1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2018-2019 Cisco Systems, Inc. All rights reserved.
9 #include <sys/socket.h>
11 #include <sys/ioctl.h>
13 #include <linux/if_ether.h>
15 #include <sys/eventfd.h>
17 #include <rte_version.h>
19 #include <rte_ether.h>
20 #include <ethdev_driver.h>
21 #include <ethdev_vdev.h>
22 #include <rte_malloc.h>
23 #include <rte_kvargs.h>
24 #include <rte_bus_vdev.h>
25 #include <rte_string_fns.h>
26 #include <rte_errno.h>
27 #include <rte_memory.h>
28 #include <rte_memzone.h>
29 #include <rte_eal_memconfig.h>
31 #include "rte_eth_memif.h"
32 #include "memif_socket.h"
34 #define ETH_MEMIF_ID_ARG "id"
35 #define ETH_MEMIF_ROLE_ARG "role"
36 #define ETH_MEMIF_PKT_BUFFER_SIZE_ARG "bsize"
37 #define ETH_MEMIF_RING_SIZE_ARG "rsize"
38 #define ETH_MEMIF_SOCKET_ARG "socket"
39 #define ETH_MEMIF_SOCKET_ABSTRACT_ARG "socket-abstract"
40 #define ETH_MEMIF_MAC_ARG "mac"
41 #define ETH_MEMIF_ZC_ARG "zero-copy"
42 #define ETH_MEMIF_SECRET_ARG "secret"
44 static const char * const valid_arguments[] = {
47 ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
48 ETH_MEMIF_RING_SIZE_ARG,
50 ETH_MEMIF_SOCKET_ABSTRACT_ARG,
57 static const struct rte_eth_link pmd_link = {
58 .link_speed = RTE_ETH_SPEED_NUM_10G,
59 .link_duplex = RTE_ETH_LINK_FULL_DUPLEX,
60 .link_status = RTE_ETH_LINK_DOWN,
61 .link_autoneg = RTE_ETH_LINK_AUTONEG
64 #define MEMIF_MP_SEND_REGION "memif_mp_send_region"
67 static int memif_region_init_zc(const struct rte_memseg_list *msl,
68 const struct rte_memseg *ms, void *arg);
73 return ("memif-" RTE_STR(MEMIF_VERSION_MAJOR) "." RTE_STR(MEMIF_VERSION_MINOR));
76 /* Message header to synchronize regions */
77 struct mp_region_msg {
78 char port_name[RTE_DEV_NAME_MAX_LEN];
79 memif_region_index_t idx;
80 memif_region_size_t size;
84 memif_mp_send_region(const struct rte_mp_msg *msg, const void *peer)
86 struct rte_eth_dev *dev;
87 struct pmd_process_private *proc_private;
88 const struct mp_region_msg *msg_param = (const struct mp_region_msg *)msg->param;
89 struct rte_mp_msg reply;
90 struct mp_region_msg *reply_param = (struct mp_region_msg *)reply.param;
92 /* Get requested port */
93 dev = rte_eth_dev_get_by_name(msg_param->port_name);
95 MIF_LOG(ERR, "Failed to get port id for %s",
96 msg_param->port_name);
99 proc_private = dev->process_private;
101 memset(&reply, 0, sizeof(reply));
102 strlcpy(reply.name, msg->name, sizeof(reply.name));
103 reply_param->idx = msg_param->idx;
104 if (proc_private->regions[msg_param->idx] != NULL) {
105 reply_param->size = proc_private->regions[msg_param->idx]->region_size;
106 reply.fds[0] = proc_private->regions[msg_param->idx]->fd;
109 reply.len_param = sizeof(*reply_param);
110 if (rte_mp_reply(&reply, peer) < 0) {
111 MIF_LOG(ERR, "Failed to reply to an add region request");
120 * Called by secondary process, when ports link status goes up.
123 memif_mp_request_regions(struct rte_eth_dev *dev)
126 struct timespec timeout = {.tv_sec = 5, .tv_nsec = 0};
127 struct rte_mp_msg msg, *reply;
128 struct rte_mp_reply replies;
129 struct mp_region_msg *msg_param = (struct mp_region_msg *)msg.param;
130 struct mp_region_msg *reply_param;
131 struct memif_region *r;
132 struct pmd_process_private *proc_private = dev->process_private;
133 struct pmd_internals *pmd = dev->data->dev_private;
134 /* in case of zero-copy client, only request region 0 */
135 uint16_t max_region_num = (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) ?
136 1 : ETH_MEMIF_MAX_REGION_NUM;
138 MIF_LOG(DEBUG, "Requesting memory regions");
140 for (i = 0; i < max_region_num; i++) {
141 /* Prepare the message */
142 memset(&msg, 0, sizeof(msg));
143 strlcpy(msg.name, MEMIF_MP_SEND_REGION, sizeof(msg.name));
144 strlcpy(msg_param->port_name, dev->data->name,
145 sizeof(msg_param->port_name));
147 msg.len_param = sizeof(*msg_param);
150 ret = rte_mp_request_sync(&msg, &replies, &timeout);
151 if (ret < 0 || replies.nb_received != 1) {
152 MIF_LOG(ERR, "Failed to send mp msg: %d",
157 reply = &replies.msgs[0];
158 reply_param = (struct mp_region_msg *)reply->param;
160 if (reply_param->size > 0) {
161 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
163 MIF_LOG(ERR, "Failed to alloc memif region.");
167 r->region_size = reply_param->size;
168 if (reply->num_fds < 1) {
169 MIF_LOG(ERR, "Missing file descriptor.");
173 r->fd = reply->fds[0];
176 proc_private->regions[reply_param->idx] = r;
177 proc_private->regions_num++;
182 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
183 ret = rte_memseg_walk(memif_region_init_zc, (void *)proc_private);
188 return memif_connect(dev);
192 memif_dev_info(struct rte_eth_dev *dev __rte_unused, struct rte_eth_dev_info *dev_info)
194 dev_info->max_mac_addrs = 1;
195 dev_info->max_rx_pktlen = RTE_ETHER_MAX_LEN;
196 dev_info->max_rx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
197 dev_info->max_tx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
198 dev_info->min_rx_bufsize = 0;
199 dev_info->tx_offload_capa = RTE_ETH_TX_OFFLOAD_MULTI_SEGS;
204 static memif_ring_t *
205 memif_get_ring(struct pmd_internals *pmd, struct pmd_process_private *proc_private,
206 memif_ring_type_t type, uint16_t ring_num)
208 /* rings only in region 0 */
209 void *p = proc_private->regions[0]->addr;
210 int ring_size = sizeof(memif_ring_t) + sizeof(memif_desc_t) *
211 (1 << pmd->run.log2_ring_size);
213 p = (uint8_t *)p + (ring_num + type * pmd->run.num_c2s_rings) * ring_size;
215 return (memif_ring_t *)p;
218 static memif_region_offset_t
219 memif_get_ring_offset(struct rte_eth_dev *dev, struct memif_queue *mq,
220 memif_ring_type_t type, uint16_t num)
222 struct pmd_internals *pmd = dev->data->dev_private;
223 struct pmd_process_private *proc_private = dev->process_private;
225 return ((uint8_t *)memif_get_ring(pmd, proc_private, type, num) -
226 (uint8_t *)proc_private->regions[mq->region]->addr);
229 static memif_ring_t *
230 memif_get_ring_from_queue(struct pmd_process_private *proc_private,
231 struct memif_queue *mq)
233 struct memif_region *r;
235 r = proc_private->regions[mq->region];
239 return (memif_ring_t *)((uint8_t *)r->addr + mq->ring_offset);
243 memif_get_buffer(struct pmd_process_private *proc_private, memif_desc_t *d)
245 return ((uint8_t *)proc_private->regions[d->region]->addr + d->offset);
248 /* Free mbufs received by server */
250 memif_free_stored_mbufs(struct pmd_process_private *proc_private, struct memif_queue *mq)
253 uint16_t mask = (1 << mq->log2_ring_size) - 1;
254 memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
256 /* FIXME: improve performance */
257 /* The ring->tail acts as a guard variable between Tx and Rx
258 * threads, so using load-acquire pairs with store-release
259 * in function eth_memif_rx for C2S queues.
