2 * Copyright (c) 2016 Solarflare Communications Inc.
5 * This software was jointly developed between OKTET Labs (under contract
6 * for Solarflare) and Solarflare Communications, Inc.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions are met:
11 * 1. Redistributions of source code must retain the above copyright notice,
12 * this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright notice,
14 * this list of conditions and the following disclaimer in the documentation
15 * and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
19 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
21 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
24 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
25 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
26 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
27 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30 #include <rte_mempool.h>
35 #include "sfc_debug.h"
39 #include "sfc_tweak.h"
42 * Maximum number of Rx queue flush attempt in the case of failure or
45 #define SFC_RX_QFLUSH_ATTEMPTS (3)
48 * Time to wait between event queue polling attempts when waiting for Rx
49 * queue flush done or failed events.
51 #define SFC_RX_QFLUSH_POLL_WAIT_MS (1)
54 * Maximum number of event queue polling attempts when waiting for Rx queue
55 * flush done or failed events. It defines Rx queue flush attempt timeout
56 * together with SFC_RX_QFLUSH_POLL_WAIT_MS.
58 #define SFC_RX_QFLUSH_POLL_ATTEMPTS (2000)
61 sfc_rx_qflush_done(struct sfc_rxq *rxq)
63 rxq->state |= SFC_RXQ_FLUSHED;
64 rxq->state &= ~SFC_RXQ_FLUSHING;
68 sfc_rx_qflush_failed(struct sfc_rxq *rxq)
70 rxq->state |= SFC_RXQ_FLUSH_FAILED;
71 rxq->state &= ~SFC_RXQ_FLUSHING;
75 sfc_rx_qrefill(struct sfc_rxq *rxq)
77 unsigned int free_space;
79 void *objs[SFC_RX_REFILL_BULK];
80 efsys_dma_addr_t addr[RTE_DIM(objs)];
81 unsigned int added = rxq->added;
84 struct sfc_rx_sw_desc *rxd;
86 uint8_t port_id = rxq->port_id;
88 free_space = EFX_RXQ_LIMIT(rxq->ptr_mask + 1) -
89 (added - rxq->completed);
90 bulks = free_space / RTE_DIM(objs);
92 id = added & rxq->ptr_mask;
94 if (rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
97 * It is hardly a safe way to increment counter
98 * from different contexts, but all PMDs do it.
100 rxq->evq->sa->eth_dev->data->rx_mbuf_alloc_failed +=
105 for (i = 0; i < RTE_DIM(objs);
106 ++i, id = (id + 1) & rxq->ptr_mask) {
109 rxd = &rxq->sw_desc[id];
112 rte_mbuf_refcnt_set(m, 1);
113 m->data_off = RTE_PKTMBUF_HEADROOM;
118 addr[i] = rte_pktmbuf_mtophys(m);
121 efx_rx_qpost(rxq->common, addr, rxq->buf_size,
122 RTE_DIM(objs), rxq->completed, added);
123 added += RTE_DIM(objs);
126 /* Push doorbell if something is posted */
127 if (rxq->added != added) {
129 efx_rx_qpush(rxq->common, added, &rxq->pushed);
134 sfc_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
136 struct sfc_rxq *rxq = rx_queue;
137 unsigned int completed;
138 unsigned int prefix_size = rxq->prefix_size;
139 unsigned int done_pkts = 0;
141 if (unlikely((rxq->state & SFC_RXQ_RUNNING) == 0))
144 sfc_ev_qpoll(rxq->evq);
146 completed = rxq->completed;
147 while (completed != rxq->pending && done_pkts < nb_pkts) {
149 struct sfc_rx_sw_desc *rxd;
151 unsigned int seg_len;
152 unsigned int desc_flags;
154 id = completed++ & rxq->ptr_mask;
155 rxd = &rxq->sw_desc[id];
157 desc_flags = rxd->flags;
159 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
162 if (desc_flags & EFX_PKT_PREFIX_LEN) {
166 rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
167 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
171 seg_len = rxd->size - prefix_size;
174 m->data_off += prefix_size;
175 rte_pktmbuf_data_len(m) = seg_len;
176 rte_pktmbuf_pkt_len(m) = seg_len;
178 m->packet_type = RTE_PTYPE_L2_ETHER;
185 rte_mempool_put(rxq->refill_mb_pool, m);
189 rxq->completed = completed;
197 sfc_rx_qpurge(struct sfc_rxq *rxq)
200 struct sfc_rx_sw_desc *rxd;
202 for (i = rxq->completed; i != rxq->added; ++i) {
203 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
204 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
210 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
213 unsigned int retry_count;
214 unsigned int wait_count;
216 rxq = sa->rxq_info[sw_index].rxq;
217 SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
220 * Retry Rx queue flushing in the case of flush failed or
221 * timeout. In the worst case it can delay for 6 seconds.
