rxqs)[i], mac_index);
2120 /* Failure, rollback. */
2122 if ((*priv->rxqs)[(--i)] != NULL)
2123 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2127 BITFIELD_SET(priv->mac_configured, mac_index);
2132 * Enable allmulti mode in a RX queue.
2135 * Pointer to RX queue structure.
2138 * 0 on success, errno value on failure.
2141 rxq_allmulticast_enable(struct rxq *rxq)
2143 struct ibv_exp_flow *flow;
2144 struct ibv_exp_flow_attr attr = {
2145 .type = IBV_EXP_FLOW_ATTR_MC_DEFAULT,
2147 .port = rxq->priv->port,
2151 DEBUG("%p: enabling allmulticast mode", (void *)rxq);
2152 if (rxq->allmulti_flow != NULL)
2155 flow = ibv_exp_create_flow(rxq->qp, &attr);
2157 /* It's not clear whether errno is always set in this case. */
2158 ERROR("%p: flow configuration failed, errno=%d: %s",
2160 (errno ? strerror(errno) : "Unknown error"));
2165 rxq->allmulti_flow = flow;
2166 DEBUG("%p: allmulticast mode enabled", (void *)rxq);
2171 * Disable allmulti mode in a RX queue.
2174 * Pointer to RX queue structure.
2177 rxq_allmulticast_disable(struct rxq *rxq)
2179 DEBUG("%p: disabling allmulticast mode", (void *)rxq);
2180 if (rxq->allmulti_flow == NULL)
2182 claim_zero(ibv_exp_destroy_flow(rxq->allmulti_flow));
2183 rxq->allmulti_flow = NULL;
2184 DEBUG("%p: allmulticast mode disabled", (void *)rxq);
2188 * Enable promiscuous mode in a RX queue.
2191 * Pointer to RX queue structure.
2194 * 0 on success, errno value on failure.
2197 rxq_promiscuous_enable(struct rxq *rxq)
2199 struct ibv_exp_flow *flow;
2200 struct ibv_exp_flow_attr attr = {
2201 .type = IBV_EXP_FLOW_ATTR_ALL_DEFAULT,
2203 .port = rxq->priv->port,
2209 DEBUG("%p: enabling promiscuous mode", (void *)rxq);
2210 if (rxq->promisc_flow != NULL)
2213 flow = ibv_exp_create_flow(rxq->qp, &attr);
2215 /* It's not clear whether errno is always set in this case. */
2216 ERROR("%p: flow configuration failed, errno=%d: %s",
2218 (errno ? strerror(errno) : "Unknown error"));
2223 rxq->promisc_flow = flow;
2224 DEBUG("%p: promiscuous mode enabled", (void *)rxq);
2229 * Disable promiscuous mode in a RX queue.
2232 * Pointer to RX queue structure.
2235 rxq_promiscuous_disable(struct rxq *rxq)
2239 DEBUG("%p: disabling promiscuous mode", (void *)rxq);
2240 if (rxq->promisc_flow == NULL)
2242 claim_zero(ibv_exp_destroy_flow(rxq->promisc_flow));
2243 rxq->promisc_flow = NULL;
2244 DEBUG("%p: promiscuous mode disabled", (void *)rxq);
2248 * Clean up a RX queue.
2250 * Destroy objects, free allocated memory and reset the structure for reuse.
2253 * Pointer to RX queue structure.
2256 rxq_cleanup(struct rxq *rxq)
2258 DEBUG("cleaning up %p", (void *)rxq);
2260 rxq_free_elts_sp(rxq);
2263 if (rxq->qp != NULL) {
2264 rxq_promiscuous_disable(rxq);
2265 rxq_allmulticast_disable(rxq);
2266 rxq_mac_addrs_del(rxq);
2267 claim_zero(ibv_destroy_qp(rxq->qp));
2269 if (rxq->cq != NULL)
2270 claim_zero(ibv_destroy_cq(rxq->cq));
2271 if (rxq->mr != NULL)
2272 claim_zero(ibv_dereg_mr(rxq->mr));
2273 memset(rxq, 0, sizeof(*rxq));
2277 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
2280 * DPDK callback for RX with scattered packets support.
2283 * Generic pointer to RX queue structure.
2285 * Array to store received packets.
2287 * Maximum number of packets in array.
2290 * Number of packets successfully received (<= pkts_n).
2293 mlx4_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2295 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2296 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
2297 struct ibv_wc wcs[pkts_n];
2298 struct ibv_recv_wr head;
2299 struct ibv_recv_wr **next = &head.next;
2300 struct ibv_recv_wr *bad_wr;
2305 if (unlikely(!rxq->sp))
2306 return mlx4_rx_burst(dpdk_rxq, pkts, pkts_n);
2307 if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
2309 wcs_n = ibv_poll_cq(rxq->cq, pkts_n, wcs);
2310 if (unlikely(wcs_n == 0))
2312 if (unlikely(wcs_n < 0)) {
2313 DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
2314 (void *)rxq, wcs_n);
2317 assert(wcs_n <= (int)pkts_n);
2318 /* For each work completion. */
2319 for (i = 0; (i != wcs_n); ++i) {
2320 struct ibv_wc *wc = &wcs[i];
2321 uint64_t wr_id = wc->wr_id;
2322 uint32_t len = wc->byte_len;
2323 struct rxq_elt_sp *elt = &(*elts)[wr_id];
2324 struct ibv_recv_wr *wr = &elt->wr;
2325 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
2326 struct rte_mbuf **pkt_buf_next = &pkt_buf;
2327 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
2330 /* Sanity checks. */
2331 assert(wr_id < rxq->elts_n);
2332 assert(wr_id == wr->wr_id);
2333 assert(wr->sg_list == elt->sges);
2334 assert(wr->num_sge == elemof(elt->sges));
2335 /* Link completed WRs together for repost. */
2338 if (unlikely(wc->status != IBV_WC_SUCCESS)) {
2339 /* Whatever, just repost the offending WR. */
2340 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work completion"
2342 (void *)rxq, wc->wr_id, wc->status,
2343 ibv_wc_status_str(wc->status));
2344 #ifdef MLX4_PMD_SOFT_COUNTERS
2345 /* Increase dropped packets counter. */
2346 ++rxq->stats.idropped;
2351 * Replace spent segments with new ones, concatenate and
2352 * return them as pkt_buf.
2355 struct ibv_sge *sge = &elt->sges[j];
2356 struct rte_mbuf *seg = elt->bufs[j];
2357 struct rte_mbuf *rep;
2358 unsigned int seg_tailroom;
2361 * Fetch initial bytes of packet descriptor into a
2362 * cacheline while allocating rep.