261 cur_tail = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
262 while (mq->last_tail != cur_tail) {
263 RTE_MBUF_PREFETCH_TO_FREE(mq->buffers[(mq->last_tail + 1) & mask]);
264 /* Decrement refcnt and free mbuf. (current segment) */
265 rte_mbuf_refcnt_update(mq->buffers[mq->last_tail & mask], -1);
266 rte_pktmbuf_free_seg(mq->buffers[mq->last_tail & mask]);
272 memif_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *cur_tail,
273 struct rte_mbuf *tail)
275 /* Check for number-of-segments-overflow */
276 if (unlikely(head->nb_segs + tail->nb_segs > RTE_MBUF_MAX_NB_SEGS))
279 /* Chain 'tail' onto the old tail */
280 cur_tail->next = tail;
282 /* accumulate number of segments and total length. */
283 head->nb_segs = (uint16_t)(head->nb_segs + tail->nb_segs);
285 tail->pkt_len = tail->data_len;
286 head->pkt_len += tail->pkt_len;
292 eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
294 struct memif_queue *mq = queue;
295 struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
296 struct pmd_process_private *proc_private =
297 rte_eth_devices[mq->in_port].process_private;
298 memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
299 uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0;
300 uint16_t n_rx_pkts = 0;
301 uint16_t mbuf_size = rte_pktmbuf_data_room_size(mq->mempool) -
302 RTE_PKTMBUF_HEADROOM;
303 uint16_t src_len, src_off, dst_len, dst_off, cp_len;
304 memif_ring_type_t type = mq->type;
306 struct rte_mbuf *mbuf, *mbuf_head, *mbuf_tail;
308 ssize_t size __rte_unused;
311 struct rte_eth_link link;
313 if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
315 if (unlikely(ring == NULL)) {
316 /* Secondary process will attempt to request regions. */
317 ret = rte_eth_link_get(mq->in_port, &link);
319 MIF_LOG(ERR, "Failed to get port %u link info: %s",
320 mq->in_port, rte_strerror(-ret));
324 /* consume interrupt */
325 if (((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) &&
326 (rte_intr_fd_get(mq->intr_handle) >= 0))
327 size = read(rte_intr_fd_get(mq->intr_handle), &b,
330 ring_size = 1 << mq->log2_ring_size;
331 mask = ring_size - 1;
333 if (type == MEMIF_RING_C2S) {
334 cur_slot = mq->last_head;
335 last_slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
337 cur_slot = mq->last_tail;
338 last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
341 if (cur_slot == last_slot)
343 n_slots = last_slot - cur_slot;
345 while (n_slots && n_rx_pkts < nb_pkts) {
346 mbuf_head = rte_pktmbuf_alloc(mq->mempool);
347 if (unlikely(mbuf_head == NULL))
350 mbuf->port = mq->in_port;
354 s0 = cur_slot & mask;
355 d0 = &ring->desc[s0];
357 src_len = d0->length;
361 dst_len = mbuf_size - dst_off;
366 /* store pointer to tail */
368 mbuf = rte_pktmbuf_alloc(mq->mempool);
369 if (unlikely(mbuf == NULL))
371 mbuf->port = mq->in_port;
372 ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
373 if (unlikely(ret < 0)) {
374 MIF_LOG(ERR, "number-of-segments-overflow");
375 rte_pktmbuf_free(mbuf);
379 cp_len = RTE_MIN(dst_len, src_len);
381 rte_pktmbuf_data_len(mbuf) += cp_len;
382 rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
383 if (mbuf != mbuf_head)
384 rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
386 rte_memcpy(rte_pktmbuf_mtod_offset(mbuf, void *,
388 (uint8_t *)memif_get_buffer(proc_private, d0) +
399 if (d0->flags & MEMIF_DESC_FLAG_NEXT)
402 mq->n_bytes += rte_pktmbuf_pkt_len(mbuf_head);
408 if (type == MEMIF_RING_C2S) {
409 __atomic_store_n(&ring->tail, cur_slot, __ATOMIC_RELEASE);
410 mq->last_head = cur_slot;
412 mq->last_tail = cur_slot;
416 if (type == MEMIF_RING_S2C) {
417 /* ring->head is updated by the receiver and this function
418 * is called in the context of receiver thread. The loads in
419 * the receiver do not need to synchronize with its own stores.