223 for (retry_count = 0;
224 ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
225 (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
227 if (efx_rx_qflush(rxq->common) != 0) {
228 rxq->state |= SFC_RXQ_FLUSH_FAILED;
231 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
232 rxq->state |= SFC_RXQ_FLUSHING;
235 * Wait for Rx queue flush done or failed event at least
236 * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
237 * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
238 * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
242 rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
243 sfc_ev_qpoll(rxq->evq);
244 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
245 (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
247 if (rxq->state & SFC_RXQ_FLUSHING)
248 sfc_err(sa, "RxQ %u flush timed out", sw_index);
250 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
251 sfc_err(sa, "RxQ %u flush failed", sw_index);
253 if (rxq->state & SFC_RXQ_FLUSHED)
254 sfc_info(sa, "RxQ %u flushed", sw_index);
261 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
263 struct sfc_rxq_info *rxq_info;
268 sfc_log_init(sa, "sw_index=%u", sw_index);
270 SFC_ASSERT(sw_index < sa->rxq_count);
272 rxq_info = &sa->rxq_info[sw_index];
274 SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
278 rc = sfc_ev_qstart(sa, evq->evq_index);
282 rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
283 &rxq->mem, rxq_info->entries,
284 0 /* not used on EF10 */, evq->common,
287 goto fail_rx_qcreate;
289 efx_rx_qenable(rxq->common);
291 rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
293 rxq->state |= (SFC_RXQ_STARTED | SFC_RXQ_RUNNING);
298 rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common,
301 goto fail_mac_filter_default_rxq_set;
304 /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
305 sa->eth_dev->data->rx_queue_state[sw_index] =
306 RTE_ETH_QUEUE_STATE_STARTED;
310 fail_mac_filter_default_rxq_set:
311 sfc_rx_qflush(sa, sw_index);
314 sfc_ev_qstop(sa, evq->evq_index);
321 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
323 struct sfc_rxq_info *rxq_info;
326 sfc_log_init(sa, "sw_index=%u", sw_index);
328 SFC_ASSERT(sw_index < sa->rxq_count);
330 rxq_info = &sa->rxq_info[sw_index];
332 SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
334 /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
335 sa->eth_dev->data->rx_queue_state[sw_index] =
336 RTE_ETH_QUEUE_STATE_STOPPED;
338 rxq->state &= ~SFC_RXQ_RUNNING;
341 efx_mac_filter_default_rxq_clear(sa->nic);
343 sfc_rx_qflush(sa, sw_index);
345 rxq->state = SFC_RXQ_INITIALIZED;
347 efx_rx_qdestroy(rxq->common);
349 sfc_ev_qstop(sa, rxq->evq->evq_index);
353 sfc_rx_qcheck_conf(struct sfc_adapter *sa,
354 const struct rte_eth_rxconf *rx_conf)
358 if (rx_conf->rx_thresh.pthresh != 0 ||
359 rx_conf->rx_thresh.hthresh != 0 ||
360 rx_conf->rx_thresh.wthresh != 0) {
362 "RxQ prefetch/host/writeback thresholds are not supported");
366 if (rx_conf->rx_free_thresh != 0) {
367 sfc_err(sa, "RxQ free threshold is not supported");
371 if (rx_conf->rx_drop_en == 0) {
372 sfc_err(sa, "RxQ drop disable is not supported");
376 if (rx_conf->rx_deferred_start != 0) {
377 sfc_err(sa, "RxQ deferred start is not supported");
385 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
390 /* The mbuf object itself is always cache line aligned */
391 order = rte_bsf32(RTE_CACHE_LINE_SIZE);
393 /* Data offset from mbuf object start */
394 data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
395 RTE_PKTMBUF_HEADROOM;
397 order = MIN(order, rte_bsf32(data_off));
399 return 1u << (order - 1);
403 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
405 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
406 const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
407 const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
409 unsigned int buf_aligned;
410 unsigned int start_alignment;
411 unsigned int end_padding_alignment;
413 /* Below it is assumed that both alignments are power of 2 */
414 SFC_ASSERT(rte_is_power_of_2(nic_align_start));
415 SFC_ASSERT(rte_is_power_of_2(nic_align_end));
418 * mbuf is always cache line aligned, double-check
419 * that it meets rx buffer start alignment requirements.
