1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2019, Intel Corporation. */
4 #include <linux/bpf_trace.h>
5 #include <net/xdp_sock_drv.h>
12 #include "ice_txrx_lib.h"
15 static struct xdp_buff **ice_xdp_buf(struct ice_rx_ring *rx_ring, u32 idx)
17 return &rx_ring->xdp_buf[idx];
21 * ice_qp_reset_stats - Resets all stats for rings of given index
22 * @vsi: VSI that contains rings of interest
23 * @q_idx: ring index in array
25 static void ice_qp_reset_stats(struct ice_vsi *vsi, u16 q_idx)
27 memset(&vsi->rx_rings[q_idx]->rx_stats, 0,
28 sizeof(vsi->rx_rings[q_idx]->rx_stats));
29 memset(&vsi->tx_rings[q_idx]->stats, 0,
30 sizeof(vsi->tx_rings[q_idx]->stats));
31 if (ice_is_xdp_ena_vsi(vsi))
32 memset(&vsi->xdp_rings[q_idx]->stats, 0,
33 sizeof(vsi->xdp_rings[q_idx]->stats));
37 * ice_qp_clean_rings - Cleans all the rings of a given index
38 * @vsi: VSI that contains rings of interest
39 * @q_idx: ring index in array
41 static void ice_qp_clean_rings(struct ice_vsi *vsi, u16 q_idx)
43 ice_clean_tx_ring(vsi->tx_rings[q_idx]);
44 if (ice_is_xdp_ena_vsi(vsi))
45 ice_clean_tx_ring(vsi->xdp_rings[q_idx]);
46 ice_clean_rx_ring(vsi->rx_rings[q_idx]);
50 * ice_qvec_toggle_napi - Enables/disables NAPI for a given q_vector
51 * @vsi: VSI that has netdev
52 * @q_vector: q_vector that has NAPI context
53 * @enable: true for enable, false for disable
56 ice_qvec_toggle_napi(struct ice_vsi *vsi, struct ice_q_vector *q_vector,
59 if (!vsi->netdev || !q_vector)
63 napi_enable(&q_vector->napi);
65 napi_disable(&q_vector->napi);
69 * ice_qvec_dis_irq - Mask off queue interrupt generation on given ring
70 * @vsi: the VSI that contains queue vector being un-configured
71 * @rx_ring: Rx ring that will have its IRQ disabled
72 * @q_vector: queue vector
75 ice_qvec_dis_irq(struct ice_vsi *vsi, struct ice_rx_ring *rx_ring,
76 struct ice_q_vector *q_vector)
78 struct ice_pf *pf = vsi->back;
79 struct ice_hw *hw = &pf->hw;
80 int base = vsi->base_vector;
84 /* QINT_TQCTL is being cleared in ice_vsi_stop_tx_ring, so handle
85 * here only QINT_RQCTL
87 reg = rx_ring->reg_idx;
88 val = rd32(hw, QINT_RQCTL(reg));
89 val &= ~QINT_RQCTL_CAUSE_ENA_M;
90 wr32(hw, QINT_RQCTL(reg), val);
93 u16 v_idx = q_vector->v_idx;
95 wr32(hw, GLINT_DYN_CTL(q_vector->reg_idx), 0);
97 synchronize_irq(pf->msix_entries[v_idx + base].vector);
102 * ice_qvec_cfg_msix - Enable IRQ for given queue vector
103 * @vsi: the VSI that contains queue vector
104 * @q_vector: queue vector
107 ice_qvec_cfg_msix(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
109 u16 reg_idx = q_vector->reg_idx;
110 struct ice_pf *pf = vsi->back;
111 struct ice_hw *hw = &pf->hw;
112 struct ice_tx_ring *tx_ring;
113 struct ice_rx_ring *rx_ring;
115 ice_cfg_itr(hw, q_vector);
117 ice_for_each_tx_ring(tx_ring, q_vector->tx)
118 ice_cfg_txq_interrupt(vsi, tx_ring->reg_idx, reg_idx,
119 q_vector->tx.itr_idx);
121 ice_for_each_rx_ring(rx_ring, q_vector->rx)
122 ice_cfg_rxq_interrupt(vsi, rx_ring->reg_idx, reg_idx,
123 q_vector->rx.itr_idx);
129 * ice_qvec_ena_irq - Enable IRQ for given queue vector
130 * @vsi: the VSI that contains queue vector
131 * @q_vector: queue vector
133 static void ice_qvec_ena_irq(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
135 struct ice_pf *pf = vsi->back;
136 struct ice_hw *hw = &pf->hw;
138 ice_irq_dynamic_ena(hw, vsi, q_vector);
144 * ice_qp_dis - Disables a queue pair
145 * @vsi: VSI of interest
146 * @q_idx: ring index in array
148 * Returns 0 on success, negative on failure.