421 head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
422 n_slots = ring_size - head + mq->last_tail;
426 d0 = &ring->desc[s0];
427 d0->length = pmd->run.pkt_buffer_size;
429 __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
432 mq->n_pkts += n_rx_pkts;
437 eth_memif_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
439 struct memif_queue *mq = queue;
440 struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
441 struct pmd_process_private *proc_private =
442 rte_eth_devices[mq->in_port].process_private;
443 memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
444 uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0, head;
445 uint16_t n_rx_pkts = 0;
447 struct rte_mbuf *mbuf, *mbuf_tail;
448 struct rte_mbuf *mbuf_head = NULL;
450 struct rte_eth_link link;
452 if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
454 if (unlikely(ring == NULL)) {
455 /* Secondary process will attempt to request regions. */
456 rte_eth_link_get(mq->in_port, &link);
460 /* consume interrupt */
461 if ((rte_intr_fd_get(mq->intr_handle) >= 0) &&
462 ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0)) {
464 ssize_t size __rte_unused;
465 size = read(rte_intr_fd_get(mq->intr_handle), &b,
469 ring_size = 1 << mq->log2_ring_size;
470 mask = ring_size - 1;
472 cur_slot = mq->last_tail;
473 /* The ring->tail acts as a guard variable between Tx and Rx
474 * threads, so using load-acquire pairs with store-release
475 * to synchronize it between threads.
477 last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
478 if (cur_slot == last_slot)
480 n_slots = last_slot - cur_slot;
482 while (n_slots && n_rx_pkts < nb_pkts) {
483 s0 = cur_slot & mask;
485 d0 = &ring->desc[s0];
486 mbuf_head = mq->buffers[s0];
490 /* prefetch next descriptor */
491 if (n_rx_pkts + 1 < nb_pkts)
492 rte_prefetch0(&ring->desc[(cur_slot + 1) & mask]);
494 mbuf->port = mq->in_port;
495 rte_pktmbuf_data_len(mbuf) = d0->length;
496 rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
498 mq->n_bytes += rte_pktmbuf_data_len(mbuf);
502 if (d0->flags & MEMIF_DESC_FLAG_NEXT) {
503 s0 = cur_slot & mask;
504 d0 = &ring->desc[s0];
506 mbuf = mq->buffers[s0];
507 ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
508 if (unlikely(ret < 0)) {
509 MIF_LOG(ERR, "number-of-segments-overflow");
519 mq->last_tail = cur_slot;
521 /* Supply server with new buffers */
523 /* ring->head is updated by the receiver and this function
524 * is called in the context of receiver thread. The loads in
525 * the receiver do not need to synchronize with its own stores.
527 head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
528 n_slots = ring_size - head + mq->last_tail;
533 ret = rte_pktmbuf_alloc_bulk(mq->mempool, &mq->buffers[head & mask], n_slots);
534 if (unlikely(ret < 0))
540 rte_prefetch0(mq->buffers[head & mask]);
541 d0 = &ring->desc[s0];
542 /* store buffer header */
543 mbuf = mq->buffers[s0];
544 /* populate descriptor */
545 d0->length = rte_pktmbuf_data_room_size(mq->mempool) -
546 RTE_PKTMBUF_HEADROOM;
548 d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
549 (uint8_t *)proc_private->regions[d0->region]->addr;
552 /* The ring->head acts as a guard variable between Tx and Rx
553 * threads, so using store-release pairs with load-acquire
554 * in function eth_memif_tx.
556 __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
558 mq->n_pkts += n_rx_pkts;
564 eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
566 struct memif_queue *mq = queue;
567 struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
568 struct pmd_process_private *proc_private =
569 rte_eth_devices[mq->in_port].process_private;
570 memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
571 uint16_t slot, saved_slot, n_free, ring_size, mask, n_tx_pkts = 0;
572 uint16_t src_len, src_off, dst_len, dst_off, cp_len, nb_segs;
573 memif_ring_type_t type = mq->type;
575 struct rte_mbuf *mbuf;
576 struct rte_mbuf *mbuf_head;
579 struct rte_eth_link link;
581 if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
583 if (unlikely(ring == NULL)) {
586 /* Secondary process will attempt to request regions. */
587 ret = rte_eth_link_get(mq->in_port, &link);
589 MIF_LOG(ERR, "Failed to get port %u link info: %s",
590 mq->in_port, rte_strerror(-ret));
594 ring_size = 1 << mq->log2_ring_size;
595 mask = ring_size - 1;
597 if (type == MEMIF_RING_C2S) {
598 /* For C2S queues ring->head is updated by the sender and
599 * this function is called in the context of sending thread.
600 * The loads in the sender do not need to synchronize with
601 * its own stores. Hence, the following load can be a
604 slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
605 n_free = ring_size - slot +
606 __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
608 /* For S2C queues ring->tail is updated by the sender and
609 * this function is called in the context of sending thread.
610 * The loads in the sender do not need to synchronize with
611 * its own stores. Hence, the following load can be a
614 slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
615 n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
618 while (n_tx_pkts < nb_pkts && n_free) {
620 nb_segs = mbuf_head->nb_segs;
624 d0 = &ring->desc[slot & mask];
626 dst_len = (type == MEMIF_RING_C2S) ?
627 pmd->run.pkt_buffer_size : d0->length;
631 src_len = rte_pktmbuf_data_len(mbuf);
638 d0->flags |= MEMIF_DESC_FLAG_NEXT;
639 d0 = &ring->desc[slot & mask];
641 dst_len = (type == MEMIF_RING_C2S) ?
