1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2019 Intel Corporation.
9 #include <rte_common.h>
10 #include <rte_lcore.h>
11 #include <rte_cycles.h>
16 #include <rte_bus_pci.h>
17 #include <rte_memzone.h>
18 #include <rte_memcpy.h>
19 #include <rte_rawdev.h>
20 #include <rte_rawdev_pmd.h>
22 #include "ntb_hw_intel.h"
23 #include "rte_pmd_ntb.h"
28 static const struct rte_pci_id pci_id_ntb_map[] = {
29 { RTE_PCI_DEVICE(NTB_INTEL_VENDOR_ID, NTB_INTEL_DEV_ID_B2B_SKX) },
30 { .vendor_id = 0, /* sentinel */ },
33 /* Align with enum ntb_xstats_idx */
34 static struct rte_rawdev_xstats_name ntb_xstats_names[] = {
42 #define NTB_XSTATS_NUM RTE_DIM(ntb_xstats_names)
45 ntb_link_cleanup(struct rte_rawdev *dev)
47 struct ntb_hw *hw = dev->dev_private;
50 if (hw->ntb_ops->spad_write == NULL ||
51 hw->ntb_ops->mw_set_trans == NULL) {
52 NTB_LOG(ERR, "Not supported to clean up link.");
56 /* Clean spad registers. */
57 for (i = 0; i < hw->spad_cnt; i++) {
58 status = (*hw->ntb_ops->spad_write)(dev, i, 0, 0);
60 NTB_LOG(ERR, "Failed to clean local spad.");
63 /* Clear mw so that peer cannot access local memory.*/
64 for (i = 0; i < hw->used_mw_num; i++) {
65 status = (*hw->ntb_ops->mw_set_trans)(dev, i, 0, 0);
67 NTB_LOG(ERR, "Failed to clean mw.");
72 ntb_handshake_work(const struct rte_rawdev *dev)
74 struct ntb_hw *hw = dev->dev_private;
78 if (hw->ntb_ops->spad_write == NULL ||
79 hw->ntb_ops->mw_set_trans == NULL) {
80 NTB_LOG(ERR, "Scratchpad/MW setting is not supported.");
84 /* Tell peer the mw info of local side. */
85 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_NUM_MWS, 1, hw->mw_cnt);
88 for (i = 0; i < hw->mw_cnt; i++) {
89 NTB_LOG(INFO, "Local %u mw size: 0x%"PRIx64"", i,
91 val = hw->mw_size[i] >> 32;
92 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_MW0_SZ_H + 2 * i,
97 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_MW0_SZ_L + 2 * i,
103 /* Tell peer about the queue info and map memory to the peer. */
104 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_Q_SZ, 1, hw->queue_size);
107 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_NUM_QPS, 1,
111 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_USED_MWS, 1,
115 for (i = 0; i < hw->used_mw_num; i++) {
116 val = (uint64_t)(size_t)(hw->mz[i]->addr) >> 32;
117 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_MW0_BA_H + 2 * i,
121 val = (uint64_t)(size_t)(hw->mz[i]->addr);
122 ret = (*hw->ntb_ops->spad_write)(dev, SPAD_MW0_BA_L + 2 * i,
128 for (i = 0; i < hw->used_mw_num; i++) {
129 ret = (*hw->ntb_ops->mw_set_trans)(dev, i, hw->mz[i]->iova,
135 /* Ring doorbell 0 to tell peer the device is ready. */
136 ret = (*hw->ntb_ops->peer_db_set)(dev, 0);
144 ntb_dev_intr_handler(void *param)
146 struct rte_rawdev *dev = (struct rte_rawdev *)param;
147 struct ntb_hw *hw = dev->dev_private;
148 uint32_t val_h, val_l;
149 uint64_t peer_mw_size;
150 uint64_t db_bits = 0;
154 if (hw->ntb_ops->db_read == NULL ||
155 hw->ntb_ops->db_clear == NULL ||
156 hw->ntb_ops->peer_db_set == NULL) {
157 NTB_LOG(ERR, "Doorbell is not supported.");
161 db_bits = (*hw->ntb_ops->db_read)(dev);
163 NTB_LOG(ERR, "No doorbells");
165 /* Doorbell 0 is for peer device ready. */
167 NTB_LOG(INFO, "DB0: Peer device is up.");
168 /* Clear received doorbell. */
169 (*hw->ntb_ops->db_clear)(dev, 1);
172 * Peer dev is already up. All mw settings are already done.
