4 * Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
5 * Copyright 2014 6WIND S.A.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * * Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * * Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * * Neither the name of Intel Corporation nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
25 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
26 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
27 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
28 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
32 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 #include <sys/queue.h>
46 #include <rte_byteorder.h>
47 #include <rte_common.h>
48 #include <rte_cycles.h>
50 #include <rte_debug.h>
51 #include <rte_interrupts.h>
53 #include <rte_memory.h>
54 #include <rte_memzone.h>
55 #include <rte_launch.h>
57 #include <rte_per_lcore.h>
58 #include <rte_lcore.h>
59 #include <rte_atomic.h>
60 #include <rte_branch_prediction.h>
62 #include <rte_mempool.h>
63 #include <rte_malloc.h>
65 #include <rte_ether.h>
66 #include <rte_ethdev.h>
67 #include <rte_prefetch.h>
71 #include <rte_string_fns.h>
72 #include <rte_errno.h>
75 #include "ixgbe_logs.h"
76 #include "base/ixgbe_api.h"
77 #include "base/ixgbe_vf.h"
78 #include "ixgbe_ethdev.h"
79 #include "base/ixgbe_dcb.h"
80 #include "base/ixgbe_common.h"
81 #include "ixgbe_rxtx.h"
83 /* Bit Mask to indicate what bits required for building TX context */
84 #define IXGBE_TX_OFFLOAD_MASK ( \
90 static inline struct rte_mbuf *
91 rte_rxmbuf_alloc(struct rte_mempool *mp)
95 m = __rte_mbuf_raw_alloc(mp);
96 __rte_mbuf_sanity_check_raw(m, 0);
102 #define RTE_PMD_USE_PREFETCH
105 #ifdef RTE_PMD_USE_PREFETCH
107 * Prefetch a cache line into all cache levels.
109 #define rte_ixgbe_prefetch(p) rte_prefetch0(p)
111 #define rte_ixgbe_prefetch(p) do {} while(0)
114 /*********************************************************************
118 **********************************************************************/
121 * Check for descriptors with their DD bit set and free mbufs.
122 * Return the total number of buffers freed.
124 static inline int __attribute__((always_inline))
125 ixgbe_tx_free_bufs(struct ixgbe_tx_queue *txq)
127 struct ixgbe_tx_entry *txep;
131 /* check DD bit on threshold descriptor */
132 status = txq->tx_ring[txq->tx_next_dd].wb.status;
133 if (! (status & IXGBE_ADVTXD_STAT_DD))
137 * first buffer to free from S/W ring is at index
138 * tx_next_dd - (tx_rs_thresh-1)
140 txep = &(txq->sw_ring[txq->tx_next_dd - (txq->tx_rs_thresh - 1)]);
142 /* free buffers one at a time */
143 if ((txq->txq_flags & (uint32_t)ETH_TXQ_FLAGS_NOREFCOUNT) != 0) {
144 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
145 txep->mbuf->next = NULL;
146 rte_mempool_put(txep->mbuf->pool, txep->mbuf);
150 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
151 rte_pktmbuf_free_seg(txep->mbuf);
156 /* buffers were freed, update counters */
157 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
158 txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
159 if (txq->tx_next_dd >= txq->nb_tx_desc)
160 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
162 return txq->tx_rs_thresh;
165 /* Populate 4 descriptors with data from 4 mbufs */
167 tx4(volatile union ixgbe_adv_tx_desc *txdp, struct rte_mbuf **pkts)
169 uint64_t buf_dma_addr;
173 for (i = 0; i < 4; ++i, ++txdp, ++pkts) {
174 buf_dma_addr = RTE_MBUF_DATA_DMA_ADDR(*pkts);
175 pkt_len = (*pkts)->data_len;
177 /* write data to descriptor */
178 txdp->read.buffer_addr = buf_dma_addr;
179 txdp->read.cmd_type_len =
180 ((uint32_t)DCMD_DTYP_FLAGS | pkt_len);
181 txdp->read.olinfo_status =
182 (pkt_len << IXGBE_ADVTXD_PAYLEN_SHIFT);
183 rte_prefetch0(&(*pkts)->pool);
187 /* Populate 1 descriptor with data from 1 mbuf */
189 tx1(volatile union ixgbe_adv_tx_desc *txdp, struct rte_mbuf **pkts)
191 uint64_t buf_dma_addr;
194 buf_dma_addr = RTE_MBUF_DATA_DMA_ADDR(*pkts);
195 pkt_len = (*pkts)->data_len;
197 /* write data to descriptor */
198 txdp->read.buffer_addr = buf_dma_addr;
199 txdp->read.cmd_type_len =
200 ((uint32_t)DCMD_DTYP_FLAGS | pkt_len);
201 txdp->read.olinfo_status =
202 (pkt_len << IXGBE_ADVTXD_PAYLEN_SHIFT);
203 rte_prefetch0(&(*pkts)->pool);
207 * Fill H/W descriptor ring with mbuf data.
208 * Copy mbuf pointers to the S/W ring.
211 ixgbe_tx_fill_hw_ring(struct ixgbe_tx_queue *txq, struct rte_mbuf **pkts,
214 volatile union ixgbe_adv_tx_desc *txdp = &(txq->tx_ring[txq->tx_tail]);
215 struct ixgbe_tx_entry *txep = &(txq->sw_ring[txq->tx_tail]);
216 const int N_PER_LOOP = 4;
217 const int N_PER_LOOP_MASK = N_PER_LOOP-1;
218 int mainpart, leftover;
222 * Process most of the packets in chunks of N pkts. Any
223 * leftover packets will get processed one at a time.
225 mainpart = (nb_pkts & ((uint32_t) ~N_PER_LOOP_MASK));
226 leftover = (nb_pkts & ((uint32_t) N_PER_LOOP_MASK));
227 for (i = 0; i < mainpart; i += N_PER_LOOP) {
228 /* Copy N mbuf pointers to the S/W ring */
229 for (j = 0; j < N_PER_LOOP; ++j) {
230 (txep + i + j)->mbuf = *(pkts + i + j);
232 tx4(txdp + i, pkts + i);
235 if (unlikely(leftover > 0)) {
236 for (i = 0; i < leftover; ++i) {
237 (txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
238 tx1(txdp + mainpart + i, pkts + mainpart + i);
243 static inline uint16_t
244 tx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
247 struct ixgbe_tx_queue *txq = (struct ixgbe_tx_queue *)tx_queue;
248 volatile union ixgbe_adv_tx_desc *tx_r = txq->tx_ring;
252 * Begin scanning the H/W ring for done descriptors when the
253 * number of available descriptors drops below tx_free_thresh. For
254 * each done descriptor, free the associated buffer.
256 if (txq->nb_tx_free < txq->tx_free_thresh)
257 ixgbe_tx_free_bufs(txq);
259 /* Only use descriptors that are available */
260 nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
261 if (unlikely(nb_pkts == 0))
264 /* Use exactly nb_pkts descriptors */
265 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
268 * At this point, we know there are enough descriptors in the
269 * ring to transmit all the packets. This assumes that each
270 * mbuf contains a single segment, and that no new offloads
271 * are expected, which would require a new context descriptor.
275 * See if we're going to wrap-around. If so, handle the top
276 * of the descriptor ring first, then do the bottom. If not,
277 * the processing looks just like the "bottom" part anyway...
279 if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
280 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
281 ixgbe_tx_fill_hw_ring(txq, tx_pkts, n);
284 * We know that the last descriptor in the ring will need to
285 * have its RS bit set because tx_rs_thresh has to be
286 * a divisor of the ring size
288 tx_r[txq->tx_next_rs].read.cmd_type_len |=
289 rte_cpu_to_le_32(IXGBE_ADVTXD_DCMD_RS);
290 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
295 /* Fill H/W descriptor ring with mbuf data */
296 ixgbe_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
297 txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
300 * Determine if RS bit should be set
301 * This is what we actually want:
302 * if ((txq->tx_tail - 1) >= txq->tx_next_rs)
303 * but instead of subtracting 1 and doing >=, we can just do
304 * greater than without subtracting.
306 if (txq->tx_tail > txq->tx_next_rs) {
307 tx_r[txq->tx_next_rs].read.cmd_type_len |=
308 rte_cpu_to_le_32(IXGBE_ADVTXD_DCMD_RS);
309 txq->tx_next_rs = (uint16_t)(txq->tx_next_rs +
311 if (txq->tx_next_rs >= txq->nb_tx_desc)
312 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
316 * Check for wrap-around. This would only happen if we used
317 * up to the last descriptor in the ring, no more, no less.
319 if (txq->tx_tail >= txq->nb_tx_desc)
322 /* update tail pointer */
324 IXGBE_PCI_REG_WRITE(txq->tdt_reg_addr, txq->tx_tail);
330 ixgbe_xmit_pkts_simple(void *tx_queue, struct rte_mbuf **tx_pkts,
335 /* Try to transmit at least chunks of TX_MAX_BURST pkts */
336 if (likely(nb_pkts <= RTE_PMD_IXGBE_TX_MAX_BURST))
337 return tx_xmit_pkts(tx_queue, tx_pkts, nb_pkts);
339 /* transmit more than the max burst, in chunks of TX_MAX_BURST */
343 n = (uint16_t)RTE_MIN(nb_pkts, RTE_PMD_IXGBE_TX_MAX_BURST);
344 ret = tx_xmit_pkts(tx_queue, &(tx_pkts[nb_tx]), n);
345 nb_tx = (uint16_t)(nb_tx + ret);
346 nb_pkts = (uint16_t)(nb_pkts - ret);
355 ixgbe_set_xmit_ctx(struct ixgbe_tx_queue *txq,
356 volatile struct ixgbe_adv_tx_context_desc *ctx_txd,
357 uint64_t ol_flags, union ixgbe_tx_offload tx_offload)
359 uint32_t type_tucmd_mlhl;
360 uint32_t mss_l4len_idx = 0;
362 uint32_t vlan_macip_lens;
363 union ixgbe_tx_offload tx_offload_mask;
365 ctx_idx = txq->ctx_curr;
366 tx_offload_mask.data = 0;
369 /* Specify which HW CTX to upload. */
370 mss_l4len_idx |= (ctx_idx << IXGBE_ADVTXD_IDX_SHIFT);
372 if (ol_flags & PKT_TX_VLAN_PKT) {
373 tx_offload_mask.vlan_tci |= ~0;
376 /* check if TCP segmentation required for this packet */
377 if (ol_flags & PKT_TX_TCP_SEG) {
378 /* implies IP cksum in IPv4 */
379 if (ol_flags & PKT_TX_IP_CKSUM)
380 type_tucmd_mlhl = IXGBE_ADVTXD_TUCMD_IPV4 |
381 IXGBE_ADVTXD_TUCMD_L4T_TCP |
382 IXGBE_ADVTXD_DTYP_CTXT | IXGBE_ADVTXD_DCMD_DEXT;
384 type_tucmd_mlhl = IXGBE_ADVTXD_TUCMD_IPV6 |
385 IXGBE_ADVTXD_TUCMD_L4T_TCP |
386 IXGBE_ADVTXD_DTYP_CTXT | IXGBE_ADVTXD_DCMD_DEXT;
388 tx_offload_mask.l2_len |= ~0;
389 tx_offload_mask.l3_len |= ~0;
390 tx_offload_mask.l4_len |= ~0;
391 tx_offload_mask.tso_segsz |= ~0;
392 mss_l4len_idx |= tx_offload.tso_segsz << IXGBE_ADVTXD_MSS_SHIFT;
393 mss_l4len_idx |= tx_offload.l4_len << IXGBE_ADVTXD_L4LEN_SHIFT;
394 } else { /* no TSO, check if hardware checksum is needed */
395 if (ol_flags & PKT_TX_IP_CKSUM) {
396 type_tucmd_mlhl = IXGBE_ADVTXD_TUCMD_IPV4;
397 tx_offload_mask.l2_len |= ~0;
398 tx_offload_mask.l3_len |= ~0;
401 switch (ol_flags & PKT_TX_L4_MASK) {
402 case PKT_TX_UDP_CKSUM:
403 type_tucmd_mlhl |= IXGBE_ADVTXD_TUCMD_L4T_UDP |
404 IXGBE_ADVTXD_DTYP_CTXT | IXGBE_ADVTXD_DCMD_DEXT;
405 mss_l4len_idx |= sizeof(struct udp_hdr) << IXGBE_ADVTXD_L4LEN_SHIFT;
406 tx_offload_mask.l2_len |= ~0;
407 tx_offload_mask.l3_len |= ~0;
409 case PKT_TX_TCP_CKSUM:
410 type_tucmd_mlhl |= IXGBE_ADVTXD_TUCMD_L4T_TCP |
411 IXGBE_ADVTXD_DTYP_CTXT | IXGBE_ADVTXD_DCMD_DEXT;
412 mss_l4len_idx |= sizeof(struct tcp_hdr) << IXGBE_ADVTXD_L4LEN_SHIFT;
413 tx_offload_mask.l2_len |= ~0;
414 tx_offload_mask.l3_len |= ~0;
415 tx_offload_mask.l4_len |= ~0;
417 case PKT_TX_SCTP_CKSUM:
418 type_tucmd_mlhl |= IXGBE_ADVTXD_TUCMD_L4T_SCTP |
419 IXGBE_ADVTXD_DTYP_CTXT | IXGBE_ADVTXD_DCMD_DEXT;
420 mss_l4len_idx |= sizeof(struct sctp_hdr) << IXGBE_ADVTXD_L4LEN_SHIFT;
421 tx_offload_mask.l2_len |= ~0;
422 tx_offload_mask.l3_len |= ~0;
425 type_tucmd_mlhl |= IXGBE_ADVTXD_TUCMD_L4T_RSV |
426 IXGBE_ADVTXD_DTYP_CTXT | IXGBE_ADVTXD_DCMD_DEXT;
431 txq->ctx_cache[ctx_idx].flags = ol_flags;
432 txq->ctx_cache[ctx_idx].tx_offload.data =
433 tx_offload_mask.data & tx_offload.data;
434 txq->ctx_cache[ctx_idx].tx_offload_mask = tx_offload_mask;
436 ctx_txd->type_tucmd_mlhl = rte_cpu_to_le_32(type_tucmd_mlhl);
437 vlan_macip_lens = tx_offload.l3_len;
438 vlan_macip_lens |= (tx_offload.l2_len << IXGBE_ADVTXD_MACLEN_SHIFT);
439 vlan_macip_lens |= ((uint32_t)tx_offload.vlan_tci << IXGBE_ADVTXD_VLAN_SHIFT);
440 ctx_txd->vlan_macip_lens = rte_cpu_to_le_32(vlan_macip_lens);
441 ctx_txd->mss_l4len_idx = rte_cpu_to_le_32(mss_l4len_idx);
442 ctx_txd->seqnum_seed = 0;
446 * Check which hardware context can be used. Use the existing match
447 * or create a new context descriptor.
449 static inline uint32_t
450 what_advctx_update(struct ixgbe_tx_queue *txq, uint64_t flags,
451 union ixgbe_tx_offload tx_offload)
453 /* If match with the current used context */
454 if (likely((txq->ctx_cache[txq->ctx_curr].flags == flags) &&
455 (txq->ctx_cache[txq->ctx_curr].tx_offload.data ==
456 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data & tx_offload.data)))) {
457 return txq->ctx_curr;
460 /* What if match with the next context */
462 if (likely((txq->ctx_cache[txq->ctx_curr].flags == flags) &&
463 (txq->ctx_cache[txq->ctx_curr].tx_offload.data ==
464 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data & tx_offload.data)))) {
465 return txq->ctx_curr;
468 /* Mismatch, use the previous context */
469 return (IXGBE_CTX_NUM);
472 static inline uint32_t
473 tx_desc_cksum_flags_to_olinfo(uint64_t ol_flags)
476 if ((ol_flags & PKT_TX_L4_MASK) != PKT_TX_L4_NO_CKSUM)
477 tmp |= IXGBE_ADVTXD_POPTS_TXSM;
478 if (ol_flags & PKT_TX_IP_CKSUM)
479 tmp |= IXGBE_ADVTXD_POPTS_IXSM;
480 if (ol_flags & PKT_TX_TCP_SEG)
481 tmp |= IXGBE_ADVTXD_POPTS_TXSM;
485 static inline uint32_t
486 tx_desc_ol_flags_to_cmdtype(uint64_t ol_flags)
488 uint32_t cmdtype = 0;
489 if (ol_flags & PKT_TX_VLAN_PKT)
490 cmdtype |= IXGBE_ADVTXD_DCMD_VLE;
491 if (ol_flags & PKT_TX_TCP_SEG)
492 cmdtype |= IXGBE_ADVTXD_DCMD_TSE;
496 /* Default RS bit threshold values */
497 #ifndef DEFAULT_TX_RS_THRESH
498 #define DEFAULT_TX_RS_THRESH 32
500 #ifndef DEFAULT_TX_FREE_THRESH
501 #define DEFAULT_TX_FREE_THRESH 32
504 /* Reset transmit descriptors after they have been used */
506 ixgbe_xmit_cleanup(struct ixgbe_tx_queue *txq)
508 struct ixgbe_tx_entry *sw_ring = txq->sw_ring;
509 volatile union ixgbe_adv_tx_desc *txr = txq->tx_ring;
510 uint16_t last_desc_cleaned = txq->last_desc_cleaned;
511 uint16_t nb_tx_desc = txq->nb_tx_desc;
512 uint16_t desc_to_clean_to;
513 uint16_t nb_tx_to_clean;
515 /* Determine the last descriptor needing to be cleaned */
516 desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
517 if (desc_to_clean_to >= nb_tx_desc)
518 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
520 /* Check to make sure the last descriptor to clean is done */
521 desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
522 if (! (txr[desc_to_clean_to].wb.status & IXGBE_TXD_STAT_DD))
524 PMD_TX_FREE_LOG(DEBUG,
525 "TX descriptor %4u is not done"
526 "(port=%d queue=%d)",
528 txq->port_id, txq->queue_id);
529 /* Failed to clean any descriptors, better luck next time */
533 /* Figure out how many descriptors will be cleaned */
534 if (last_desc_cleaned > desc_to_clean_to)
535 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
538 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
541 PMD_TX_FREE_LOG(DEBUG,
542 "Cleaning %4u TX descriptors: %4u to %4u "
543 "(port=%d queue=%d)",
544 nb_tx_to_clean, last_desc_cleaned, desc_to_clean_to,
545 txq->port_id, txq->queue_id);
548 * The last descriptor to clean is done, so that means all the
549 * descriptors from the last descriptor that was cleaned
550 * up to the last descriptor with the RS bit set
551 * are done. Only reset the threshold descriptor.
553 txr[desc_to_clean_to].wb.status = 0;
555 /* Update the txq to reflect the last descriptor that was cleaned */
556 txq->last_desc_cleaned = desc_to_clean_to;
557 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
564 ixgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
567 struct ixgbe_tx_queue *txq;
568 struct ixgbe_tx_entry *sw_ring;
569 struct ixgbe_tx_entry *txe, *txn;
570 volatile union ixgbe_adv_tx_desc *txr;
571 volatile union ixgbe_adv_tx_desc *txd;
572 struct rte_mbuf *tx_pkt;
573 struct rte_mbuf *m_seg;
574 uint64_t buf_dma_addr;
575 uint32_t olinfo_status;
576 uint32_t cmd_type_len;
587 union ixgbe_tx_offload tx_offload = {0};
590 sw_ring = txq->sw_ring;
592 tx_id = txq->tx_tail;
593 txe = &sw_ring[tx_id];
595 /* Determine if the descriptor ring needs to be cleaned. */
596 if ((txq->nb_tx_desc - txq->nb_tx_free) > txq->tx_free_thresh) {
597 ixgbe_xmit_cleanup(txq);
600 rte_prefetch0(&txe->mbuf->pool);
603 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
606 pkt_len = tx_pkt->pkt_len;
609 * Determine how many (if any) context descriptors
610 * are needed for offload functionality.
612 ol_flags = tx_pkt->ol_flags;
614 /* If hardware offload required */
615 tx_ol_req = ol_flags & IXGBE_TX_OFFLOAD_MASK;
617 tx_offload.l2_len = tx_pkt->l2_len;
618 tx_offload.l3_len = tx_pkt->l3_len;
619 tx_offload.l4_len = tx_pkt->l4_len;
620 tx_offload.vlan_tci = tx_pkt->vlan_tci;
621 tx_offload.tso_segsz = tx_pkt->tso_segsz;
623 /* If new context need be built or reuse the exist ctx. */
624 ctx = what_advctx_update(txq, tx_ol_req,
626 /* Only allocate context descriptor if required*/
627 new_ctx = (ctx == IXGBE_CTX_NUM);
632 * Keep track of how many descriptors are used this loop
633 * This will always be the number of segments + the number of
634 * Context descriptors required to transmit the packet
636 nb_used = (uint16_t)(tx_pkt->nb_segs + new_ctx);
639 * The number of descriptors that must be allocated for a
640 * packet is the number of segments of that packet, plus 1
641 * Context Descriptor for the hardware offload, if any.