2365 rep = __rte_mbuf_raw_alloc(rxq->mp);
2366 if (unlikely(rep == NULL)) {
2368 * Unable to allocate a replacement mbuf,
2371 DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
2372 " can't allocate a new mbuf",
2373 (void *)rxq, wr_id);
2374 if (pkt_buf != NULL)
2375 rte_pktmbuf_free(pkt_buf);
2376 /* Increase out of memory counters. */
2377 ++rxq->stats.rx_nombuf;
2378 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2382 /* Poison user-modifiable fields in rep. */
2383 NEXT(rep) = (void *)((uintptr_t)-1);
2384 SET_DATA_OFF(rep, 0xdead);
2385 DATA_LEN(rep) = 0xd00d;
2386 PKT_LEN(rep) = 0xdeadd00d;
2387 NB_SEGS(rep) = 0x2a;
2391 assert(rep->buf_len == seg->buf_len);
2392 assert(rep->buf_len == rxq->mb_len);
2393 /* Reconfigure sge to use rep instead of seg. */
2394 assert(sge->lkey == rxq->mr->lkey);
2395 sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
2398 /* Update pkt_buf if it's the first segment, or link
2399 * seg to the previous one and update pkt_buf_next. */
2400 *pkt_buf_next = seg;
2401 pkt_buf_next = &NEXT(seg);
2402 /* Update seg information. */
2403 seg_tailroom = (seg->buf_len - seg_headroom);
2404 assert(sge->length == seg_tailroom);
2405 SET_DATA_OFF(seg, seg_headroom);
2406 if (likely(len <= seg_tailroom)) {
2408 DATA_LEN(seg) = len;
2411 assert(rte_pktmbuf_headroom(seg) ==
2413 assert(rte_pktmbuf_tailroom(seg) ==
2414 (seg_tailroom - len));
2417 DATA_LEN(seg) = seg_tailroom;
2418 PKT_LEN(seg) = seg_tailroom;
2420 assert(rte_pktmbuf_headroom(seg) == seg_headroom);
2421 assert(rte_pktmbuf_tailroom(seg) == 0);
2422 /* Fix len and clear headroom for next segments. */
2423 len -= seg_tailroom;
2426 /* Update head and tail segments. */
2427 *pkt_buf_next = NULL;
2428 assert(pkt_buf != NULL);
2430 NB_SEGS(pkt_buf) = j;
2431 PORT(pkt_buf) = rxq->port_id;
2432 PKT_LEN(pkt_buf) = wc->byte_len;
2433 pkt_buf->ol_flags = 0;
2435 /* Return packet. */
2436 *(pkts++) = pkt_buf;
2438 #ifdef MLX4_PMD_SOFT_COUNTERS
2439 /* Increase bytes counter. */
2440 rxq->stats.ibytes += wc->byte_len;
2448 DEBUG("%p: reposting %d WRs starting from %" PRIu64 " (%p)",
2449 (void *)rxq, wcs_n, wcs[0].wr_id, (void *)head.next);
2451 i = ibv_post_recv(rxq->qp, head.next, &bad_wr);
2453 /* Inability to repost WRs is fatal. */
2454 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
2460 #ifdef MLX4_PMD_SOFT_COUNTERS
2461 /* Increase packets counter. */
2462 rxq->stats.ipackets += ret;
2468 * DPDK callback for RX.
2470 * The following function is the same as mlx4_rx_burst_sp(), except it doesn't
2471 * manage scattered packets. Improves performance when MRU is lower than the
2472 * size of the first segment.
2475 * Generic pointer to RX queue structure.
2477 * Array to store received packets.
2479 * Maximum number of packets in array.
2482 * Number of packets successfully received (<= pkts_n).
2485 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2487 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2488 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
2489 struct ibv_wc wcs[pkts_n];
2490 struct ibv_recv_wr head;
2491 struct ibv_recv_wr **next = &head.next;
2492 struct ibv_recv_wr *bad_wr;
2497 if (unlikely(rxq->sp))
2498 return mlx4_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
2499 wcs_n = ibv_poll_cq(rxq->cq, pkts_n, wcs);
2500 if (unlikely(wcs_n == 0))
2502 if (unlikely(wcs_n < 0)) {
2503 DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
2504 (void *)rxq, wcs_n);
2507 assert(wcs_n <= (int)pkts_n);
2508 /* For each work completion. */
2509 for (i = 0; (i != wcs_n); ++i) {
2510 struct ibv_wc *wc = &wcs[i];
2511 uint64_t wr_id = wc->wr_id;
2512 uint32_t len = wc->byte_len;
2513 struct rxq_elt *elt = &(*elts)[WR_ID(wr_id).id];
2514 struct ibv_recv_wr *wr = &elt->wr;
2515 struct rte_mbuf *seg = (void *)((uintptr_t)elt->sge.addr -
2516 WR_ID(wr_id).offset);
2517 struct rte_mbuf *rep;
2519 /* Sanity checks. */
2520 assert(WR_ID(wr_id).id < rxq->elts_n);
2521 assert(wr_id == wr->wr_id);
2522 assert(wr->sg_list == &elt->sge);
2523 assert(wr->num_sge == 1);
2524 /* Link completed WRs together for repost. */
2527 if (unlikely(wc->status != IBV_WC_SUCCESS)) {
2528 /* Whatever, just repost the offending WR. */
2529 DEBUG("rxq=%p, wr_id=%" PRIu32 ": bad work completion"
2531 (void *)rxq, WR_ID(wr_id).id, wc->status,
2532 ibv_wc_status_str(wc->status));
2533 #ifdef MLX4_PMD_SOFT_COUNTERS
2534 /* Increase dropped packets counter. */
2535 ++rxq->stats.idropped;
2540 * Fetch initial bytes of packet descriptor into a
2541 * cacheline while allocating rep.
2544 rep = __rte_mbuf_raw_alloc(rxq->mp);
2545 if (unlikely(rep == NULL)) {
2547 * Unable to allocate a replacement mbuf,
2550 DEBUG("rxq=%p, wr_id=%" PRIu32 ":"
2551 " can't allocate a new mbuf",
2552 (void *)rxq, WR_ID(wr_id).id);
2553 /* Increase out of memory counters. */
2554 ++rxq->stats.rx_nombuf;
2555 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2559 /* Reconfigure sge to use rep instead of seg. */
2560 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
2561 assert(elt->sge.lkey == rxq->mr->lkey);
2562 WR_ID(wr->wr_id).offset =
2563 (((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
2565 assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
2567 /* Update seg information. */
2568 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
2570 PORT(seg) = rxq->port_id;
2573 DATA_LEN(seg) = len;
2576 /* Return packet. */
2579 #ifdef MLX4_PMD_SOFT_COUNTERS
2580 /* Increase bytes counter. */
2581 rxq->stats.ibytes += wc->byte_len;
2589 DEBUG("%p: reposting %d WRs starting from %" PRIu32 " (%p)",
2590 (void *)rxq, wcs_n, WR_ID(wcs[0].wr_id).id, (void *)head.next);
2592 i = ibv_post_recv(rxq->qp, head.next, &bad_wr);
2594 /* Inability to repost WRs is fatal. */
2595 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
2601 #ifdef MLX4_PMD_SOFT_COUNTERS
2602 /* Increase packets counter. */
2603 rxq->stats.ipackets += ret;
2611 * Allocate a Queue Pair in case inline receive is supported.
2614 * Pointer to private structure.
2616 * Completion queue to associate with QP.
2618 * Number of descriptors in QP (hint only).
2621 * QP pointer or NULL in case of error.
2623 static struct ibv_qp *
2624 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
2626 struct ibv_exp_qp_init_attr attr = {
2627 /* CQ to be associated with the send queue. */
2629 /* CQ to be associated with the receive queue. */
2631 .max_inl_recv = priv->inl_recv_size,
2633 /* Max number of outstanding WRs. */
2634 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2635 priv->device_attr.max_qp_wr :
2637 /* Max number of scatter/gather elements in a WR. */
2638 .max_recv_sge = ((priv->device_attr.max_sge <
2639 MLX4_PMD_SGE_WR_N) ?
2640 priv->device_attr.max_sge :
2643 .qp_type = IBV_QPT_RAW_PACKET,
2647 attr.comp_mask = IBV_EXP_QP_INIT_ATTR_PD;
2648 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
2650 return ibv_exp_create_qp(priv->ctx, &attr);
2653 #else /* INLINE_RECV */
2656 * Allocate a Queue Pair.
2659 * Pointer to private structure.
2661 * Completion queue to associate with QP.
2663 * Number of descriptors in QP (hint only).
2666 * QP pointer or NULL in case of error.
2668 static struct ibv_qp *
2669 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
2671 struct ibv_qp_init_attr attr = {
2672 /* CQ to be associated with the send queue. */
2674 /* CQ to be associated with the receive queue. */
2677 /* Max number of outstanding WRs. */
2678 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2679 priv->device_attr.max_qp_wr :
2681 /* Max number of scatter/gather elements in a WR. */
2682 .max_recv_sge = ((priv->device_attr.max_sge <
2683 MLX4_PMD_SGE_WR_N) ?
2684 priv->device_attr.max_sge :
2687 .qp_type = IBV_QPT_RAW_PACKET
2690 return ibv_create_qp(priv->pd, &attr);
2693 #endif /* INLINE_RECV */
2698 * Allocate a RSS Queue Pair.
2701 * Pointer to private structure.
2703 * Completion queue to associate with QP.
2705 * Number of descriptors in QP (hint only).
2707 * If nonzero, create a parent QP, otherwise a child.
2710 * QP pointer or NULL in case of error.