150 static int ice_qp_dis(struct ice_vsi *vsi, u16 q_idx)
152 struct ice_txq_meta txq_meta = { };
153 struct ice_q_vector *q_vector;
154 struct ice_tx_ring *tx_ring;
155 struct ice_rx_ring *rx_ring;
159 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
162 tx_ring = vsi->tx_rings[q_idx];
163 rx_ring = vsi->rx_rings[q_idx];
164 q_vector = rx_ring->q_vector;
166 while (test_and_set_bit(ICE_CFG_BUSY, vsi->state)) {
170 usleep_range(1000, 2000);
172 netif_tx_stop_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
174 ice_qvec_dis_irq(vsi, rx_ring, q_vector);
176 ice_fill_txq_meta(vsi, tx_ring, &txq_meta);
177 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, tx_ring, &txq_meta);
180 if (ice_is_xdp_ena_vsi(vsi)) {
181 struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
183 memset(&txq_meta, 0, sizeof(txq_meta));
184 ice_fill_txq_meta(vsi, xdp_ring, &txq_meta);
185 err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, xdp_ring,
190 err = ice_vsi_ctrl_one_rx_ring(vsi, false, q_idx, true);
194 ice_qvec_toggle_napi(vsi, q_vector, false);
195 ice_qp_clean_rings(vsi, q_idx);
196 ice_qp_reset_stats(vsi, q_idx);
202 * ice_qp_ena - Enables a queue pair
203 * @vsi: VSI of interest
204 * @q_idx: ring index in array
206 * Returns 0 on success, negative on failure.
208 static int ice_qp_ena(struct ice_vsi *vsi, u16 q_idx)
210 struct ice_aqc_add_tx_qgrp *qg_buf;
211 struct ice_q_vector *q_vector;
212 struct ice_tx_ring *tx_ring;
213 struct ice_rx_ring *rx_ring;
217 if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
220 size = struct_size(qg_buf, txqs, 1);
221 qg_buf = kzalloc(size, GFP_KERNEL);
225 qg_buf->num_txqs = 1;
227 tx_ring = vsi->tx_rings[q_idx];
228 rx_ring = vsi->rx_rings[q_idx];
229 q_vector = rx_ring->q_vector;
231 err = ice_vsi_cfg_txq(vsi, tx_ring, qg_buf);
235 if (ice_is_xdp_ena_vsi(vsi)) {
236 struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
238 memset(qg_buf, 0, size);
239 qg_buf->num_txqs = 1;
240 err = ice_vsi_cfg_txq(vsi, xdp_ring, qg_buf);
243 ice_set_ring_xdp(xdp_ring);
244 xdp_ring->xsk_pool = ice_tx_xsk_pool(xdp_ring);
247 err = ice_vsi_cfg_rxq(rx_ring);
251 ice_qvec_cfg_msix(vsi, q_vector);
253 err = ice_vsi_ctrl_one_rx_ring(vsi, true, q_idx, true);
257 clear_bit(ICE_CFG_BUSY, vsi->state);
258 ice_qvec_toggle_napi(vsi, q_vector, true);
259 ice_qvec_ena_irq(vsi, q_vector);
261 netif_tx_start_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
268 * ice_xsk_pool_disable - disable a buffer pool region
272 * Returns 0 on success, negative on failure
274 static int ice_xsk_pool_disable(struct ice_vsi *vsi, u16 qid)
276 struct xsk_buff_pool *pool = xsk_get_pool_from_qid(vsi->netdev, qid);
281 clear_bit(qid, vsi->af_xdp_zc_qps);
282 xsk_pool_dma_unmap(pool, ICE_RX_DMA_ATTR);
288 * ice_xsk_pool_enable - enable a buffer pool region
