Linux Audio

Check our new training course

Loading...
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0
   2/* Copyright (c) 2019, Intel Corporation. */
   3
   4#include <linux/bpf_trace.h>
   5#include <net/xdp_sock_drv.h>
   6#include <net/xdp.h>
   7#include "ice.h"
   8#include "ice_base.h"
   9#include "ice_type.h"
  10#include "ice_xsk.h"
  11#include "ice_txrx.h"
  12#include "ice_txrx_lib.h"
  13#include "ice_lib.h"
  14
  15static struct xdp_buff **ice_xdp_buf(struct ice_rx_ring *rx_ring, u32 idx)
  16{
  17	return &rx_ring->xdp_buf[idx];
  18}
  19
  20/**
  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
  24 */
  25static void ice_qp_reset_stats(struct ice_vsi *vsi, u16 q_idx)
  26{
  27	struct ice_vsi_stats *vsi_stat;
  28	struct ice_pf *pf;
  29
  30	pf = vsi->back;
  31	if (!pf->vsi_stats)
  32		return;
  33
  34	vsi_stat = pf->vsi_stats[vsi->idx];
  35	if (!vsi_stat)
  36		return;
  37
  38	memset(&vsi_stat->rx_ring_stats[q_idx]->rx_stats, 0,
  39	       sizeof(vsi_stat->rx_ring_stats[q_idx]->rx_stats));
  40	memset(&vsi_stat->tx_ring_stats[q_idx]->stats, 0,
  41	       sizeof(vsi_stat->tx_ring_stats[q_idx]->stats));
  42	if (vsi->xdp_rings)
  43		memset(&vsi->xdp_rings[q_idx]->ring_stats->stats, 0,
  44		       sizeof(vsi->xdp_rings[q_idx]->ring_stats->stats));
  45}
  46
  47/**
  48 * ice_qp_clean_rings - Cleans all the rings of a given index
  49 * @vsi: VSI that contains rings of interest
  50 * @q_idx: ring index in array
  51 */
  52static void ice_qp_clean_rings(struct ice_vsi *vsi, u16 q_idx)
  53{
  54	ice_clean_tx_ring(vsi->tx_rings[q_idx]);
  55	if (vsi->xdp_rings)
 
  56		ice_clean_tx_ring(vsi->xdp_rings[q_idx]);
 
  57	ice_clean_rx_ring(vsi->rx_rings[q_idx]);
  58}
  59
  60/**
  61 * ice_qvec_toggle_napi - Enables/disables NAPI for a given q_vector
  62 * @vsi: VSI that has netdev
  63 * @q_vector: q_vector that has NAPI context
  64 * @enable: true for enable, false for disable
  65 */
  66static void
  67ice_qvec_toggle_napi(struct ice_vsi *vsi, struct ice_q_vector *q_vector,
  68		     bool enable)
  69{
  70	if (!vsi->netdev || !q_vector)
  71		return;
  72
  73	if (enable)
  74		napi_enable(&q_vector->napi);
  75	else
  76		napi_disable(&q_vector->napi);
  77}
  78
  79/**
  80 * ice_qvec_dis_irq - Mask off queue interrupt generation on given ring
  81 * @vsi: the VSI that contains queue vector being un-configured
  82 * @rx_ring: Rx ring that will have its IRQ disabled
  83 * @q_vector: queue vector
  84 */
  85static void
  86ice_qvec_dis_irq(struct ice_vsi *vsi, struct ice_rx_ring *rx_ring,
  87		 struct ice_q_vector *q_vector)
  88{
  89	struct ice_pf *pf = vsi->back;
  90	struct ice_hw *hw = &pf->hw;
 
  91	u16 reg;
  92	u32 val;
  93
  94	/* QINT_TQCTL is being cleared in ice_vsi_stop_tx_ring, so handle
  95	 * here only QINT_RQCTL
  96	 */
  97	reg = rx_ring->reg_idx;
  98	val = rd32(hw, QINT_RQCTL(reg));
  99	val &= ~QINT_RQCTL_CAUSE_ENA_M;
 100	wr32(hw, QINT_RQCTL(reg), val);
 101
 102	if (q_vector) {
 
 
 103		wr32(hw, GLINT_DYN_CTL(q_vector->reg_idx), 0);
 104		ice_flush(hw);
 105		synchronize_irq(q_vector->irq.virq);
 106	}
 107}
 108
 109/**
 110 * ice_qvec_cfg_msix - Enable IRQ for given queue vector
 111 * @vsi: the VSI that contains queue vector
 112 * @q_vector: queue vector
 113 * @qid: queue index
 114 */
 115static void
 116ice_qvec_cfg_msix(struct ice_vsi *vsi, struct ice_q_vector *q_vector, u16 qid)
 117{
 118	u16 reg_idx = q_vector->reg_idx;
 119	struct ice_pf *pf = vsi->back;
 120	struct ice_hw *hw = &pf->hw;
 121	int q, _qid = qid;
 
