1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2013-2016 Intel Corporation
7 #include <ethdev_driver.h>
8 #include <rte_common.h>
11 #include "base/fm10k_type.h"
13 #ifdef RTE_PMD_PACKET_PREFETCH
14 #define rte_packet_prefetch(p) rte_prefetch1(p)
16 #define rte_packet_prefetch(p) do {} while (0)
19 #ifdef RTE_ETHDEV_DEBUG_RX
20 static inline void dump_rxd(union fm10k_rx_desc *rxd)
22 PMD_RX_LOG(DEBUG, "+----------------|----------------+");
23 PMD_RX_LOG(DEBUG, "| GLORT | PKT HDR & TYPE |");
24 PMD_RX_LOG(DEBUG, "| 0x%08x | 0x%08x |", rxd->d.glort,
26 PMD_RX_LOG(DEBUG, "+----------------|----------------+");
27 PMD_RX_LOG(DEBUG, "| VLAN & LEN | STATUS |");
28 PMD_RX_LOG(DEBUG, "| 0x%08x | 0x%08x |", rxd->d.vlan_len,
30 PMD_RX_LOG(DEBUG, "+----------------|----------------+");
31 PMD_RX_LOG(DEBUG, "| RESERVED | RSS_HASH |");
32 PMD_RX_LOG(DEBUG, "| 0x%08x | 0x%08x |", 0, rxd->d.rss);
33 PMD_RX_LOG(DEBUG, "+----------------|----------------+");
34 PMD_RX_LOG(DEBUG, "| TIME TAG |");
35 PMD_RX_LOG(DEBUG, "| 0x%016"PRIx64" |", rxd->q.timestamp);
36 PMD_RX_LOG(DEBUG, "+----------------|----------------+");
40 #define FM10K_TX_OFFLOAD_MASK ( \
48 #define FM10K_TX_OFFLOAD_NOTSUP_MASK \
49 (PKT_TX_OFFLOAD_MASK ^ FM10K_TX_OFFLOAD_MASK)
51 /* @note: When this function is changed, make corresponding change to
52 * fm10k_dev_supported_ptypes_get()
55 rx_desc_to_ol_flags(struct rte_mbuf *m, const union fm10k_rx_desc *d)
58 ptype_table[FM10K_RXD_PKTTYPE_MASK >> FM10K_RXD_PKTTYPE_SHIFT]
59 __rte_cache_aligned = {
60 [FM10K_PKTTYPE_OTHER] = RTE_PTYPE_L2_ETHER,
61 [FM10K_PKTTYPE_IPV4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4,
62 [FM10K_PKTTYPE_IPV4_EX] = RTE_PTYPE_L2_ETHER |
63 RTE_PTYPE_L3_IPV4_EXT,
64 [FM10K_PKTTYPE_IPV6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6,
65 [FM10K_PKTTYPE_IPV6_EX] = RTE_PTYPE_L2_ETHER |
66 RTE_PTYPE_L3_IPV6_EXT,
67 [FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
68 RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP,
69 [FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_TCP] = RTE_PTYPE_L2_ETHER |
70 RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_TCP,
71 [FM10K_PKTTYPE_IPV4 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
72 RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP,
73 [FM10K_PKTTYPE_IPV6 | FM10K_PKTTYPE_UDP] = RTE_PTYPE_L2_ETHER |
74 RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_UDP,
77 m->packet_type = ptype_table[(d->w.pkt_info & FM10K_RXD_PKTTYPE_MASK)
78 >> FM10K_RXD_PKTTYPE_SHIFT];
80 if (d->w.pkt_info & FM10K_RXD_RSSTYPE_MASK)
81 m->ol_flags |= PKT_RX_RSS_HASH;
83 if (unlikely((d->d.staterr &
84 (FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)) ==
85 (FM10K_RXD_STATUS_IPCS | FM10K_RXD_STATUS_IPE)))
86 m->ol_flags |= PKT_RX_IP_CKSUM_BAD;
88 m->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
90 if (unlikely((d->d.staterr &
91 (FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)) ==
92 (FM10K_RXD_STATUS_L4CS | FM10K_RXD_STATUS_L4E)))
93 m->ol_flags |= PKT_RX_L4_CKSUM_BAD;
95 m->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
99 fm10k_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
102 struct rte_mbuf *mbuf;
103 union fm10k_rx_desc desc;
104 struct fm10k_rx_queue *q = rx_queue;
110 next_dd = q->next_dd;
112 nb_pkts = RTE_MIN(nb_pkts, q->alloc_thresh);
113 for (count = 0; count < nb_pkts; ++count) {
114 if (!(q->hw_ring[next_dd].d.staterr & FM10K_RXD_STATUS_DD))
116 mbuf = q->sw_ring[next_dd];
117 desc = q->hw_ring[next_dd];
118 #ifdef RTE_ETHDEV_DEBUG_RX
121 rte_pktmbuf_pkt_len(mbuf) = desc.w.length;
122 rte_pktmbuf_data_len(mbuf) = desc.w.length;
125 #ifdef RTE_LIBRTE_FM10K_RX_OLFLAGS_ENABLE
126 rx_desc_to_ol_flags(mbuf, &desc);
129 mbuf->hash.rss = desc.d.rss;
131 * Packets in fm10k device always carry at least one VLAN tag.
