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_free < txq->tx_free_thresh)
597 ixgbe_xmit_cleanup(txq);
599 rte_prefetch0(&txe->mbuf->pool);
602 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
605 pkt_len = tx_pkt->pkt_len;
608 * Determine how many (if any) context descriptors
609 * are needed for offload functionality.
611 ol_flags = tx_pkt->ol_flags;
613 /* If hardware offload required */
614 tx_ol_req = ol_flags & IXGBE_TX_OFFLOAD_MASK;
616 tx_offload.l2_len = tx_pkt->l2_len;
617 tx_offload.l3_len = tx_pkt->l3_len;
618 tx_offload.l4_len = tx_pkt->l4_len;
619 tx_offload.vlan_tci = tx_pkt->vlan_tci;
620 tx_offload.tso_segsz = tx_pkt->tso_segsz;
622 /* If new context need be built or reuse the exist ctx. */
623 ctx = what_advctx_update(txq, tx_ol_req,
625 /* Only allocate context descriptor if required*/
626 new_ctx = (ctx == IXGBE_CTX_NUM);
631 * Keep track of how many descriptors are used this loop
632 * This will always be the number of segments + the number of
633 * Context descriptors required to transmit the packet
635 nb_used = (uint16_t)(tx_pkt->nb_segs + new_ctx);
638 * The number of descriptors that must be allocated for a
639 * packet is the number of segments of that packet, plus 1
640 * Context Descriptor for the hardware offload, if any.
641 * Determine the last TX descriptor to allocate in the TX ring
642 * for the packet, starting from the current position (tx_id)
645 tx_last = (uint16_t) (tx_id + nb_used - 1);
648 if (tx_last >= txq->nb_tx_desc)
649 tx_last = (uint16_t) (tx_last - txq->nb_tx_desc);
651 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u pktlen=%u"
652 " tx_first=%u tx_last=%u",
653 (unsigned) txq->port_id,
654 (unsigned) txq->queue_id,
660 * Make sure there are enough TX descriptors available to
661 * transmit the entire packet.
662 * nb_used better be less than or equal to txq->tx_rs_thresh
664 if (nb_used > txq->nb_tx_free) {
665 PMD_TX_FREE_LOG(DEBUG,
666 "Not enough free TX descriptors "
667 "nb_used=%4u nb_free=%4u "
668 "(port=%d queue=%d)",
669 nb_used, txq->nb_tx_free,
670 txq->port_id, txq->queue_id);
672 if (ixgbe_xmit_cleanup(txq) != 0) {
673 /* Could not clean any descriptors */
679 /* nb_used better be <= txq->tx_rs_thresh */
680 if (unlikely(nb_used > txq->tx_rs_thresh)) {
681 PMD_TX_FREE_LOG(DEBUG,
682 "The number of descriptors needed to "
683 "transmit the packet exceeds the "
684 "RS bit threshold. This will impact "
686 "nb_used=%4u nb_free=%4u "
688 "(port=%d queue=%d)",
689 nb_used, txq->nb_tx_free,
691 txq->port_id, txq->queue_id);
693 * Loop here until there are enough TX
694 * descriptors or until the ring cannot be
697 while (nb_used > txq->nb_tx_free) {
698 if (ixgbe_xmit_cleanup(txq) != 0) {
700 * Could not clean any
712 * By now there are enough free TX descriptors to transmit
717 * Set common flags of all TX Data Descriptors.
719 * The following bits must be set in all Data Descriptors:
720 * - IXGBE_ADVTXD_DTYP_DATA
721 * - IXGBE_ADVTXD_DCMD_DEXT
723 * The following bits must be set in the first Data Descriptor
724 * and are ignored in the other ones:
725 * - IXGBE_ADVTXD_DCMD_IFCS
726 * - IXGBE_ADVTXD_MAC_1588
727 * - IXGBE_ADVTXD_DCMD_VLE
729 * The following bits must only be set in the last Data
731 * - IXGBE_TXD_CMD_EOP
733 * The following bits can be set in any Data Descriptor, but
734 * are only set in the last Data Descriptor:
737 cmd_type_len = IXGBE_ADVTXD_DTYP_DATA |
738 IXGBE_ADVTXD_DCMD_IFCS | IXGBE_ADVTXD_DCMD_DEXT;
740 #ifdef RTE_LIBRTE_IEEE1588
741 if (ol_flags & PKT_TX_IEEE1588_TMST)
742 cmd_type_len |= IXGBE_ADVTXD_MAC_1588;
748 if (ol_flags & PKT_TX_TCP_SEG) {
749 /* when TSO is on, paylen in descriptor is the
750 * not the packet len but the tcp payload len */
751 pkt_len -= (tx_offload.l2_len +
752 tx_offload.l3_len + tx_offload.l4_len);
756 * Setup the TX Advanced Context Descriptor if required
759 volatile struct ixgbe_adv_tx_context_desc *
762 ctx_txd = (volatile struct
763 ixgbe_adv_tx_context_desc *)
766 txn = &sw_ring[txe->next_id];
767 rte_prefetch0(&txn->mbuf->pool);
769 if (txe->mbuf != NULL) {
770 rte_pktmbuf_free_seg(txe->mbuf);
774 ixgbe_set_xmit_ctx(txq, ctx_txd, tx_ol_req,
777 txe->last_id = tx_last;
778 tx_id = txe->next_id;
783 * Setup the TX Advanced Data Descriptor,
784 * This path will go through
785 * whatever new/reuse the context descriptor
787 cmd_type_len |= tx_desc_ol_flags_to_cmdtype(ol_flags);
788 olinfo_status |= tx_desc_cksum_flags_to_olinfo(ol_flags);
789 olinfo_status |= ctx << IXGBE_ADVTXD_IDX_SHIFT;
792 olinfo_status |= (pkt_len << IXGBE_ADVTXD_PAYLEN_SHIFT);
797 txn = &sw_ring[txe->next_id];
798 rte_prefetch0(&txn->mbuf->pool);
800 if (txe->mbuf != NULL)
801 rte_pktmbuf_free_seg(txe->mbuf);
805 * Set up Transmit Data Descriptor.
807 slen = m_seg->data_len;
808 buf_dma_addr = RTE_MBUF_DATA_DMA_ADDR(m_seg);
809 txd->read.buffer_addr =
810 rte_cpu_to_le_64(buf_dma_addr);
811 txd->read.cmd_type_len =
812 rte_cpu_to_le_32(cmd_type_len | slen);
813 txd->read.olinfo_status =
814 rte_cpu_to_le_32(olinfo_status);
815 txe->last_id = tx_last;
816 tx_id = txe->next_id;
819 } while (m_seg != NULL);
822 * The last packet data descriptor needs End Of Packet (EOP)
824 cmd_type_len |= IXGBE_TXD_CMD_EOP;
825 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
826 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
828 /* Set RS bit only on threshold packets' last descriptor */
829 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
830 PMD_TX_FREE_LOG(DEBUG,
831 "Setting RS bit on TXD id="
832 "%4u (port=%d queue=%d)",
833 tx_last, txq->port_id, txq->queue_id);
835 cmd_type_len |= IXGBE_TXD_CMD_RS;
837 /* Update txq RS bit counters */
840 txd->read.cmd_type_len |= rte_cpu_to_le_32(cmd_type_len);
846 * Set the Transmit Descriptor Tail (TDT)
848 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
849 (unsigned) txq->port_id, (unsigned) txq->queue_id,
850 (unsigned) tx_id, (unsigned) nb_tx);
851 IXGBE_PCI_REG_WRITE(txq->tdt_reg_addr, tx_id);
852 txq->tx_tail = tx_id;
857 /*********************************************************************
861 **********************************************************************/
862 static inline uint64_t
863 rx_desc_hlen_type_rss_to_pkt_flags(uint32_t hl_tp_rs)
867 static const uint64_t ip_pkt_types_map[16] = {
868 0, PKT_RX_IPV4_HDR, PKT_RX_IPV4_HDR_EXT, PKT_RX_IPV4_HDR_EXT,
869 PKT_RX_IPV6_HDR, 0, 0, 0,
870 PKT_RX_IPV6_HDR_EXT, 0, 0, 0,
871 PKT_RX_IPV6_HDR_EXT, 0, 0, 0,
874 static const uint64_t ip_rss_types_map[16] = {
875 0, PKT_RX_RSS_HASH, PKT_RX_RSS_HASH, PKT_RX_RSS_HASH,
876 0, PKT_RX_RSS_HASH, 0, PKT_RX_RSS_HASH,
877 PKT_RX_RSS_HASH, 0, 0, 0,
878 0, 0, 0, PKT_RX_FDIR,
881 #ifdef RTE_LIBRTE_IEEE1588
882 static uint64_t ip_pkt_etqf_map[8] = {
883 0, 0, 0, PKT_RX_IEEE1588_PTP,
887 pkt_flags = (hl_tp_rs & IXGBE_RXDADV_PKTTYPE_ETQF) ?
888 ip_pkt_etqf_map[(hl_tp_rs >> 4) & 0x07] :
889 ip_pkt_types_map[(hl_tp_rs >> 4) & 0x0F];
891 pkt_flags = (hl_tp_rs & IXGBE_RXDADV_PKTTYPE_ETQF) ? 0 :
892 ip_pkt_types_map[(hl_tp_rs >> 4) & 0x0F];
895 return pkt_flags | ip_rss_types_map[hl_tp_rs & 0xF];
898 static inline uint64_t
899 rx_desc_status_to_pkt_flags(uint32_t rx_status)
904 * Check if VLAN present only.
905 * Do not check whether L3/L4 rx checksum done by NIC or not,
906 * That can be found from rte_eth_rxmode.hw_ip_checksum flag
908 pkt_flags = (rx_status & IXGBE_RXD_STAT_VP) ? PKT_RX_VLAN_PKT : 0;
910 #ifdef RTE_LIBRTE_IEEE1588
911 if (rx_status & IXGBE_RXD_STAT_TMST)
912 pkt_flags = pkt_flags | PKT_RX_IEEE1588_TMST;
917 static inline uint64_t
918 rx_desc_error_to_pkt_flags(uint32_t rx_status)
921 * Bit 31: IPE, IPv4 checksum error
922 * Bit 30: L4I, L4I integrity error
924 static uint64_t error_to_pkt_flags_map[4] = {
925 0, PKT_RX_L4_CKSUM_BAD, PKT_RX_IP_CKSUM_BAD,
926 PKT_RX_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD
928 return error_to_pkt_flags_map[(rx_status >>
929 IXGBE_RXDADV_ERR_CKSUM_BIT) & IXGBE_RXDADV_ERR_CKSUM_MSK];
932 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
934 * LOOK_AHEAD defines how many desc statuses to check beyond the
935 * current descriptor.
936 * It must be a pound define for optimal performance.
937 * Do not change the value of LOOK_AHEAD, as the ixgbe_rx_scan_hw_ring
938 * function only works with LOOK_AHEAD=8.
941 #if (LOOK_AHEAD != 8)
942 #error "PMD IXGBE: LOOK_AHEAD must be 8\n"
945 ixgbe_rx_scan_hw_ring(struct ixgbe_rx_queue *rxq)
947 volatile union ixgbe_adv_rx_desc *rxdp;
948 struct ixgbe_rx_entry *rxep;
952 int s[LOOK_AHEAD], nb_dd;
956 /* get references to current descriptor and S/W ring entry */
957 rxdp = &rxq->rx_ring[rxq->rx_tail];
958 rxep = &rxq->sw_ring[rxq->rx_tail];
960 /* check to make sure there is at least 1 packet to receive */
961 if (! (rxdp->wb.upper.status_error & IXGBE_RXDADV_STAT_DD))
965 * Scan LOOK_AHEAD descriptors at a time to determine which descriptors
966 * reference packets that are ready to be received.
968 for (i = 0; i < RTE_PMD_IXGBE_RX_MAX_BURST;
969 i += LOOK_AHEAD, rxdp += LOOK_AHEAD, rxep += LOOK_AHEAD)
971 /* Read desc statuses backwards to avoid race condition */
972 for (j = LOOK_AHEAD-1; j >= 0; --j)
973 s[j] = rxdp[j].wb.upper.status_error;
975 /* Compute how many status bits were set */
977 for (j = 0; j < LOOK_AHEAD; ++j)
978 nb_dd += s[j] & IXGBE_RXDADV_STAT_DD;
982 /* Translate descriptor info to mbuf format */
983 for (j = 0; j < nb_dd; ++j) {
985 pkt_len = (uint16_t)(rxdp[j].wb.upper.length - rxq->crc_len);
986 mb->data_len = pkt_len;
987 mb->pkt_len = pkt_len;
988 mb->vlan_tci = rxdp[j].wb.upper.vlan;
989 mb->vlan_tci = rte_le_to_cpu_16(rxdp[j].wb.upper.vlan);
991 /* convert descriptor fields to rte mbuf flags */
992 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(
993 rxdp[j].wb.lower.lo_dword.data);
994 /* reuse status field from scan list */
995 pkt_flags |= rx_desc_status_to_pkt_flags(s[j]);
996 pkt_flags |= rx_desc_error_to_pkt_flags(s[j]);
997 mb->ol_flags = pkt_flags;
999 if (likely(pkt_flags & PKT_RX_RSS_HASH))
1000 mb->hash.rss = rxdp[j].wb.lower.hi_dword.rss;
1001 else if (pkt_flags & PKT_RX_FDIR) {
1002 mb->hash.fdir.hash =
1003 (uint16_t)((rxdp[j].wb.lower.hi_dword.csum_ip.csum)
1004 & IXGBE_ATR_HASH_MASK);
1005 mb->hash.fdir.id = rxdp[j].wb.lower.hi_dword.csum_ip.ip_id;
1009 /* Move mbuf pointers from the S/W ring to the stage */
1010 for (j = 0; j < LOOK_AHEAD; ++j) {
1011 rxq->rx_stage[i + j] = rxep[j].mbuf;
1014 /* stop if all requested packets could not be received */
1015 if (nb_dd != LOOK_AHEAD)
1019 /* clear software ring entries so we can cleanup correctly */
1020 for (i = 0; i < nb_rx; ++i) {
1021 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
1029 ixgbe_rx_alloc_bufs(struct ixgbe_rx_queue *rxq, bool reset_mbuf)
1031 volatile union ixgbe_adv_rx_desc *rxdp;
1032 struct ixgbe_rx_entry *rxep;
1033 struct rte_mbuf *mb;
1038 /* allocate buffers in bulk directly into the S/W ring */
1039 alloc_idx = rxq->rx_free_trigger - (rxq->rx_free_thresh - 1);
1040 rxep = &rxq->sw_ring[alloc_idx];
1041 diag = rte_mempool_get_bulk(rxq->mb_pool, (void *)rxep,
1042 rxq->rx_free_thresh);
1043 if (unlikely(diag != 0))
1046 rxdp = &rxq->rx_ring[alloc_idx];
1047 for (i = 0; i < rxq->rx_free_thresh; ++i) {
1048 /* populate the static rte mbuf fields */
1053 mb->port = rxq->port_id;
1056 rte_mbuf_refcnt_set(mb, 1);
1057 mb->data_off = RTE_PKTMBUF_HEADROOM;
1059 /* populate the descriptors */
1060 dma_addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mb));
1061 rxdp[i].read.hdr_addr = dma_addr;
1062 rxdp[i].read.pkt_addr = dma_addr;
1065 /* update state of internal queue structure */
1066 rxq->rx_free_trigger = rxq->rx_free_trigger + rxq->rx_free_thresh;
1067 if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
1068 rxq->rx_free_trigger = rxq->rx_free_thresh - 1;
1074 static inline uint16_t
1075 ixgbe_rx_fill_from_stage(struct ixgbe_rx_queue *rxq, struct rte_mbuf **rx_pkts,
1078 struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
1081 /* how many packets are ready to return? */
1082 nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
1084 /* copy mbuf pointers to the application's packet list */
1085 for (i = 0; i < nb_pkts; ++i)
1086 rx_pkts[i] = stage[i];
1088 /* update internal queue state */
1089 rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
1090 rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
1095 static inline uint16_t
1096 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
1099 struct ixgbe_rx_queue *rxq = (struct ixgbe_rx_queue *)rx_queue;
1102 /* Any previously recv'd pkts will be returned from the Rx stage */
1103 if (rxq->rx_nb_avail)
1104 return ixgbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
1106 /* Scan the H/W ring for packets to receive */
1107 nb_rx = (uint16_t)ixgbe_rx_scan_hw_ring(rxq);
1109 /* update internal queue state */
1110 rxq->rx_next_avail = 0;
1111 rxq->rx_nb_avail = nb_rx;
1112 rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
1114 /* if required, allocate new buffers to replenish descriptors */
1115 if (rxq->rx_tail > rxq->rx_free_trigger) {
1116 uint16_t cur_free_trigger = rxq->rx_free_trigger;
1118 if (ixgbe_rx_alloc_bufs(rxq, true) != 0) {
1120 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
1121 "queue_id=%u", (unsigned) rxq->port_id,
1122 (unsigned) rxq->queue_id);
1124 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
1125 rxq->rx_free_thresh;
1128 * Need to rewind any previous receives if we cannot
1129 * allocate new buffers to replenish the old ones.
1131 rxq->rx_nb_avail = 0;
1132 rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
1133 for (i = 0, j = rxq->rx_tail; i < nb_rx; ++i, ++j)
1134 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
1139 /* update tail pointer */
1141 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, cur_free_trigger);
1144 if (rxq->rx_tail >= rxq->nb_rx_desc)
1147 /* received any packets this loop? */
1148 if (rxq->rx_nb_avail)
1149 return ixgbe_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
1154 /* split requests into chunks of size RTE_PMD_IXGBE_RX_MAX_BURST */
1156 ixgbe_recv_pkts_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1161 if (unlikely(nb_pkts == 0))
1164 if (likely(nb_pkts <= RTE_PMD_IXGBE_RX_MAX_BURST))
1165 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
1167 /* request is relatively large, chunk it up */
1171 n = (uint16_t)RTE_MIN(nb_pkts, RTE_PMD_IXGBE_RX_MAX_BURST);
1172 ret = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
1173 nb_rx = (uint16_t)(nb_rx + ret);
1174 nb_pkts = (uint16_t)(nb_pkts - ret);
1184 /* Stub to avoid extra ifdefs */
1186 ixgbe_recv_pkts_bulk_alloc(__rte_unused void *rx_queue,
1187 __rte_unused struct rte_mbuf **rx_pkts, __rte_unused uint16_t nb_pkts)
1193 ixgbe_rx_alloc_bufs(__rte_unused struct ixgbe_rx_queue *rxq,
1194 __rte_unused bool reset_mbuf)
1198 #endif /* RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC */
1201 ixgbe_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
1204 struct ixgbe_rx_queue *rxq;
1205 volatile union ixgbe_adv_rx_desc *rx_ring;
1206 volatile union ixgbe_adv_rx_desc *rxdp;
1207 struct ixgbe_rx_entry *sw_ring;
1208 struct ixgbe_rx_entry *rxe;
1209 struct rte_mbuf *rxm;
1210 struct rte_mbuf *nmb;
1211 union ixgbe_adv_rx_desc rxd;
1214 uint32_t hlen_type_rss;
1224 rx_id = rxq->rx_tail;
1225 rx_ring = rxq->rx_ring;
1226 sw_ring = rxq->sw_ring;
1227 while (nb_rx < nb_pkts) {
1229 * The order of operations here is important as the DD status
1230 * bit must not be read after any other descriptor fields.