642 * Determine the last TX descriptor to allocate in the TX ring
643 * for the packet, starting from the current position (tx_id)
646 tx_last = (uint16_t) (tx_id + nb_used - 1);
649 if (tx_last >= txq->nb_tx_desc)
650 tx_last = (uint16_t) (tx_last - txq->nb_tx_desc);
652 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u pktlen=%u"
653 " tx_first=%u tx_last=%u",
654 (unsigned) txq->port_id,
655 (unsigned) txq->queue_id,
661 * Make sure there are enough TX descriptors available to
662 * transmit the entire packet.
663 * nb_used better be less than or equal to txq->tx_rs_thresh
665 if (nb_used > txq->nb_tx_free) {
666 PMD_TX_FREE_LOG(DEBUG,
667 "Not enough free TX descriptors "
668 "nb_used=%4u nb_free=%4u "
669 "(port=%d queue=%d)",
670 nb_used, txq->nb_tx_free,
671 txq->port_id, txq->queue_id);
673 if (ixgbe_xmit_cleanup(txq) != 0) {
674 /* Could not clean any descriptors */
680 /* nb_used better be <= txq->tx_rs_thresh */
681 if (unlikely(nb_used > txq->tx_rs_thresh)) {
682 PMD_TX_FREE_LOG(DEBUG,
683 "The number of descriptors needed to "
684 "transmit the packet exceeds the "
685 "RS bit threshold. This will impact "
687 "nb_used=%4u nb_free=%4u "
689 "(port=%d queue=%d)",
690 nb_used, txq->nb_tx_free,
692 txq->port_id, txq->queue_id);
694 * Loop here until there are enough TX
695 * descriptors or until the ring cannot be
698 while (nb_used > txq->nb_tx_free) {
699 if (ixgbe_xmit_cleanup(txq) != 0) {
701 * Could not clean any
713 * By now there are enough free TX descriptors to transmit
718 * Set common flags of all TX Data Descriptors.
720 * The following bits must be set in all Data Descriptors:
721 * - IXGBE_ADVTXD_DTYP_DATA
722 * - IXGBE_ADVTXD_DCMD_DEXT
724 * The following bits must be set in the first Data Descriptor
725 * and are ignored in the other ones:
726 * - IXGBE_ADVTXD_DCMD_IFCS
727 * - IXGBE_ADVTXD_MAC_1588
728 * - IXGBE_ADVTXD_DCMD_VLE
730 * The following bits must only be set in the last Data
732 * - IXGBE_TXD_CMD_EOP
734 * The following bits can be set in any Data Descriptor, but
735 * are only set in the last Data Descriptor:
738 cmd_type_len = IXGBE_ADVTXD_DTYP_DATA |
739 IXGBE_ADVTXD_DCMD_IFCS | IXGBE_ADVTXD_DCMD_DEXT;
741 #ifdef RTE_LIBRTE_IEEE1588
742 if (ol_flags & PKT_TX_IEEE1588_TMST)
743 cmd_type_len |= IXGBE_ADVTXD_MAC_1588;
749 if (ol_flags & PKT_TX_TCP_SEG) {
750 /* when TSO is on, paylen in descriptor is the
751 * not the packet len but the tcp payload len */
752 pkt_len -= (tx_offload.l2_len +
753 tx_offload.l3_len + tx_offload.l4_len);
757 * Setup the TX Advanced Context Descriptor if required
760 volatile struct ixgbe_adv_tx_context_desc *
763 ctx_txd = (volatile struct
764 ixgbe_adv_tx_context_desc *)
767 txn = &sw_ring[txe->next_id];
768 rte_prefetch0(&txn->mbuf->pool);
770 if (txe->mbuf != NULL) {
771 rte_pktmbuf_free_seg(txe->mbuf);
775 ixgbe_set_xmit_ctx(txq, ctx_txd, tx_ol_req,
778 txe->last_id = tx_last;
779 tx_id = txe->next_id;
784 * Setup the TX Advanced Data Descriptor,
785 * This path will go through
786 * whatever new/reuse the context descriptor
788 cmd_type_len |= tx_desc_ol_flags_to_cmdtype(ol_flags);
789 olinfo_status |= tx_desc_cksum_flags_to_olinfo(ol_flags);
790 olinfo_status |= ctx << IXGBE_ADVTXD_IDX_SHIFT;
793 olinfo_status |= (pkt_len << IXGBE_ADVTXD_PAYLEN_SHIFT);
798 txn = &sw_ring[txe->next_id];
799 rte_prefetch0(&txn->mbuf->pool);
801 if (txe->mbuf != NULL)
802 rte_pktmbuf_free_seg(txe->mbuf);
806 * Set up Transmit Data Descriptor.
808 slen = m_seg->data_len;
809 buf_dma_addr = RTE_MBUF_DATA_DMA_ADDR(m_seg);
810 txd->read.buffer_addr =
811 rte_cpu_to_le_64(buf_dma_addr);
812 txd->read.cmd_type_len =
813 rte_cpu_to_le_32(cmd_type_len | slen);
814 txd->read.olinfo_status =
815 rte_cpu_to_le_32(olinfo_status);
816 txe->last_id = tx_last;
817 tx_id = txe->next_id;
820 } while (m_seg != NULL);
823 * The last packet data descriptor needs End Of Packet (EOP)
825 cmd_type_len |= IXGBE_TXD_CMD_EOP;
826 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
827 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
829 /* Set RS bit only on threshold packets' last descriptor */
830 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
831 PMD_TX_FREE_LOG(DEBUG,
832 "Setting RS bit on TXD id="
833 "%4u (port=%d queue=%d)",
834 tx_last, txq->port_id, txq->queue_id);
836 cmd_type_len |= IXGBE_TXD_CMD_RS;
838 /* Update txq RS bit counters */
841 txd->read.cmd_type_len |= rte_cpu_to_le_32(cmd_type_len);
847 * Set the Transmit Descriptor Tail (TDT)
849 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
850 (unsigned) txq->port_id, (unsigned) txq->queue_id,
851 (unsigned) tx_id, (unsigned) nb_tx);
852 IXGBE_PCI_REG_WRITE(txq->tdt_reg_addr, tx_id);
853 txq->tx_tail = tx_id;
858 /*********************************************************************
862 **********************************************************************/
863 static inline uint64_t
864 rx_desc_hlen_type_rss_to_pkt_flags(uint32_t hl_tp_rs)
868 static const uint64_t ip_pkt_types_map[16] = {
869 0, PKT_RX_IPV4_HDR, PKT_RX_IPV4_HDR_EXT, PKT_RX_IPV4_HDR_EXT,
870 PKT_RX_IPV6_HDR, 0, 0, 0,
871 PKT_RX_IPV6_HDR_EXT, 0, 0, 0,
872 PKT_RX_IPV6_HDR_EXT, 0, 0, 0,
875 static const uint64_t ip_rss_types_map[16] = {
876 0, PKT_RX_RSS_HASH, PKT_RX_RSS_HASH, PKT_RX_RSS_HASH,
877 0, PKT_RX_RSS_HASH, 0, PKT_RX_RSS_HASH,
878 PKT_RX_RSS_HASH, 0, 0, 0,
879 0, 0, 0, PKT_RX_FDIR,
882 #ifdef RTE_LIBRTE_IEEE1588
883 static uint64_t ip_pkt_etqf_map[8] = {
884 0, 0, 0, PKT_RX_IEEE1588_PTP,
888 pkt_flags = (hl_tp_rs & IXGBE_RXDADV_PKTTYPE_ETQF) ?
889 ip_pkt_etqf_map[(hl_tp_rs >> 4) & 0x07] :
890 ip_pkt_types_map[(hl_tp_rs >> 4) & 0x0F];
892 pkt_flags = (hl_tp_rs & IXGBE_RXDADV_PKTTYPE_ETQF) ? 0 :
893 ip_pkt_types_map[(hl_tp_rs >> 4) & 0x0F];
896 return pkt_flags | ip_rss_types_map[hl_tp_rs & 0xF];
899 static inline uint64_t
900 rx_desc_status_to_pkt_flags(uint32_t rx_status)
905 * Check if VLAN present only.
906 * Do not check whether L3/L4 rx checksum done by NIC or not,
907 * That can be found from rte_eth_rxmode.hw_ip_checksum flag
909 pkt_flags = (rx_status & IXGBE_RXD_STAT_VP) ? PKT_RX_VLAN_PKT : 0;
911 #ifdef RTE_LIBRTE_IEEE1588
912 if (rx_status & IXGBE_RXD_STAT_TMST)
913 pkt_flags = pkt_flags | PKT_RX_IEEE1588_TMST;
918 static inline uint64_t
919 rx_desc_error_to_pkt_flags(uint32_t rx_status)
922 * Bit 31: IPE, IPv4 checksum error
923 * Bit 30: L4I, L4I integrity error
925 static uint64_t error_to_pkt_flags_map[4] = {
926 0, PKT_RX_L4_CKSUM_BAD, PKT_RX_IP_CKSUM_BAD,
927 PKT_RX_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD
929 return error_to_pkt_flags_map[(rx_status >>
930 IXGBE_RXDADV_ERR_CKSUM_BIT) & IXGBE_RXDADV_ERR_CKSUM_MSK];
933 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
935 * LOOK_AHEAD defines how many desc statuses to check beyond the
936 * current descriptor.
937 * It must be a pound define for optimal performance.
938 * Do not change the value of LOOK_AHEAD, as the ixgbe_rx_scan_hw_ring
939 * function only works with LOOK_AHEAD=8.
942 #if (LOOK_AHEAD != 8)
943 #error "PMD IXGBE: LOOK_AHEAD must be 8\n"
946 ixgbe_rx_scan_hw_ring(struct ixgbe_rx_queue *rxq)
948 volatile union ixgbe_adv_rx_desc *rxdp;
949 struct ixgbe_rx_entry *rxep;
953 int s[LOOK_AHEAD], nb_dd;
957 /* get references to current descriptor and S/W ring entry */
958 rxdp = &rxq->rx_ring[rxq->rx_tail];
959 rxep = &rxq->sw_ring[rxq->rx_tail];
961 /* check to make sure there is at least 1 packet to receive */
962 if (! (rxdp->wb.upper.status_error & IXGBE_RXDADV_STAT_DD))
966 * Scan LOOK_AHEAD descriptors at a time to determine which descriptors
967 * reference packets that are ready to be received.
969 for (i = 0; i < RTE_PMD_IXGBE_RX_MAX_BURST;
970 i += LOOK_AHEAD, rxdp += LOOK_AHEAD, rxep += LOOK_AHEAD)
972 /* Read desc statuses backwards to avoid race condition */
973 for (j = LOOK_AHEAD-1; j >= 0; --j)
974 s[j] = rxdp[j].wb.upper.status_error;
976 /* Compute how many status bits were set */
978 for (j = 0; j < LOOK_AHEAD; ++j)
979 nb_dd += s[j] & IXGBE_RXDADV_STAT_DD;
983 /* Translate descriptor info to mbuf format */
984 for (j = 0; j < nb_dd; ++j) {
986 pkt_len = (uint16_t)(rxdp[j].wb.upper.length - rxq->crc_len);
987 mb->data_len = pkt_len;
988 mb->pkt_len = pkt_len;
989 mb->vlan_tci = rxdp[j].wb.upper.vlan;
990 mb->vlan_tci = rte_le_to_cpu_16(rxdp[j].wb.upper.vlan);
992 /* convert descriptor fields to rte mbuf flags */
993 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(
994 rxdp[j].wb.lower.lo_dword.data);
995 /* reuse status field from scan list */
996 pkt_flags |= rx_desc_status_to_pkt_flags(s[j]);
997 pkt_flags |= rx_desc_error_to_pkt_flags(s[j]);
998 mb->ol_flags = pkt_flags;
1000 if (likely(pkt_flags & PKT_RX_RSS_HASH))
1001 mb->hash.rss = rxdp[j].wb.lower.hi_dword.rss;
1002 else if (pkt_flags & PKT_RX_FDIR) {
1003 mb->hash.fdir.hash =
1004 (uint16_t)((rxdp[j].wb.lower.hi_dword.csum_ip.csum)
1005 & IXGBE_ATR_HASH_MASK);
1006 mb->hash.fdir.id = rxdp[j].wb.lower.hi_dword.csum_ip.ip_id;
1010 /* Move mbuf pointers from the S/W ring to the stage */
1011 for (j = 0; j < LOOK_AHEAD; ++j) {
1012 rxq->rx_stage[i + j] = rxep[j].mbuf;
1015 /* stop if all requested packets could not be received */
1016 if (nb_dd != LOOK_AHEAD)
1020 /* clear software ring entries so we can cleanup correctly */
1021 for (i = 0; i < nb_rx; ++i) {
1022 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
1030 ixgbe_rx_alloc_bufs(struct ixgbe_rx_queue *rxq, bool reset_mbuf)
1032 volatile union ixgbe_adv_rx_desc *rxdp;
1033 struct ixgbe_rx_entry *rxep;
1034 struct rte_mbuf *mb;
1039 /* allocate buffers in bulk directly into the S/W ring */
1040 alloc_idx = rxq->rx_free_trigger - (rxq->rx_free_thresh - 1);
1041 rxep = &rxq->sw_ring[alloc_idx];
1042 diag = rte_mempool_get_bulk(rxq->mb_pool, (void *)rxep,
1043 rxq->rx_free_thresh);
1044 if (unlikely(diag != 0))
1047 rxdp = &rxq->rx_ring[alloc_idx];
1048 for (i = 0; i < rxq->rx_free_thresh; ++i) {
1049 /* populate the static rte mbuf fields */
1054 mb->port = rxq->port_id;
1057 rte_mbuf_refcnt_set(mb, 1);
1058 mb->data_off = RTE_PKTMBUF_HEADROOM;
1060 /* populate the descriptors */
1061 dma_addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mb));
1062 rxdp[i].read.hdr_addr = dma_addr;
1063 rxdp[i].read.pkt_addr = dma_addr;
1066 /* update state of internal queue structure */
1067 rxq->rx_free_trigger = rxq->rx_free_trigger + rxq->rx_free_thresh;
1068 if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
1069 rxq->rx_free_trigger = rxq->rx_free_thresh - 1;
1075 static inline uint16_t
1076 ixgbe_rx_fill_from_stage(struct ixgbe_rx_queue *rxq, struct rte_mbuf **rx_pkts,
1079 struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
1082 /* how many packets are ready to return? */
1083 nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
1085 /* copy mbuf pointers to the application's packet list */
1086 for (i = 0; i < nb_pkts; ++i)
1087 rx_pkts[i] = stage[i];
1089 /* update internal queue state */
1090 rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
1091 rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
1096 static inline uint16_t
1097 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
1100 struct ixgbe_rx_queue *rxq = (struct ixgbe_rx_queue *)rx_queue;
1103 /* Any previously recv'd pkts will be returned from the Rx stage */
1104 if (rxq->rx_nb_avail)
1105 return ixgbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
1107 /* Scan the H/W ring for packets to receive */
1108 nb_rx = (uint16_t)ixgbe_rx_scan_hw_ring(rxq);
1110 /* update internal queue state */
1111 rxq->rx_next_avail = 0;
1112 rxq->rx_nb_avail = nb_rx;
1113 rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
1115 /* if required, allocate new buffers to replenish descriptors */
1116 if (rxq->rx_tail > rxq->rx_free_trigger) {
1117 uint16_t cur_free_trigger = rxq->rx_free_trigger;
1119 if (ixgbe_rx_alloc_bufs(rxq, true) != 0) {
1121 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
1122 "queue_id=%u", (unsigned) rxq->port_id,
1123 (unsigned) rxq->queue_id);
1125 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
1126 rxq->rx_free_thresh;
1129 * Need to rewind any previous receives if we cannot
1130 * allocate new buffers to replenish the old ones.
1132 rxq->rx_nb_avail = 0;
1133 rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
1134 for (i = 0, j = rxq->rx_tail; i < nb_rx; ++i, ++j)
1135 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
1140 /* update tail pointer */
1142 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, cur_free_trigger);
1145 if (rxq->rx_tail >= rxq->nb_rx_desc)
1148 /* received any packets this loop? */
1149 if (rxq->rx_nb_avail)
1150 return ixgbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
1155 /* split requests into chunks of size RTE_PMD_IXGBE_RX_MAX_BURST */
1157 ixgbe_recv_pkts_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1162 if (unlikely(nb_pkts == 0))
1165 if (likely(nb_pkts <= RTE_PMD_IXGBE_RX_MAX_BURST))
1166 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
1168 /* request is relatively large, chunk it up */
1172 n = (uint16_t)RTE_MIN(nb_pkts, RTE_PMD_IXGBE_RX_MAX_BURST);
1173 ret = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
1174 nb_rx = (uint16_t)(nb_rx + ret);
1175 nb_pkts = (uint16_t)(nb_pkts - ret);
1185 /* Stub to avoid extra ifdefs */
1187 ixgbe_recv_pkts_bulk_alloc(__rte_unused void *rx_queue,
1188 __rte_unused struct rte_mbuf **rx_pkts, __rte_unused uint16_t nb_pkts)
1194 ixgbe_rx_alloc_bufs(__rte_unused struct ixgbe_rx_queue *rxq,
1195 __rte_unused bool reset_mbuf)
1199 #endif /* RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC */
1202 ixgbe_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
1205 struct ixgbe_rx_queue *rxq;
1206 volatile union ixgbe_adv_rx_desc *rx_ring;
1207 volatile union ixgbe_adv_rx_desc *rxdp;
1208 struct ixgbe_rx_entry *sw_ring;
1209 struct ixgbe_rx_entry *rxe;
1210 struct rte_mbuf *rxm;
1211 struct rte_mbuf *nmb;
1212 union ixgbe_adv_rx_desc rxd;
1215 uint32_t hlen_type_rss;
1225 rx_id = rxq->rx_tail;
1226 rx_ring = rxq->rx_ring;
1227 sw_ring = rxq->sw_ring;
1228 while (nb_rx < nb_pkts) {
1230 * The order of operations here is important as the DD status
1231 * bit must not be read after any other descriptor fields.
1232 * rx_ring and rxdp are pointing to volatile data so the order
1233 * of accesses cannot be reordered by the compiler. If they were
1234 * not volatile, they could be reordered which could lead to
1235 * using invalid descriptor fields when read from rxd.
1237 rxdp = &rx_ring[rx_id];
1238 staterr = rxdp->wb.upper.status_error;
1239 if (! (staterr & rte_cpu_to_le_32(IXGBE_RXDADV_STAT_DD)))
1246 * If the IXGBE_RXDADV_STAT_EOP flag is not set, the RX packet
1247 * is likely to be invalid and to be dropped by the various
1248 * validation checks performed by the network stack.
1250 * Allocate a new mbuf to replenish the RX ring descriptor.
1251 * If the allocation fails:
1252 * - arrange for that RX descriptor to be the first one
1253 * being parsed the next time the receive function is
1254 * invoked [on the same queue].
1256 * - Stop parsing the RX ring and return immediately.
1258 * This policy do not drop the packet received in the RX
1259 * descriptor for which the allocation of a new mbuf failed.
1260 * Thus, it allows that packet to be later retrieved if
1261 * mbuf have been freed in the mean time.
1262 * As a side effect, holding RX descriptors instead of
1263 * systematically giving them back to the NIC may lead to
1264 * RX ring exhaustion situations.
1265 * However, the NIC can gracefully prevent such situations
1266 * to happen by sending specific "back-pressure" flow control
1267 * frames to its peer(s).
1269 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
1270 "ext_err_stat=0x%08x pkt_len=%u",
1271 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1272 (unsigned) rx_id, (unsigned) staterr,
1273 (unsigned) rte_le_to_cpu_16(rxd.wb.upper.length));
1275 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
1277 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
1278 "queue_id=%u", (unsigned) rxq->port_id,
1279 (unsigned) rxq->queue_id);
1280 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
1285 rxe = &sw_ring[rx_id];
1287 if (rx_id == rxq->nb_rx_desc)
1290 /* Prefetch next mbuf while processing current one. */
1291 rte_ixgbe_prefetch(sw_ring[rx_id].mbuf);
1294 * When next RX descriptor is on a cache-line boundary,
1295 * prefetch the next 4 RX descriptors and the next 8 pointers
1298 if ((rx_id & 0x3) == 0) {
1299 rte_ixgbe_prefetch(&rx_ring[rx_id]);
1300 rte_ixgbe_prefetch(&sw_ring[rx_id]);
1306 rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
1307 rxdp->read.hdr_addr = dma_addr;
1308 rxdp->read.pkt_addr = dma_addr;
1311 * Initialize the returned mbuf.
1312 * 1) setup generic mbuf fields:
1313 * - number of segments,
1316 * - RX port identifier.
