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 *)(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 (elt->sges[0].addr - WR_ID(wr->wr_id).offset);
1072 /* Faster than rte_pktmbuf_free(). */
1074 struct rte_mbuf *next = NEXT(tmp);
1076 rte_pktmbuf_free_seg(tmp);
1078 } while (tmp != NULL);
1082 WR_ID(wr->wr_id).offset = 0;
1083 for (j = 0; ((int)j < wr->num_sge); ++j) {
1084 elt->sges[j].addr = 0;
1085 elt->sges[j].length = 0;
1086 elt->sges[j].lkey = 0;
1091 /* Sanity checks, most of which are only relevant with
1092 * debugging enabled. */
1093 assert(WR_ID(wr->wr_id).id == elts_head);
1094 assert(WR_ID(wr->wr_id).offset == 0);
1095 assert(wr->next == NULL);
1096 assert(wr->sg_list == &elt->sges[0]);
1097 assert(wr->num_sge == 0);
1098 assert(wr->opcode == IBV_WR_SEND);
1099 /* When there are too many segments, extra segments are
1100 * linearized in the last SGE. */
1101 if (unlikely(segs > elemof(elt->sges))) {
1102 segs = (elemof(elt->sges) - 1);
1105 /* Set WR fields. */
1106 assert(((uintptr_t)rte_pktmbuf_mtod(buf, char *) -
1107 (uintptr_t)buf) <= 0xffff);
1108 WR_ID(wr->wr_id).offset =
1109 ((uintptr_t)rte_pktmbuf_mtod(buf, char *) -
1112 /* Register segments as SGEs. */
1113 for (j = 0; (j != segs); ++j) {
1114 struct ibv_sge *sge = &elt->sges[j];
1117 /* Retrieve Memory Region key for this memory pool. */
1118 lkey = txq_mp2mr(txq, buf->pool);
1119 if (unlikely(lkey == (uint32_t)-1)) {
1120 /* MR does not exist. */
1121 DEBUG("%p: unable to get MP <-> MR"
1122 " association", (void *)txq);
1123 /* Clean up TX element. */
1124 WR_ID(elt->wr.wr_id).offset = 0;
1138 /* Sanity checks, only relevant with debugging
1140 assert(sge->addr == 0);
1141 assert(sge->length == 0);
1142 assert(sge->lkey == 0);
1144 sge->addr = (uintptr_t)rte_pktmbuf_mtod(buf, char *);
1146 rte_prefetch0((volatile void *)sge->addr);
1147 sge->length = DATA_LEN(buf);
1149 #if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
1150 sent_size += sge->length;
1154 /* If buf is not NULL here and is not going to be linearized,
1155 * nb_segs is not valid. */
1157 assert((buf == NULL) || (linearize));
1158 /* Linearize extra segments. */
1160 struct ibv_sge *sge = &elt->sges[segs];
1161 linear_t *linear = &(*txq->elts_linear)[elts_head];
1162 unsigned int size = linearize_mbuf(linear, buf);
1164 assert(segs == (elemof(elt->sges) - 1));
1166 /* Invalid packet. */
1167 DEBUG("%p: packet too large to be linearized.",
1169 /* Clean up TX element. */
1170 WR_ID(elt->wr.wr_id).offset = 0;
1184 /* If MLX4_PMD_SGE_WR_N is 1, free mbuf immediately
1185 * and clear offset from WR ID. */
1186 if (elemof(elt->sges) == 1) {
1188 struct rte_mbuf *next = NEXT(buf);
1190 rte_pktmbuf_free_seg(buf);
1192 } while (buf != NULL);
1193 WR_ID(wr->wr_id).offset = 0;
1195 /* Set WR fields and fill SGE with linear buffer. */
1197 /* Sanity checks, only relevant with debugging
1199 assert(sge->addr == 0);
1200 assert(sge->length == 0);
1201 assert(sge->lkey == 0);
1203 sge->addr = (uintptr_t)&(*linear)[0];
1205 sge->lkey = txq->mr_linear->lkey;
1206 #if (MLX4_PMD_MAX_INLINE > 0) || defined(MLX4_PMD_SOFT_COUNTERS)
1210 /* Link WRs together for ibv_post_send(). */
1212 wr_next = &wr->next;
1213 #if MLX4_PMD_MAX_INLINE > 0
1214 if (sent_size <= txq->max_inline)
1215 wr->send_flags = IBV_SEND_INLINE;
1219 if (++elts_head >= elts_n)
1221 #ifdef MLX4_PMD_SOFT_COUNTERS
1222 /* Increment sent bytes counter. */
1223 txq->stats.obytes += sent_size;
1227 /* Take a shortcut if nothing must be sent. */
1228 if (unlikely(i == 0))
1230 #ifdef MLX4_PMD_SOFT_COUNTERS
1231 /* Increment sent packets counter. */
1232 txq->stats.opackets += i;
1235 /* The last WR is the only one asking for a completion event. */
1236 containerof(wr_next, mlx4_send_wr_t, next)->
1237 send_flags |= IBV_SEND_SIGNALED;
1238 err = mlx4_post_send(txq->qp, head.next, &bad_wr);
1239 if (unlikely(err)) {
1240 unsigned int unsent = 0;
1242 /* An error occurred, completion event is lost. Fix counters. */
1243 while (bad_wr != NULL) {
1244 struct txq_elt *elt =
1245 containerof(bad_wr, struct txq_elt, wr);
1246 mlx4_send_wr_t *wr = &elt->wr;
1247 mlx4_send_wr_t *next = wr->next;
1248 #if defined(MLX4_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1252 assert(wr == bad_wr);
1253 /* Clean up TX element without freeing it, caller
1254 * should take care of this. */
1255 WR_ID(elt->wr.wr_id).offset = 0;
1256 #ifdef MLX4_PMD_SOFT_COUNTERS
1257 for (j = 0; ((int)j < wr->num_sge); ++j)
1258 txq->stats.obytes -= wr->sg_list[j].length;
1263 for (j = 0; ((int)j < wr->num_sge); ++j) {
1264 elt->sges[j].addr = 0;
1265 elt->sges[j].length = 0;
1266 elt->sges[j].lkey = 0;
1273 #ifdef MLX4_PMD_SOFT_COUNTERS
1274 txq->stats.opackets -= unsent;
1276 assert(i >= unsent);
1278 /* "Unsend" remaining packets. */
1279 elts_head -= unsent;
1280 if (elts_head >= elts_n)
1281 elts_head += elts_n;
1282 assert(elts_head < elts_n);
1283 DEBUG("%p: mlx4_post_send() failed, %u unprocessed WRs: %s",
1284 (void *)txq, unsent,
1285 ((err <= -1) ? "Internal error" : strerror(err)));
1288 txq->elts_head = elts_head;
1293 * Configure a TX queue.
1296 * Pointer to Ethernet device structure.
1298 * Pointer to TX queue structure.
1300 * Number of descriptors to configure in queue.
1302 * NUMA socket on which memory must be allocated.
1304 * Thresholds parameters.
1307 * 0 on success, errno value on failure.
1310 txq_setup(struct rte_eth_dev *dev, struct txq *txq, uint16_t desc,
1311 unsigned int socket, const struct rte_eth_txconf *conf)
1313 struct priv *priv = dev->data->dev_private;
1319 struct ibv_qp_init_attr init;
1320 struct ibv_exp_qp_attr mod;
1324 (void)conf; /* Thresholds configuration (ignored). */
1325 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
1326 ERROR("%p: invalid number of TX descriptors (must be a"
1327 " multiple of %d)", (void *)dev, desc);
1330 desc /= MLX4_PMD_SGE_WR_N;
1331 /* MRs will be registered in mp2mr[] later. */
1332 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
1333 if (tmpl.cq == NULL) {
1335 ERROR("%p: CQ creation failure: %s",
1336 (void *)dev, strerror(ret));
1339 DEBUG("priv->device_attr.max_qp_wr is %d",
1340 priv->device_attr.max_qp_wr);
1341 DEBUG("priv->device_attr.max_sge is %d",
1342 priv->device_attr.max_sge);
1343 attr.init = (struct ibv_qp_init_attr){
1344 /* CQ to be associated with the send queue. */
1346 /* CQ to be associated with the receive queue. */
1349 /* Max number of outstanding WRs. */
1350 .max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
1351 priv->device_attr.max_qp_wr :
1353 /* Max number of scatter/gather elements in a WR. */
1354 .max_send_sge = ((priv->device_attr.max_sge <
1355 MLX4_PMD_SGE_WR_N) ?
1356 priv->device_attr.max_sge :
1358 #if MLX4_PMD_MAX_INLINE > 0
1359 .max_inline_data = MLX4_PMD_MAX_INLINE,
1362 .qp_type = IBV_QPT_RAW_PACKET,
1363 /* Do *NOT* enable this, completions events are managed per
1367 tmpl.qp = ibv_create_qp(priv->pd, &attr.init);
1368 if (tmpl.qp == NULL) {
1369 ret = (errno ? errno : EINVAL);
1370 ERROR("%p: QP creation failure: %s",
1371 (void *)dev, strerror(ret));
1374 #if MLX4_PMD_MAX_INLINE > 0
1375 /* ibv_create_qp() updates this value. */
1376 tmpl.max_inline = attr.init.cap.max_inline_data;
1378 attr.mod = (struct ibv_exp_qp_attr){
1379 /* Move the QP to this state. */
1380 .qp_state = IBV_QPS_INIT,
1381 /* Primary port number. */
1382 .port_num = priv->port
1384 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod,
1385 (IBV_EXP_QP_STATE | IBV_EXP_QP_PORT));
1387 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
1388 (void *)dev, strerror(ret));
1391 ret = txq_alloc_elts(&tmpl, desc);
1393 ERROR("%p: TXQ allocation failed: %s",
1394 (void *)dev, strerror(ret));
1397 attr.mod = (struct ibv_exp_qp_attr){
1398 .qp_state = IBV_QPS_RTR
1400 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1402 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
1403 (void *)dev, strerror(ret));
1406 attr.mod.qp_state = IBV_QPS_RTS;
1407 ret = ibv_exp_modify_qp(tmpl.qp, &attr.mod, IBV_EXP_QP_STATE);
1409 ERROR("%p: QP state to IBV_QPS_RTS failed: %s",
1410 (void *)dev, strerror(ret));
1413 /* Clean up txq in case we're reinitializing it. */
1414 DEBUG("%p: cleaning-up old txq just in case", (void *)txq);
1417 DEBUG("%p: txq updated with %p", (void *)txq, (void *)&tmpl);
1427 * DPDK callback to configure a TX queue.
1430 * Pointer to Ethernet device structure.
1434 * Number of descriptors to configure in queue.
1436 * NUMA socket on which memory must be allocated.
1438 * Thresholds parameters.
1441 * 0 on success, negative errno value on failure.
1444 mlx4_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
1445 unsigned int socket, const struct rte_eth_txconf *conf)
1447 struct priv *priv = dev->data->dev_private;
1448 struct txq *txq = (*priv->txqs)[idx];
1452 DEBUG("%p: configuring queue %u for %u descriptors",
1453 (void *)dev, idx, desc);
1454 if (idx >= priv->txqs_n) {
1455 ERROR("%p: queue index out of range (%u >= %u)",
1456 (void *)dev, idx, priv->txqs_n);
1461 DEBUG("%p: reusing already allocated queue index %u (%p)",
1462 (void *)dev, idx, (void *)txq);
1463 if (priv->started) {
1467 (*priv->txqs)[idx] = NULL;
1470 txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0, socket);
1472 ERROR("%p: unable to allocate queue index %u",
1478 ret = txq_setup(dev, txq, desc, socket, conf);
1482 txq->stats.idx = idx;
1483 DEBUG("%p: adding TX queue %p to list",
1484 (void *)dev, (void *)txq);
1485 (*priv->txqs)[idx] = txq;
1486 /* Update send callback. */
1487 dev->tx_pkt_burst = mlx4_tx_burst;
1494 * DPDK callback to release a TX queue.
1497 * Generic TX queue pointer.