178 if (hw->ntb_ops->spad_read == NULL) {
179 NTB_LOG(ERR, "Scratchpad read is not supported.");
183 /* Check if mw setting on the peer is the same as local. */
184 peer_mw_cnt = (*hw->ntb_ops->spad_read)(dev, SPAD_NUM_MWS, 0);
185 if (peer_mw_cnt != hw->mw_cnt) {
186 NTB_LOG(ERR, "Both mw cnt must be the same.");
190 for (i = 0; i < hw->mw_cnt; i++) {
191 val_h = (*hw->ntb_ops->spad_read)
192 (dev, SPAD_MW0_SZ_H + 2 * i, 0);
193 val_l = (*hw->ntb_ops->spad_read)
194 (dev, SPAD_MW0_SZ_L + 2 * i, 0);
195 peer_mw_size = ((uint64_t)val_h << 32) | val_l;
196 NTB_LOG(DEBUG, "Peer %u mw size: 0x%"PRIx64"", i,
198 if (peer_mw_size != hw->mw_size[i]) {
199 NTB_LOG(ERR, "Mw config must be the same.");
207 * Handshake with peer. Spad_write & mw_set_trans only works
208 * when both devices are up. So write spad again when db is
209 * received. And set db again for the later device who may miss
212 if (ntb_handshake_work(dev) < 0) {
213 NTB_LOG(ERR, "Handshake work failed.");
217 /* To get the link info. */
218 if (hw->ntb_ops->get_link_status == NULL) {
219 NTB_LOG(ERR, "Not supported to get link status.");
222 (*hw->ntb_ops->get_link_status)(dev);
223 NTB_LOG(INFO, "Link is up. Link speed: %u. Link width: %u",
224 hw->link_speed, hw->link_width);
228 if (db_bits & (1 << 1)) {
229 NTB_LOG(INFO, "DB1: Peer device is down.");
230 /* Clear received doorbell. */
231 (*hw->ntb_ops->db_clear)(dev, 2);
233 /* Peer device will be down, So clean local side too. */
234 ntb_link_cleanup(dev);
237 /* Response peer's dev_stop request. */
238 (*hw->ntb_ops->peer_db_set)(dev, 2);
242 if (db_bits & (1 << 2)) {
243 NTB_LOG(INFO, "DB2: Peer device agrees dev to be down.");
244 /* Clear received doorbell. */
245 (*hw->ntb_ops->db_clear)(dev, (1 << 2));
252 ntb_queue_conf_get(struct rte_rawdev *dev,
254 rte_rawdev_obj_t queue_conf)
256 struct ntb_queue_conf *q_conf = queue_conf;
257 struct ntb_hw *hw = dev->dev_private;
259 q_conf->tx_free_thresh = hw->tx_queues[queue_id]->tx_free_thresh;
260 q_conf->nb_desc = hw->rx_queues[queue_id]->nb_rx_desc;
261 q_conf->rx_mp = hw->rx_queues[queue_id]->mpool;
265 ntb_rxq_release_mbufs(struct ntb_rx_queue *q)
269 if (!q || !q->sw_ring) {
270 NTB_LOG(ERR, "Pointer to rxq or sw_ring is NULL");
274 for (i = 0; i < q->nb_rx_desc; i++) {
275 if (q->sw_ring[i].mbuf) {
276 rte_pktmbuf_free_seg(q->sw_ring[i].mbuf);
277 q->sw_ring[i].mbuf = NULL;
283 ntb_rxq_release(struct ntb_rx_queue *rxq)
286 NTB_LOG(ERR, "Pointer to rxq is NULL");
290 ntb_rxq_release_mbufs(rxq);
292 rte_free(rxq->sw_ring);
297 ntb_rxq_setup(struct rte_rawdev *dev,
299 rte_rawdev_obj_t queue_conf)
301 struct ntb_queue_conf *rxq_conf = queue_conf;
302 struct ntb_hw *hw = dev->dev_private;
303 struct ntb_rx_queue *rxq;
305 /* Allocate the rx queue data structure */
306 rxq = rte_zmalloc_socket("ntb rx queue",
307 sizeof(struct ntb_rx_queue),
311 NTB_LOG(ERR, "Failed to allocate memory for "
312 "rx queue data structure.");
316 if (rxq_conf->rx_mp == NULL) {
317 NTB_LOG(ERR, "Invalid null mempool pointer.");
320 rxq->nb_rx_desc = rxq_conf->nb_desc;
321 rxq->mpool = rxq_conf->rx_mp;
322 rxq->port_id = dev->dev_id;
323 rxq->queue_id = qp_id;
326 /* Allocate the software ring. */
328 rte_zmalloc_socket("ntb rx sw ring",
329 sizeof(struct ntb_rx_entry) *
334 ntb_rxq_release(rxq);
336 NTB_LOG(ERR, "Failed to allocate memory for SW ring");
340 hw->rx_queues[qp_id] = rxq;
346 ntb_txq_release_mbufs(struct ntb_tx_queue *q)
350 if (!q || !q->sw_ring) {
351 NTB_LOG(ERR, "Pointer to txq or sw_ring is NULL");
355 for (i = 0; i < q->nb_tx_desc; i++) {
356 if (q->sw_ring[i].mbuf) {
357 rte_pktmbuf_free_seg(q->sw_ring[i].mbuf);
358 q->sw_ring[i].mbuf = NULL;
364 ntb_txq_release(struct ntb_tx_queue *txq)
367 NTB_LOG(ERR, "Pointer to txq is NULL");
371 ntb_txq_release_mbufs(txq);
373 rte_free(txq->sw_ring);
378 ntb_txq_setup(struct rte_rawdev *dev,
380 rte_rawdev_obj_t queue_conf)
382 struct ntb_queue_conf *txq_conf = queue_conf;
383 struct ntb_hw *hw = dev->dev_private;
384 struct ntb_tx_queue *txq;
387 /* Allocate the TX queue data structure. */
388 txq = rte_zmalloc_socket("ntb tx queue",
389 sizeof(struct ntb_tx_queue),
393 NTB_LOG(ERR, "Failed to allocate memory for "
394 "tx queue structure");
398 txq->nb_tx_desc = txq_conf->nb_desc;
399 txq->port_id = dev->dev_id;
400 txq->queue_id = qp_id;
403 /* Allocate software ring */
405 rte_zmalloc_socket("ntb tx sw ring",
406 sizeof(struct ntb_tx_entry) *
411 ntb_txq_release(txq);
413 NTB_LOG(ERR, "Failed to allocate memory for SW TX ring");
417 prev = txq->nb_tx_desc - 1;
418 for (i = 0; i < txq->nb_tx_desc; i++) {
419 txq->sw_ring[i].mbuf = NULL;
420 txq->sw_ring[i].last_id = i;
421 txq->sw_ring[prev].next_id = i;
425 txq->tx_free_thresh = txq_conf->tx_free_thresh ?