1317 * 2) integrate hardware offload data, if any:
1318 * - RSS flag & hash,
1319 * - IP checksum flag,
1320 * - VLAN TCI, if any,
1323 pkt_len = (uint16_t) (rte_le_to_cpu_16(rxd.wb.upper.length) -
1325 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1326 rte_packet_prefetch((char *)rxm->buf_addr + rxm->data_off);
1329 rxm->pkt_len = pkt_len;
1330 rxm->data_len = pkt_len;
1331 rxm->port = rxq->port_id;
1333 hlen_type_rss = rte_le_to_cpu_32(rxd.wb.lower.lo_dword.data);
1334 /* Only valid if PKT_RX_VLAN_PKT set in pkt_flags */
1335 rxm->vlan_tci = rte_le_to_cpu_16(rxd.wb.upper.vlan);
1337 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(hlen_type_rss);
1338 pkt_flags = pkt_flags | rx_desc_status_to_pkt_flags(staterr);
1339 pkt_flags = pkt_flags | rx_desc_error_to_pkt_flags(staterr);
1340 rxm->ol_flags = pkt_flags;
1342 if (likely(pkt_flags & PKT_RX_RSS_HASH))
1343 rxm->hash.rss = rxd.wb.lower.hi_dword.rss;
1344 else if (pkt_flags & PKT_RX_FDIR) {
1345 rxm->hash.fdir.hash =
1346 (uint16_t)((rxd.wb.lower.hi_dword.csum_ip.csum)
1347 & IXGBE_ATR_HASH_MASK);
1348 rxm->hash.fdir.id = rxd.wb.lower.hi_dword.csum_ip.ip_id;
1351 * Store the mbuf address into the next entry of the array
1352 * of returned packets.
1354 rx_pkts[nb_rx++] = rxm;
1356 rxq->rx_tail = rx_id;
1359 * If the number of free RX descriptors is greater than the RX free
1360 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1362 * Update the RDT with the value of the last processed RX descriptor
1363 * minus 1, to guarantee that the RDT register is never equal to the
1364 * RDH register, which creates a "full" ring situtation from the
1365 * hardware point of view...
1367 nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
1368 if (nb_hold > rxq->rx_free_thresh) {
1369 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1370 "nb_hold=%u nb_rx=%u",
1371 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1372 (unsigned) rx_id, (unsigned) nb_hold,
1374 rx_id = (uint16_t) ((rx_id == 0) ?
1375 (rxq->nb_rx_desc - 1) : (rx_id - 1));
1376 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
1379 rxq->nb_rx_hold = nb_hold;
1384 * Detect an RSC descriptor.
1386 static inline uint32_t
1387 ixgbe_rsc_count(union ixgbe_adv_rx_desc *rx)
1389 return (rte_le_to_cpu_32(rx->wb.lower.lo_dword.data) &
1390 IXGBE_RXDADV_RSCCNT_MASK) >> IXGBE_RXDADV_RSCCNT_SHIFT;
1394 * ixgbe_fill_cluster_head_buf - fill the first mbuf of the returned packet
1396 * Fill the following info in the HEAD buffer of the Rx cluster:
1397 * - RX port identifier
1398 * - hardware offload data, if any:
1400 * - IP checksum flag
1401 * - VLAN TCI, if any
1403 * @head HEAD of the packet cluster
1404 * @desc HW descriptor to get data from
1405 * @port_id Port ID of the Rx queue
1408 ixgbe_fill_cluster_head_buf(
1409 struct rte_mbuf *head,
1410 union ixgbe_adv_rx_desc *desc,
1414 uint32_t hlen_type_rss;
1417 head->port = port_id;
1420 * The vlan_tci field is only valid when PKT_RX_VLAN_PKT is
1421 * set in the pkt_flags field.
1423 head->vlan_tci = rte_le_to_cpu_16(desc->wb.upper.vlan);
1424 hlen_type_rss = rte_le_to_cpu_32(desc->wb.lower.lo_dword.data);
1425 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(hlen_type_rss);
1426 pkt_flags |= rx_desc_status_to_pkt_flags(staterr);
1427 pkt_flags |= rx_desc_error_to_pkt_flags(staterr);
1428 head->ol_flags = pkt_flags;
1430 if (likely(pkt_flags & PKT_RX_RSS_HASH))
1431 head->hash.rss = rte_le_to_cpu_32(desc->wb.lower.hi_dword.rss);
1432 else if (pkt_flags & PKT_RX_FDIR) {
1433 head->hash.fdir.hash =
1434 rte_le_to_cpu_16(desc->wb.lower.hi_dword.csum_ip.csum)
1435 & IXGBE_ATR_HASH_MASK;
1436 head->hash.fdir.id =
1437 rte_le_to_cpu_16(desc->wb.lower.hi_dword.csum_ip.ip_id);
1442 * ixgbe_recv_pkts_lro - receive handler for and LRO case.
1444 * @rx_queue Rx queue handle
1445 * @rx_pkts table of received packets
1446 * @nb_pkts size of rx_pkts table
1447 * @bulk_alloc if TRUE bulk allocation is used for a HW ring refilling
1449 * Handles the Rx HW ring completions when RSC feature is configured. Uses an
1450 * additional ring of ixgbe_rsc_entry's that will hold the relevant RSC info.
1452 * We use the same logic as in Linux and in FreeBSD ixgbe drivers:
1453 * 1) When non-EOP RSC completion arrives:
1454 * a) Update the HEAD of the current RSC aggregation cluster with the new
1455 * segment's data length.
1456 * b) Set the "next" pointer of the current segment to point to the segment
1457 * at the NEXTP index.
1458 * c) Pass the HEAD of RSC aggregation cluster on to the next NEXTP entry
1459 * in the sw_rsc_ring.
1460 * 2) When EOP arrives we just update the cluster's total length and offload
1461 * flags and deliver the cluster up to the upper layers. In our case - put it
1462 * in the rx_pkts table.
1464 * Returns the number of received packets/clusters (according to the "bulk
1465 * receive" interface).
1467 static inline uint16_t
1468 ixgbe_recv_pkts_lro(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts,
1471 struct ixgbe_rx_queue *rxq = rx_queue;
1472 volatile union ixgbe_adv_rx_desc *rx_ring = rxq->rx_ring;
1473 struct ixgbe_rx_entry *sw_ring = rxq->sw_ring;
1474 struct ixgbe_scattered_rx_entry *sw_sc_ring = rxq->sw_sc_ring;
1475 uint16_t rx_id = rxq->rx_tail;
1477 uint16_t nb_hold = rxq->nb_rx_hold;
1478 uint16_t prev_id = rxq->rx_tail;
1480 while (nb_rx < nb_pkts) {
1482 struct ixgbe_rx_entry *rxe;
1483 struct ixgbe_scattered_rx_entry *sc_entry;
1484 struct ixgbe_scattered_rx_entry *next_sc_entry;
1485 struct ixgbe_rx_entry *next_rxe;
1486 struct rte_mbuf *first_seg;
1487 struct rte_mbuf *rxm;
1488 struct rte_mbuf *nmb;
1489 union ixgbe_adv_rx_desc rxd;
1492 volatile union ixgbe_adv_rx_desc *rxdp;
1497 * The code in this whole file uses the volatile pointer to
1498 * ensure the read ordering of the status and the rest of the
1499 * descriptor fields (on the compiler level only!!!). This is so
1500 * UGLY - why not to just use the compiler barrier instead? DPDK
1501 * even has the rte_compiler_barrier() for that.
1503 * But most importantly this is just wrong because this doesn't
1504 * ensure memory ordering in a general case at all. For
1505 * instance, DPDK is supposed to work on Power CPUs where
1506 * compiler barrier may just not be enough!
1508 * I tried to write only this function properly to have a
1509 * starting point (as a part of an LRO/RSC series) but the
1510 * compiler cursed at me when I tried to cast away the
1511 * "volatile" from rx_ring (yes, it's volatile too!!!). So, I'm
1512 * keeping it the way it is for now.
1514 * The code in this file is broken in so many other places and
1515 * will just not work on a big endian CPU anyway therefore the
1516 * lines below will have to be revisited together with the rest
1520 * - Get rid of "volatile" crap and let the compiler do its
1522 * - Use the proper memory barrier (rte_rmb()) to ensure the
1523 * memory ordering below.
1525 rxdp = &rx_ring[rx_id];
1526 staterr = rte_le_to_cpu_32(rxdp->wb.upper.status_error);
1528 if (!(staterr & IXGBE_RXDADV_STAT_DD))
1533 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
1534 "staterr=0x%x data_len=%u",
1535 rxq->port_id, rxq->queue_id, rx_id, staterr,
1536 rte_le_to_cpu_16(rxd.wb.upper.length));
1539 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
1541 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed "
1542 "port_id=%u queue_id=%u",
1543 rxq->port_id, rxq->queue_id);
1545 rte_eth_devices[rxq->port_id].data->
1546 rx_mbuf_alloc_failed++;
1549 } else if (nb_hold > rxq->rx_free_thresh) {
1550 uint16_t next_rdt = rxq->rx_free_trigger;
1552 if (!ixgbe_rx_alloc_bufs(rxq, false)) {
1554 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr,
1556 nb_hold -= rxq->rx_free_thresh;
1558 PMD_RX_LOG(DEBUG, "RX bulk alloc failed "
1559 "port_id=%u queue_id=%u",
1560 rxq->port_id, rxq->queue_id);
1562 rte_eth_devices[rxq->port_id].data->
1563 rx_mbuf_alloc_failed++;
1569 rxe = &sw_ring[rx_id];
1570 eop = staterr & IXGBE_RXDADV_STAT_EOP;
1572 next_id = rx_id + 1;
1573 if (next_id == rxq->nb_rx_desc)
1576 /* Prefetch next mbuf while processing current one. */
1577 rte_ixgbe_prefetch(sw_ring[next_id].mbuf);
1580 * When next RX descriptor is on a cache-line boundary,
1581 * prefetch the next 4 RX descriptors and the next 4 pointers
1584 if ((next_id & 0x3) == 0) {
1585 rte_ixgbe_prefetch(&rx_ring[next_id]);
1586 rte_ixgbe_prefetch(&sw_ring[next_id]);
1593 rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
1595 * Update RX descriptor with the physical address of the
1596 * new data buffer of the new allocated mbuf.
1600 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1601 rxdp->read.hdr_addr = dma;
1602 rxdp->read.pkt_addr = dma;
1607 * Set data length & data buffer address of mbuf.
1609 data_len = rte_le_to_cpu_16(rxd.wb.upper.length);
1610 rxm->data_len = data_len;
1615 * Get next descriptor index:
1616 * - For RSC it's in the NEXTP field.
1617 * - For a scattered packet - it's just a following
1620 if (ixgbe_rsc_count(&rxd))
1622 (staterr & IXGBE_RXDADV_NEXTP_MASK) >>
1623 IXGBE_RXDADV_NEXTP_SHIFT;
1627 next_sc_entry = &sw_sc_ring[nextp_id];
1628 next_rxe = &sw_ring[nextp_id];
1629 rte_ixgbe_prefetch(next_rxe);
1632 sc_entry = &sw_sc_ring[rx_id];
1633 first_seg = sc_entry->fbuf;
1634 sc_entry->fbuf = NULL;
1637 * If this is the first buffer of the received packet,
1638 * set the pointer to the first mbuf of the packet and
1639 * initialize its context.
1640 * Otherwise, update the total length and the number of segments
1641 * of the current scattered packet, and update the pointer to
1642 * the last mbuf of the current packet.
1644 if (first_seg == NULL) {
1646 first_seg->pkt_len = data_len;
1647 first_seg->nb_segs = 1;
1649 first_seg->pkt_len += data_len;
1650 first_seg->nb_segs++;
1657 * If this is not the last buffer of the received packet, update
1658 * the pointer to the first mbuf at the NEXTP entry in the
1659 * sw_sc_ring and continue to parse the RX ring.
1662 rxm->next = next_rxe->mbuf;
1663 next_sc_entry->fbuf = first_seg;
1668 * This is the last buffer of the received packet - return
1669 * the current cluster to the user.
1673 /* Initialize the first mbuf of the returned packet */
1674 ixgbe_fill_cluster_head_buf(first_seg, &rxd, rxq->port_id,
1677 /* Prefetch data of first segment, if configured to do so. */
1678 rte_packet_prefetch((char *)first_seg->buf_addr +
1679 first_seg->data_off);
1682 * Store the mbuf address into the next entry of the array
1683 * of returned packets.
1685 rx_pkts[nb_rx++] = first_seg;
1689 * Record index of the next RX descriptor to probe.
1691 rxq->rx_tail = rx_id;
1694 * If the number of free RX descriptors is greater than the RX free
1695 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1697 * Update the RDT with the value of the last processed RX descriptor
1698 * minus 1, to guarantee that the RDT register is never equal to the
1699 * RDH register, which creates a "full" ring situtation from the
1700 * hardware point of view...
1702 if (!bulk_alloc && nb_hold > rxq->rx_free_thresh) {
1703 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1704 "nb_hold=%u nb_rx=%u",
1705 rxq->port_id, rxq->queue_id, rx_id, nb_hold, nb_rx);
1708 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, prev_id);
1712 rxq->nb_rx_hold = nb_hold;
1717 ixgbe_recv_pkts_lro_single_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1720 return ixgbe_recv_pkts_lro(rx_queue, rx_pkts, nb_pkts, false);
1724 ixgbe_recv_pkts_lro_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1727 return ixgbe_recv_pkts_lro(rx_queue, rx_pkts, nb_pkts, true);
1730 /*********************************************************************
1732 * Queue management functions
1734 **********************************************************************/
1737 * Rings setup and release.
1739 * TDBA/RDBA should be aligned on 16 byte boundary. But TDLEN/RDLEN should be
1740 * multiple of 128 bytes. So we align TDBA/RDBA on 128 byte boundary. This will
1741 * also optimize cache line size effect. H/W supports up to cache line size 128.
1743 #define IXGBE_ALIGN 128
1746 * Maximum number of Ring Descriptors.
1748 * Since RDLEN/TDLEN should be multiple of 128 bytes, the number of ring
1749 * descriptors should meet the following condition:
1750 * (num_ring_desc * sizeof(rx/tx descriptor)) % 128 == 0
1752 #define IXGBE_MIN_RING_DESC 32
1753 #define IXGBE_MAX_RING_DESC 4096
1756 * Create memzone for HW rings. malloc can't be used as the physical address is
1757 * needed. If the memzone is already created, then this function returns a ptr
1760 static const struct rte_memzone * __attribute__((cold))
1761 ring_dma_zone_reserve(struct rte_eth_dev *dev, const char *ring_name,
1762 uint16_t queue_id, uint32_t ring_size, int socket_id)
1764 char z_name[RTE_MEMZONE_NAMESIZE];
1765 const struct rte_memzone *mz;
1767 snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
1768 dev->driver->pci_drv.name, ring_name,
1769 dev->data->port_id, queue_id);
1771 mz = rte_memzone_lookup(z_name);
1775 #ifdef RTE_LIBRTE_XEN_DOM0
1776 return rte_memzone_reserve_bounded(z_name, ring_size,
1777 socket_id, 0, IXGBE_ALIGN, RTE_PGSIZE_2M);
1779 return rte_memzone_reserve_aligned(z_name, ring_size,
1780 socket_id, 0, IXGBE_ALIGN);
1784 static void __attribute__((cold))
1785 ixgbe_tx_queue_release_mbufs(struct ixgbe_tx_queue *txq)
1789 if (txq->sw_ring != NULL) {
1790 for (i = 0; i < txq->nb_tx_desc; i++) {
1791 if (txq->sw_ring[i].mbuf != NULL) {
1792 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
1793 txq->sw_ring[i].mbuf = NULL;
1799 static void __attribute__((cold))
1800 ixgbe_tx_free_swring(struct ixgbe_tx_queue *txq)
1803 txq->sw_ring != NULL)
1804 rte_free(txq->sw_ring);
1807 static void __attribute__((cold))
1808 ixgbe_tx_queue_release(struct ixgbe_tx_queue *txq)
1810 if (txq != NULL && txq->ops != NULL) {
1811 txq->ops->release_mbufs(txq);
1812 txq->ops->free_swring(txq);
1817 void __attribute__((cold))
1818 ixgbe_dev_tx_queue_release(void *txq)
1820 ixgbe_tx_queue_release(txq);
1823 /* (Re)set dynamic ixgbe_tx_queue fields to defaults */
1824 static void __attribute__((cold))
1825 ixgbe_reset_tx_queue(struct ixgbe_tx_queue *txq)
1827 static const union ixgbe_adv_tx_desc zeroed_desc = {{0}};
1828 struct ixgbe_tx_entry *txe = txq->sw_ring;
1831 /* Zero out HW ring memory */
1832 for (i = 0; i < txq->nb_tx_desc; i++) {
1833 txq->tx_ring[i] = zeroed_desc;
1836 /* Initialize SW ring entries */
1837 prev = (uint16_t) (txq->nb_tx_desc - 1);
1838 for (i = 0; i < txq->nb_tx_desc; i++) {
1839 volatile union ixgbe_adv_tx_desc *txd = &txq->tx_ring[i];
1840 txd->wb.status = IXGBE_TXD_STAT_DD;
1843 txe[prev].next_id = i;
1847 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1848 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1851 txq->nb_tx_used = 0;
1853 * Always allow 1 descriptor to be un-allocated to avoid
1854 * a H/W race condition
1856 txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
1857 txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
1859 memset((void*)&txq->ctx_cache, 0,
1860 IXGBE_CTX_NUM * sizeof(struct ixgbe_advctx_info));
1863 static const struct ixgbe_txq_ops def_txq_ops = {
1864 .release_mbufs = ixgbe_tx_queue_release_mbufs,
1865 .free_swring = ixgbe_tx_free_swring,
1866 .reset = ixgbe_reset_tx_queue,
1869 /* Takes an ethdev and a queue and sets up the tx function to be used based on
1870 * the queue parameters. Used in tx_queue_setup by primary process and then
1871 * in dev_init by secondary process when attaching to an existing ethdev.
1873 void __attribute__((cold))
1874 ixgbe_set_tx_function(struct rte_eth_dev *dev, struct ixgbe_tx_queue *txq)
1876 /* Use a simple Tx queue (no offloads, no multi segs) if possible */
1877 if (((txq->txq_flags & IXGBE_SIMPLE_FLAGS) == IXGBE_SIMPLE_FLAGS)
1878 && (txq->tx_rs_thresh >= RTE_PMD_IXGBE_TX_MAX_BURST)) {
1879 PMD_INIT_LOG(INFO, "Using simple tx code path");
1880 #ifdef RTE_IXGBE_INC_VECTOR
1881 if (txq->tx_rs_thresh <= RTE_IXGBE_TX_MAX_FREE_BUF_SZ &&
1882 (rte_eal_process_type() != RTE_PROC_PRIMARY ||
1883 ixgbe_txq_vec_setup(txq) == 0)) {
1884 PMD_INIT_LOG(INFO, "Vector tx enabled.");
1885 dev->tx_pkt_burst = ixgbe_xmit_pkts_vec;
1888 dev->tx_pkt_burst = ixgbe_xmit_pkts_simple;
1890 PMD_INIT_LOG(INFO, "Using full-featured tx code path");
1892 " - txq_flags = %lx " "[IXGBE_SIMPLE_FLAGS=%lx]",
1893 (unsigned long)txq->txq_flags,
1894 (unsigned long)IXGBE_SIMPLE_FLAGS);
1896 " - tx_rs_thresh = %lu " "[RTE_PMD_IXGBE_TX_MAX_BURST=%lu]",
1897 (unsigned long)txq->tx_rs_thresh,
1898 (unsigned long)RTE_PMD_IXGBE_TX_MAX_BURST);
1899 dev->tx_pkt_burst = ixgbe_xmit_pkts;
1903 int __attribute__((cold))
1904 ixgbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
1907 unsigned int socket_id,
1908 const struct rte_eth_txconf *tx_conf)
1910 const struct rte_memzone *tz;
1911 struct ixgbe_tx_queue *txq;
1912 struct ixgbe_hw *hw;
1913 uint16_t tx_rs_thresh, tx_free_thresh;
1915 PMD_INIT_FUNC_TRACE();
1916 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1919 * Validate number of transmit descriptors.
1920 * It must not exceed hardware maximum, and must be multiple
1923 if (((nb_desc * sizeof(union ixgbe_adv_tx_desc)) % IXGBE_ALIGN) != 0 ||
1924 (nb_desc > IXGBE_MAX_RING_DESC) ||
1925 (nb_desc < IXGBE_MIN_RING_DESC)) {
1930 * The following two parameters control the setting of the RS bit on
1931 * transmit descriptors.
1932 * TX descriptors will have their RS bit set after txq->tx_rs_thresh
1933 * descriptors have been used.
1934 * The TX descriptor ring will be cleaned after txq->tx_free_thresh
1935 * descriptors are used or if the number of descriptors required
1936 * to transmit a packet is greater than the number of free TX
1938 * The following constraints must be satisfied:
1939 * tx_rs_thresh must be greater than 0.
1940 * tx_rs_thresh must be less than the size of the ring minus 2.
1941 * tx_rs_thresh must be less than or equal to tx_free_thresh.
1942 * tx_rs_thresh must be a divisor of the ring size.
1943 * tx_free_thresh must be greater than 0.
1944 * tx_free_thresh must be less than the size of the ring minus 3.
1945 * One descriptor in the TX ring is used as a sentinel to avoid a
1946 * H/W race condition, hence the maximum threshold constraints.
1947 * When set to zero use default values.
1949 tx_rs_thresh = (uint16_t)((tx_conf->tx_rs_thresh) ?