 122
 123	ice_cfg_itr(hw, q_vector);
 124
 125	for (q = 0; q < q_vector->num_ring_tx; q++) {
 126		ice_cfg_txq_interrupt(vsi, _qid, reg_idx, q_vector->tx.itr_idx);
 127		_qid++;
 128	}
 129
 130	_qid = qid;
 131
 132	for (q = 0; q < q_vector->num_ring_rx; q++) {
 133		ice_cfg_rxq_interrupt(vsi, _qid, reg_idx, q_vector->rx.itr_idx);
 134		_qid++;
 135	}
 136
 137	ice_flush(hw);
 138}
 139
 140/**
 141 * ice_qvec_ena_irq - Enable IRQ for given queue vector
 142 * @vsi: the VSI that contains queue vector
 143 * @q_vector: queue vector
 144 */
 145static void ice_qvec_ena_irq(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
 146{
 147	struct ice_pf *pf = vsi->back;
 148	struct ice_hw *hw = &pf->hw;
 149
 150	ice_irq_dynamic_ena(hw, vsi, q_vector);
 151
 152	ice_flush(hw);
 153}
 154
 155/**
 156 * ice_qp_dis - Disables a queue pair
 157 * @vsi: VSI of interest
 158 * @q_idx: ring index in array
 159 *
 160 * Returns 0 on success, negative on failure.
 161 */
 162static int ice_qp_dis(struct ice_vsi *vsi, u16 q_idx)
 163{
 164	struct ice_txq_meta txq_meta = { };
 165	struct ice_q_vector *q_vector;
 166	struct ice_tx_ring *tx_ring;
 167	struct ice_rx_ring *rx_ring;
 168	int fail = 0;
 169	int err;
 170
 171	if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
 172		return -EINVAL;
 173
 174	tx_ring = vsi->tx_rings[q_idx];
 175	rx_ring = vsi->rx_rings[q_idx];
 176	q_vector = rx_ring->q_vector;
 177
 178	synchronize_net();
 179	netif_carrier_off(vsi->netdev);
 
 
 
 
 180	netif_tx_stop_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
 181
 182	ice_qvec_dis_irq(vsi, rx_ring, q_vector);
 183	ice_qvec_toggle_napi(vsi, q_vector, false);
 184
 185	ice_fill_txq_meta(vsi, tx_ring, &txq_meta);
 186	err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, tx_ring, &txq_meta);
 187	if (!fail)
 188		fail = err;
 189	if (vsi->xdp_rings) {
 190		struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
 191
 192		memset(&txq_meta, 0, sizeof(txq_meta));
 193		ice_fill_txq_meta(vsi, xdp_ring, &txq_meta);
 194		err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, xdp_ring,
 195					   &txq_meta);
 196		if (!fail)
 197			fail = err;
 198	}
 
 
 
 
 199
 200	ice_vsi_ctrl_one_rx_ring(vsi, false, q_idx, false);
 201	ice_qp_clean_rings(vsi, q_idx);
 202	ice_qp_reset_stats(vsi, q_idx);
 203
 204	return fail;
 205}
 206
 207/**
 208 * ice_qp_ena - Enables a queue pair
 209 * @vsi: VSI of interest
 210 * @q_idx: ring index in array
 211 *
 212 * Returns 0 on success, negative on failure.
 213 */
 214static int ice_qp_ena(struct ice_vsi *vsi, u16 q_idx)
 215{
 
 216	struct ice_q_vector *q_vector;
 217	int fail = 0;
 218	bool link_up;
 
 219	int err;
 220
 221	err = ice_vsi_cfg_single_txq(vsi, vsi->tx_rings, q_idx);
 222	if (!fail)
 223		fail = err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 224
 225	if (ice_is_xdp_ena_vsi(vsi)) {
 226		struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
 227
 228		err = ice_vsi_cfg_single_txq(vsi, vsi->xdp_rings, q_idx);
 229		if (!fail)
 230			fail = err;
 
 
 231		ice_set_ring_xdp(xdp_ring);
 232		ice_tx_xsk_pool(vsi, q_idx);
 233	}
 234
 235	err = ice_vsi_cfg_single_rxq(vsi, q_idx);
 236	if (!fail)
 237		fail = err;
 238
 239	q_vector = vsi->rx_rings[q_idx]->q_vector;
 240	ice_qvec_cfg_msix(vsi, q_vector, q_idx);
 241
 242	err = ice_vsi_ctrl_one_rx_ring(vsi, true, q_idx, true);
 243	if (!fail)
 244		fail = err;
 245
 
 246	ice_qvec_toggle_napi(vsi, q_vector, true);
 247	ice_qvec_ena_irq(vsi, q_vector);
 248
 249	/* make sure NAPI sees updated ice_{t,x}_ring::xsk_pool */
 250	synchronize_net();
 251	ice_get_link_status(vsi->port_info, &link_up);
 252	if (link_up) {
 253		netif_tx_start_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
 254		netif_carrier_on(vsi->netdev);
 255	}
 256
 257	return fail;
 258}
 259
 260/**
 261 * ice_xsk_pool_disable - disable a buffer pool region
 262 * @vsi: Current VSI
 263 * @qid: queue ID
 264 *
 265 * Returns 0 on success, negative on failure
 266 */
 267static int ice_xsk_pool_disable(struct ice_vsi *vsi, u16 qid)
 268{
 269	struct xsk_buff_pool *pool = xsk_get_pool_from_qid(vsi->netdev, qid);
 270
 271	if (!pool)
 272		return -EINVAL;
 273
 