642 pmd->run.pkt_buffer_size : d0->length;
649 cp_len = RTE_MIN(dst_len, src_len);
651 rte_memcpy((uint8_t *)memif_get_buffer(proc_private,
653 rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
656 mq->n_bytes += cp_len;
662 d0->length = dst_off;
673 rte_pktmbuf_free(mbuf_head);
677 if (type == MEMIF_RING_C2S)
678 __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
680 __atomic_store_n(&ring->tail, slot, __ATOMIC_RELEASE);
682 if (((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) &&
683 (rte_intr_fd_get(mq->intr_handle) >= 0)) {
685 size = write(rte_intr_fd_get(mq->intr_handle), &a,
687 if (unlikely(size < 0)) {
689 "Failed to send interrupt. %s", strerror(errno));
693 mq->n_pkts += n_tx_pkts;
699 memif_tx_one_zc(struct pmd_process_private *proc_private, struct memif_queue *mq,
700 memif_ring_t *ring, struct rte_mbuf *mbuf, const uint16_t mask,
701 uint16_t slot, uint16_t n_free)
704 uint16_t nb_segs = mbuf->nb_segs;
708 /* store pointer to mbuf to free it later */
709 mq->buffers[slot & mask] = mbuf;
710 /* Increment refcnt to make sure the buffer is not freed before server
711 * receives it. (current segment)
713 rte_mbuf_refcnt_update(mbuf, 1);
714 /* populate descriptor */
715 d0 = &ring->desc[slot & mask];
716 d0->length = rte_pktmbuf_data_len(mbuf);
717 mq->n_bytes += rte_pktmbuf_data_len(mbuf);
718 /* FIXME: get region index */
720 d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
721 (uint8_t *)proc_private->regions[d0->region]->addr;
724 /* check if buffer is chained */
728 /* mark buffer as chained */
729 d0->flags |= MEMIF_DESC_FLAG_NEXT;
732 /* update counters */
742 eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
744 struct memif_queue *mq = queue;
745 struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
746 struct pmd_process_private *proc_private =
747 rte_eth_devices[mq->in_port].process_private;
748 memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
749 uint16_t slot, n_free, ring_size, mask, n_tx_pkts = 0;
750 struct rte_eth_link link;
752 if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
754 if (unlikely(ring == NULL)) {
755 /* Secondary process will attempt to request regions. */
756 rte_eth_link_get(mq->in_port, &link);
760 ring_size = 1 << mq->log2_ring_size;
761 mask = ring_size - 1;
763 /* free mbufs received by server */
764 memif_free_stored_mbufs(proc_private, mq);
766 /* ring type always MEMIF_RING_C2S */
767 /* For C2S queues ring->head is updated by the sender and
768 * this function is called in the context of sending thread.
769 * The loads in the sender do not need to synchronize with
770 * its own stores. Hence, the following load can be a
773 slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
774 n_free = ring_size - slot + mq->last_tail;
778 while (n_free && (n_tx_pkts < nb_pkts)) {
779 while ((n_free > 4) && ((nb_pkts - n_tx_pkts) > 4)) {
780 if ((nb_pkts - n_tx_pkts) > 8) {
781 rte_prefetch0(*bufs + 4);
782 rte_prefetch0(*bufs + 5);
783 rte_prefetch0(*bufs + 6);
784 rte_prefetch0(*bufs + 7);
786 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
788 if (unlikely(used_slots < 1))
792 n_free -= used_slots;
794 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
796 if (unlikely(used_slots < 1))
800 n_free -= used_slots;
802 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
804 if (unlikely(used_slots < 1))
808 n_free -= used_slots;
810 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
812 if (unlikely(used_slots < 1))
816 n_free -= used_slots;
818 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
820 if (unlikely(used_slots < 1))
824 n_free -= used_slots;
828 /* ring type always MEMIF_RING_C2S */
829 /* The ring->head acts as a guard variable between Tx and Rx
830 * threads, so using store-release pairs with load-acquire
831 * in function eth_memif_rx for C2S rings.
833 __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
835 /* Send interrupt, if enabled. */
836 if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
838 if (rte_intr_fd_get(mq->intr_handle) < 0)
841 ssize_t size = write(rte_intr_fd_get(mq->intr_handle),
843 if (unlikely(size < 0)) {
845 "Failed to send interrupt. %s", strerror(errno));
849 /* increment queue counters */
850 mq->n_pkts += n_tx_pkts;
856 memif_free_regions(struct rte_eth_dev *dev)
858 struct pmd_process_private *proc_private = dev->process_private;
859 struct pmd_internals *pmd = dev->data->dev_private;
861 struct memif_region *r;
863 /* regions are allocated contiguously, so it's
864 * enough to loop until 'proc_private->regions_num'
866 for (i = 0; i < proc_private->regions_num; i++) {
867 r = proc_private->regions[i];
869 /* This is memzone */
870 if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
875 if (r->addr != NULL) {
876 munmap(r->addr, r->region_size);
883 proc_private->regions[i] = NULL;
886 proc_private->regions_num = 0;
890 memif_region_init_zc(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
893 struct pmd_process_private *proc_private = (struct pmd_process_private *)arg;
894 struct memif_region *r;
896 if (proc_private->regions_num < 1) {
897 MIF_LOG(ERR, "Missing descriptor region");
901 r = proc_private->regions[proc_private->regions_num - 1];
903 if (r->addr != msl->base_va)
904 r = proc_private->regions[++proc_private->regions_num - 1];
907 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
909 MIF_LOG(ERR, "Failed to alloc memif region.");
913 r->addr = msl->base_va;
914 r->region_size = ms->len;
915 r->fd = rte_memseg_get_fd(ms);
918 r->pkt_buffer_offset = 0;
920 proc_private->regions[proc_private->regions_num - 1] = r;
922 r->region_size += ms->len;
929 memif_region_init_shm(struct rte_eth_dev *dev, uint8_t has_buffers)
931 struct pmd_internals *pmd = dev->data->dev_private;
932 struct pmd_process_private *proc_private = dev->process_private;
933 char shm_name[ETH_MEMIF_SHM_NAME_SIZE];
935 struct memif_region *r;
937 if (proc_private->regions_num >= ETH_MEMIF_MAX_REGION_NUM) {
938 MIF_LOG(ERR, "Too many regions.");
942 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
944 MIF_LOG(ERR, "Failed to alloc memif region.");
948 /* calculate buffer offset */
949 r->pkt_buffer_offset = (pmd->run.num_c2s_rings + pmd->run.num_s2c_rings) *
950 (sizeof(memif_ring_t) + sizeof(memif_desc_t) *
951 (1 << pmd->run.log2_ring_size));
953 r->region_size = r->pkt_buffer_offset;
954 /* if region has buffers, add buffers size to region_size */
955 if (has_buffers == 1)
956 r->region_size += (uint32_t)(pmd->run.pkt_buffer_size *
957 (1 << pmd->run.log2_ring_size) *
958 (pmd->run.num_c2s_rings +
959 pmd->run.num_s2c_rings));
961 memset(shm_name, 0, sizeof(char) * ETH_MEMIF_SHM_NAME_SIZE);
962 snprintf(shm_name, ETH_MEMIF_SHM_NAME_SIZE, "memif_region_%d",
963 proc_private->regions_num);
965 r->fd = memfd_create(shm_name, MFD_ALLOW_SEALING);
967 MIF_LOG(ERR, "Failed to create shm file: %s.", strerror(errno));
972 ret = fcntl(r->fd, F_ADD_SEALS, F_SEAL_SHRINK);
974 MIF_LOG(ERR, "Failed to add seals to shm file: %s.", strerror(errno));
978 ret = ftruncate(r->fd, r->region_size);
980 MIF_LOG(ERR, "Failed to truncate shm file: %s.", strerror(errno));
984 r->addr = mmap(NULL, r->region_size, PROT_READ |
985 PROT_WRITE, MAP_SHARED, r->fd, 0);
986 if (r->addr == MAP_FAILED) {
987 MIF_LOG(ERR, "Failed to mmap shm region: %s.", strerror(ret));
992 proc_private->regions[proc_private->regions_num] = r;
993 proc_private->regions_num++;
1006 memif_regions_init(struct rte_eth_dev *dev)
1008 struct pmd_internals *pmd = dev->data->dev_private;
1012 * Zero-copy exposes dpdk memory.