290 * @pool: pointer to a requested buffer pool region
293 * Returns 0 on success, negative on failure
296 ice_xsk_pool_enable(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
300 if (vsi->type != ICE_VSI_PF)
303 if (qid >= vsi->netdev->real_num_rx_queues ||
304 qid >= vsi->netdev->real_num_tx_queues)
307 err = xsk_pool_dma_map(pool, ice_pf_to_dev(vsi->back),
312 set_bit(qid, vsi->af_xdp_zc_qps);
318 * ice_xsk_pool_setup - enable/disable a buffer pool region depending on its state
320 * @pool: buffer pool to enable/associate to a ring, NULL to disable
323 * Returns 0 on success, negative on failure
325 int ice_xsk_pool_setup(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
327 bool if_running, pool_present = !!pool;
328 int ret = 0, pool_failure = 0;
330 if (!is_power_of_2(vsi->rx_rings[qid]->count) ||
331 !is_power_of_2(vsi->tx_rings[qid]->count)) {
332 netdev_err(vsi->netdev, "Please align ring sizes to power of 2\n");
333 pool_failure = -EINVAL;
337 if_running = netif_running(vsi->netdev) && ice_is_xdp_ena_vsi(vsi);
340 ret = ice_qp_dis(vsi, qid);
342 netdev_err(vsi->netdev, "ice_qp_dis error = %d\n", ret);
347 pool_failure = pool_present ? ice_xsk_pool_enable(vsi, pool, qid) :
348 ice_xsk_pool_disable(vsi, qid);
352 ret = ice_qp_ena(vsi, qid);
353 if (!ret && pool_present)
354 napi_schedule(&vsi->xdp_rings[qid]->q_vector->napi);
356 netdev_err(vsi->netdev, "ice_qp_ena error = %d\n", ret);
361 netdev_err(vsi->netdev, "Could not %sable buffer pool, error = %d\n",
362 pool_present ? "en" : "dis", pool_failure);
370 * ice_fill_rx_descs - pick buffers from XSK buffer pool and use it
371 * @pool: XSK Buffer pool to pull the buffers from
372 * @xdp: SW ring of xdp_buff that will hold the buffers
373 * @rx_desc: Pointer to Rx descriptors that will be filled
374 * @count: The number of buffers to allocate
376 * This function allocates a number of Rx buffers from the fill ring
377 * or the internal recycle mechanism and places them on the Rx ring.
379 * Note that ring wrap should be handled by caller of this function.
381 * Returns the amount of allocated Rx descriptors
383 static u16 ice_fill_rx_descs(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
384 union ice_32b_rx_flex_desc *rx_desc, u16 count)
390 buffs = xsk_buff_alloc_batch(pool, xdp, count);
391 for (i = 0; i < buffs; i++) {
392 dma = xsk_buff_xdp_get_dma(*xdp);
393 rx_desc->read.pkt_addr = cpu_to_le64(dma);
394 rx_desc->wb.status_error0 = 0;
404 * __ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
406 * @count: The number of buffers to allocate
408 * Place the @count of descriptors onto Rx ring. Handle the ring wrap
409 * for case where space from next_to_use up to the end of ring is less
410 * than @count. Finally do a tail bump.
412 * Returns true if all allocations were successful, false if any fail.