 274	xsk_pool_dma_unmap(pool, ICE_RX_DMA_ATTR);
 275
 276	return 0;
 277}
 278
 279/**
 280 * ice_xsk_pool_enable - enable a buffer pool region
 281 * @vsi: Current VSI
 282 * @pool: pointer to a requested buffer pool region
 283 * @qid: queue ID
 284 *
 285 * Returns 0 on success, negative on failure
 286 */
 287static int
 288ice_xsk_pool_enable(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
 289{
 290	int err;
 291
 292	if (vsi->type != ICE_VSI_PF && vsi->type != ICE_VSI_SF)
 293		return -EINVAL;
 294
 295	if (qid >= vsi->netdev->real_num_rx_queues ||
 296	    qid >= vsi->netdev->real_num_tx_queues)
 297		return -EINVAL;
 298
 299	err = xsk_pool_dma_map(pool, ice_pf_to_dev(vsi->back),
 300			       ICE_RX_DMA_ATTR);
 301	if (err)
 302		return err;
 303
 
 
 304	return 0;
 305}
 306
 307/**
 308 * ice_realloc_rx_xdp_bufs - reallocate for either XSK or normal buffer
 309 * @rx_ring: Rx ring
 310 * @pool_present: is pool for XSK present
 311 *
 312 * Try allocating memory and return ENOMEM, if failed to allocate.
 313 * If allocation was successful, substitute buffer with allocated one.
 314 * Returns 0 on success, negative on failure
 315 */
 316static int
 317ice_realloc_rx_xdp_bufs(struct ice_rx_ring *rx_ring, bool pool_present)
 318{
 319	size_t elem_size = pool_present ? sizeof(*rx_ring->xdp_buf) :
 320					  sizeof(*rx_ring->rx_buf);
 321	void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
 322
 323	if (!sw_ring)
 324		return -ENOMEM;
 325
 326	if (pool_present) {
 327		kfree(rx_ring->rx_buf);
 328		rx_ring->rx_buf = NULL;
 329		rx_ring->xdp_buf = sw_ring;
 330	} else {
 331		kfree(rx_ring->xdp_buf);
 332		rx_ring->xdp_buf = NULL;
 333		rx_ring->rx_buf = sw_ring;
 334	}
 335
 336	return 0;
 337}
 338
 339/**
 340 * ice_realloc_zc_buf - reallocate XDP ZC queue pairs
 341 * @vsi: Current VSI
 342 * @zc: is zero copy set
 343 *
 344 * Reallocate buffer for rx_rings that might be used by XSK.
 345 * XDP requires more memory, than rx_buf provides.
 346 * Returns 0 on success, negative on failure
 347 */
 348int ice_realloc_zc_buf(struct ice_vsi *vsi, bool zc)
 349{
 350	struct ice_rx_ring *rx_ring;
 351	uint i;
 352
 353	ice_for_each_rxq(vsi, i) {
 354		rx_ring = vsi->rx_rings[i];
 355		if (!rx_ring->xsk_pool)
 356			continue;
 357
 
 
 