2712 static struct ibv_qp *
2713 rxq_setup_qp_rss(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
2716 struct ibv_exp_qp_init_attr attr = {
2717 /* CQ to be associated with the send queue. */
2719 /* CQ to be associated with the receive queue. */
2722 .max_inl_recv = priv->inl_recv_size,
2725 /* Max number of outstanding WRs. */
2726 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2727 priv->device_attr.max_qp_wr :
2729 /* Max number of scatter/gather elements in a WR. */
2730 .max_recv_sge = ((priv->device_attr.max_sge <
2731 MLX4_PMD_SGE_WR_N) ?
2732 priv->device_attr.max_sge :
2735 .qp_type = IBV_QPT_RAW_PACKET,
2736 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
2737 IBV_EXP_QP_INIT_ATTR_QPG),
2742 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
2745 attr.qpg.qpg_type = IBV_EXP_QPG_PARENT;
2746 /* TSS isn't necessary. */
2747 attr.qpg.parent_attrib.tss_child_count = 0;
2748 attr.qpg.parent_attrib.rss_child_count = priv->rxqs_n;
2749 DEBUG("initializing parent RSS queue");
2751 attr.qpg.qpg_type = IBV_EXP_QPG_CHILD_RX;
2752 attr.qpg.qpg_parent = priv->rxq_parent.qp;
2753 DEBUG("initializing child RSS queue");
2755 return ibv_exp_create_qp(priv->ctx, &attr);
2758 #endif /* RSS_SUPPORT */
2761 * Reconfigure a RX queue with new parameters.
2763 * rxq_rehash() does not allocate mbufs, which, if not done from the right
2764 * thread (such as a control thread), may corrupt the pool.
2765 * In case of failure, the queue is left untouched.
2768 * Pointer to Ethernet device structure.
2773 * 0 on success, errno value on failure.
2776 rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
2778 struct priv *priv = rxq->priv;
2779 struct rxq tmpl = *rxq;
2780 unsigned int mbuf_n;
2781 unsigned int desc_n;
2782 struct rte_mbuf **pool;
2784 struct ibv_exp_qp_attr mod;
2785 struct ibv_recv_wr *bad_wr;
2787 int parent = (rxq == &priv->rxq_parent);
2790 ERROR("%p: cannot rehash parent queue %p",
2791 (void *)dev, (void *)rxq);
2794 DEBUG("%p: rehashing queue %p", (void *)dev, (void *)rxq);
2795 /* Number of descriptors and mbufs currently allocated. */
2796 desc_n = (tmpl.elts_n * (tmpl.sp ? MLX4_PMD_SGE_WR_N : 1));
2798 /* Enable scattered packets support for this queue if necessary. */
2799 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
2800 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
2801 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
2803 desc_n /= MLX4_PMD_SGE_WR_N;
2806 DEBUG("%p: %s scattered packets support (%u WRs)",
2807 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc_n);
2808 /* If scatter mode is the same as before, nothing to do. */
2809 if (tmpl.sp == rxq->sp) {
2810 DEBUG("%p: nothing to do", (void *)dev);
2813 /* Remove attached flows if RSS is disabled (no parent queue). */
2815 rxq_allmulticast_disable(&tmpl);
2816 rxq_promiscuous_disable(&tmpl);
2817 rxq_mac_addrs_del(&tmpl);
2818 /* Update original queue in case of failure. */
2819 rxq->allmulti_flow = tmpl.allmulti_flow;
2820 rxq->promisc_flow = tmpl.promisc_flow;
2821 memcpy(rxq->mac_configured, tmpl.mac_configured,
2822 sizeof(rxq->mac_configured));
2823 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
2825 /* From now on, any failure will render the queue unusable.
2826 * Reinitialize QP. */
2827 mod = (struct ibv_exp_qp_attr){ .qp_state = IBV_QPS_RESET };
2828 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
2830 ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
2834 err = ibv_resize_cq(tmpl.cq, desc_n);
2836 ERROR("%p: cannot resize CQ: %s", (void *)dev, strerror(err));
2840 mod = (struct ibv_exp_qp_attr){
2841 /* Move the QP to this state. */
2842 .qp_state = IBV_QPS_INIT,
2843 /* Primary port number. */
2844 .port_num = priv->port
2846 err = ibv_exp_modify_qp(tmpl.qp, &mod,
2849 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
2850 #endif /* RSS_SUPPORT */
2853 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
2854 (void *)dev, strerror(err));
2858 /* Reconfigure flows. Do not care for errors. */
2860 rxq_mac_addrs_add(&tmpl);
2862 rxq_promiscuous_enable(&tmpl);
2864 rxq_allmulticast_enable(&tmpl);
2865 /* Update original queue in case of failure. */
2866 rxq->allmulti_flow = tmpl.allmulti_flow;
2867 rxq->promisc_flow = tmpl.promisc_flow;
2868 memcpy(rxq->mac_configured, tmpl.mac_configured,
2869 sizeof(rxq->mac_configured));
2870 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
2872 /* Allocate pool. */
2873 pool = rte_malloc(__func__, (mbuf_n * sizeof(*pool)), 0);
2875 ERROR("%p: cannot allocate memory", (void *)dev);
2878 /* Snatch mbufs from original queue. */
2881 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
2883 for (i = 0; (i != elemof(*elts)); ++i) {
2884 struct rxq_elt_sp *elt = &(*elts)[i];
2887 for (j = 0; (j != elemof(elt->bufs)); ++j) {
2888 assert(elt->bufs[j] != NULL);
2889 pool[k++] = elt->bufs[j];
2893 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
2895 for (i = 0; (i != elemof(*elts)); ++i) {
2896 struct rxq_elt *elt = &(*elts)[i];
2897 struct rte_mbuf *buf = (void *)
2898 ((uintptr_t)elt->sge.addr -
2899 WR_ID(elt->wr.wr_id).offset);
2901 assert(WR_ID(elt->wr.wr_id).id == i);
2905 assert(k == mbuf_n);
2907 tmpl.elts.sp = NULL;
2908 assert((void *)&tmpl.elts.sp == (void *)&tmpl.elts.no_sp);
2910 rxq_alloc_elts_sp(&tmpl, desc_n, pool) :
2911 rxq_alloc_elts(&tmpl, desc_n, pool));
2913 ERROR("%p: cannot reallocate WRs, aborting", (void *)dev);
2918 assert(tmpl.elts_n == desc_n);
2919 assert(tmpl.elts.sp != NULL);
2921 /* Clean up original data. */
2923 rte_free(rxq->elts.sp);
2924 rxq->elts.sp = NULL;
2926 err = ibv_post_recv(tmpl.qp,
2928 &(*tmpl.elts.sp)[0].wr :
2929 &(*tmpl.elts.no_sp)[0].wr),
2932 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
2938 mod = (struct ibv_exp_qp_attr){
2939 .qp_state = IBV_QPS_RTR
2941 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
2943 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
2944 (void *)dev, strerror(err));
2952 * Configure a RX queue.
2955 * Pointer to Ethernet device structure.
2957 * Pointer to RX queue structure.
2959 * Number of descriptors to configure in queue.
2961 * NUMA socket on which memory must be allocated.
2963 * Thresholds parameters.
2965 * Memory pool for buffer allocations.
2968 * 0 on success, errno value on failure.
2971 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
2972 unsigned int socket, const struct rte_eth_rxconf *conf,
2973 struct rte_mempool *mp)
2975 struct priv *priv = dev->data->dev_private;
2981 struct ibv_exp_qp_attr mod;
2982 struct ibv_recv_wr *bad_wr;
2983 struct rte_mbuf *buf;
2985 int parent = (rxq == &priv->rxq_parent);
2987 (void)conf; /* Thresholds configuration (ignored). */
2989 * If this is a parent queue, hardware must support RSS and
2990 * RSS must be enabled.