132 * For those packets coming in without VLAN tag,
133 * the port default VLAN tag will be used.
134 * So, always PKT_RX_VLAN flag is set and vlan_tci
135 * is valid for each RX packet's mbuf.
137 mbuf->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
138 mbuf->vlan_tci = desc.w.vlan;
140 * mbuf->vlan_tci_outer is an idle field in fm10k driver,
141 * so it can be selected to store sglort value.
144 mbuf->vlan_tci_outer = rte_le_to_cpu_16(desc.w.sglort);
146 rx_pkts[count] = mbuf;
147 if (++next_dd == q->nb_desc) {
152 /* Prefetch next mbuf while processing current one. */
153 rte_prefetch0(q->sw_ring[next_dd]);
156 * When next RX descriptor is on a cache-line boundary,
157 * prefetch the next 4 RX descriptors and the next 8 pointers
160 if ((next_dd & 0x3) == 0) {
161 rte_prefetch0(&q->hw_ring[next_dd]);
162 rte_prefetch0(&q->sw_ring[next_dd]);
166 q->next_dd = next_dd;
168 if ((q->next_dd > q->next_trigger) || (alloc == 1)) {
169 ret = rte_mempool_get_bulk(q->mp,
170 (void **)&q->sw_ring[q->next_alloc],
173 if (unlikely(ret != 0)) {
174 uint16_t port = q->port_id;
175 PMD_RX_LOG(ERR, "Failed to alloc mbuf");
177 * Need to restore next_dd if we cannot allocate new
178 * buffers to replenish the old ones.
180 q->next_dd = (q->next_dd + q->nb_desc - count) %
182 rte_eth_devices[port].data->rx_mbuf_alloc_failed++;
186 for (; q->next_alloc <= q->next_trigger; ++q->next_alloc) {
187 mbuf = q->sw_ring[q->next_alloc];
189 /* setup static mbuf fields */
190 fm10k_pktmbuf_reset(mbuf, q->port_id);
192 /* write descriptor */
193 desc.q.pkt_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
194 desc.q.hdr_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
195 q->hw_ring[q->next_alloc] = desc;
197 FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_trigger);
198 q->next_trigger += q->alloc_thresh;
199 if (q->next_trigger >= q->nb_desc) {
200 q->next_trigger = q->alloc_thresh - 1;
209 fm10k_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
212 struct rte_mbuf *mbuf;
213 union fm10k_rx_desc desc;
214 struct fm10k_rx_queue *q = rx_queue;
216 uint16_t nb_rcv, nb_seg;
219 struct rte_mbuf *first_seg = q->pkt_first_seg;
220 struct rte_mbuf *last_seg = q->pkt_last_seg;
223 next_dd = q->next_dd;
226 nb_seg = RTE_MIN(nb_pkts, q->alloc_thresh);
227 for (count = 0; count < nb_seg; count++) {
228 if (!(q->hw_ring[next_dd].d.staterr & FM10K_RXD_STATUS_DD))
230 mbuf = q->sw_ring[next_dd];
231 desc = q->hw_ring[next_dd];
232 #ifdef RTE_ETHDEV_DEBUG_RX
236 if (++next_dd == q->nb_desc) {
241 /* Prefetch next mbuf while processing current one. */
242 rte_prefetch0(q->sw_ring[next_dd]);
245 * When next RX descriptor is on a cache-line boundary,
246 * prefetch the next 4 RX descriptors and the next 8 pointers
249 if ((next_dd & 0x3) == 0) {
250 rte_prefetch0(&q->hw_ring[next_dd]);
251 rte_prefetch0(&q->sw_ring[next_dd]);
254 /* Fill data length */
255 rte_pktmbuf_data_len(mbuf) = desc.w.length;
258 * If this is the first buffer of the received packet,
259 * set the pointer to the first mbuf of the packet and
260 * initialize its context.