426 txq_conf->tx_free_thresh :
427 NTB_DFLT_TX_FREE_THRESH;
428 if (txq->tx_free_thresh >= txq->nb_tx_desc - 3) {
429 NTB_LOG(ERR, "tx_free_thresh must be less than nb_desc - 3. "
430 "(tx_free_thresh=%u qp_id=%u)", txq->tx_free_thresh,
435 hw->tx_queues[qp_id] = txq;
442 ntb_queue_setup(struct rte_rawdev *dev,
444 rte_rawdev_obj_t queue_conf)
446 struct ntb_hw *hw = dev->dev_private;
449 if (queue_id >= hw->queue_pairs)
452 ret = ntb_txq_setup(dev, queue_id, queue_conf);
456 ret = ntb_rxq_setup(dev, queue_id, queue_conf);
462 ntb_queue_release(struct rte_rawdev *dev, uint16_t queue_id)
464 struct ntb_hw *hw = dev->dev_private;
466 if (queue_id >= hw->queue_pairs)
469 ntb_txq_release(hw->tx_queues[queue_id]);
470 hw->tx_queues[queue_id] = NULL;
471 ntb_rxq_release(hw->rx_queues[queue_id]);
472 hw->rx_queues[queue_id] = NULL;
478 ntb_queue_count(struct rte_rawdev *dev)
480 struct ntb_hw *hw = dev->dev_private;
481 return hw->queue_pairs;
485 ntb_queue_init(struct rte_rawdev *dev, uint16_t qp_id)
487 struct ntb_hw *hw = dev->dev_private;
488 struct ntb_rx_queue *rxq = hw->rx_queues[qp_id];
489 struct ntb_tx_queue *txq = hw->tx_queues[qp_id];
490 volatile struct ntb_header *local_hdr;
491 struct ntb_header *remote_hdr;
492 uint16_t q_size = hw->queue_size;
497 if (hw->ntb_ops->get_peer_mw_addr == NULL) {
498 NTB_LOG(ERR, "Getting peer mw addr is not supported.");
502 /* Put queue info into the start of shared memory. */
503 hdr_offset = hw->hdr_size_per_queue * qp_id;
504 local_hdr = (volatile struct ntb_header *)
505 ((size_t)hw->mz[0]->addr + hdr_offset);
506 bar_addr = (*hw->ntb_ops->get_peer_mw_addr)(dev, 0);
507 if (bar_addr == NULL)
509 remote_hdr = (struct ntb_header *)
510 ((size_t)bar_addr + hdr_offset);
513 rxq->rx_desc_ring = (struct ntb_desc *)
514 (&remote_hdr->desc_ring);
515 rxq->rx_used_ring = (volatile struct ntb_used *)
516 (&local_hdr->desc_ring[q_size]);
517 rxq->avail_cnt = &remote_hdr->avail_cnt;
518 rxq->used_cnt = &local_hdr->used_cnt;
520 for (i = 0; i < rxq->nb_rx_desc - 1; i++) {
521 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mpool);
522 if (unlikely(!mbuf)) {
523 NTB_LOG(ERR, "Failed to allocate mbuf for RX");
526 mbuf->port = dev->dev_id;
528 rxq->sw_ring[i].mbuf = mbuf;
530 rxq->rx_desc_ring[i].addr = rte_pktmbuf_mtod(mbuf, size_t);
531 rxq->rx_desc_ring[i].len = mbuf->buf_len - RTE_PKTMBUF_HEADROOM;
534 *rxq->avail_cnt = rxq->nb_rx_desc - 1;
535 rxq->last_avail = rxq->nb_rx_desc - 1;
539 txq->tx_desc_ring = (volatile struct ntb_desc *)
540 (&local_hdr->desc_ring);
541 txq->tx_used_ring = (struct ntb_used *)
542 (&remote_hdr->desc_ring[q_size]);
543 txq->avail_cnt = &local_hdr->avail_cnt;
544 txq->used_cnt = &remote_hdr->used_cnt;
550 txq->nb_tx_free = txq->nb_tx_desc - 1;
552 /* Set per queue stats. */
553 for (i = 0; i < NTB_XSTATS_NUM; i++) {
554 hw->ntb_xstats[i + NTB_XSTATS_NUM * (qp_id + 1)] = 0;
555 hw->ntb_xstats_off[i + NTB_XSTATS_NUM * (qp_id + 1)] = 0;
562 ntb_enqueue_cleanup(struct ntb_tx_queue *txq)
564 struct ntb_tx_entry *sw_ring = txq->sw_ring;
565 uint16_t tx_free = txq->last_avail;
566 uint16_t nb_to_clean, i;
568 /* avail_cnt + 1 represents where to rx next in the peer. */
569 nb_to_clean = (*txq->avail_cnt - txq->last_avail + 1 +
570 txq->nb_tx_desc) & (txq->nb_tx_desc - 1);
571 nb_to_clean = RTE_MIN(nb_to_clean, txq->tx_free_thresh);
572 for (i = 0; i < nb_to_clean; i++) {
573 if (sw_ring[tx_free].mbuf)
574 rte_pktmbuf_free_seg(sw_ring[tx_free].mbuf);
575 tx_free = (tx_free + 1) & (txq->nb_tx_desc - 1);
578 txq->nb_tx_free += nb_to_clean;
579 txq->last_avail = tx_free;
583 ntb_enqueue_bufs(struct rte_rawdev *dev,
584 struct rte_rawdev_buf **buffers,
586 rte_rawdev_obj_t context)
588 struct ntb_hw *hw = dev->dev_private;
589 struct ntb_tx_queue *txq = hw->tx_queues[(size_t)context];
590 struct ntb_tx_entry *sw_ring = txq->sw_ring;
591 struct rte_mbuf *txm;
592 struct ntb_used tx_used[NTB_MAX_DESC_SIZE];
593 volatile struct ntb_desc *tx_item;
594 uint16_t tx_last, nb_segs, off, last_used, avail_cnt;
595 uint16_t nb_mbufs = 0;
601 if (unlikely(hw->ntb_ops->ioremap == NULL)) {
602 NTB_LOG(ERR, "Ioremap not supported.");
606 if (unlikely(dev->started == 0 || hw->peer_dev_up == 0)) {
607 NTB_LOG(DEBUG, "Link is not up.");
611 if (txq->nb_tx_free < txq->tx_free_thresh)
612 ntb_enqueue_cleanup(txq);
614 off = NTB_XSTATS_NUM * ((size_t)context + 1);
615 last_used = txq->last_used;
616 avail_cnt = *txq->avail_cnt;/* Where to alloc next. */
617 for (nb_tx = 0; nb_tx < count; nb_tx++) {
618 txm = (struct rte_mbuf *)(buffers[nb_tx]->buf_addr);
619 if (txm == NULL || txq->nb_tx_free < txm->nb_segs)
622 tx_last = (txq->last_used + txm->nb_segs - 1) &
623 (txq->nb_tx_desc - 1);
624 nb_segs = txm->nb_segs;
625 for (i = 0; i < nb_segs; i++) {
626 /* Not enough ring space for tx. */
627 if (txq->last_used == avail_cnt)
629 sw_ring[txq->last_used].mbuf = txm;
630 tx_item = txq->tx_desc_ring + txq->last_used;
633 (hw->ntb_xstats[NTB_TX_ERRS_ID + off])++;