 358		if (ice_realloc_rx_xdp_bufs(rx_ring, zc))
 359			return -ENOMEM;
 360	}
 361
 362	return 0;
 363}
 364
 365/**
 366 * ice_xsk_pool_setup - enable/disable a buffer pool region depending on its state
 367 * @vsi: Current VSI
 368 * @pool: buffer pool to enable/associate to a ring, NULL to disable
 369 * @qid: queue ID
 370 *
 371 * Returns 0 on success, negative on failure
 372 */
 373int ice_xsk_pool_setup(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
 374{
 375	bool if_running, pool_present = !!pool;
 376	int ret = 0, pool_failure = 0;
 377
 378	if (qid >= vsi->num_rxq || qid >= vsi->num_txq) {
 379		netdev_err(vsi->netdev, "Please use queue id in scope of combined queues count\n");
 380		pool_failure = -EINVAL;
 381		goto failure;
 382	}
 383
 384	if_running = !test_bit(ICE_VSI_DOWN, vsi->state) &&
 385		     ice_is_xdp_ena_vsi(vsi);
 386
 387	if (if_running) {
 388		struct ice_rx_ring *rx_ring = vsi->rx_rings[qid];
 389
 390		ret = ice_qp_dis(vsi, qid);
 391		if (ret) {
 392			netdev_err(vsi->netdev, "ice_qp_dis error = %d\n", ret);
 393			goto xsk_pool_if_up;
 394		}
 395
 396		ret = ice_realloc_rx_xdp_bufs(rx_ring, pool_present);
 397		if (ret)
 398			goto xsk_pool_if_up;
 399	}
 400
 401	pool_failure = pool_present ? ice_xsk_pool_enable(vsi, pool, qid) :
 402				      ice_xsk_pool_disable(vsi, qid);
 403
 404xsk_pool_if_up:
 405	if (if_running) {
 406		ret = ice_qp_ena(vsi, qid);
 407		if (!ret && pool_present)
 408			napi_schedule(&vsi->rx_rings[qid]->xdp_ring->q_vector->napi);
 409		else if (ret)
 410			netdev_err(vsi->netdev, "ice_qp_ena error = %d\n", ret);
 411	}
 412
 413failure:
 414	if (pool_failure) {
 415		netdev_err(vsi->netdev, "Could not %sable buffer pool, error = %d\n",
 416			   pool_present ? "en" : "dis", pool_failure);
 417		return pool_failure;
 418	}
 419
 420	return ret;
 421}
 422
 423/**
 424 * ice_fill_rx_descs - pick buffers from XSK buffer pool and use it
 425 * @pool: XSK Buffer pool to pull the buffers from
 426 * @xdp: SW ring of xdp_buff that will hold the buffers
 427 * @rx_desc: Pointer to Rx descriptors that will be filled
 428 * @count: The number of buffers to allocate
 429 *
 430 * This function allocates a number of Rx buffers from the fill ring
 431 * or the internal recycle mechanism and places them on the Rx ring.
 432 *
 433 * Note that ring wrap should be handled by caller of this function.
 434 *
 435 * Returns the amount of allocated Rx descriptors
 436 */
 437static u16 ice_fill_rx_descs(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
 438			     union ice_32b_rx_flex_desc *rx_desc, u16 count)
 439{
 440	dma_addr_t dma;
 441	u16 buffs;
 442	int i;
 443
 444	buffs = xsk_buff_alloc_batch(pool, xdp, count);
 445	for (i = 0; i < buffs; i++) {
 446		dma = xsk_buff_xdp_get_dma(*xdp);
 447		rx_desc->read.pkt_addr = cpu_to_le64(dma);
 448		rx_desc->wb.status_error0 = 0;
 449
 450		/* Put private info that changes on a per-packet basis
 451		 * into xdp_buff_xsk->cb.
 452		 */
 453		ice_xdp_meta_set_desc(*xdp, rx_desc);
 454
 455		rx_desc++;
 456		xdp++;
 457	}
 458
 459	return buffs;
 460}
 461
 462/**
 463 * __ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
 464 * @rx_ring: Rx ring
 465 * @xsk_pool: XSK buffer pool to pick buffers to be filled by HW
 466 * @count: The number of buffers to allocate
 467 *
 468 * Place the @count of descriptors onto Rx ring. Handle the ring wrap
 469 * for case where space from next_to_use up to the end of ring is less
 470 * than @count. Finally do a tail bump.
 471 *
 472 * Returns true if all allocations were successful, false if any fail.
 473 */
 474static bool __ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring,
 475				   struct xsk_buff_pool *xsk_pool, u16 count)
 476{
 477	u32 nb_buffs_extra = 0, nb_buffs = 0;
 478	union ice_32b_rx_flex_desc *rx_desc;
 479	u16 ntu = rx_ring->next_to_use;
 480	u16 total_count = count;
 481	struct xdp_buff **xdp;
 482
 483	rx_desc = ICE_RX_DESC(rx_ring, ntu);
 484	xdp = ice_xdp_buf(rx_ring, ntu);
 485
 486	if (ntu + count >= rx_ring->count) {
 487		nb_buffs_extra = ice_fill_rx_descs(xsk_pool, xdp, rx_desc,
 
 488						   rx_ring->count - ntu);
 489		if (nb_buffs_extra != rx_ring->count - ntu) {
 490			ntu += nb_buffs_extra;
 491			goto exit;
 492		}
 493		rx_desc = ICE_RX_DESC(rx_ring, 0);
 494		xdp = ice_xdp_buf(rx_ring, 0);
 495		ntu = 0;
 496		count -= nb_buffs_extra;
 497		ice_release_rx_desc(rx_ring, 0);
 498	}
 499
 500	nb_buffs = ice_fill_rx_descs(xsk_pool, xdp, rx_desc, count);
 501
 502	ntu += nb_buffs;
 503	if (ntu == rx_ring->count)
 504		ntu = 0;
 505
 506exit:
 507	if (rx_ring->next_to_use != ntu)
 508		ice_release_rx_desc(rx_ring, ntu);
 509
 510	return total_count == (nb_buffs_extra + nb_buffs);
 511}
 512
 513/**
 514 * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
 515 * @rx_ring: Rx ring
 516 * @xsk_pool: XSK buffer pool to pick buffers to be filled by HW
 517 * @count: The number of buffers to allocate
 518 *
 519 * Wrapper for internal allocation routine; figure out how many tail
 520 * bumps should take place based on the given threshold
 521 *
 522 * Returns true if all calls to internal alloc routine succeeded
 523 */
 524bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring,
 525			  struct xsk_buff_pool *xsk_pool, u16 count)
 526{
 527	u16 rx_thresh = ICE_RING_QUARTER(rx_ring);
 528	u16 leftover, i, tail_bumps;
 529
 530	tail_bumps = count / rx_thresh;
 531	leftover = count - (tail_bumps * rx_thresh);
 532
 533	for (i = 0; i < tail_bumps; i++)
 534		if (!__ice_alloc_rx_bufs_zc(rx_ring, xsk_pool, rx_thresh))
 535			return false;
 536	return __ice_alloc_rx_bufs_zc(rx_ring, xsk_pool, leftover);
 