1231 * rx_ring and rxdp are pointing to volatile data so the order
1232 * of accesses cannot be reordered by the compiler. If they were
1233 * not volatile, they could be reordered which could lead to
1234 * using invalid descriptor fields when read from rxd.
1236 rxdp = &rx_ring[rx_id];
1237 staterr = rxdp->wb.upper.status_error;
1238 if (! (staterr & rte_cpu_to_le_32(IXGBE_RXDADV_STAT_DD)))
1245 * If the IXGBE_RXDADV_STAT_EOP flag is not set, the RX packet
1246 * is likely to be invalid and to be dropped by the various
1247 * validation checks performed by the network stack.
1249 * Allocate a new mbuf to replenish the RX ring descriptor.
1250 * If the allocation fails:
1251 * - arrange for that RX descriptor to be the first one
1252 * being parsed the next time the receive function is
1253 * invoked [on the same queue].
1255 * - Stop parsing the RX ring and return immediately.
1257 * This policy do not drop the packet received in the RX
1258 * descriptor for which the allocation of a new mbuf failed.
1259 * Thus, it allows that packet to be later retrieved if
1260 * mbuf have been freed in the mean time.
1261 * As a side effect, holding RX descriptors instead of
1262 * systematically giving them back to the NIC may lead to
1263 * RX ring exhaustion situations.
1264 * However, the NIC can gracefully prevent such situations
1265 * to happen by sending specific "back-pressure" flow control
1266 * frames to its peer(s).
1268 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
1269 "ext_err_stat=0x%08x pkt_len=%u",
1270 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1271 (unsigned) rx_id, (unsigned) staterr,
1272 (unsigned) rte_le_to_cpu_16(rxd.wb.upper.length));
1274 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
1276 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
1277 "queue_id=%u", (unsigned) rxq->port_id,
1278 (unsigned) rxq->queue_id);
1279 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
1284 rxe = &sw_ring[rx_id];
1286 if (rx_id == rxq->nb_rx_desc)
1289 /* Prefetch next mbuf while processing current one. */
1290 rte_ixgbe_prefetch(sw_ring[rx_id].mbuf);
1293 * When next RX descriptor is on a cache-line boundary,
1294 * prefetch the next 4 RX descriptors and the next 8 pointers
1297 if ((rx_id & 0x3) == 0) {
1298 rte_ixgbe_prefetch(&rx_ring[rx_id]);
1299 rte_ixgbe_prefetch(&sw_ring[rx_id]);
1305 rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
1306 rxdp->read.hdr_addr = dma_addr;
1307 rxdp->read.pkt_addr = dma_addr;
1310 * Initialize the returned mbuf.
1311 * 1) setup generic mbuf fields:
1312 * - number of segments,
1315 * - RX port identifier.
1316 * 2) integrate hardware offload data, if any:
1317 * - RSS flag & hash,
1318 * - IP checksum flag,
1319 * - VLAN TCI, if any,
1322 pkt_len = (uint16_t) (rte_le_to_cpu_16(rxd.wb.upper.length) -
1324 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1325 rte_packet_prefetch((char *)rxm->buf_addr + rxm->data_off);
1328 rxm->pkt_len = pkt_len;
1329 rxm->data_len = pkt_len;
1330 rxm->port = rxq->port_id;
1332 hlen_type_rss = rte_le_to_cpu_32(rxd.wb.lower.lo_dword.data);
1333 /* Only valid if PKT_RX_VLAN_PKT set in pkt_flags */
1334 rxm->vlan_tci = rte_le_to_cpu_16(rxd.wb.upper.vlan);
1336 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(hlen_type_rss);
1337 pkt_flags = pkt_flags | rx_desc_status_to_pkt_flags(staterr);
1338 pkt_flags = pkt_flags | rx_desc_error_to_pkt_flags(staterr);
1339 rxm->ol_flags = pkt_flags;
1341 if (likely(pkt_flags & PKT_RX_RSS_HASH))
1342 rxm->hash.rss = rxd.wb.lower.hi_dword.rss;
1343 else if (pkt_flags & PKT_RX_FDIR) {
1344 rxm->hash.fdir.hash =
1345 (uint16_t)((rxd.wb.lower.hi_dword.csum_ip.csum)
1346 & IXGBE_ATR_HASH_MASK);
1347 rxm->hash.fdir.id = rxd.wb.lower.hi_dword.csum_ip.ip_id;
1350 * Store the mbuf address into the next entry of the array
1351 * of returned packets.
1353 rx_pkts[nb_rx++] = rxm;
1355 rxq->rx_tail = rx_id;
1358 * If the number of free RX descriptors is greater than the RX free
1359 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1361 * Update the RDT with the value of the last processed RX descriptor
1362 * minus 1, to guarantee that the RDT register is never equal to the
1363 * RDH register, which creates a "full" ring situtation from the
1364 * hardware point of view...
1366 nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
1367 if (nb_hold > rxq->rx_free_thresh) {
1368 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1369 "nb_hold=%u nb_rx=%u",
1370 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1371 (unsigned) rx_id, (unsigned) nb_hold,
1373 rx_id = (uint16_t) ((rx_id == 0) ?
1374 (rxq->nb_rx_desc - 1) : (rx_id - 1));
1375 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
1378 rxq->nb_rx_hold = nb_hold;
1383 * Detect an RSC descriptor.
1385 static inline uint32_t
1386 ixgbe_rsc_count(union ixgbe_adv_rx_desc *rx)
1388 return (rte_le_to_cpu_32(rx->wb.lower.lo_dword.data) &
1389 IXGBE_RXDADV_RSCCNT_MASK) >> IXGBE_RXDADV_RSCCNT_SHIFT;
1393 * ixgbe_fill_cluster_head_buf - fill the first mbuf of the returned packet
1395 * Fill the following info in the HEAD buffer of the Rx cluster:
1396 * - RX port identifier
1397 * - hardware offload data, if any:
1399 * - IP checksum flag
1400 * - VLAN TCI, if any
1402 * @head HEAD of the packet cluster
1403 * @desc HW descriptor to get data from
1404 * @port_id Port ID of the Rx queue
1407 ixgbe_fill_cluster_head_buf(
1408 struct rte_mbuf *head,
1409 union ixgbe_adv_rx_desc *desc,
1413 uint32_t hlen_type_rss;
1416 head->port = port_id;
1419 * The vlan_tci field is only valid when PKT_RX_VLAN_PKT is
1420 * set in the pkt_flags field.
1422 head->vlan_tci = rte_le_to_cpu_16(desc->wb.upper.vlan);
1423 hlen_type_rss = rte_le_to_cpu_32(desc->wb.lower.lo_dword.data);
1424 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(hlen_type_rss);
1425 pkt_flags |= rx_desc_status_to_pkt_flags(staterr);
1426 pkt_flags |= rx_desc_error_to_pkt_flags(staterr);
1427 head->ol_flags = pkt_flags;
1429 if (likely(pkt_flags & PKT_RX_RSS_HASH))
1430 head->hash.rss = rte_le_to_cpu_32(desc->wb.lower.hi_dword.rss);
1431 else if (pkt_flags & PKT_RX_FDIR) {
1432 head->hash.fdir.hash =
1433 rte_le_to_cpu_16(desc->wb.lower.hi_dword.csum_ip.csum)
1434 & IXGBE_ATR_HASH_MASK;
1435 head->hash.fdir.id =
1436 rte_le_to_cpu_16(desc->wb.lower.hi_dword.csum_ip.ip_id);
1441 * ixgbe_recv_pkts_lro - receive handler for and LRO case.
1443 * @rx_queue Rx queue handle
1444 * @rx_pkts table of received packets
1445 * @nb_pkts size of rx_pkts table
1446 * @bulk_alloc if TRUE bulk allocation is used for a HW ring refilling
1448 * Handles the Rx HW ring completions when RSC feature is configured. Uses an
1449 * additional ring of ixgbe_rsc_entry's that will hold the relevant RSC info.
1451 * We use the same logic as in Linux and in FreeBSD ixgbe drivers:
1452 * 1) When non-EOP RSC completion arrives:
1453 * a) Update the HEAD of the current RSC aggregation cluster with the new
1454 * segment's data length.
1455 * b) Set the "next" pointer of the current segment to point to the segment
1456 * at the NEXTP index.
1457 * c) Pass the HEAD of RSC aggregation cluster on to the next NEXTP entry
1458 * in the sw_rsc_ring.
1459 * 2) When EOP arrives we just update the cluster's total length and offload
1460 * flags and deliver the cluster up to the upper layers. In our case - put it
1461 * in the rx_pkts table.
1463 * Returns the number of received packets/clusters (according to the "bulk
1464 * receive" interface).
1466 static inline uint16_t
1467 ixgbe_recv_pkts_lro(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts,
1470 struct ixgbe_rx_queue *rxq = rx_queue;
1471 volatile union ixgbe_adv_rx_desc *rx_ring = rxq->rx_ring;
1472 struct ixgbe_rx_entry *sw_ring = rxq->sw_ring;
1473 struct ixgbe_scattered_rx_entry *sw_sc_ring = rxq->sw_sc_ring;
1474 uint16_t rx_id = rxq->rx_tail;
1476 uint16_t nb_hold = rxq->nb_rx_hold;
1477 uint16_t prev_id = rxq->rx_tail;
1479 while (nb_rx < nb_pkts) {
1481 struct ixgbe_rx_entry *rxe;
1482 struct ixgbe_scattered_rx_entry *sc_entry;
1483 struct ixgbe_scattered_rx_entry *next_sc_entry;
1484 struct ixgbe_rx_entry *next_rxe;
1485 struct rte_mbuf *first_seg;
1486 struct rte_mbuf *rxm;
1487 struct rte_mbuf *nmb;
1488 union ixgbe_adv_rx_desc rxd;
1491 volatile union ixgbe_adv_rx_desc *rxdp;
1496 * The code in this whole file uses the volatile pointer to
1497 * ensure the read ordering of the status and the rest of the
1498 * descriptor fields (on the compiler level only!!!). This is so
1499 * UGLY - why not to just use the compiler barrier instead? DPDK
1500 * even has the rte_compiler_barrier() for that.
1502 * But most importantly this is just wrong because this doesn't
1503 * ensure memory ordering in a general case at all. For
1504 * instance, DPDK is supposed to work on Power CPUs where
1505 * compiler barrier may just not be enough!
1507 * I tried to write only this function properly to have a
1508 * starting point (as a part of an LRO/RSC series) but the
1509 * compiler cursed at me when I tried to cast away the
1510 * "volatile" from rx_ring (yes, it's volatile too!!!). So, I'm
1511 * keeping it the way it is for now.
1513 * The code in this file is broken in so many other places and
1514 * will just not work on a big endian CPU anyway therefore the
1515 * lines below will have to be revisited together with the rest
1519 * - Get rid of "volatile" crap and let the compiler do its
1521 * - Use the proper memory barrier (rte_rmb()) to ensure the
1522 * memory ordering below.
1524 rxdp = &rx_ring[rx_id];
1525 staterr = rte_le_to_cpu_32(rxdp->wb.upper.status_error);
1527 if (!(staterr & IXGBE_RXDADV_STAT_DD))
1532 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
1533 "staterr=0x%x data_len=%u",
1534 rxq->port_id, rxq->queue_id, rx_id, staterr,
1535 rte_le_to_cpu_16(rxd.wb.upper.length));
1538 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
1540 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed "
1541 "port_id=%u queue_id=%u",
1542 rxq->port_id, rxq->queue_id);
1544 rte_eth_devices[rxq->port_id].data->
1545 rx_mbuf_alloc_failed++;
1548 } else if (nb_hold > rxq->rx_free_thresh) {
1549 uint16_t next_rdt = rxq->rx_free_trigger;
1551 if (!ixgbe_rx_alloc_bufs(rxq, false)) {
1553 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr,
1555 nb_hold -= rxq->rx_free_thresh;
1557 PMD_RX_LOG(DEBUG, "RX bulk alloc failed "
1558 "port_id=%u queue_id=%u",
1559 rxq->port_id, rxq->queue_id);
1561 rte_eth_devices[rxq->port_id].data->
1562 rx_mbuf_alloc_failed++;
1568 rxe = &sw_ring[rx_id];
1569 eop = staterr & IXGBE_RXDADV_STAT_EOP;
1571 next_id = rx_id + 1;
1572 if (next_id == rxq->nb_rx_desc)
1575 /* Prefetch next mbuf while processing current one. */
1576 rte_ixgbe_prefetch(sw_ring[next_id].mbuf);
1579 * When next RX descriptor is on a cache-line boundary,
1580 * prefetch the next 4 RX descriptors and the next 4 pointers
1583 if ((next_id & 0x3) == 0) {
1584 rte_ixgbe_prefetch(&rx_ring[next_id]);
1585 rte_ixgbe_prefetch(&sw_ring[next_id]);
1592 rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
1594 * Update RX descriptor with the physical address of the
1595 * new data buffer of the new allocated mbuf.
1599 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1600 rxdp->read.hdr_addr = dma;
1601 rxdp->read.pkt_addr = dma;
1606 * Set data length & data buffer address of mbuf.
1608 data_len = rte_le_to_cpu_16(rxd.wb.upper.length);
1609 rxm->data_len = data_len;
1614 * Get next descriptor index:
1615 * - For RSC it's in the NEXTP field.
1616 * - For a scattered packet - it's just a following
1619 if (ixgbe_rsc_count(&rxd))
1621 (staterr & IXGBE_RXDADV_NEXTP_MASK) >>
1622 IXGBE_RXDADV_NEXTP_SHIFT;
1626 next_sc_entry = &sw_sc_ring[nextp_id];
1627 next_rxe = &sw_ring[nextp_id];
1628 rte_ixgbe_prefetch(next_rxe);
1631 sc_entry = &sw_sc_ring[rx_id];
1632 first_seg = sc_entry->fbuf;
1633 sc_entry->fbuf = NULL;
1636 * If this is the first buffer of the received packet,
1637 * set the pointer to the first mbuf of the packet and
1638 * initialize its context.
1639 * Otherwise, update the total length and the number of segments
1640 * of the current scattered packet, and update the pointer to
1641 * the last mbuf of the current packet.
1643 if (first_seg == NULL) {
1645 first_seg->pkt_len = data_len;
1646 first_seg->nb_segs = 1;
1648 first_seg->pkt_len += data_len;
1649 first_seg->nb_segs++;
1656 * If this is not the last buffer of the received packet, update
1657 * the pointer to the first mbuf at the NEXTP entry in the
1658 * sw_sc_ring and continue to parse the RX ring.
1661 rxm->next = next_rxe->mbuf;
1662 next_sc_entry->fbuf = first_seg;
1667 * This is the last buffer of the received packet - return
1668 * the current cluster to the user.
1672 /* Initialize the first mbuf of the returned packet */
1673 ixgbe_fill_cluster_head_buf(first_seg, &rxd, rxq->port_id,
1676 /* Prefetch data of first segment, if configured to do so. */
1677 rte_packet_prefetch((char *)first_seg->buf_addr +
1678 first_seg->data_off);
1681 * Store the mbuf address into the next entry of the array
1682 * of returned packets.
1684 rx_pkts[nb_rx++] = first_seg;
1688 * Record index of the next RX descriptor to probe.
1690 rxq->rx_tail = rx_id;
1693 * If the number of free RX descriptors is greater than the RX free
1694 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1696 * Update the RDT with the value of the last processed RX descriptor
1697 * minus 1, to guarantee that the RDT register is never equal to the
1698 * RDH register, which creates a "full" ring situtation from the
1699 * hardware point of view...
1701 if (!bulk_alloc && nb_hold > rxq->rx_free_thresh) {
1702 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1703 "nb_hold=%u nb_rx=%u",
1704 rxq->port_id, rxq->queue_id, rx_id, nb_hold, nb_rx);
1707 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, prev_id);
1711 rxq->nb_rx_hold = nb_hold;
1716 ixgbe_recv_pkts_lro_single_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1719 return ixgbe_recv_pkts_lro(rx_queue, rx_pkts, nb_pkts, false);
1723 ixgbe_recv_pkts_lro_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
1726 return ixgbe_recv_pkts_lro(rx_queue, rx_pkts, nb_pkts, true);
1729 /*********************************************************************
1731 * Queue management functions
1733 **********************************************************************/
1736 * Rings setup and release.
1738 * TDBA/RDBA should be aligned on 16 byte boundary. But TDLEN/RDLEN should be
1739 * multiple of 128 bytes. So we align TDBA/RDBA on 128 byte boundary. This will
1740 * also optimize cache line size effect. H/W supports up to cache line size 128.
1742 #define IXGBE_ALIGN 128
1745 * Maximum number of Ring Descriptors.