1500 mlx4_tx_queue_release(void *dpdk_txq)
1502 struct txq *txq = (struct txq *)dpdk_txq;
1510 for (i = 0; (i != priv->txqs_n); ++i)
1511 if ((*priv->txqs)[i] == txq) {
1512 DEBUG("%p: removing TX queue %p from list",
1513 (void *)priv->dev, (void *)txq);
1514 (*priv->txqs)[i] = NULL;
1522 /* RX queues handling. */
1525 * Allocate RX queue elements with scattered packets support.
1528 * Pointer to RX queue structure.
1530 * Number of elements to allocate.
1532 * If not NULL, fetch buffers from this array instead of allocating them
1533 * with rte_pktmbuf_alloc().
1536 * 0 on success, errno value on failure.
1539 rxq_alloc_elts_sp(struct rxq *rxq, unsigned int elts_n,
1540 struct rte_mbuf **pool)
1543 struct rxq_elt_sp (*elts)[elts_n] =
1544 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1549 ERROR("%p: can't allocate packets array", (void *)rxq);
1553 /* For each WR (packet). */
1554 for (i = 0; (i != elts_n); ++i) {
1556 struct rxq_elt_sp *elt = &(*elts)[i];
1557 struct ibv_recv_wr *wr = &elt->wr;
1558 struct ibv_sge (*sges)[(elemof(elt->sges))] = &elt->sges;
1560 /* These two arrays must have the same size. */
1561 assert(elemof(elt->sges) == elemof(elt->bufs));
1564 wr->next = &(*elts)[(i + 1)].wr;
1565 wr->sg_list = &(*sges)[0];
1566 wr->num_sge = elemof(*sges);
1567 /* For each SGE (segment). */
1568 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1569 struct ibv_sge *sge = &(*sges)[j];
1570 struct rte_mbuf *buf;
1574 assert(buf != NULL);
1575 rte_pktmbuf_reset(buf);
1577 buf = rte_pktmbuf_alloc(rxq->mp);
1579 assert(pool == NULL);
1580 ERROR("%p: empty mbuf pool", (void *)rxq);
1585 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1586 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1587 /* Buffer is supposed to be empty. */
1588 assert(rte_pktmbuf_data_len(buf) == 0);
1589 assert(rte_pktmbuf_pkt_len(buf) == 0);
1590 /* sge->addr must be able to store a pointer. */
1591 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1593 /* The first SGE keeps its headroom. */
1594 sge->addr = (uintptr_t)rte_pktmbuf_mtod(buf,
1596 sge->length = (buf->buf_len -
1597 RTE_PKTMBUF_HEADROOM);
1599 /* Subsequent SGEs lose theirs. */
1600 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1601 SET_DATA_OFF(buf, 0);
1602 sge->addr = (uintptr_t)buf->buf_addr;
1603 sge->length = buf->buf_len;
1605 sge->lkey = rxq->mr->lkey;
1606 /* Redundant check for tailroom. */
1607 assert(sge->length == rte_pktmbuf_tailroom(buf));
1610 /* The last WR pointer must be NULL. */
1611 (*elts)[(i - 1)].wr.next = NULL;
1612 DEBUG("%p: allocated and configured %u WRs (%zu segments)",
1613 (void *)rxq, elts_n, (elts_n * elemof((*elts)[0].sges)));
1614 rxq->elts_n = elts_n;
1615 rxq->elts.sp = elts;
1620 assert(pool == NULL);
1621 for (i = 0; (i != elemof(*elts)); ++i) {
1623 struct rxq_elt_sp *elt = &(*elts)[i];
1625 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1626 struct rte_mbuf *buf = elt->bufs[j];
1629 rte_pktmbuf_free_seg(buf);
1634 DEBUG("%p: failed, freed everything", (void *)rxq);
1640 * Free RX queue elements with scattered packets support.
1643 * Pointer to RX queue structure.
1646 rxq_free_elts_sp(struct rxq *rxq)
1649 unsigned int elts_n = rxq->elts_n;
1650 struct rxq_elt_sp (*elts)[elts_n] = rxq->elts.sp;
1652 DEBUG("%p: freeing WRs", (void *)rxq);
1654 rxq->elts.sp = NULL;
1657 for (i = 0; (i != elemof(*elts)); ++i) {
1659 struct rxq_elt_sp *elt = &(*elts)[i];
1661 for (j = 0; (j != elemof(elt->bufs)); ++j) {
1662 struct rte_mbuf *buf = elt->bufs[j];
1665 rte_pktmbuf_free_seg(buf);
1672 * Allocate RX queue elements.
1675 * Pointer to RX queue structure.
1677 * Number of elements to allocate.
1679 * If not NULL, fetch buffers from this array instead of allocating them
1680 * with rte_pktmbuf_alloc().
1683 * 0 on success, errno value on failure.
1686 rxq_alloc_elts(struct rxq *rxq, unsigned int elts_n, struct rte_mbuf **pool)
1689 struct rxq_elt (*elts)[elts_n] =
1690 rte_calloc_socket("RXQ elements", 1, sizeof(*elts), 0,
1695 ERROR("%p: can't allocate packets array", (void *)rxq);
1699 /* For each WR (packet). */
1700 for (i = 0; (i != elts_n); ++i) {
1701 struct rxq_elt *elt = &(*elts)[i];
1702 struct ibv_recv_wr *wr = &elt->wr;
1703 struct ibv_sge *sge = &(*elts)[i].sge;
1704 struct rte_mbuf *buf;
1708 assert(buf != NULL);
1709 rte_pktmbuf_reset(buf);
1711 buf = rte_pktmbuf_alloc(rxq->mp);
1713 assert(pool == NULL);
1714 ERROR("%p: empty mbuf pool", (void *)rxq);
1718 /* Configure WR. Work request ID contains its own index in
1719 * the elts array and the offset between SGE buffer header and
1721 WR_ID(wr->wr_id).id = i;
1722 WR_ID(wr->wr_id).offset =
1723 (((uintptr_t)buf->buf_addr + RTE_PKTMBUF_HEADROOM) -
1725 wr->next = &(*elts)[(i + 1)].wr;
1728 /* Headroom is reserved by rte_pktmbuf_alloc(). */
1729 assert(DATA_OFF(buf) == RTE_PKTMBUF_HEADROOM);
1730 /* Buffer is supposed to be empty. */
1731 assert(rte_pktmbuf_data_len(buf) == 0);
1732 assert(rte_pktmbuf_pkt_len(buf) == 0);
1733 /* sge->addr must be able to store a pointer. */
1734 assert(sizeof(sge->addr) >= sizeof(uintptr_t));
1735 /* SGE keeps its headroom. */
1736 sge->addr = (uintptr_t)
1737 ((uint8_t *)buf->buf_addr + RTE_PKTMBUF_HEADROOM);
1738 sge->length = (buf->buf_len - RTE_PKTMBUF_HEADROOM);
1739 sge->lkey = rxq->mr->lkey;
1740 /* Redundant check for tailroom. */
1741 assert(sge->length == rte_pktmbuf_tailroom(buf));
1742 /* Make sure elts index and SGE mbuf pointer can be deduced
1744 if ((WR_ID(wr->wr_id).id != i) ||
1745 ((void *)(sge->addr - WR_ID(wr->wr_id).offset) != buf)) {
1746 ERROR("%p: cannot store index and offset in WR ID",
1749 rte_pktmbuf_free(buf);
1754 /* The last WR pointer must be NULL. */
1755 (*elts)[(i - 1)].wr.next = NULL;
1756 DEBUG("%p: allocated and configured %u single-segment WRs",
1757 (void *)rxq, elts_n);
1758 rxq->elts_n = elts_n;
1759 rxq->elts.no_sp = elts;
1764 assert(pool == NULL);
1765 for (i = 0; (i != elemof(*elts)); ++i) {
1766 struct rxq_elt *elt = &(*elts)[i];
1767 struct rte_mbuf *buf;
1769 if (elt->sge.addr == 0)
1771 assert(WR_ID(elt->wr.wr_id).id == i);
1773 (elt->sge.addr - WR_ID(elt->wr.wr_id).offset);
1774 rte_pktmbuf_free_seg(buf);
1778 DEBUG("%p: failed, freed everything", (void *)rxq);
1784 * Free RX queue elements.
1787 * Pointer to RX queue structure.
1790 rxq_free_elts(struct rxq *rxq)
1793 unsigned int elts_n = rxq->elts_n;
1794 struct rxq_elt (*elts)[elts_n] = rxq->elts.no_sp;
1796 DEBUG("%p: freeing WRs", (void *)rxq);
1798 rxq->elts.no_sp = NULL;
1801 for (i = 0; (i != elemof(*elts)); ++i) {
1802 struct rxq_elt *elt = &(*elts)[i];
1803 struct rte_mbuf *buf;
1805 if (elt->sge.addr == 0)
1807 assert(WR_ID(elt->wr.wr_id).id == i);
1808 buf = (void *)(elt->sge.addr - WR_ID(elt->wr.wr_id).offset);
1809 rte_pktmbuf_free_seg(buf);
1815 * Unregister a MAC address from a RX queue.
1818 * Pointer to RX queue structure.
1820 * MAC address index.
1823 rxq_mac_addr_del(struct rxq *rxq, unsigned int mac_index)
1826 struct priv *priv = rxq->priv;
1827 const uint8_t (*mac)[ETHER_ADDR_LEN] =
1828 (const uint8_t (*)[ETHER_ADDR_LEN])
1829 priv->mac[mac_index].addr_bytes;
1832 assert(mac_index < elemof(priv->mac));
1833 if (!BITFIELD_ISSET(rxq->mac_configured, mac_index)) {
1834 assert(rxq->mac_flow[mac_index] == NULL);
1837 DEBUG("%p: removing MAC address %02x:%02x:%02x:%02x:%02x:%02x"
1840 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
1842 assert(rxq->mac_flow[mac_index] != NULL);
1843 claim_zero(ibv_exp_destroy_flow(rxq->mac_flow[mac_index]));
1844 rxq->mac_flow[mac_index] = NULL;
1845 BITFIELD_RESET(rxq->mac_configured, mac_index);
1849 * Unregister all MAC addresses from a RX queue.
1852 * Pointer to RX queue structure.
1855 rxq_mac_addrs_del(struct rxq *rxq)
1857 struct priv *priv = rxq->priv;
1860 for (i = 0; (i != elemof(priv->mac)); ++i)
1861 rxq_mac_addr_del(rxq, i);
1864 static int rxq_promiscuous_enable(struct rxq *);
1865 static void rxq_promiscuous_disable(struct rxq *);
1868 * Register a MAC address in a RX queue.
1871 * Pointer to RX queue structure.
1873 * MAC address index to register.
1876 * 0 on success, errno value on failure.
1879 rxq_mac_addr_add(struct rxq *rxq, unsigned int mac_index)
1881 struct priv *priv = rxq->priv;
1882 const uint8_t (*mac)[ETHER_ADDR_LEN] =
1883 (const uint8_t (*)[ETHER_ADDR_LEN])
1884 priv->mac[mac_index].addr_bytes;
1885 unsigned int vlans = 0;
1886 unsigned int specs = 0;
1888 struct ibv_exp_flow *flow;
1890 assert(mac_index < elemof(priv->mac));
1891 if (BITFIELD_ISSET(rxq->mac_configured, mac_index))
1892 rxq_mac_addr_del(rxq, mac_index);
1893 /* Number of configured VLANs. */
1894 for (i = 0; (i != elemof(priv->vlan_filter)); ++i)
1895 if (priv->vlan_filter[i].enabled)
1897 specs = (vlans ? vlans : 1);
1899 /* Allocate flow specification on the stack. */
1900 struct ibv_exp_flow_attr data
1902 (sizeof(struct ibv_exp_flow_spec_eth[specs]) /
1903 sizeof(struct ibv_exp_flow_attr)) +
1904 !!(sizeof(struct ibv_exp_flow_spec_eth[specs]) %
1905 sizeof(struct ibv_exp_flow_attr))];
1906 struct ibv_exp_flow_attr *attr = (void *)&data[0];
1907 struct ibv_exp_flow_spec_eth *spec = (void *)&data[1];
1910 * No padding must be inserted by the compiler between attr and spec.
1911 * This layout is expected by libibverbs.