 
 
 
 
 
 
 
 
 
 
 
 
 537}
 538
 539/**
 540 * ice_construct_skb_zc - Create an sk_buff from zero-copy buffer
 541 * @rx_ring: Rx ring
 542 * @xdp: Pointer to XDP buffer
 543 *
 544 * This function allocates a new skb from a zero-copy Rx buffer.
 545 *
 546 * Returns the skb on success, NULL on failure.
 547 */
 548static struct sk_buff *
 549ice_construct_skb_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp)
 550{
 551	unsigned int totalsize = xdp->data_end - xdp->data_meta;
 552	unsigned int metasize = xdp->data - xdp->data_meta;
 553	struct skb_shared_info *sinfo = NULL;
 554	struct sk_buff *skb;
 555	u32 nr_frags = 0;
 556
 557	if (unlikely(xdp_buff_has_frags(xdp))) {
 558		sinfo = xdp_get_shared_info_from_buff(xdp);
 559		nr_frags = sinfo->nr_frags;
 560	}
 561	net_prefetch(xdp->data_meta);
 562
 563	skb = napi_alloc_skb(&rx_ring->q_vector->napi, totalsize);
 
 564	if (unlikely(!skb))
 565		return NULL;
 566
 567	memcpy(__skb_put(skb, totalsize), xdp->data_meta,
 568	       ALIGN(totalsize, sizeof(long)));
 569
 570	if (metasize) {
 571		skb_metadata_set(skb, metasize);
 572		__skb_pull(skb, metasize);
 573	}
 574
 575	if (likely(!xdp_buff_has_frags(xdp)))
 576		goto out;
 577
 578	for (int i = 0; i < nr_frags; i++) {
 579		struct skb_shared_info *skinfo = skb_shinfo(skb);
 580		skb_frag_t *frag = &sinfo->frags[i];
 581		struct page *page;
 582		void *addr;
 583
 584		page = dev_alloc_page();
 585		if (!page) {
 586			dev_kfree_skb(skb);
 587			return NULL;
 588		}
 589		addr = page_to_virt(page);
 590
 591		memcpy(addr, skb_frag_page(frag), skb_frag_size(frag));
 592
 593		__skb_fill_page_desc_noacc(skinfo, skinfo->nr_frags++,
 594					   addr, 0, skb_frag_size(frag));
 595	}
 596
 597out:
 598	xsk_buff_free(xdp);
 599	return skb;
 600}
 601
 602/**
 603 * ice_clean_xdp_irq_zc - produce AF_XDP descriptors to CQ
 604 * @xdp_ring: XDP Tx ring
 605 * @xsk_pool: AF_XDP buffer pool pointer
 606 */
 607static u32 ice_clean_xdp_irq_zc(struct ice_tx_ring *xdp_ring,
 608				struct xsk_buff_pool *xsk_pool)
 609{
 610	u16 ntc = xdp_ring->next_to_clean;
 611	struct ice_tx_desc *tx_desc;
 612	u16 cnt = xdp_ring->count;
 613	struct ice_tx_buf *tx_buf;
 614	u16 completed_frames = 0;
 615	u16 xsk_frames = 0;
 616	u16 last_rs;
 617	int i;
 618
 619	last_rs = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : cnt - 1;
 620	tx_desc = ICE_TX_DESC(xdp_ring, last_rs);
 621	if (tx_desc->cmd_type_offset_bsz &
 622	    cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE)) {
 623		if (last_rs >= ntc)
 624			completed_frames = last_rs - ntc + 1;
 625		else
 626			completed_frames = last_rs + cnt - ntc + 1;
 627	}
 628
 629	if (!completed_frames)
 630		return 0;
 631
 632	if (likely(!xdp_ring->xdp_tx_active)) {
 633		xsk_frames = completed_frames;
 634		goto skip;
 635	}
 636
 637	ntc = xdp_ring->next_to_clean;
 638	for (i = 0; i < completed_frames; i++) {
 639		tx_buf = &xdp_ring->tx_buf[ntc];
 640
 641		if (tx_buf->type == ICE_TX_BUF_XSK_TX) {
 642			tx_buf->type = ICE_TX_BUF_EMPTY;
 643			xsk_buff_free(tx_buf->xdp);
 644			xdp_ring->xdp_tx_active--;
 645		} else {
 646			xsk_frames++;
 647		}
 648
 649		ntc++;
 650		if (ntc >= xdp_ring->count)
 651			ntc = 0;
 652	}
 653skip:
 654	tx_desc->cmd_type_offset_bsz = 0;
 655	xdp_ring->next_to_clean += completed_frames;
 656	if (xdp_ring->next_to_clean >= cnt)
 657		xdp_ring->next_to_clean -= cnt;
 658	if (xsk_frames)
 659		xsk_tx_completed(xsk_pool, xsk_frames);
 660
 661	return completed_frames;
 662}
 663
 664/**
 665 * ice_xmit_xdp_tx_zc - AF_XDP ZC handler for XDP_TX
 666 * @xdp: XDP buffer to xmit
 667 * @xdp_ring: XDP ring to produce descriptor onto
 668 * @xsk_pool: AF_XDP buffer pool pointer
 669 *
 670 * note that this function works directly on xdp_buff, no need to convert
 671 * it to xdp_frame. xdp_buff pointer is stored to ice_tx_buf so that cleaning
 672 * side will be able to xsk_buff_free() it.
 673 *
 674 * Returns ICE_XDP_TX for successfully produced desc, ICE_XDP_CONSUMED if there
 675 * was not enough space on XDP ring
 676 */
 677static int ice_xmit_xdp_tx_zc(struct xdp_buff *xdp,
 678			      struct ice_tx_ring *xdp_ring,
 679			      struct xsk_buff_pool *xsk_pool)
 680{
 681	struct skb_shared_info *sinfo = NULL;
 682	u32 size = xdp->data_end - xdp->data;
 683	u32 ntu = xdp_ring->next_to_use;
 684	struct ice_tx_desc *tx_desc;
 685	struct ice_tx_buf *tx_buf;
 686	struct xdp_buff *head;
 687	u32 nr_frags = 0;
 688	u32 free_space;
 689	u32 frag = 0;
 690
 691	free_space = ICE_DESC_UNUSED(xdp_ring);
 692	if (free_space < ICE_RING_QUARTER(xdp_ring))
 693		free_space += ice_clean_xdp_irq_zc(xdp_ring, xsk_pool);
 694
 695	if (unlikely(!free_space))
 696		goto busy;
 697
 698	if (unlikely(xdp_buff_has_frags(xdp))) {
 699		sinfo = xdp_get_shared_info_from_buff(xdp);
 700		nr_frags = sinfo->nr_frags;