414 static bool __ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
416 union ice_32b_rx_flex_desc *rx_desc;
417 u32 nb_buffs_extra = 0, nb_buffs;
418 u16 ntu = rx_ring->next_to_use;
419 u16 total_count = count;
420 struct xdp_buff **xdp;
422 rx_desc = ICE_RX_DESC(rx_ring, ntu);
423 xdp = ice_xdp_buf(rx_ring, ntu);
425 if (ntu + count >= rx_ring->count) {
426 nb_buffs_extra = ice_fill_rx_descs(rx_ring->xsk_pool, xdp,
428 rx_ring->count - ntu);
429 rx_desc = ICE_RX_DESC(rx_ring, 0);
430 xdp = ice_xdp_buf(rx_ring, 0);
432 count -= nb_buffs_extra;
433 ice_release_rx_desc(rx_ring, 0);
436 nb_buffs = ice_fill_rx_descs(rx_ring->xsk_pool, xdp, rx_desc, count);
439 if (ntu == rx_ring->count)
442 if (rx_ring->next_to_use != ntu)
443 ice_release_rx_desc(rx_ring, ntu);
445 return total_count == (nb_buffs_extra + nb_buffs);
449 * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
451 * @count: The number of buffers to allocate
453 * Wrapper for internal allocation routine; figure out how many tail
454 * bumps should take place based on the given threshold
456 * Returns true if all calls to internal alloc routine succeeded
458 bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
460 u16 rx_thresh = ICE_RING_QUARTER(rx_ring);
461 u16 batched, leftover, i, tail_bumps;
463 batched = ALIGN_DOWN(count, rx_thresh);
464 tail_bumps = batched / rx_thresh;
465 leftover = count & (rx_thresh - 1);
467 for (i = 0; i < tail_bumps; i++)
468 if (!__ice_alloc_rx_bufs_zc(rx_ring, rx_thresh))
470 return __ice_alloc_rx_bufs_zc(rx_ring, leftover);
474 * ice_bump_ntc - Bump the next_to_clean counter of an Rx ring
477 static void ice_bump_ntc(struct ice_rx_ring *rx_ring)
479 int ntc = rx_ring->next_to_clean + 1;
481 ntc = (ntc < rx_ring->count) ? ntc : 0;
482 rx_ring->next_to_clean = ntc;
483 prefetch(ICE_RX_DESC(rx_ring, ntc));
487 * ice_construct_skb_zc - Create an sk_buff from zero-copy buffer
489 * @xdp: Pointer to XDP buffer
491 * This function allocates a new skb from a zero-copy Rx buffer.
493 * Returns the skb on success, NULL on failure.
495 static struct sk_buff *
496 ice_construct_skb_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp)
498 unsigned int totalsize = xdp->data_end - xdp->data_meta;
499 unsigned int metasize = xdp->data - xdp->data_meta;
502 net_prefetch(xdp->data_meta);
504 skb = __napi_alloc_skb(&rx_ring->q_vector->napi, totalsize,
505 GFP_ATOMIC | __GFP_NOWARN);
509 memcpy(__skb_put(skb, totalsize), xdp->data_meta,
510 ALIGN(totalsize, sizeof(long)));
513 skb_metadata_set(skb, metasize);
514 __skb_pull(skb, metasize);
522 * ice_run_xdp_zc - Executes an XDP program in zero-copy path
524 * @xdp: xdp_buff used as input to the XDP program
525 * @xdp_prog: XDP program to run
526 * @xdp_ring: ring to be used for XDP_TX action
528 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
531 ice_run_xdp_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
532 struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring)
534 int err, result = ICE_XDP_PASS;
537 act = bpf_prog_run_xdp(xdp_prog, xdp);
539 if (likely(act == XDP_REDIRECT)) {
540 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
543 return ICE_XDP_REDIR;
550 result = ice_xmit_xdp_buff(xdp, xdp_ring);
551 if (result == ICE_XDP_CONSUMED)
555 bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
559 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
562 result = ICE_XDP_CONSUMED;
570 * ice_clean_rx_irq_zc - consumes packets from the hardware ring
571 * @rx_ring: AF_XDP Rx ring
572 * @budget: NAPI budget
574 * Returns number of processed packets on success, remaining budget on failure.
576 int ice_clean_rx_irq_zc(struct ice_rx_ring *rx_ring, int budget)
578 unsigned int total_rx_bytes = 0, total_rx_packets = 0;
579 struct ice_tx_ring *xdp_ring;
580 unsigned int xdp_xmit = 0;
581 struct bpf_prog *xdp_prog;
582 bool failure = false;
584 /* ZC patch is enabled only when XDP program is set,
585 * so here it can not be NULL
587 xdp_prog = READ_ONCE(rx_ring->xdp_prog);
588 xdp_ring = rx_ring->xdp_ring;
590 while (likely(total_rx_packets < (unsigned int)budget)) {
591 union ice_32b_rx_flex_desc *rx_desc;
592 unsigned int size, xdp_res = 0;
593 struct xdp_buff *xdp;
599 rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean);
601 stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
602 if (!ice_test_staterr(rx_desc->wb.status_error0, stat_err_bits))
605 /* This memory barrier is needed to keep us from reading
606 * any other fields out of the rx_desc until we have
607 * verified the descriptor has been written back.