1950 tx_conf->tx_rs_thresh : DEFAULT_TX_RS_THRESH);
1951 tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
1952 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
1953 if (tx_rs_thresh >= (nb_desc - 2)) {
1954 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the number "
1955 "of TX descriptors minus 2. (tx_rs_thresh=%u "
1956 "port=%d queue=%d)", (unsigned int)tx_rs_thresh,
1957 (int)dev->data->port_id, (int)queue_idx);
1960 if (tx_free_thresh >= (nb_desc - 3)) {
1961 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
1962 "tx_free_thresh must be less than the number of "
1963 "TX descriptors minus 3. (tx_free_thresh=%u "
1964 "port=%d queue=%d)",
1965 (unsigned int)tx_free_thresh,
1966 (int)dev->data->port_id, (int)queue_idx);
1969 if (tx_rs_thresh > tx_free_thresh) {
1970 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or equal to "
1971 "tx_free_thresh. (tx_free_thresh=%u "
1972 "tx_rs_thresh=%u port=%d queue=%d)",
1973 (unsigned int)tx_free_thresh,
1974 (unsigned int)tx_rs_thresh,
1975 (int)dev->data->port_id,
1979 if ((nb_desc % tx_rs_thresh) != 0) {
1980 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
1981 "number of TX descriptors. (tx_rs_thresh=%u "
1982 "port=%d queue=%d)", (unsigned int)tx_rs_thresh,
1983 (int)dev->data->port_id, (int)queue_idx);
1988 * If rs_bit_thresh is greater than 1, then TX WTHRESH should be
1989 * set to 0. If WTHRESH is greater than zero, the RS bit is ignored
1990 * by the NIC and all descriptors are written back after the NIC
1991 * accumulates WTHRESH descriptors.
1993 if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
1994 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
1995 "tx_rs_thresh is greater than 1. (tx_rs_thresh=%u "
1996 "port=%d queue=%d)", (unsigned int)tx_rs_thresh,
1997 (int)dev->data->port_id, (int)queue_idx);
2001 /* Free memory prior to re-allocation if needed... */
2002 if (dev->data->tx_queues[queue_idx] != NULL) {
2003 ixgbe_tx_queue_release(dev->data->tx_queues[queue_idx]);
2004 dev->data->tx_queues[queue_idx] = NULL;
2007 /* First allocate the tx queue data structure */
2008 txq = rte_zmalloc_socket("ethdev TX queue", sizeof(struct ixgbe_tx_queue),
2009 RTE_CACHE_LINE_SIZE, socket_id);
2014 * Allocate TX ring hardware descriptors. A memzone large enough to
2015 * handle the maximum ring size is allocated in order to allow for
2016 * resizing in later calls to the queue setup function.
2018 tz = ring_dma_zone_reserve(dev, "tx_ring", queue_idx,
2019 sizeof(union ixgbe_adv_tx_desc) * IXGBE_MAX_RING_DESC,
2022 ixgbe_tx_queue_release(txq);
2026 txq->nb_tx_desc = nb_desc;
2027 txq->tx_rs_thresh = tx_rs_thresh;
2028 txq->tx_free_thresh = tx_free_thresh;
2029 txq->pthresh = tx_conf->tx_thresh.pthresh;
2030 txq->hthresh = tx_conf->tx_thresh.hthresh;
2031 txq->wthresh = tx_conf->tx_thresh.wthresh;
2032 txq->queue_id = queue_idx;
2033 txq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
2034 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
2035 txq->port_id = dev->data->port_id;
2036 txq->txq_flags = tx_conf->txq_flags;
2037 txq->ops = &def_txq_ops;
2038 txq->tx_deferred_start = tx_conf->tx_deferred_start;
2041 * Modification to set VFTDT for virtual function if vf is detected
2043 if (hw->mac.type == ixgbe_mac_82599_vf ||
2044 hw->mac.type == ixgbe_mac_X540_vf ||
2045 hw->mac.type == ixgbe_mac_X550_vf ||
2046 hw->mac.type == ixgbe_mac_X550EM_x_vf)
2047 txq->tdt_reg_addr = IXGBE_PCI_REG_ADDR(hw, IXGBE_VFTDT(queue_idx));
2049 txq->tdt_reg_addr = IXGBE_PCI_REG_ADDR(hw, IXGBE_TDT(txq->reg_idx));
2050 #ifndef RTE_LIBRTE_XEN_DOM0
2051 txq->tx_ring_phys_addr = (uint64_t) tz->phys_addr;
2053 txq->tx_ring_phys_addr = rte_mem_phy2mch(tz->memseg_id, tz->phys_addr);
2055 txq->tx_ring = (union ixgbe_adv_tx_desc *) tz->addr;
2057 /* Allocate software ring */
2058 txq->sw_ring = rte_zmalloc_socket("txq->sw_ring",
2059 sizeof(struct ixgbe_tx_entry) * nb_desc,
2060 RTE_CACHE_LINE_SIZE, socket_id);
2061 if (txq->sw_ring == NULL) {
2062 ixgbe_tx_queue_release(txq);
2065 PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64,
2066 txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr);
2068 /* set up vector or scalar TX function as appropriate */
2069 ixgbe_set_tx_function(dev, txq);
2071 txq->ops->reset(txq);
2073 dev->data->tx_queues[queue_idx] = txq;
2080 * ixgbe_free_sc_cluster - free the not-yet-completed scattered cluster
2082 * The "next" pointer of the last segment of (not-yet-completed) RSC clusters
2083 * in the sw_rsc_ring is not set to NULL but rather points to the next
2084 * mbuf of this RSC aggregation (that has not been completed yet and still
2085 * resides on the HW ring). So, instead of calling for rte_pktmbuf_free() we
2086 * will just free first "nb_segs" segments of the cluster explicitly by calling
2087 * an rte_pktmbuf_free_seg().
2089 * @m scattered cluster head
2091 static void __attribute__((cold))
2092 ixgbe_free_sc_cluster(struct rte_mbuf *m)
2094 uint8_t i, nb_segs = m->nb_segs;
2095 struct rte_mbuf *next_seg;
2097 for (i = 0; i < nb_segs; i++) {
2099 rte_pktmbuf_free_seg(m);
2104 static void __attribute__((cold))
2105 ixgbe_rx_queue_release_mbufs(struct ixgbe_rx_queue *rxq)
2109 if (rxq->sw_ring != NULL) {
2110 for (i = 0; i < rxq->nb_rx_desc; i++) {
2111 if (rxq->sw_ring[i].mbuf != NULL &&
2112 rte_mbuf_refcnt_read(rxq->sw_ring[i].mbuf)) {
2113 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2114 rxq->sw_ring[i].mbuf = NULL;
2117 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2118 if (rxq->rx_nb_avail) {
2119 for (i = 0; i < rxq->rx_nb_avail; ++i) {
2120 struct rte_mbuf *mb;
2121 mb = rxq->rx_stage[rxq->rx_next_avail + i];
2122 rte_pktmbuf_free_seg(mb);
2124 rxq->rx_nb_avail = 0;
2129 if (rxq->sw_sc_ring)
2130 for (i = 0; i < rxq->nb_rx_desc; i++)
2131 if (rxq->sw_sc_ring[i].fbuf) {
2132 ixgbe_free_sc_cluster(rxq->sw_sc_ring[i].fbuf);
2133 rxq->sw_sc_ring[i].fbuf = NULL;
2137 static void __attribute__((cold))
2138 ixgbe_rx_queue_release(struct ixgbe_rx_queue *rxq)
2141 ixgbe_rx_queue_release_mbufs(rxq);
2142 rte_free(rxq->sw_ring);
2143 rte_free(rxq->sw_sc_ring);
2148 void __attribute__((cold))
2149 ixgbe_dev_rx_queue_release(void *rxq)
2151 ixgbe_rx_queue_release(rxq);
2155 * Check if Rx Burst Bulk Alloc function can be used.
2157 * 0: the preconditions are satisfied and the bulk allocation function
2159 * -EINVAL: the preconditions are NOT satisfied and the default Rx burst
2160 * function must be used.
2162 static inline int __attribute__((cold))
2163 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2164 check_rx_burst_bulk_alloc_preconditions(struct ixgbe_rx_queue *rxq)
2166 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct ixgbe_rx_queue *rxq)
2172 * Make sure the following pre-conditions are satisfied:
2173 * rxq->rx_free_thresh >= RTE_PMD_IXGBE_RX_MAX_BURST
2174 * rxq->rx_free_thresh < rxq->nb_rx_desc
2175 * (rxq->nb_rx_desc % rxq->rx_free_thresh) == 0
2176 * rxq->nb_rx_desc<(IXGBE_MAX_RING_DESC-RTE_PMD_IXGBE_RX_MAX_BURST)
2177 * Scattered packets are not supported. This should be checked
2178 * outside of this function.
2180 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2181 if (!(rxq->rx_free_thresh >= RTE_PMD_IXGBE_RX_MAX_BURST)) {
2182 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2183 "rxq->rx_free_thresh=%d, "
2184 "RTE_PMD_IXGBE_RX_MAX_BURST=%d",
2185 rxq->rx_free_thresh, RTE_PMD_IXGBE_RX_MAX_BURST);
2187 } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
2188 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2189 "rxq->rx_free_thresh=%d, "
2190 "rxq->nb_rx_desc=%d",
2191 rxq->rx_free_thresh, rxq->nb_rx_desc);
2193 } else if (!((rxq->nb_rx_desc % rxq->rx_free_thresh) == 0)) {
2194 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2195 "rxq->nb_rx_desc=%d, "
2196 "rxq->rx_free_thresh=%d",
2197 rxq->nb_rx_desc, rxq->rx_free_thresh);
2199 } else if (!(rxq->nb_rx_desc <
2200 (IXGBE_MAX_RING_DESC - RTE_PMD_IXGBE_RX_MAX_BURST))) {
2201 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2202 "rxq->nb_rx_desc=%d, "
2203 "IXGBE_MAX_RING_DESC=%d, "
2204 "RTE_PMD_IXGBE_RX_MAX_BURST=%d",
2205 rxq->nb_rx_desc, IXGBE_MAX_RING_DESC,
2206 RTE_PMD_IXGBE_RX_MAX_BURST);
2216 /* Reset dynamic ixgbe_rx_queue fields back to defaults */
2217 static void __attribute__((cold))
2218 ixgbe_reset_rx_queue(struct ixgbe_adapter *adapter, struct ixgbe_rx_queue *rxq)
2220 static const union ixgbe_adv_rx_desc zeroed_desc = {{0}};
2222 uint16_t len = rxq->nb_rx_desc;
2225 * By default, the Rx queue setup function allocates enough memory for
2226 * IXGBE_MAX_RING_DESC. The Rx Burst bulk allocation function requires
2227 * extra memory at the end of the descriptor ring to be zero'd out. A
2228 * pre-condition for using the Rx burst bulk alloc function is that the
2229 * number of descriptors is less than or equal to
2230 * (IXGBE_MAX_RING_DESC - RTE_PMD_IXGBE_RX_MAX_BURST). Check all the
2231 * constraints here to see if we need to zero out memory after the end
2232 * of the H/W descriptor ring.
2234 if (adapter->rx_bulk_alloc_allowed)
2235 /* zero out extra memory */
2236 len += RTE_PMD_IXGBE_RX_MAX_BURST;
2239 * Zero out HW ring memory. Zero out extra memory at the end of
2240 * the H/W ring so look-ahead logic in Rx Burst bulk alloc function
2241 * reads extra memory as zeros.
2243 for (i = 0; i < len; i++) {
2244 rxq->rx_ring[i] = zeroed_desc;
2247 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2249 * initialize extra software ring entries. Space for these extra
2250 * entries is always allocated
2252 memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2253 for (i = rxq->nb_rx_desc; i < len; ++i) {
2254 rxq->sw_ring[i].mbuf = &rxq->fake_mbuf;
2257 rxq->rx_nb_avail = 0;
2258 rxq->rx_next_avail = 0;
2259 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2260 #endif /* RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC */
2262 rxq->nb_rx_hold = 0;
2263 rxq->pkt_first_seg = NULL;
2264 rxq->pkt_last_seg = NULL;
2267 int __attribute__((cold))
2268 ixgbe_dev_rx_queue_setup(struct rte_eth_dev *dev,
2271 unsigned int socket_id,
2272 const struct rte_eth_rxconf *rx_conf,
2273 struct rte_mempool *mp)
2275 const struct rte_memzone *rz;
2276 struct ixgbe_rx_queue *rxq;
2277 struct ixgbe_hw *hw;
2279 struct ixgbe_adapter *adapter =
2280 (struct ixgbe_adapter *)dev->data->dev_private;
2282 PMD_INIT_FUNC_TRACE();
2283 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2286 * Validate number of receive descriptors.
2287 * It must not exceed hardware maximum, and must be multiple
2290 if (((nb_desc * sizeof(union ixgbe_adv_rx_desc)) % IXGBE_ALIGN) != 0 ||
2291 (nb_desc > IXGBE_MAX_RING_DESC) ||
2292 (nb_desc < IXGBE_MIN_RING_DESC)) {
2296 /* Free memory prior to re-allocation if needed... */
2297 if (dev->data->rx_queues[queue_idx] != NULL) {
2298 ixgbe_rx_queue_release(dev->data->rx_queues[queue_idx]);
2299 dev->data->rx_queues[queue_idx] = NULL;
2302 /* First allocate the rx queue data structure */
2303 rxq = rte_zmalloc_socket("ethdev RX queue", sizeof(struct ixgbe_rx_queue),
2304 RTE_CACHE_LINE_SIZE, socket_id);
2308 rxq->nb_rx_desc = nb_desc;
2309 rxq->rx_free_thresh = rx_conf->rx_free_thresh;
2310 rxq->queue_id = queue_idx;
2311 rxq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
2312 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
2313 rxq->port_id = dev->data->port_id;
2314 rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ?
2316 rxq->drop_en = rx_conf->rx_drop_en;
2317 rxq->rx_deferred_start = rx_conf->rx_deferred_start;
2320 * Allocate RX ring hardware descriptors. A memzone large enough to
2321 * handle the maximum ring size is allocated in order to allow for
2322 * resizing in later calls to the queue setup function.
2324 rz = ring_dma_zone_reserve(dev, "rx_ring", queue_idx,
2325 RX_RING_SZ, socket_id);
2327 ixgbe_rx_queue_release(rxq);
2332 * Zero init all the descriptors in the ring.
2334 memset (rz->addr, 0, RX_RING_SZ);
2337 * Modified to setup VFRDT for Virtual Function
2339 if (hw->mac.type == ixgbe_mac_82599_vf ||
2340 hw->mac.type == ixgbe_mac_X540_vf ||
2341 hw->mac.type == ixgbe_mac_X550_vf ||
2342 hw->mac.type == ixgbe_mac_X550EM_x_vf) {
2344 IXGBE_PCI_REG_ADDR(hw, IXGBE_VFRDT(queue_idx));
2346 IXGBE_PCI_REG_ADDR(hw, IXGBE_VFRDH(queue_idx));
2350 IXGBE_PCI_REG_ADDR(hw, IXGBE_RDT(rxq->reg_idx));
2352 IXGBE_PCI_REG_ADDR(hw, IXGBE_RDH(rxq->reg_idx));
2354 #ifndef RTE_LIBRTE_XEN_DOM0
2355 rxq->rx_ring_phys_addr = (uint64_t) rz->phys_addr;
2357 rxq->rx_ring_phys_addr = rte_mem_phy2mch(rz->memseg_id, rz->phys_addr);
2359 rxq->rx_ring = (union ixgbe_adv_rx_desc *) rz->addr;
2362 * Certain constraints must be met in order to use the bulk buffer
2363 * allocation Rx burst function. If any of Rx queues doesn't meet them
2364 * the feature should be disabled for the whole port.
2366 if (check_rx_burst_bulk_alloc_preconditions(rxq)) {
2367 PMD_INIT_LOG(DEBUG, "queue[%d] doesn't meet Rx Bulk Alloc "
2368 "preconditions - canceling the feature for "
2369 "the whole port[%d]",
2370 rxq->queue_id, rxq->port_id);
2371 adapter->rx_bulk_alloc_allowed = false;
2375 * Allocate software ring. Allow for space at the end of the
2376 * S/W ring to make sure look-ahead logic in bulk alloc Rx burst
2377 * function does not access an invalid memory region.
2380 if (adapter->rx_bulk_alloc_allowed)
2381 len += RTE_PMD_IXGBE_RX_MAX_BURST;
2383 rxq->sw_ring = rte_zmalloc_socket("rxq->sw_ring",
2384 sizeof(struct ixgbe_rx_entry) * len,
2385 RTE_CACHE_LINE_SIZE, socket_id);
2386 if (!rxq->sw_ring) {
2387 ixgbe_rx_queue_release(rxq);
2392 * Always allocate even if it's not going to be needed in order to
2393 * simplify the code.
2395 * This ring is used in LRO and Scattered Rx cases and Scattered Rx may
2396 * be requested in ixgbe_dev_rx_init(), which is called later from
2400 rte_zmalloc_socket("rxq->sw_sc_ring",
2401 sizeof(struct ixgbe_scattered_rx_entry) * len,
2402 RTE_CACHE_LINE_SIZE, socket_id);
2403 if (!rxq->sw_sc_ring) {
2404 ixgbe_rx_queue_release(rxq);
2408 PMD_INIT_LOG(DEBUG, "sw_ring=%p sw_sc_ring=%p hw_ring=%p "
2409 "dma_addr=0x%"PRIx64,
2410 rxq->sw_ring, rxq->sw_sc_ring, rxq->rx_ring,
2411 rxq->rx_ring_phys_addr);
2413 if (!rte_is_power_of_2(nb_desc)) {
2414 PMD_INIT_LOG(DEBUG, "queue[%d] doesn't meet Vector Rx "
2415 "preconditions - canceling the feature for "
2416 "the whole port[%d]",
2417 rxq->queue_id, rxq->port_id);
2418 adapter->rx_vec_allowed = false;
2420 ixgbe_rxq_vec_setup(rxq);
2422 dev->data->rx_queues[queue_idx] = rxq;
2424 ixgbe_reset_rx_queue(adapter, rxq);
2430 ixgbe_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
2432 #define IXGBE_RXQ_SCAN_INTERVAL 4
2433 volatile union ixgbe_adv_rx_desc *rxdp;
2434 struct ixgbe_rx_queue *rxq;
2437 if (rx_queue_id >= dev->data->nb_rx_queues) {
2438 PMD_RX_LOG(ERR, "Invalid RX queue id=%d", rx_queue_id);
2442 rxq = dev->data->rx_queues[rx_queue_id];
2443 rxdp = &(rxq->rx_ring[rxq->rx_tail]);
2445 while ((desc < rxq->nb_rx_desc) &&
2446 (rxdp->wb.upper.status_error & IXGBE_RXDADV_STAT_DD)) {
2447 desc += IXGBE_RXQ_SCAN_INTERVAL;
2448 rxdp += IXGBE_RXQ_SCAN_INTERVAL;
2449 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
2450 rxdp = &(rxq->rx_ring[rxq->rx_tail +
2451 desc - rxq->nb_rx_desc]);
2458 ixgbe_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
2460 volatile union ixgbe_adv_rx_desc *rxdp;
2461 struct ixgbe_rx_queue *rxq = rx_queue;
2464 if (unlikely(offset >= rxq->nb_rx_desc))
2466 desc = rxq->rx_tail + offset;
2467 if (desc >= rxq->nb_rx_desc)
2468 desc -= rxq->nb_rx_desc;
2470 rxdp = &rxq->rx_ring[desc];
2471 return !!(rxdp->wb.upper.status_error & IXGBE_RXDADV_STAT_DD);
2474 void __attribute__((cold))
2475 ixgbe_dev_clear_queues(struct rte_eth_dev *dev)
2478 struct ixgbe_adapter *adapter =
2479 (struct ixgbe_adapter *)dev->data->dev_private;
2481 PMD_INIT_FUNC_TRACE();
2483 for (i = 0; i < dev->data->nb_tx_queues; i++) {
2484 struct ixgbe_tx_queue *txq = dev->data->tx_queues[i];
2486 txq->ops->release_mbufs(txq);
2487 txq->ops->reset(txq);
2491 for (i = 0; i < dev->data->nb_rx_queues; i++) {
2492 struct ixgbe_rx_queue *rxq = dev->data->rx_queues[i];
2494 ixgbe_rx_queue_release_mbufs(rxq);
2495 ixgbe_reset_rx_queue(adapter, rxq);
2500 /*********************************************************************
2502 * Device RX/TX init functions
2504 **********************************************************************/
2507 * Receive Side Scaling (RSS)
2508 * See section 7.1.2.8 in the following document:
2509 * "Intel 82599 10 GbE Controller Datasheet" - Revision 2.1 October 2009
2512 * The source and destination IP addresses of the IP header and the source
2513 * and destination ports of TCP/UDP headers, if any, of received packets are
2514 * hashed against a configurable random key to compute a 32-bit RSS hash result.
2515 * The seven (7) LSBs of the 32-bit hash result are used as an index into a
2516 * 128-entry redirection table (RETA). Each entry of the RETA provides a 3-bit
2517 * RSS output index which is used as the RX queue index where to store the
2519 * The following output is supplied in the RX write-back descriptor:
2520 * - 32-bit result of the Microsoft RSS hash function,
2521 * - 4-bit RSS type field.