636 if (txm->data_len > tx_item->len) {
637 NTB_LOG(ERR, "Data length exceeds buf length."
638 " Only %u data would be transmitted.",
640 txm->data_len = tx_item->len;
643 /* translate remote virtual addr to bar virtual addr */
644 buf_addr = (*hw->ntb_ops->ioremap)(dev, tx_item->addr);
645 if (buf_addr == NULL) {
646 (hw->ntb_xstats[NTB_TX_ERRS_ID + off])++;
647 NTB_LOG(ERR, "Null remap addr.");
650 rte_memcpy(buf_addr, rte_pktmbuf_mtod(txm, void *),
653 tx_used[nb_mbufs].len = txm->data_len;
654 tx_used[nb_mbufs++].flags = (txq->last_used ==
659 bytes += txm->data_len;
663 sw_ring[txq->last_used].next_id = (txq->last_used + 1) &
664 (txq->nb_tx_desc - 1);
665 sw_ring[txq->last_used].last_id = tx_last;
666 txq->last_used = (txq->last_used + 1) &
667 (txq->nb_tx_desc - 1);
669 txq->nb_tx_free -= nb_segs;
675 if (nb_mbufs > txq->nb_tx_desc - last_used) {
676 nb1 = txq->nb_tx_desc - last_used;
677 nb2 = nb_mbufs - txq->nb_tx_desc + last_used;
682 rte_memcpy(txq->tx_used_ring + last_used, tx_used,
683 sizeof(struct ntb_used) * nb1);
684 rte_memcpy(txq->tx_used_ring, tx_used + nb1,
685 sizeof(struct ntb_used) * nb2);
686 *txq->used_cnt = txq->last_used;
689 /* update queue stats */
690 hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes;
691 hw->ntb_xstats[NTB_TX_PKTS_ID + off] += nb_tx;
698 ntb_dequeue_bufs(struct rte_rawdev *dev,
699 struct rte_rawdev_buf **buffers,
701 rte_rawdev_obj_t context)
703 struct ntb_hw *hw = dev->dev_private;
704 struct ntb_rx_queue *rxq = hw->rx_queues[(size_t)context];
705 struct ntb_rx_entry *sw_ring = rxq->sw_ring;
706 struct ntb_desc rx_desc[NTB_MAX_DESC_SIZE];
707 struct rte_mbuf *first, *rxm_t;
708 struct rte_mbuf *prev = NULL;
709 volatile struct ntb_used *rx_item;
710 uint16_t nb_mbufs = 0;
713 uint16_t off, last_avail, used_cnt, used_nb;
716 if (unlikely(dev->started == 0 || hw->peer_dev_up == 0)) {
717 NTB_LOG(DEBUG, "Link is not up");
721 used_cnt = *rxq->used_cnt;
723 if (rxq->last_used == used_cnt)
726 last_avail = rxq->last_avail;
727 used_nb = (used_cnt - rxq->last_used) & (rxq->nb_rx_desc - 1);
728 count = RTE_MIN(count, used_nb);
729 for (nb_rx = 0; nb_rx < count; nb_rx++) {
732 rx_item = rxq->rx_used_ring + rxq->last_used;
733 rxm_t = sw_ring[rxq->last_used].mbuf;
734 rxm_t->data_len = rx_item->len;
735 rxm_t->data_off = RTE_PKTMBUF_HEADROOM;
736 rxm_t->port = rxq->port_id;
742 buffers[nb_rx]->buf_addr = rxm_t;
749 first->pkt_len += prev->data_len;
750 rxq->last_used = (rxq->last_used + 1) &
751 (rxq->nb_rx_desc - 1);
754 rxm_t = rte_mbuf_raw_alloc(rxq->mpool);
755 if (unlikely(rxm_t == NULL)) {
756 NTB_LOG(ERR, "recv alloc mbuf failed.");
759 rxm_t->port = rxq->port_id;
760 sw_ring[rxq->last_avail].mbuf = rxm_t;
764 rx_desc[nb_mbufs].addr =
765 rte_pktmbuf_mtod(rxm_t, size_t);
766 rx_desc[nb_mbufs++].len = rxm_t->buf_len -
767 RTE_PKTMBUF_HEADROOM;
768 rxq->last_avail = (rxq->last_avail + 1) &
769 (rxq->nb_rx_desc - 1);
771 if (rx_item->flags & NTB_FLAG_EOP)
775 bytes += first->pkt_len;
781 if (nb_mbufs > rxq->nb_rx_desc - last_avail) {
782 nb1 = rxq->nb_rx_desc - last_avail;
783 nb2 = nb_mbufs - rxq->nb_rx_desc + last_avail;
788 rte_memcpy(rxq->rx_desc_ring + last_avail, rx_desc,
789 sizeof(struct ntb_desc) * nb1);
790 rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
791 sizeof(struct ntb_desc) * nb2);
792 *rxq->avail_cnt = rxq->last_avail;
795 /* update queue stats */
796 off = NTB_XSTATS_NUM * ((size_t)context + 1);
797 hw->ntb_xstats[NTB_RX_BYTES_ID + off] += bytes;
798 hw->ntb_xstats[NTB_RX_PKTS_ID + off] += nb_rx;
799 hw->ntb_xstats[NTB_RX_MISS_ID + off] += (count - nb_rx);
806 ntb_dev_info_get(struct rte_rawdev *dev, rte_rawdev_obj_t dev_info)
808 struct ntb_hw *hw = dev->dev_private;
809 struct ntb_dev_info *info = dev_info;
811 info->mw_cnt = hw->mw_cnt;
812 info->mw_size = hw->mw_size;
815 * Intel hardware requires that mapped memory base address should be
816 * aligned with EMBARSZ and needs continuous memzone.