1747 * Since RDLEN/TDLEN should be multiple of 128 bytes, the number of ring
1748 * descriptors should meet the following condition:
1749 * (num_ring_desc * sizeof(rx/tx descriptor)) % 128 == 0
1751 #define IXGBE_MIN_RING_DESC 32
1752 #define IXGBE_MAX_RING_DESC 4096
1755 * Create memzone for HW rings. malloc can't be used as the physical address is
1756 * needed. If the memzone is already created, then this function returns a ptr
1759 static const struct rte_memzone * __attribute__((cold))
1760 ring_dma_zone_reserve(struct rte_eth_dev *dev, const char *ring_name,
1761 uint16_t queue_id, uint32_t ring_size, int socket_id)
1763 char z_name[RTE_MEMZONE_NAMESIZE];
1764 const struct rte_memzone *mz;
1766 snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
1767 dev->driver->pci_drv.name, ring_name,
1768 dev->data->port_id, queue_id);
1770 mz = rte_memzone_lookup(z_name);
1774 #ifdef RTE_LIBRTE_XEN_DOM0
1775 return rte_memzone_reserve_bounded(z_name, ring_size,
1776 socket_id, 0, IXGBE_ALIGN, RTE_PGSIZE_2M);
1778 return rte_memzone_reserve_aligned(z_name, ring_size,
1779 socket_id, 0, IXGBE_ALIGN);
1783 static void __attribute__((cold))
1784 ixgbe_tx_queue_release_mbufs(struct ixgbe_tx_queue *txq)
1788 if (txq->sw_ring != NULL) {
1789 for (i = 0; i < txq->nb_tx_desc; i++) {
1790 if (txq->sw_ring[i].mbuf != NULL) {
1791 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
1792 txq->sw_ring[i].mbuf = NULL;
1798 static void __attribute__((cold))
1799 ixgbe_tx_free_swring(struct ixgbe_tx_queue *txq)
1802 txq->sw_ring != NULL)
1803 rte_free(txq->sw_ring);
1806 static void __attribute__((cold))
1807 ixgbe_tx_queue_release(struct ixgbe_tx_queue *txq)
1809 if (txq != NULL && txq->ops != NULL) {
1810 txq->ops->release_mbufs(txq);
1811 txq->ops->free_swring(txq);
1816 void __attribute__((cold))
1817 ixgbe_dev_tx_queue_release(void *txq)
1819 ixgbe_tx_queue_release(txq);
1822 /* (Re)set dynamic ixgbe_tx_queue fields to defaults */
1823 static void __attribute__((cold))
1824 ixgbe_reset_tx_queue(struct ixgbe_tx_queue *txq)
1826 static const union ixgbe_adv_tx_desc zeroed_desc = {{0}};
1827 struct ixgbe_tx_entry *txe = txq->sw_ring;
1830 /* Zero out HW ring memory */
1831 for (i = 0; i < txq->nb_tx_desc; i++) {
1832 txq->tx_ring[i] = zeroed_desc;
1835 /* Initialize SW ring entries */
1836 prev = (uint16_t) (txq->nb_tx_desc - 1);
1837 for (i = 0; i < txq->nb_tx_desc; i++) {
1838 volatile union ixgbe_adv_tx_desc *txd = &txq->tx_ring[i];
1839 txd->wb.status = IXGBE_TXD_STAT_DD;
1842 txe[prev].next_id = i;
1846 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1847 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1850 txq->nb_tx_used = 0;
1852 * Always allow 1 descriptor to be un-allocated to avoid
1853 * a H/W race condition
1855 txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
1856 txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
1858 memset((void*)&txq->ctx_cache, 0,
1859 IXGBE_CTX_NUM * sizeof(struct ixgbe_advctx_info));
1862 static const struct ixgbe_txq_ops def_txq_ops = {
1863 .release_mbufs = ixgbe_tx_queue_release_mbufs,
1864 .free_swring = ixgbe_tx_free_swring,
1865 .reset = ixgbe_reset_tx_queue,
1868 /* Takes an ethdev and a queue and sets up the tx function to be used based on
1869 * the queue parameters. Used in tx_queue_setup by primary process and then
1870 * in dev_init by secondary process when attaching to an existing ethdev.
1872 void __attribute__((cold))
1873 ixgbe_set_tx_function(struct rte_eth_dev *dev, struct ixgbe_tx_queue *txq)
1875 /* Use a simple Tx queue (no offloads, no multi segs) if possible */
1876 if (((txq->txq_flags & IXGBE_SIMPLE_FLAGS) == IXGBE_SIMPLE_FLAGS)
1877 && (txq->tx_rs_thresh >= RTE_PMD_IXGBE_TX_MAX_BURST)) {
1878 PMD_INIT_LOG(INFO, "Using simple tx code path");
1879 #ifdef RTE_IXGBE_INC_VECTOR
1880 if (txq->tx_rs_thresh <= RTE_IXGBE_TX_MAX_FREE_BUF_SZ &&
1881 (rte_eal_process_type() != RTE_PROC_PRIMARY ||
1882 ixgbe_txq_vec_setup(txq) == 0)) {
1883 PMD_INIT_LOG(INFO, "Vector tx enabled.");
1884 dev->tx_pkt_burst = ixgbe_xmit_pkts_vec;
1887 dev->tx_pkt_burst = ixgbe_xmit_pkts_simple;
1889 PMD_INIT_LOG(INFO, "Using full-featured tx code path");
1891 " - txq_flags = %lx " "[IXGBE_SIMPLE_FLAGS=%lx]",
1892 (unsigned long)txq->txq_flags,
1893 (unsigned long)IXGBE_SIMPLE_FLAGS);
1895 " - tx_rs_thresh = %lu " "[RTE_PMD_IXGBE_TX_MAX_BURST=%lu]",
1896 (unsigned long)txq->tx_rs_thresh,
1897 (unsigned long)RTE_PMD_IXGBE_TX_MAX_BURST);
1898 dev->tx_pkt_burst = ixgbe_xmit_pkts;
1902 int __attribute__((cold))
1903 ixgbe_dev_tx_queue_setup(struct rte_eth_dev *dev,
1906 unsigned int socket_id,
1907 const struct rte_eth_txconf *tx_conf)
1909 const struct rte_memzone *tz;
1910 struct ixgbe_tx_queue *txq;
1911 struct ixgbe_hw *hw;
1912 uint16_t tx_rs_thresh, tx_free_thresh;
1914 PMD_INIT_FUNC_TRACE();
1915 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1918 * Validate number of transmit descriptors.
1919 * It must not exceed hardware maximum, and must be multiple
1922 if (((nb_desc * sizeof(union ixgbe_adv_tx_desc)) % IXGBE_ALIGN) != 0 ||
1923 (nb_desc > IXGBE_MAX_RING_DESC) ||
1924 (nb_desc < IXGBE_MIN_RING_DESC)) {
1929 * The following two parameters control the setting of the RS bit on
1930 * transmit descriptors.
1931 * TX descriptors will have their RS bit set after txq->tx_rs_thresh
1932 * descriptors have been used.
1933 * The TX descriptor ring will be cleaned after txq->tx_free_thresh
1934 * descriptors are used or if the number of descriptors required
1935 * to transmit a packet is greater than the number of free TX
1937 * The following constraints must be satisfied:
1938 * tx_rs_thresh must be greater than 0.
1939 * tx_rs_thresh must be less than the size of the ring minus 2.
1940 * tx_rs_thresh must be less than or equal to tx_free_thresh.
1941 * tx_rs_thresh must be a divisor of the ring size.
1942 * tx_free_thresh must be greater than 0.
1943 * tx_free_thresh must be less than the size of the ring minus 3.
1944 * One descriptor in the TX ring is used as a sentinel to avoid a
1945 * H/W race condition, hence the maximum threshold constraints.
1946 * When set to zero use default values.
1948 tx_rs_thresh = (uint16_t)((tx_conf->tx_rs_thresh) ?
1949 tx_conf->tx_rs_thresh : DEFAULT_TX_RS_THRESH);
1950 tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
1951 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
1952 if (tx_rs_thresh >= (nb_desc - 2)) {
1953 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the number "
1954 "of TX descriptors minus 2. (tx_rs_thresh=%u "
1955 "port=%d queue=%d)", (unsigned int)tx_rs_thresh,
1956 (int)dev->data->port_id, (int)queue_idx);
1959 if (tx_free_thresh >= (nb_desc - 3)) {
1960 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
1961 "tx_free_thresh must be less than the number of "
1962 "TX descriptors minus 3. (tx_free_thresh=%u "
1963 "port=%d queue=%d)",
1964 (unsigned int)tx_free_thresh,
1965 (int)dev->data->port_id, (int)queue_idx);
1968 if (tx_rs_thresh > tx_free_thresh) {
1969 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or equal to "
1970 "tx_free_thresh. (tx_free_thresh=%u "
1971 "tx_rs_thresh=%u port=%d queue=%d)",
1972 (unsigned int)tx_free_thresh,
1973 (unsigned int)tx_rs_thresh,
1974 (int)dev->data->port_id,
1978 if ((nb_desc % tx_rs_thresh) != 0) {
1979 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
1980 "number of TX descriptors. (tx_rs_thresh=%u "
1981 "port=%d queue=%d)", (unsigned int)tx_rs_thresh,
1982 (int)dev->data->port_id, (int)queue_idx);
1987 * If rs_bit_thresh is greater than 1, then TX WTHRESH should be
1988 * set to 0. If WTHRESH is greater than zero, the RS bit is ignored
1989 * by the NIC and all descriptors are written back after the NIC
1990 * accumulates WTHRESH descriptors.
1992 if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
1993 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
1994 "tx_rs_thresh is greater than 1. (tx_rs_thresh=%u "
1995 "port=%d queue=%d)", (unsigned int)tx_rs_thresh,
1996 (int)dev->data->port_id, (int)queue_idx);
2000 /* Free memory prior to re-allocation if needed... */
2001 if (dev->data->tx_queues[queue_idx] != NULL) {
2002 ixgbe_tx_queue_release(dev->data->tx_queues[queue_idx]);
2003 dev->data->tx_queues[queue_idx] = NULL;
2006 /* First allocate the tx queue data structure */
2007 txq = rte_zmalloc_socket("ethdev TX queue", sizeof(struct ixgbe_tx_queue),
2008 RTE_CACHE_LINE_SIZE, socket_id);
2013 * Allocate TX ring hardware descriptors. A memzone large enough to
2014 * handle the maximum ring size is allocated in order to allow for
2015 * resizing in later calls to the queue setup function.
2017 tz = ring_dma_zone_reserve(dev, "tx_ring", queue_idx,
2018 sizeof(union ixgbe_adv_tx_desc) * IXGBE_MAX_RING_DESC,
2021 ixgbe_tx_queue_release(txq);
2025 txq->nb_tx_desc = nb_desc;
2026 txq->tx_rs_thresh = tx_rs_thresh;
2027 txq->tx_free_thresh = tx_free_thresh;
2028 txq->pthresh = tx_conf->tx_thresh.pthresh;
2029 txq->hthresh = tx_conf->tx_thresh.hthresh;
2030 txq->wthresh = tx_conf->tx_thresh.wthresh;
2031 txq->queue_id = queue_idx;
2032 txq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
2033 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
2034 txq->port_id = dev->data->port_id;
2035 txq->txq_flags = tx_conf->txq_flags;
2036 txq->ops = &def_txq_ops;
2037 txq->tx_deferred_start = tx_conf->tx_deferred_start;
2040 * Modification to set VFTDT for virtual function if vf is detected
2042 if (hw->mac.type == ixgbe_mac_82599_vf ||
2043 hw->mac.type == ixgbe_mac_X540_vf ||
2044 hw->mac.type == ixgbe_mac_X550_vf ||
2045 hw->mac.type == ixgbe_mac_X550EM_x_vf)
2046 txq->tdt_reg_addr = IXGBE_PCI_REG_ADDR(hw, IXGBE_VFTDT(queue_idx));
2048 txq->tdt_reg_addr = IXGBE_PCI_REG_ADDR(hw, IXGBE_TDT(txq->reg_idx));
2049 #ifndef RTE_LIBRTE_XEN_DOM0
2050 txq->tx_ring_phys_addr = (uint64_t) tz->phys_addr;
2052 txq->tx_ring_phys_addr = rte_mem_phy2mch(tz->memseg_id, tz->phys_addr);
2054 txq->tx_ring = (union ixgbe_adv_tx_desc *) tz->addr;
2056 /* Allocate software ring */
2057 txq->sw_ring = rte_zmalloc_socket("txq->sw_ring",
2058 sizeof(struct ixgbe_tx_entry) * nb_desc,
2059 RTE_CACHE_LINE_SIZE, socket_id);
2060 if (txq->sw_ring == NULL) {
2061 ixgbe_tx_queue_release(txq);
2064 PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64,
2065 txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr);
2067 /* set up vector or scalar TX function as appropriate */
2068 ixgbe_set_tx_function(dev, txq);
2070 txq->ops->reset(txq);
2072 dev->data->tx_queues[queue_idx] = txq;
2079 * ixgbe_free_sc_cluster - free the not-yet-completed scattered cluster
2081 * The "next" pointer of the last segment of (not-yet-completed) RSC clusters
2082 * in the sw_rsc_ring is not set to NULL but rather points to the next
2083 * mbuf of this RSC aggregation (that has not been completed yet and still
2084 * resides on the HW ring). So, instead of calling for rte_pktmbuf_free() we
2085 * will just free first "nb_segs" segments of the cluster explicitly by calling
2086 * an rte_pktmbuf_free_seg().
2088 * @m scattered cluster head
2090 static void __attribute__((cold))
2091 ixgbe_free_sc_cluster(struct rte_mbuf *m)
2093 uint8_t i, nb_segs = m->nb_segs;
2094 struct rte_mbuf *next_seg;
2096 for (i = 0; i < nb_segs; i++) {
2098 rte_pktmbuf_free_seg(m);
2103 static void __attribute__((cold))
2104 ixgbe_rx_queue_release_mbufs(struct ixgbe_rx_queue *rxq)
2108 if (rxq->sw_ring != NULL) {
2109 for (i = 0; i < rxq->nb_rx_desc; i++) {
2110 if (rxq->sw_ring[i].mbuf != NULL &&
2111 rte_mbuf_refcnt_read(rxq->sw_ring[i].mbuf)) {
2112 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2113 rxq->sw_ring[i].mbuf = NULL;
2116 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2117 if (rxq->rx_nb_avail) {
2118 for (i = 0; i < rxq->rx_nb_avail; ++i) {
2119 struct rte_mbuf *mb;
2120 mb = rxq->rx_stage[rxq->rx_next_avail + i];
2121 rte_pktmbuf_free_seg(mb);
2123 rxq->rx_nb_avail = 0;
2128 if (rxq->sw_sc_ring)
2129 for (i = 0; i < rxq->nb_rx_desc; i++)
2130 if (rxq->sw_sc_ring[i].fbuf) {
2131 ixgbe_free_sc_cluster(rxq->sw_sc_ring[i].fbuf);
2132 rxq->sw_sc_ring[i].fbuf = NULL;
2136 static void __attribute__((cold))
2137 ixgbe_rx_queue_release(struct ixgbe_rx_queue *rxq)
2140 ixgbe_rx_queue_release_mbufs(rxq);
2141 rte_free(rxq->sw_ring);
2142 rte_free(rxq->sw_sc_ring);
2147 void __attribute__((cold))
2148 ixgbe_dev_rx_queue_release(void *rxq)
2150 ixgbe_rx_queue_release(rxq);
2154 * Check if Rx Burst Bulk Alloc function can be used.
2156 * 0: the preconditions are satisfied and the bulk allocation function
2158 * -EINVAL: the preconditions are NOT satisfied and the default Rx burst
2159 * function must be used.
2161 static inline int __attribute__((cold))
2162 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2163 check_rx_burst_bulk_alloc_preconditions(struct ixgbe_rx_queue *rxq)
2165 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct ixgbe_rx_queue *rxq)
2171 * Make sure the following pre-conditions are satisfied:
2172 * rxq->rx_free_thresh >= RTE_PMD_IXGBE_RX_MAX_BURST
2173 * rxq->rx_free_thresh < rxq->nb_rx_desc
2174 * (rxq->nb_rx_desc % rxq->rx_free_thresh) == 0
2175 * rxq->nb_rx_desc<(IXGBE_MAX_RING_DESC-RTE_PMD_IXGBE_RX_MAX_BURST)
2176 * Scattered packets are not supported. This should be checked
2177 * outside of this function.
2179 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2180 if (!(rxq->rx_free_thresh >= RTE_PMD_IXGBE_RX_MAX_BURST)) {
2181 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2182 "rxq->rx_free_thresh=%d, "
2183 "RTE_PMD_IXGBE_RX_MAX_BURST=%d",
2184 rxq->rx_free_thresh, RTE_PMD_IXGBE_RX_MAX_BURST);
2186 } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
2187 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2188 "rxq->rx_free_thresh=%d, "
2189 "rxq->nb_rx_desc=%d",
2190 rxq->rx_free_thresh, rxq->nb_rx_desc);
2192 } else if (!((rxq->nb_rx_desc % rxq->rx_free_thresh) == 0)) {
2193 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2194 "rxq->nb_rx_desc=%d, "
2195 "rxq->rx_free_thresh=%d",
2196 rxq->nb_rx_desc, rxq->rx_free_thresh);
2198 } else if (!(rxq->nb_rx_desc <
2199 (IXGBE_MAX_RING_DESC - RTE_PMD_IXGBE_RX_MAX_BURST))) {
2200 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
2201 "rxq->nb_rx_desc=%d, "
2202 "IXGBE_MAX_RING_DESC=%d, "
2203 "RTE_PMD_IXGBE_RX_MAX_BURST=%d",
2204 rxq->nb_rx_desc, IXGBE_MAX_RING_DESC,
2205 RTE_PMD_IXGBE_RX_MAX_BURST);
2215 /* Reset dynamic ixgbe_rx_queue fields back to defaults */
2216 static void __attribute__((cold))
2217 ixgbe_reset_rx_queue(struct ixgbe_adapter *adapter, struct ixgbe_rx_queue *rxq)
2219 static const union ixgbe_adv_rx_desc zeroed_desc = {{0}};
2221 uint16_t len = rxq->nb_rx_desc;
2224 * By default, the Rx queue setup function allocates enough memory for
2225 * IXGBE_MAX_RING_DESC. The Rx Burst bulk allocation function requires
2226 * extra memory at the end of the descriptor ring to be zero'd out. A
2227 * pre-condition for using the Rx burst bulk alloc function is that the
2228 * number of descriptors is less than or equal to
2229 * (IXGBE_MAX_RING_DESC - RTE_PMD_IXGBE_RX_MAX_BURST). Check all the
2230 * constraints here to see if we need to zero out memory after the end
2231 * of the H/W descriptor ring.
2233 if (adapter->rx_bulk_alloc_allowed)
2234 /* zero out extra memory */
2235 len += RTE_PMD_IXGBE_RX_MAX_BURST;
2238 * Zero out HW ring memory. Zero out extra memory at the end of
2239 * the H/W ring so look-ahead logic in Rx Burst bulk alloc function
2240 * reads extra memory as zeros.
2242 for (i = 0; i < len; i++) {
2243 rxq->rx_ring[i] = zeroed_desc;
2246 #ifdef RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC
2248 * initialize extra software ring entries. Space for these extra
2249 * entries is always allocated
2251 memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2252 for (i = rxq->nb_rx_desc; i < len; ++i) {
2253 rxq->sw_ring[i].mbuf = &rxq->fake_mbuf;
2256 rxq->rx_nb_avail = 0;
2257 rxq->rx_next_avail = 0;
2258 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2259 #endif /* RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC */
2261 rxq->nb_rx_hold = 0;
2262 rxq->pkt_first_seg = NULL;
2263 rxq->pkt_last_seg = NULL;
2266 int __attribute__((cold))
2267 ixgbe_dev_rx_queue_setup(struct rte_eth_dev *dev,
2270 unsigned int socket_id,
2271 const struct rte_eth_rxconf *rx_conf,
2272 struct rte_mempool *mp)
2274 const struct rte_memzone *rz;
2275 struct ixgbe_rx_queue *rxq;
2276 struct ixgbe_hw *hw;
2278 struct ixgbe_adapter *adapter =
2279 (struct ixgbe_adapter *)dev->data->dev_private;
2281 PMD_INIT_FUNC_TRACE();
2282 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2285 * Validate number of receive descriptors.