1013 * Each memseg list will be represented by memif region.
1014 * Zero-copy regions indexing: memseg list idx + 1,
1015 * as we already have region 0 reserved for descriptors.
1017 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1018 /* create region idx 0 containing descriptors */
1019 ret = memif_region_init_shm(dev, 0);
1022 ret = rte_memseg_walk(memif_region_init_zc, (void *)dev->process_private);
1026 /* create one memory region containing rings and buffers */
1027 ret = memif_region_init_shm(dev, /* has buffers */ 1);
1036 memif_init_rings(struct rte_eth_dev *dev)
1038 struct pmd_internals *pmd = dev->data->dev_private;
1039 struct pmd_process_private *proc_private = dev->process_private;
1044 for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1045 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_C2S, i);
1046 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1047 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1048 ring->cookie = MEMIF_COOKIE;
1051 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1054 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1055 slot = i * (1 << pmd->run.log2_ring_size) + j;
1056 ring->desc[j].region = 0;
1057 ring->desc[j].offset =
1058 proc_private->regions[0]->pkt_buffer_offset +
1059 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1060 ring->desc[j].length = pmd->run.pkt_buffer_size;
1064 for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1065 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_S2C, i);
1066 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1067 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1068 ring->cookie = MEMIF_COOKIE;
1071 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1074 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1075 slot = (i + pmd->run.num_c2s_rings) *
1076 (1 << pmd->run.log2_ring_size) + j;
1077 ring->desc[j].region = 0;
1078 ring->desc[j].offset =
1079 proc_private->regions[0]->pkt_buffer_offset +
1080 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1081 ring->desc[j].length = pmd->run.pkt_buffer_size;
1086 /* called only by client */
1088 memif_init_queues(struct rte_eth_dev *dev)
1090 struct pmd_internals *pmd = dev->data->dev_private;
1091 struct memif_queue *mq;
1094 for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1095 mq = dev->data->tx_queues[i];
1096 mq->log2_ring_size = pmd->run.log2_ring_size;
1097 /* queues located only in region 0 */
1099 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_C2S, i);
1102 if (rte_intr_fd_set(mq->intr_handle, eventfd(0, EFD_NONBLOCK)))
1105 if (rte_intr_fd_get(mq->intr_handle) < 0) {
1107 "Failed to create eventfd for tx queue %d: %s.", i,
1111 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1112 mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1113 (1 << mq->log2_ring_size), 0);
1114 if (mq->buffers == NULL)
1119 for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1120 mq = dev->data->rx_queues[i];
1121 mq->log2_ring_size = pmd->run.log2_ring_size;
1122 /* queues located only in region 0 */
1124 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_S2C, i);
1127 if (rte_intr_fd_set(mq->intr_handle, eventfd(0, EFD_NONBLOCK)))
1129 if (rte_intr_fd_get(mq->intr_handle) < 0) {
1131 "Failed to create eventfd for rx queue %d: %s.", i,
1135 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1136 mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1137 (1 << mq->log2_ring_size), 0);
1138 if (mq->buffers == NULL)
1146 memif_init_regions_and_queues(struct rte_eth_dev *dev)
1150 ret = memif_regions_init(dev);
1154 memif_init_rings(dev);
1156 ret = memif_init_queues(dev);
1164 memif_connect(struct rte_eth_dev *dev)
1166 struct pmd_internals *pmd = dev->data->dev_private;
1167 struct pmd_process_private *proc_private = dev->process_private;
1168 struct memif_region *mr;
1169 struct memif_queue *mq;
1173 for (i = 0; i < proc_private->regions_num; i++) {
1174 mr = proc_private->regions[i];
1176 if (mr->addr == NULL) {
1179 mr->addr = mmap(NULL, mr->region_size,
1180 PROT_READ | PROT_WRITE,
1181 MAP_SHARED, mr->fd, 0);
1182 if (mr->addr == MAP_FAILED) {
1183 MIF_LOG(ERR, "mmap failed: %s\n",
1188 if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
1189 /* close memseg file */
1196 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1197 for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1198 mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1199 dev->data->tx_queues[i] : dev->data->rx_queues[i];
1200 ring = memif_get_ring_from_queue(proc_private, mq);
1201 if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1202 MIF_LOG(ERR, "Wrong ring");
1205 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1206 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1209 /* enable polling mode */
1210 if (pmd->role == MEMIF_ROLE_SERVER)
1211 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1213 for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1214 mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1215 dev->data->rx_queues[i] : dev->data->tx_queues[i];
1216 ring = memif_get_ring_from_queue(proc_private, mq);
1217 if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1218 MIF_LOG(ERR, "Wrong ring");
1221 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1222 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1225 /* enable polling mode */
1226 if (pmd->role == MEMIF_ROLE_CLIENT)