611 xdp = *ice_xdp_buf(rx_ring, rx_ring->next_to_clean);
613 size = le16_to_cpu(rx_desc->wb.pkt_len) &
614 ICE_RX_FLX_DESC_PKT_LEN_M;
617 xdp->data_end = NULL;
618 xdp->data_hard_start = NULL;
619 xdp->data_meta = NULL;
623 xsk_buff_set_size(xdp, size);
624 xsk_buff_dma_sync_for_cpu(xdp, rx_ring->xsk_pool);
626 xdp_res = ice_run_xdp_zc(rx_ring, xdp, xdp_prog, xdp_ring);
628 if (xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR))
633 total_rx_bytes += size;
636 ice_bump_ntc(rx_ring);
641 skb = ice_construct_skb_zc(rx_ring, xdp);
643 rx_ring->rx_stats.alloc_buf_failed++;
647 ice_bump_ntc(rx_ring);
649 if (eth_skb_pad(skb)) {
654 total_rx_bytes += skb->len;
657 vlan_tag = ice_get_vlan_tag_from_rx_desc(rx_desc);
659 rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) &
660 ICE_RX_FLEX_DESC_PTYPE_M;
662 ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
663 ice_receive_skb(rx_ring, skb, vlan_tag);
666 failure = !ice_alloc_rx_bufs_zc(rx_ring, ICE_DESC_UNUSED(rx_ring));
668 ice_finalize_xdp_rx(xdp_ring, xdp_xmit);
669 ice_update_rx_ring_stats(rx_ring, total_rx_packets, total_rx_bytes);
671 if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
672 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
673 xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
675 xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
677 return (int)total_rx_packets;
680 return failure ? budget : (int)total_rx_packets;
684 * ice_clean_xdp_tx_buf - Free and unmap XDP Tx buffer
685 * @xdp_ring: XDP Tx ring
686 * @tx_buf: Tx buffer to clean
689 ice_clean_xdp_tx_buf(struct ice_tx_ring *xdp_ring, struct ice_tx_buf *tx_buf)
691 xdp_return_frame((struct xdp_frame *)tx_buf->raw_buf);
692 xdp_ring->xdp_tx_active--;
693 dma_unmap_single(xdp_ring->dev, dma_unmap_addr(tx_buf, dma),
694 dma_unmap_len(tx_buf, len), DMA_TO_DEVICE);
695 dma_unmap_len_set(tx_buf, len, 0);
699 * ice_clean_xdp_irq_zc - Reclaim resources after transmit completes on XDP ring
700 * @xdp_ring: XDP ring to clean
701 * @napi_budget: amount of descriptors that NAPI allows us to clean
703 * Returns count of cleaned descriptors
705 static u16 ice_clean_xdp_irq_zc(struct ice_tx_ring *xdp_ring, int napi_budget)
707 u16 tx_thresh = ICE_RING_QUARTER(xdp_ring);
708 int budget = napi_budget / tx_thresh;
709 u16 next_dd = xdp_ring->next_dd;
710 u16 ntc, cleared_dds = 0;
713 struct ice_tx_desc *next_dd_desc;
714 u16 desc_cnt = xdp_ring->count;
715 struct ice_tx_buf *tx_buf;
719 next_dd_desc = ICE_TX_DESC(xdp_ring, next_dd);
720 if (!(next_dd_desc->cmd_type_offset_bsz &
721 cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)))
726 if (likely(!xdp_ring->xdp_tx_active)) {
727 xsk_frames = tx_thresh;
731 ntc = xdp_ring->next_to_clean;
733 for (i = 0; i < tx_thresh; i++) {
734 tx_buf = &xdp_ring->tx_buf[ntc];
736 if (tx_buf->raw_buf) {
737 ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
738 tx_buf->raw_buf = NULL;
744 if (ntc >= xdp_ring->count)
748 xdp_ring->next_to_clean += tx_thresh;
749 if (xdp_ring->next_to_clean >= desc_cnt)
750 xdp_ring->next_to_clean -= desc_cnt;
752 xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
753 next_dd_desc->cmd_type_offset_bsz = 0;
754 next_dd = next_dd + tx_thresh;
755 if (next_dd >= desc_cnt)
756 next_dd = tx_thresh - 1;
759 xdp_ring->next_dd = next_dd;
761 return cleared_dds * tx_thresh;
765 * ice_xmit_pkt - produce a single HW Tx descriptor out of AF_XDP descriptor