2286 * It must not exceed hardware maximum, and must be multiple
2289 if (((nb_desc * sizeof(union ixgbe_adv_rx_desc)) % IXGBE_ALIGN) != 0 ||
2290 (nb_desc > IXGBE_MAX_RING_DESC) ||
2291 (nb_desc < IXGBE_MIN_RING_DESC)) {
2295 /* Free memory prior to re-allocation if needed... */
2296 if (dev->data->rx_queues[queue_idx] != NULL) {
2297 ixgbe_rx_queue_release(dev->data->rx_queues[queue_idx]);
2298 dev->data->rx_queues[queue_idx] = NULL;
2301 /* First allocate the rx queue data structure */
2302 rxq = rte_zmalloc_socket("ethdev RX queue", sizeof(struct ixgbe_rx_queue),
2303 RTE_CACHE_LINE_SIZE, socket_id);
2307 rxq->nb_rx_desc = nb_desc;
2308 rxq->rx_free_thresh = rx_conf->rx_free_thresh;
2309 rxq->queue_id = queue_idx;
2310 rxq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
2311 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
2312 rxq->port_id = dev->data->port_id;
2313 rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ?
2315 rxq->drop_en = rx_conf->rx_drop_en;
2316 rxq->rx_deferred_start = rx_conf->rx_deferred_start;
2319 * Allocate RX ring hardware descriptors. A memzone large enough to
2320 * handle the maximum ring size is allocated in order to allow for
2321 * resizing in later calls to the queue setup function.
2323 rz = ring_dma_zone_reserve(dev, "rx_ring", queue_idx,
2324 RX_RING_SZ, socket_id);
2326 ixgbe_rx_queue_release(rxq);
2331 * Zero init all the descriptors in the ring.
2333 memset (rz->addr, 0, RX_RING_SZ);
2336 * Modified to setup VFRDT for Virtual Function
2338 if (hw->mac.type == ixgbe_mac_82599_vf ||
2339 hw->mac.type == ixgbe_mac_X540_vf ||
2340 hw->mac.type == ixgbe_mac_X550_vf ||
2341 hw->mac.type == ixgbe_mac_X550EM_x_vf) {
2343 IXGBE_PCI_REG_ADDR(hw, IXGBE_VFRDT(queue_idx));
2345 IXGBE_PCI_REG_ADDR(hw, IXGBE_VFRDH(queue_idx));
2349 IXGBE_PCI_REG_ADDR(hw, IXGBE_RDT(rxq->reg_idx));
2351 IXGBE_PCI_REG_ADDR(hw, IXGBE_RDH(rxq->reg_idx));
2353 #ifndef RTE_LIBRTE_XEN_DOM0
2354 rxq->rx_ring_phys_addr = (uint64_t) rz->phys_addr;
2356 rxq->rx_ring_phys_addr = rte_mem_phy2mch(rz->memseg_id, rz->phys_addr);
2358 rxq->rx_ring = (union ixgbe_adv_rx_desc *) rz->addr;
2361 * Certain constraints must be met in order to use the bulk buffer
2362 * allocation Rx burst function. If any of Rx queues doesn't meet them
2363 * the feature should be disabled for the whole port.
2365 if (check_rx_burst_bulk_alloc_preconditions(rxq)) {
2366 PMD_INIT_LOG(DEBUG, "queue[%d] doesn't meet Rx Bulk Alloc "
2367 "preconditions - canceling the feature for "
2368 "the whole port[%d]",
2369 rxq->queue_id, rxq->port_id);
2370 adapter->rx_bulk_alloc_allowed = false;
2374 * Allocate software ring. Allow for space at the end of the
2375 * S/W ring to make sure look-ahead logic in bulk alloc Rx burst
2376 * function does not access an invalid memory region.
2379 if (adapter->rx_bulk_alloc_allowed)
2380 len += RTE_PMD_IXGBE_RX_MAX_BURST;
2382 rxq->sw_ring = rte_zmalloc_socket("rxq->sw_ring",
2383 sizeof(struct ixgbe_rx_entry) * len,
2384 RTE_CACHE_LINE_SIZE, socket_id);
2385 if (!rxq->sw_ring) {
2386 ixgbe_rx_queue_release(rxq);
2391 * Always allocate even if it's not going to be needed in order to
2392 * simplify the code.
2394 * This ring is used in LRO and Scattered Rx cases and Scattered Rx may
2395 * be requested in ixgbe_dev_rx_init(), which is called later from
2399 rte_zmalloc_socket("rxq->sw_sc_ring",
2400 sizeof(struct ixgbe_scattered_rx_entry) * len,
2401 RTE_CACHE_LINE_SIZE, socket_id);
2402 if (!rxq->sw_sc_ring) {
2403 ixgbe_rx_queue_release(rxq);
2407 PMD_INIT_LOG(DEBUG, "sw_ring=%p sw_sc_ring=%p hw_ring=%p "
2408 "dma_addr=0x%"PRIx64,
2409 rxq->sw_ring, rxq->sw_sc_ring, rxq->rx_ring,
2410 rxq->rx_ring_phys_addr);
2412 if (!rte_is_power_of_2(nb_desc)) {
2413 PMD_INIT_LOG(DEBUG, "queue[%d] doesn't meet Vector Rx "
2414 "preconditions - canceling the feature for "
2415 "the whole port[%d]",
2416 rxq->queue_id, rxq->port_id);
2417 adapter->rx_vec_allowed = false;
2419 ixgbe_rxq_vec_setup(rxq);
2421 dev->data->rx_queues[queue_idx] = rxq;
2423 ixgbe_reset_rx_queue(adapter, rxq);
2429 ixgbe_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
2431 #define IXGBE_RXQ_SCAN_INTERVAL 4
2432 volatile union ixgbe_adv_rx_desc *rxdp;
2433 struct ixgbe_rx_queue *rxq;
2436 if (rx_queue_id >= dev->data->nb_rx_queues) {
2437 PMD_RX_LOG(ERR, "Invalid RX queue id=%d", rx_queue_id);
2441 rxq = dev->data->rx_queues[rx_queue_id];
2442 rxdp = &(rxq->rx_ring[rxq->rx_tail]);
2444 while ((desc < rxq->nb_rx_desc) &&
2445 (rxdp->wb.upper.status_error & IXGBE_RXDADV_STAT_DD)) {
2446 desc += IXGBE_RXQ_SCAN_INTERVAL;
2447 rxdp += IXGBE_RXQ_SCAN_INTERVAL;
2448 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
2449 rxdp = &(rxq->rx_ring[rxq->rx_tail +
2450 desc - rxq->nb_rx_desc]);
2457 ixgbe_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
2459 volatile union ixgbe_adv_rx_desc *rxdp;
2460 struct ixgbe_rx_queue *rxq = rx_queue;
2463 if (unlikely(offset >= rxq->nb_rx_desc))
2465 desc = rxq->rx_tail + offset;
2466 if (desc >= rxq->nb_rx_desc)
2467 desc -= rxq->nb_rx_desc;
2469 rxdp = &rxq->rx_ring[desc];
2470 return !!(rxdp->wb.upper.status_error & IXGBE_RXDADV_STAT_DD);
2473 void __attribute__((cold))
2474 ixgbe_dev_clear_queues(struct rte_eth_dev *dev)
2477 struct ixgbe_adapter *adapter =
2478 (struct ixgbe_adapter *)dev->data->dev_private;
2480 PMD_INIT_FUNC_TRACE();
2482 for (i = 0; i < dev->data->nb_tx_queues; i++) {
2483 struct ixgbe_tx_queue *txq = dev->data->tx_queues[i];
2485 txq->ops->release_mbufs(txq);
2486 txq->ops->reset(txq);
2490 for (i = 0; i < dev->data->nb_rx_queues; i++) {
2491 struct ixgbe_rx_queue *rxq = dev->data->rx_queues[i];
2493 ixgbe_rx_queue_release_mbufs(rxq);
2494 ixgbe_reset_rx_queue(adapter, rxq);
2499 /*********************************************************************
2501 * Device RX/TX init functions
2503 **********************************************************************/
2506 * Receive Side Scaling (RSS)
2507 * See section 7.1.2.8 in the following document:
2508 * "Intel 82599 10 GbE Controller Datasheet" - Revision 2.1 October 2009
2511 * The source and destination IP addresses of the IP header and the source
2512 * and destination ports of TCP/UDP headers, if any, of received packets are
2513 * hashed against a configurable random key to compute a 32-bit RSS hash result.
2514 * The seven (7) LSBs of the 32-bit hash result are used as an index into a
2515 * 128-entry redirection table (RETA). Each entry of the RETA provides a 3-bit
2516 * RSS output index which is used as the RX queue index where to store the
2518 * The following output is supplied in the RX write-back descriptor:
2519 * - 32-bit result of the Microsoft RSS hash function,
2520 * - 4-bit RSS type field.
2524 * RSS random key supplied in section 7.1.2.8.3 of the Intel 82599 datasheet.
2525 * Used as the default key.
2527 static uint8_t rss_intel_key[40] = {
2528 0x6D, 0x5A, 0x56, 0xDA, 0x25, 0x5B, 0x0E, 0xC2,
2529 0x41, 0x67, 0x25, 0x3D, 0x43, 0xA3, 0x8F, 0xB0,
2530 0xD0, 0xCA, 0x2B, 0xCB, 0xAE, 0x7B, 0x30, 0xB4,
2531 0x77, 0xCB, 0x2D, 0xA3, 0x80, 0x30, 0xF2, 0x0C,
2532 0x6A, 0x42, 0xB7, 0x3B, 0xBE, 0xAC, 0x01, 0xFA,
2536 ixgbe_rss_disable(struct rte_eth_dev *dev)
2538 struct ixgbe_hw *hw;
2541 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2542 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
2543 mrqc &= ~IXGBE_MRQC_RSSEN;
2544 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
2548 ixgbe_hw_rss_hash_set(struct ixgbe_hw *hw, struct rte_eth_rss_conf *rss_conf)
2556 hash_key = rss_conf->rss_key;
2557 if (hash_key != NULL) {
2558 /* Fill in RSS hash key */
2559 for (i = 0; i < 10; i++) {
2560 rss_key = hash_key[(i * 4)];
2561 rss_key |= hash_key[(i * 4) + 1] << 8;
2562 rss_key |= hash_key[(i * 4) + 2] << 16;
2563 rss_key |= hash_key[(i * 4) + 3] << 24;
2564 IXGBE_WRITE_REG_ARRAY(hw, IXGBE_RSSRK(0), i, rss_key);
2568 /* Set configured hashing protocols in MRQC register */
2569 rss_hf = rss_conf->rss_hf;
2570 mrqc = IXGBE_MRQC_RSSEN; /* Enable RSS */
2571 if (rss_hf & ETH_RSS_IPV4)
2572 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV4;
2573 if (rss_hf & ETH_RSS_NONFRAG_IPV4_TCP)
2574 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV4_TCP;
2575 if (rss_hf & ETH_RSS_IPV6)
2576 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6;
2577 if (rss_hf & ETH_RSS_IPV6_EX)
2578 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_EX;
2579 if (rss_hf & ETH_RSS_NONFRAG_IPV6_TCP)
2580 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_TCP;
2581 if (rss_hf & ETH_RSS_IPV6_TCP_EX)
2582 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_EX_TCP;
2583 if (rss_hf & ETH_RSS_NONFRAG_IPV4_UDP)
2584 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV4_UDP;
2585 if (rss_hf & ETH_RSS_NONFRAG_IPV6_UDP)
2586 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_UDP;
2587 if (rss_hf & ETH_RSS_IPV6_UDP_EX)
2588 mrqc |= IXGBE_MRQC_RSS_FIELD_IPV6_EX_UDP;
2589 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
2593 ixgbe_dev_rss_hash_update(struct rte_eth_dev *dev,
2594 struct rte_eth_rss_conf *rss_conf)
2596 struct ixgbe_hw *hw;
2600 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2603 * Excerpt from section 7.1.2.8 Receive-Side Scaling (RSS):
2604 * "RSS enabling cannot be done dynamically while it must be
2605 * preceded by a software reset"
2606 * Before changing anything, first check that the update RSS operation
2607 * does not attempt to disable RSS, if RSS was enabled at
2608 * initialization time, or does not attempt to enable RSS, if RSS was
2609 * disabled at initialization time.
2611 rss_hf = rss_conf->rss_hf & IXGBE_RSS_OFFLOAD_ALL;
2612 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
2613 if (!(mrqc & IXGBE_MRQC_RSSEN)) { /* RSS disabled */
2614 if (rss_hf != 0) /* Enable RSS */
2616 return 0; /* Nothing to do */
2619 if (rss_hf == 0) /* Disable RSS */
2621 ixgbe_hw_rss_hash_set(hw, rss_conf);
2626 ixgbe_dev_rss_hash_conf_get(struct rte_eth_dev *dev,
2627 struct rte_eth_rss_conf *rss_conf)
2629 struct ixgbe_hw *hw;
2636 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2637 hash_key = rss_conf->rss_key;
2638 if (hash_key != NULL) {
2639 /* Return RSS hash key */
2640 for (i = 0; i < 10; i++) {
2641 rss_key = IXGBE_READ_REG_ARRAY(hw, IXGBE_RSSRK(0), i);
2642 hash_key[(i * 4)] = rss_key & 0x000000FF;
2643 hash_key[(i * 4) + 1] = (rss_key >> 8) & 0x000000FF;
2644 hash_key[(i * 4) + 2] = (rss_key >> 16) & 0x000000FF;
2645 hash_key[(i * 4) + 3] = (rss_key >> 24) & 0x000000FF;
2649 /* Get RSS functions configured in MRQC register */
2650 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
2651 if ((mrqc & IXGBE_MRQC_RSSEN) == 0) { /* RSS is disabled */
2652 rss_conf->rss_hf = 0;
2656 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV4)
2657 rss_hf |= ETH_RSS_IPV4;
2658 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV4_TCP)
2659 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
2660 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6)
2661 rss_hf |= ETH_RSS_IPV6;
2662 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_EX)
2663 rss_hf |= ETH_RSS_IPV6_EX;
2664 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_TCP)
2665 rss_hf |= ETH_RSS_NONFRAG_IPV6_TCP;
2666 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_EX_TCP)
2667 rss_hf |= ETH_RSS_IPV6_TCP_EX;
2668 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV4_UDP)
2669 rss_hf |= ETH_RSS_NONFRAG_IPV4_UDP;
2670 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_UDP)
2671 rss_hf |= ETH_RSS_NONFRAG_IPV6_UDP;
2672 if (mrqc & IXGBE_MRQC_RSS_FIELD_IPV6_EX_UDP)
2673 rss_hf |= ETH_RSS_IPV6_UDP_EX;
2674 rss_conf->rss_hf = rss_hf;
2679 ixgbe_rss_configure(struct rte_eth_dev *dev)
2681 struct rte_eth_rss_conf rss_conf;
2682 struct ixgbe_hw *hw;
2687 PMD_INIT_FUNC_TRACE();
2688 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2691 * Fill in redirection table
2692 * The byte-swap is needed because NIC registers are in
2693 * little-endian order.
2696 for (i = 0, j = 0; i < 128; i++, j++) {
2697 if (j == dev->data->nb_rx_queues)
2699 reta = (reta << 8) | j;
2701 IXGBE_WRITE_REG(hw, IXGBE_RETA(i >> 2),
2706 * Configure the RSS key and the RSS protocols used to compute
2707 * the RSS hash of input packets.