1913 assert(((uint8_t *)attr + sizeof(*attr)) == (uint8_t *)spec);
1914 *attr = (struct ibv_exp_flow_attr){
1915 .type = IBV_EXP_FLOW_ATTR_NORMAL,
1916 .num_of_specs = specs,
1920 *spec = (struct ibv_exp_flow_spec_eth){
1921 .type = IBV_EXP_FLOW_SPEC_ETH,
1922 .size = sizeof(*spec),
1925 (*mac)[0], (*mac)[1], (*mac)[2],
1926 (*mac)[3], (*mac)[4], (*mac)[5]
1930 .dst_mac = "\xff\xff\xff\xff\xff\xff",
1931 .vlan_tag = (vlans ? htons(0xfff) : 0)
1934 /* Fill VLAN specifications. */
1935 for (i = 0, j = 0; (i != elemof(priv->vlan_filter)); ++i) {
1936 if (!priv->vlan_filter[i].enabled)
1941 spec[j].val.vlan_tag = htons(priv->vlan_filter[i].id);
1944 DEBUG("%p: adding MAC address %02x:%02x:%02x:%02x:%02x:%02x index %u"
1945 " (%u VLAN(s) configured)",
1947 (*mac)[0], (*mac)[1], (*mac)[2], (*mac)[3], (*mac)[4], (*mac)[5],
1950 /* Create related flow. */
1952 flow = ibv_exp_create_flow(rxq->qp, attr);
1956 /* Flow creation failure is not fatal when in DMFS A0 mode.
1957 * Ignore error if promiscuity is already enabled or can be
1959 if (priv->promisc_ok)
1961 if ((rxq->promisc_flow != NULL) ||
1962 (rxq_promiscuous_enable(rxq) == 0)) {
1963 if (rxq->promisc_flow != NULL)
1964 rxq_promiscuous_disable(rxq);
1965 WARN("cannot configure normal flow but promiscuous"
1966 " mode is fine, assuming promiscuous optimization"
1968 " (options mlx4_core log_num_mgm_entry_size=-7)");
1969 priv->promisc_ok = 1;
1973 /* It's not clear whether errno is always set in this case. */
1974 ERROR("%p: flow configuration failed, errno=%d: %s",
1976 (errno ? strerror(errno) : "Unknown error"));
1981 assert(rxq->mac_flow[mac_index] == NULL);
1982 rxq->mac_flow[mac_index] = flow;
1983 BITFIELD_SET(rxq->mac_configured, mac_index);
1988 * Register all MAC addresses in a RX queue.
1991 * Pointer to RX queue structure.
1994 * 0 on success, errno value on failure.
1997 rxq_mac_addrs_add(struct rxq *rxq)
1999 struct priv *priv = rxq->priv;
2003 for (i = 0; (i != elemof(priv->mac)); ++i) {
2004 if (!BITFIELD_ISSET(priv->mac_configured, i))
2006 ret = rxq_mac_addr_add(rxq, i);
2009 /* Failure, rollback. */
2011 rxq_mac_addr_del(rxq, --i);
2019 * Unregister a MAC address.
2021 * In RSS mode, the MAC address is unregistered from the parent queue,
2022 * otherwise it is unregistered from each queue directly.
2025 * Pointer to private structure.
2027 * MAC address index.
2030 priv_mac_addr_del(struct priv *priv, unsigned int mac_index)
2034 assert(mac_index < elemof(priv->mac));
2035 if (!BITFIELD_ISSET(priv->mac_configured, mac_index))
2038 rxq_mac_addr_del(&priv->rxq_parent, mac_index);
2041 for (i = 0; (i != priv->dev->data->nb_rx_queues); ++i)
2042 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2044 BITFIELD_RESET(priv->mac_configured, mac_index);
2048 * Register a MAC address.
2050 * In RSS mode, the MAC address is registered in the parent queue,
2051 * otherwise it is registered in each queue directly.
2054 * Pointer to private structure.
2056 * MAC address index to use.
2058 * MAC address to register.
2061 * 0 on success, errno value on failure.
2064 priv_mac_addr_add(struct priv *priv, unsigned int mac_index,
2065 const uint8_t (*mac)[ETHER_ADDR_LEN])
2070 assert(mac_index < elemof(priv->mac));
2071 /* First, make sure this address isn't already configured. */
2072 for (i = 0; (i != elemof(priv->mac)); ++i) {
2073 /* Skip this index, it's going to be reconfigured. */
2076 if (!BITFIELD_ISSET(priv->mac_configured, i))
2078 if (memcmp(priv->mac[i].addr_bytes, *mac, sizeof(*mac)))
2080 /* Address already configured elsewhere, return with error. */
2083 if (BITFIELD_ISSET(priv->mac_configured, mac_index))
2084 priv_mac_addr_del(priv, mac_index);
2085 priv->mac[mac_index] = (struct ether_addr){
2087 (*mac)[0], (*mac)[1], (*mac)[2],
2088 (*mac)[3], (*mac)[4], (*mac)[5]
2091 /* If device isn't started, this is all we need to do. */
2092 if (!priv->started) {
2094 /* Verify that all queues have this index disabled. */
2095 for (i = 0; (i != priv->rxqs_n); ++i) {
2096 if ((*priv->rxqs)[i] == NULL)
2098 assert(!BITFIELD_ISSET
2099 ((*priv->rxqs)[i]->mac_configured, mac_index));
2105 ret = rxq_mac_addr_add(&priv->rxq_parent, mac_index);
2110 for (i = 0; (i != priv->rxqs_n); ++i) {
2111 if ((*priv->rxqs)[i] == NULL)
2113 ret = rxq_mac_addr_add((*priv->rxqs)[i], mac_index);
2116 /* Failure, rollback. */
2118 if ((*priv->rxqs)[(--i)] != NULL)
2119 rxq_mac_addr_del((*priv->rxqs)[i], mac_index);
2123 BITFIELD_SET(priv->mac_configured, mac_index);
2128 * Enable allmulti mode in a RX queue.
2131 * Pointer to RX queue structure.
2134 * 0 on success, errno value on failure.
2137 rxq_allmulticast_enable(struct rxq *rxq)
2139 struct ibv_exp_flow *flow;
2140 struct ibv_exp_flow_attr attr = {
2141 .type = IBV_EXP_FLOW_ATTR_MC_DEFAULT,
2143 .port = rxq->priv->port,
2147 DEBUG("%p: enabling allmulticast mode", (void *)rxq);
2148 if (rxq->allmulti_flow != NULL)
2151 flow = ibv_exp_create_flow(rxq->qp, &attr);
2153 /* It's not clear whether errno is always set in this case. */
2154 ERROR("%p: flow configuration failed, errno=%d: %s",
2156 (errno ? strerror(errno) : "Unknown error"));
2161 rxq->allmulti_flow = flow;
2162 DEBUG("%p: allmulticast mode enabled", (void *)rxq);
2167 * Disable allmulti mode in a RX queue.
2170 * Pointer to RX queue structure.
2173 rxq_allmulticast_disable(struct rxq *rxq)
2175 DEBUG("%p: disabling allmulticast mode", (void *)rxq);
2176 if (rxq->allmulti_flow == NULL)
2178 claim_zero(ibv_exp_destroy_flow(rxq->allmulti_flow));
2179 rxq->allmulti_flow = NULL;
2180 DEBUG("%p: allmulticast mode disabled", (void *)rxq);
2184 * Enable promiscuous mode in a RX queue.
2187 * Pointer to RX queue structure.
2190 * 0 on success, errno value on failure.
2193 rxq_promiscuous_enable(struct rxq *rxq)
2195 struct ibv_exp_flow *flow;
2196 struct ibv_exp_flow_attr attr = {
2197 .type = IBV_EXP_FLOW_ATTR_ALL_DEFAULT,
2199 .port = rxq->priv->port,
2205 DEBUG("%p: enabling promiscuous mode", (void *)rxq);
2206 if (rxq->promisc_flow != NULL)
2209 flow = ibv_exp_create_flow(rxq->qp, &attr);
2211 /* It's not clear whether errno is always set in this case. */
2212 ERROR("%p: flow configuration failed, errno=%d: %s",
2214 (errno ? strerror(errno) : "Unknown error"));
2219 rxq->promisc_flow = flow;
2220 DEBUG("%p: promiscuous mode enabled", (void *)rxq);
2225 * Disable promiscuous mode in a RX queue.
2228 * Pointer to RX queue structure.
2231 rxq_promiscuous_disable(struct rxq *rxq)
2235 DEBUG("%p: disabling promiscuous mode", (void *)rxq);
2236 if (rxq->promisc_flow == NULL)
2238 claim_zero(ibv_exp_destroy_flow(rxq->promisc_flow));
2239 rxq->promisc_flow = NULL;
2240 DEBUG("%p: promiscuous mode disabled", (void *)rxq);
2244 * Clean up a RX queue.
2246 * Destroy objects, free allocated memory and reset the structure for reuse.
2249 * Pointer to RX queue structure.
2252 rxq_cleanup(struct rxq *rxq)
2254 DEBUG("cleaning up %p", (void *)rxq);
2256 rxq_free_elts_sp(rxq);
2259 if (rxq->qp != NULL) {
2260 rxq_promiscuous_disable(rxq);
2261 rxq_allmulticast_disable(rxq);
2262 rxq_mac_addrs_del(rxq);
2263 claim_zero(ibv_destroy_qp(rxq->qp));
2265 if (rxq->cq != NULL)
2266 claim_zero(ibv_destroy_cq(rxq->cq));
2267 if (rxq->mr != NULL)
2268 claim_zero(ibv_dereg_mr(rxq->mr));
2269 memset(rxq, 0, sizeof(*rxq));
2273 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n);
2276 * DPDK callback for RX with scattered packets support.
2279 * Generic pointer to RX queue structure.
2281 * Array to store received packets.
2283 * Maximum number of packets in array.
2286 * Number of packets successfully received (<= pkts_n).
2289 mlx4_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2291 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2292 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
2293 struct ibv_wc wcs[pkts_n];
2294 struct ibv_recv_wr head;
2295 struct ibv_recv_wr **next = &head.next;
2296 struct ibv_recv_wr *bad_wr;
2301 if (unlikely(!rxq->sp))
2302 return mlx4_rx_burst(dpdk_rxq, pkts, pkts_n);
2303 if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
2305 wcs_n = ibv_poll_cq(rxq->cq, pkts_n, wcs);
2306 if (unlikely(wcs_n == 0))
2308 if (unlikely(wcs_n < 0)) {
2309 DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
2310 (void *)rxq, wcs_n);
2313 assert(wcs_n <= (int)pkts_n);
2314 /* For each work completion. */
2315 for (i = 0; (i != wcs_n); ++i) {
2316 struct ibv_wc *wc = &wcs[i];
2317 uint64_t wr_id = wc->wr_id;
2318 uint32_t len = wc->byte_len;
2319 struct rxq_elt_sp *elt = &(*elts)[wr_id];
2320 struct ibv_recv_wr *wr = &elt->wr;
2321 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
2322 struct rte_mbuf **pkt_buf_next = &pkt_buf;
2323 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
2326 /* Sanity checks. */
2327 assert(wr_id < rxq->elts_n);
2328 assert(wr_id == wr->wr_id);
2329 assert(wr->sg_list == elt->sges);
2330 assert(wr->num_sge == elemof(elt->sges));
2331 /* Link completed WRs together for repost. */
2334 if (unlikely(wc->status != IBV_WC_SUCCESS)) {
2335 /* Whatever, just repost the offending WR. */
2336 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work completion"
2338 (void *)rxq, wc->wr_id, wc->status,
2339 ibv_wc_status_str(wc->status));
2340 #ifdef MLX4_PMD_SOFT_COUNTERS
2341 /* Increase dropped packets counter. */
2342 ++rxq->stats.idropped;
2347 * Replace spent segments with new ones, concatenate and
2348 * return them as pkt_buf.
2351 struct ibv_sge *sge = &elt->sges[j];
2352 struct rte_mbuf *seg = elt->bufs[j];
2353 struct rte_mbuf *rep;
2354 unsigned int seg_tailroom;
2357 * Fetch initial bytes of packet descriptor into a
2358 * cacheline while allocating rep.