2525 * RSS random key supplied in section 7.1.2.8.3 of the Intel 82599 datasheet.
2526 * Used as the default key.
2528 static uint8_t rss_intel_key[40] = {
2529 0x6D, 0x5A, 0x56, 0xDA, 0x25, 0x5B, 0x0E, 0xC2,
2530 0x41, 0x67, 0x25, 0x3D, 0x43, 0xA3, 0x8F, 0xB0,
2531 0xD0, 0xCA, 0x2B, 0xCB, 0xAE, 0x7B, 0x30, 0xB4,
2532 0x77, 0xCB, 0x2D, 0xA3, 0x80, 0x30, 0xF2, 0x0C,
2533 0x6A, 0x42, 0xB7, 0x3B, 0xBE, 0xAC, 0x01, 0xFA,
2537 ixgbe_rss_disable(struct rte_eth_dev *dev)
2539 struct ixgbe_hw *hw;
2542 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2543 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
2544 mrqc &= ~IXGBE_MRQC_RSSEN;
2545 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
2549 ixgbe_hw_rss_hash_set(struct ixgbe_hw *hw, struct rte_eth_rss_conf *rss_conf)
2557 hash_key = rss_conf->rss_key;
2558 if (hash_key != NULL) {
2559 /* Fill in RSS hash key */
2560 for (i = 0; i < 10; i++) {
2561 rss_key = hash_key[(i * 4)];
2562 rss_key |= hash_key[(i * 4) + 1] << 8;
2563 rss_key |= hash_key[(i * 4) + 2] << 16;
2564 rss_key |= hash_key[(i * 4) + 3] << 24;
2565 IXGBE_WRITE_REG_ARRAY(hw, IXGBE_RSSRK(0), i, rss_key);
2569 /* Set configured hashing protocols in MRQC register */
2570 rss_hf = rss_conf->rss_hf;
2571 mrqc = IXGBE_MRQC_RSSEN; /* Enable RSS */
2572 if (rss_hf & ETH_RSS_IPV4)
2573 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV4;
2574 if (rss_hf & ETH_RSS_NONFRAG_IPV4_TCP)
2575 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV4_TCP;
2576 if (rss_hf & ETH_RSS_IPV6)
2577 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6;
2578 if (rss_hf & ETH_RSS_IPV6_EX)
2579 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_EX;
2580 if (rss_hf & ETH_RSS_NONFRAG_IPV6_TCP)
2581 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_TCP;
2582 if (rss_hf & ETH_RSS_IPV6_TCP_EX)
2583 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_EX_TCP;
2584 if (rss_hf & ETH_RSS_NONFRAG_IPV4_UDP)
2585 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV4_UDP;
2586 if (rss_hf & ETH_RSS_NONFRAG_IPV6_UDP)
2587 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_UDP;
2588 if (rss_hf & ETH_RSS_IPV6_UDP_EX)
2589 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_EX_UDP;
2590 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
2594 ixgbe_dev_rss_hash_update(struct rte_eth_dev *dev,
2595 struct rte_eth_rss_conf *rss_conf)
2597 struct ixgbe_hw *hw;
2601 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2604 * Excerpt from section 7.1.2.8 Receive-Side Scaling (RSS):
2605 * "RSS enabling cannot be done dynamically while it must be
2606 * preceded by a software reset"
2607 * Before changing anything, first check that the update RSS operation
2608 * does not attempt to disable RSS, if RSS was enabled at
2609 * initialization time, or does not attempt to enable RSS, if RSS was
2610 * disabled at initialization time.
2612 rss_hf = rss_conf->rss_hf & IXGBE_RSS_OFFLOAD_ALL;
2613 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
2614 if (!(mrqc & IXGBE_MRQC_RSSEN)) { /* RSS disabled */
2615 if (rss_hf != 0) /* Enable RSS */
2617 return 0; /* Nothing to do */
2620 if (rss_hf == 0) /* Disable RSS */
2622 ixgbe_hw_rss_hash_set(hw, rss_conf);
2627 ixgbe_dev_rss_hash_conf_get(struct rte_eth_dev *dev,
2628 struct rte_eth_rss_conf *rss_conf)
2630 struct ixgbe_hw *hw;
2637 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2638 hash_key = rss_conf->rss_key;
2639 if (hash_key != NULL) {
2640 /* Return RSS hash key */
2641 for (i = 0; i < 10; i++) {
2642 rss_key = IXGBE_READ_REG_ARRAY(hw, IXGBE_RSSRK(0), i);
2643 hash_key[(i * 4)] = rss_key & 0x000000FF;
2644 hash_key[(i * 4) + 1] = (rss_key >> 8) & 0x000000FF;
2645 hash_key[(i * 4) + 2] = (rss_key >> 16) & 0x000000FF;
2646 hash_key[(i * 4) + 3] = (rss_key >> 24) & 0x000000FF;
2650 /* Get RSS functions configured in MRQC register */
2651 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
2652 if ((mrqc & IXGBE_MRQC_RSSEN) == 0) { /* RSS is disabled */
2653 rss_conf->rss_hf = 0;
2657 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV4)
2658 rss_hf |= ETH_RSS_IPV4;
2659 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV4_TCP)
2660 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
2661 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6)
2662 rss_hf |= ETH_RSS_IPV6;
2663 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_EX)
2664 rss_hf |= ETH_RSS_IPV6_EX;
2665 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_TCP)
2666 rss_hf |= ETH_RSS_NONFRAG_IPV6_TCP;
2667 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_EX_TCP)
2668 rss_hf |= ETH_RSS_IPV6_TCP_EX;
2669 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV4_UDP)
2670 rss_hf |= ETH_RSS_NONFRAG_IPV4_UDP;
2671 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_UDP)
2672 rss_hf |= ETH_RSS_NONFRAG_IPV6_UDP;
2673 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_EX_UDP)
2674 rss_hf |= ETH_RSS_IPV6_UDP_EX;
2675 rss_conf->rss_hf = rss_hf;
2680 ixgbe_rss_configure(struct rte_eth_dev *dev)
2682 struct rte_eth_rss_conf rss_conf;
2683 struct ixgbe_hw *hw;
2688 PMD_INIT_FUNC_TRACE();
2689 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2692 * Fill in redirection table
2693 * The byte-swap is needed because NIC registers are in
2694 * little-endian order.
2697 for (i = 0, j = 0; i < 128; i++, j++) {
2698 if (j == dev->data->nb_rx_queues)
2700 reta = (reta << 8) | j;
2702 IXGBE_WRITE_REG(hw, IXGBE_RETA(i >> 2),
2707 * Configure the RSS key and the RSS protocols used to compute
2708 * the RSS hash of input packets.
2710 rss_conf = dev->data->dev_conf.rx_adv_conf.rss_conf;
2711 if ((rss_conf.rss_hf & IXGBE_RSS_OFFLOAD_ALL) == 0) {
2712 ixgbe_rss_disable(dev);
2715 if (rss_conf.rss_key == NULL)
2716 rss_conf.rss_key = rss_intel_key; /* Default hash key */
2717 ixgbe_hw_rss_hash_set(hw, &rss_conf);
2720 #define NUM_VFTA_REGISTERS 128
2721 #define NIC_RX_BUFFER_SIZE 0x200
2724 ixgbe_vmdq_dcb_configure(struct rte_eth_dev *dev)
2726 struct rte_eth_vmdq_dcb_conf *cfg;
2727 struct ixgbe_hw *hw;
2728 enum rte_eth_nb_pools num_pools;
2729 uint32_t mrqc, vt_ctl, queue_mapping, vlanctrl;
2731 uint8_t nb_tcs; /* number of traffic classes */
2734 PMD_INIT_FUNC_TRACE();
2735 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2736 cfg = &dev->data->dev_conf.rx_adv_conf.vmdq_dcb_conf;
2737 num_pools = cfg->nb_queue_pools;
2738 /* Check we have a valid number of pools */
2739 if (num_pools != ETH_16_POOLS && num_pools != ETH_32_POOLS) {
2740 ixgbe_rss_disable(dev);
2743 /* 16 pools -> 8 traffic classes, 32 pools -> 4 traffic classes */
2744 nb_tcs = (uint8_t)(ETH_VMDQ_DCB_NUM_QUEUES / (int)num_pools);
2748 * split rx buffer up into sections, each for 1 traffic class
2750 pbsize = (uint16_t)(NIC_RX_BUFFER_SIZE / nb_tcs);
2751 for (i = 0 ; i < nb_tcs; i++) {
2752 uint32_t rxpbsize = IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(i));
2753 rxpbsize &= (~(0x3FF << IXGBE_RXPBSIZE_SHIFT));
2754 /* clear 10 bits. */
2755 rxpbsize |= (pbsize << IXGBE_RXPBSIZE_SHIFT); /* set value */
2756 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), rxpbsize);
2758 /* zero alloc all unused TCs */
2759 for (i = nb_tcs; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2760 uint32_t rxpbsize = IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(i));
2761 rxpbsize &= (~( 0x3FF << IXGBE_RXPBSIZE_SHIFT ));
2762 /* clear 10 bits. */
2763 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), rxpbsize);
2766 /* MRQC: enable vmdq and dcb */
2767 mrqc = ((num_pools == ETH_16_POOLS) ? \
2768 IXGBE_MRQC_VMDQRT8TCEN : IXGBE_MRQC_VMDQRT4TCEN );
2769 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
2771 /* PFVTCTL: turn on virtualisation and set the default pool */
2772 vt_ctl = IXGBE_VT_CTL_VT_ENABLE | IXGBE_VT_CTL_REPLEN;
2773 if (cfg->enable_default_pool) {
2774 vt_ctl |= (cfg->default_pool << IXGBE_VT_CTL_POOL_SHIFT);
2776 vt_ctl |= IXGBE_VT_CTL_DIS_DEFPL;
2779 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, vt_ctl);
2781 /* RTRUP2TC: mapping user priorities to traffic classes (TCs) */
2783 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++)
2785 * mapping is done with 3 bits per priority,
2786 * so shift by i*3 each time
2788 queue_mapping |= ((cfg->dcb_queue[i] & 0x07) << (i * 3));
2790 IXGBE_WRITE_REG(hw, IXGBE_RTRUP2TC, queue_mapping);
2792 /* RTRPCS: DCB related */
2793 IXGBE_WRITE_REG(hw, IXGBE_RTRPCS, IXGBE_RMCS_RRM);
2795 /* VLNCTRL: enable vlan filtering and allow all vlan tags through */
2796 vlanctrl = IXGBE_READ_REG(hw, IXGBE_VLNCTRL);
2797 vlanctrl |= IXGBE_VLNCTRL_VFE ; /* enable vlan filters */
2798 IXGBE_WRITE_REG(hw, IXGBE_VLNCTRL, vlanctrl);
2800 /* VFTA - enable all vlan filters */
2801 for (i = 0; i < NUM_VFTA_REGISTERS; i++) {
2802 IXGBE_WRITE_REG(hw, IXGBE_VFTA(i), 0xFFFFFFFF);
2805 /* VFRE: pool enabling for receive - 16 or 32 */
2806 IXGBE_WRITE_REG(hw, IXGBE_VFRE(0), \
2807 num_pools == ETH_16_POOLS ? 0xFFFF : 0xFFFFFFFF);
2810 * MPSAR - allow pools to read specific mac addresses
2811 * In this case, all pools should be able to read from mac addr 0
2813 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_LO(0), 0xFFFFFFFF);
2814 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_HI(0), 0xFFFFFFFF);
2816 /* PFVLVF, PFVLVFB: set up filters for vlan tags as configured */
2817 for (i = 0; i < cfg->nb_pool_maps; i++) {
2818 /* set vlan id in VF register and set the valid bit */
2819 IXGBE_WRITE_REG(hw, IXGBE_VLVF(i), (IXGBE_VLVF_VIEN | \
2820 (cfg->pool_map[i].vlan_id & 0xFFF)));
2822 * Put the allowed pools in VFB reg. As we only have 16 or 32
2823 * pools, we only need to use the first half of the register
2826 IXGBE_WRITE_REG(hw, IXGBE_VLVFB(i*2), cfg->pool_map[i].pools);
2831 * ixgbe_dcb_config_tx_hw_config - Configure general DCB TX parameters
2832 * @hw: pointer to hardware structure
2833 * @dcb_config: pointer to ixgbe_dcb_config structure
2836 ixgbe_dcb_tx_hw_config(struct ixgbe_hw *hw,
2837 struct ixgbe_dcb_config *dcb_config)
2842 PMD_INIT_FUNC_TRACE();
2843 if (hw->mac.type != ixgbe_mac_82598EB) {
2844 /* Disable the Tx desc arbiter so that MTQC can be changed */
2845 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
2846 reg |= IXGBE_RTTDCS_ARBDIS;
2847 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
2849 /* Enable DCB for Tx with 8 TCs */
2850 if (dcb_config->num_tcs.pg_tcs == 8) {
2851 reg = IXGBE_MTQC_RT_ENA | IXGBE_MTQC_8TC_8TQ;
2854 reg = IXGBE_MTQC_RT_ENA | IXGBE_MTQC_4TC_4TQ;
2856 if (dcb_config->vt_mode)
2857 reg |= IXGBE_MTQC_VT_ENA;
2858 IXGBE_WRITE_REG(hw, IXGBE_MTQC, reg);
2860 /* Disable drop for all queues */
2861 for (q = 0; q < 128; q++)
2862 IXGBE_WRITE_REG(hw, IXGBE_QDE,
2863 (IXGBE_QDE_WRITE | (q << IXGBE_QDE_IDX_SHIFT)));
2865 /* Enable the Tx desc arbiter */
2866 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
2867 reg &= ~IXGBE_RTTDCS_ARBDIS;
2868 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
2870 /* Enable Security TX Buffer IFG for DCB */
2871 reg = IXGBE_READ_REG(hw, IXGBE_SECTXMINIFG);
2872 reg |= IXGBE_SECTX_DCB;
2873 IXGBE_WRITE_REG(hw, IXGBE_SECTXMINIFG, reg);
2879 * ixgbe_vmdq_dcb_hw_tx_config - Configure general VMDQ+DCB TX parameters
2880 * @dev: pointer to rte_eth_dev structure
2881 * @dcb_config: pointer to ixgbe_dcb_config structure
2884 ixgbe_vmdq_dcb_hw_tx_config(struct rte_eth_dev *dev,
2885 struct ixgbe_dcb_config *dcb_config)
2887 struct rte_eth_vmdq_dcb_tx_conf *vmdq_tx_conf =
2888 &dev->data->dev_conf.tx_adv_conf.vmdq_dcb_tx_conf;
2889 struct ixgbe_hw *hw =
2890 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2892 PMD_INIT_FUNC_TRACE();
2893 if (hw->mac.type != ixgbe_mac_82598EB)
2894 /*PF VF Transmit Enable*/
2895 IXGBE_WRITE_REG(hw, IXGBE_VFTE(0),
2896 vmdq_tx_conf->nb_queue_pools == ETH_16_POOLS ? 0xFFFF : 0xFFFFFFFF);
2898 /*Configure general DCB TX parameters*/
2899 ixgbe_dcb_tx_hw_config(hw,dcb_config);
2904 ixgbe_vmdq_dcb_rx_config(struct rte_eth_dev *dev,
2905 struct ixgbe_dcb_config *dcb_config)
2907 struct rte_eth_vmdq_dcb_conf *vmdq_rx_conf =
2908 &dev->data->dev_conf.rx_adv_conf.vmdq_dcb_conf;
2909 struct ixgbe_dcb_tc_config *tc;
2912 /* convert rte_eth_conf.rx_adv_conf to struct ixgbe_dcb_config */
2913 if (vmdq_rx_conf->nb_queue_pools == ETH_16_POOLS ) {
2914 dcb_config->num_tcs.pg_tcs = ETH_8_TCS;
2915 dcb_config->num_tcs.pfc_tcs = ETH_8_TCS;
2918 dcb_config->num_tcs.pg_tcs = ETH_4_TCS;
2919 dcb_config->num_tcs.pfc_tcs = ETH_4_TCS;
2921 /* User Priority to Traffic Class mapping */
2922 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2923 j = vmdq_rx_conf->dcb_queue[i];
2924 tc = &dcb_config->tc_config[j];
2925 tc->path[IXGBE_DCB_RX_CONFIG].up_to_tc_bitmap =
2931 ixgbe_dcb_vt_tx_config(struct rte_eth_dev *dev,
2932 struct ixgbe_dcb_config *dcb_config)
2934 struct rte_eth_vmdq_dcb_tx_conf *vmdq_tx_conf =
2935 &dev->data->dev_conf.tx_adv_conf.vmdq_dcb_tx_conf;
2936 struct ixgbe_dcb_tc_config *tc;
2939 /* convert rte_eth_conf.rx_adv_conf to struct ixgbe_dcb_config */
2940 if (vmdq_tx_conf->nb_queue_pools == ETH_16_POOLS ) {
2941 dcb_config->num_tcs.pg_tcs = ETH_8_TCS;
2942 dcb_config->num_tcs.pfc_tcs = ETH_8_TCS;
2945 dcb_config->num_tcs.pg_tcs = ETH_4_TCS;
2946 dcb_config->num_tcs.pfc_tcs = ETH_4_TCS;
2949 /* User Priority to Traffic Class mapping */
2950 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2951 j = vmdq_tx_conf->dcb_queue[i];
2952 tc = &dcb_config->tc_config[j];
2953 tc->path[IXGBE_DCB_TX_CONFIG].up_to_tc_bitmap =
2960 ixgbe_dcb_rx_config(struct rte_eth_dev *dev,
2961 struct ixgbe_dcb_config *dcb_config)
2963 struct rte_eth_dcb_rx_conf *rx_conf =
2964 &dev->data->dev_conf.rx_adv_conf.dcb_rx_conf;
2965 struct ixgbe_dcb_tc_config *tc;
2968 dcb_config->num_tcs.pg_tcs = (uint8_t)rx_conf->nb_tcs;
2969 dcb_config->num_tcs.pfc_tcs = (uint8_t)rx_conf->nb_tcs;
2971 /* User Priority to Traffic Class mapping */
2972 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2973 j = rx_conf->dcb_queue[i];
2974 tc = &dcb_config->tc_config[j];
2975 tc->path[IXGBE_DCB_RX_CONFIG].up_to_tc_bitmap =
2981 ixgbe_dcb_tx_config(struct rte_eth_dev *dev,
2982 struct ixgbe_dcb_config *dcb_config)
2984 struct rte_eth_dcb_tx_conf *tx_conf =
2985 &dev->data->dev_conf.tx_adv_conf.dcb_tx_conf;
2986 struct ixgbe_dcb_tc_config *tc;
2989 dcb_config->num_tcs.pg_tcs = (uint8_t)tx_conf->nb_tcs;
2990 dcb_config->num_tcs.pfc_tcs = (uint8_t)tx_conf->nb_tcs;
2992 /* User Priority to Traffic Class mapping */
2993 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2994 j = tx_conf->dcb_queue[i];
2995 tc = &dcb_config->tc_config[j];
2996 tc->path[IXGBE_DCB_TX_CONFIG].up_to_tc_bitmap =
3002 * ixgbe_dcb_rx_hw_config - Configure general DCB RX HW parameters
3003 * @hw: pointer to hardware structure
3004 * @dcb_config: pointer to ixgbe_dcb_config structure
3007 ixgbe_dcb_rx_hw_config(struct ixgbe_hw *hw,
3008 struct ixgbe_dcb_config *dcb_config)
3014 PMD_INIT_FUNC_TRACE();
3016 * Disable the arbiter before changing parameters
3017 * (always enable recycle mode; WSP)
3019 reg = IXGBE_RTRPCS_RRM | IXGBE_RTRPCS_RAC | IXGBE_RTRPCS_ARBDIS;
3020 IXGBE_WRITE_REG(hw, IXGBE_RTRPCS, reg);
3022 if (hw->mac.type != ixgbe_mac_82598EB) {
3023 reg = IXGBE_READ_REG(hw, IXGBE_MRQC);
3024 if (dcb_config->num_tcs.pg_tcs == 4) {
3025 if (dcb_config->vt_mode)
3026 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3027 IXGBE_MRQC_VMDQRT4TCEN;
3029 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, 0);
3030 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3034 if (dcb_config->num_tcs.pg_tcs == 8) {
3035 if (dcb_config->vt_mode)
3036 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3037 IXGBE_MRQC_VMDQRT8TCEN;
3039 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, 0);
3040 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3045 IXGBE_WRITE_REG(hw, IXGBE_MRQC, reg);
3048 /* VLNCTRL: enable vlan filtering and allow all vlan tags through */
3049 vlanctrl = IXGBE_READ_REG(hw, IXGBE_VLNCTRL);
3050 vlanctrl |= IXGBE_VLNCTRL_VFE ; /* enable vlan filters */
3051 IXGBE_WRITE_REG(hw, IXGBE_VLNCTRL, vlanctrl);
3053 /* VFTA - enable all vlan filters */
3054 for (i = 0; i < NUM_VFTA_REGISTERS; i++) {
3055 IXGBE_WRITE_REG(hw, IXGBE_VFTA(i), 0xFFFFFFFF);