818 info->mw_size_align = (uint8_t)(hw->pci_dev->id.vendor_id ==
819 NTB_INTEL_VENDOR_ID);
821 if (!hw->queue_size || !hw->queue_pairs) {
822 NTB_LOG(ERR, "No queue size and queue num assigned.");
826 hw->hdr_size_per_queue = RTE_ALIGN(sizeof(struct ntb_header) +
827 hw->queue_size * sizeof(struct ntb_desc) +
828 hw->queue_size * sizeof(struct ntb_used),
829 RTE_CACHE_LINE_SIZE);
830 info->ntb_hdr_size = hw->hdr_size_per_queue * hw->queue_pairs;
834 ntb_dev_configure(const struct rte_rawdev *dev, rte_rawdev_obj_t config)
836 struct ntb_dev_config *conf = config;
837 struct ntb_hw *hw = dev->dev_private;
841 hw->queue_pairs = conf->num_queues;
842 hw->queue_size = conf->queue_size;
843 hw->used_mw_num = conf->mz_num;
844 hw->mz = conf->mz_list;
845 hw->rx_queues = rte_zmalloc("ntb_rx_queues",
846 sizeof(struct ntb_rx_queue *) * hw->queue_pairs, 0);
847 hw->tx_queues = rte_zmalloc("ntb_tx_queues",
848 sizeof(struct ntb_tx_queue *) * hw->queue_pairs, 0);
849 /* First total stats, then per queue stats. */
850 xstats_num = (hw->queue_pairs + 1) * NTB_XSTATS_NUM;
851 hw->ntb_xstats = rte_zmalloc("ntb_xstats", xstats_num *
852 sizeof(uint64_t), 0);
853 hw->ntb_xstats_off = rte_zmalloc("ntb_xstats_off", xstats_num *
854 sizeof(uint64_t), 0);
856 /* Start handshake with the peer. */
857 ret = ntb_handshake_work(dev);
859 rte_free(hw->rx_queues);
860 rte_free(hw->tx_queues);
861 hw->rx_queues = NULL;
862 hw->tx_queues = NULL;
870 ntb_dev_start(struct rte_rawdev *dev)
872 struct ntb_hw *hw = dev->dev_private;
873 uint32_t peer_base_l, peer_val;
874 uint64_t peer_base_h;
878 if (!hw->link_status || !hw->peer_dev_up)
881 /* Set total stats. */
882 for (i = 0; i < NTB_XSTATS_NUM; i++) {
883 hw->ntb_xstats[i] = 0;
884 hw->ntb_xstats_off[i] = 0;
887 for (i = 0; i < hw->queue_pairs; i++) {
888 ret = ntb_queue_init(dev, i);
890 NTB_LOG(ERR, "Failed to init queue.");
895 hw->peer_mw_base = rte_zmalloc("ntb_peer_mw_base", hw->mw_cnt *
896 sizeof(uint64_t), 0);
898 if (hw->ntb_ops->spad_read == NULL) {
903 peer_val = (*hw->ntb_ops->spad_read)(dev, SPAD_Q_SZ, 0);
904 if (peer_val != hw->queue_size) {
905 NTB_LOG(ERR, "Inconsistent queue size! (local: %u peer: %u)",
906 hw->queue_size, peer_val);
911 peer_val = (*hw->ntb_ops->spad_read)(dev, SPAD_NUM_QPS, 0);
912 if (peer_val != hw->queue_pairs) {
913 NTB_LOG(ERR, "Inconsistent number of queues! (local: %u peer:"
914 " %u)", hw->queue_pairs, peer_val);
919 hw->peer_used_mws = (*hw->ntb_ops->spad_read)(dev, SPAD_USED_MWS, 0);
921 for (i = 0; i < hw->peer_used_mws; i++) {
922 peer_base_h = (*hw->ntb_ops->spad_read)(dev,
923 SPAD_MW0_BA_H + 2 * i, 0);
924 peer_base_l = (*hw->ntb_ops->spad_read)(dev,
925 SPAD_MW0_BA_L + 2 * i, 0);
926 hw->peer_mw_base[i] = (peer_base_h << 32) + peer_base_l;
934 rte_free(hw->peer_mw_base);
936 for (i = 0; i < hw->queue_pairs; i++) {
937 ntb_rxq_release_mbufs(hw->rx_queues[i]);
938 ntb_txq_release_mbufs(hw->tx_queues[i]);
945 ntb_dev_stop(struct rte_rawdev *dev)
947 struct ntb_hw *hw = dev->dev_private;
951 if (!hw->peer_dev_up)
954 ntb_link_cleanup(dev);
956 /* Notify the peer that device will be down. */
957 if (hw->ntb_ops->peer_db_set == NULL) {
958 NTB_LOG(ERR, "Peer doorbell setting is not supported.");
961 status = (*hw->ntb_ops->peer_db_set)(dev, 1);
963 NTB_LOG(ERR, "Failed to tell peer device is down.");
968 * Set time out as 1s in case that the peer is stopped accidently
969 * without any notification.