2992 assert((!parent) || ((priv->hw_rss) && (priv->rss)));
2994 /* Even if unused, ibv_create_cq() requires at least one
2999 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
3000 ERROR("%p: invalid number of RX descriptors (must be a"
3001 " multiple of %d)", (void *)dev, desc);
3004 /* Get mbuf length. */
3005 buf = rte_pktmbuf_alloc(mp);
3007 ERROR("%p: unable to allocate mbuf", (void *)dev);
3010 tmpl.mb_len = buf->buf_len;
3011 assert((rte_pktmbuf_headroom(buf) +
3012 rte_pktmbuf_tailroom(buf)) == tmpl.mb_len);
3013 assert(rte_pktmbuf_headroom(buf) == RTE_PKTMBUF_HEADROOM);
3014 rte_pktmbuf_free(buf);
3015 /* Enable scattered packets support for this queue if necessary. */
3016 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3017 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3018 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3020 desc /= MLX4_PMD_SGE_WR_N;
3022 DEBUG("%p: %s scattered packets support (%u WRs)",
3023 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc);
3024 /* Use the entire RX mempool as the memory region. */
3025 tmpl.mr = ibv_reg_mr(priv->pd,
3026 (void *)mp->elt_va_start,
3027 (mp->elt_va_end - mp->elt_va_start),
3028 (IBV_ACCESS_LOCAL_WRITE |
3029 IBV_ACCESS_REMOTE_WRITE));
3030 if (tmpl.mr == NULL) {
3032 ERROR("%p: MR creation failure: %s",
3033 (void *)dev, strerror(ret));
3037 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
3038 if (tmpl.cq == NULL) {
3040 ERROR("%p: CQ creation failure: %s",
3041 (void *)dev, strerror(ret));
3044 DEBUG("priv->device_attr.max_qp_wr is %d",
3045 priv->device_attr.max_qp_wr);
3046 DEBUG("priv->device_attr.max_sge is %d",
3047 priv->device_attr.max_sge);
3050 tmpl.qp = rxq_setup_qp_rss(priv, tmpl.cq, desc, parent);
3052 #endif /* RSS_SUPPORT */
3053 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc);
3054 if (tmpl.qp == NULL) {
3055 ret = (errno ? errno : EINVAL);
3056 ERROR("%p: QP creation failure: %s",
3057 (void *)dev, strerror(ret));
3060 mod = (struct ibv_exp_qp_attr){
3061 /* Move the QP to this state. */
3062 .qp_state = IBV_QPS_INIT,
3063 /* Primary port number. */
3064 .port_num = priv->port
3066 ret = ibv_exp_modify_qp(tmpl.qp, &mod,
3069 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3070 #endif /* RSS_SUPPORT */
3073 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3074 (void *)dev, strerror(ret));
3077 if ((parent) || (!priv->rss)) {
3078 /* Configure MAC and broadcast addresses. */
3079 ret = rxq_mac_addrs_add(&tmpl);
3081 ERROR("%p: QP flow attachment failed: %s",
3082 (void *)dev, strerror(ret));
3086 /* Allocate descriptors for RX queues, except for the RSS parent. */
3090 ret = rxq_alloc_elts_sp(&tmpl, desc, NULL);
3092 ret = rxq_alloc_elts(&tmpl, desc, NULL);
3094 ERROR("%p: RXQ allocation failed: %s",
3095 (void *)dev, strerror(ret));
3098 ret = ibv_post_recv(tmpl.qp,
3100 &(*tmpl.elts.sp)[0].wr :
3101 &(*tmpl.elts.no_sp)[0].wr),
3104 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3111 mod = (struct ibv_exp_qp_attr){
3112 .qp_state = IBV_QPS_RTR
3114 ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3116 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3117 (void *)dev, strerror(ret));
3121 tmpl.port_id = dev->data->port_id;
3122 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
3123 /* Clean up rxq in case we're reinitializing it. */
3124 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
3127 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
3137 * DPDK callback to configure a RX queue.
3140 * Pointer to Ethernet device structure.
3144 * Number of descriptors to configure in queue.
3146 * NUMA socket on which memory must be allocated.
3148 * Thresholds parameters.
3150 * Memory pool for buffer allocations.
3153 * 0 on success, negative errno value on failure.
3156 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
3157 unsigned int socket, const struct rte_eth_rxconf *conf,
3158 struct rte_mempool *mp)
3160 struct priv *priv = dev->data->dev_private;
3161 struct rxq *rxq = (*priv->rxqs)[idx];
3165 DEBUG("%p: configuring queue %u for %u descriptors",
3166 (void *)dev, idx, desc);
3167 if (idx >= priv->rxqs_n) {
3168 ERROR("%p: queue index out of range (%u >= %u)",
3169 (void *)dev, idx, priv->rxqs_n);
3174 DEBUG("%p: reusing already allocated queue index %u (%p)",
3175 (void *)dev, idx, (void *)rxq);
3176 if (priv->started) {
3180 (*priv->rxqs)[idx] = NULL;
3183 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
3185 ERROR("%p: unable to allocate queue index %u",
3191 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
3195 rxq->stats.idx = idx;
3196 DEBUG("%p: adding RX queue %p to list",
3197 (void *)dev, (void *)rxq);
3198 (*priv->rxqs)[idx] = rxq;
3199 /* Update receive callback. */
3201 dev->rx_pkt_burst = mlx4_rx_burst_sp;
3203 dev->rx_pkt_burst = mlx4_rx_burst;
3210 * DPDK callback to release a RX queue.
3213 * Generic RX queue pointer.
3216 mlx4_rx_queue_release(void *dpdk_rxq)
3218 struct rxq *rxq = (struct rxq *)dpdk_rxq;
3226 assert(rxq != &priv->rxq_parent);
3227 for (i = 0; (i != priv->rxqs_n); ++i)
3228 if ((*priv->rxqs)[i] == rxq) {
3229 DEBUG("%p: removing RX queue %p from list",
3230 (void *)priv->dev, (void *)rxq);
3231 (*priv->rxqs)[i] = NULL;
3240 * DPDK callback to start the device.
3242 * Simulate device start by attaching all configured flows.
3245 * Pointer to Ethernet device structure.
3248 * 0 on success, negative errno value on failure.
3251 mlx4_dev_start(struct rte_eth_dev *dev)
3253 struct priv *priv = dev->data->dev_private;
3259 if (priv->started) {
3263 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
3266 rxq = &priv->rxq_parent;
3269 rxq = (*priv->rxqs)[0];
3272 /* Iterate only once when RSS is enabled. */
3276 /* Ignore nonexistent RX queues. */
3279 ret = rxq_mac_addrs_add(rxq);
3280 if (!ret && priv->promisc)
3281 ret = rxq_promiscuous_enable(rxq);
3282 if (!ret && priv->allmulti)
3283 ret = rxq_allmulticast_enable(rxq);
3286 WARN("%p: QP flow attachment failed: %s",
3287 (void *)dev, strerror(ret));
3290 rxq = (*priv->rxqs)[--i];
3292 rxq_allmulticast_disable(rxq);
3293 rxq_promiscuous_disable(rxq);
3294 rxq_mac_addrs_del(rxq);
3299 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3305 * DPDK callback to stop the device.
3307 * Simulate device stop by detaching all configured flows.
3310 * Pointer to Ethernet device structure.
3313 mlx4_dev_stop(struct rte_eth_dev *dev)
3315 struct priv *priv = dev->data->dev_private;
3321 if (!priv->started) {
3325 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
3328 rxq = &priv->rxq_parent;
3331 rxq = (*priv->rxqs)[0];
3334 /* Iterate only once when RSS is enabled. */
3336 /* Ignore nonexistent RX queues. */
3339 rxq_allmulticast_disable(rxq);
3340 rxq_promiscuous_disable(rxq);
3341 rxq_mac_addrs_del(rxq);
3342 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3347 * Dummy DPDK callback for TX.
3349 * This function is used to temporarily replace the real callback during
3350 * unsafe control operations on the queue, or in case of error.
3353 * Generic pointer to TX queue structure.
3355 * Packets to transmit.
3357 * Number of packets in array.
3360 * Number of packets successfully transmitted (<= pkts_n).
3363 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
3372 * Dummy DPDK callback for RX.
3374 * This function is used to temporarily replace the real callback during
3375 * unsafe control operations on the queue, or in case of error.
3378 * Generic pointer to RX queue structure.
3380 * Array to store received packets.
3382 * Maximum number of packets in array.
3385 * Number of packets successfully received (<= pkts_n).
3388 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
3397 * DPDK callback to close the device.