261 * Otherwise, update the total length and the number of segments
262 * of the current scattered packet, and update the pointer to
263 * the last mbuf of the current packet.
267 first_seg->pkt_len = desc.w.length;
270 (uint16_t)(first_seg->pkt_len +
271 rte_pktmbuf_data_len(mbuf));
272 first_seg->nb_segs++;
273 last_seg->next = mbuf;
277 * If this is not the last buffer of the received packet,
278 * update the pointer to the last mbuf of the current scattered
279 * packet and continue to parse the RX ring.
281 if (!(desc.d.staterr & FM10K_RXD_STATUS_EOP)) {
286 first_seg->ol_flags = 0;
287 #ifdef RTE_LIBRTE_FM10K_RX_OLFLAGS_ENABLE
288 rx_desc_to_ol_flags(first_seg, &desc);
290 first_seg->hash.rss = desc.d.rss;
292 * Packets in fm10k device always carry at least one VLAN tag.
293 * For those packets coming in without VLAN tag,
294 * the port default VLAN tag will be used.
295 * So, always PKT_RX_VLAN flag is set and vlan_tci
296 * is valid for each RX packet's mbuf.
298 first_seg->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
299 first_seg->vlan_tci = desc.w.vlan;
301 * mbuf->vlan_tci_outer is an idle field in fm10k driver,
302 * so it can be selected to store sglort value.
305 first_seg->vlan_tci_outer =
306 rte_le_to_cpu_16(desc.w.sglort);
308 /* Prefetch data of first segment, if configured to do so. */
309 rte_packet_prefetch((char *)first_seg->buf_addr +
310 first_seg->data_off);
313 * Store the mbuf address into the next entry of the array
314 * of returned packets.
316 rx_pkts[nb_rcv++] = first_seg;
319 * Setup receipt context for a new packet.
324 q->next_dd = next_dd;
326 if ((q->next_dd > q->next_trigger) || (alloc == 1)) {
327 ret = rte_mempool_get_bulk(q->mp,
328 (void **)&q->sw_ring[q->next_alloc],
331 if (unlikely(ret != 0)) {
332 uint16_t port = q->port_id;
333 PMD_RX_LOG(ERR, "Failed to alloc mbuf");
335 * Need to restore next_dd if we cannot allocate new
336 * buffers to replenish the old ones.
338 q->next_dd = (q->next_dd + q->nb_desc - count) %
340 rte_eth_devices[port].data->rx_mbuf_alloc_failed++;
344 for (; q->next_alloc <= q->next_trigger; ++q->next_alloc) {
345 mbuf = q->sw_ring[q->next_alloc];
347 /* setup static mbuf fields */
348 fm10k_pktmbuf_reset(mbuf, q->port_id);
350 /* write descriptor */
351 desc.q.pkt_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
352 desc.q.hdr_addr = MBUF_DMA_ADDR_DEFAULT(mbuf);
353 q->hw_ring[q->next_alloc] = desc;
355 FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_trigger);
356 q->next_trigger += q->alloc_thresh;
357 if (q->next_trigger >= q->nb_desc) {
358 q->next_trigger = q->alloc_thresh - 1;
363 q->pkt_first_seg = first_seg;
364 q->pkt_last_seg = last_seg;
370 fm10k_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
372 #define FM10K_RXQ_SCAN_INTERVAL 4
373 volatile union fm10k_rx_desc *rxdp;
374 struct fm10k_rx_queue *rxq;
377 rxq = dev->data->rx_queues[rx_queue_id];
378 rxdp = &rxq->hw_ring[rxq->next_dd];
379 while ((desc < rxq->nb_desc) &&
380 rxdp->w.status & rte_cpu_to_le_16(FM10K_RXD_STATUS_DD)) {
382 * Check the DD bit of a rx descriptor of each group of 4 desc,
383 * to avoid checking too frequently and downgrading performance
386 desc += FM10K_RXQ_SCAN_INTERVAL;
387 rxdp += FM10K_RXQ_SCAN_INTERVAL;
388 if (rxq->next_dd + desc >= rxq->nb_desc)
389 rxdp = &rxq->hw_ring[rxq->next_dd + desc -
397 fm10k_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