3059 * Configure Rx packet plane (recycle mode; WSP) and
3062 reg = IXGBE_RTRPCS_RRM | IXGBE_RTRPCS_RAC;
3063 IXGBE_WRITE_REG(hw, IXGBE_RTRPCS, reg);
3069 ixgbe_dcb_hw_arbite_rx_config(struct ixgbe_hw *hw, uint16_t *refill,
3070 uint16_t *max,uint8_t *bwg_id, uint8_t *tsa, uint8_t *map)
3072 switch (hw->mac.type) {
3073 case ixgbe_mac_82598EB:
3074 ixgbe_dcb_config_rx_arbiter_82598(hw, refill, max, tsa);
3076 case ixgbe_mac_82599EB:
3077 case ixgbe_mac_X540:
3078 case ixgbe_mac_X550:
3079 case ixgbe_mac_X550EM_x:
3080 ixgbe_dcb_config_rx_arbiter_82599(hw, refill, max, bwg_id,
3089 ixgbe_dcb_hw_arbite_tx_config(struct ixgbe_hw *hw, uint16_t *refill, uint16_t *max,
3090 uint8_t *bwg_id, uint8_t *tsa, uint8_t *map)
3092 switch (hw->mac.type) {
3093 case ixgbe_mac_82598EB:
3094 ixgbe_dcb_config_tx_desc_arbiter_82598(hw, refill, max, bwg_id,tsa);
3095 ixgbe_dcb_config_tx_data_arbiter_82598(hw, refill, max, bwg_id,tsa);
3097 case ixgbe_mac_82599EB:
3098 case ixgbe_mac_X540:
3099 case ixgbe_mac_X550:
3100 case ixgbe_mac_X550EM_x:
3101 ixgbe_dcb_config_tx_desc_arbiter_82599(hw, refill, max, bwg_id,tsa);
3102 ixgbe_dcb_config_tx_data_arbiter_82599(hw, refill, max, bwg_id,tsa, map);
3109 #define DCB_RX_CONFIG 1
3110 #define DCB_TX_CONFIG 1
3111 #define DCB_TX_PB 1024
3113 * ixgbe_dcb_hw_configure - Enable DCB and configure
3114 * general DCB in VT mode and non-VT mode parameters
3115 * @dev: pointer to rte_eth_dev structure
3116 * @dcb_config: pointer to ixgbe_dcb_config structure
3119 ixgbe_dcb_hw_configure(struct rte_eth_dev *dev,
3120 struct ixgbe_dcb_config *dcb_config)
3123 uint8_t i,pfc_en,nb_tcs;
3125 uint8_t config_dcb_rx = 0;
3126 uint8_t config_dcb_tx = 0;
3127 uint8_t tsa[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3128 uint8_t bwgid[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3129 uint16_t refill[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3130 uint16_t max[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3131 uint8_t map[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3132 struct ixgbe_dcb_tc_config *tc;
3133 uint32_t max_frame = dev->data->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN;
3134 struct ixgbe_hw *hw =
3135 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3137 switch(dev->data->dev_conf.rxmode.mq_mode){
3138 case ETH_MQ_RX_VMDQ_DCB:
3139 dcb_config->vt_mode = true;
3140 if (hw->mac.type != ixgbe_mac_82598EB) {
3141 config_dcb_rx = DCB_RX_CONFIG;
3143 *get dcb and VT rx configuration parameters
3146 ixgbe_vmdq_dcb_rx_config(dev,dcb_config);
3147 /*Configure general VMDQ and DCB RX parameters*/
3148 ixgbe_vmdq_dcb_configure(dev);
3152 dcb_config->vt_mode = false;
3153 config_dcb_rx = DCB_RX_CONFIG;
3154 /* Get dcb TX configuration parameters from rte_eth_conf */
3155 ixgbe_dcb_rx_config(dev,dcb_config);
3156 /*Configure general DCB RX parameters*/
3157 ixgbe_dcb_rx_hw_config(hw, dcb_config);
3160 PMD_INIT_LOG(ERR, "Incorrect DCB RX mode configuration");
3163 switch (dev->data->dev_conf.txmode.mq_mode) {
3164 case ETH_MQ_TX_VMDQ_DCB:
3165 dcb_config->vt_mode = true;
3166 config_dcb_tx = DCB_TX_CONFIG;
3167 /* get DCB and VT TX configuration parameters from rte_eth_conf */
3168 ixgbe_dcb_vt_tx_config(dev,dcb_config);
3169 /*Configure general VMDQ and DCB TX parameters*/
3170 ixgbe_vmdq_dcb_hw_tx_config(dev,dcb_config);
3174 dcb_config->vt_mode = false;
3175 config_dcb_tx = DCB_TX_CONFIG;
3176 /*get DCB TX configuration parameters from rte_eth_conf*/
3177 ixgbe_dcb_tx_config(dev,dcb_config);
3178 /*Configure general DCB TX parameters*/
3179 ixgbe_dcb_tx_hw_config(hw, dcb_config);
3182 PMD_INIT_LOG(ERR, "Incorrect DCB TX mode configuration");
3186 nb_tcs = dcb_config->num_tcs.pfc_tcs;
3188 ixgbe_dcb_unpack_map_cee(dcb_config, IXGBE_DCB_RX_CONFIG, map);
3189 if(nb_tcs == ETH_4_TCS) {
3190 /* Avoid un-configured priority mapping to TC0 */
3192 uint8_t mask = 0xFF;
3193 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES - 4; i++)
3194 mask = (uint8_t)(mask & (~ (1 << map[i])));
3195 for (i = 0; mask && (i < IXGBE_DCB_MAX_TRAFFIC_CLASS); i++) {
3196 if ((mask & 0x1) && (j < ETH_DCB_NUM_USER_PRIORITIES))
3200 /* Re-configure 4 TCs BW */
3201 for (i = 0; i < nb_tcs; i++) {
3202 tc = &dcb_config->tc_config[i];
3203 tc->path[IXGBE_DCB_TX_CONFIG].bwg_percent =
3204 (uint8_t)(100 / nb_tcs);
3205 tc->path[IXGBE_DCB_RX_CONFIG].bwg_percent =
3206 (uint8_t)(100 / nb_tcs);
3208 for (; i < IXGBE_DCB_MAX_TRAFFIC_CLASS; i++) {
3209 tc = &dcb_config->tc_config[i];
3210 tc->path[IXGBE_DCB_TX_CONFIG].bwg_percent = 0;
3211 tc->path[IXGBE_DCB_RX_CONFIG].bwg_percent = 0;
3216 /* Set RX buffer size */
3217 pbsize = (uint16_t)(NIC_RX_BUFFER_SIZE / nb_tcs);
3218 uint32_t rxpbsize = pbsize << IXGBE_RXPBSIZE_SHIFT;
3219 for (i = 0 ; i < nb_tcs; i++) {
3220 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), rxpbsize);
3222 /* zero alloc all unused TCs */
3223 for (; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
3224 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), 0);
3228 /* Only support an equally distributed Tx packet buffer strategy. */
3229 uint32_t txpktsize = IXGBE_TXPBSIZE_MAX / nb_tcs;
3230 uint32_t txpbthresh = (txpktsize / DCB_TX_PB) - IXGBE_TXPKT_SIZE_MAX;
3231 for (i = 0; i < nb_tcs; i++) {
3232 IXGBE_WRITE_REG(hw, IXGBE_TXPBSIZE(i), txpktsize);
3233 IXGBE_WRITE_REG(hw, IXGBE_TXPBTHRESH(i), txpbthresh);
3235 /* Clear unused TCs, if any, to zero buffer size*/
3236 for (; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
3237 IXGBE_WRITE_REG(hw, IXGBE_TXPBSIZE(i), 0);
3238 IXGBE_WRITE_REG(hw, IXGBE_TXPBTHRESH(i), 0);
3242 /*Calculates traffic class credits*/
3243 ixgbe_dcb_calculate_tc_credits_cee(hw, dcb_config,max_frame,
3244 IXGBE_DCB_TX_CONFIG);
3245 ixgbe_dcb_calculate_tc_credits_cee(hw, dcb_config,max_frame,
3246 IXGBE_DCB_RX_CONFIG);
3249 /* Unpack CEE standard containers */
3250 ixgbe_dcb_unpack_refill_cee(dcb_config, IXGBE_DCB_RX_CONFIG, refill);
3251 ixgbe_dcb_unpack_max_cee(dcb_config, max);
3252 ixgbe_dcb_unpack_bwgid_cee(dcb_config, IXGBE_DCB_RX_CONFIG, bwgid);
3253 ixgbe_dcb_unpack_tsa_cee(dcb_config, IXGBE_DCB_RX_CONFIG, tsa);
3254 /* Configure PG(ETS) RX */
3255 ixgbe_dcb_hw_arbite_rx_config(hw,refill,max,bwgid,tsa,map);
3259 /* Unpack CEE standard containers */
3260 ixgbe_dcb_unpack_refill_cee(dcb_config, IXGBE_DCB_TX_CONFIG, refill);
3261 ixgbe_dcb_unpack_max_cee(dcb_config, max);
3262 ixgbe_dcb_unpack_bwgid_cee(dcb_config, IXGBE_DCB_TX_CONFIG, bwgid);
3263 ixgbe_dcb_unpack_tsa_cee(dcb_config, IXGBE_DCB_TX_CONFIG, tsa);
3264 /* Configure PG(ETS) TX */
3265 ixgbe_dcb_hw_arbite_tx_config(hw,refill,max,bwgid,tsa,map);
3268 /*Configure queue statistics registers*/
3269 ixgbe_dcb_config_tc_stats_82599(hw, dcb_config);
3271 /* Check if the PFC is supported */
3272 if(dev->data->dev_conf.dcb_capability_en & ETH_DCB_PFC_SUPPORT) {
3273 pbsize = (uint16_t) (NIC_RX_BUFFER_SIZE / nb_tcs);
3274 for (i = 0; i < nb_tcs; i++) {
3276 * If the TC count is 8,and the default high_water is 48,
3277 * the low_water is 16 as default.
3279 hw->fc.high_water[i] = (pbsize * 3 ) / 4;
3280 hw->fc.low_water[i] = pbsize / 4;
3281 /* Enable pfc for this TC */
3282 tc = &dcb_config->tc_config[i];
3283 tc->pfc = ixgbe_dcb_pfc_enabled;
3285 ixgbe_dcb_unpack_pfc_cee(dcb_config, map, &pfc_en);
3286 if(dcb_config->num_tcs.pfc_tcs == ETH_4_TCS)
3288 ret = ixgbe_dcb_config_pfc(hw, pfc_en, map);
3295 * ixgbe_configure_dcb - Configure DCB Hardware
3296 * @dev: pointer to rte_eth_dev
3298 void ixgbe_configure_dcb(struct rte_eth_dev *dev)
3300 struct ixgbe_dcb_config *dcb_cfg =
3301 IXGBE_DEV_PRIVATE_TO_DCB_CFG(dev->data->dev_private);
3302 struct rte_eth_conf *dev_conf = &(dev->data->dev_conf);
3304 PMD_INIT_FUNC_TRACE();
3306 /* check support mq_mode for DCB */
3307 if ((dev_conf->rxmode.mq_mode != ETH_MQ_RX_VMDQ_DCB) &&
3308 (dev_conf->rxmode.mq_mode != ETH_MQ_RX_DCB))
3311 if (dev->data->nb_rx_queues != ETH_DCB_NUM_QUEUES)
3314 /** Configure DCB hardware **/
3315 ixgbe_dcb_hw_configure(dev,dcb_cfg);
3321 * VMDq only support for 10 GbE NIC.
3324 ixgbe_vmdq_rx_hw_configure(struct rte_eth_dev *dev)
3326 struct rte_eth_vmdq_rx_conf *cfg;
3327 struct ixgbe_hw *hw;
3328 enum rte_eth_nb_pools num_pools;
3329 uint32_t mrqc, vt_ctl, vlanctrl;
3333 PMD_INIT_FUNC_TRACE();
3334 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3335 cfg = &dev->data->dev_conf.rx_adv_conf.vmdq_rx_conf;
3336 num_pools = cfg->nb_queue_pools;
3338 ixgbe_rss_disable(dev);
3340 /* MRQC: enable vmdq */
3341 mrqc = IXGBE_MRQC_VMDQEN;
3342 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
3344 /* PFVTCTL: turn on virtualisation and set the default pool */
3345 vt_ctl = IXGBE_VT_CTL_VT_ENABLE | IXGBE_VT_CTL_REPLEN;
3346 if (cfg->enable_default_pool)
3347 vt_ctl |= (cfg->default_pool << IXGBE_VT_CTL_POOL_SHIFT);
3349 vt_ctl |= IXGBE_VT_CTL_DIS_DEFPL;
3351 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, vt_ctl);
3353 for (i = 0; i < (int)num_pools; i++) {
3354 vmolr = ixgbe_convert_vm_rx_mask_to_val(cfg->rx_mode, vmolr);
3355 IXGBE_WRITE_REG(hw, IXGBE_VMOLR(i), vmolr);
3358 /* VLNCTRL: enable vlan filtering and allow all vlan tags through */
3359 vlanctrl = IXGBE_READ_REG(hw, IXGBE_VLNCTRL);
3360 vlanctrl |= IXGBE_VLNCTRL_VFE ; /* enable vlan filters */
3361 IXGBE_WRITE_REG(hw, IXGBE_VLNCTRL, vlanctrl);
3363 /* VFTA - enable all vlan filters */
3364 for (i = 0; i < NUM_VFTA_REGISTERS; i++)
3365 IXGBE_WRITE_REG(hw, IXGBE_VFTA(i), UINT32_MAX);
3367 /* VFRE: pool enabling for receive - 64 */
3368 IXGBE_WRITE_REG(hw, IXGBE_VFRE(0), UINT32_MAX);
3369 if (num_pools == ETH_64_POOLS)
3370 IXGBE_WRITE_REG(hw, IXGBE_VFRE(1), UINT32_MAX);
3373 * MPSAR - allow pools to read specific mac addresses
3374 * In this case, all pools should be able to read from mac addr 0
3376 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_LO(0), UINT32_MAX);
3377 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_HI(0), UINT32_MAX);
3379 /* PFVLVF, PFVLVFB: set up filters for vlan tags as configured */
3380 for (i = 0; i < cfg->nb_pool_maps; i++) {
3381 /* set vlan id in VF register and set the valid bit */
3382 IXGBE_WRITE_REG(hw, IXGBE_VLVF(i), (IXGBE_VLVF_VIEN | \
3383 (cfg->pool_map[i].vlan_id & IXGBE_RXD_VLAN_ID_MASK)));
3385 * Put the allowed pools in VFB reg. As we only have 16 or 64
3386 * pools, we only need to use the first half of the register
3389 if (((cfg->pool_map[i].pools >> 32) & UINT32_MAX) == 0)
3390 IXGBE_WRITE_REG(hw, IXGBE_VLVFB(i*2), \
3391 (cfg->pool_map[i].pools & UINT32_MAX));
3393 IXGBE_WRITE_REG(hw, IXGBE_VLVFB((i*2+1)), \
3394 ((cfg->pool_map[i].pools >> 32) \
3399 /* PFDMA Tx General Switch Control Enables VMDQ loopback */
3400 if (cfg->enable_loop_back) {
3401 IXGBE_WRITE_REG(hw, IXGBE_PFDTXGSWC, IXGBE_PFDTXGSWC_VT_LBEN);
3402 for (i = 0; i < RTE_IXGBE_VMTXSW_REGISTER_COUNT; i++)
3403 IXGBE_WRITE_REG(hw, IXGBE_VMTXSW(i), UINT32_MAX);
3406 IXGBE_WRITE_FLUSH(hw);
3410 * ixgbe_dcb_config_tx_hw_config - Configure general VMDq TX parameters
3411 * @hw: pointer to hardware structure
3414 ixgbe_vmdq_tx_hw_configure(struct ixgbe_hw *hw)
3419 PMD_INIT_FUNC_TRACE();
3420 /*PF VF Transmit Enable*/
3421 IXGBE_WRITE_REG(hw, IXGBE_VFTE(0), UINT32_MAX);
3422 IXGBE_WRITE_REG(hw, IXGBE_VFTE(1), UINT32_MAX);
3424 /* Disable the Tx desc arbiter so that MTQC can be changed */
3425 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
3426 reg |= IXGBE_RTTDCS_ARBDIS;
3427 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
3429 reg = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_64VF;
3430 IXGBE_WRITE_REG(hw, IXGBE_MTQC, reg);
3432 /* Disable drop for all queues */
3433 for (q = 0; q < IXGBE_MAX_RX_QUEUE_NUM; q++)
3434 IXGBE_WRITE_REG(hw, IXGBE_QDE,
3435 (IXGBE_QDE_WRITE | (q << IXGBE_QDE_IDX_SHIFT)));
3437 /* Enable the Tx desc arbiter */
3438 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
3439 reg &= ~IXGBE_RTTDCS_ARBDIS;
3440 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
3442 IXGBE_WRITE_FLUSH(hw);
3447 static int __attribute__((cold))
3448 ixgbe_alloc_rx_queue_mbufs(struct ixgbe_rx_queue *rxq)
3450 struct ixgbe_rx_entry *rxe = rxq->sw_ring;
3454 /* Initialize software ring entries */
3455 for (i = 0; i < rxq->nb_rx_desc; i++) {
3456 volatile union ixgbe_adv_rx_desc *rxd;
3457 struct rte_mbuf *mbuf = rte_rxmbuf_alloc(rxq->mb_pool);
3459 PMD_INIT_LOG(ERR, "RX mbuf alloc failed queue_id=%u",
3460 (unsigned) rxq->queue_id);
3464 rte_mbuf_refcnt_set(mbuf, 1);
3466 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
3468 mbuf->port = rxq->port_id;
3471 rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mbuf));
3472 rxd = &rxq->rx_ring[i];
3473 rxd->read.hdr_addr = dma_addr;
3474 rxd->read.pkt_addr = dma_addr;
3482 ixgbe_config_vf_rss(struct rte_eth_dev *dev)
3484 struct ixgbe_hw *hw;
3487 ixgbe_rss_configure(dev);
3489 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3491 /* MRQC: enable VF RSS */
3492 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
3493 mrqc &= ~IXGBE_MRQC_MRQE_MASK;
3494 switch (RTE_ETH_DEV_SRIOV(dev).active) {
3496 mrqc |= IXGBE_MRQC_VMDQRSS64EN;
3500 mrqc |= IXGBE_MRQC_VMDQRSS32EN;
3504 PMD_INIT_LOG(ERR, "Invalid pool number in IOV mode with VMDQ RSS");
3508 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
3514 ixgbe_config_vf_default(struct rte_eth_dev *dev)
3516 struct ixgbe_hw *hw =
3517 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3519 switch (RTE_ETH_DEV_SRIOV(dev).active) {
3521 IXGBE_WRITE_REG(hw, IXGBE_MRQC,
3526 IXGBE_WRITE_REG(hw, IXGBE_MRQC,
3527 IXGBE_MRQC_VMDQRT4TCEN);
3531 IXGBE_WRITE_REG(hw, IXGBE_MRQC,
3532 IXGBE_MRQC_VMDQRT8TCEN);
3536 "invalid pool number in IOV mode");
3543 ixgbe_dev_mq_rx_configure(struct rte_eth_dev *dev)
3545 struct ixgbe_hw *hw =
3546 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3548 if (hw->mac.type == ixgbe_mac_82598EB)
3551 if (RTE_ETH_DEV_SRIOV(dev).active == 0) {
3553 * SRIOV inactive scheme
3554 * any DCB/RSS w/o VMDq multi-queue setting
3556 switch (dev->data->dev_conf.rxmode.mq_mode) {
3558 ixgbe_rss_configure(dev);
3561 case ETH_MQ_RX_VMDQ_DCB:
3562 ixgbe_vmdq_dcb_configure(dev);
3565 case ETH_MQ_RX_VMDQ_ONLY:
3566 ixgbe_vmdq_rx_hw_configure(dev);
3569 case ETH_MQ_RX_NONE:
3570 /* if mq_mode is none, disable rss mode.*/
3571 default: ixgbe_rss_disable(dev);
3575 * SRIOV active scheme
3576 * Support RSS together with VMDq & SRIOV
3578 switch (dev->data->dev_conf.rxmode.mq_mode) {
3580 case ETH_MQ_RX_VMDQ_RSS:
3581 ixgbe_config_vf_rss(dev);
3584 /* FIXME if support DCB/RSS together with VMDq & SRIOV */
3585 case ETH_MQ_RX_VMDQ_DCB:
3586 case ETH_MQ_RX_VMDQ_DCB_RSS:
3588 "Could not support DCB with VMDq & SRIOV");
3591 ixgbe_config_vf_default(dev);
3600 ixgbe_dev_mq_tx_configure(struct rte_eth_dev *dev)
3602 struct ixgbe_hw *hw =
3603 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3607 if (hw->mac.type == ixgbe_mac_82598EB)
3610 /* disable arbiter before setting MTQC */
3611 rttdcs = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
3612 rttdcs |= IXGBE_RTTDCS_ARBDIS;
3613 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, rttdcs);
3615 if (RTE_ETH_DEV_SRIOV(dev).active == 0) {
3617 * SRIOV inactive scheme
3618 * any DCB w/o VMDq multi-queue setting
3620 if (dev->data->dev_conf.txmode.mq_mode == ETH_MQ_TX_VMDQ_ONLY)
3621 ixgbe_vmdq_tx_hw_configure(hw);
3623 mtqc = IXGBE_MTQC_64Q_1PB;
3624 IXGBE_WRITE_REG(hw, IXGBE_MTQC, mtqc);
3627 switch (RTE_ETH_DEV_SRIOV(dev).active) {
3630 * SRIOV active scheme
3631 * FIXME if support DCB together with VMDq & SRIOV
3634 mtqc = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_64VF;
3637 mtqc = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_32VF;
3640 mtqc = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_RT_ENA |
3644 mtqc = IXGBE_MTQC_64Q_1PB;
3645 PMD_INIT_LOG(ERR, "invalid pool number in IOV mode");
3647 IXGBE_WRITE_REG(hw, IXGBE_MTQC, mtqc);
3650 /* re-enable arbiter */
3651 rttdcs &= ~IXGBE_RTTDCS_ARBDIS;
3652 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, rttdcs);
3658 * ixgbe_get_rscctl_maxdesc - Calculate the RSCCTL[n].MAXDESC for PF
3660 * Return the RSCCTL[n].MAXDESC for 82599 and x540 PF devices according to the
3661 * spec rev. 3.0 chapter 8.2.3.8.13.