2361 rep = __rte_mbuf_raw_alloc(rxq->mp);
2362 if (unlikely(rep == NULL)) {
2364 * Unable to allocate a replacement mbuf,
2367 DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
2368 " can't allocate a new mbuf",
2369 (void *)rxq, wr_id);
2370 if (pkt_buf != NULL)
2371 rte_pktmbuf_free(pkt_buf);
2372 /* Increase out of memory counters. */
2373 ++rxq->stats.rx_nombuf;
2374 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2378 /* Poison user-modifiable fields in rep. */
2379 NEXT(rep) = (void *)((uintptr_t)-1);
2380 SET_DATA_OFF(rep, 0xdead);
2381 DATA_LEN(rep) = 0xd00d;
2382 PKT_LEN(rep) = 0xdeadd00d;
2383 NB_SEGS(rep) = 0x2a;
2387 assert(rep->buf_len == seg->buf_len);
2388 assert(rep->buf_len == rxq->mb_len);
2389 /* Reconfigure sge to use rep instead of seg. */
2390 assert(sge->lkey == rxq->mr->lkey);
2391 sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
2394 /* Update pkt_buf if it's the first segment, or link
2395 * seg to the previous one and update pkt_buf_next. */
2396 *pkt_buf_next = seg;
2397 pkt_buf_next = &NEXT(seg);
2398 /* Update seg information. */
2399 seg_tailroom = (seg->buf_len - seg_headroom);
2400 assert(sge->length == seg_tailroom);
2401 SET_DATA_OFF(seg, seg_headroom);
2402 if (likely(len <= seg_tailroom)) {
2404 DATA_LEN(seg) = len;
2407 assert(rte_pktmbuf_headroom(seg) ==
2409 assert(rte_pktmbuf_tailroom(seg) ==
2410 (seg_tailroom - len));
2413 DATA_LEN(seg) = seg_tailroom;
2414 PKT_LEN(seg) = seg_tailroom;
2416 assert(rte_pktmbuf_headroom(seg) == seg_headroom);
2417 assert(rte_pktmbuf_tailroom(seg) == 0);
2418 /* Fix len and clear headroom for next segments. */
2419 len -= seg_tailroom;
2422 /* Update head and tail segments. */
2423 *pkt_buf_next = NULL;
2424 assert(pkt_buf != NULL);
2426 NB_SEGS(pkt_buf) = j;
2427 PORT(pkt_buf) = rxq->port_id;
2428 PKT_LEN(pkt_buf) = wc->byte_len;
2429 pkt_buf->ol_flags = 0;
2431 /* Return packet. */
2432 *(pkts++) = pkt_buf;
2434 #ifdef MLX4_PMD_SOFT_COUNTERS
2435 /* Increase bytes counter. */
2436 rxq->stats.ibytes += wc->byte_len;
2444 DEBUG("%p: reposting %d WRs starting from %" PRIu64 " (%p)",
2445 (void *)rxq, wcs_n, wcs[0].wr_id, (void *)head.next);
2447 i = ibv_post_recv(rxq->qp, head.next, &bad_wr);
2449 /* Inability to repost WRs is fatal. */
2450 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
2456 #ifdef MLX4_PMD_SOFT_COUNTERS
2457 /* Increase packets counter. */
2458 rxq->stats.ipackets += ret;
2464 * DPDK callback for RX.
2466 * The following function is the same as mlx4_rx_burst_sp(), except it doesn't
2467 * manage scattered packets. Improves performance when MRU is lower than the
2468 * size of the first segment.
2471 * Generic pointer to RX queue structure.
2473 * Array to store received packets.
2475 * Maximum number of packets in array.
2478 * Number of packets successfully received (<= pkts_n).
2481 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2483 struct rxq *rxq = (struct rxq *)dpdk_rxq;
2484 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
2485 struct ibv_wc wcs[pkts_n];
2486 struct ibv_recv_wr head;
2487 struct ibv_recv_wr **next = &head.next;
2488 struct ibv_recv_wr *bad_wr;
2493 if (unlikely(rxq->sp))
2494 return mlx4_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
2495 wcs_n = ibv_poll_cq(rxq->cq, pkts_n, wcs);
2496 if (unlikely(wcs_n == 0))
2498 if (unlikely(wcs_n < 0)) {
2499 DEBUG("rxq=%p, ibv_poll_cq() failed (wc_n=%d)",
2500 (void *)rxq, wcs_n);
2503 assert(wcs_n <= (int)pkts_n);
2504 /* For each work completion. */
2505 for (i = 0; (i != wcs_n); ++i) {
2506 struct ibv_wc *wc = &wcs[i];
2507 uint64_t wr_id = wc->wr_id;
2508 uint32_t len = wc->byte_len;
2509 struct rxq_elt *elt = &(*elts)[WR_ID(wr_id).id];
2510 struct ibv_recv_wr *wr = &elt->wr;
2511 struct rte_mbuf *seg =
2512 (void *)(elt->sge.addr - WR_ID(wr_id).offset);
2513 struct rte_mbuf *rep;
2515 /* Sanity checks. */
2516 assert(WR_ID(wr_id).id < rxq->elts_n);
2517 assert(wr_id == wr->wr_id);
2518 assert(wr->sg_list == &elt->sge);
2519 assert(wr->num_sge == 1);
2520 /* Link completed WRs together for repost. */
2523 if (unlikely(wc->status != IBV_WC_SUCCESS)) {
2524 /* Whatever, just repost the offending WR. */
2525 DEBUG("rxq=%p, wr_id=%" PRIu32 ": bad work completion"
2527 (void *)rxq, WR_ID(wr_id).id, wc->status,
2528 ibv_wc_status_str(wc->status));
2529 #ifdef MLX4_PMD_SOFT_COUNTERS
2530 /* Increase dropped packets counter. */
2531 ++rxq->stats.idropped;
2536 * Fetch initial bytes of packet descriptor into a
2537 * cacheline while allocating rep.
2540 rep = __rte_mbuf_raw_alloc(rxq->mp);
2541 if (unlikely(rep == NULL)) {
2543 * Unable to allocate a replacement mbuf,
2546 DEBUG("rxq=%p, wr_id=%" PRIu32 ":"
2547 " can't allocate a new mbuf",
2548 (void *)rxq, WR_ID(wr_id).id);
2549 /* Increase out of memory counters. */
2550 ++rxq->stats.rx_nombuf;
2551 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
2555 /* Reconfigure sge to use rep instead of seg. */
2556 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
2557 assert(elt->sge.lkey == rxq->mr->lkey);
2558 WR_ID(wr->wr_id).offset =
2559 (((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
2561 assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
2563 /* Update seg information. */
2564 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
2566 PORT(seg) = rxq->port_id;
2569 DATA_LEN(seg) = len;
2572 /* Return packet. */
2575 #ifdef MLX4_PMD_SOFT_COUNTERS
2576 /* Increase bytes counter. */
2577 rxq->stats.ibytes += wc->byte_len;
2585 DEBUG("%p: reposting %d WRs starting from %" PRIu32 " (%p)",
2586 (void *)rxq, wcs_n, WR_ID(wcs[0].wr_id).id, (void *)head.next);
2588 i = ibv_post_recv(rxq->qp, head.next, &bad_wr);
2590 /* Inability to repost WRs is fatal. */
2591 DEBUG("%p: ibv_post_recv(): failed for WR %p: %s",
2597 #ifdef MLX4_PMD_SOFT_COUNTERS
2598 /* Increase packets counter. */
2599 rxq->stats.ipackets += ret;
2607 * Allocate a Queue Pair in case inline receive is supported.
2610 * Pointer to private structure.
2612 * Completion queue to associate with QP.
2614 * Number of descriptors in QP (hint only).
2617 * QP pointer or NULL in case of error.
2619 static struct ibv_qp *
2620 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
2622 struct ibv_exp_qp_init_attr attr = {
2623 /* CQ to be associated with the send queue. */
2625 /* CQ to be associated with the receive queue. */
2627 .max_inl_recv = priv->inl_recv_size,
2629 /* Max number of outstanding WRs. */
2630 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2631 priv->device_attr.max_qp_wr :
2633 /* Max number of scatter/gather elements in a WR. */
2634 .max_recv_sge = ((priv->device_attr.max_sge <
2635 MLX4_PMD_SGE_WR_N) ?
2636 priv->device_attr.max_sge :
2639 .qp_type = IBV_QPT_RAW_PACKET,
2643 attr.comp_mask = IBV_EXP_QP_INIT_ATTR_PD;
2644 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
2646 return ibv_exp_create_qp(priv->ctx, &attr);
2649 #else /* INLINE_RECV */
2652 * Allocate a Queue Pair.
2655 * Pointer to private structure.
2657 * Completion queue to associate with QP.
2659 * Number of descriptors in QP (hint only).
2662 * QP pointer or NULL in case of error.
2664 static struct ibv_qp *
2665 rxq_setup_qp(struct priv *priv, struct ibv_cq *cq, uint16_t desc)
2667 struct ibv_qp_init_attr attr = {
2668 /* CQ to be associated with the send queue. */
2670 /* CQ to be associated with the receive queue. */
2673 /* Max number of outstanding WRs. */
2674 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2675 priv->device_attr.max_qp_wr :
2677 /* Max number of scatter/gather elements in a WR. */
2678 .max_recv_sge = ((priv->device_attr.max_sge <
2679 MLX4_PMD_SGE_WR_N) ?
2680 priv->device_attr.max_sge :
2683 .qp_type = IBV_QPT_RAW_PACKET
2686 return ibv_create_qp(priv->pd, &attr);
2689 #endif /* INLINE_RECV */
2694 * Allocate a RSS Queue Pair.
2697 * Pointer to private structure.
2699 * Completion queue to associate with QP.
2701 * Number of descriptors in QP (hint only).
2703 * If nonzero, create a parent QP, otherwise a child.
2706 * QP pointer or NULL in case of error.
2708 static struct ibv_qp *
2709 rxq_setup_qp_rss(struct priv *priv, struct ibv_cq *cq, uint16_t desc,
2712 struct ibv_exp_qp_init_attr attr = {
2713 /* CQ to be associated with the send queue. */
2715 /* CQ to be associated with the receive queue. */
2718 .max_inl_recv = priv->inl_recv_size,
2721 /* Max number of outstanding WRs. */
2722 .max_recv_wr = ((priv->device_attr.max_qp_wr < desc) ?
2723 priv->device_attr.max_qp_wr :
2725 /* Max number of scatter/gather elements in a WR. */
2726 .max_recv_sge = ((priv->device_attr.max_sge <
2727 MLX4_PMD_SGE_WR_N) ?
2728 priv->device_attr.max_sge :
2731 .qp_type = IBV_QPT_RAW_PACKET,
2732 .comp_mask = (IBV_EXP_QP_INIT_ATTR_PD |
2733 IBV_EXP_QP_INIT_ATTR_QPG),
2738 attr.comp_mask |= IBV_EXP_QP_INIT_ATTR_INL_RECV;
2741 attr.qpg.qpg_type = IBV_EXP_QPG_PARENT;
2742 /* TSS isn't necessary. */
2743 attr.qpg.parent_attrib.tss_child_count = 0;
2744 attr.qpg.parent_attrib.rss_child_count = priv->rxqs_n;
2745 DEBUG("initializing parent RSS queue");
2747 attr.qpg.qpg_type = IBV_EXP_QPG_CHILD_RX;
2748 attr.qpg.qpg_parent = priv->rxq_parent.qp;
2749 DEBUG("initializing child RSS queue");
2751 return ibv_exp_create_qp(priv->ctx, &attr);
2754 #endif /* RSS_SUPPORT */
2757 * Reconfigure a RX queue with new parameters.
2759 * rxq_rehash() does not allocate mbufs, which, if not done from the right
2760 * thread (such as a control thread), may corrupt the pool.
2761 * In case of failure, the queue is left untouched.
2764 * Pointer to Ethernet device structure.
2769 * 0 on success, errno value on failure.