3399 * Destroy all queues and objects, free memory.
3402 * Pointer to Ethernet device structure.
3405 mlx4_dev_close(struct rte_eth_dev *dev)
3407 struct priv *priv = dev->data->dev_private;
3412 DEBUG("%p: closing device \"%s\"",
3414 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
3415 /* Prevent crashes when queues are still in use. This is unfortunately
3416 * still required for DPDK 1.3 because some programs (such as testpmd)
3417 * never release them before closing the device. */
3418 dev->rx_pkt_burst = removed_rx_burst;
3419 dev->tx_pkt_burst = removed_tx_burst;
3420 if (priv->rxqs != NULL) {
3421 /* XXX race condition if mlx4_rx_burst() is still running. */
3423 for (i = 0; (i != priv->rxqs_n); ++i) {
3424 tmp = (*priv->rxqs)[i];
3427 (*priv->rxqs)[i] = NULL;
3434 if (priv->txqs != NULL) {
3435 /* XXX race condition if mlx4_tx_burst() is still running. */
3437 for (i = 0; (i != priv->txqs_n); ++i) {
3438 tmp = (*priv->txqs)[i];
3441 (*priv->txqs)[i] = NULL;
3449 rxq_cleanup(&priv->rxq_parent);
3450 if (priv->pd != NULL) {
3451 assert(priv->ctx != NULL);
3452 claim_zero(ibv_dealloc_pd(priv->pd));
3453 claim_zero(ibv_close_device(priv->ctx));
3455 assert(priv->ctx == NULL);
3457 memset(priv, 0, sizeof(*priv));
3461 * DPDK callback to get information about the device.
3464 * Pointer to Ethernet device structure.
3466 * Info structure output buffer.
3469 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
3471 struct priv *priv = dev->data->dev_private;
3475 /* FIXME: we should ask the device for these values. */
3476 info->min_rx_bufsize = 32;
3477 info->max_rx_pktlen = 65536;
3479 * Since we need one CQ per QP, the limit is the minimum number
3480 * between the two values.
3482 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
3483 priv->device_attr.max_qp : priv->device_attr.max_cq);
3484 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
3487 info->max_rx_queues = max;
3488 info->max_tx_queues = max;
3489 info->max_mac_addrs = elemof(priv->mac);
3494 * DPDK callback to get device statistics.
3497 * Pointer to Ethernet device structure.
3499 * Stats structure output buffer.
3502 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
3504 struct priv *priv = dev->data->dev_private;
3505 struct rte_eth_stats tmp = {0};
3510 /* Add software counters. */
3511 for (i = 0; (i != priv->rxqs_n); ++i) {
3512 struct rxq *rxq = (*priv->rxqs)[i];
3516 idx = rxq->stats.idx;
3517 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3518 #ifdef MLX4_PMD_SOFT_COUNTERS
3519 tmp.q_ipackets[idx] += rxq->stats.ipackets;
3520 tmp.q_ibytes[idx] += rxq->stats.ibytes;
3522 tmp.q_errors[idx] += (rxq->stats.idropped +
3523 rxq->stats.rx_nombuf);
3525 #ifdef MLX4_PMD_SOFT_COUNTERS
3526 tmp.ipackets += rxq->stats.ipackets;
3527 tmp.ibytes += rxq->stats.ibytes;
3529 tmp.ierrors += rxq->stats.idropped;
3530 tmp.rx_nombuf += rxq->stats.rx_nombuf;
3532 for (i = 0; (i != priv->txqs_n); ++i) {
3533 struct txq *txq = (*priv->txqs)[i];
3537 idx = txq->stats.idx;
3538 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3539 #ifdef MLX4_PMD_SOFT_COUNTERS
3540 tmp.q_opackets[idx] += txq->stats.opackets;
3541 tmp.q_obytes[idx] += txq->stats.obytes;
3543 tmp.q_errors[idx] += txq->stats.odropped;
3545 #ifdef MLX4_PMD_SOFT_COUNTERS
3546 tmp.opackets += txq->stats.opackets;
3547 tmp.obytes += txq->stats.obytes;
3549 tmp.oerrors += txq->stats.odropped;
3551 #ifndef MLX4_PMD_SOFT_COUNTERS
3552 /* FIXME: retrieve and add hardware counters. */
3559 * DPDK callback to clear device statistics.
3562 * Pointer to Ethernet device structure.
3565 mlx4_stats_reset(struct rte_eth_dev *dev)
3567 struct priv *priv = dev->data->dev_private;
3572 for (i = 0; (i != priv->rxqs_n); ++i) {
3573 if ((*priv->rxqs)[i] == NULL)
3575 idx = (*priv->rxqs)[i]->stats.idx;
3576 (*priv->rxqs)[i]->stats =
3577 (struct mlx4_rxq_stats){ .idx = idx };
3579 for (i = 0; (i != priv->txqs_n); ++i) {
3580 if ((*priv->txqs)[i] == NULL)
3582 idx = (*priv->rxqs)[i]->stats.idx;
3583 (*priv->txqs)[i]->stats =
3584 (struct mlx4_txq_stats){ .idx = idx };
3586 #ifndef MLX4_PMD_SOFT_COUNTERS
3587 /* FIXME: reset hardware counters. */
3593 * DPDK callback to remove a MAC address.
3596 * Pointer to Ethernet device structure.
3598 * MAC address index.
3601 mlx4_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
3603 struct priv *priv = dev->data->dev_private;
3606 DEBUG("%p: removing MAC address from index %" PRIu32,
3607 (void *)dev, index);
3608 if (index >= MLX4_MAX_MAC_ADDRESSES)
3610 /* Refuse to remove the broadcast address, this one is special. */
3611 if (!memcmp(priv->mac[index].addr_bytes, "\xff\xff\xff\xff\xff\xff",
3614 priv_mac_addr_del(priv, index);
3620 * DPDK callback to add a MAC address.
3623 * Pointer to Ethernet device structure.
3625 * MAC address to register.
3627 * MAC address index.
3629 * VMDq pool index to associate address with (ignored).
3632 mlx4_mac_addr_add(struct rte_eth_dev *dev, struct ether_addr *mac_addr,
3633 uint32_t index, uint32_t vmdq)
3635 struct priv *priv = dev->data->dev_private;
3639 DEBUG("%p: adding MAC address at index %" PRIu32,
3640 (void *)dev, index);
3641 if (index >= MLX4_MAX_MAC_ADDRESSES)
3643 /* Refuse to add the broadcast address, this one is special. */
3644 if (!memcmp(mac_addr->addr_bytes, "\xff\xff\xff\xff\xff\xff",
3647 priv_mac_addr_add(priv, index,
3648 (const uint8_t (*)[ETHER_ADDR_LEN])
3649 mac_addr->addr_bytes);
3655 * DPDK callback to enable promiscuous mode.
3658 * Pointer to Ethernet device structure.
3661 mlx4_promiscuous_enable(struct rte_eth_dev *dev)
3663 struct priv *priv = dev->data->dev_private;
3668 if (priv->promisc) {
3672 /* If device isn't started, this is all we need to do. */
3676 ret = rxq_promiscuous_enable(&priv->rxq_parent);
3683 for (i = 0; (i != priv->rxqs_n); ++i) {
3684 if ((*priv->rxqs)[i] == NULL)
3686 ret = rxq_promiscuous_enable((*priv->rxqs)[i]);
3689 /* Failure, rollback. */
3691 if ((*priv->rxqs)[--i] != NULL)
3692 rxq_promiscuous_disable((*priv->rxqs)[i]);
3702 * DPDK callback to disable promiscuous mode.
3705 * Pointer to Ethernet device structure.
3708 mlx4_promiscuous_disable(struct rte_eth_dev *dev)
3710 struct priv *priv = dev->data->dev_private;
3714 if (!priv->promisc) {
3719 rxq_promiscuous_disable(&priv->rxq_parent);
3722 for (i = 0; (i != priv->rxqs_n); ++i)
3723 if ((*priv->rxqs)[i] != NULL)
3724 rxq_promiscuous_disable((*priv->rxqs)[i]);
3731 * DPDK callback to enable allmulti mode.
3734 * Pointer to Ethernet device structure.