 701		if (free_space < nr_frags + 1)
 702			goto busy;
 703	}
 704
 705	tx_desc = ICE_TX_DESC(xdp_ring, ntu);
 706	tx_buf = &xdp_ring->tx_buf[ntu];
 707	head = xdp;
 708
 709	for (;;) {
 710		dma_addr_t dma;
 711
 712		dma = xsk_buff_xdp_get_dma(xdp);
 713		xsk_buff_raw_dma_sync_for_device(xsk_pool, dma, size);
 714
 715		tx_buf->xdp = xdp;
 716		tx_buf->type = ICE_TX_BUF_XSK_TX;
 717		tx_desc->buf_addr = cpu_to_le64(dma);
 718		tx_desc->cmd_type_offset_bsz = ice_build_ctob(0, 0, size, 0);
 719		/* account for each xdp_buff from xsk_buff_pool */
 720		xdp_ring->xdp_tx_active++;
 721
 722		if (++ntu == xdp_ring->count)
 723			ntu = 0;
 724
 725		if (frag == nr_frags)
 726			break;
 727
 728		tx_desc = ICE_TX_DESC(xdp_ring, ntu);
 729		tx_buf = &xdp_ring->tx_buf[ntu];
 730
 731		xdp = xsk_buff_get_frag(head);
 732		size = skb_frag_size(&sinfo->frags[frag]);
 733		frag++;
 734	}
 735
 736	xdp_ring->next_to_use = ntu;
 737	/* update last descriptor from a frame with EOP */
 738	tx_desc->cmd_type_offset_bsz |=
 739		cpu_to_le64(ICE_TX_DESC_CMD_EOP << ICE_TXD_QW1_CMD_S);
 740
 741	return ICE_XDP_TX;
 742
 743busy:
 744	xdp_ring->ring_stats->tx_stats.tx_busy++;
 745
 746	return ICE_XDP_CONSUMED;
 747}
 748
 749/**
 750 * ice_run_xdp_zc - Executes an XDP program in zero-copy path
 751 * @rx_ring: Rx ring
 752 * @xdp: xdp_buff used as input to the XDP program
 753 * @xdp_prog: XDP program to run
 754 * @xdp_ring: ring to be used for XDP_TX action
 755 * @xsk_pool: AF_XDP buffer pool pointer
 756 *
 757 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
 758 */
 759static int
 760ice_run_xdp_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
 761	       struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring,
 762	       struct xsk_buff_pool *xsk_pool)
 763{
 764	int err, result = ICE_XDP_PASS;
 765	u32 act;
 766
 767	act = bpf_prog_run_xdp(xdp_prog, xdp);
 768
 769	if (likely(act == XDP_REDIRECT)) {
 770		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
 771		if (!err)
 772			return ICE_XDP_REDIR;
 773		if (xsk_uses_need_wakeup(xsk_pool) && err == -ENOBUFS)
 774			result = ICE_XDP_EXIT;
 775		else
 776			result = ICE_XDP_CONSUMED;
 777		goto out_failure;
 778	}
 779
 780	switch (act) {
 781	case XDP_PASS:
 782		break;
 783	case XDP_TX:
 784		result = ice_xmit_xdp_tx_zc(xdp, xdp_ring, xsk_pool);
 785		if (result == ICE_XDP_CONSUMED)
 786			goto out_failure;
 787		break;
 788	case XDP_DROP:
 789		result = ICE_XDP_CONSUMED;
 790		break;
 791	default:
 792		bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
 793		fallthrough;
 794	case XDP_ABORTED:
 795		result = ICE_XDP_CONSUMED;
 796out_failure:
 797		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
 798		break;
 799	}
 800
 801	return result;
 802}
 803
 804static int
 805ice_add_xsk_frag(struct ice_rx_ring *rx_ring, struct xdp_buff *first,
 806		 struct xdp_buff *xdp, const unsigned int size)
 807{
 808	struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(first);
 809
 810	if (!size)
 811		return 0;
 812
 813	if (!xdp_buff_has_frags(first)) {
 814		sinfo->nr_frags = 0;
 815		sinfo->xdp_frags_size = 0;
 816		xdp_buff_set_frags_flag(first);
 817	}
 818
 819	if (unlikely(sinfo->nr_frags == MAX_SKB_FRAGS)) {
 820		xsk_buff_free(first);
 821		return -ENOMEM;
 822	}
 823
 824	__skb_fill_page_desc_noacc(sinfo, sinfo->nr_frags++,
 825				   virt_to_page(xdp->data_hard_start),
 826				   XDP_PACKET_HEADROOM, size);
 827	sinfo->xdp_frags_size += size;
 828	xsk_buff_add_frag(xdp);
 829
 830	return 0;
 831}
 832
 833/**
 834 * ice_clean_rx_irq_zc - consumes packets from the hardware ring
 835 * @rx_ring: AF_XDP Rx ring
 836 * @xsk_pool: AF_XDP buffer pool pointer
 837 * @budget: NAPI budget
 838 *
 839 * Returns number of processed packets on success, remaining budget on failure.
 840 */
 841int ice_clean_rx_irq_zc(struct ice_rx_ring *rx_ring,
 842			struct xsk_buff_pool *xsk_pool,
 843			int budget)
 844{
 845	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
 846	u32 ntc = rx_ring->next_to_clean;
 847	u32 ntu = rx_ring->next_to_use;
 848	struct xdp_buff *first = NULL;
 849	struct ice_tx_ring *xdp_ring;
 850	unsigned int xdp_xmit = 0;
 851	struct bpf_prog *xdp_prog;
 852	u32 cnt = rx_ring->count;
 853	bool failure = false;
 854	int entries_to_alloc;
 855
 856	/* ZC patch is enabled only when XDP program is set,
 857	 * so here it can not be NULL
 858	 */
 859	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
 860	xdp_ring = rx_ring->xdp_ring;
 861
 862	if (ntc != rx_ring->first_desc)
 863		first = *ice_xdp_buf(rx_ring, rx_ring->first_desc);
 864
 865	while (likely(total_rx_packets < (unsigned int)budget)) {
 866		union ice_32b_rx_flex_desc *rx_desc;
 867		unsigned int size, xdp_res = 0;
 868		struct xdp_buff *xdp;
 869		struct sk_buff *skb;
 870		u16 stat_err_bits;
 871		u16 vlan_tci;
 