3663 * @pool Memory pool of the Rx queue
3665 static inline uint32_t
3666 ixgbe_get_rscctl_maxdesc(struct rte_mempool *pool)
3668 struct rte_pktmbuf_pool_private *mp_priv = rte_mempool_get_priv(pool);
3670 /* MAXDESC * SRRCTL.BSIZEPKT must not exceed 64 KB minus one */
3673 (mp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM);
3676 return IXGBE_RSCCTL_MAXDESC_16;
3677 else if (maxdesc >= 8)
3678 return IXGBE_RSCCTL_MAXDESC_8;
3679 else if (maxdesc >= 4)
3680 return IXGBE_RSCCTL_MAXDESC_4;
3682 return IXGBE_RSCCTL_MAXDESC_1;
3686 * ixgbe_set_ivar - Setup the correct IVAR register for a particular MSIX
3689 * (Taken from FreeBSD tree)
3690 * (yes this is all very magic and confusing :)
3693 * @entry the register array entry
3694 * @vector the MSIX vector for this queue
3698 ixgbe_set_ivar(struct rte_eth_dev *dev, u8 entry, u8 vector, s8 type)
3700 struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3703 vector |= IXGBE_IVAR_ALLOC_VAL;
3705 switch (hw->mac.type) {
3707 case ixgbe_mac_82598EB:
3709 entry = IXGBE_IVAR_OTHER_CAUSES_INDEX;
3711 entry += (type * 64);
3712 index = (entry >> 2) & 0x1F;
3713 ivar = IXGBE_READ_REG(hw, IXGBE_IVAR(index));
3714 ivar &= ~(0xFF << (8 * (entry & 0x3)));
3715 ivar |= (vector << (8 * (entry & 0x3)));
3716 IXGBE_WRITE_REG(hw, IXGBE_IVAR(index), ivar);
3719 case ixgbe_mac_82599EB:
3720 case ixgbe_mac_X540:
3721 if (type == -1) { /* MISC IVAR */
3722 index = (entry & 1) * 8;
3723 ivar = IXGBE_READ_REG(hw, IXGBE_IVAR_MISC);
3724 ivar &= ~(0xFF << index);
3725 ivar |= (vector << index);
3726 IXGBE_WRITE_REG(hw, IXGBE_IVAR_MISC, ivar);
3727 } else { /* RX/TX IVARS */
3728 index = (16 * (entry & 1)) + (8 * type);
3729 ivar = IXGBE_READ_REG(hw, IXGBE_IVAR(entry >> 1));
3730 ivar &= ~(0xFF << index);
3731 ivar |= (vector << index);
3732 IXGBE_WRITE_REG(hw, IXGBE_IVAR(entry >> 1), ivar);
3742 void __attribute__((cold))
3743 ixgbe_set_rx_function(struct rte_eth_dev *dev)
3745 struct ixgbe_adapter *adapter =
3746 (struct ixgbe_adapter *)dev->data->dev_private;
3749 * In order to allow Vector Rx there are a few configuration
3750 * conditions to be met and Rx Bulk Allocation should be allowed.
3752 if (ixgbe_rx_vec_dev_conf_condition_check(dev) ||
3753 !adapter->rx_bulk_alloc_allowed) {
3754 PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet Vector Rx "
3755 "preconditions or RTE_IXGBE_INC_VECTOR is "
3757 dev->data->port_id);
3759 adapter->rx_vec_allowed = false;
3763 * Initialize the appropriate LRO callback.
3765 * If all queues satisfy the bulk allocation preconditions
3766 * (hw->rx_bulk_alloc_allowed is TRUE) then we may use bulk allocation.
3767 * Otherwise use a single allocation version.
3769 if (dev->data->lro) {
3770 if (adapter->rx_bulk_alloc_allowed) {
3771 PMD_INIT_LOG(INFO, "LRO is requested. Using a bulk "
3772 "allocation version");
3773 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_bulk_alloc;
3775 PMD_INIT_LOG(INFO, "LRO is requested. Using a single "
3776 "allocation version");
3777 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_single_alloc;
3779 } else if (dev->data->scattered_rx) {
3781 * Set the non-LRO scattered callback: there are Vector and
3782 * single allocation versions.
3784 if (adapter->rx_vec_allowed) {
3785 PMD_INIT_LOG(DEBUG, "Using Vector Scattered Rx "
3786 "callback (port=%d).",
3787 dev->data->port_id);
3789 dev->rx_pkt_burst = ixgbe_recv_scattered_pkts_vec;
3790 } else if (adapter->rx_bulk_alloc_allowed) {
3791 PMD_INIT_LOG(INFO, "Using a Scattered with bulk "
3792 "allocation callback (port=%d).",
3793 dev->data->port_id);
3794 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_bulk_alloc;
3796 PMD_INIT_LOG(DEBUG, "Using Regualr (non-vector, "
3797 "single allocation) "
3798 "Scattered Rx callback "
3800 dev->data->port_id);
3802 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_single_alloc;
3805 * Below we set "simple" callbacks according to port/queues parameters.
3806 * If parameters allow we are going to choose between the following
3810 * - Single buffer allocation (the simplest one)
3812 } else if (adapter->rx_vec_allowed) {
3813 PMD_INIT_LOG(INFO, "Vector rx enabled, please make sure RX "
3814 "burst size no less than 32.");
3816 dev->rx_pkt_burst = ixgbe_recv_pkts_vec;
3817 } else if (adapter->rx_bulk_alloc_allowed) {
3818 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
3819 "satisfied. Rx Burst Bulk Alloc function "
3820 "will be used on port=%d.",
3821 dev->data->port_id);
3823 dev->rx_pkt_burst = ixgbe_recv_pkts_bulk_alloc;
3825 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are not "
3826 "satisfied, or Scattered Rx is requested, "
3827 "or RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC "
3828 "is not enabled (port=%d).",
3829 dev->data->port_id);
3831 dev->rx_pkt_burst = ixgbe_recv_pkts;
3836 * ixgbe_set_rsc - configure RSC related port HW registers
3838 * Configures the port's RSC related registers according to the 4.6.7.2 chapter
3839 * of 82599 Spec (x540 configuration is virtually the same).
3843 * Returns 0 in case of success or a non-zero error code
3846 ixgbe_set_rsc(struct rte_eth_dev *dev)
3848 struct rte_eth_rxmode *rx_conf = &dev->data->dev_conf.rxmode;
3849 struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3850 struct rte_eth_dev_info dev_info = { 0 };
3851 bool rsc_capable = false;
3856 dev->dev_ops->dev_infos_get(dev, &dev_info);
3857 if (dev_info.rx_offload_capa & DEV_RX_OFFLOAD_TCP_LRO)
3860 if (!rsc_capable && rx_conf->enable_lro) {
3861 PMD_INIT_LOG(CRIT, "LRO is requested on HW that doesn't "
3866 /* RSC global configuration (chapter 4.6.7.2.1 of 82599 Spec) */
3868 if (!rx_conf->hw_strip_crc && rx_conf->enable_lro) {
3870 * According to chapter of 4.6.7.2.1 of the Spec Rev.
3871 * 3.0 RSC configuration requires HW CRC stripping being
3872 * enabled. If user requested both HW CRC stripping off
3873 * and RSC on - return an error.
3875 PMD_INIT_LOG(CRIT, "LRO can't be enabled when HW CRC "
3880 /* RFCTL configuration */
3882 uint32_t rfctl = IXGBE_READ_REG(hw, IXGBE_RFCTL);
3883 if (rx_conf->enable_lro)
3885 * Since NFS packets coalescing is not supported - clear
3886 * RFCTL.NFSW_DIS and RFCTL.NFSR_DIS when RSC is
3889 rfctl &= ~(IXGBE_RFCTL_RSC_DIS | IXGBE_RFCTL_NFSW_DIS |
3890 IXGBE_RFCTL_NFSR_DIS);
3892 rfctl |= IXGBE_RFCTL_RSC_DIS;
3894 IXGBE_WRITE_REG(hw, IXGBE_RFCTL, rfctl);
3897 /* If LRO hasn't been requested - we are done here. */
3898 if (!rx_conf->enable_lro)
3901 /* Set RDRXCTL.RSCACKC bit */
3902 rdrxctl = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
3903 rdrxctl |= IXGBE_RDRXCTL_RSCACKC;
3904 IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rdrxctl);
3906 /* Per-queue RSC configuration (chapter 4.6.7.2.2 of 82599 Spec) */
3907 for (i = 0; i < dev->data->nb_rx_queues; i++) {
3908 struct ixgbe_rx_queue *rxq = dev->data->rx_queues[i];
3910 IXGBE_READ_REG(hw, IXGBE_SRRCTL(rxq->reg_idx));
3912 IXGBE_READ_REG(hw, IXGBE_RSCCTL(rxq->reg_idx));
3914 IXGBE_READ_REG(hw, IXGBE_PSRTYPE(rxq->reg_idx));
3916 IXGBE_READ_REG(hw, IXGBE_EITR(rxq->reg_idx));
3919 * ixgbe PMD doesn't support header-split at the moment.
3921 * Following the 4.6.7.2.1 chapter of the 82599/x540
3922 * Spec if RSC is enabled the SRRCTL[n].BSIZEHEADER
3923 * should be configured even if header split is not
3924 * enabled. We will configure it 128 bytes following the
3925 * recommendation in the spec.
3927 srrctl &= ~IXGBE_SRRCTL_BSIZEHDR_MASK;
3928 srrctl |= (128 << IXGBE_SRRCTL_BSIZEHDRSIZE_SHIFT) &
3929 IXGBE_SRRCTL_BSIZEHDR_MASK;
3932 * TODO: Consider setting the Receive Descriptor Minimum
3933 * Threshold Size for an RSC case. This is not an obviously
3934 * beneficiary option but the one worth considering...
3937 rscctl |= IXGBE_RSCCTL_RSCEN;
3938 rscctl |= ixgbe_get_rscctl_maxdesc(rxq->mb_pool);
3939 psrtype |= IXGBE_PSRTYPE_TCPHDR;
3942 * RSC: Set ITR interval corresponding to 2K ints/s.
3944 * Full-sized RSC aggregations for a 10Gb/s link will
3945 * arrive at about 20K aggregation/s rate.
3947 * 2K inst/s rate will make only 10% of the
3948 * aggregations to be closed due to the interrupt timer
3949 * expiration for a streaming at wire-speed case.
3951 * For a sparse streaming case this setting will yield
3952 * at most 500us latency for a single RSC aggregation.
3954 eitr &= ~IXGBE_EITR_ITR_INT_MASK;
3955 eitr |= IXGBE_EITR_INTERVAL_US(500) | IXGBE_EITR_CNT_WDIS;
3957 IXGBE_WRITE_REG(hw, IXGBE_SRRCTL(rxq->reg_idx), srrctl);
3958 IXGBE_WRITE_REG(hw, IXGBE_RSCCTL(rxq->reg_idx), rscctl);
3959 IXGBE_WRITE_REG(hw, IXGBE_PSRTYPE(rxq->reg_idx), psrtype);
3960 IXGBE_WRITE_REG(hw, IXGBE_EITR(rxq->reg_idx), eitr);
3963 * RSC requires the mapping of the queue to the
3966 ixgbe_set_ivar(dev, rxq->reg_idx, i, 0);
3971 PMD_INIT_LOG(INFO, "enabling LRO mode");
3977 * Initializes Receive Unit.
3979 int __attribute__((cold))
3980 ixgbe_dev_rx_init(struct rte_eth_dev *dev)
3982 struct ixgbe_hw *hw;
3983 struct ixgbe_rx_queue *rxq;
3994 struct rte_eth_rxmode *rx_conf = &dev->data->dev_conf.rxmode;
3997 PMD_INIT_FUNC_TRACE();
3998 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4001 * Make sure receives are disabled while setting
4002 * up the RX context (registers, descriptor rings, etc.).
4004 rxctrl = IXGBE_READ_REG(hw, IXGBE_RXCTRL);
4005 IXGBE_WRITE_REG(hw, IXGBE_RXCTRL, rxctrl & ~IXGBE_RXCTRL_RXEN);
4007 /* Enable receipt of broadcasted frames */
4008 fctrl = IXGBE_READ_REG(hw, IXGBE_FCTRL);
4009 fctrl |= IXGBE_FCTRL_BAM;
4010 fctrl |= IXGBE_FCTRL_DPF;
4011 fctrl |= IXGBE_FCTRL_PMCF;
4012 IXGBE_WRITE_REG(hw, IXGBE_FCTRL, fctrl);
4015 * Configure CRC stripping, if any.
4017 hlreg0 = IXGBE_READ_REG(hw, IXGBE_HLREG0);
4018 if (rx_conf->hw_strip_crc)
4019 hlreg0 |= IXGBE_HLREG0_RXCRCSTRP;
4021 hlreg0 &= ~IXGBE_HLREG0_RXCRCSTRP;
4024 * Configure jumbo frame support, if any.
4026 if (rx_conf->jumbo_frame == 1) {
4027 hlreg0 |= IXGBE_HLREG0_JUMBOEN;
4028 maxfrs = IXGBE_READ_REG(hw, IXGBE_MAXFRS);
4029 maxfrs &= 0x0000FFFF;
4030 maxfrs |= (rx_conf->max_rx_pkt_len << 16);
4031 IXGBE_WRITE_REG(hw, IXGBE_MAXFRS, maxfrs);
4033 hlreg0 &= ~IXGBE_HLREG0_JUMBOEN;
4036 * If loopback mode is configured for 82599, set LPBK bit.
4038 if (hw->mac.type == ixgbe_mac_82599EB &&
4039 dev->data->dev_conf.lpbk_mode == IXGBE_LPBK_82599_TX_RX)
4040 hlreg0 |= IXGBE_HLREG0_LPBK;
4042 hlreg0 &= ~IXGBE_HLREG0_LPBK;
4044 IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hlreg0);
4046 /* Setup RX queues */
4047 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4048 rxq = dev->data->rx_queues[i];
4051 * Reset crc_len in case it was changed after queue setup by a
4052 * call to configure.
4054 rxq->crc_len = rx_conf->hw_strip_crc ? 0 : ETHER_CRC_LEN;
4056 /* Setup the Base and Length of the Rx Descriptor Rings */
4057 bus_addr = rxq->rx_ring_phys_addr;
4058 IXGBE_WRITE_REG(hw, IXGBE_RDBAL(rxq->reg_idx),
4059 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4060 IXGBE_WRITE_REG(hw, IXGBE_RDBAH(rxq->reg_idx),
4061 (uint32_t)(bus_addr >> 32));
4062 IXGBE_WRITE_REG(hw, IXGBE_RDLEN(rxq->reg_idx),
4063 rxq->nb_rx_desc * sizeof(union ixgbe_adv_rx_desc));
4064 IXGBE_WRITE_REG(hw, IXGBE_RDH(rxq->reg_idx), 0);
4065 IXGBE_WRITE_REG(hw, IXGBE_RDT(rxq->reg_idx), 0);
4067 /* Configure the SRRCTL register */
4068 #ifdef RTE_HEADER_SPLIT_ENABLE
4070 * Configure Header Split
4072 if (rx_conf->header_split) {
4073 if (hw->mac.type == ixgbe_mac_82599EB) {
4074 /* Must setup the PSRTYPE register */
4076 psrtype = IXGBE_PSRTYPE_TCPHDR |
4077 IXGBE_PSRTYPE_UDPHDR |
4078 IXGBE_PSRTYPE_IPV4HDR |
4079 IXGBE_PSRTYPE_IPV6HDR;
4080 IXGBE_WRITE_REG(hw, IXGBE_PSRTYPE(rxq->reg_idx), psrtype);
4082 srrctl = ((rx_conf->split_hdr_size <<
4083 IXGBE_SRRCTL_BSIZEHDRSIZE_SHIFT) &
4084 IXGBE_SRRCTL_BSIZEHDR_MASK);
4085 srrctl |= IXGBE_SRRCTL_DESCTYPE_HDR_SPLIT_ALWAYS;
4088 srrctl = IXGBE_SRRCTL_DESCTYPE_ADV_ONEBUF;
4090 /* Set if packets are dropped when no descriptors available */
4092 srrctl |= IXGBE_SRRCTL_DROP_EN;
4095 * Configure the RX buffer size in the BSIZEPACKET field of
4096 * the SRRCTL register of the queue.
4097 * The value is in 1 KB resolution. Valid values can be from
4100 buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
4101 RTE_PKTMBUF_HEADROOM);
4102 srrctl |= ((buf_size >> IXGBE_SRRCTL_BSIZEPKT_SHIFT) &
4103 IXGBE_SRRCTL_BSIZEPKT_MASK);
4105 IXGBE_WRITE_REG(hw, IXGBE_SRRCTL(rxq->reg_idx), srrctl);
4107 buf_size = (uint16_t) ((srrctl & IXGBE_SRRCTL_BSIZEPKT_MASK) <<
4108 IXGBE_SRRCTL_BSIZEPKT_SHIFT);
4110 /* It adds dual VLAN length for supporting dual VLAN */
4111 if (dev->data->dev_conf.rxmode.max_rx_pkt_len +
4112 2 * IXGBE_VLAN_TAG_SIZE > buf_size)
4113 dev->data->scattered_rx = 1;
4116 if (rx_conf->enable_scatter)
4117 dev->data->scattered_rx = 1;
4120 * Device configured with multiple RX queues.
4122 ixgbe_dev_mq_rx_configure(dev);
4125 * Setup the Checksum Register.
4126 * Disable Full-Packet Checksum which is mutually exclusive with RSS.
4127 * Enable IP/L4 checkum computation by hardware if requested to do so.
4129 rxcsum = IXGBE_READ_REG(hw, IXGBE_RXCSUM);
4130 rxcsum |= IXGBE_RXCSUM_PCSD;
4131 if (rx_conf->hw_ip_checksum)
4132 rxcsum |= IXGBE_RXCSUM_IPPCSE;
4134 rxcsum &= ~IXGBE_RXCSUM_IPPCSE;
4136 IXGBE_WRITE_REG(hw, IXGBE_RXCSUM, rxcsum);
4138 if (hw->mac.type == ixgbe_mac_82599EB ||
4139 hw->mac.type == ixgbe_mac_X540) {
4140 rdrxctl = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
4141 if (rx_conf->hw_strip_crc)
4142 rdrxctl |= IXGBE_RDRXCTL_CRCSTRIP;
4144 rdrxctl &= ~IXGBE_RDRXCTL_CRCSTRIP;
4145 rdrxctl &= ~IXGBE_RDRXCTL_RSCFRSTSIZE;
4146 IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rdrxctl);
4149 rc = ixgbe_set_rsc(dev);
4153 ixgbe_set_rx_function(dev);
4159 * Initializes Transmit Unit.
4161 void __attribute__((cold))
4162 ixgbe_dev_tx_init(struct rte_eth_dev *dev)
4164 struct ixgbe_hw *hw;
4165 struct ixgbe_tx_queue *txq;
4171 PMD_INIT_FUNC_TRACE();
4172 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4174 /* Enable TX CRC (checksum offload requirement) and hw padding
4175 * (TSO requirement) */
4176 hlreg0 = IXGBE_READ_REG(hw, IXGBE_HLREG0);
4177 hlreg0 |= (IXGBE_HLREG0_TXCRCEN | IXGBE_HLREG0_TXPADEN);
4178 IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hlreg0);
4180 /* Setup the Base and Length of the Tx Descriptor Rings */
4181 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4182 txq = dev->data->tx_queues[i];
4184 bus_addr = txq->tx_ring_phys_addr;
4185 IXGBE_WRITE_REG(hw, IXGBE_TDBAL(txq->reg_idx),
4186 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4187 IXGBE_WRITE_REG(hw, IXGBE_TDBAH(txq->reg_idx),
4188 (uint32_t)(bus_addr >> 32));
4189 IXGBE_WRITE_REG(hw, IXGBE_TDLEN(txq->reg_idx),
4190 txq->nb_tx_desc * sizeof(union ixgbe_adv_tx_desc));
4191 /* Setup the HW Tx Head and TX Tail descriptor pointers */
4192 IXGBE_WRITE_REG(hw, IXGBE_TDH(txq->reg_idx), 0);
4193 IXGBE_WRITE_REG(hw, IXGBE_TDT(txq->reg_idx), 0);
4196 * Disable Tx Head Writeback RO bit, since this hoses
4197 * bookkeeping if things aren't delivered in order.