2709 rss_conf = dev->data->dev_conf.rx_adv_conf.rss_conf;
2710 if ((rss_conf.rss_hf & IXGBE_RSS_OFFLOAD_ALL) == 0) {
2711 ixgbe_rss_disable(dev);
2714 if (rss_conf.rss_key == NULL)
2715 rss_conf.rss_key = rss_intel_key; /* Default hash key */
2716 ixgbe_hw_rss_hash_set(hw, &rss_conf);
2719 #define NUM_VFTA_REGISTERS 128
2720 #define NIC_RX_BUFFER_SIZE 0x200
2723 ixgbe_vmdq_dcb_configure(struct rte_eth_dev *dev)
2725 struct rte_eth_vmdq_dcb_conf *cfg;
2726 struct ixgbe_hw *hw;
2727 enum rte_eth_nb_pools num_pools;
2728 uint32_t mrqc, vt_ctl, queue_mapping, vlanctrl;
2730 uint8_t nb_tcs; /* number of traffic classes */
2733 PMD_INIT_FUNC_TRACE();
2734 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2735 cfg = &dev->data->dev_conf.rx_adv_conf.vmdq_dcb_conf;
2736 num_pools = cfg->nb_queue_pools;
2737 /* Check we have a valid number of pools */
2738 if (num_pools != ETH_16_POOLS && num_pools != ETH_32_POOLS) {
2739 ixgbe_rss_disable(dev);
2742 /* 16 pools -> 8 traffic classes, 32 pools -> 4 traffic classes */
2743 nb_tcs = (uint8_t)(ETH_VMDQ_DCB_NUM_QUEUES / (int)num_pools);
2747 * split rx buffer up into sections, each for 1 traffic class
2749 pbsize = (uint16_t)(NIC_RX_BUFFER_SIZE / nb_tcs);
2750 for (i = 0 ; i < nb_tcs; i++) {
2751 uint32_t rxpbsize = IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(i));
2752 rxpbsize &= (~(0x3FF << IXGBE_RXPBSIZE_SHIFT));
2753 /* clear 10 bits. */
2754 rxpbsize |= (pbsize << IXGBE_RXPBSIZE_SHIFT); /* set value */
2755 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), rxpbsize);
2757 /* zero alloc all unused TCs */
2758 for (i = nb_tcs; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2759 uint32_t rxpbsize = IXGBE_READ_REG(hw, IXGBE_RXPBSIZE(i));
2760 rxpbsize &= (~( 0x3FF << IXGBE_RXPBSIZE_SHIFT ));
2761 /* clear 10 bits. */
2762 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), rxpbsize);
2765 /* MRQC: enable vmdq and dcb */
2766 mrqc = ((num_pools == ETH_16_POOLS) ? \
2767 IXGBE_MRQC_VMDQRT8TCEN : IXGBE_MRQC_VMDQRT4TCEN );
2768 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
2770 /* PFVTCTL: turn on virtualisation and set the default pool */
2771 vt_ctl = IXGBE_VT_CTL_VT_ENABLE | IXGBE_VT_CTL_REPLEN;
2772 if (cfg->enable_default_pool) {
2773 vt_ctl |= (cfg->default_pool << IXGBE_VT_CTL_POOL_SHIFT);
2775 vt_ctl |= IXGBE_VT_CTL_DIS_DEFPL;
2778 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, vt_ctl);
2780 /* RTRUP2TC: mapping user priorities to traffic classes (TCs) */
2782 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++)
2784 * mapping is done with 3 bits per priority,
2785 * so shift by i*3 each time
2787 queue_mapping |= ((cfg->dcb_queue[i] & 0x07) << (i * 3));
2789 IXGBE_WRITE_REG(hw, IXGBE_RTRUP2TC, queue_mapping);
2791 /* RTRPCS: DCB related */
2792 IXGBE_WRITE_REG(hw, IXGBE_RTRPCS, IXGBE_RMCS_RRM);
2794 /* VLNCTRL: enable vlan filtering and allow all vlan tags through */
2795 vlanctrl = IXGBE_READ_REG(hw, IXGBE_VLNCTRL);
2796 vlanctrl |= IXGBE_VLNCTRL_VFE ; /* enable vlan filters */
2797 IXGBE_WRITE_REG(hw, IXGBE_VLNCTRL, vlanctrl);
2799 /* VFTA - enable all vlan filters */
2800 for (i = 0; i < NUM_VFTA_REGISTERS; i++) {
2801 IXGBE_WRITE_REG(hw, IXGBE_VFTA(i), 0xFFFFFFFF);
2804 /* VFRE: pool enabling for receive - 16 or 32 */
2805 IXGBE_WRITE_REG(hw, IXGBE_VFRE(0), \
2806 num_pools == ETH_16_POOLS ? 0xFFFF : 0xFFFFFFFF);
2809 * MPSAR - allow pools to read specific mac addresses
2810 * In this case, all pools should be able to read from mac addr 0
2812 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_LO(0), 0xFFFFFFFF);
2813 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_HI(0), 0xFFFFFFFF);
2815 /* PFVLVF, PFVLVFB: set up filters for vlan tags as configured */
2816 for (i = 0; i < cfg->nb_pool_maps; i++) {
2817 /* set vlan id in VF register and set the valid bit */
2818 IXGBE_WRITE_REG(hw, IXGBE_VLVF(i), (IXGBE_VLVF_VIEN | \
2819 (cfg->pool_map[i].vlan_id & 0xFFF)));
2821 * Put the allowed pools in VFB reg. As we only have 16 or 32
2822 * pools, we only need to use the first half of the register
2825 IXGBE_WRITE_REG(hw, IXGBE_VLVFB(i*2), cfg->pool_map[i].pools);
2830 * ixgbe_dcb_config_tx_hw_config - Configure general DCB TX parameters
2831 * @hw: pointer to hardware structure
2832 * @dcb_config: pointer to ixgbe_dcb_config structure
2835 ixgbe_dcb_tx_hw_config(struct ixgbe_hw *hw,
2836 struct ixgbe_dcb_config *dcb_config)
2841 PMD_INIT_FUNC_TRACE();
2842 if (hw->mac.type != ixgbe_mac_82598EB) {
2843 /* Disable the Tx desc arbiter so that MTQC can be changed */
2844 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
2845 reg |= IXGBE_RTTDCS_ARBDIS;
2846 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
2848 /* Enable DCB for Tx with 8 TCs */
2849 if (dcb_config->num_tcs.pg_tcs == 8) {
2850 reg = IXGBE_MTQC_RT_ENA | IXGBE_MTQC_8TC_8TQ;
2853 reg = IXGBE_MTQC_RT_ENA | IXGBE_MTQC_4TC_4TQ;
2855 if (dcb_config->vt_mode)
2856 reg |= IXGBE_MTQC_VT_ENA;
2857 IXGBE_WRITE_REG(hw, IXGBE_MTQC, reg);
2859 /* Disable drop for all queues */
2860 for (q = 0; q < 128; q++)
2861 IXGBE_WRITE_REG(hw, IXGBE_QDE,
2862 (IXGBE_QDE_WRITE | (q << IXGBE_QDE_IDX_SHIFT)));
2864 /* Enable the Tx desc arbiter */
2865 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
2866 reg &= ~IXGBE_RTTDCS_ARBDIS;
2867 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
2869 /* Enable Security TX Buffer IFG for DCB */
2870 reg = IXGBE_READ_REG(hw, IXGBE_SECTXMINIFG);
2871 reg |= IXGBE_SECTX_DCB;
2872 IXGBE_WRITE_REG(hw, IXGBE_SECTXMINIFG, reg);
2878 * ixgbe_vmdq_dcb_hw_tx_config - Configure general VMDQ+DCB TX parameters
2879 * @dev: pointer to rte_eth_dev structure
2880 * @dcb_config: pointer to ixgbe_dcb_config structure
2883 ixgbe_vmdq_dcb_hw_tx_config(struct rte_eth_dev *dev,
2884 struct ixgbe_dcb_config *dcb_config)
2886 struct rte_eth_vmdq_dcb_tx_conf *vmdq_tx_conf =
2887 &dev->data->dev_conf.tx_adv_conf.vmdq_dcb_tx_conf;
2888 struct ixgbe_hw *hw =
2889 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2891 PMD_INIT_FUNC_TRACE();
2892 if (hw->mac.type != ixgbe_mac_82598EB)
2893 /*PF VF Transmit Enable*/
2894 IXGBE_WRITE_REG(hw, IXGBE_VFTE(0),
2895 vmdq_tx_conf->nb_queue_pools == ETH_16_POOLS ? 0xFFFF : 0xFFFFFFFF);
2897 /*Configure general DCB TX parameters*/
2898 ixgbe_dcb_tx_hw_config(hw,dcb_config);
2903 ixgbe_vmdq_dcb_rx_config(struct rte_eth_dev *dev,
2904 struct ixgbe_dcb_config *dcb_config)
2906 struct rte_eth_vmdq_dcb_conf *vmdq_rx_conf =
2907 &dev->data->dev_conf.rx_adv_conf.vmdq_dcb_conf;
2908 struct ixgbe_dcb_tc_config *tc;
2911 /* convert rte_eth_conf.rx_adv_conf to struct ixgbe_dcb_config */
2912 if (vmdq_rx_conf->nb_queue_pools == ETH_16_POOLS ) {
2913 dcb_config->num_tcs.pg_tcs = ETH_8_TCS;
2914 dcb_config->num_tcs.pfc_tcs = ETH_8_TCS;
2917 dcb_config->num_tcs.pg_tcs = ETH_4_TCS;
2918 dcb_config->num_tcs.pfc_tcs = ETH_4_TCS;
2920 /* User Priority to Traffic Class mapping */
2921 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2922 j = vmdq_rx_conf->dcb_queue[i];
2923 tc = &dcb_config->tc_config[j];
2924 tc->path[IXGBE_DCB_RX_CONFIG].up_to_tc_bitmap =
2930 ixgbe_dcb_vt_tx_config(struct rte_eth_dev *dev,
2931 struct ixgbe_dcb_config *dcb_config)
2933 struct rte_eth_vmdq_dcb_tx_conf *vmdq_tx_conf =
2934 &dev->data->dev_conf.tx_adv_conf.vmdq_dcb_tx_conf;
2935 struct ixgbe_dcb_tc_config *tc;
2938 /* convert rte_eth_conf.rx_adv_conf to struct ixgbe_dcb_config */
2939 if (vmdq_tx_conf->nb_queue_pools == ETH_16_POOLS ) {
2940 dcb_config->num_tcs.pg_tcs = ETH_8_TCS;
2941 dcb_config->num_tcs.pfc_tcs = ETH_8_TCS;
2944 dcb_config->num_tcs.pg_tcs = ETH_4_TCS;
2945 dcb_config->num_tcs.pfc_tcs = ETH_4_TCS;
2948 /* User Priority to Traffic Class mapping */
2949 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2950 j = vmdq_tx_conf->dcb_queue[i];
2951 tc = &dcb_config->tc_config[j];
2952 tc->path[IXGBE_DCB_TX_CONFIG].up_to_tc_bitmap =
2959 ixgbe_dcb_rx_config(struct rte_eth_dev *dev,
2960 struct ixgbe_dcb_config *dcb_config)
2962 struct rte_eth_dcb_rx_conf *rx_conf =
2963 &dev->data->dev_conf.rx_adv_conf.dcb_rx_conf;
2964 struct ixgbe_dcb_tc_config *tc;
2967 dcb_config->num_tcs.pg_tcs = (uint8_t)rx_conf->nb_tcs;
2968 dcb_config->num_tcs.pfc_tcs = (uint8_t)rx_conf->nb_tcs;
2970 /* User Priority to Traffic Class mapping */
2971 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2972 j = rx_conf->dcb_queue[i];
2973 tc = &dcb_config->tc_config[j];
2974 tc->path[IXGBE_DCB_RX_CONFIG].up_to_tc_bitmap =
2980 ixgbe_dcb_tx_config(struct rte_eth_dev *dev,
2981 struct ixgbe_dcb_config *dcb_config)
2983 struct rte_eth_dcb_tx_conf *tx_conf =
2984 &dev->data->dev_conf.tx_adv_conf.dcb_tx_conf;
2985 struct ixgbe_dcb_tc_config *tc;
2988 dcb_config->num_tcs.pg_tcs = (uint8_t)tx_conf->nb_tcs;
2989 dcb_config->num_tcs.pfc_tcs = (uint8_t)tx_conf->nb_tcs;
2991 /* User Priority to Traffic Class mapping */
2992 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
2993 j = tx_conf->dcb_queue[i];
2994 tc = &dcb_config->tc_config[j];
2995 tc->path[IXGBE_DCB_TX_CONFIG].up_to_tc_bitmap =
3001 * ixgbe_dcb_rx_hw_config - Configure general DCB RX HW parameters
3002 * @hw: pointer to hardware structure
3003 * @dcb_config: pointer to ixgbe_dcb_config structure
3006 ixgbe_dcb_rx_hw_config(struct ixgbe_hw *hw,
3007 struct ixgbe_dcb_config *dcb_config)
3013 PMD_INIT_FUNC_TRACE();
3015 * Disable the arbiter before changing parameters
3016 * (always enable recycle mode; WSP)
3018 reg = IXGBE_RTRPCS_RRM | IXGBE_RTRPCS_RAC | IXGBE_RTRPCS_ARBDIS;
3019 IXGBE_WRITE_REG(hw, IXGBE_RTRPCS, reg);
3021 if (hw->mac.type != ixgbe_mac_82598EB) {
3022 reg = IXGBE_READ_REG(hw, IXGBE_MRQC);
3023 if (dcb_config->num_tcs.pg_tcs == 4) {
3024 if (dcb_config->vt_mode)
3025 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3026 IXGBE_MRQC_VMDQRT4TCEN;
3028 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, 0);
3029 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3033 if (dcb_config->num_tcs.pg_tcs == 8) {
3034 if (dcb_config->vt_mode)
3035 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3036 IXGBE_MRQC_VMDQRT8TCEN;
3038 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, 0);
3039 reg = (reg & ~IXGBE_MRQC_MRQE_MASK) |
3044 IXGBE_WRITE_REG(hw, IXGBE_MRQC, reg);
3047 /* VLNCTRL: enable vlan filtering and allow all vlan tags through */
3048 vlanctrl = IXGBE_READ_REG(hw, IXGBE_VLNCTRL);
3049 vlanctrl |= IXGBE_VLNCTRL_VFE ; /* enable vlan filters */
3050 IXGBE_WRITE_REG(hw, IXGBE_VLNCTRL, vlanctrl);
3052 /* VFTA - enable all vlan filters */
3053 for (i = 0; i < NUM_VFTA_REGISTERS; i++) {
3054 IXGBE_WRITE_REG(hw, IXGBE_VFTA(i), 0xFFFFFFFF);
3058 * Configure Rx packet plane (recycle mode; WSP) and
3061 reg = IXGBE_RTRPCS_RRM | IXGBE_RTRPCS_RAC;
3062 IXGBE_WRITE_REG(hw, IXGBE_RTRPCS, reg);
3068 ixgbe_dcb_hw_arbite_rx_config(struct ixgbe_hw *hw, uint16_t *refill,
3069 uint16_t *max,uint8_t *bwg_id, uint8_t *tsa, uint8_t *map)
3071 switch (hw->mac.type) {
3072 case ixgbe_mac_82598EB:
3073 ixgbe_dcb_config_rx_arbiter_82598(hw, refill, max, tsa);
3075 case ixgbe_mac_82599EB:
3076 case ixgbe_mac_X540:
3077 case ixgbe_mac_X550:
3078 case ixgbe_mac_X550EM_x:
3079 ixgbe_dcb_config_rx_arbiter_82599(hw, refill, max, bwg_id,
3088 ixgbe_dcb_hw_arbite_tx_config(struct ixgbe_hw *hw, uint16_t *refill, uint16_t *max,
3089 uint8_t *bwg_id, uint8_t *tsa, uint8_t *map)
3091 switch (hw->mac.type) {
3092 case ixgbe_mac_82598EB:
3093 ixgbe_dcb_config_tx_desc_arbiter_82598(hw, refill, max, bwg_id,tsa);
3094 ixgbe_dcb_config_tx_data_arbiter_82598(hw, refill, max, bwg_id,tsa);
3096 case ixgbe_mac_82599EB:
3097 case ixgbe_mac_X540:
3098 case ixgbe_mac_X550:
3099 case ixgbe_mac_X550EM_x:
3100 ixgbe_dcb_config_tx_desc_arbiter_82599(hw, refill, max, bwg_id,tsa);
3101 ixgbe_dcb_config_tx_data_arbiter_82599(hw, refill, max, bwg_id,tsa, map);
3108 #define DCB_RX_CONFIG 1
3109 #define DCB_TX_CONFIG 1
3110 #define DCB_TX_PB 1024
3112 * ixgbe_dcb_hw_configure - Enable DCB and configure
3113 * general DCB in VT mode and non-VT mode parameters
3114 * @dev: pointer to rte_eth_dev structure
3115 * @dcb_config: pointer to ixgbe_dcb_config structure
3118 ixgbe_dcb_hw_configure(struct rte_eth_dev *dev,
3119 struct ixgbe_dcb_config *dcb_config)
3122 uint8_t i,pfc_en,nb_tcs;
3124 uint8_t config_dcb_rx = 0;
3125 uint8_t config_dcb_tx = 0;
3126 uint8_t tsa[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3127 uint8_t bwgid[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3128 uint16_t refill[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3129 uint16_t max[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3130 uint8_t map[IXGBE_DCB_MAX_TRAFFIC_CLASS] = {0};
3131 struct ixgbe_dcb_tc_config *tc;
3132 uint32_t max_frame = dev->data->mtu + ETHER_HDR_LEN + ETHER_CRC_LEN;
3133 struct ixgbe_hw *hw =
3134 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3136 switch(dev->data->dev_conf.rxmode.mq_mode){
3137 case ETH_MQ_RX_VMDQ_DCB:
3138 dcb_config->vt_mode = true;
3139 if (hw->mac.type != ixgbe_mac_82598EB) {
3140 config_dcb_rx = DCB_RX_CONFIG;
3142 *get dcb and VT rx configuration parameters
3145 ixgbe_vmdq_dcb_rx_config(dev,dcb_config);
3146 /*Configure general VMDQ and DCB RX parameters*/
3147 ixgbe_vmdq_dcb_configure(dev);
3151 dcb_config->vt_mode = false;
3152 config_dcb_rx = DCB_RX_CONFIG;
3153 /* Get dcb TX configuration parameters from rte_eth_conf */
3154 ixgbe_dcb_rx_config(dev,dcb_config);
3155 /*Configure general DCB RX parameters*/
3156 ixgbe_dcb_rx_hw_config(hw, dcb_config);
3159 PMD_INIT_LOG(ERR, "Incorrect DCB RX mode configuration");
3162 switch (dev->data->dev_conf.txmode.mq_mode) {
3163 case ETH_MQ_TX_VMDQ_DCB:
3164 dcb_config->vt_mode = true;
3165 config_dcb_tx = DCB_TX_CONFIG;
3166 /* get DCB and VT TX configuration parameters from rte_eth_conf */
3167 ixgbe_dcb_vt_tx_config(dev,dcb_config);
3168 /*Configure general VMDQ and DCB TX parameters*/
3169 ixgbe_vmdq_dcb_hw_tx_config(dev,dcb_config);
3173 dcb_config->vt_mode = false;
3174 config_dcb_tx = DCB_TX_CONFIG;
3175 /*get DCB TX configuration parameters from rte_eth_conf*/
3176 ixgbe_dcb_tx_config(dev,dcb_config);
3177 /*Configure general DCB TX parameters*/
3178 ixgbe_dcb_tx_hw_config(hw, dcb_config);
3181 PMD_INIT_LOG(ERR, "Incorrect DCB TX mode configuration");
3185 nb_tcs = dcb_config->num_tcs.pfc_tcs;
3187 ixgbe_dcb_unpack_map_cee(dcb_config, IXGBE_DCB_RX_CONFIG, map);
3188 if(nb_tcs == ETH_4_TCS) {
3189 /* Avoid un-configured priority mapping to TC0 */
3191 uint8_t mask = 0xFF;
3192 for (i = 0; i < ETH_DCB_NUM_USER_PRIORITIES - 4; i++)
3193 mask = (uint8_t)(mask & (~ (1 << map[i])));
3194 for (i = 0; mask && (i < IXGBE_DCB_MAX_TRAFFIC_CLASS); i++) {
3195 if ((mask & 0x1) && (j < ETH_DCB_NUM_USER_PRIORITIES))
3199 /* Re-configure 4 TCs BW */
3200 for (i = 0; i < nb_tcs; i++) {
3201 tc = &dcb_config->tc_config[i];
3202 tc->path[IXGBE_DCB_TX_CONFIG].bwg_percent =
3203 (uint8_t)(100 / nb_tcs);
3204 tc->path[IXGBE_DCB_RX_CONFIG].bwg_percent =
3205 (uint8_t)(100 / nb_tcs);
3207 for (; i < IXGBE_DCB_MAX_TRAFFIC_CLASS; i++) {
3208 tc = &dcb_config->tc_config[i];
3209 tc->path[IXGBE_DCB_TX_CONFIG].bwg_percent = 0;
3210 tc->path[IXGBE_DCB_RX_CONFIG].bwg_percent = 0;
3215 /* Set RX buffer size */
3216 pbsize = (uint16_t)(NIC_RX_BUFFER_SIZE / nb_tcs);
3217 uint32_t rxpbsize = pbsize << IXGBE_RXPBSIZE_SHIFT;
3218 for (i = 0 ; i < nb_tcs; i++) {
3219 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), rxpbsize);
3221 /* zero alloc all unused TCs */
3222 for (; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
3223 IXGBE_WRITE_REG(hw, IXGBE_RXPBSIZE(i), 0);
3227 /* Only support an equally distributed Tx packet buffer strategy. */
3228 uint32_t txpktsize = IXGBE_TXPBSIZE_MAX / nb_tcs;
3229 uint32_t txpbthresh = (txpktsize / DCB_TX_PB) - IXGBE_TXPKT_SIZE_MAX;
3230 for (i = 0; i < nb_tcs; i++) {
3231 IXGBE_WRITE_REG(hw, IXGBE_TXPBSIZE(i), txpktsize);
3232 IXGBE_WRITE_REG(hw, IXGBE_TXPBTHRESH(i), txpbthresh);
3234 /* Clear unused TCs, if any, to zero buffer size*/
3235 for (; i < ETH_DCB_NUM_USER_PRIORITIES; i++) {
3236 IXGBE_WRITE_REG(hw, IXGBE_TXPBSIZE(i), 0);
3237 IXGBE_WRITE_REG(hw, IXGBE_TXPBTHRESH(i), 0);
3241 /*Calculates traffic class credits*/
3242 ixgbe_dcb_calculate_tc_credits_cee(hw, dcb_config,max_frame,
3243 IXGBE_DCB_TX_CONFIG);
3244 ixgbe_dcb_calculate_tc_credits_cee(hw, dcb_config,max_frame,
3245 IXGBE_DCB_RX_CONFIG);
3248 /* Unpack CEE standard containers */
3249 ixgbe_dcb_unpack_refill_cee(dcb_config, IXGBE_DCB_RX_CONFIG, refill);
3250 ixgbe_dcb_unpack_max_cee(dcb_config, max);
3251 ixgbe_dcb_unpack_bwgid_cee(dcb_config, IXGBE_DCB_RX_CONFIG, bwgid);
3252 ixgbe_dcb_unpack_tsa_cee(dcb_config, IXGBE_DCB_RX_CONFIG, tsa);
3253 /* Configure PG(ETS) RX */
3254 ixgbe_dcb_hw_arbite_rx_config(hw,refill,max,bwgid,tsa,map);
3258 /* Unpack CEE standard containers */
3259 ixgbe_dcb_unpack_refill_cee(dcb_config, IXGBE_DCB_TX_CONFIG, refill);
3260 ixgbe_dcb_unpack_max_cee(dcb_config, max);
3261 ixgbe_dcb_unpack_bwgid_cee(dcb_config, IXGBE_DCB_TX_CONFIG, bwgid);
3262 ixgbe_dcb_unpack_tsa_cee(dcb_config, IXGBE_DCB_TX_CONFIG, tsa);
3263 /* Configure PG(ETS) TX */
3264 ixgbe_dcb_hw_arbite_tx_config(hw,refill,max,bwgid,tsa,map);
3267 /*Configure queue statistics registers*/
3268 ixgbe_dcb_config_tc_stats_82599(hw, dcb_config);
3270 /* Check if the PFC is supported */
3271 if(dev->data->dev_conf.dcb_capability_en & ETH_DCB_PFC_SUPPORT) {
3272 pbsize = (uint16_t) (NIC_RX_BUFFER_SIZE / nb_tcs);
3273 for (i = 0; i < nb_tcs; i++) {
3275 * If the TC count is 8,and the default high_water is 48,
3276 * the low_water is 16 as default.