766 * @xdp_ring: XDP ring to produce the HW Tx descriptor on
767 * @desc: AF_XDP descriptor to pull the DMA address and length from
768 * @total_bytes: bytes accumulator that will be used for stats update
770 static void ice_xmit_pkt(struct ice_tx_ring *xdp_ring, struct xdp_desc *desc,
771 unsigned int *total_bytes)
773 struct ice_tx_desc *tx_desc;
776 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
777 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
779 tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
780 tx_desc->buf_addr = cpu_to_le64(dma);
781 tx_desc->cmd_type_offset_bsz = ice_build_ctob(ICE_TX_DESC_CMD_EOP,
784 *total_bytes += desc->len;
788 * ice_xmit_pkt_batch - produce a batch of HW Tx descriptors out of AF_XDP descriptors
789 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
790 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
791 * @total_bytes: bytes accumulator that will be used for stats update
793 static void ice_xmit_pkt_batch(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
794 unsigned int *total_bytes)
796 u16 tx_thresh = ICE_RING_QUARTER(xdp_ring);
797 u16 ntu = xdp_ring->next_to_use;
798 struct ice_tx_desc *tx_desc;
801 loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
804 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, descs[i].addr);
805 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, descs[i].len);
807 tx_desc = ICE_TX_DESC(xdp_ring, ntu++);
808 tx_desc->buf_addr = cpu_to_le64(dma);
809 tx_desc->cmd_type_offset_bsz = ice_build_ctob(ICE_TX_DESC_CMD_EOP,
812 *total_bytes += descs[i].len;
815 xdp_ring->next_to_use = ntu;
817 if (xdp_ring->next_to_use > xdp_ring->next_rs) {
818 tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_rs);
819 tx_desc->cmd_type_offset_bsz |=
820 cpu_to_le64(ICE_TX_DESC_CMD_RS << ICE_TXD_QW1_CMD_S);
821 xdp_ring->next_rs += tx_thresh;
826 * ice_fill_tx_hw_ring - produce the number of Tx descriptors onto ring
827 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
828 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
829 * @nb_pkts: count of packets to be send
830 * @total_bytes: bytes accumulator that will be used for stats update
832 static void ice_fill_tx_hw_ring(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
833 u32 nb_pkts, unsigned int *total_bytes)
835 u16 tx_thresh = ICE_RING_QUARTER(xdp_ring);
836 u32 batched, leftover, i;
838 batched = ALIGN_DOWN(nb_pkts, PKTS_PER_BATCH);
839 leftover = nb_pkts & (PKTS_PER_BATCH - 1);
840 for (i = 0; i < batched; i += PKTS_PER_BATCH)
841 ice_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
842 for (; i < batched + leftover; i++)
843 ice_xmit_pkt(xdp_ring, &descs[i], total_bytes);
845 if (xdp_ring->next_to_use > xdp_ring->next_rs) {
846 struct ice_tx_desc *tx_desc;
848 tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_rs);
849 tx_desc->cmd_type_offset_bsz |=
850 cpu_to_le64(ICE_TX_DESC_CMD_RS << ICE_TXD_QW1_CMD_S);
851 xdp_ring->next_rs += tx_thresh;
856 * ice_xmit_zc - take entries from XSK Tx ring and place them onto HW Tx ring
857 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
858 * @budget: number of free descriptors on HW Tx ring that can be used
859 * @napi_budget: amount of descriptors that NAPI allows us to clean
861 * Returns true if there is no more work that needs to be done, false otherwise