1227 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1230 pmd->flags &= ~ETH_MEMIF_FLAG_CONNECTING;
1231 pmd->flags |= ETH_MEMIF_FLAG_CONNECTED;
1232 dev->data->dev_link.link_status = RTE_ETH_LINK_UP;
1234 MIF_LOG(INFO, "Connected.");
1239 memif_dev_start(struct rte_eth_dev *dev)
1241 struct pmd_internals *pmd = dev->data->dev_private;
1244 switch (pmd->role) {
1245 case MEMIF_ROLE_CLIENT:
1246 ret = memif_connect_client(dev);
1248 case MEMIF_ROLE_SERVER:
1249 ret = memif_connect_server(dev);
1252 MIF_LOG(ERR, "Unknown role: %d.", pmd->role);
1261 memif_dev_stop(struct rte_eth_dev *dev)
1263 memif_disconnect(dev);
1268 memif_dev_close(struct rte_eth_dev *dev)
1270 struct pmd_internals *pmd = dev->data->dev_private;
1273 if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1274 memif_msg_enq_disconnect(pmd->cc, "Device closed", 0);
1276 for (i = 0; i < dev->data->nb_rx_queues; i++)
1277 (*dev->dev_ops->rx_queue_release)(dev, i);
1278 for (i = 0; i < dev->data->nb_tx_queues; i++)
1279 (*dev->dev_ops->tx_queue_release)(dev, i);
1281 memif_socket_remove_device(dev);
1284 rte_free(dev->process_private);
1290 memif_dev_configure(struct rte_eth_dev *dev)
1292 struct pmd_internals *pmd = dev->data->dev_private;
1298 pmd->cfg.num_c2s_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1299 dev->data->nb_tx_queues : dev->data->nb_rx_queues;
1305 pmd->cfg.num_s2c_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1306 dev->data->nb_rx_queues : dev->data->nb_tx_queues;
1312 memif_tx_queue_setup(struct rte_eth_dev *dev,
1314 uint16_t nb_tx_desc __rte_unused,
1315 unsigned int socket_id __rte_unused,
1316 const struct rte_eth_txconf *tx_conf __rte_unused)
1318 struct pmd_internals *pmd = dev->data->dev_private;
1319 struct memif_queue *mq;
1321 mq = rte_zmalloc("tx-queue", sizeof(struct memif_queue), 0);
1323 MIF_LOG(ERR, "Failed to allocate tx queue id: %u", qid);
1327 /* Allocate interrupt instance */
1328 mq->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
1329 if (mq->intr_handle == NULL) {
1330 MIF_LOG(ERR, "Failed to allocate intr handle");
1335 (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_C2S : MEMIF_RING_S2C;
1339 if (rte_intr_fd_set(mq->intr_handle, -1))
1342 if (rte_intr_type_set(mq->intr_handle, RTE_INTR_HANDLE_EXT))
1345 mq->in_port = dev->data->port_id;
1346 dev->data->tx_queues[qid] = mq;
1352 memif_rx_queue_setup(struct rte_eth_dev *dev,
1354 uint16_t nb_rx_desc __rte_unused,
1355 unsigned int socket_id __rte_unused,
1356 const struct rte_eth_rxconf *rx_conf __rte_unused,
1357 struct rte_mempool *mb_pool)
1359 struct pmd_internals *pmd = dev->data->dev_private;
1360 struct memif_queue *mq;
1362 mq = rte_zmalloc("rx-queue", sizeof(struct memif_queue), 0);
1364 MIF_LOG(ERR, "Failed to allocate rx queue id: %u", qid);
1368 /* Allocate interrupt instance */
1369 mq->intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
1370 if (mq->intr_handle == NULL) {
1371 MIF_LOG(ERR, "Failed to allocate intr handle");
1375 mq->type = (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_S2C : MEMIF_RING_C2S;
1379 if (rte_intr_fd_set(mq->intr_handle, -1))
1382 if (rte_intr_type_set(mq->intr_handle, RTE_INTR_HANDLE_EXT))
1385 mq->mempool = mb_pool;
1386 mq->in_port = dev->data->port_id;
1387 dev->data->rx_queues[qid] = mq;
1393 memif_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
1395 struct memif_queue *mq = dev->data->rx_queues[qid];
1400 rte_intr_instance_free(mq->intr_handle);
1405 memif_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
1407 struct memif_queue *mq = dev->data->tx_queues[qid];
1416 memif_link_update(struct rte_eth_dev *dev,
1417 int wait_to_complete __rte_unused)
1419 struct pmd_process_private *proc_private;
1421 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1422 proc_private = dev->process_private;
1423 if (dev->data->dev_link.link_status == RTE_ETH_LINK_UP &&
1424 proc_private->regions_num == 0) {
1425 memif_mp_request_regions(dev);
1426 } else if (dev->data->dev_link.link_status == RTE_ETH_LINK_DOWN &&
1427 proc_private->regions_num > 0) {
1428 memif_free_regions(dev);
1435 memif_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1437 struct pmd_internals *pmd = dev->data->dev_private;
1438 struct memif_queue *mq;
1442 stats->ipackets = 0;
1444 stats->opackets = 0;
1447 tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_c2s_rings :
1448 pmd->run.num_s2c_rings;
1449 nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1450 RTE_ETHDEV_QUEUE_STAT_CNTRS;
1453 for (i = 0; i < nq; i++) {
1454 mq = dev->data->rx_queues[i];
1455 stats->q_ipackets[i] = mq->n_pkts;
1456 stats->q_ibytes[i] = mq->n_bytes;
1457 stats->ipackets += mq->n_pkts;
1458 stats->ibytes += mq->n_bytes;
1461 tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_s2c_rings :
1462 pmd->run.num_c2s_rings;
1463 nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1464 RTE_ETHDEV_QUEUE_STAT_CNTRS;
1467 for (i = 0; i < nq; i++) {
1468 mq = dev->data->tx_queues[i];
1469 stats->q_opackets[i] = mq->n_pkts;
1470 stats->q_obytes[i] = mq->n_bytes;
1471 stats->opackets += mq->n_pkts;
1472 stats->obytes += mq->n_bytes;
1478 memif_stats_reset(struct rte_eth_dev *dev)
1480 struct pmd_internals *pmd = dev->data->dev_private;
1482 struct memif_queue *mq;
1484 for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1485 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->tx_queues[i] :
1486 dev->data->rx_queues[i];