973 /* Wait for cleanup work down before db mask clear. */
974 while (hw->peer_dev_up && time_out) {
980 /* Clear doorbells mask. */
981 if (hw->ntb_ops->db_set_mask == NULL) {
982 NTB_LOG(ERR, "Doorbell mask setting is not supported.");
985 status = (*hw->ntb_ops->db_set_mask)(dev,
986 (((uint64_t)1 << hw->db_cnt) - 1));
988 NTB_LOG(ERR, "Failed to clear doorbells.");
990 for (i = 0; i < hw->queue_pairs; i++) {
991 ntb_rxq_release_mbufs(hw->rx_queues[i]);
992 ntb_txq_release_mbufs(hw->tx_queues[i]);
999 ntb_dev_close(struct rte_rawdev *dev)
1001 struct ntb_hw *hw = dev->dev_private;
1002 struct rte_intr_handle *intr_handle;
1009 for (i = 0; i < hw->queue_pairs; i++)
1010 ntb_queue_release(dev, i);
1011 hw->queue_pairs = 0;
1013 intr_handle = &hw->pci_dev->intr_handle;
1014 /* Clean datapath event and vec mapping */
1015 rte_intr_efd_disable(intr_handle);
1016 if (intr_handle->intr_vec) {
1017 rte_free(intr_handle->intr_vec);
1018 intr_handle->intr_vec = NULL;
1020 /* Disable uio intr before callback unregister */
1021 rte_intr_disable(intr_handle);
1023 /* Unregister callback func to eal lib */
1024 rte_intr_callback_unregister(intr_handle,
1025 ntb_dev_intr_handler, dev);
1031 ntb_dev_reset(struct rte_rawdev *rawdev __rte_unused)
1037 ntb_attr_set(struct rte_rawdev *dev, const char *attr_name,
1038 uint64_t attr_value)
1043 if (dev == NULL || attr_name == NULL) {
1044 NTB_LOG(ERR, "Invalid arguments for setting attributes");
1048 hw = dev->dev_private;
1050 if (!strncmp(attr_name, NTB_SPAD_USER, NTB_SPAD_USER_LEN)) {
1051 if (hw->ntb_ops->spad_write == NULL)
1053 index = atoi(&attr_name[NTB_SPAD_USER_LEN]);
1054 (*hw->ntb_ops->spad_write)(dev, hw->spad_user_list[index],
1056 NTB_LOG(DEBUG, "Set attribute (%s) Value (%" PRIu64 ")",
1057 attr_name, attr_value);
1061 if (!strncmp(attr_name, NTB_QUEUE_SZ_NAME, NTB_ATTR_NAME_LEN)) {
1062 hw->queue_size = attr_value;
1063 NTB_LOG(DEBUG, "Set attribute (%s) Value (%" PRIu64 ")",
1064 attr_name, attr_value);
1068 if (!strncmp(attr_name, NTB_QUEUE_NUM_NAME, NTB_ATTR_NAME_LEN)) {
1069 hw->queue_pairs = attr_value;
1070 NTB_LOG(DEBUG, "Set attribute (%s) Value (%" PRIu64 ")",
1071 attr_name, attr_value);
1075 /* Attribute not found. */
1076 NTB_LOG(ERR, "Attribute not found.");
1081 ntb_attr_get(struct rte_rawdev *dev, const char *attr_name,
1082 uint64_t *attr_value)
1087 if (dev == NULL || attr_name == NULL || attr_value == NULL) {
1088 NTB_LOG(ERR, "Invalid arguments for getting attributes");
1092 hw = dev->dev_private;
1094 if (!strncmp(attr_name, NTB_TOPO_NAME, NTB_ATTR_NAME_LEN)) {
1095 *attr_value = hw->topo;
1096 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1097 attr_name, *attr_value);
1101 if (!strncmp(attr_name, NTB_LINK_STATUS_NAME, NTB_ATTR_NAME_LEN)) {
1102 /* hw->link_status only indicates hw link status. */
1103 *attr_value = hw->link_status && hw->peer_dev_up;
1104 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1105 attr_name, *attr_value);
1109 if (!strncmp(attr_name, NTB_SPEED_NAME, NTB_ATTR_NAME_LEN)) {
1110 *attr_value = hw->link_speed;
1111 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1112 attr_name, *attr_value);
1116 if (!strncmp(attr_name, NTB_WIDTH_NAME, NTB_ATTR_NAME_LEN)) {
1117 *attr_value = hw->link_width;
1118 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1119 attr_name, *attr_value);
1123 if (!strncmp(attr_name, NTB_MW_CNT_NAME, NTB_ATTR_NAME_LEN)) {
1124 *attr_value = hw->mw_cnt;
1125 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1126 attr_name, *attr_value);
1130 if (!strncmp(attr_name, NTB_DB_CNT_NAME, NTB_ATTR_NAME_LEN)) {
1131 *attr_value = hw->db_cnt;
1132 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1133 attr_name, *attr_value);
1137 if (!strncmp(attr_name, NTB_SPAD_CNT_NAME, NTB_ATTR_NAME_LEN)) {
1138 *attr_value = hw->spad_cnt;
1139 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1140 attr_name, *attr_value);
1144 if (!strncmp(attr_name, NTB_SPAD_USER, NTB_SPAD_USER_LEN)) {
1145 if (hw->ntb_ops->spad_read == NULL)
1147 index = atoi(&attr_name[NTB_SPAD_USER_LEN]);
1148 *attr_value = (*hw->ntb_ops->spad_read)(dev,
1149 hw->spad_user_list[index], 0);
1150 NTB_LOG(DEBUG, "Attribute (%s) Value (%" PRIu64 ")",
1151 attr_name, *attr_value);
1155 /* Attribute not found. */
1156 NTB_LOG(ERR, "Attribute not found.");
1160 static inline uint64_t
1161 ntb_stats_update(uint64_t offset, uint64_t stat)
1164 return (stat - offset);
1166 return (uint64_t)(((uint64_t)-1) - offset + stat + 1);
1170 ntb_xstats_get(const struct rte_rawdev *dev,
1171 const unsigned int ids[],
1175 struct ntb_hw *hw = dev->dev_private;
1176 uint32_t i, j, off, xstats_num;
1178 /* Calculate total stats of all queues. */