3278 hw->fc.high_water[i] = (pbsize * 3 ) / 4;
3279 hw->fc.low_water[i] = pbsize / 4;
3280 /* Enable pfc for this TC */
3281 tc = &dcb_config->tc_config[i];
3282 tc->pfc = ixgbe_dcb_pfc_enabled;
3284 ixgbe_dcb_unpack_pfc_cee(dcb_config, map, &pfc_en);
3285 if(dcb_config->num_tcs.pfc_tcs == ETH_4_TCS)
3287 ret = ixgbe_dcb_config_pfc(hw, pfc_en, map);
3294 * ixgbe_configure_dcb - Configure DCB Hardware
3295 * @dev: pointer to rte_eth_dev
3297 void ixgbe_configure_dcb(struct rte_eth_dev *dev)
3299 struct ixgbe_dcb_config *dcb_cfg =
3300 IXGBE_DEV_PRIVATE_TO_DCB_CFG(dev->data->dev_private);
3301 struct rte_eth_conf *dev_conf = &(dev->data->dev_conf);
3303 PMD_INIT_FUNC_TRACE();
3305 /* check support mq_mode for DCB */
3306 if ((dev_conf->rxmode.mq_mode != ETH_MQ_RX_VMDQ_DCB) &&
3307 (dev_conf->rxmode.mq_mode != ETH_MQ_RX_DCB))
3310 if (dev->data->nb_rx_queues != ETH_DCB_NUM_QUEUES)
3313 /** Configure DCB hardware **/
3314 ixgbe_dcb_hw_configure(dev,dcb_cfg);
3320 * VMDq only support for 10 GbE NIC.
3323 ixgbe_vmdq_rx_hw_configure(struct rte_eth_dev *dev)
3325 struct rte_eth_vmdq_rx_conf *cfg;
3326 struct ixgbe_hw *hw;
3327 enum rte_eth_nb_pools num_pools;
3328 uint32_t mrqc, vt_ctl, vlanctrl;
3332 PMD_INIT_FUNC_TRACE();
3333 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3334 cfg = &dev->data->dev_conf.rx_adv_conf.vmdq_rx_conf;
3335 num_pools = cfg->nb_queue_pools;
3337 ixgbe_rss_disable(dev);
3339 /* MRQC: enable vmdq */
3340 mrqc = IXGBE_MRQC_VMDQEN;
3341 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
3343 /* PFVTCTL: turn on virtualisation and set the default pool */
3344 vt_ctl = IXGBE_VT_CTL_VT_ENABLE | IXGBE_VT_CTL_REPLEN;
3345 if (cfg->enable_default_pool)
3346 vt_ctl |= (cfg->default_pool << IXGBE_VT_CTL_POOL_SHIFT);
3348 vt_ctl |= IXGBE_VT_CTL_DIS_DEFPL;
3350 IXGBE_WRITE_REG(hw, IXGBE_VT_CTL, vt_ctl);
3352 for (i = 0; i < (int)num_pools; i++) {
3353 vmolr = ixgbe_convert_vm_rx_mask_to_val(cfg->rx_mode, vmolr);
3354 IXGBE_WRITE_REG(hw, IXGBE_VMOLR(i), vmolr);
3357 /* VLNCTRL: enable vlan filtering and allow all vlan tags through */
3358 vlanctrl = IXGBE_READ_REG(hw, IXGBE_VLNCTRL);
3359 vlanctrl |= IXGBE_VLNCTRL_VFE ; /* enable vlan filters */
3360 IXGBE_WRITE_REG(hw, IXGBE_VLNCTRL, vlanctrl);
3362 /* VFTA - enable all vlan filters */
3363 for (i = 0; i < NUM_VFTA_REGISTERS; i++)
3364 IXGBE_WRITE_REG(hw, IXGBE_VFTA(i), UINT32_MAX);
3366 /* VFRE: pool enabling for receive - 64 */
3367 IXGBE_WRITE_REG(hw, IXGBE_VFRE(0), UINT32_MAX);
3368 if (num_pools == ETH_64_POOLS)
3369 IXGBE_WRITE_REG(hw, IXGBE_VFRE(1), UINT32_MAX);
3372 * MPSAR - allow pools to read specific mac addresses
3373 * In this case, all pools should be able to read from mac addr 0
3375 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_LO(0), UINT32_MAX);
3376 IXGBE_WRITE_REG(hw, IXGBE_MPSAR_HI(0), UINT32_MAX);
3378 /* PFVLVF, PFVLVFB: set up filters for vlan tags as configured */
3379 for (i = 0; i < cfg->nb_pool_maps; i++) {
3380 /* set vlan id in VF register and set the valid bit */
3381 IXGBE_WRITE_REG(hw, IXGBE_VLVF(i), (IXGBE_VLVF_VIEN | \
3382 (cfg->pool_map[i].vlan_id & IXGBE_RXD_VLAN_ID_MASK)));
3384 * Put the allowed pools in VFB reg. As we only have 16 or 64
3385 * pools, we only need to use the first half of the register
3388 if (((cfg->pool_map[i].pools >> 32) & UINT32_MAX) == 0)
3389 IXGBE_WRITE_REG(hw, IXGBE_VLVFB(i*2), \
3390 (cfg->pool_map[i].pools & UINT32_MAX));
3392 IXGBE_WRITE_REG(hw, IXGBE_VLVFB((i*2+1)), \
3393 ((cfg->pool_map[i].pools >> 32) \
3398 /* PFDMA Tx General Switch Control Enables VMDQ loopback */
3399 if (cfg->enable_loop_back) {
3400 IXGBE_WRITE_REG(hw, IXGBE_PFDTXGSWC, IXGBE_PFDTXGSWC_VT_LBEN);
3401 for (i = 0; i < RTE_IXGBE_VMTXSW_REGISTER_COUNT; i++)
3402 IXGBE_WRITE_REG(hw, IXGBE_VMTXSW(i), UINT32_MAX);
3405 IXGBE_WRITE_FLUSH(hw);
3409 * ixgbe_dcb_config_tx_hw_config - Configure general VMDq TX parameters
3410 * @hw: pointer to hardware structure
3413 ixgbe_vmdq_tx_hw_configure(struct ixgbe_hw *hw)
3418 PMD_INIT_FUNC_TRACE();
3419 /*PF VF Transmit Enable*/
3420 IXGBE_WRITE_REG(hw, IXGBE_VFTE(0), UINT32_MAX);
3421 IXGBE_WRITE_REG(hw, IXGBE_VFTE(1), UINT32_MAX);
3423 /* Disable the Tx desc arbiter so that MTQC can be changed */
3424 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
3425 reg |= IXGBE_RTTDCS_ARBDIS;
3426 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
3428 reg = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_64VF;
3429 IXGBE_WRITE_REG(hw, IXGBE_MTQC, reg);
3431 /* Disable drop for all queues */
3432 for (q = 0; q < IXGBE_MAX_RX_QUEUE_NUM; q++)
3433 IXGBE_WRITE_REG(hw, IXGBE_QDE,
3434 (IXGBE_QDE_WRITE | (q << IXGBE_QDE_IDX_SHIFT)));
3436 /* Enable the Tx desc arbiter */
3437 reg = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
3438 reg &= ~IXGBE_RTTDCS_ARBDIS;
3439 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, reg);
3441 IXGBE_WRITE_FLUSH(hw);
3446 static int __attribute__((cold))
3447 ixgbe_alloc_rx_queue_mbufs(struct ixgbe_rx_queue *rxq)
3449 struct ixgbe_rx_entry *rxe = rxq->sw_ring;
3453 /* Initialize software ring entries */
3454 for (i = 0; i < rxq->nb_rx_desc; i++) {
3455 volatile union ixgbe_adv_rx_desc *rxd;
3456 struct rte_mbuf *mbuf = rte_rxmbuf_alloc(rxq->mb_pool);
3458 PMD_INIT_LOG(ERR, "RX mbuf alloc failed queue_id=%u",
3459 (unsigned) rxq->queue_id);
3463 rte_mbuf_refcnt_set(mbuf, 1);
3465 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
3467 mbuf->port = rxq->port_id;
3470 rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mbuf));
3471 rxd = &rxq->rx_ring[i];
3472 rxd->read.hdr_addr = dma_addr;
3473 rxd->read.pkt_addr = dma_addr;
3481 ixgbe_config_vf_rss(struct rte_eth_dev *dev)
3483 struct ixgbe_hw *hw;
3486 ixgbe_rss_configure(dev);
3488 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3490 /* MRQC: enable VF RSS */
3491 mrqc = IXGBE_READ_REG(hw, IXGBE_MRQC);
3492 mrqc &= ~IXGBE_MRQC_MRQE_MASK;
3493 switch (RTE_ETH_DEV_SRIOV(dev).active) {
3495 mrqc |= IXGBE_MRQC_VMDQRSS64EN;
3499 mrqc |= IXGBE_MRQC_VMDQRSS32EN;
3503 PMD_INIT_LOG(ERR, "Invalid pool number in IOV mode with VMDQ RSS");
3507 IXGBE_WRITE_REG(hw, IXGBE_MRQC, mrqc);
3513 ixgbe_config_vf_default(struct rte_eth_dev *dev)
3515 struct ixgbe_hw *hw =
3516 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3518 switch (RTE_ETH_DEV_SRIOV(dev).active) {
3520 IXGBE_WRITE_REG(hw, IXGBE_MRQC,
3525 IXGBE_WRITE_REG(hw, IXGBE_MRQC,
3526 IXGBE_MRQC_VMDQRT4TCEN);
3530 IXGBE_WRITE_REG(hw, IXGBE_MRQC,
3531 IXGBE_MRQC_VMDQRT8TCEN);
3535 "invalid pool number in IOV mode");
3542 ixgbe_dev_mq_rx_configure(struct rte_eth_dev *dev)
3544 struct ixgbe_hw *hw =
3545 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3547 if (hw->mac.type == ixgbe_mac_82598EB)
3550 if (RTE_ETH_DEV_SRIOV(dev).active == 0) {
3552 * SRIOV inactive scheme
3553 * any DCB/RSS w/o VMDq multi-queue setting
3555 switch (dev->data->dev_conf.rxmode.mq_mode) {
3557 ixgbe_rss_configure(dev);
3560 case ETH_MQ_RX_VMDQ_DCB:
3561 ixgbe_vmdq_dcb_configure(dev);
3564 case ETH_MQ_RX_VMDQ_ONLY:
3565 ixgbe_vmdq_rx_hw_configure(dev);
3568 case ETH_MQ_RX_NONE:
3569 /* if mq_mode is none, disable rss mode.*/
3570 default: ixgbe_rss_disable(dev);
3574 * SRIOV active scheme
3575 * Support RSS together with VMDq & SRIOV
3577 switch (dev->data->dev_conf.rxmode.mq_mode) {
3579 case ETH_MQ_RX_VMDQ_RSS:
3580 ixgbe_config_vf_rss(dev);
3583 /* FIXME if support DCB/RSS together with VMDq & SRIOV */
3584 case ETH_MQ_RX_VMDQ_DCB:
3585 case ETH_MQ_RX_VMDQ_DCB_RSS:
3587 "Could not support DCB with VMDq & SRIOV");
3590 ixgbe_config_vf_default(dev);
3599 ixgbe_dev_mq_tx_configure(struct rte_eth_dev *dev)
3601 struct ixgbe_hw *hw =
3602 IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3606 if (hw->mac.type == ixgbe_mac_82598EB)
3609 /* disable arbiter before setting MTQC */
3610 rttdcs = IXGBE_READ_REG(hw, IXGBE_RTTDCS);
3611 rttdcs |= IXGBE_RTTDCS_ARBDIS;
3612 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, rttdcs);
3614 if (RTE_ETH_DEV_SRIOV(dev).active == 0) {
3616 * SRIOV inactive scheme
3617 * any DCB w/o VMDq multi-queue setting
3619 if (dev->data->dev_conf.txmode.mq_mode == ETH_MQ_TX_VMDQ_ONLY)
3620 ixgbe_vmdq_tx_hw_configure(hw);
3622 mtqc = IXGBE_MTQC_64Q_1PB;
3623 IXGBE_WRITE_REG(hw, IXGBE_MTQC, mtqc);
3626 switch (RTE_ETH_DEV_SRIOV(dev).active) {
3629 * SRIOV active scheme
3630 * FIXME if support DCB together with VMDq & SRIOV
3633 mtqc = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_64VF;
3636 mtqc = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_32VF;
3639 mtqc = IXGBE_MTQC_VT_ENA | IXGBE_MTQC_RT_ENA |
3643 mtqc = IXGBE_MTQC_64Q_1PB;
3644 PMD_INIT_LOG(ERR, "invalid pool number in IOV mode");
3646 IXGBE_WRITE_REG(hw, IXGBE_MTQC, mtqc);
3649 /* re-enable arbiter */
3650 rttdcs &= ~IXGBE_RTTDCS_ARBDIS;
3651 IXGBE_WRITE_REG(hw, IXGBE_RTTDCS, rttdcs);
3657 * ixgbe_get_rscctl_maxdesc - Calculate the RSCCTL[n].MAXDESC for PF
3659 * Return the RSCCTL[n].MAXDESC for 82599 and x540 PF devices according to the
3660 * spec rev. 3.0 chapter 8.2.3.8.13.
3662 * @pool Memory pool of the Rx queue
3664 static inline uint32_t
3665 ixgbe_get_rscctl_maxdesc(struct rte_mempool *pool)
3667 struct rte_pktmbuf_pool_private *mp_priv = rte_mempool_get_priv(pool);
3669 /* MAXDESC * SRRCTL.BSIZEPKT must not exceed 64 KB minus one */
3672 (mp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM);
3675 return IXGBE_RSCCTL_MAXDESC_16;
3676 else if (maxdesc >= 8)
3677 return IXGBE_RSCCTL_MAXDESC_8;
3678 else if (maxdesc >= 4)
3679 return IXGBE_RSCCTL_MAXDESC_4;
3681 return IXGBE_RSCCTL_MAXDESC_1;
3685 * ixgbe_set_ivar - Setup the correct IVAR register for a particular MSIX
3688 * (Taken from FreeBSD tree)
3689 * (yes this is all very magic and confusing :)
3692 * @entry the register array entry
3693 * @vector the MSIX vector for this queue
3697 ixgbe_set_ivar(struct rte_eth_dev *dev, u8 entry, u8 vector, s8 type)
3699 struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3702 vector |= IXGBE_IVAR_ALLOC_VAL;
3704 switch (hw->mac.type) {
3706 case ixgbe_mac_82598EB:
3708 entry = IXGBE_IVAR_OTHER_CAUSES_INDEX;
3710 entry += (type * 64);
3711 index = (entry >> 2) & 0x1F;
3712 ivar = IXGBE_READ_REG(hw, IXGBE_IVAR(index));
3713 ivar &= ~(0xFF << (8 * (entry & 0x3)));
3714 ivar |= (vector << (8 * (entry & 0x3)));
3715 IXGBE_WRITE_REG(hw, IXGBE_IVAR(index), ivar);
3718 case ixgbe_mac_82599EB:
3719 case ixgbe_mac_X540:
3720 if (type == -1) { /* MISC IVAR */
3721 index = (entry & 1) * 8;
3722 ivar = IXGBE_READ_REG(hw, IXGBE_IVAR_MISC);
3723 ivar &= ~(0xFF << index);
3724 ivar |= (vector << index);
3725 IXGBE_WRITE_REG(hw, IXGBE_IVAR_MISC, ivar);
3726 } else { /* RX/TX IVARS */
3727 index = (16 * (entry & 1)) + (8 * type);
3728 ivar = IXGBE_READ_REG(hw, IXGBE_IVAR(entry >> 1));
3729 ivar &= ~(0xFF << index);
3730 ivar |= (vector << index);
3731 IXGBE_WRITE_REG(hw, IXGBE_IVAR(entry >> 1), ivar);
3741 void __attribute__((cold))
3742 ixgbe_set_rx_function(struct rte_eth_dev *dev)
3744 struct ixgbe_adapter *adapter =
3745 (struct ixgbe_adapter *)dev->data->dev_private;
3748 * In order to allow Vector Rx there are a few configuration
3749 * conditions to be met and Rx Bulk Allocation should be allowed.
3751 if (ixgbe_rx_vec_dev_conf_condition_check(dev) ||
3752 !adapter->rx_bulk_alloc_allowed) {
3753 PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet Vector Rx "
3754 "preconditions or RTE_IXGBE_INC_VECTOR is "
3756 dev->data->port_id);
3758 adapter->rx_vec_allowed = false;
3762 * Initialize the appropriate LRO callback.
3764 * If all queues satisfy the bulk allocation preconditions
3765 * (hw->rx_bulk_alloc_allowed is TRUE) then we may use bulk allocation.
3766 * Otherwise use a single allocation version.