3737 mlx4_allmulticast_enable(struct rte_eth_dev *dev)
3739 struct priv *priv = dev->data->dev_private;
3744 if (priv->allmulti) {
3748 /* If device isn't started, this is all we need to do. */
3752 ret = rxq_allmulticast_enable(&priv->rxq_parent);
3759 for (i = 0; (i != priv->rxqs_n); ++i) {
3760 if ((*priv->rxqs)[i] == NULL)
3762 ret = rxq_allmulticast_enable((*priv->rxqs)[i]);
3765 /* Failure, rollback. */
3767 if ((*priv->rxqs)[--i] != NULL)
3768 rxq_allmulticast_disable((*priv->rxqs)[i]);
3778 * DPDK callback to disable allmulti mode.
3781 * Pointer to Ethernet device structure.
3784 mlx4_allmulticast_disable(struct rte_eth_dev *dev)
3786 struct priv *priv = dev->data->dev_private;
3790 if (!priv->allmulti) {
3795 rxq_allmulticast_disable(&priv->rxq_parent);
3798 for (i = 0; (i != priv->rxqs_n); ++i)
3799 if ((*priv->rxqs)[i] != NULL)
3800 rxq_allmulticast_disable((*priv->rxqs)[i]);
3807 * DPDK callback to retrieve physical link information (unlocked version).
3810 * Pointer to Ethernet device structure.
3811 * @param wait_to_complete
3812 * Wait for request completion (ignored).
3815 mlx4_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
3817 struct priv *priv = dev->data->dev_private;
3818 struct ibv_port_attr port_attr;
3819 static const uint8_t width_mult[] = {
3820 /* Multiplier values taken from devinfo.c in libibverbs. */
3821 0, 1, 4, 0, 8, 0, 0, 0, 12, 0
3824 (void)wait_to_complete;
3825 errno = ibv_query_port(priv->ctx, priv->port, &port_attr);
3827 WARN("port query failed: %s", strerror(errno));
3830 dev->data->dev_link = (struct rte_eth_link){
3831 .link_speed = (ibv_rate_to_mbps(mult_to_ibv_rate
3832 (port_attr.active_speed)) *
3833 width_mult[(port_attr.active_width %
3834 sizeof(width_mult))]),
3835 .link_duplex = ETH_LINK_FULL_DUPLEX,
3836 .link_status = (port_attr.state == IBV_PORT_ACTIVE)
3838 if (memcmp(&port_attr, &priv->port_attr, sizeof(port_attr))) {
3839 /* Link status changed. */
3840 priv->port_attr = port_attr;
3843 /* Link status is still the same. */
3848 * DPDK callback to retrieve physical link information.
3851 * Pointer to Ethernet device structure.
3852 * @param wait_to_complete
3853 * Wait for request completion (ignored).
3856 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
3858 struct priv *priv = dev->data->dev_private;
3862 ret = mlx4_link_update_unlocked(dev, wait_to_complete);
3868 * DPDK callback to change the MTU.
3870 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
3871 * received). Use this as a hint to enable/disable scattered packets support
3872 * and improve performance when not needed.
3873 * Since failure is not an option, reconfiguring queues on the fly is not
3877 * Pointer to Ethernet device structure.
3882 * 0 on success, negative errno value on failure.
3885 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
3887 struct priv *priv = dev->data->dev_private;
3890 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
3894 /* Set kernel interface MTU first. */
3895 if (priv_set_mtu(priv, mtu)) {
3897 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
3901 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
3903 /* Temporarily replace RX handler with a fake one, assuming it has not
3904 * been copied elsewhere. */
3905 dev->rx_pkt_burst = removed_rx_burst;
3906 /* Make sure everyone has left mlx4_rx_burst() and uses
3907 * removed_rx_burst() instead. */
3910 /* Reconfigure each RX queue. */
3911 for (i = 0; (i != priv->rxqs_n); ++i) {
3912 struct rxq *rxq = (*priv->rxqs)[i];
3913 unsigned int max_frame_len;
3918 /* Calculate new maximum frame length according to MTU and
3919 * toggle scattered support (sp) if necessary. */
3920 max_frame_len = (priv->mtu + ETHER_HDR_LEN +
3921 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
3922 sp = (max_frame_len > (rxq->mb_len - RTE_PKTMBUF_HEADROOM));
3923 /* Provide new values to rxq_setup(). */
3924 dev->data->dev_conf.rxmode.jumbo_frame = sp;
3925 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
3926 ret = rxq_rehash(dev, rxq);
3928 /* Force SP RX if that queue requires it and abort. */
3930 rx_func = mlx4_rx_burst_sp;
3933 /* Reenable non-RSS queue attributes. No need to check
3934 * for errors at this stage. */
3936 rxq_mac_addrs_add(rxq);
3938 rxq_promiscuous_enable(rxq);
3940 rxq_allmulticast_enable(rxq);
3942 /* Scattered burst function takes priority. */
3944 rx_func = mlx4_rx_burst_sp;
3946 /* Burst functions can now be called again. */
3948 dev->rx_pkt_burst = rx_func;
3956 * DPDK callback to get flow control status.
3959 * Pointer to Ethernet device structure.
3960 * @param[out] fc_conf
3961 * Flow control output buffer.
3964 * 0 on success, negative errno value on failure.
3967 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
3969 struct priv *priv = dev->data->dev_private;
3971 struct ethtool_pauseparam ethpause = {
3972 .cmd = ETHTOOL_GPAUSEPARAM
3976 ifr.ifr_data = ðpause;
3978 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
3980 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
3986 fc_conf->autoneg = ethpause.autoneg;
3987 if (ethpause.rx_pause && ethpause.tx_pause)
3988 fc_conf->mode = RTE_FC_FULL;
3989 else if (ethpause.rx_pause)
3990 fc_conf->mode = RTE_FC_RX_PAUSE;
3991 else if (ethpause.tx_pause)
3992 fc_conf->mode = RTE_FC_TX_PAUSE;
3994 fc_conf->mode = RTE_FC_NONE;
4004 * DPDK callback to modify flow control parameters.
4007 * Pointer to Ethernet device structure.
4008 * @param[in] fc_conf
4009 * Flow control parameters.
4012 * 0 on success, negative errno value on failure.
4015 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4017 struct priv *priv = dev->data->dev_private;
4019 struct ethtool_pauseparam ethpause = {
4020 .cmd = ETHTOOL_SPAUSEPARAM
4024 ifr.ifr_data = ðpause;
4025 ethpause.autoneg = fc_conf->autoneg;
4026 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4027 (fc_conf->mode & RTE_FC_RX_PAUSE))
4028 ethpause.rx_pause = 1;
4030 ethpause.rx_pause = 0;
4032 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4033 (fc_conf->mode & RTE_FC_TX_PAUSE))
4034 ethpause.tx_pause = 1;
4036 ethpause.tx_pause = 0;
4039 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4041 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
4055 * Configure a VLAN filter.
4058 * Pointer to Ethernet device structure.
4060 * VLAN ID to filter.
4065 * 0 on success, errno value on failure.
4068 vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4070 struct priv *priv = dev->data->dev_private;
4072 unsigned int j = -1;
4074 DEBUG("%p: %s VLAN filter ID %" PRIu16,
4075 (void *)dev, (on ? "enable" : "disable"), vlan_id);
4076 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
4077 if (!priv->vlan_filter[i].enabled) {
4078 /* Unused index, remember it. */
4082 if (priv->vlan_filter[i].id != vlan_id)
4084 /* This VLAN ID is already known, use its index. */
4088 /* Check if there's room for another VLAN filter. */
4089 if (j == (unsigned int)-1)
4092 * VLAN filters apply to all configured MAC addresses, flow
4093 * specifications must be reconfigured accordingly.
4095 priv->vlan_filter[j].id = vlan_id;
4096 if ((on) && (!priv->vlan_filter[j].enabled)) {
4098 * Filter is disabled, enable it.
4099 * Rehashing flows in all RX queues is necessary.