4199 switch (hw->mac.type) {
4200 case ixgbe_mac_82598EB:
4201 txctrl = IXGBE_READ_REG(hw,
4202 IXGBE_DCA_TXCTRL(txq->reg_idx));
4203 txctrl &= ~IXGBE_DCA_TXCTRL_DESC_WRO_EN;
4204 IXGBE_WRITE_REG(hw, IXGBE_DCA_TXCTRL(txq->reg_idx),
4208 case ixgbe_mac_82599EB:
4209 case ixgbe_mac_X540:
4210 case ixgbe_mac_X550:
4211 case ixgbe_mac_X550EM_x:
4213 txctrl = IXGBE_READ_REG(hw,
4214 IXGBE_DCA_TXCTRL_82599(txq->reg_idx));
4215 txctrl &= ~IXGBE_DCA_TXCTRL_DESC_WRO_EN;
4216 IXGBE_WRITE_REG(hw, IXGBE_DCA_TXCTRL_82599(txq->reg_idx),
4222 /* Device configured with multiple TX queues. */
4223 ixgbe_dev_mq_tx_configure(dev);
4227 * Set up link for 82599 loopback mode Tx->Rx.
4229 static inline void __attribute__((cold))
4230 ixgbe_setup_loopback_link_82599(struct ixgbe_hw *hw)
4232 PMD_INIT_FUNC_TRACE();
4234 if (ixgbe_verify_lesm_fw_enabled_82599(hw)) {
4235 if (hw->mac.ops.acquire_swfw_sync(hw, IXGBE_GSSR_MAC_CSR_SM) !=
4237 PMD_INIT_LOG(ERR, "Could not enable loopback mode");
4246 IXGBE_AUTOC_LMS_10G_LINK_NO_AN | IXGBE_AUTOC_FLU);
4247 ixgbe_reset_pipeline_82599(hw);
4249 hw->mac.ops.release_swfw_sync(hw, IXGBE_GSSR_MAC_CSR_SM);
4255 * Start Transmit and Receive Units.
4257 int __attribute__((cold))
4258 ixgbe_dev_rxtx_start(struct rte_eth_dev *dev)
4260 struct ixgbe_hw *hw;
4261 struct ixgbe_tx_queue *txq;
4262 struct ixgbe_rx_queue *rxq;
4269 PMD_INIT_FUNC_TRACE();
4270 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4272 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4273 txq = dev->data->tx_queues[i];
4274 /* Setup Transmit Threshold Registers */
4275 txdctl = IXGBE_READ_REG(hw, IXGBE_TXDCTL(txq->reg_idx));
4276 txdctl |= txq->pthresh & 0x7F;
4277 txdctl |= ((txq->hthresh & 0x7F) << 8);
4278 txdctl |= ((txq->wthresh & 0x7F) << 16);
4279 IXGBE_WRITE_REG(hw, IXGBE_TXDCTL(txq->reg_idx), txdctl);
4282 if (hw->mac.type != ixgbe_mac_82598EB) {
4283 dmatxctl = IXGBE_READ_REG(hw, IXGBE_DMATXCTL);
4284 dmatxctl |= IXGBE_DMATXCTL_TE;
4285 IXGBE_WRITE_REG(hw, IXGBE_DMATXCTL, dmatxctl);
4288 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4289 txq = dev->data->tx_queues[i];
4290 if (!txq->tx_deferred_start) {
4291 ret = ixgbe_dev_tx_queue_start(dev, i);
4297 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4298 rxq = dev->data->rx_queues[i];
4299 if (!rxq->rx_deferred_start) {
4300 ret = ixgbe_dev_rx_queue_start(dev, i);
4306 /* Enable Receive engine */
4307 rxctrl = IXGBE_READ_REG(hw, IXGBE_RXCTRL);
4308 if (hw->mac.type == ixgbe_mac_82598EB)
4309 rxctrl |= IXGBE_RXCTRL_DMBYPS;
4310 rxctrl |= IXGBE_RXCTRL_RXEN;
4311 hw->mac.ops.enable_rx_dma(hw, rxctrl);
4313 /* If loopback mode is enabled for 82599, set up the link accordingly */
4314 if (hw->mac.type == ixgbe_mac_82599EB &&
4315 dev->data->dev_conf.lpbk_mode == IXGBE_LPBK_82599_TX_RX)
4316 ixgbe_setup_loopback_link_82599(hw);
4322 * Start Receive Units for specified queue.
4324 int __attribute__((cold))
4325 ixgbe_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
4327 struct ixgbe_hw *hw;
4328 struct ixgbe_rx_queue *rxq;
4332 PMD_INIT_FUNC_TRACE();
4333 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4335 if (rx_queue_id < dev->data->nb_rx_queues) {
4336 rxq = dev->data->rx_queues[rx_queue_id];
4338 /* Allocate buffers for descriptor rings */
4339 if (ixgbe_alloc_rx_queue_mbufs(rxq) != 0) {
4340 PMD_INIT_LOG(ERR, "Could not alloc mbuf for queue:%d",
4344 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4345 rxdctl |= IXGBE_RXDCTL_ENABLE;
4346 IXGBE_WRITE_REG(hw, IXGBE_RXDCTL(rxq->reg_idx), rxdctl);
4348 /* Wait until RX Enable ready */
4349 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4352 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4353 } while (--poll_ms && !(rxdctl & IXGBE_RXDCTL_ENABLE));
4355 PMD_INIT_LOG(ERR, "Could not enable Rx Queue %d",
4358 IXGBE_WRITE_REG(hw, IXGBE_RDH(rxq->reg_idx), 0);
4359 IXGBE_WRITE_REG(hw, IXGBE_RDT(rxq->reg_idx), rxq->nb_rx_desc - 1);
4367 * Stop Receive Units for specified queue.
4369 int __attribute__((cold))
4370 ixgbe_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
4372 struct ixgbe_hw *hw;
4373 struct ixgbe_adapter *adapter =
4374 (struct ixgbe_adapter *)dev->data->dev_private;
4375 struct ixgbe_rx_queue *rxq;
4379 PMD_INIT_FUNC_TRACE();
4380 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4382 if (rx_queue_id < dev->data->nb_rx_queues) {
4383 rxq = dev->data->rx_queues[rx_queue_id];
4385 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4386 rxdctl &= ~IXGBE_RXDCTL_ENABLE;
4387 IXGBE_WRITE_REG(hw, IXGBE_RXDCTL(rxq->reg_idx), rxdctl);
4389 /* Wait until RX Enable ready */
4390 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4393 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4394 } while (--poll_ms && (rxdctl | IXGBE_RXDCTL_ENABLE));
4396 PMD_INIT_LOG(ERR, "Could not disable Rx Queue %d",
4399 rte_delay_us(RTE_IXGBE_WAIT_100_US);
4401 ixgbe_rx_queue_release_mbufs(rxq);
4402 ixgbe_reset_rx_queue(adapter, rxq);
4411 * Start Transmit Units for specified queue.
4413 int __attribute__((cold))
4414 ixgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
4416 struct ixgbe_hw *hw;
4417 struct ixgbe_tx_queue *txq;
4421 PMD_INIT_FUNC_TRACE();
4422 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4424 if (tx_queue_id < dev->data->nb_tx_queues) {
4425 txq = dev->data->tx_queues[tx_queue_id];
4426 txdctl = IXGBE_READ_REG(hw, IXGBE_TXDCTL(txq->reg_idx));
4427 txdctl |= IXGBE_TXDCTL_ENABLE;
4428 IXGBE_WRITE_REG(hw, IXGBE_TXDCTL(txq->reg_idx), txdctl);
4430 /* Wait until TX Enable ready */
4431 if (hw->mac.type == ixgbe_mac_82599EB) {
4432 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4435 txdctl = IXGBE_READ_REG(hw,
4436 IXGBE_TXDCTL(txq->reg_idx));
4437 } while (--poll_ms && !(txdctl & IXGBE_TXDCTL_ENABLE));
4439 PMD_INIT_LOG(ERR, "Could not enable "
4440 "Tx Queue %d", tx_queue_id);
4443 IXGBE_WRITE_REG(hw, IXGBE_TDH(txq->reg_idx), 0);
4444 IXGBE_WRITE_REG(hw, IXGBE_TDT(txq->reg_idx), 0);
4452 * Stop Transmit Units for specified queue.
4454 int __attribute__((cold))
4455 ixgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
4457 struct ixgbe_hw *hw;
4458 struct ixgbe_tx_queue *txq;
4460 uint32_t txtdh, txtdt;
4463 PMD_INIT_FUNC_TRACE();
4464 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4466 if (tx_queue_id < dev->data->nb_tx_queues) {
4467 txq = dev->data->tx_queues[tx_queue_id];
4469 /* Wait until TX queue is empty */
4470 if (hw->mac.type == ixgbe_mac_82599EB) {
4471 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4473 rte_delay_us(RTE_IXGBE_WAIT_100_US);
4474 txtdh = IXGBE_READ_REG(hw,
4475 IXGBE_TDH(txq->reg_idx));
4476 txtdt = IXGBE_READ_REG(hw,
4477 IXGBE_TDT(txq->reg_idx));
4478 } while (--poll_ms && (txtdh != txtdt));
4480 PMD_INIT_LOG(ERR, "Tx Queue %d is not empty "
4481 "when stopping.", tx_queue_id);
4484 txdctl = IXGBE_READ_REG(hw, IXGBE_TXDCTL(txq->reg_idx));
4485 txdctl &= ~IXGBE_TXDCTL_ENABLE;
4486 IXGBE_WRITE_REG(hw, IXGBE_TXDCTL(txq->reg_idx), txdctl);
4488 /* Wait until TX Enable ready */
4489 if (hw->mac.type == ixgbe_mac_82599EB) {
4490 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4493 txdctl = IXGBE_READ_REG(hw,
4494 IXGBE_TXDCTL(txq->reg_idx));
4495 } while (--poll_ms && (txdctl | IXGBE_TXDCTL_ENABLE));
4497 PMD_INIT_LOG(ERR, "Could not disable "
4498 "Tx Queue %d", tx_queue_id);
4501 if (txq->ops != NULL) {
4502 txq->ops->release_mbufs(txq);
4503 txq->ops->reset(txq);
4512 * [VF] Initializes Receive Unit.
4514 int __attribute__((cold))
4515 ixgbevf_dev_rx_init(struct rte_eth_dev *dev)
4517 struct ixgbe_hw *hw;
4518 struct ixgbe_rx_queue *rxq;
4520 uint32_t srrctl, psrtype = 0;
4525 PMD_INIT_FUNC_TRACE();
4526 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4528 if (rte_is_power_of_2(dev->data->nb_rx_queues) == 0) {
4529 PMD_INIT_LOG(ERR, "The number of Rx queue invalid, "
4530 "it should be power of 2");
4534 if (dev->data->nb_rx_queues > hw->mac.max_rx_queues) {
4535 PMD_INIT_LOG(ERR, "The number of Rx queue invalid, "
4536 "it should be equal to or less than %d",
4537 hw->mac.max_rx_queues);
4542 * When the VF driver issues a IXGBE_VF_RESET request, the PF driver
4543 * disables the VF receipt of packets if the PF MTU is > 1500.
4544 * This is done to deal with 82599 limitations that imposes
4545 * the PF and all VFs to share the same MTU.
4546 * Then, the PF driver enables again the VF receipt of packet when
4547 * the VF driver issues a IXGBE_VF_SET_LPE request.
4548 * In the meantime, the VF device cannot be used, even if the VF driver
4549 * and the Guest VM network stack are ready to accept packets with a
4550 * size up to the PF MTU.
4551 * As a work-around to this PF behaviour, force the call to
4552 * ixgbevf_rlpml_set_vf even if jumbo frames are not used. This way,
4553 * VF packets received can work in all cases.
4555 ixgbevf_rlpml_set_vf(hw,
4556 (uint16_t)dev->data->dev_conf.rxmode.max_rx_pkt_len);
4558 /* Setup RX queues */
4559 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4560 rxq = dev->data->rx_queues[i];
4562 /* Allocate buffers for descriptor rings */
4563 ret = ixgbe_alloc_rx_queue_mbufs(rxq);
4567 /* Setup the Base and Length of the Rx Descriptor Rings */
4568 bus_addr = rxq->rx_ring_phys_addr;
4570 IXGBE_WRITE_REG(hw, IXGBE_VFRDBAL(i),
4571 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4572 IXGBE_WRITE_REG(hw, IXGBE_VFRDBAH(i),
4573 (uint32_t)(bus_addr >> 32));
4574 IXGBE_WRITE_REG(hw, IXGBE_VFRDLEN(i),
4575 rxq->nb_rx_desc * sizeof(union ixgbe_adv_rx_desc));
4576 IXGBE_WRITE_REG(hw, IXGBE_VFRDH(i), 0);
4577 IXGBE_WRITE_REG(hw, IXGBE_VFRDT(i), 0);
4580 /* Configure the SRRCTL register */
4581 #ifdef RTE_HEADER_SPLIT_ENABLE
4583 * Configure Header Split
4585 if (dev->data->dev_conf.rxmode.header_split) {
4586 srrctl = ((dev->data->dev_conf.rxmode.split_hdr_size <<
4587 IXGBE_SRRCTL_BSIZEHDRSIZE_SHIFT) &
4588 IXGBE_SRRCTL_BSIZEHDR_MASK);
4589 srrctl |= IXGBE_SRRCTL_DESCTYPE_HDR_SPLIT_ALWAYS;
4592 srrctl = IXGBE_SRRCTL_DESCTYPE_ADV_ONEBUF;
4594 /* Set if packets are dropped when no descriptors available */
4596 srrctl |= IXGBE_SRRCTL_DROP_EN;
4599 * Configure the RX buffer size in the BSIZEPACKET field of
4600 * the SRRCTL register of the queue.
4601 * The value is in 1 KB resolution. Valid values can be from
4604 buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
4605 RTE_PKTMBUF_HEADROOM);
4606 srrctl |= ((buf_size >> IXGBE_SRRCTL_BSIZEPKT_SHIFT) &
4607 IXGBE_SRRCTL_BSIZEPKT_MASK);
4610 * VF modification to write virtual function SRRCTL register
4612 IXGBE_WRITE_REG(hw, IXGBE_VFSRRCTL(i), srrctl);
4614 buf_size = (uint16_t) ((srrctl & IXGBE_SRRCTL_BSIZEPKT_MASK) <<
4615 IXGBE_SRRCTL_BSIZEPKT_SHIFT);
4617 if (dev->data->dev_conf.rxmode.enable_scatter ||
4618 /* It adds dual VLAN length for supporting dual VLAN */
4619 (dev->data->dev_conf.rxmode.max_rx_pkt_len +
4620 2 * IXGBE_VLAN_TAG_SIZE) > buf_size) {
4621 if (!dev->data->scattered_rx)
4622 PMD_INIT_LOG(DEBUG, "forcing scatter mode");
4623 dev->data->scattered_rx = 1;
4627 #ifdef RTE_HEADER_SPLIT_ENABLE
4628 if (dev->data->dev_conf.rxmode.header_split)
4629 /* Must setup the PSRTYPE register */
4630 psrtype = IXGBE_PSRTYPE_TCPHDR |
4631 IXGBE_PSRTYPE_UDPHDR |
4632 IXGBE_PSRTYPE_IPV4HDR |
4633 IXGBE_PSRTYPE_IPV6HDR;
4636 /* Set RQPL for VF RSS according to max Rx queue */
4637 psrtype |= (dev->data->nb_rx_queues >> 1) <<
4638 IXGBE_PSRTYPE_RQPL_SHIFT;
4639 IXGBE_WRITE_REG(hw, IXGBE_VFPSRTYPE, psrtype);
4641 ixgbe_set_rx_function(dev);
4647 * [VF] Initializes Transmit Unit.
4649 void __attribute__((cold))
4650 ixgbevf_dev_tx_init(struct rte_eth_dev *dev)
4652 struct ixgbe_hw *hw;
4653 struct ixgbe_tx_queue *txq;
4658 PMD_INIT_FUNC_TRACE();
4659 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4661 /* Setup the Base and Length of the Tx Descriptor Rings */
4662 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4663 txq = dev->data->tx_queues[i];
4664 bus_addr = txq->tx_ring_phys_addr;
4665 IXGBE_WRITE_REG(hw, IXGBE_VFTDBAL(i),
4666 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4667 IXGBE_WRITE_REG(hw, IXGBE_VFTDBAH(i),
4668 (uint32_t)(bus_addr >> 32));
4669 IXGBE_WRITE_REG(hw, IXGBE_VFTDLEN(i),
4670 txq->nb_tx_desc * sizeof(union ixgbe_adv_tx_desc));
4671 /* Setup the HW Tx Head and TX Tail descriptor pointers */
4672 IXGBE_WRITE_REG(hw, IXGBE_VFTDH(i), 0);
4673 IXGBE_WRITE_REG(hw, IXGBE_VFTDT(i), 0);
4676 * Disable Tx Head Writeback RO bit, since this hoses
4677 * bookkeeping if things aren't delivered in order.
4679 txctrl = IXGBE_READ_REG(hw,
4680 IXGBE_VFDCA_TXCTRL(i));
4681 txctrl &= ~IXGBE_DCA_TXCTRL_DESC_WRO_EN;
4682 IXGBE_WRITE_REG(hw, IXGBE_VFDCA_TXCTRL(i),
4688 * [VF] Start Transmit and Receive Units.
4690 void __attribute__((cold))
4691 ixgbevf_dev_rxtx_start(struct rte_eth_dev *dev)
4693 struct ixgbe_hw *hw;
4694 struct ixgbe_tx_queue *txq;
4695 struct ixgbe_rx_queue *rxq;
4701 PMD_INIT_FUNC_TRACE();
4702 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4704 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4705 txq = dev->data->tx_queues[i];
4706 /* Setup Transmit Threshold Registers */
4707 txdctl = IXGBE_READ_REG(hw, IXGBE_VFTXDCTL(i));
4708 txdctl |= txq->pthresh & 0x7F;
4709 txdctl |= ((txq->hthresh & 0x7F) << 8);
4710 txdctl |= ((txq->wthresh & 0x7F) << 16);
4711 IXGBE_WRITE_REG(hw, IXGBE_VFTXDCTL(i), txdctl);
4714 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4716 txdctl = IXGBE_READ_REG(hw, IXGBE_VFTXDCTL(i));
4717 txdctl |= IXGBE_TXDCTL_ENABLE;
4718 IXGBE_WRITE_REG(hw, IXGBE_VFTXDCTL(i), txdctl);
4721 /* Wait until TX Enable ready */
4724 txdctl = IXGBE_READ_REG(hw, IXGBE_VFTXDCTL(i));
4725 } while (--poll_ms && !(txdctl & IXGBE_TXDCTL_ENABLE));
4727 PMD_INIT_LOG(ERR, "Could not enable Tx Queue %d", i);
4729 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4731 rxq = dev->data->rx_queues[i];
4733 rxdctl = IXGBE_READ_REG(hw, IXGBE_VFRXDCTL(i));
4734 rxdctl |= IXGBE_RXDCTL_ENABLE;
4735 IXGBE_WRITE_REG(hw, IXGBE_VFRXDCTL(i), rxdctl);
4737 /* Wait until RX Enable ready */
4741 rxdctl = IXGBE_READ_REG(hw, IXGBE_VFRXDCTL(i));
4742 } while (--poll_ms && !(rxdctl & IXGBE_RXDCTL_ENABLE));
4744 PMD_INIT_LOG(ERR, "Could not enable Rx Queue %d", i);
4746 IXGBE_WRITE_REG(hw, IXGBE_VFRDT(i), rxq->nb_rx_desc - 1);
4751 /* Stubs needed for linkage when CONFIG_RTE_IXGBE_INC_VECTOR is set to 'n' */
4752 int __attribute__((weak))
4753 ixgbe_rx_vec_dev_conf_condition_check(struct rte_eth_dev __rte_unused *dev)
4758 uint16_t __attribute__((weak))
4759 ixgbe_recv_pkts_vec(
4760 void __rte_unused *rx_queue,
4761 struct rte_mbuf __rte_unused **rx_pkts,
4762 uint16_t __rte_unused nb_pkts)
4767 uint16_t __attribute__((weak))
4768 ixgbe_recv_scattered_pkts_vec(
4769 void __rte_unused *rx_queue,
4770 struct rte_mbuf __rte_unused **rx_pkts,
4771 uint16_t __rte_unused nb_pkts)
4776 int __attribute__((weak))
4777 ixgbe_rxq_vec_setup(struct ixgbe_rx_queue __rte_unused *rxq)