3768 if (dev->data->lro) {
3769 if (adapter->rx_bulk_alloc_allowed) {
3770 PMD_INIT_LOG(INFO, "LRO is requested. Using a bulk "
3771 "allocation version");
3772 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_bulk_alloc;
3774 PMD_INIT_LOG(INFO, "LRO is requested. Using a single "
3775 "allocation version");
3776 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_single_alloc;
3778 } else if (dev->data->scattered_rx) {
3780 * Set the non-LRO scattered callback: there are Vector and
3781 * single allocation versions.
3783 if (adapter->rx_vec_allowed) {
3784 PMD_INIT_LOG(DEBUG, "Using Vector Scattered Rx "
3785 "callback (port=%d).",
3786 dev->data->port_id);
3788 dev->rx_pkt_burst = ixgbe_recv_scattered_pkts_vec;
3789 } else if (adapter->rx_bulk_alloc_allowed) {
3790 PMD_INIT_LOG(INFO, "Using a Scattered with bulk "
3791 "allocation callback (port=%d).",
3792 dev->data->port_id);
3793 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_bulk_alloc;
3795 PMD_INIT_LOG(DEBUG, "Using Regualr (non-vector, "
3796 "single allocation) "
3797 "Scattered Rx callback "
3799 dev->data->port_id);
3801 dev->rx_pkt_burst = ixgbe_recv_pkts_lro_single_alloc;
3804 * Below we set "simple" callbacks according to port/queues parameters.
3805 * If parameters allow we are going to choose between the following
3809 * - Single buffer allocation (the simplest one)
3811 } else if (adapter->rx_vec_allowed) {
3812 PMD_INIT_LOG(INFO, "Vector rx enabled, please make sure RX "
3813 "burst size no less than 32.");
3815 dev->rx_pkt_burst = ixgbe_recv_pkts_vec;
3816 } else if (adapter->rx_bulk_alloc_allowed) {
3817 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
3818 "satisfied. Rx Burst Bulk Alloc function "
3819 "will be used on port=%d.",
3820 dev->data->port_id);
3822 dev->rx_pkt_burst = ixgbe_recv_pkts_bulk_alloc;
3824 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are not "
3825 "satisfied, or Scattered Rx is requested, "
3826 "or RTE_LIBRTE_IXGBE_RX_ALLOW_BULK_ALLOC "
3827 "is not enabled (port=%d).",
3828 dev->data->port_id);
3830 dev->rx_pkt_burst = ixgbe_recv_pkts;
3835 * ixgbe_set_rsc - configure RSC related port HW registers
3837 * Configures the port's RSC related registers according to the 4.6.7.2 chapter
3838 * of 82599 Spec (x540 configuration is virtually the same).
3842 * Returns 0 in case of success or a non-zero error code
3845 ixgbe_set_rsc(struct rte_eth_dev *dev)
3847 struct rte_eth_rxmode *rx_conf = &dev->data->dev_conf.rxmode;
3848 struct ixgbe_hw *hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3849 struct rte_eth_dev_info dev_info = { 0 };
3850 bool rsc_capable = false;
3855 dev->dev_ops->dev_infos_get(dev, &dev_info);
3856 if (dev_info.rx_offload_capa & DEV_RX_OFFLOAD_TCP_LRO)
3859 if (!rsc_capable && rx_conf->enable_lro) {
3860 PMD_INIT_LOG(CRIT, "LRO is requested on HW that doesn't "
3865 /* RSC global configuration (chapter 4.6.7.2.1 of 82599 Spec) */
3867 if (!rx_conf->hw_strip_crc && rx_conf->enable_lro) {
3869 * According to chapter of 4.6.7.2.1 of the Spec Rev.
3870 * 3.0 RSC configuration requires HW CRC stripping being
3871 * enabled. If user requested both HW CRC stripping off
3872 * and RSC on - return an error.
3874 PMD_INIT_LOG(CRIT, "LRO can't be enabled when HW CRC "
3879 /* RFCTL configuration */
3881 uint32_t rfctl = IXGBE_READ_REG(hw, IXGBE_RFCTL);
3882 if (rx_conf->enable_lro)
3884 * Since NFS packets coalescing is not supported - clear
3885 * RFCTL.NFSW_DIS and RFCTL.NFSR_DIS when RSC is
3888 rfctl &= ~(IXGBE_RFCTL_RSC_DIS | IXGBE_RFCTL_NFSW_DIS |
3889 IXGBE_RFCTL_NFSR_DIS);
3891 rfctl |= IXGBE_RFCTL_RSC_DIS;
3893 IXGBE_WRITE_REG(hw, IXGBE_RFCTL, rfctl);
3896 /* If LRO hasn't been requested - we are done here. */
3897 if (!rx_conf->enable_lro)
3900 /* Set RDRXCTL.RSCACKC bit */
3901 rdrxctl = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
3902 rdrxctl |= IXGBE_RDRXCTL_RSCACKC;
3903 IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rdrxctl);
3905 /* Per-queue RSC configuration (chapter 4.6.7.2.2 of 82599 Spec) */
3906 for (i = 0; i < dev->data->nb_rx_queues; i++) {
3907 struct ixgbe_rx_queue *rxq = dev->data->rx_queues[i];
3909 IXGBE_READ_REG(hw, IXGBE_SRRCTL(rxq->reg_idx));
3911 IXGBE_READ_REG(hw, IXGBE_RSCCTL(rxq->reg_idx));
3913 IXGBE_READ_REG(hw, IXGBE_PSRTYPE(rxq->reg_idx));
3915 IXGBE_READ_REG(hw, IXGBE_EITR(rxq->reg_idx));
3918 * ixgbe PMD doesn't support header-split at the moment.
3920 * Following the 4.6.7.2.1 chapter of the 82599/x540
3921 * Spec if RSC is enabled the SRRCTL[n].BSIZEHEADER
3922 * should be configured even if header split is not
3923 * enabled. We will configure it 128 bytes following the
3924 * recommendation in the spec.
3926 srrctl &= ~IXGBE_SRRCTL_BSIZEHDR_MASK;
3927 srrctl |= (128 << IXGBE_SRRCTL_BSIZEHDRSIZE_SHIFT) &
3928 IXGBE_SRRCTL_BSIZEHDR_MASK;
3931 * TODO: Consider setting the Receive Descriptor Minimum
3932 * Threshold Size for an RSC case. This is not an obviously
3933 * beneficiary option but the one worth considering...
3936 rscctl |= IXGBE_RSCCTL_RSCEN;
3937 rscctl |= ixgbe_get_rscctl_maxdesc(rxq->mb_pool);
3938 psrtype |= IXGBE_PSRTYPE_TCPHDR;
3941 * RSC: Set ITR interval corresponding to 2K ints/s.
3943 * Full-sized RSC aggregations for a 10Gb/s link will
3944 * arrive at about 20K aggregation/s rate.
3946 * 2K inst/s rate will make only 10% of the
3947 * aggregations to be closed due to the interrupt timer
3948 * expiration for a streaming at wire-speed case.
3950 * For a sparse streaming case this setting will yield
3951 * at most 500us latency for a single RSC aggregation.
3953 eitr &= ~IXGBE_EITR_ITR_INT_MASK;
3954 eitr |= IXGBE_EITR_INTERVAL_US(500) | IXGBE_EITR_CNT_WDIS;
3956 IXGBE_WRITE_REG(hw, IXGBE_SRRCTL(rxq->reg_idx), srrctl);
3957 IXGBE_WRITE_REG(hw, IXGBE_RSCCTL(rxq->reg_idx), rscctl);
3958 IXGBE_WRITE_REG(hw, IXGBE_PSRTYPE(rxq->reg_idx), psrtype);
3959 IXGBE_WRITE_REG(hw, IXGBE_EITR(rxq->reg_idx), eitr);
3962 * RSC requires the mapping of the queue to the
3965 ixgbe_set_ivar(dev, rxq->reg_idx, i, 0);
3970 PMD_INIT_LOG(INFO, "enabling LRO mode");
3976 * Initializes Receive Unit.
3978 int __attribute__((cold))
3979 ixgbe_dev_rx_init(struct rte_eth_dev *dev)
3981 struct ixgbe_hw *hw;
3982 struct ixgbe_rx_queue *rxq;
3993 struct rte_eth_rxmode *rx_conf = &dev->data->dev_conf.rxmode;
3996 PMD_INIT_FUNC_TRACE();
3997 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4000 * Make sure receives are disabled while setting
4001 * up the RX context (registers, descriptor rings, etc.).
4003 rxctrl = IXGBE_READ_REG(hw, IXGBE_RXCTRL);
4004 IXGBE_WRITE_REG(hw, IXGBE_RXCTRL, rxctrl & ~IXGBE_RXCTRL_RXEN);
4006 /* Enable receipt of broadcasted frames */
4007 fctrl = IXGBE_READ_REG(hw, IXGBE_FCTRL);
4008 fctrl |= IXGBE_FCTRL_BAM;
4009 fctrl |= IXGBE_FCTRL_DPF;
4010 fctrl |= IXGBE_FCTRL_PMCF;
4011 IXGBE_WRITE_REG(hw, IXGBE_FCTRL, fctrl);
4014 * Configure CRC stripping, if any.
4016 hlreg0 = IXGBE_READ_REG(hw, IXGBE_HLREG0);
4017 if (rx_conf->hw_strip_crc)
4018 hlreg0 |= IXGBE_HLREG0_RXCRCSTRP;
4020 hlreg0 &= ~IXGBE_HLREG0_RXCRCSTRP;
4023 * Configure jumbo frame support, if any.
4025 if (rx_conf->jumbo_frame == 1) {
4026 hlreg0 |= IXGBE_HLREG0_JUMBOEN;
4027 maxfrs = IXGBE_READ_REG(hw, IXGBE_MAXFRS);
4028 maxfrs &= 0x0000FFFF;
4029 maxfrs |= (rx_conf->max_rx_pkt_len << 16);
4030 IXGBE_WRITE_REG(hw, IXGBE_MAXFRS, maxfrs);
4032 hlreg0 &= ~IXGBE_HLREG0_JUMBOEN;
4035 * If loopback mode is configured for 82599, set LPBK bit.
4037 if (hw->mac.type == ixgbe_mac_82599EB &&
4038 dev->data->dev_conf.lpbk_mode == IXGBE_LPBK_82599_TX_RX)
4039 hlreg0 |= IXGBE_HLREG0_LPBK;
4041 hlreg0 &= ~IXGBE_HLREG0_LPBK;
4043 IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hlreg0);
4045 /* Setup RX queues */
4046 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4047 rxq = dev->data->rx_queues[i];
4050 * Reset crc_len in case it was changed after queue setup by a
4051 * call to configure.
4053 rxq->crc_len = rx_conf->hw_strip_crc ? 0 : ETHER_CRC_LEN;
4055 /* Setup the Base and Length of the Rx Descriptor Rings */
4056 bus_addr = rxq->rx_ring_phys_addr;
4057 IXGBE_WRITE_REG(hw, IXGBE_RDBAL(rxq->reg_idx),
4058 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4059 IXGBE_WRITE_REG(hw, IXGBE_RDBAH(rxq->reg_idx),
4060 (uint32_t)(bus_addr >> 32));
4061 IXGBE_WRITE_REG(hw, IXGBE_RDLEN(rxq->reg_idx),
4062 rxq->nb_rx_desc * sizeof(union ixgbe_adv_rx_desc));
4063 IXGBE_WRITE_REG(hw, IXGBE_RDH(rxq->reg_idx), 0);
4064 IXGBE_WRITE_REG(hw, IXGBE_RDT(rxq->reg_idx), 0);
4066 /* Configure the SRRCTL register */
4067 #ifdef RTE_HEADER_SPLIT_ENABLE
4069 * Configure Header Split
4071 if (rx_conf->header_split) {
4072 if (hw->mac.type == ixgbe_mac_82599EB) {
4073 /* Must setup the PSRTYPE register */
4075 psrtype = IXGBE_PSRTYPE_TCPHDR |
4076 IXGBE_PSRTYPE_UDPHDR |
4077 IXGBE_PSRTYPE_IPV4HDR |
4078 IXGBE_PSRTYPE_IPV6HDR;
4079 IXGBE_WRITE_REG(hw, IXGBE_PSRTYPE(rxq->reg_idx), psrtype);
4081 srrctl = ((rx_conf->split_hdr_size <<
4082 IXGBE_SRRCTL_BSIZEHDRSIZE_SHIFT) &
4083 IXGBE_SRRCTL_BSIZEHDR_MASK);
4084 srrctl |= IXGBE_SRRCTL_DESCTYPE_HDR_SPLIT_ALWAYS;
4087 srrctl = IXGBE_SRRCTL_DESCTYPE_ADV_ONEBUF;
4089 /* Set if packets are dropped when no descriptors available */
4091 srrctl |= IXGBE_SRRCTL_DROP_EN;
4094 * Configure the RX buffer size in the BSIZEPACKET field of
4095 * the SRRCTL register of the queue.
4096 * The value is in 1 KB resolution. Valid values can be from
4099 buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
4100 RTE_PKTMBUF_HEADROOM);
4101 srrctl |= ((buf_size >> IXGBE_SRRCTL_BSIZEPKT_SHIFT) &
4102 IXGBE_SRRCTL_BSIZEPKT_MASK);
4104 IXGBE_WRITE_REG(hw, IXGBE_SRRCTL(rxq->reg_idx), srrctl);
4106 buf_size = (uint16_t) ((srrctl & IXGBE_SRRCTL_BSIZEPKT_MASK) <<
4107 IXGBE_SRRCTL_BSIZEPKT_SHIFT);
4109 /* It adds dual VLAN length for supporting dual VLAN */
4110 if (dev->data->dev_conf.rxmode.max_rx_pkt_len +
4111 2 * IXGBE_VLAN_TAG_SIZE > buf_size)
4112 dev->data->scattered_rx = 1;
4115 if (rx_conf->enable_scatter)
4116 dev->data->scattered_rx = 1;
4119 * Device configured with multiple RX queues.
4121 ixgbe_dev_mq_rx_configure(dev);
4124 * Setup the Checksum Register.
4125 * Disable Full-Packet Checksum which is mutually exclusive with RSS.
4126 * Enable IP/L4 checkum computation by hardware if requested to do so.
4128 rxcsum = IXGBE_READ_REG(hw, IXGBE_RXCSUM);
4129 rxcsum |= IXGBE_RXCSUM_PCSD;
4130 if (rx_conf->hw_ip_checksum)
4131 rxcsum |= IXGBE_RXCSUM_IPPCSE;
4133 rxcsum &= ~IXGBE_RXCSUM_IPPCSE;
4135 IXGBE_WRITE_REG(hw, IXGBE_RXCSUM, rxcsum);
4137 if (hw->mac.type == ixgbe_mac_82599EB ||
4138 hw->mac.type == ixgbe_mac_X540) {
4139 rdrxctl = IXGBE_READ_REG(hw, IXGBE_RDRXCTL);
4140 if (rx_conf->hw_strip_crc)
4141 rdrxctl |= IXGBE_RDRXCTL_CRCSTRIP;
4143 rdrxctl &= ~IXGBE_RDRXCTL_CRCSTRIP;
4144 rdrxctl &= ~IXGBE_RDRXCTL_RSCFRSTSIZE;
4145 IXGBE_WRITE_REG(hw, IXGBE_RDRXCTL, rdrxctl);
4148 rc = ixgbe_set_rsc(dev);
4152 ixgbe_set_rx_function(dev);
4158 * Initializes Transmit Unit.
4160 void __attribute__((cold))
4161 ixgbe_dev_tx_init(struct rte_eth_dev *dev)
4163 struct ixgbe_hw *hw;
4164 struct ixgbe_tx_queue *txq;
4170 PMD_INIT_FUNC_TRACE();
4171 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4173 /* Enable TX CRC (checksum offload requirement) and hw padding
4174 * (TSO requirement) */
4175 hlreg0 = IXGBE_READ_REG(hw, IXGBE_HLREG0);
4176 hlreg0 |= (IXGBE_HLREG0_TXCRCEN | IXGBE_HLREG0_TXPADEN);
4177 IXGBE_WRITE_REG(hw, IXGBE_HLREG0, hlreg0);
4179 /* Setup the Base and Length of the Tx Descriptor Rings */
4180 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4181 txq = dev->data->tx_queues[i];
4183 bus_addr = txq->tx_ring_phys_addr;
4184 IXGBE_WRITE_REG(hw, IXGBE_TDBAL(txq->reg_idx),
4185 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4186 IXGBE_WRITE_REG(hw, IXGBE_TDBAH(txq->reg_idx),
4187 (uint32_t)(bus_addr >> 32));
4188 IXGBE_WRITE_REG(hw, IXGBE_TDLEN(txq->reg_idx),
4189 txq->nb_tx_desc * sizeof(union ixgbe_adv_tx_desc));
4190 /* Setup the HW Tx Head and TX Tail descriptor pointers */
4191 IXGBE_WRITE_REG(hw, IXGBE_TDH(txq->reg_idx), 0);
4192 IXGBE_WRITE_REG(hw, IXGBE_TDT(txq->reg_idx), 0);
4195 * Disable Tx Head Writeback RO bit, since this hoses
4196 * bookkeeping if things aren't delivered in order.
4198 switch (hw->mac.type) {
4199 case ixgbe_mac_82598EB:
4200 txctrl = IXGBE_READ_REG(hw,
4201 IXGBE_DCA_TXCTRL(txq->reg_idx));
4202 txctrl &= ~IXGBE_DCA_TXCTRL_DESC_WRO_EN;
4203 IXGBE_WRITE_REG(hw, IXGBE_DCA_TXCTRL(txq->reg_idx),
4207 case ixgbe_mac_82599EB:
4208 case ixgbe_mac_X540:
4209 case ixgbe_mac_X550:
4210 case ixgbe_mac_X550EM_x:
4212 txctrl = IXGBE_READ_REG(hw,
4213 IXGBE_DCA_TXCTRL_82599(txq->reg_idx));
4214 txctrl &= ~IXGBE_DCA_TXCTRL_DESC_WRO_EN;
4215 IXGBE_WRITE_REG(hw, IXGBE_DCA_TXCTRL_82599(txq->reg_idx),
4221 /* Device configured with multiple TX queues. */
4222 ixgbe_dev_mq_tx_configure(dev);
4226 * Set up link for 82599 loopback mode Tx->Rx.
4228 static inline void __attribute__((cold))
4229 ixgbe_setup_loopback_link_82599(struct ixgbe_hw *hw)
4231 PMD_INIT_FUNC_TRACE();
4233 if (ixgbe_verify_lesm_fw_enabled_82599(hw)) {
4234 if (hw->mac.ops.acquire_swfw_sync(hw, IXGBE_GSSR_MAC_CSR_SM) !=
4236 PMD_INIT_LOG(ERR, "Could not enable loopback mode");
4245 IXGBE_AUTOC_LMS_10G_LINK_NO_AN | IXGBE_AUTOC_FLU);
4246 ixgbe_reset_pipeline_82599(hw);
4248 hw->mac.ops.release_swfw_sync(hw, IXGBE_GSSR_MAC_CSR_SM);
4254 * Start Transmit and Receive Units.