1490 for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1491 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->rx_queues[i] :
1492 dev->data->tx_queues[i];
1500 static const struct eth_dev_ops ops = {
1501 .dev_start = memif_dev_start,
1502 .dev_stop = memif_dev_stop,
1503 .dev_close = memif_dev_close,
1504 .dev_infos_get = memif_dev_info,
1505 .dev_configure = memif_dev_configure,
1506 .tx_queue_setup = memif_tx_queue_setup,
1507 .rx_queue_setup = memif_rx_queue_setup,
1508 .rx_queue_release = memif_rx_queue_release,
1509 .tx_queue_release = memif_tx_queue_release,
1510 .link_update = memif_link_update,
1511 .stats_get = memif_stats_get,
1512 .stats_reset = memif_stats_reset,
1516 memif_create(struct rte_vdev_device *vdev, enum memif_role_t role,
1517 memif_interface_id_t id, uint32_t flags,
1518 const char *socket_filename,
1519 memif_log2_ring_size_t log2_ring_size,
1520 uint16_t pkt_buffer_size, const char *secret,
1521 struct rte_ether_addr *ether_addr)
1524 struct rte_eth_dev *eth_dev;
1525 struct rte_eth_dev_data *data;
1526 struct pmd_internals *pmd;
1527 struct pmd_process_private *process_private;
1528 const unsigned int numa_node = vdev->device.numa_node;
1529 const char *name = rte_vdev_device_name(vdev);
1531 eth_dev = rte_eth_vdev_allocate(vdev, sizeof(*pmd));
1532 if (eth_dev == NULL) {
1533 MIF_LOG(ERR, "%s: Unable to allocate device struct.", name);
1537 process_private = (struct pmd_process_private *)
1538 rte_zmalloc(name, sizeof(struct pmd_process_private),
1539 RTE_CACHE_LINE_SIZE);
1541 if (process_private == NULL) {
1542 MIF_LOG(ERR, "Failed to alloc memory for process private");
1545 eth_dev->process_private = process_private;
1547 pmd = eth_dev->data->dev_private;
1548 memset(pmd, 0, sizeof(*pmd));
1552 pmd->flags |= ETH_MEMIF_FLAG_DISABLED;
1554 /* Zero-copy flag irelevant to server. */
1555 if (pmd->role == MEMIF_ROLE_SERVER)
1556 pmd->flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1558 ret = memif_socket_init(eth_dev, socket_filename);
1562 memset(pmd->secret, 0, sizeof(char) * ETH_MEMIF_SECRET_SIZE);
1564 strlcpy(pmd->secret, secret, sizeof(pmd->secret));
1566 pmd->cfg.log2_ring_size = log2_ring_size;
1567 /* set in .dev_configure() */
1568 pmd->cfg.num_c2s_rings = 0;
1569 pmd->cfg.num_s2c_rings = 0;
1571 pmd->cfg.pkt_buffer_size = pkt_buffer_size;
1572 rte_spinlock_init(&pmd->cc_lock);
1574 data = eth_dev->data;
1575 data->dev_private = pmd;
1576 data->numa_node = numa_node;
1577 data->dev_link = pmd_link;
1578 data->mac_addrs = ether_addr;
1579 data->promiscuous = 1;
1580 data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1582 eth_dev->dev_ops = &ops;
1583 eth_dev->device = &vdev->device;
1584 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1585 eth_dev->rx_pkt_burst = eth_memif_rx_zc;
1586 eth_dev->tx_pkt_burst = eth_memif_tx_zc;
1588 eth_dev->rx_pkt_burst = eth_memif_rx;
1589 eth_dev->tx_pkt_burst = eth_memif_tx;
1592 rte_eth_dev_probing_finish(eth_dev);
1598 memif_set_role(const char *key __rte_unused, const char *value,
1601 enum memif_role_t *role = (enum memif_role_t *)extra_args;
1603 if (strstr(value, "server") != NULL) {
1604 *role = MEMIF_ROLE_SERVER;
1605 } else if (strstr(value, "client") != NULL) {
1606 *role = MEMIF_ROLE_CLIENT;
1607 } else if (strstr(value, "master") != NULL) {
1608 MIF_LOG(NOTICE, "Role argument \"master\" is deprecated, use \"server\"");
1609 *role = MEMIF_ROLE_SERVER;
1610 } else if (strstr(value, "slave") != NULL) {
1611 MIF_LOG(NOTICE, "Role argument \"slave\" is deprecated, use \"client\"");
1612 *role = MEMIF_ROLE_CLIENT;
1614 MIF_LOG(ERR, "Unknown role: %s.", value);
1621 memif_set_zc(const char *key __rte_unused, const char *value, void *extra_args)
1623 uint32_t *flags = (uint32_t *)extra_args;
1625 if (strstr(value, "yes") != NULL) {
1626 if (!rte_mcfg_get_single_file_segments()) {
1627 MIF_LOG(ERR, "Zero-copy doesn't support multi-file segments.");
1630 *flags |= ETH_MEMIF_FLAG_ZERO_COPY;
1631 } else if (strstr(value, "no") != NULL) {
1632 *flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1634 MIF_LOG(ERR, "Failed to parse zero-copy param: %s.", value);
1641 memif_set_id(const char *key __rte_unused, const char *value, void *extra_args)
1643 memif_interface_id_t *id = (memif_interface_id_t *)extra_args;
1645 /* even if parsing fails, 0 is a valid id */
1646 *id = strtoul(value, NULL, 10);
1651 memif_set_bs(const char *key __rte_unused, const char *value, void *extra_args)
1654 uint16_t *pkt_buffer_size = (uint16_t *)extra_args;
1656 tmp = strtoul(value, NULL, 10);
1657 if (tmp == 0 || tmp > 0xFFFF) {
1658 MIF_LOG(ERR, "Invalid buffer size: %s.", value);
1661 *pkt_buffer_size = tmp;
1666 memif_set_rs(const char *key __rte_unused, const char *value, void *extra_args)
1669 memif_log2_ring_size_t *log2_ring_size =
1670 (memif_log2_ring_size_t *)extra_args;
1672 tmp = strtoul(value, NULL, 10);
1673 if (tmp == 0 || tmp > ETH_MEMIF_MAX_LOG2_RING_SIZE) {
1674 MIF_LOG(ERR, "Invalid ring size: %s (max %u).",
1675 value, ETH_MEMIF_MAX_LOG2_RING_SIZE);
1678 *log2_ring_size = tmp;
1682 /* check if directory exists and if we have permission to read/write */
1684 memif_check_socket_filename(const char *filename)
1686 char *dir = NULL, *tmp;
1690 if (strlen(filename) >= MEMIF_SOCKET_UN_SIZE) {
1691 MIF_LOG(ERR, "Unix socket address too long (max 108).");
1695 tmp = strrchr(filename, '/');
1697 idx = tmp - filename;
1698 dir = rte_zmalloc("memif_tmp", sizeof(char) * (idx + 1), 0);