4102 rxq_mac_addrs_del(&priv->rxq_parent);
4104 for (i = 0; (i != priv->rxqs_n); ++i)
4105 if ((*priv->rxqs)[i] != NULL)
4106 rxq_mac_addrs_del((*priv->rxqs)[i]);
4107 priv->vlan_filter[j].enabled = 1;
4108 if (priv->started) {
4110 rxq_mac_addrs_add(&priv->rxq_parent);
4112 for (i = 0; (i != priv->rxqs_n); ++i) {
4113 if ((*priv->rxqs)[i] == NULL)
4115 rxq_mac_addrs_add((*priv->rxqs)[i]);
4118 } else if ((!on) && (priv->vlan_filter[j].enabled)) {
4120 * Filter is enabled, disable it.
4121 * Rehashing flows in all RX queues is necessary.
4124 rxq_mac_addrs_del(&priv->rxq_parent);
4126 for (i = 0; (i != priv->rxqs_n); ++i)
4127 if ((*priv->rxqs)[i] != NULL)
4128 rxq_mac_addrs_del((*priv->rxqs)[i]);
4129 priv->vlan_filter[j].enabled = 0;
4130 if (priv->started) {
4132 rxq_mac_addrs_add(&priv->rxq_parent);
4134 for (i = 0; (i != priv->rxqs_n); ++i) {
4135 if ((*priv->rxqs)[i] == NULL)
4137 rxq_mac_addrs_add((*priv->rxqs)[i]);
4145 * DPDK callback to configure a VLAN filter.
4148 * Pointer to Ethernet device structure.
4150 * VLAN ID to filter.
4155 * 0 on success, negative errno value on failure.
4158 mlx4_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4160 struct priv *priv = dev->data->dev_private;
4164 ret = vlan_filter_set(dev, vlan_id, on);
4170 static const struct eth_dev_ops mlx4_dev_ops = {
4171 .dev_configure = mlx4_dev_configure,
4172 .dev_start = mlx4_dev_start,
4173 .dev_stop = mlx4_dev_stop,
4174 .dev_close = mlx4_dev_close,
4175 .promiscuous_enable = mlx4_promiscuous_enable,
4176 .promiscuous_disable = mlx4_promiscuous_disable,
4177 .allmulticast_enable = mlx4_allmulticast_enable,
4178 .allmulticast_disable = mlx4_allmulticast_disable,
4179 .link_update = mlx4_link_update,
4180 .stats_get = mlx4_stats_get,
4181 .stats_reset = mlx4_stats_reset,
4182 .queue_stats_mapping_set = NULL,
4183 .dev_infos_get = mlx4_dev_infos_get,
4184 .vlan_filter_set = mlx4_vlan_filter_set,
4185 .vlan_tpid_set = NULL,
4186 .vlan_strip_queue_set = NULL,
4187 .vlan_offload_set = NULL,
4188 .rx_queue_setup = mlx4_rx_queue_setup,
4189 .tx_queue_setup = mlx4_tx_queue_setup,
4190 .rx_queue_release = mlx4_rx_queue_release,
4191 .tx_queue_release = mlx4_tx_queue_release,
4193 .dev_led_off = NULL,
4194 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
4195 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
4196 .priority_flow_ctrl_set = NULL,
4197 .mac_addr_remove = mlx4_mac_addr_remove,
4198 .mac_addr_add = mlx4_mac_addr_add,
4199 .mtu_set = mlx4_dev_set_mtu,
4200 .fdir_add_signature_filter = NULL,
4201 .fdir_update_signature_filter = NULL,
4202 .fdir_remove_signature_filter = NULL,
4203 .fdir_add_perfect_filter = NULL,
4204 .fdir_update_perfect_filter = NULL,
4205 .fdir_remove_perfect_filter = NULL,
4206 .fdir_set_masks = NULL
4210 * Get PCI information from struct ibv_device.
4213 * Pointer to Ethernet device structure.
4214 * @param[out] pci_addr
4215 * PCI bus address output buffer.
4218 * 0 on success, -1 on failure and errno is set.
4221 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
4222 struct rte_pci_addr *pci_addr)
4226 MKSTR(path, "%s/device/uevent", device->ibdev_path);
4228 file = fopen(path, "rb");
4231 while (fgets(line, sizeof(line), file) == line) {
4232 size_t len = strlen(line);
4235 /* Truncate long lines. */
4236 if (len == (sizeof(line) - 1))
4237 while (line[(len - 1)] != '\n') {
4241 line[(len - 1)] = ret;
4243 /* Extract information. */
4246 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
4250 &pci_addr->function) == 4) {
4260 * Derive MAC address from port GID.
4263 * MAC address output buffer.
4265 * Physical port number.
4270 mac_from_gid(uint8_t (*mac)[ETHER_ADDR_LEN], uint32_t port, uint8_t *gid)
4272 memcpy(&(*mac)[0], gid + 8, 3);
4273 memcpy(&(*mac)[3], gid + 13, 3);
4278 /* Support up to 32 adapters. */
4280 struct rte_pci_addr pci_addr; /* associated PCI address */
4281 uint32_t ports; /* physical ports bitfield. */
4285 * Get device index in mlx4_dev[] from PCI bus address.
4287 * @param[in] pci_addr
4288 * PCI bus address to look for.
4291 * mlx4_dev[] index on success, -1 on failure.
4294 mlx4_dev_idx(struct rte_pci_addr *pci_addr)
4299 assert(pci_addr != NULL);
4300 for (i = 0; (i != elemof(mlx4_dev)); ++i) {
4301 if ((mlx4_dev[i].pci_addr.domain == pci_addr->domain) &&
4302 (mlx4_dev[i].pci_addr.bus == pci_addr->bus) &&
4303 (mlx4_dev[i].pci_addr.devid == pci_addr->devid) &&
4304 (mlx4_dev[i].pci_addr.function == pci_addr->function))
4306 if ((mlx4_dev[i].ports == 0) && (ret == -1))
4313 * Retrieve integer value from environment variable.
4316 * Environment variable name.
4319 * Integer value, 0 if the variable is not set.
4322 mlx4_getenv_int(const char *name)
4324 const char *val = getenv(name);
4331 static struct eth_driver mlx4_driver;
4334 * DPDK callback to register a PCI device.
4336 * This function creates an Ethernet device for each port of a given
4339 * @param[in] pci_drv
4340 * PCI driver structure (mlx4_driver).
4341 * @param[in] pci_dev
4342 * PCI device information.
4345 * 0 on success, negative errno value on failure.
4348 mlx4_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
4350 struct ibv_device **list;
4351 struct ibv_device *ibv_dev;
4353 struct ibv_context *attr_ctx = NULL;
4354 struct ibv_device_attr device_attr;
4360 assert(pci_drv == &mlx4_driver.pci_drv);
4361 /* Get mlx4_dev[] index. */
4362 idx = mlx4_dev_idx(&pci_dev->addr);
4364 ERROR("this driver cannot support any more adapters");
4367 DEBUG("using driver device index %d", idx);
4369 /* Save PCI address. */
4370 mlx4_dev[idx].pci_addr = pci_dev->addr;
4371 list = ibv_get_device_list(&i);
4374 if (errno == ENOSYS) {
4375 WARN("cannot list devices, is ib_uverbs loaded?");
4382 * For each listed device, check related sysfs entry against
4383 * the provided PCI ID.