2772 rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
2774 struct priv *priv = rxq->priv;
2775 struct rxq tmpl = *rxq;
2776 unsigned int mbuf_n;
2777 unsigned int desc_n;
2778 struct rte_mbuf **pool;
2780 struct ibv_exp_qp_attr mod;
2781 struct ibv_recv_wr *bad_wr;
2783 int parent = (rxq == &priv->rxq_parent);
2786 ERROR("%p: cannot rehash parent queue %p",
2787 (void *)dev, (void *)rxq);
2790 DEBUG("%p: rehashing queue %p", (void *)dev, (void *)rxq);
2791 /* Number of descriptors and mbufs currently allocated. */
2792 desc_n = (tmpl.elts_n * (tmpl.sp ? MLX4_PMD_SGE_WR_N : 1));
2794 /* Enable scattered packets support for this queue if necessary. */
2795 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
2796 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
2797 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
2799 desc_n /= MLX4_PMD_SGE_WR_N;
2802 DEBUG("%p: %s scattered packets support (%u WRs)",
2803 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc_n);
2804 /* If scatter mode is the same as before, nothing to do. */
2805 if (tmpl.sp == rxq->sp) {
2806 DEBUG("%p: nothing to do", (void *)dev);
2809 /* Remove attached flows if RSS is disabled (no parent queue). */
2811 rxq_allmulticast_disable(&tmpl);
2812 rxq_promiscuous_disable(&tmpl);
2813 rxq_mac_addrs_del(&tmpl);
2814 /* Update original queue in case of failure. */
2815 rxq->allmulti_flow = tmpl.allmulti_flow;
2816 rxq->promisc_flow = tmpl.promisc_flow;
2817 memcpy(rxq->mac_configured, tmpl.mac_configured,
2818 sizeof(rxq->mac_configured));
2819 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
2821 /* From now on, any failure will render the queue unusable.
2822 * Reinitialize QP. */
2823 mod = (struct ibv_exp_qp_attr){ .qp_state = IBV_QPS_RESET };
2824 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
2826 ERROR("%p: cannot reset QP: %s", (void *)dev, strerror(err));
2830 err = ibv_resize_cq(tmpl.cq, desc_n);
2832 ERROR("%p: cannot resize CQ: %s", (void *)dev, strerror(err));
2836 mod = (struct ibv_exp_qp_attr){
2837 /* Move the QP to this state. */
2838 .qp_state = IBV_QPS_INIT,
2839 /* Primary port number. */
2840 .port_num = priv->port
2842 err = ibv_exp_modify_qp(tmpl.qp, &mod,
2845 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
2846 #endif /* RSS_SUPPORT */
2849 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
2850 (void *)dev, strerror(err));
2854 /* Reconfigure flows. Do not care for errors. */
2856 rxq_mac_addrs_add(&tmpl);
2858 rxq_promiscuous_enable(&tmpl);
2860 rxq_allmulticast_enable(&tmpl);
2861 /* Update original queue in case of failure. */
2862 rxq->allmulti_flow = tmpl.allmulti_flow;
2863 rxq->promisc_flow = tmpl.promisc_flow;
2864 memcpy(rxq->mac_configured, tmpl.mac_configured,
2865 sizeof(rxq->mac_configured));
2866 memcpy(rxq->mac_flow, tmpl.mac_flow, sizeof(rxq->mac_flow));
2868 /* Allocate pool. */
2869 pool = rte_malloc(__func__, (mbuf_n * sizeof(*pool)), 0);
2871 ERROR("%p: cannot allocate memory", (void *)dev);
2874 /* Snatch mbufs from original queue. */
2877 struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
2879 for (i = 0; (i != elemof(*elts)); ++i) {
2880 struct rxq_elt_sp *elt = &(*elts)[i];
2883 for (j = 0; (j != elemof(elt->bufs)); ++j) {
2884 assert(elt->bufs[j] != NULL);
2885 pool[k++] = elt->bufs[j];
2889 struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
2891 for (i = 0; (i != elemof(*elts)); ++i) {
2892 struct rxq_elt *elt = &(*elts)[i];
2893 struct rte_mbuf *buf = (void *)
2894 (elt->sge.addr - WR_ID(elt->wr.wr_id).offset);
2896 assert(WR_ID(elt->wr.wr_id).id == i);
2900 assert(k == mbuf_n);
2902 tmpl.elts.sp = NULL;
2903 assert((void *)&tmpl.elts.sp == (void *)&tmpl.elts.no_sp);
2905 rxq_alloc_elts_sp(&tmpl, desc_n, pool) :
2906 rxq_alloc_elts(&tmpl, desc_n, pool));
2908 ERROR("%p: cannot reallocate WRs, aborting", (void *)dev);
2913 assert(tmpl.elts_n == desc_n);
2914 assert(tmpl.elts.sp != NULL);
2916 /* Clean up original data. */
2918 rte_free(rxq->elts.sp);
2919 rxq->elts.sp = NULL;
2921 err = ibv_post_recv(tmpl.qp,
2923 &(*tmpl.elts.sp)[0].wr :
2924 &(*tmpl.elts.no_sp)[0].wr),
2927 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
2933 mod = (struct ibv_exp_qp_attr){
2934 .qp_state = IBV_QPS_RTR
2936 err = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
2938 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
2939 (void *)dev, strerror(err));
2947 * Configure a RX queue.
2950 * Pointer to Ethernet device structure.
2952 * Pointer to RX queue structure.
2954 * Number of descriptors to configure in queue.
2956 * NUMA socket on which memory must be allocated.
2958 * Thresholds parameters.
2960 * Memory pool for buffer allocations.
2963 * 0 on success, errno value on failure.
2966 rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
2967 unsigned int socket, const struct rte_eth_rxconf *conf,
2968 struct rte_mempool *mp)
2970 struct priv *priv = dev->data->dev_private;
2976 struct ibv_exp_qp_attr mod;
2977 struct ibv_recv_wr *bad_wr;
2978 struct rte_mbuf *buf;
2980 int parent = (rxq == &priv->rxq_parent);
2982 (void)conf; /* Thresholds configuration (ignored). */
2984 * If this is a parent queue, hardware must support RSS and
2985 * RSS must be enabled.
2987 assert((!parent) || ((priv->hw_rss) && (priv->rss)));
2989 /* Even if unused, ibv_create_cq() requires at least one
2994 if ((desc == 0) || (desc % MLX4_PMD_SGE_WR_N)) {
2995 ERROR("%p: invalid number of RX descriptors (must be a"
2996 " multiple of %d)", (void *)dev, desc);
2999 /* Get mbuf length. */
3000 buf = rte_pktmbuf_alloc(mp);
3002 ERROR("%p: unable to allocate mbuf", (void *)dev);
3005 tmpl.mb_len = buf->buf_len;
3006 assert((rte_pktmbuf_headroom(buf) +
3007 rte_pktmbuf_tailroom(buf)) == tmpl.mb_len);
3008 assert(rte_pktmbuf_headroom(buf) == RTE_PKTMBUF_HEADROOM);
3009 rte_pktmbuf_free(buf);
3010 /* Enable scattered packets support for this queue if necessary. */
3011 if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
3012 (dev->data->dev_conf.rxmode.max_rx_pkt_len >
3013 (tmpl.mb_len - RTE_PKTMBUF_HEADROOM))) {
3015 desc /= MLX4_PMD_SGE_WR_N;
3017 DEBUG("%p: %s scattered packets support (%u WRs)",
3018 (void *)dev, (tmpl.sp ? "enabling" : "disabling"), desc);
3019 /* Use the entire RX mempool as the memory region. */
3020 tmpl.mr = ibv_reg_mr(priv->pd,
3021 (void *)mp->elt_va_start,
3022 (mp->elt_va_end - mp->elt_va_start),
3023 (IBV_ACCESS_LOCAL_WRITE |
3024 IBV_ACCESS_REMOTE_WRITE));
3025 if (tmpl.mr == NULL) {
3027 ERROR("%p: MR creation failure: %s",
3028 (void *)dev, strerror(ret));
3032 tmpl.cq = ibv_create_cq(priv->ctx, desc, NULL, NULL, 0);
3033 if (tmpl.cq == NULL) {
3035 ERROR("%p: CQ creation failure: %s",
3036 (void *)dev, strerror(ret));
3039 DEBUG("priv->device_attr.max_qp_wr is %d",
3040 priv->device_attr.max_qp_wr);
3041 DEBUG("priv->device_attr.max_sge is %d",
3042 priv->device_attr.max_sge);
3045 tmpl.qp = rxq_setup_qp_rss(priv, tmpl.cq, desc, parent);
3047 #endif /* RSS_SUPPORT */
3048 tmpl.qp = rxq_setup_qp(priv, tmpl.cq, desc);
3049 if (tmpl.qp == NULL) {
3050 ret = (errno ? errno : EINVAL);
3051 ERROR("%p: QP creation failure: %s",
3052 (void *)dev, strerror(ret));
3055 mod = (struct ibv_exp_qp_attr){
3056 /* Move the QP to this state. */
3057 .qp_state = IBV_QPS_INIT,
3058 /* Primary port number. */
3059 .port_num = priv->port
3061 ret = ibv_exp_modify_qp(tmpl.qp, &mod,
3064 (parent ? IBV_EXP_QP_GROUP_RSS : 0) |
3065 #endif /* RSS_SUPPORT */
3068 ERROR("%p: QP state to IBV_QPS_INIT failed: %s",
3069 (void *)dev, strerror(ret));
3072 if ((parent) || (!priv->rss)) {
3073 /* Configure MAC and broadcast addresses. */
3074 ret = rxq_mac_addrs_add(&tmpl);
3076 ERROR("%p: QP flow attachment failed: %s",
3077 (void *)dev, strerror(ret));
3081 /* Allocate descriptors for RX queues, except for the RSS parent. */
3085 ret = rxq_alloc_elts_sp(&tmpl, desc, NULL);
3087 ret = rxq_alloc_elts(&tmpl, desc, NULL);
3089 ERROR("%p: RXQ allocation failed: %s",
3090 (void *)dev, strerror(ret));
3093 ret = ibv_post_recv(tmpl.qp,
3095 &(*tmpl.elts.sp)[0].wr :
3096 &(*tmpl.elts.no_sp)[0].wr),
3099 ERROR("%p: ibv_post_recv() failed for WR %p: %s",
3106 mod = (struct ibv_exp_qp_attr){
3107 .qp_state = IBV_QPS_RTR
3109 ret = ibv_exp_modify_qp(tmpl.qp, &mod, IBV_EXP_QP_STATE);
3111 ERROR("%p: QP state to IBV_QPS_RTR failed: %s",
3112 (void *)dev, strerror(ret));
3116 tmpl.port_id = dev->data->port_id;
3117 DEBUG("%p: RTE port ID: %u", (void *)rxq, tmpl.port_id);
3118 /* Clean up rxq in case we're reinitializing it. */
3119 DEBUG("%p: cleaning-up old rxq just in case", (void *)rxq);
3122 DEBUG("%p: rxq updated with %p", (void *)rxq, (void *)&tmpl);
3132 * DPDK callback to configure a RX queue.
3135 * Pointer to Ethernet device structure.
3139 * Number of descriptors to configure in queue.
3141 * NUMA socket on which memory must be allocated.
3143 * Thresholds parameters.
3145 * Memory pool for buffer allocations.
3148 * 0 on success, negative errno value on failure.
3151 mlx4_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
3152 unsigned int socket, const struct rte_eth_rxconf *conf,
3153 struct rte_mempool *mp)
3155 struct priv *priv = dev->data->dev_private;
3156 struct rxq *rxq = (*priv->rxqs)[idx];
3160 DEBUG("%p: configuring queue %u for %u descriptors",
3161 (void *)dev, idx, desc);
3162 if (idx >= priv->rxqs_n) {
3163 ERROR("%p: queue index out of range (%u >= %u)",
3164 (void *)dev, idx, priv->rxqs_n);
3169 DEBUG("%p: reusing already allocated queue index %u (%p)",
3170 (void *)dev, idx, (void *)rxq);
3171 if (priv->started) {
3175 (*priv->rxqs)[idx] = NULL;
3178 rxq = rte_calloc_socket("RXQ", 1, sizeof(*rxq), 0, socket);
3180 ERROR("%p: unable to allocate queue index %u",
3186 ret = rxq_setup(dev, rxq, desc, socket, conf, mp);
3190 rxq->stats.idx = idx;
3191 DEBUG("%p: adding RX queue %p to list",
3192 (void *)dev, (void *)rxq);
3193 (*priv->rxqs)[idx] = rxq;
3194 /* Update receive callback. */
3196 dev->rx_pkt_burst = mlx4_rx_burst_sp;
3198 dev->rx_pkt_burst = mlx4_rx_burst;
3205 * DPDK callback to release a RX queue.