1700 MIF_LOG(ERR, "Failed to allocate memory.");
1703 strlcpy(dir, filename, sizeof(char) * (idx + 1));
1706 if (dir == NULL || (faccessat(-1, dir, F_OK | R_OK |
1707 W_OK, AT_EACCESS) < 0)) {
1708 MIF_LOG(ERR, "Invalid socket directory.");
1718 memif_set_socket_filename(const char *key __rte_unused, const char *value,
1721 const char **socket_filename = (const char **)extra_args;
1723 *socket_filename = value;
1728 memif_set_is_socket_abstract(const char *key __rte_unused, const char *value, void *extra_args)
1730 uint32_t *flags = (uint32_t *)extra_args;
1732 if (strstr(value, "yes") != NULL) {
1733 *flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1734 } else if (strstr(value, "no") != NULL) {
1735 *flags &= ~ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1737 MIF_LOG(ERR, "Failed to parse socket-abstract param: %s.", value);
1744 memif_set_mac(const char *key __rte_unused, const char *value, void *extra_args)
1746 struct rte_ether_addr *ether_addr = (struct rte_ether_addr *)extra_args;
1748 if (rte_ether_unformat_addr(value, ether_addr) < 0)
1749 MIF_LOG(WARNING, "Failed to parse mac '%s'.", value);
1754 memif_set_secret(const char *key __rte_unused, const char *value, void *extra_args)
1756 const char **secret = (const char **)extra_args;
1763 rte_pmd_memif_probe(struct rte_vdev_device *vdev)
1765 RTE_BUILD_BUG_ON(sizeof(memif_msg_t) != 128);
1766 RTE_BUILD_BUG_ON(sizeof(memif_desc_t) != 16);
1768 struct rte_kvargs *kvlist;
1769 const char *name = rte_vdev_device_name(vdev);
1770 enum memif_role_t role = MEMIF_ROLE_CLIENT;
1771 memif_interface_id_t id = 0;
1772 uint16_t pkt_buffer_size = ETH_MEMIF_DEFAULT_PKT_BUFFER_SIZE;
1773 memif_log2_ring_size_t log2_ring_size = ETH_MEMIF_DEFAULT_RING_SIZE;
1774 const char *socket_filename = ETH_MEMIF_DEFAULT_SOCKET_FILENAME;
1776 const char *secret = NULL;
1777 struct rte_ether_addr *ether_addr = rte_zmalloc("",
1778 sizeof(struct rte_ether_addr), 0);
1779 struct rte_eth_dev *eth_dev;
1781 rte_eth_random_addr(ether_addr->addr_bytes);
1783 MIF_LOG(INFO, "Initialize MEMIF: %s.", name);
1785 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1786 eth_dev = rte_eth_dev_attach_secondary(name);
1788 MIF_LOG(ERR, "Failed to probe %s", name);
1792 eth_dev->dev_ops = &ops;
1793 eth_dev->device = &vdev->device;
1794 eth_dev->rx_pkt_burst = eth_memif_rx;
1795 eth_dev->tx_pkt_burst = eth_memif_tx;
1797 if (!rte_eal_primary_proc_alive(NULL)) {
1798 MIF_LOG(ERR, "Primary process is missing");
1802 eth_dev->process_private = (struct pmd_process_private *)
1804 sizeof(struct pmd_process_private),
1805 RTE_CACHE_LINE_SIZE);
1806 if (eth_dev->process_private == NULL) {
1808 "Failed to alloc memory for process private");
1812 rte_eth_dev_probing_finish(eth_dev);
1817 ret = rte_mp_action_register(MEMIF_MP_SEND_REGION, memif_mp_send_region);
1819 * Primary process can continue probing, but secondary process won't
1820 * be able to get memory regions information
1822 if (ret < 0 && rte_errno != EEXIST)
1823 MIF_LOG(WARNING, "Failed to register mp action callback: %s",
1824 strerror(rte_errno));
1826 /* use abstract address by default */
1827 flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1829 kvlist = rte_kvargs_parse(rte_vdev_device_args(vdev), valid_arguments);
1831 /* parse parameters */
1832 if (kvlist != NULL) {
1833 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ROLE_ARG,
1834 &memif_set_role, &role);
1837 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ID_ARG,
1838 &memif_set_id, &id);
1841 ret = rte_kvargs_process(kvlist, ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
1842 &memif_set_bs, &pkt_buffer_size);
1845 ret = rte_kvargs_process(kvlist, ETH_MEMIF_RING_SIZE_ARG,
1846 &memif_set_rs, &log2_ring_size);
1849 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ARG,
1850 &memif_set_socket_filename,
1851 (void *)(&socket_filename));
1854 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ABSTRACT_ARG,
1855 &memif_set_is_socket_abstract, &flags);
1858 ret = rte_kvargs_process(kvlist, ETH_MEMIF_MAC_ARG,
1859 &memif_set_mac, ether_addr);
1862 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ZC_ARG,
1863 &memif_set_zc, &flags);
1866 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SECRET_ARG,
1867 &memif_set_secret, (void *)(&secret));
1872 if (!(flags & ETH_MEMIF_FLAG_SOCKET_ABSTRACT)) {
1873 ret = memif_check_socket_filename(socket_filename);
1878 /* create interface */
1879 ret = memif_create(vdev, role, id, flags, socket_filename,
1880 log2_ring_size, pkt_buffer_size, secret, ether_addr);
1883 rte_kvargs_free(kvlist);
1888 rte_pmd_memif_remove(struct rte_vdev_device *vdev)
1890 struct rte_eth_dev *eth_dev;
1892 eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(vdev));
1893 if (eth_dev == NULL)
1896 return rte_eth_dev_close(eth_dev->data->port_id);
1899 static struct rte_vdev_driver pmd_memif_drv = {
1900 .probe = rte_pmd_memif_probe,
1901 .remove = rte_pmd_memif_remove,
1904 RTE_PMD_REGISTER_VDEV(net_memif, pmd_memif_drv);
1906 RTE_PMD_REGISTER_PARAM_STRING(net_memif,
1907 ETH_MEMIF_ID_ARG "=<int>"
1908 ETH_MEMIF_ROLE_ARG "=server|client"
1909 ETH_MEMIF_PKT_BUFFER_SIZE_ARG "=<int>"
1910 ETH_MEMIF_RING_SIZE_ARG "=<int>"
1911 ETH_MEMIF_SOCKET_ARG "=<string>"
1912 ETH_MEMIF_SOCKET_ABSTRACT_ARG "=yes|no"
1913 ETH_MEMIF_MAC_ARG "=xx:xx:xx:xx:xx:xx"
1914 ETH_MEMIF_ZC_ARG "=yes|no"
1915 ETH_MEMIF_SECRET_ARG "=<string>");
1917 RTE_LOG_REGISTER_DEFAULT(memif_logtype, NOTICE);