422 /* Start from mbuf pool data room size */
423 buf_size = rte_pktmbuf_data_room_size(mb_pool);
425 /* Remove headroom */
426 if (buf_size <= RTE_PKTMBUF_HEADROOM) {
428 "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
429 mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
432 buf_size -= RTE_PKTMBUF_HEADROOM;
434 /* Calculate guaranteed data start alignment */
435 buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
437 /* Reserve space for start alignment */
438 if (buf_aligned < nic_align_start) {
439 start_alignment = nic_align_start - buf_aligned;
440 if (buf_size <= start_alignment) {
442 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
444 rte_pktmbuf_data_room_size(mb_pool),
445 RTE_PKTMBUF_HEADROOM, start_alignment);
448 buf_aligned = nic_align_start;
449 buf_size -= start_alignment;
454 /* Make sure that end padding does not write beyond the buffer */
455 if (buf_aligned < nic_align_end) {
457 * Estimate space which can be lost. If guarnteed buffer
458 * size is odd, lost space is (nic_align_end - 1). More
459 * accurate formula is below.
461 end_padding_alignment = nic_align_end -
462 MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
463 if (buf_size <= end_padding_alignment) {
465 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u, buffer start alignment %u and end padding alignment %u required by NIC",
467 rte_pktmbuf_data_room_size(mb_pool),
468 RTE_PKTMBUF_HEADROOM, start_alignment,
469 end_padding_alignment);
472 buf_size -= end_padding_alignment;
475 * Start is aligned the same or better than end,
478 buf_size = P2ALIGN(buf_size, nic_align_end);
485 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
486 uint16_t nb_rx_desc, unsigned int socket_id,
487 const struct rte_eth_rxconf *rx_conf,
488 struct rte_mempool *mb_pool)
490 const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
493 struct sfc_rxq_info *rxq_info;
494 unsigned int evq_index;
498 rc = sfc_rx_qcheck_conf(sa, rx_conf);
502 buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
504 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
510 if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
511 !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
512 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
513 "object size is too small", sw_index);
514 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
515 "PDU size %u plus Rx prefix %u bytes",
516 sw_index, buf_size, (unsigned int)sa->port.pdu,
517 encp->enc_rx_prefix_size);
522 SFC_ASSERT(sw_index < sa->rxq_count);
523 rxq_info = &sa->rxq_info[sw_index];
525 SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
526 rxq_info->entries = nb_rx_desc;
527 rxq_info->type = EFX_RXQ_TYPE_DEFAULT;
529 evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
531 rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
535 evq = sa->evq_info[evq_index].evq;
538 rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
543 rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
544 socket_id, &rxq->mem);
549 rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
550 sizeof(*rxq->sw_desc),
551 RTE_CACHE_LINE_SIZE, socket_id);
552 if (rxq->sw_desc == NULL)
553 goto fail_desc_alloc;
557 rxq->ptr_mask = rxq_info->entries - 1;
558 rxq->refill_mb_pool = mb_pool;
559 rxq->buf_size = buf_size;
560 rxq->hw_index = sw_index;
561 rxq->port_id = sa->eth_dev->data->port_id;
563 /* Cache limits required on datapath in RxQ structure */
564 rxq->batch_max = encp->enc_rx_batch_max;