 872
 873		rx_desc = ICE_RX_DESC(rx_ring, ntc);
 874
 875		stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
 876		if (!ice_test_staterr(rx_desc->wb.status_error0, stat_err_bits))
 877			break;
 878
 879		/* This memory barrier is needed to keep us from reading
 880		 * any other fields out of the rx_desc until we have
 881		 * verified the descriptor has been written back.
 882		 */
 883		dma_rmb();
 884
 885		if (unlikely(ntc == ntu))
 886			break;
 887
 888		xdp = *ice_xdp_buf(rx_ring, ntc);
 889
 890		size = le16_to_cpu(rx_desc->wb.pkt_len) &
 891				   ICE_RX_FLX_DESC_PKT_LEN_M;
 892
 893		xsk_buff_set_size(xdp, size);
 894		xsk_buff_dma_sync_for_cpu(xdp);
 895
 896		if (!first) {
 897			first = xdp;
 898		} else if (ice_add_xsk_frag(rx_ring, first, xdp, size)) {
 899			break;
 900		}
 901
 902		if (++ntc == cnt)
 903			ntc = 0;
 904
 905		if (ice_is_non_eop(rx_ring, rx_desc))
 906			continue;
 907
 908		xdp_res = ice_run_xdp_zc(rx_ring, first, xdp_prog, xdp_ring,
 909					 xsk_pool);
 910		if (likely(xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR))) {
 911			xdp_xmit |= xdp_res;
 912		} else if (xdp_res == ICE_XDP_EXIT) {
 913			failure = true;
 914			first = NULL;
 915			rx_ring->first_desc = ntc;
 916			break;
 917		} else if (xdp_res == ICE_XDP_CONSUMED) {
 918			xsk_buff_free(first);
 919		} else if (xdp_res == ICE_XDP_PASS) {
 920			goto construct_skb;
 921		}
 922
 923		total_rx_bytes += xdp_get_buff_len(first);
 924		total_rx_packets++;
 925
 926		first = NULL;
 927		rx_ring->first_desc = ntc;
 928		continue;
 929
 930construct_skb:
 931		/* XDP_PASS path */
 932		skb = ice_construct_skb_zc(rx_ring, first);
 933		if (!skb) {
 934			rx_ring->ring_stats->rx_stats.alloc_buf_failed++;
 935			break;
 936		}
 937
 938		first = NULL;
 939		rx_ring->first_desc = ntc;
 940
 941		if (eth_skb_pad(skb)) {
 942			skb = NULL;
 943			continue;
 944		}
 945
 946		total_rx_bytes += skb->len;
 947		total_rx_packets++;
 948
 949		vlan_tci = ice_get_vlan_tci(rx_desc);
 