4386 struct rte_pci_addr pci_addr;
4389 DEBUG("checking device \"%s\"", list[i]->name);
4390 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
4392 if ((pci_dev->addr.domain != pci_addr.domain) ||
4393 (pci_dev->addr.bus != pci_addr.bus) ||
4394 (pci_dev->addr.devid != pci_addr.devid) ||
4395 (pci_dev->addr.function != pci_addr.function))
4397 vf = (pci_dev->id.device_id ==
4398 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
4399 INFO("PCI information matches, using device \"%s\" (VF: %s)",
4400 list[i]->name, (vf ? "true" : "false"));
4401 attr_ctx = ibv_open_device(list[i]);
4405 if (attr_ctx == NULL) {
4406 ibv_free_device_list(list);
4409 WARN("cannot access device, is mlx4_ib loaded?");
4412 WARN("cannot use device, are drivers up to date?");
4420 DEBUG("device opened");
4421 if (ibv_query_device(attr_ctx, &device_attr))
4423 INFO("%u port(s) detected", device_attr.phys_port_cnt);
4425 for (i = 0; i < device_attr.phys_port_cnt; i++) {
4426 uint32_t port = i + 1; /* ports are indexed from one */
4427 uint32_t test = (1 << i);
4428 struct ibv_context *ctx = NULL;
4429 struct ibv_port_attr port_attr;
4430 struct ibv_pd *pd = NULL;
4431 struct priv *priv = NULL;
4432 struct rte_eth_dev *eth_dev;
4433 #if defined(INLINE_RECV) || defined(RSS_SUPPORT)
4434 struct ibv_exp_device_attr exp_device_attr;
4436 struct ether_addr mac;
4437 union ibv_gid temp_gid;
4440 exp_device_attr.comp_mask =
4441 (IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
4442 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ);
4443 #endif /* RSS_SUPPORT */
4445 DEBUG("using port %u (%08" PRIx32 ")", port, test);
4447 ctx = ibv_open_device(ibv_dev);
4451 /* Check port status. */
4452 err = ibv_query_port(ctx, port, &port_attr);
4454 ERROR("port query failed: %s", strerror(err));
4457 if (port_attr.state != IBV_PORT_ACTIVE)
4458 WARN("bad state for port %d: \"%s\" (%d)",
4459 port, ibv_port_state_str(port_attr.state),
4462 /* Allocate protection domain. */
4463 pd = ibv_alloc_pd(ctx);
4465 ERROR("PD allocation failure");
4470 mlx4_dev[idx].ports |= test;
4472 /* from rte_ethdev.c */
4473 priv = rte_zmalloc("ethdev private structure",
4475 RTE_CACHE_LINE_SIZE);
4477 ERROR("priv allocation failure");
4483 priv->device_attr = device_attr;
4484 priv->port_attr = port_attr;
4487 priv->mtu = ETHER_MTU;
4489 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
4490 INFO("experimental ibv_exp_query_device");
4493 if ((exp_device_attr.exp_device_cap_flags &
4494 IBV_EXP_DEVICE_QPG) &&
4495 (exp_device_attr.exp_device_cap_flags &
4496 IBV_EXP_DEVICE_UD_RSS) &&
4497 (exp_device_attr.comp_mask &
4498 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ) &&
4499 (exp_device_attr.max_rss_tbl_sz > 0)) {
4502 priv->max_rss_tbl_sz = exp_device_attr.max_rss_tbl_sz;
4506 priv->max_rss_tbl_sz = 0;
4508 priv->hw_tss = !!(exp_device_attr.exp_device_cap_flags &
4509 IBV_EXP_DEVICE_UD_TSS);
4510 DEBUG("device flags: %s%s%s",
4511 (priv->hw_qpg ? "IBV_DEVICE_QPG " : ""),
4512 (priv->hw_tss ? "IBV_DEVICE_TSS " : ""),
4513 (priv->hw_rss ? "IBV_DEVICE_RSS " : ""));
4515 DEBUG("maximum RSS indirection table size: %u",
4516 exp_device_attr.max_rss_tbl_sz);
4517 #endif /* RSS_SUPPORT */
4520 priv->inl_recv_size = mlx4_getenv_int("MLX4_INLINE_RECV_SIZE");
4522 if (priv->inl_recv_size) {
4523 exp_device_attr.comp_mask =
4524 IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ;
4525 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
4526 INFO("Couldn't query device for inline-receive"
4528 priv->inl_recv_size = 0;
4530 if ((unsigned)exp_device_attr.inline_recv_sz <
4531 priv->inl_recv_size) {
4532 INFO("Max inline-receive (%d) <"
4533 " requested inline-receive (%u)",
4534 exp_device_attr.inline_recv_sz,
4535 priv->inl_recv_size);
4536 priv->inl_recv_size =
4537 exp_device_attr.inline_recv_sz;
4540 INFO("Set inline receive size to %u",
4541 priv->inl_recv_size);
4543 #endif /* INLINE_RECV */
4545 (void)mlx4_getenv_int;
4547 if (ibv_query_gid(ctx, port, 0, &temp_gid)) {
4548 ERROR("ibv_query_gid() failure");
4551 /* Configure the first MAC address by default. */
4552 mac_from_gid(&mac.addr_bytes, port, temp_gid.raw);
4553 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
4555 mac.addr_bytes[0], mac.addr_bytes[1],
4556 mac.addr_bytes[2], mac.addr_bytes[3],
4557 mac.addr_bytes[4], mac.addr_bytes[5]);
4558 /* Register MAC and broadcast addresses. */
4559 claim_zero(priv_mac_addr_add(priv, 0,
4560 (const uint8_t (*)[ETHER_ADDR_LEN])
4562 claim_zero(priv_mac_addr_add(priv, 1,
4563 &(const uint8_t [ETHER_ADDR_LEN])
4564 { "\xff\xff\xff\xff\xff\xff" }));
4567 char ifname[IF_NAMESIZE];
4569 if (priv_get_ifname(priv, &ifname) == 0)
4570 DEBUG("port %u ifname is \"%s\"",
4571 priv->port, ifname);
4573 DEBUG("port %u ifname is unknown", priv->port);
4576 /* Get actual MTU if possible. */
4577 priv_get_mtu(priv, &priv->mtu);
4578 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
4580 /* from rte_ethdev.c */
4582 char name[RTE_ETH_NAME_MAX_LEN];
4584 snprintf(name, sizeof(name), "%s port %u",
4585 ibv_get_device_name(ibv_dev), port);
4586 eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
4588 if (eth_dev == NULL) {
4589 ERROR("can not allocate rte ethdev");
4594 eth_dev->data->dev_private = priv;
4595 eth_dev->pci_dev = pci_dev;
4596 eth_dev->driver = &mlx4_driver;
4597 eth_dev->data->rx_mbuf_alloc_failed = 0;
4598 eth_dev->data->mtu = ETHER_MTU;
4600 priv->dev = eth_dev;
4601 eth_dev->dev_ops = &mlx4_dev_ops;
4602 eth_dev->data->mac_addrs = priv->mac;
4604 /* Bring Ethernet device up. */
4605 DEBUG("forcing Ethernet interface up");
4606 priv_set_flags(priv, ~IFF_UP, IFF_UP);
4612 claim_zero(ibv_dealloc_pd(pd));
4614 claim_zero(ibv_close_device(ctx));
4619 * XXX if something went wrong in the loop above, there is a resource
4620 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
4621 * long as the dpdk does not provide a way to deallocate a ethdev and a
4622 * way to enumerate the registered ethdevs to free the previous ones.
4625 /* no port found, complain */
4626 if (!mlx4_dev[idx].ports) {
4633 claim_zero(ibv_close_device(attr_ctx));
4635 ibv_free_device_list(list);
4640 static const struct rte_pci_id mlx4_pci_id_map[] = {
4642 .vendor_id = PCI_VENDOR_ID_MELLANOX,
4643 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3,
4644 .subsystem_vendor_id = PCI_ANY_ID,
4645 .subsystem_device_id = PCI_ANY_ID
4648 .vendor_id = PCI_VENDOR_ID_MELLANOX,
4649 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO,
4650 .subsystem_vendor_id = PCI_ANY_ID,
4651 .subsystem_device_id = PCI_ANY_ID
4654 .vendor_id = PCI_VENDOR_ID_MELLANOX,
4655 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3VF,
4656 .subsystem_vendor_id = PCI_ANY_ID,
4657 .subsystem_device_id = PCI_ANY_ID
4664 static struct eth_driver mlx4_driver = {
4666 .name = MLX4_DRIVER_NAME,
4667 .id_table = mlx4_pci_id_map,
4668 .devinit = mlx4_pci_devinit,
4670 .dev_private_size = sizeof(struct priv)
4674 * Driver initialization routine.
4677 rte_mlx4_pmd_init(const char *name, const char *args)
4681 rte_eal_pci_register(&mlx4_driver.pci_drv);
4685 static struct rte_driver rte_mlx4_driver = {
4687 .name = MLX4_DRIVER_NAME,
4688 .init = rte_mlx4_pmd_init,
4691 PMD_REGISTER_DRIVER(rte_mlx4_driver)