3208 * Generic RX queue pointer.
3211 mlx4_rx_queue_release(void *dpdk_rxq)
3213 struct rxq *rxq = (struct rxq *)dpdk_rxq;
3221 assert(rxq != &priv->rxq_parent);
3222 for (i = 0; (i != priv->rxqs_n); ++i)
3223 if ((*priv->rxqs)[i] == rxq) {
3224 DEBUG("%p: removing RX queue %p from list",
3225 (void *)priv->dev, (void *)rxq);
3226 (*priv->rxqs)[i] = NULL;
3235 * DPDK callback to start the device.
3237 * Simulate device start by attaching all configured flows.
3240 * Pointer to Ethernet device structure.
3243 * 0 on success, negative errno value on failure.
3246 mlx4_dev_start(struct rte_eth_dev *dev)
3248 struct priv *priv = dev->data->dev_private;
3254 if (priv->started) {
3258 DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
3261 rxq = &priv->rxq_parent;
3264 rxq = (*priv->rxqs)[0];
3267 /* Iterate only once when RSS is enabled. */
3271 /* Ignore nonexistent RX queues. */
3274 ret = rxq_mac_addrs_add(rxq);
3275 if (!ret && priv->promisc)
3276 ret = rxq_promiscuous_enable(rxq);
3277 if (!ret && priv->allmulti)
3278 ret = rxq_allmulticast_enable(rxq);
3281 WARN("%p: QP flow attachment failed: %s",
3282 (void *)dev, strerror(ret));
3285 rxq = (*priv->rxqs)[--i];
3287 rxq_allmulticast_disable(rxq);
3288 rxq_promiscuous_disable(rxq);
3289 rxq_mac_addrs_del(rxq);
3294 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3300 * DPDK callback to stop the device.
3302 * Simulate device stop by detaching all configured flows.
3305 * Pointer to Ethernet device structure.
3308 mlx4_dev_stop(struct rte_eth_dev *dev)
3310 struct priv *priv = dev->data->dev_private;
3316 if (!priv->started) {
3320 DEBUG("%p: detaching flows from all RX queues", (void *)dev);
3323 rxq = &priv->rxq_parent;
3326 rxq = (*priv->rxqs)[0];
3329 /* Iterate only once when RSS is enabled. */
3331 /* Ignore nonexistent RX queues. */
3334 rxq_allmulticast_disable(rxq);
3335 rxq_promiscuous_disable(rxq);
3336 rxq_mac_addrs_del(rxq);
3337 } while ((--r) && ((rxq = (*priv->rxqs)[++i]), i));
3342 * Dummy DPDK callback for TX.
3344 * This function is used to temporarily replace the real callback during
3345 * unsafe control operations on the queue, or in case of error.
3348 * Generic pointer to TX queue structure.
3350 * Packets to transmit.
3352 * Number of packets in array.
3355 * Number of packets successfully transmitted (<= pkts_n).
3358 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
3367 * Dummy DPDK callback for RX.
3369 * This function is used to temporarily replace the real callback during
3370 * unsafe control operations on the queue, or in case of error.
3373 * Generic pointer to RX queue structure.
3375 * Array to store received packets.
3377 * Maximum number of packets in array.
3380 * Number of packets successfully received (<= pkts_n).
3383 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
3392 * DPDK callback to close the device.
3394 * Destroy all queues and objects, free memory.
3397 * Pointer to Ethernet device structure.
3400 mlx4_dev_close(struct rte_eth_dev *dev)
3402 struct priv *priv = dev->data->dev_private;
3407 DEBUG("%p: closing device \"%s\"",
3409 ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
3410 /* Prevent crashes when queues are still in use. This is unfortunately
3411 * still required for DPDK 1.3 because some programs (such as testpmd)
3412 * never release them before closing the device. */
3413 dev->rx_pkt_burst = removed_rx_burst;
3414 dev->tx_pkt_burst = removed_tx_burst;
3415 if (priv->rxqs != NULL) {
3416 /* XXX race condition if mlx4_rx_burst() is still running. */
3418 for (i = 0; (i != priv->rxqs_n); ++i) {
3419 tmp = (*priv->rxqs)[i];
3422 (*priv->rxqs)[i] = NULL;
3429 if (priv->txqs != NULL) {
3430 /* XXX race condition if mlx4_tx_burst() is still running. */
3432 for (i = 0; (i != priv->txqs_n); ++i) {
3433 tmp = (*priv->txqs)[i];
3436 (*priv->txqs)[i] = NULL;
3444 rxq_cleanup(&priv->rxq_parent);
3445 if (priv->pd != NULL) {
3446 assert(priv->ctx != NULL);
3447 claim_zero(ibv_dealloc_pd(priv->pd));
3448 claim_zero(ibv_close_device(priv->ctx));
3450 assert(priv->ctx == NULL);
3452 memset(priv, 0, sizeof(*priv));
3456 * DPDK callback to get information about the device.
3459 * Pointer to Ethernet device structure.
3461 * Info structure output buffer.
3464 mlx4_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
3466 struct priv *priv = dev->data->dev_private;
3470 /* FIXME: we should ask the device for these values. */
3471 info->min_rx_bufsize = 32;
3472 info->max_rx_pktlen = 65536;
3474 * Since we need one CQ per QP, the limit is the minimum number
3475 * between the two values.
3477 max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
3478 priv->device_attr.max_qp : priv->device_attr.max_cq);
3479 /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
3482 info->max_rx_queues = max;
3483 info->max_tx_queues = max;
3484 info->max_mac_addrs = elemof(priv->mac);
3489 * DPDK callback to get device statistics.
3492 * Pointer to Ethernet device structure.
3494 * Stats structure output buffer.
3497 mlx4_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
3499 struct priv *priv = dev->data->dev_private;
3500 struct rte_eth_stats tmp = { .ipackets = 0 };
3505 /* Add software counters. */
3506 for (i = 0; (i != priv->rxqs_n); ++i) {
3507 struct rxq *rxq = (*priv->rxqs)[i];
3511 idx = rxq->stats.idx;
3512 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3513 #ifdef MLX4_PMD_SOFT_COUNTERS
3514 tmp.q_ipackets[idx] += rxq->stats.ipackets;
3515 tmp.q_ibytes[idx] += rxq->stats.ibytes;
3517 tmp.q_errors[idx] += (rxq->stats.idropped +
3518 rxq->stats.rx_nombuf);
3520 #ifdef MLX4_PMD_SOFT_COUNTERS
3521 tmp.ipackets += rxq->stats.ipackets;
3522 tmp.ibytes += rxq->stats.ibytes;
3524 tmp.ierrors += rxq->stats.idropped;
3525 tmp.rx_nombuf += rxq->stats.rx_nombuf;
3527 for (i = 0; (i != priv->txqs_n); ++i) {
3528 struct txq *txq = (*priv->txqs)[i];
3532 idx = txq->stats.idx;
3533 if (idx < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
3534 #ifdef MLX4_PMD_SOFT_COUNTERS
3535 tmp.q_opackets[idx] += txq->stats.opackets;
3536 tmp.q_obytes[idx] += txq->stats.obytes;
3538 tmp.q_errors[idx] += txq->stats.odropped;
3540 #ifdef MLX4_PMD_SOFT_COUNTERS
3541 tmp.opackets += txq->stats.opackets;
3542 tmp.obytes += txq->stats.obytes;
3544 tmp.oerrors += txq->stats.odropped;
3546 #ifndef MLX4_PMD_SOFT_COUNTERS
3547 /* FIXME: retrieve and add hardware counters. */
3554 * DPDK callback to clear device statistics.
3557 * Pointer to Ethernet device structure.
3560 mlx4_stats_reset(struct rte_eth_dev *dev)
3562 struct priv *priv = dev->data->dev_private;
3567 for (i = 0; (i != priv->rxqs_n); ++i) {
3568 if ((*priv->rxqs)[i] == NULL)
3570 idx = (*priv->rxqs)[i]->stats.idx;
3571 (*priv->rxqs)[i]->stats =
3572 (struct mlx4_rxq_stats){ .idx = idx };
3574 for (i = 0; (i != priv->txqs_n); ++i) {
3575 if ((*priv->txqs)[i] == NULL)
3577 idx = (*priv->rxqs)[i]->stats.idx;
3578 (*priv->txqs)[i]->stats =
3579 (struct mlx4_txq_stats){ .idx = idx };
3581 #ifndef MLX4_PMD_SOFT_COUNTERS
3582 /* FIXME: reset hardware counters. */
3588 * DPDK callback to remove a MAC address.
3591 * Pointer to Ethernet device structure.
3593 * MAC address index.
3596 mlx4_mac_addr_remove(struct rte_eth_dev *dev, uint32_t index)
3598 struct priv *priv = dev->data->dev_private;
3601 DEBUG("%p: removing MAC address from index %" PRIu32,
3602 (void *)dev, index);
3603 if (index >= MLX4_MAX_MAC_ADDRESSES)
3605 /* Refuse to remove the broadcast address, this one is special. */
3606 if (!memcmp(priv->mac[index].addr_bytes, "\xff\xff\xff\xff\xff\xff",
3609 priv_mac_addr_del(priv, index);
3615 * DPDK callback to add a MAC address.
3618 * Pointer to Ethernet device structure.
3620 * MAC address to register.
3622 * MAC address index.
3624 * VMDq pool index to associate address with (ignored).
3627 mlx4_mac_addr_add(struct rte_eth_dev *dev, struct ether_addr *mac_addr,
3628 uint32_t index, uint32_t vmdq)
3630 struct priv *priv = dev->data->dev_private;
3634 DEBUG("%p: adding MAC address at index %" PRIu32,
3635 (void *)dev, index);
3636 if (index >= MLX4_MAX_MAC_ADDRESSES)
3638 /* Refuse to add the broadcast address, this one is special. */
3639 if (!memcmp(mac_addr->addr_bytes, "\xff\xff\xff\xff\xff\xff",
3642 priv_mac_addr_add(priv, index,
3643 (const uint8_t (*)[ETHER_ADDR_LEN])
3644 mac_addr->addr_bytes);
3650 * DPDK callback to enable promiscuous mode.
3653 * Pointer to Ethernet device structure.
3656 mlx4_promiscuous_enable(struct rte_eth_dev *dev)
3658 struct priv *priv = dev->data->dev_private;
3663 if (priv->promisc) {
3667 /* If device isn't started, this is all we need to do. */
3671 ret = rxq_promiscuous_enable(&priv->rxq_parent);
3678 for (i = 0; (i != priv->rxqs_n); ++i) {
3679 if ((*priv->rxqs)[i] == NULL)
3681 ret = rxq_promiscuous_enable((*priv->rxqs)[i]);
3684 /* Failure, rollback. */
3686 if ((*priv->rxqs)[--i] != NULL)
3687 rxq_promiscuous_disable((*priv->rxqs)[i]);
3697 * DPDK callback to disable promiscuous mode.
3700 * Pointer to Ethernet device structure.
3703 mlx4_promiscuous_disable(struct rte_eth_dev *dev)
3705 struct priv *priv = dev->data->dev_private;
3709 if (!priv->promisc) {
3714 rxq_promiscuous_disable(&priv->rxq_parent);
3717 for (i = 0; (i != priv->rxqs_n); ++i)
3718 if ((*priv->rxqs)[i] != NULL)
3719 rxq_promiscuous_disable((*priv->rxqs)[i]);
3726 * DPDK callback to enable allmulti mode.
3729 * Pointer to Ethernet device structure.
3732 mlx4_allmulticast_enable(struct rte_eth_dev *dev)
3734 struct priv *priv = dev->data->dev_private;
3739 if (priv->allmulti) {
3743 /* If device isn't started, this is all we need to do. */
3747 ret = rxq_allmulticast_enable(&priv->rxq_parent);
3754 for (i = 0; (i != priv->rxqs_n); ++i) {
3755 if ((*priv->rxqs)[i] == NULL)
3757 ret = rxq_allmulticast_enable((*priv->rxqs)[i]);
3760 /* Failure, rollback. */
3762 if ((*priv->rxqs)[--i] != NULL)
3763 rxq_allmulticast_disable((*priv->rxqs)[i]);
3773 * DPDK callback to disable allmulti mode.
3776 * Pointer to Ethernet device structure.