4256 int __attribute__((cold))
4257 ixgbe_dev_rxtx_start(struct rte_eth_dev *dev)
4259 struct ixgbe_hw *hw;
4260 struct ixgbe_tx_queue *txq;
4261 struct ixgbe_rx_queue *rxq;
4268 PMD_INIT_FUNC_TRACE();
4269 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4271 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4272 txq = dev->data->tx_queues[i];
4273 /* Setup Transmit Threshold Registers */
4274 txdctl = IXGBE_READ_REG(hw, IXGBE_TXDCTL(txq->reg_idx));
4275 txdctl |= txq->pthresh & 0x7F;
4276 txdctl |= ((txq->hthresh & 0x7F) << 8);
4277 txdctl |= ((txq->wthresh & 0x7F) << 16);
4278 IXGBE_WRITE_REG(hw, IXGBE_TXDCTL(txq->reg_idx), txdctl);
4281 if (hw->mac.type != ixgbe_mac_82598EB) {
4282 dmatxctl = IXGBE_READ_REG(hw, IXGBE_DMATXCTL);
4283 dmatxctl |= IXGBE_DMATXCTL_TE;
4284 IXGBE_WRITE_REG(hw, IXGBE_DMATXCTL, dmatxctl);
4287 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4288 txq = dev->data->tx_queues[i];
4289 if (!txq->tx_deferred_start) {
4290 ret = ixgbe_dev_tx_queue_start(dev, i);
4296 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4297 rxq = dev->data->rx_queues[i];
4298 if (!rxq->rx_deferred_start) {
4299 ret = ixgbe_dev_rx_queue_start(dev, i);
4305 /* Enable Receive engine */
4306 rxctrl = IXGBE_READ_REG(hw, IXGBE_RXCTRL);
4307 if (hw->mac.type == ixgbe_mac_82598EB)
4308 rxctrl |= IXGBE_RXCTRL_DMBYPS;
4309 rxctrl |= IXGBE_RXCTRL_RXEN;
4310 hw->mac.ops.enable_rx_dma(hw, rxctrl);
4312 /* If loopback mode is enabled for 82599, set up the link accordingly */
4313 if (hw->mac.type == ixgbe_mac_82599EB &&
4314 dev->data->dev_conf.lpbk_mode == IXGBE_LPBK_82599_TX_RX)
4315 ixgbe_setup_loopback_link_82599(hw);
4321 * Start Receive Units for specified queue.
4323 int __attribute__((cold))
4324 ixgbe_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
4326 struct ixgbe_hw *hw;
4327 struct ixgbe_rx_queue *rxq;
4331 PMD_INIT_FUNC_TRACE();
4332 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4334 if (rx_queue_id < dev->data->nb_rx_queues) {
4335 rxq = dev->data->rx_queues[rx_queue_id];
4337 /* Allocate buffers for descriptor rings */
4338 if (ixgbe_alloc_rx_queue_mbufs(rxq) != 0) {
4339 PMD_INIT_LOG(ERR, "Could not alloc mbuf for queue:%d",
4343 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4344 rxdctl |= IXGBE_RXDCTL_ENABLE;
4345 IXGBE_WRITE_REG(hw, IXGBE_RXDCTL(rxq->reg_idx), rxdctl);
4347 /* Wait until RX Enable ready */
4348 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4351 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4352 } while (--poll_ms && !(rxdctl & IXGBE_RXDCTL_ENABLE));
4354 PMD_INIT_LOG(ERR, "Could not enable Rx Queue %d",
4357 IXGBE_WRITE_REG(hw, IXGBE_RDH(rxq->reg_idx), 0);
4358 IXGBE_WRITE_REG(hw, IXGBE_RDT(rxq->reg_idx), rxq->nb_rx_desc - 1);
4366 * Stop Receive Units for specified queue.
4368 int __attribute__((cold))
4369 ixgbe_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
4371 struct ixgbe_hw *hw;
4372 struct ixgbe_adapter *adapter =
4373 (struct ixgbe_adapter *)dev->data->dev_private;
4374 struct ixgbe_rx_queue *rxq;
4378 PMD_INIT_FUNC_TRACE();
4379 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4381 if (rx_queue_id < dev->data->nb_rx_queues) {
4382 rxq = dev->data->rx_queues[rx_queue_id];
4384 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4385 rxdctl &= ~IXGBE_RXDCTL_ENABLE;
4386 IXGBE_WRITE_REG(hw, IXGBE_RXDCTL(rxq->reg_idx), rxdctl);
4388 /* Wait until RX Enable ready */
4389 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4392 rxdctl = IXGBE_READ_REG(hw, IXGBE_RXDCTL(rxq->reg_idx));
4393 } while (--poll_ms && (rxdctl | IXGBE_RXDCTL_ENABLE));
4395 PMD_INIT_LOG(ERR, "Could not disable Rx Queue %d",
4398 rte_delay_us(RTE_IXGBE_WAIT_100_US);
4400 ixgbe_rx_queue_release_mbufs(rxq);
4401 ixgbe_reset_rx_queue(adapter, rxq);
4410 * Start Transmit Units for specified queue.
4412 int __attribute__((cold))
4413 ixgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
4415 struct ixgbe_hw *hw;
4416 struct ixgbe_tx_queue *txq;
4420 PMD_INIT_FUNC_TRACE();
4421 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4423 if (tx_queue_id < dev->data->nb_tx_queues) {
4424 txq = dev->data->tx_queues[tx_queue_id];
4425 txdctl = IXGBE_READ_REG(hw, IXGBE_TXDCTL(txq->reg_idx));
4426 txdctl |= IXGBE_TXDCTL_ENABLE;
4427 IXGBE_WRITE_REG(hw, IXGBE_TXDCTL(txq->reg_idx), txdctl);
4429 /* Wait until TX Enable ready */
4430 if (hw->mac.type == ixgbe_mac_82599EB) {
4431 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4434 txdctl = IXGBE_READ_REG(hw,
4435 IXGBE_TXDCTL(txq->reg_idx));
4436 } while (--poll_ms && !(txdctl & IXGBE_TXDCTL_ENABLE));
4438 PMD_INIT_LOG(ERR, "Could not enable "
4439 "Tx Queue %d", tx_queue_id);
4442 IXGBE_WRITE_REG(hw, IXGBE_TDH(txq->reg_idx), 0);
4443 IXGBE_WRITE_REG(hw, IXGBE_TDT(txq->reg_idx), 0);
4451 * Stop Transmit Units for specified queue.
4453 int __attribute__((cold))
4454 ixgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
4456 struct ixgbe_hw *hw;
4457 struct ixgbe_tx_queue *txq;
4459 uint32_t txtdh, txtdt;
4462 PMD_INIT_FUNC_TRACE();
4463 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4465 if (tx_queue_id < dev->data->nb_tx_queues) {
4466 txq = dev->data->tx_queues[tx_queue_id];
4468 /* Wait until TX queue is empty */
4469 if (hw->mac.type == ixgbe_mac_82599EB) {
4470 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4472 rte_delay_us(RTE_IXGBE_WAIT_100_US);
4473 txtdh = IXGBE_READ_REG(hw,
4474 IXGBE_TDH(txq->reg_idx));
4475 txtdt = IXGBE_READ_REG(hw,
4476 IXGBE_TDT(txq->reg_idx));
4477 } while (--poll_ms && (txtdh != txtdt));
4479 PMD_INIT_LOG(ERR, "Tx Queue %d is not empty "
4480 "when stopping.", tx_queue_id);
4483 txdctl = IXGBE_READ_REG(hw, IXGBE_TXDCTL(txq->reg_idx));
4484 txdctl &= ~IXGBE_TXDCTL_ENABLE;
4485 IXGBE_WRITE_REG(hw, IXGBE_TXDCTL(txq->reg_idx), txdctl);
4487 /* Wait until TX Enable ready */
4488 if (hw->mac.type == ixgbe_mac_82599EB) {
4489 poll_ms = RTE_IXGBE_REGISTER_POLL_WAIT_10_MS;
4492 txdctl = IXGBE_READ_REG(hw,
4493 IXGBE_TXDCTL(txq->reg_idx));
4494 } while (--poll_ms && (txdctl | IXGBE_TXDCTL_ENABLE));
4496 PMD_INIT_LOG(ERR, "Could not disable "
4497 "Tx Queue %d", tx_queue_id);
4500 if (txq->ops != NULL) {
4501 txq->ops->release_mbufs(txq);
4502 txq->ops->reset(txq);
4511 * [VF] Initializes Receive Unit.
4513 int __attribute__((cold))
4514 ixgbevf_dev_rx_init(struct rte_eth_dev *dev)
4516 struct ixgbe_hw *hw;
4517 struct ixgbe_rx_queue *rxq;
4519 uint32_t srrctl, psrtype = 0;
4524 PMD_INIT_FUNC_TRACE();
4525 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4527 if (rte_is_power_of_2(dev->data->nb_rx_queues) == 0) {
4528 PMD_INIT_LOG(ERR, "The number of Rx queue invalid, "
4529 "it should be power of 2");
4533 if (dev->data->nb_rx_queues > hw->mac.max_rx_queues) {
4534 PMD_INIT_LOG(ERR, "The number of Rx queue invalid, "
4535 "it should be equal to or less than %d",
4536 hw->mac.max_rx_queues);
4541 * When the VF driver issues a IXGBE_VF_RESET request, the PF driver
4542 * disables the VF receipt of packets if the PF MTU is > 1500.
4543 * This is done to deal with 82599 limitations that imposes
4544 * the PF and all VFs to share the same MTU.
4545 * Then, the PF driver enables again the VF receipt of packet when
4546 * the VF driver issues a IXGBE_VF_SET_LPE request.
4547 * In the meantime, the VF device cannot be used, even if the VF driver
4548 * and the Guest VM network stack are ready to accept packets with a
4549 * size up to the PF MTU.
4550 * As a work-around to this PF behaviour, force the call to
4551 * ixgbevf_rlpml_set_vf even if jumbo frames are not used. This way,
4552 * VF packets received can work in all cases.
4554 ixgbevf_rlpml_set_vf(hw,
4555 (uint16_t)dev->data->dev_conf.rxmode.max_rx_pkt_len);
4557 /* Setup RX queues */
4558 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4559 rxq = dev->data->rx_queues[i];
4561 /* Allocate buffers for descriptor rings */
4562 ret = ixgbe_alloc_rx_queue_mbufs(rxq);
4566 /* Setup the Base and Length of the Rx Descriptor Rings */
4567 bus_addr = rxq->rx_ring_phys_addr;
4569 IXGBE_WRITE_REG(hw, IXGBE_VFRDBAL(i),
4570 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4571 IXGBE_WRITE_REG(hw, IXGBE_VFRDBAH(i),
4572 (uint32_t)(bus_addr >> 32));
4573 IXGBE_WRITE_REG(hw, IXGBE_VFRDLEN(i),
4574 rxq->nb_rx_desc * sizeof(union ixgbe_adv_rx_desc));
4575 IXGBE_WRITE_REG(hw, IXGBE_VFRDH(i), 0);
4576 IXGBE_WRITE_REG(hw, IXGBE_VFRDT(i), 0);
4579 /* Configure the SRRCTL register */
4580 #ifdef RTE_HEADER_SPLIT_ENABLE
4582 * Configure Header Split
4584 if (dev->data->dev_conf.rxmode.header_split) {
4585 srrctl = ((dev->data->dev_conf.rxmode.split_hdr_size <<
4586 IXGBE_SRRCTL_BSIZEHDRSIZE_SHIFT) &
4587 IXGBE_SRRCTL_BSIZEHDR_MASK);
4588 srrctl |= IXGBE_SRRCTL_DESCTYPE_HDR_SPLIT_ALWAYS;
4591 srrctl = IXGBE_SRRCTL_DESCTYPE_ADV_ONEBUF;
4593 /* Set if packets are dropped when no descriptors available */
4595 srrctl |= IXGBE_SRRCTL_DROP_EN;
4598 * Configure the RX buffer size in the BSIZEPACKET field of
4599 * the SRRCTL register of the queue.
4600 * The value is in 1 KB resolution. Valid values can be from
4603 buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
4604 RTE_PKTMBUF_HEADROOM);
4605 srrctl |= ((buf_size >> IXGBE_SRRCTL_BSIZEPKT_SHIFT) &
4606 IXGBE_SRRCTL_BSIZEPKT_MASK);
4609 * VF modification to write virtual function SRRCTL register
4611 IXGBE_WRITE_REG(hw, IXGBE_VFSRRCTL(i), srrctl);
4613 buf_size = (uint16_t) ((srrctl & IXGBE_SRRCTL_BSIZEPKT_MASK) <<
4614 IXGBE_SRRCTL_BSIZEPKT_SHIFT);
4616 if (dev->data->dev_conf.rxmode.enable_scatter ||
4617 /* It adds dual VLAN length for supporting dual VLAN */
4618 (dev->data->dev_conf.rxmode.max_rx_pkt_len +
4619 2 * IXGBE_VLAN_TAG_SIZE) > buf_size) {
4620 if (!dev->data->scattered_rx)
4621 PMD_INIT_LOG(DEBUG, "forcing scatter mode");
4622 dev->data->scattered_rx = 1;
4626 #ifdef RTE_HEADER_SPLIT_ENABLE
4627 if (dev->data->dev_conf.rxmode.header_split)
4628 /* Must setup the PSRTYPE register */
4629 psrtype = IXGBE_PSRTYPE_TCPHDR |
4630 IXGBE_PSRTYPE_UDPHDR |
4631 IXGBE_PSRTYPE_IPV4HDR |
4632 IXGBE_PSRTYPE_IPV6HDR;
4635 /* Set RQPL for VF RSS according to max Rx queue */
4636 psrtype |= (dev->data->nb_rx_queues >> 1) <<
4637 IXGBE_PSRTYPE_RQPL_SHIFT;
4638 IXGBE_WRITE_REG(hw, IXGBE_VFPSRTYPE, psrtype);
4640 ixgbe_set_rx_function(dev);
4646 * [VF] Initializes Transmit Unit.
4648 void __attribute__((cold))
4649 ixgbevf_dev_tx_init(struct rte_eth_dev *dev)
4651 struct ixgbe_hw *hw;
4652 struct ixgbe_tx_queue *txq;
4657 PMD_INIT_FUNC_TRACE();
4658 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4660 /* Setup the Base and Length of the Tx Descriptor Rings */
4661 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4662 txq = dev->data->tx_queues[i];
4663 bus_addr = txq->tx_ring_phys_addr;
4664 IXGBE_WRITE_REG(hw, IXGBE_VFTDBAL(i),
4665 (uint32_t)(bus_addr & 0x00000000ffffffffULL));
4666 IXGBE_WRITE_REG(hw, IXGBE_VFTDBAH(i),
4667 (uint32_t)(bus_addr >> 32));
4668 IXGBE_WRITE_REG(hw, IXGBE_VFTDLEN(i),
4669 txq->nb_tx_desc * sizeof(union ixgbe_adv_tx_desc));
4670 /* Setup the HW Tx Head and TX Tail descriptor pointers */
4671 IXGBE_WRITE_REG(hw, IXGBE_VFTDH(i), 0);
4672 IXGBE_WRITE_REG(hw, IXGBE_VFTDT(i), 0);
4675 * Disable Tx Head Writeback RO bit, since this hoses
4676 * bookkeeping if things aren't delivered in order.
4678 txctrl = IXGBE_READ_REG(hw,
4679 IXGBE_VFDCA_TXCTRL(i));
4680 txctrl &= ~IXGBE_DCA_TXCTRL_DESC_WRO_EN;
4681 IXGBE_WRITE_REG(hw, IXGBE_VFDCA_TXCTRL(i),
4687 * [VF] Start Transmit and Receive Units.
4689 void __attribute__((cold))
4690 ixgbevf_dev_rxtx_start(struct rte_eth_dev *dev)
4692 struct ixgbe_hw *hw;
4693 struct ixgbe_tx_queue *txq;
4694 struct ixgbe_rx_queue *rxq;
4700 PMD_INIT_FUNC_TRACE();
4701 hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
4703 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4704 txq = dev->data->tx_queues[i];
4705 /* Setup Transmit Threshold Registers */
4706 txdctl = IXGBE_READ_REG(hw, IXGBE_VFTXDCTL(i));
4707 txdctl |= txq->pthresh & 0x7F;
4708 txdctl |= ((txq->hthresh & 0x7F) << 8);
4709 txdctl |= ((txq->wthresh & 0x7F) << 16);
4710 IXGBE_WRITE_REG(hw, IXGBE_VFTXDCTL(i), txdctl);
4713 for (i = 0; i < dev->data->nb_tx_queues; i++) {
4715 txdctl = IXGBE_READ_REG(hw, IXGBE_VFTXDCTL(i));
4716 txdctl |= IXGBE_TXDCTL_ENABLE;
4717 IXGBE_WRITE_REG(hw, IXGBE_VFTXDCTL(i), txdctl);
4720 /* Wait until TX Enable ready */
4723 txdctl = IXGBE_READ_REG(hw, IXGBE_VFTXDCTL(i));
4724 } while (--poll_ms && !(txdctl & IXGBE_TXDCTL_ENABLE));
4726 PMD_INIT_LOG(ERR, "Could not enable Tx Queue %d", i);
4728 for (i = 0; i < dev->data->nb_rx_queues; i++) {
4730 rxq = dev->data->rx_queues[i];
4732 rxdctl = IXGBE_READ_REG(hw, IXGBE_VFRXDCTL(i));
4733 rxdctl |= IXGBE_RXDCTL_ENABLE;
4734 IXGBE_WRITE_REG(hw, IXGBE_VFRXDCTL(i), rxdctl);
4736 /* Wait until RX Enable ready */
4740 rxdctl = IXGBE_READ_REG(hw, IXGBE_VFRXDCTL(i));
4741 } while (--poll_ms && !(rxdctl & IXGBE_RXDCTL_ENABLE));
4743 PMD_INIT_LOG(ERR, "Could not enable Rx Queue %d", i);
4745 IXGBE_WRITE_REG(hw, IXGBE_VFRDT(i), rxq->nb_rx_desc - 1);
4750 /* Stubs needed for linkage when CONFIG_RTE_IXGBE_INC_VECTOR is set to 'n' */
4751 int __attribute__((weak))
4752 ixgbe_rx_vec_dev_conf_condition_check(struct rte_eth_dev __rte_unused *dev)
4757 uint16_t __attribute__((weak))
4758 ixgbe_recv_pkts_vec(
4759 void __rte_unused *rx_queue,
4760 struct rte_mbuf __rte_unused **rx_pkts,
4761 uint16_t __rte_unused nb_pkts)
4766 uint16_t __attribute__((weak))
4767 ixgbe_recv_scattered_pkts_vec(
4768 void __rte_unused *rx_queue,
4769 struct rte_mbuf __rte_unused **rx_pkts,
4770 uint16_t __rte_unused nb_pkts)
4775 int __attribute__((weak))
4776 ixgbe_rxq_vec_setup(struct ixgbe_rx_queue __rte_unused *rxq)