1179 for (i = 0; i < NTB_XSTATS_NUM; i++) {
1180 hw->ntb_xstats[i] = 0;
1181 for (j = 0; j < hw->queue_pairs; j++) {
1182 off = NTB_XSTATS_NUM * (j + 1) + i;
1183 hw->ntb_xstats[i] +=
1184 ntb_stats_update(hw->ntb_xstats_off[off],
1185 hw->ntb_xstats[off]);
1189 xstats_num = NTB_XSTATS_NUM * (hw->queue_pairs + 1);
1190 for (i = 0; i < n && ids[i] < xstats_num; i++) {
1191 if (ids[i] < NTB_XSTATS_NUM)
1192 values[i] = hw->ntb_xstats[ids[i]];
1195 ntb_stats_update(hw->ntb_xstats_off[ids[i]],
1196 hw->ntb_xstats[ids[i]]);
1203 ntb_xstats_get_names(const struct rte_rawdev *dev,
1204 struct rte_rawdev_xstats_name *xstats_names,
1207 struct ntb_hw *hw = dev->dev_private;
1208 uint32_t xstats_num, i, j, off;
1210 xstats_num = NTB_XSTATS_NUM * (hw->queue_pairs + 1);
1211 if (xstats_names == NULL || size < xstats_num)
1214 /* Total stats names */
1215 memcpy(xstats_names, ntb_xstats_names, sizeof(ntb_xstats_names));
1217 /* Queue stats names */
1218 for (i = 0; i < hw->queue_pairs; i++) {
1219 for (j = 0; j < NTB_XSTATS_NUM; j++) {
1220 off = j + (i + 1) * NTB_XSTATS_NUM;
1221 snprintf(xstats_names[off].name,
1222 sizeof(xstats_names[0].name),
1223 "%s_q%u", ntb_xstats_names[j].name, i);
1231 ntb_xstats_get_by_name(const struct rte_rawdev *dev,
1232 const char *name, unsigned int *id)
1234 struct rte_rawdev_xstats_name *xstats_names;
1235 struct ntb_hw *hw = dev->dev_private;
1236 uint32_t xstats_num, i, j, off;
1241 xstats_num = NTB_XSTATS_NUM * (hw->queue_pairs + 1);
1242 xstats_names = rte_zmalloc("ntb_stats_name",
1243 sizeof(struct rte_rawdev_xstats_name) *
1245 ntb_xstats_get_names(dev, xstats_names, xstats_num);
1247 /* Calculate total stats of all queues. */
1248 for (i = 0; i < NTB_XSTATS_NUM; i++) {
1249 for (j = 0; j < hw->queue_pairs; j++) {
1250 off = NTB_XSTATS_NUM * (j + 1) + i;
1251 hw->ntb_xstats[i] +=
1252 ntb_stats_update(hw->ntb_xstats_off[off],
1253 hw->ntb_xstats[off]);
1257 for (i = 0; i < xstats_num; i++) {
1258 if (!strncmp(name, xstats_names[i].name,
1259 RTE_RAW_DEV_XSTATS_NAME_SIZE)) {
1261 rte_free(xstats_names);
1262 if (i < NTB_XSTATS_NUM)
1263 return hw->ntb_xstats[i];
1265 return ntb_stats_update(hw->ntb_xstats_off[i],
1270 NTB_LOG(ERR, "Cannot find the xstats name.");
1276 ntb_xstats_reset(struct rte_rawdev *dev,
1277 const uint32_t ids[],
1280 struct ntb_hw *hw = dev->dev_private;
1281 uint32_t i, j, off, xstats_num;
1283 xstats_num = NTB_XSTATS_NUM * (hw->queue_pairs + 1);
1284 for (i = 0; i < nb_ids && ids[i] < xstats_num; i++) {
1285 if (ids[i] < NTB_XSTATS_NUM) {
1286 for (j = 0; j < hw->queue_pairs; j++) {
1287 off = NTB_XSTATS_NUM * (j + 1) + ids[i];
1288 hw->ntb_xstats_off[off] = hw->ntb_xstats[off];
1291 hw->ntb_xstats_off[ids[i]] = hw->ntb_xstats[ids[i]];
1298 static const struct rte_rawdev_ops ntb_ops = {
1299 .dev_info_get = ntb_dev_info_get,
1300 .dev_configure = ntb_dev_configure,
1301 .dev_start = ntb_dev_start,
1302 .dev_stop = ntb_dev_stop,
1303 .dev_close = ntb_dev_close,
1304 .dev_reset = ntb_dev_reset,
1306 .queue_def_conf = ntb_queue_conf_get,
1307 .queue_setup = ntb_queue_setup,
1308 .queue_release = ntb_queue_release,
1309 .queue_count = ntb_queue_count,
1311 .enqueue_bufs = ntb_enqueue_bufs,
1312 .dequeue_bufs = ntb_dequeue_bufs,
1314 .attr_get = ntb_attr_get,
1315 .attr_set = ntb_attr_set,
1317 .xstats_get = ntb_xstats_get,
1318 .xstats_get_names = ntb_xstats_get_names,
1319 .xstats_get_by_name = ntb_xstats_get_by_name,
1320 .xstats_reset = ntb_xstats_reset,
1324 ntb_init_hw(struct rte_rawdev *dev, struct rte_pci_device *pci_dev)
1326 struct ntb_hw *hw = dev->dev_private;
1327 struct rte_intr_handle *intr_handle;
1330 hw->pci_dev = pci_dev;
1331 hw->peer_dev_up = 0;
1332 hw->link_status = NTB_LINK_DOWN;
1333 hw->link_speed = NTB_SPEED_NONE;
1334 hw->link_width = NTB_WIDTH_NONE;
1336 switch (pci_dev->id.device_id) {
1337 case NTB_INTEL_DEV_ID_B2B_SKX:
1338 hw->ntb_ops = &intel_ntb_ops;
1341 NTB_LOG(ERR, "Not supported device.");
1345 if (hw->ntb_ops->ntb_dev_init == NULL)
1347 ret = (*hw->ntb_ops->ntb_dev_init)(dev);
1349 NTB_LOG(ERR, "Unable to init ntb dev.");
1353 if (hw->ntb_ops->set_link == NULL)
1355 ret = (*hw->ntb_ops->set_link)(dev, 1);
1359 /* Init doorbell. */
1360 hw->db_valid_mask = RTE_LEN2MASK(hw->db_cnt, uint64_t);
1362 intr_handle = &pci_dev->intr_handle;
1363 /* Register callback func to eal lib */
1364 rte_intr_callback_register(intr_handle,
1365 ntb_dev_intr_handler, dev);
1367 ret = rte_intr_efd_enable(intr_handle, hw->db_cnt);
1371 /* To clarify, the interrupt for each doorbell is already mapped
1372 * by default for intel gen3. They are mapped to msix vec 1-32,
1373 * and hardware intr is mapped to 0. Map all to 0 for uio.