399 volatile union fm10k_rx_desc *rxdp;
400 struct fm10k_rx_queue *rxq = rx_queue;
401 uint16_t nb_hold, trigger_last;
405 if (unlikely(offset >= rxq->nb_desc)) {
406 PMD_DRV_LOG(ERR, "Invalid RX descriptor offset %u", offset);
410 if (rxq->next_trigger < rxq->alloc_thresh)
411 trigger_last = rxq->next_trigger +
412 rxq->nb_desc - rxq->alloc_thresh;
414 trigger_last = rxq->next_trigger - rxq->alloc_thresh;
416 if (rxq->next_dd < trigger_last)
417 nb_hold = rxq->next_dd + rxq->nb_desc - trigger_last;
419 nb_hold = rxq->next_dd - trigger_last;
421 if (offset >= rxq->nb_desc - nb_hold)
422 return RTE_ETH_RX_DESC_UNAVAIL;
424 desc = rxq->next_dd + offset;
425 if (desc >= rxq->nb_desc)
426 desc -= rxq->nb_desc;
428 rxdp = &rxq->hw_ring[desc];
430 ret = !!(rxdp->w.status &
431 rte_cpu_to_le_16(FM10K_RXD_STATUS_DD));
437 fm10k_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
439 volatile struct fm10k_tx_desc *txdp;
440 struct fm10k_tx_queue *txq = tx_queue;
442 uint16_t next_rs = txq->nb_desc;
443 struct fifo rs_tracker = txq->rs_tracker;
444 struct fifo *r = &rs_tracker;
446 if (unlikely(offset >= txq->nb_desc))
449 desc = txq->next_free + offset;
450 /* go to next desc that has the RS bit */
451 desc = (desc / txq->rs_thresh + 1) *
454 if (desc >= txq->nb_desc) {
455 desc -= txq->nb_desc;
456 if (desc >= txq->nb_desc)
457 desc -= txq->nb_desc;
461 for ( ; r->head != r->endp; ) {
462 if (*r->head >= desc && *r->head < next_rs)
467 txdp = &txq->hw_ring[next_rs];
468 if (txdp->flags & FM10K_TXD_FLAG_DONE)
469 return RTE_ETH_TX_DESC_DONE;
471 return RTE_ETH_TX_DESC_FULL;
475 * Free multiple TX mbuf at a time if they are in the same pool
477 * @txep: software desc ring index that starts to free
478 * @num: number of descs to free
481 static inline void tx_free_bulk_mbuf(struct rte_mbuf **txep, int num)
483 struct rte_mbuf *m, *free[RTE_FM10K_TX_MAX_FREE_BUF_SZ];
487 if (unlikely(num == 0))
490 m = rte_pktmbuf_prefree_seg(txep[0]);
491 if (likely(m != NULL)) {
494 for (i = 1; i < num; i++) {
495 m = rte_pktmbuf_prefree_seg(txep[i]);
496 if (likely(m != NULL)) {
497 if (likely(m->pool == free[0]->pool))
500 rte_mempool_put_bulk(free[0]->pool,
501 (void *)free, nb_free);
508 rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
510 for (i = 1; i < num; i++) {
511 m = rte_pktmbuf_prefree_seg(txep[i]);
513 rte_mempool_put(m->pool, m);
519 static inline void tx_free_descriptors(struct fm10k_tx_queue *q)
521 uint16_t next_rs, count = 0;
523 next_rs = fifo_peek(&q->rs_tracker);
524 if (!(q->hw_ring[next_rs].flags & FM10K_TXD_FLAG_DONE))
527 /* the DONE flag is set on this descriptor so remove the ID
528 * from the RS bit tracker and free the buffers */
529 fifo_remove(&q->rs_tracker);
531 /* wrap around? if so, free buffers from last_free up to but NOT
532 * including nb_desc */
533 if (q->last_free > next_rs) {
534 count = q->nb_desc - q->last_free;
535 tx_free_bulk_mbuf(&q->sw_ring[q->last_free], count);
539 /* adjust free descriptor count before the next loop */
540 q->nb_free += count + (next_rs + 1 - q->last_free);
542 /* free buffers from last_free, up to and including next_rs */
543 if (q->last_free <= next_rs) {
544 count = next_rs - q->last_free + 1;
545 tx_free_bulk_mbuf(&q->sw_ring[q->last_free], count);
546 q->last_free += count;
549 if (q->last_free == q->nb_desc)
553 static inline void tx_xmit_pkt(struct fm10k_tx_queue *q, struct rte_mbuf *mb)
556 uint8_t flags, hdrlen;