863 bool ice_xmit_zc(struct ice_tx_ring *xdp_ring, u32 budget, int napi_budget)
865 struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
866 u16 tx_thresh = ICE_RING_QUARTER(xdp_ring);
867 u32 nb_pkts, nb_processed = 0;
868 unsigned int total_bytes = 0;
870 if (budget < tx_thresh)
871 budget += ice_clean_xdp_irq_zc(xdp_ring, napi_budget);
873 nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
877 if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
878 struct ice_tx_desc *tx_desc;
880 nb_processed = xdp_ring->count - xdp_ring->next_to_use;
881 ice_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
882 tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_rs);
883 tx_desc->cmd_type_offset_bsz |=
884 cpu_to_le64(ICE_TX_DESC_CMD_RS << ICE_TXD_QW1_CMD_S);
885 xdp_ring->next_rs = tx_thresh - 1;
886 xdp_ring->next_to_use = 0;
889 ice_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
892 ice_xdp_ring_update_tail(xdp_ring);
893 ice_update_tx_ring_stats(xdp_ring, nb_pkts, total_bytes);
895 if (xsk_uses_need_wakeup(xdp_ring->xsk_pool))
896 xsk_set_tx_need_wakeup(xdp_ring->xsk_pool);
898 return nb_pkts < budget;
902 * ice_xsk_wakeup - Implements ndo_xsk_wakeup
903 * @netdev: net_device
904 * @queue_id: queue to wake up
905 * @flags: ignored in our case, since we have Rx and Tx in the same NAPI
907 * Returns negative on error, zero otherwise.
910 ice_xsk_wakeup(struct net_device *netdev, u32 queue_id,
911 u32 __always_unused flags)
913 struct ice_netdev_priv *np = netdev_priv(netdev);
914 struct ice_q_vector *q_vector;
915 struct ice_vsi *vsi = np->vsi;
916 struct ice_tx_ring *ring;
918 if (test_bit(ICE_DOWN, vsi->state))
921 if (!ice_is_xdp_ena_vsi(vsi))
924 if (queue_id >= vsi->num_txq)
927 if (!vsi->xdp_rings[queue_id]->xsk_pool)
930 ring = vsi->xdp_rings[queue_id];
932 /* The idea here is that if NAPI is running, mark a miss, so
933 * it will run again. If not, trigger an interrupt and
934 * schedule the NAPI from interrupt context. If NAPI would be
935 * scheduled here, the interrupt affinity would not be
938 q_vector = ring->q_vector;
939 if (!napi_if_scheduled_mark_missed(&q_vector->napi))
940 ice_trigger_sw_intr(&vsi->back->hw, q_vector);
946 * ice_xsk_any_rx_ring_ena - Checks if Rx rings have AF_XDP buff pool attached
947 * @vsi: VSI to be checked
949 * Returns true if any of the Rx rings has an AF_XDP buff pool attached
951 bool ice_xsk_any_rx_ring_ena(struct ice_vsi *vsi)
955 ice_for_each_rxq(vsi, i) {
956 if (xsk_get_pool_from_qid(vsi->netdev, i))
964 * ice_xsk_clean_rx_ring - clean buffer pool queues connected to a given Rx ring
965 * @rx_ring: ring to be cleaned
967 void ice_xsk_clean_rx_ring(struct ice_rx_ring *rx_ring)
969 u16 count_mask = rx_ring->count - 1;
970 u16 ntc = rx_ring->next_to_clean;
971 u16 ntu = rx_ring->next_to_use;
973 for ( ; ntc != ntu; ntc = (ntc + 1) & count_mask) {
974 struct xdp_buff *xdp = *ice_xdp_buf(rx_ring, ntc);
981 * ice_xsk_clean_xdp_ring - Clean the XDP Tx ring and its buffer pool queues
982 * @xdp_ring: XDP_Tx ring
984 void ice_xsk_clean_xdp_ring(struct ice_tx_ring *xdp_ring)
986 u16 ntc = xdp_ring->next_to_clean, ntu = xdp_ring->next_to_use;
990 struct ice_tx_buf *tx_buf = &xdp_ring->tx_buf[ntc];
993 ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
997 tx_buf->raw_buf = NULL;
1000 if (ntc >= xdp_ring->count)
1005 xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);