1375 if (!rte_intr_cap_multiple(intr_handle)) {
1376 for (i = 0; i < hw->db_cnt; i++) {
1377 if (hw->ntb_ops->vector_bind == NULL)
1379 ret = (*hw->ntb_ops->vector_bind)(dev, i, 0);
1385 if (hw->ntb_ops->db_set_mask == NULL ||
1386 hw->ntb_ops->peer_db_set == NULL) {
1387 NTB_LOG(ERR, "Doorbell is not supported.");
1391 ret = (*hw->ntb_ops->db_set_mask)(dev, hw->db_mask);
1393 NTB_LOG(ERR, "Unable to enable intr for all dbs.");
1397 /* enable uio intr after callback register */
1398 rte_intr_enable(intr_handle);
1404 ntb_create(struct rte_pci_device *pci_dev, int socket_id)
1406 char name[RTE_RAWDEV_NAME_MAX_LEN];
1407 struct rte_rawdev *rawdev = NULL;
1410 if (pci_dev == NULL) {
1411 NTB_LOG(ERR, "Invalid pci_dev.");
1415 memset(name, 0, sizeof(name));
1416 snprintf(name, RTE_RAWDEV_NAME_MAX_LEN, "NTB:%x:%02x.%x",
1417 pci_dev->addr.bus, pci_dev->addr.devid,
1418 pci_dev->addr.function);
1420 NTB_LOG(INFO, "Init %s on NUMA node %d", name, socket_id);
1422 /* Allocate device structure. */
1423 rawdev = rte_rawdev_pmd_allocate(name, sizeof(struct ntb_hw),
1425 if (rawdev == NULL) {
1426 NTB_LOG(ERR, "Unable to allocate rawdev.");
1430 rawdev->dev_ops = &ntb_ops;
1431 rawdev->device = &pci_dev->device;
1432 rawdev->driver_name = pci_dev->driver->driver.name;
1434 ret = ntb_init_hw(rawdev, pci_dev);
1436 NTB_LOG(ERR, "Unable to init ntb hw.");
1444 rte_rawdev_pmd_release(rawdev);
1450 ntb_destroy(struct rte_pci_device *pci_dev)
1452 char name[RTE_RAWDEV_NAME_MAX_LEN];
1453 struct rte_rawdev *rawdev;
1456 if (pci_dev == NULL) {
1457 NTB_LOG(ERR, "Invalid pci_dev.");
1462 memset(name, 0, sizeof(name));
1463 snprintf(name, RTE_RAWDEV_NAME_MAX_LEN, "NTB:%x:%02x.%x",
1464 pci_dev->addr.bus, pci_dev->addr.devid,
1465 pci_dev->addr.function);
1467 NTB_LOG(INFO, "Closing %s on NUMA node %d", name, rte_socket_id());
1469 rawdev = rte_rawdev_pmd_get_named_dev(name);
1470 if (rawdev == NULL) {
1471 NTB_LOG(ERR, "Invalid device name (%s)", name);
1476 ret = rte_rawdev_pmd_release(rawdev);
1478 NTB_LOG(ERR, "Failed to destroy ntb rawdev.");
1484 ntb_probe(struct rte_pci_driver *pci_drv __rte_unused,
1485 struct rte_pci_device *pci_dev)
1487 return ntb_create(pci_dev, rte_socket_id());
1491 ntb_remove(struct rte_pci_device *pci_dev)
1493 return ntb_destroy(pci_dev);
1497 static struct rte_pci_driver rte_ntb_pmd = {
1498 .id_table = pci_id_ntb_map,
1499 .drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_WC_ACTIVATE,
1501 .remove = ntb_remove,
1504 RTE_PMD_REGISTER_PCI(raw_ntb, rte_ntb_pmd);
1505 RTE_PMD_REGISTER_PCI_TABLE(raw_ntb, pci_id_ntb_map);
1506 RTE_PMD_REGISTER_KMOD_DEP(raw_ntb, "* igb_uio | uio_pci_generic | vfio-pci");
1508 RTE_INIT(ntb_init_log)
1510 ntb_logtype = rte_log_register("pmd.raw.ntb");
1511 if (ntb_logtype >= 0)
1512 rte_log_set_level(ntb_logtype, RTE_LOG_INFO);