3779 mlx4_allmulticast_disable(struct rte_eth_dev *dev)
3781 struct priv *priv = dev->data->dev_private;
3785 if (!priv->allmulti) {
3790 rxq_allmulticast_disable(&priv->rxq_parent);
3793 for (i = 0; (i != priv->rxqs_n); ++i)
3794 if ((*priv->rxqs)[i] != NULL)
3795 rxq_allmulticast_disable((*priv->rxqs)[i]);
3802 * DPDK callback to retrieve physical link information (unlocked version).
3805 * Pointer to Ethernet device structure.
3806 * @param wait_to_complete
3807 * Wait for request completion (ignored).
3810 mlx4_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
3812 struct priv *priv = dev->data->dev_private;
3813 struct ibv_port_attr port_attr;
3814 static const uint8_t width_mult[] = {
3815 /* Multiplier values taken from devinfo.c in libibverbs. */
3816 0, 1, 4, 0, 8, 0, 0, 0, 12, 0
3819 (void)wait_to_complete;
3820 errno = ibv_query_port(priv->ctx, priv->port, &port_attr);
3822 WARN("port query failed: %s", strerror(errno));
3825 dev->data->dev_link = (struct rte_eth_link){
3826 .link_speed = (ibv_rate_to_mbps(mult_to_ibv_rate
3827 (port_attr.active_speed)) *
3828 width_mult[(port_attr.active_width %
3829 sizeof(width_mult))]),
3830 .link_duplex = ETH_LINK_FULL_DUPLEX,
3831 .link_status = (port_attr.state == IBV_PORT_ACTIVE)
3833 if (memcmp(&port_attr, &priv->port_attr, sizeof(port_attr))) {
3834 /* Link status changed. */
3835 priv->port_attr = port_attr;
3838 /* Link status is still the same. */
3843 * DPDK callback to retrieve physical link information.
3846 * Pointer to Ethernet device structure.
3847 * @param wait_to_complete
3848 * Wait for request completion (ignored).
3851 mlx4_link_update(struct rte_eth_dev *dev, int wait_to_complete)
3853 struct priv *priv = dev->data->dev_private;
3857 ret = mlx4_link_update_unlocked(dev, wait_to_complete);
3863 * DPDK callback to change the MTU.
3865 * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
3866 * received). Use this as a hint to enable/disable scattered packets support
3867 * and improve performance when not needed.
3868 * Since failure is not an option, reconfiguring queues on the fly is not
3872 * Pointer to Ethernet device structure.
3877 * 0 on success, negative errno value on failure.
3880 mlx4_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
3882 struct priv *priv = dev->data->dev_private;
3885 uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
3889 /* Set kernel interface MTU first. */
3890 if (priv_set_mtu(priv, mtu)) {
3892 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
3896 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
3898 /* Temporarily replace RX handler with a fake one, assuming it has not
3899 * been copied elsewhere. */
3900 dev->rx_pkt_burst = removed_rx_burst;
3901 /* Make sure everyone has left mlx4_rx_burst() and uses
3902 * removed_rx_burst() instead. */
3905 /* Reconfigure each RX queue. */
3906 for (i = 0; (i != priv->rxqs_n); ++i) {
3907 struct rxq *rxq = (*priv->rxqs)[i];
3908 unsigned int max_frame_len;
3913 /* Calculate new maximum frame length according to MTU and
3914 * toggle scattered support (sp) if necessary. */
3915 max_frame_len = (priv->mtu + ETHER_HDR_LEN +
3916 (ETHER_MAX_VLAN_FRAME_LEN - ETHER_MAX_LEN));
3917 sp = (max_frame_len > (rxq->mb_len - RTE_PKTMBUF_HEADROOM));
3918 /* Provide new values to rxq_setup(). */
3919 dev->data->dev_conf.rxmode.jumbo_frame = sp;
3920 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
3921 ret = rxq_rehash(dev, rxq);
3923 /* Force SP RX if that queue requires it and abort. */
3925 rx_func = mlx4_rx_burst_sp;
3928 /* Reenable non-RSS queue attributes. No need to check
3929 * for errors at this stage. */
3931 rxq_mac_addrs_add(rxq);
3933 rxq_promiscuous_enable(rxq);
3935 rxq_allmulticast_enable(rxq);
3937 /* Scattered burst function takes priority. */
3939 rx_func = mlx4_rx_burst_sp;
3941 /* Burst functions can now be called again. */
3943 dev->rx_pkt_burst = rx_func;
3951 * DPDK callback to get flow control status.
3954 * Pointer to Ethernet device structure.
3955 * @param[out] fc_conf
3956 * Flow control output buffer.
3959 * 0 on success, negative errno value on failure.
3962 mlx4_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
3964 struct priv *priv = dev->data->dev_private;
3966 struct ethtool_pauseparam ethpause = {
3967 .cmd = ETHTOOL_GPAUSEPARAM
3971 ifr.ifr_data = ðpause;
3973 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
3975 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
3981 fc_conf->autoneg = ethpause.autoneg;
3982 if (ethpause.rx_pause && ethpause.tx_pause)
3983 fc_conf->mode = RTE_FC_FULL;
3984 else if (ethpause.rx_pause)
3985 fc_conf->mode = RTE_FC_RX_PAUSE;
3986 else if (ethpause.tx_pause)
3987 fc_conf->mode = RTE_FC_TX_PAUSE;
3989 fc_conf->mode = RTE_FC_NONE;
3999 * DPDK callback to modify flow control parameters.
4002 * Pointer to Ethernet device structure.
4003 * @param[in] fc_conf
4004 * Flow control parameters.
4007 * 0 on success, negative errno value on failure.
4010 mlx4_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
4012 struct priv *priv = dev->data->dev_private;
4014 struct ethtool_pauseparam ethpause = {
4015 .cmd = ETHTOOL_SPAUSEPARAM
4019 ifr.ifr_data = ðpause;
4020 ethpause.autoneg = fc_conf->autoneg;
4021 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4022 (fc_conf->mode & RTE_FC_RX_PAUSE))
4023 ethpause.rx_pause = 1;
4025 ethpause.rx_pause = 0;
4027 if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
4028 (fc_conf->mode & RTE_FC_TX_PAUSE))
4029 ethpause.tx_pause = 1;
4031 ethpause.tx_pause = 0;
4034 if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
4036 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
4050 * Configure a VLAN filter.
4053 * Pointer to Ethernet device structure.
4055 * VLAN ID to filter.
4060 * 0 on success, errno value on failure.
4063 vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4065 struct priv *priv = dev->data->dev_private;
4067 unsigned int j = -1;
4069 DEBUG("%p: %s VLAN filter ID %" PRIu16,
4070 (void *)dev, (on ? "enable" : "disable"), vlan_id);
4071 for (i = 0; (i != elemof(priv->vlan_filter)); ++i) {
4072 if (!priv->vlan_filter[i].enabled) {
4073 /* Unused index, remember it. */
4077 if (priv->vlan_filter[i].id != vlan_id)
4079 /* This VLAN ID is already known, use its index. */
4083 /* Check if there's room for another VLAN filter. */
4084 if (j == (unsigned int)-1)
4087 * VLAN filters apply to all configured MAC addresses, flow
4088 * specifications must be reconfigured accordingly.
4090 priv->vlan_filter[j].id = vlan_id;
4091 if ((on) && (!priv->vlan_filter[j].enabled)) {
4093 * Filter is disabled, enable it.
4094 * Rehashing flows in all RX queues is necessary.
4097 rxq_mac_addrs_del(&priv->rxq_parent);
4099 for (i = 0; (i != priv->rxqs_n); ++i)
4100 if ((*priv->rxqs)[i] != NULL)
4101 rxq_mac_addrs_del((*priv->rxqs)[i]);
4102 priv->vlan_filter[j].enabled = 1;
4103 if (priv->started) {
4105 rxq_mac_addrs_add(&priv->rxq_parent);
4107 for (i = 0; (i != priv->rxqs_n); ++i) {
4108 if ((*priv->rxqs)[i] == NULL)
4110 rxq_mac_addrs_add((*priv->rxqs)[i]);
4113 } else if ((!on) && (priv->vlan_filter[j].enabled)) {
4115 * Filter is enabled, disable it.
4116 * Rehashing flows in all RX queues is necessary.
4119 rxq_mac_addrs_del(&priv->rxq_parent);
4121 for (i = 0; (i != priv->rxqs_n); ++i)
4122 if ((*priv->rxqs)[i] != NULL)
4123 rxq_mac_addrs_del((*priv->rxqs)[i]);
4124 priv->vlan_filter[j].enabled = 0;
4125 if (priv->started) {
4127 rxq_mac_addrs_add(&priv->rxq_parent);
4129 for (i = 0; (i != priv->rxqs_n); ++i) {
4130 if ((*priv->rxqs)[i] == NULL)
4132 rxq_mac_addrs_add((*priv->rxqs)[i]);
4140 * DPDK callback to configure a VLAN filter.
4143 * Pointer to Ethernet device structure.
4145 * VLAN ID to filter.
4150 * 0 on success, negative errno value on failure.
4153 mlx4_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vlan_id, int on)
4155 struct priv *priv = dev->data->dev_private;
4159 ret = vlan_filter_set(dev, vlan_id, on);
4165 static struct eth_dev_ops mlx4_dev_ops = {
4166 .dev_configure = mlx4_dev_configure,
4167 .dev_start = mlx4_dev_start,
4168 .dev_stop = mlx4_dev_stop,
4169 .dev_close = mlx4_dev_close,
4170 .promiscuous_enable = mlx4_promiscuous_enable,
4171 .promiscuous_disable = mlx4_promiscuous_disable,
4172 .allmulticast_enable = mlx4_allmulticast_enable,
4173 .allmulticast_disable = mlx4_allmulticast_disable,
4174 .link_update = mlx4_link_update,
4175 .stats_get = mlx4_stats_get,
4176 .stats_reset = mlx4_stats_reset,
4177 .queue_stats_mapping_set = NULL,
4178 .dev_infos_get = mlx4_dev_infos_get,
4179 .vlan_filter_set = mlx4_vlan_filter_set,
4180 .vlan_tpid_set = NULL,
4181 .vlan_strip_queue_set = NULL,
4182 .vlan_offload_set = NULL,
4183 .rx_queue_setup = mlx4_rx_queue_setup,
4184 .tx_queue_setup = mlx4_tx_queue_setup,
4185 .rx_queue_release = mlx4_rx_queue_release,
4186 .tx_queue_release = mlx4_tx_queue_release,
4188 .dev_led_off = NULL,
4189 .flow_ctrl_get = mlx4_dev_get_flow_ctrl,
4190 .flow_ctrl_set = mlx4_dev_set_flow_ctrl,
4191 .priority_flow_ctrl_set = NULL,
4192 .mac_addr_remove = mlx4_mac_addr_remove,
4193 .mac_addr_add = mlx4_mac_addr_add,
4194 .mtu_set = mlx4_dev_set_mtu,
4195 .fdir_add_signature_filter = NULL,
4196 .fdir_update_signature_filter = NULL,
4197 .fdir_remove_signature_filter = NULL,
4198 .fdir_add_perfect_filter = NULL,
4199 .fdir_update_perfect_filter = NULL,
4200 .fdir_remove_perfect_filter = NULL,
4201 .fdir_set_masks = NULL
4205 * Get PCI information from struct ibv_device.
4208 * Pointer to Ethernet device structure.
4209 * @param[out] pci_addr
4210 * PCI bus address output buffer.
4213 * 0 on success, -1 on failure and errno is set.
4216 mlx4_ibv_device_to_pci_addr(const struct ibv_device *device,
4217 struct rte_pci_addr *pci_addr)
4221 MKSTR(path, "%s/device/uevent", device->ibdev_path);
4223 file = fopen(path, "rb");
4226 while (fgets(line, sizeof(line), file) == line) {
4227 size_t len = strlen(line);
4230 /* Truncate long lines. */
4231 if (len == (sizeof(line) - 1))
4232 while (line[(len - 1)] != '\n') {
4236 line[(len - 1)] = ret;
4238 /* Extract information. */
4241 "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
4245 &pci_addr->function) == 4) {
4255 * Derive MAC address from port GID.
4258 * MAC address output buffer.
4260 * Physical port number.