558 /* always set the LAST flag on the last descriptor used to
559 * transmit the packet */
560 flags = FM10K_TXD_FLAG_LAST;
561 last_id = q->next_free + mb->nb_segs - 1;
562 if (last_id >= q->nb_desc)
563 last_id = last_id - q->nb_desc;
565 /* but only set the RS flag on the last descriptor if rs_thresh
566 * descriptors will be used since the RS flag was last set */
567 if ((q->nb_used + mb->nb_segs) >= q->rs_thresh) {
568 flags |= FM10K_TXD_FLAG_RS;
569 fifo_insert(&q->rs_tracker, last_id);
572 q->nb_used = q->nb_used + mb->nb_segs;
575 q->nb_free -= mb->nb_segs;
577 q->hw_ring[q->next_free].flags = 0;
579 q->hw_ring[q->next_free].flags |= FM10K_TXD_FLAG_FTAG;
580 /* set checksum flags on first descriptor of packet. SCTP checksum
581 * offload is not supported, but we do not explicitly check for this
582 * case in favor of greatly simplified processing. */
583 if (mb->ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK | PKT_TX_TCP_SEG))
584 q->hw_ring[q->next_free].flags |= FM10K_TXD_FLAG_CSUM;
586 /* set vlan if requested */
587 if (mb->ol_flags & PKT_TX_VLAN_PKT)
588 q->hw_ring[q->next_free].vlan = mb->vlan_tci;
590 q->hw_ring[q->next_free].vlan = 0;
592 q->sw_ring[q->next_free] = mb;
593 q->hw_ring[q->next_free].buffer_addr =
594 rte_cpu_to_le_64(MBUF_DMA_ADDR(mb));
595 q->hw_ring[q->next_free].buflen =
596 rte_cpu_to_le_16(rte_pktmbuf_data_len(mb));
598 if (mb->ol_flags & PKT_TX_TCP_SEG) {
599 hdrlen = mb->l2_len + mb->l3_len + mb->l4_len;
600 hdrlen += (mb->ol_flags & PKT_TX_TUNNEL_MASK) ?
601 mb->outer_l2_len + mb->outer_l3_len : 0;
602 if (q->hw_ring[q->next_free].flags & FM10K_TXD_FLAG_FTAG)
603 hdrlen += sizeof(struct fm10k_ftag);
605 if (likely((hdrlen >= FM10K_TSO_MIN_HEADERLEN) &&
606 (hdrlen <= FM10K_TSO_MAX_HEADERLEN) &&
607 (mb->tso_segsz >= FM10K_TSO_MINMSS))) {
608 q->hw_ring[q->next_free].mss = mb->tso_segsz;
609 q->hw_ring[q->next_free].hdrlen = hdrlen;
613 if (++q->next_free == q->nb_desc)
616 /* fill up the rings */
617 for (mb = mb->next; mb != NULL; mb = mb->next) {
618 q->sw_ring[q->next_free] = mb;
619 q->hw_ring[q->next_free].buffer_addr =
620 rte_cpu_to_le_64(MBUF_DMA_ADDR(mb));
621 q->hw_ring[q->next_free].buflen =
622 rte_cpu_to_le_16(rte_pktmbuf_data_len(mb));
623 q->hw_ring[q->next_free].flags = 0;
624 if (++q->next_free == q->nb_desc)
628 q->hw_ring[last_id].flags |= flags;
632 fm10k_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
635 struct fm10k_tx_queue *q = tx_queue;
639 for (count = 0; count < nb_pkts; ++count) {
642 /* running low on descriptors? try to free some... */
643 if (q->nb_free < q->free_thresh)
644 tx_free_descriptors(q);
646 /* make sure there are enough free descriptors to transmit the
647 * entire packet before doing anything */
648 if (q->nb_free < mb->nb_segs)
651 /* sanity check to make sure the mbuf is valid */
652 if ((mb->nb_segs == 0) ||
653 ((mb->nb_segs > 1) && (mb->next == NULL)))
656 /* process the packet */
660 /* update the tail pointer if any packets were processed */
661 if (likely(count > 0))
662 FM10K_PCI_REG_WRITE(q->tail_ptr, q->next_free);
668 fm10k_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
674 for (i = 0; i < nb_pkts; i++) {
677 if ((m->ol_flags & PKT_TX_TCP_SEG) &&
678 (m->tso_segsz < FM10K_TSO_MINMSS)) {
683 if (m->ol_flags & FM10K_TX_OFFLOAD_NOTSUP_MASK) {
688 #ifdef RTE_ETHDEV_DEBUG_TX
689 ret = rte_validate_tx_offload(m);
695 ret = rte_net_intel_cksum_prepare(m);