565 rxq->prefix_size = encp->enc_rx_prefix_size;
567 rxq->state = SFC_RXQ_INITIALIZED;
574 sfc_dma_free(sa, &rxq->mem);
580 sfc_ev_qfini(sa, evq_index);
583 rxq_info->entries = 0;
586 sfc_log_init(sa, "failed %d", rc);
591 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
593 struct sfc_rxq_info *rxq_info;
596 SFC_ASSERT(sw_index < sa->rxq_count);
598 rxq_info = &sa->rxq_info[sw_index];
601 SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
603 rxq_info->rxq = NULL;
604 rxq_info->entries = 0;
606 rte_free(rxq->sw_desc);
607 sfc_dma_free(sa, &rxq->mem);
612 sfc_rx_start(struct sfc_adapter *sa)
614 unsigned int sw_index;
617 sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
619 rc = efx_rx_init(sa->nic);
623 for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
624 rc = sfc_rx_qstart(sa, sw_index);
632 while (sw_index-- > 0)
633 sfc_rx_qstop(sa, sw_index);
635 efx_rx_fini(sa->nic);
638 sfc_log_init(sa, "failed %d", rc);
643 sfc_rx_stop(struct sfc_adapter *sa)
645 unsigned int sw_index;
647 sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
649 sw_index = sa->rxq_count;
650 while (sw_index-- > 0) {
651 if (sa->rxq_info[sw_index].rxq != NULL)
652 sfc_rx_qstop(sa, sw_index);
655 efx_rx_fini(sa->nic);
659 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
661 struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
662 unsigned int max_entries;
664 max_entries = EFX_RXQ_MAXNDESCS;
665 SFC_ASSERT(rte_is_power_of_2(max_entries));
667 rxq_info->max_entries = max_entries;
673 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
677 switch (rxmode->mq_mode) {
679 /* No special checks are required */
682 sfc_err(sa, "Rx multi-queue mode %u not supported",
687 if (rxmode->header_split) {
688 sfc_err(sa, "Header split on Rx not supported");
692 if (rxmode->hw_vlan_filter) {
693 sfc_err(sa, "HW VLAN filtering not supported");
697 if (rxmode->hw_vlan_strip) {
698 sfc_err(sa, "HW VLAN stripping not supported");
702 if (rxmode->hw_vlan_extend) {
704 "Q-in-Q HW VLAN stripping not supported");
708 if (!rxmode->hw_strip_crc) {
710 "FCS stripping control not supported - always stripped");
711 rxmode->hw_strip_crc = 1;
714 if (rxmode->enable_scatter) {
715 sfc_err(sa, "Scatter on Rx not supported");
719 if (rxmode->enable_lro) {
720 sfc_err(sa, "LRO not supported");
728 * Initialize Rx subsystem.
730 * Called at device configuration stage when number of receive queues is
731 * specified together with other device level receive configuration.
733 * It should be used to allocate NUMA-unaware resources.
736 sfc_rx_init(struct sfc_adapter *sa)
738 struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
739 unsigned int sw_index;
742 rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
744 goto fail_check_mode;
746 sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
749 sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
750 sizeof(struct sfc_rxq_info), 0,
752 if (sa->rxq_info == NULL)
753 goto fail_rxqs_alloc;
755 for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
756 rc = sfc_rx_qinit_info(sa, sw_index);
758 goto fail_rx_qinit_info;
764 rte_free(sa->rxq_info);
770 sfc_log_init(sa, "failed %d", rc);
775 * Shutdown Rx subsystem.
777 * Called at device close stage, for example, before device
778 * reconfiguration or shutdown.
781 sfc_rx_fini(struct sfc_adapter *sa)
783 unsigned int sw_index;
785 sw_index = sa->rxq_count;
786 while (sw_index-- > 0) {
787 if (sa->rxq_info[sw_index].rxq != NULL)
788 sfc_rx_qfini(sa, sw_index);
791 rte_free(sa->rxq_info);