4265 mac_from_gid(uint8_t (*mac)[ETHER_ADDR_LEN], uint32_t port, uint8_t *gid)
4267 memcpy(&(*mac)[0], gid + 8, 3);
4268 memcpy(&(*mac)[3], gid + 13, 3);
4273 /* Support up to 32 adapters. */
4275 struct rte_pci_addr pci_addr; /* associated PCI address */
4276 uint32_t ports; /* physical ports bitfield. */
4280 * Get device index in mlx4_dev[] from PCI bus address.
4282 * @param[in] pci_addr
4283 * PCI bus address to look for.
4286 * mlx4_dev[] index on success, -1 on failure.
4289 mlx4_dev_idx(struct rte_pci_addr *pci_addr)
4294 assert(pci_addr != NULL);
4295 for (i = 0; (i != elemof(mlx4_dev)); ++i) {
4296 if ((mlx4_dev[i].pci_addr.domain == pci_addr->domain) &&
4297 (mlx4_dev[i].pci_addr.bus == pci_addr->bus) &&
4298 (mlx4_dev[i].pci_addr.devid == pci_addr->devid) &&
4299 (mlx4_dev[i].pci_addr.function == pci_addr->function))
4301 if ((mlx4_dev[i].ports == 0) && (ret == -1))
4308 * Retrieve integer value from environment variable.
4311 * Environment variable name.
4314 * Integer value, 0 if the variable is not set.
4317 mlx4_getenv_int(const char *name)
4319 const char *val = getenv(name);
4326 static struct eth_driver mlx4_driver;
4329 * DPDK callback to register a PCI device.
4331 * This function creates an Ethernet device for each port of a given
4334 * @param[in] pci_drv
4335 * PCI driver structure (mlx4_driver).
4336 * @param[in] pci_dev
4337 * PCI device information.
4340 * 0 on success, negative errno value on failure.
4343 mlx4_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
4345 struct ibv_device **list;
4346 struct ibv_device *ibv_dev;
4348 struct ibv_context *attr_ctx = NULL;
4349 struct ibv_device_attr device_attr;
4355 assert(pci_drv == &mlx4_driver.pci_drv);
4356 /* Get mlx4_dev[] index. */
4357 idx = mlx4_dev_idx(&pci_dev->addr);
4359 ERROR("this driver cannot support any more adapters");
4362 DEBUG("using driver device index %d", idx);
4364 /* Save PCI address. */
4365 mlx4_dev[idx].pci_addr = pci_dev->addr;
4366 list = ibv_get_device_list(&i);
4369 if (errno == ENOSYS) {
4370 WARN("cannot list devices, is ib_uverbs loaded?");
4377 * For each listed device, check related sysfs entry against
4378 * the provided PCI ID.
4381 struct rte_pci_addr pci_addr;
4384 DEBUG("checking device \"%s\"", list[i]->name);
4385 if (mlx4_ibv_device_to_pci_addr(list[i], &pci_addr))
4387 if ((pci_dev->addr.domain != pci_addr.domain) ||
4388 (pci_dev->addr.bus != pci_addr.bus) ||
4389 (pci_dev->addr.devid != pci_addr.devid) ||
4390 (pci_dev->addr.function != pci_addr.function))
4392 vf = (pci_dev->id.device_id ==
4393 PCI_DEVICE_ID_MELLANOX_CONNECTX3VF);
4394 INFO("PCI information matches, using device \"%s\" (VF: %s)",
4395 list[i]->name, (vf ? "true" : "false"));
4396 attr_ctx = ibv_open_device(list[i]);
4400 if (attr_ctx == NULL) {
4401 ibv_free_device_list(list);
4404 WARN("cannot access device, is mlx4_ib loaded?");
4407 WARN("cannot use device, are drivers up to date?");
4415 DEBUG("device opened");
4416 if (ibv_query_device(attr_ctx, &device_attr))
4418 INFO("%u port(s) detected", device_attr.phys_port_cnt);
4420 for (i = 0; i < device_attr.phys_port_cnt; i++) {
4421 uint32_t port = i + 1; /* ports are indexed from one */
4422 uint32_t test = (1 << i);
4423 struct ibv_context *ctx = NULL;
4424 struct ibv_port_attr port_attr;
4425 struct ibv_pd *pd = NULL;
4426 struct priv *priv = NULL;
4427 struct rte_eth_dev *eth_dev;
4428 #if defined(INLINE_RECV) || defined(RSS_SUPPORT)
4429 struct ibv_exp_device_attr exp_device_attr;
4431 struct ether_addr mac;
4432 union ibv_gid temp_gid;
4435 exp_device_attr.comp_mask =
4436 (IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
4437 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ);
4438 #endif /* RSS_SUPPORT */
4440 DEBUG("using port %u (%08" PRIx32 ")", port, test);
4442 ctx = ibv_open_device(ibv_dev);
4446 /* Check port status. */
4447 err = ibv_query_port(ctx, port, &port_attr);
4449 ERROR("port query failed: %s", strerror(err));
4452 if (port_attr.state != IBV_PORT_ACTIVE)
4453 WARN("bad state for port %d: \"%s\" (%d)",
4454 port, ibv_port_state_str(port_attr.state),
4457 /* Allocate protection domain. */
4458 pd = ibv_alloc_pd(ctx);
4460 ERROR("PD allocation failure");
4465 mlx4_dev[idx].ports |= test;
4467 /* from rte_ethdev.c */
4468 priv = rte_zmalloc("ethdev private structure",
4470 RTE_CACHE_LINE_SIZE);
4472 ERROR("priv allocation failure");
4478 priv->device_attr = device_attr;
4479 priv->port_attr = port_attr;
4482 priv->mtu = ETHER_MTU;
4484 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
4485 INFO("experimental ibv_exp_query_device");
4488 if ((exp_device_attr.exp_device_cap_flags &
4489 IBV_EXP_DEVICE_QPG) &&
4490 (exp_device_attr.exp_device_cap_flags &
4491 IBV_EXP_DEVICE_UD_RSS) &&
4492 (exp_device_attr.comp_mask &
4493 IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ) &&
4494 (exp_device_attr.max_rss_tbl_sz > 0)) {
4497 priv->max_rss_tbl_sz = exp_device_attr.max_rss_tbl_sz;
4501 priv->max_rss_tbl_sz = 0;
4503 priv->hw_tss = !!(exp_device_attr.exp_device_cap_flags &
4504 IBV_EXP_DEVICE_UD_TSS);
4505 DEBUG("device flags: %s%s%s",
4506 (priv->hw_qpg ? "IBV_DEVICE_QPG " : ""),
4507 (priv->hw_tss ? "IBV_DEVICE_TSS " : ""),
4508 (priv->hw_rss ? "IBV_DEVICE_RSS " : ""));
4510 DEBUG("maximum RSS indirection table size: %u",
4511 exp_device_attr.max_rss_tbl_sz);
4512 #endif /* RSS_SUPPORT */
4515 priv->inl_recv_size = mlx4_getenv_int("MLX4_INLINE_RECV_SIZE");
4517 if (priv->inl_recv_size) {
4518 exp_device_attr.comp_mask =
4519 IBV_EXP_DEVICE_ATTR_INLINE_RECV_SZ;
4520 if (ibv_exp_query_device(ctx, &exp_device_attr)) {
4521 INFO("Couldn't query device for inline-receive"
4523 priv->inl_recv_size = 0;
4525 if ((unsigned)exp_device_attr.inline_recv_sz <
4526 priv->inl_recv_size) {
4527 INFO("Max inline-receive (%d) <"
4528 " requested inline-receive (%u)",
4529 exp_device_attr.inline_recv_sz,
4530 priv->inl_recv_size);
4531 priv->inl_recv_size =
4532 exp_device_attr.inline_recv_sz;
4535 INFO("Set inline receive size to %u",
4536 priv->inl_recv_size);
4538 #endif /* INLINE_RECV */
4540 (void)mlx4_getenv_int;
4542 if (ibv_query_gid(ctx, port, 0, &temp_gid)) {
4543 ERROR("ibv_query_gid() failure");
4546 /* Configure the first MAC address by default. */
4547 mac_from_gid(&mac.addr_bytes, port, temp_gid.raw);
4548 INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
4550 mac.addr_bytes[0], mac.addr_bytes[1],
4551 mac.addr_bytes[2], mac.addr_bytes[3],
4552 mac.addr_bytes[4], mac.addr_bytes[5]);
4553 /* Register MAC and broadcast addresses. */
4554 claim_zero(priv_mac_addr_add(priv, 0,
4555 (const uint8_t (*)[ETHER_ADDR_LEN])
4557 claim_zero(priv_mac_addr_add(priv, 1,
4558 &(const uint8_t [ETHER_ADDR_LEN])
4559 { "\xff\xff\xff\xff\xff\xff" }));
4562 char ifname[IF_NAMESIZE];
4564 if (priv_get_ifname(priv, &ifname) == 0)
4565 DEBUG("port %u ifname is \"%s\"",
4566 priv->port, ifname);
4568 DEBUG("port %u ifname is unknown", priv->port);
4571 /* Get actual MTU if possible. */
4572 priv_get_mtu(priv, &priv->mtu);
4573 DEBUG("port %u MTU is %u", priv->port, priv->mtu);
4575 /* from rte_ethdev.c */
4577 char name[RTE_ETH_NAME_MAX_LEN];
4579 snprintf(name, sizeof(name), "%s port %u",
4580 ibv_get_device_name(ibv_dev), port);
4581 eth_dev = rte_eth_dev_allocate(name, RTE_ETH_DEV_PCI);
4583 if (eth_dev == NULL) {
4584 ERROR("can not allocate rte ethdev");
4589 eth_dev->data->dev_private = priv;
4590 eth_dev->pci_dev = pci_dev;
4591 eth_dev->driver = &mlx4_driver;
4592 eth_dev->data->rx_mbuf_alloc_failed = 0;
4593 eth_dev->data->mtu = ETHER_MTU;
4595 priv->dev = eth_dev;
4596 eth_dev->dev_ops = &mlx4_dev_ops;
4597 eth_dev->data->mac_addrs = priv->mac;
4599 /* Bring Ethernet device up. */
4600 DEBUG("forcing Ethernet interface up");
4601 priv_set_flags(priv, ~IFF_UP, IFF_UP);
4607 claim_zero(ibv_dealloc_pd(pd));
4609 claim_zero(ibv_close_device(ctx));
4614 * XXX if something went wrong in the loop above, there is a resource
4615 * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
4616 * long as the dpdk does not provide a way to deallocate a ethdev and a
4617 * way to enumerate the registered ethdevs to free the previous ones.
4620 /* no port found, complain */
4621 if (!mlx4_dev[idx].ports) {
4628 claim_zero(ibv_close_device(attr_ctx));
4630 ibv_free_device_list(list);
4635 static struct rte_pci_id mlx4_pci_id_map[] = {
4637 .vendor_id = PCI_VENDOR_ID_MELLANOX,
4638 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3,
4639 .subsystem_vendor_id = PCI_ANY_ID,
4640 .subsystem_device_id = PCI_ANY_ID
4643 .vendor_id = PCI_VENDOR_ID_MELLANOX,
4644 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO,
4645 .subsystem_vendor_id = PCI_ANY_ID,
4646 .subsystem_device_id = PCI_ANY_ID
4649 .vendor_id = PCI_VENDOR_ID_MELLANOX,
4650 .device_id = PCI_DEVICE_ID_MELLANOX_CONNECTX3VF,
4651 .subsystem_vendor_id = PCI_ANY_ID,
4652 .subsystem_device_id = PCI_ANY_ID
4659 static struct eth_driver mlx4_driver = {
4661 .name = MLX4_DRIVER_NAME,
4662 .id_table = mlx4_pci_id_map,
4663 .devinit = mlx4_pci_devinit,
4665 .dev_private_size = sizeof(struct priv)
4669 * Driver initialization routine.
4672 rte_mlx4_pmd_init(const char *name, const char *args)
4676 rte_eal_pci_register(&mlx4_driver.pci_drv);
4680 static struct rte_driver rte_mlx4_driver = {
4682 .name = MLX4_DRIVER_NAME,
4683 .init = rte_mlx4_pmd_init,
4686 PMD_REGISTER_DRIVER(rte_mlx4_driver)