 
 
 950
 951		ice_process_skb_fields(rx_ring, rx_desc, skb);
 952		ice_receive_skb(rx_ring, skb, vlan_tci);
 953	}
 954
 955	rx_ring->next_to_clean = ntc;
 956	entries_to_alloc = ICE_RX_DESC_UNUSED(rx_ring);
 957	if (entries_to_alloc > ICE_RING_QUARTER(rx_ring))
 958		failure |= !ice_alloc_rx_bufs_zc(rx_ring, xsk_pool,
 959						 entries_to_alloc);
 960
 961	ice_finalize_xdp_rx(xdp_ring, xdp_xmit, 0);
 962	ice_update_rx_ring_stats(rx_ring, total_rx_packets, total_rx_bytes);
 963
 964	if (xsk_uses_need_wakeup(xsk_pool)) {
 965		/* ntu could have changed when allocating entries above, so
 966		 * use rx_ring value instead of stack based one
 967		 */
 968		if (failure || ntc == rx_ring->next_to_use)
 969			xsk_set_rx_need_wakeup(xsk_pool);
 970		else
 971			xsk_clear_rx_need_wakeup(xsk_pool);
 972
 973		return (int)total_rx_packets;
 974	}
 975
 976	return failure ? budget : (int)total_rx_packets;
 977}
 978
 979/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 980 * ice_xmit_pkt - produce a single HW Tx descriptor out of AF_XDP descriptor
 981 * @xdp_ring: XDP ring to produce the HW Tx descriptor on
 982 * @xsk_pool: XSK buffer pool to pick buffers to be consumed by HW
 983 * @desc: AF_XDP descriptor to pull the DMA address and length from
 984 * @total_bytes: bytes accumulator that will be used for stats update
 985 */
 986static void ice_xmit_pkt(struct ice_tx_ring *xdp_ring,
 987			 struct xsk_buff_pool *xsk_pool, struct xdp_desc *desc,
 988			 unsigned int *total_bytes)
 989{
 990	struct ice_tx_desc *tx_desc;
 991	dma_addr_t dma;
 992
 993	dma = xsk_buff_raw_get_dma(xsk_pool, desc->addr);
 994	xsk_buff_raw_dma_sync_for_device(xsk_pool, dma, desc->len);
 995
 996	tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
 997	tx_desc->buf_addr = cpu_to_le64(dma);
 998	tx_desc->cmd_type_offset_bsz = ice_build_ctob(xsk_is_eop_desc(desc),
 999						      0, desc->len, 0);
1000
1001	*total_bytes += desc->len;
1002}
1003
1004/**
1005 * ice_xmit_pkt_batch - produce a batch of HW Tx descriptors out of AF_XDP descriptors
1006 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
1007 * @xsk_pool: XSK buffer pool to pick buffers to be consumed by HW
1008 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
1009 * @total_bytes: bytes accumulator that will be used for stats update
1010 */
1011static void ice_xmit_pkt_batch(struct ice_tx_ring *xdp_ring,
1012			       struct xsk_buff_pool *xsk_pool,
1013			       struct xdp_desc *descs,
1014			       unsigned int *total_bytes)
1015{
1016	u16 ntu = xdp_ring->next_to_use;
1017	struct ice_tx_desc *tx_desc;
1018	u32 i;
1019
1020	loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
1021		dma_addr_t dma;
1022
1023		dma = xsk_buff_raw_get_dma(xsk_pool, descs[i].addr);
1024		xsk_buff_raw_dma_sync_for_device(xsk_pool, dma, descs[i].len);
1025
1026		tx_desc = ICE_TX_DESC(xdp_ring, ntu++);
1027		tx_desc->buf_addr = cpu_to_le64(dma);
1028		tx_desc->cmd_type_offset_bsz = ice_build_ctob(xsk_is_eop_desc(&descs[i]),
1029							      0, descs[i].len, 0);
1030
1031		*total_bytes += descs[i].len;
1032	}
1033
1034	xdp_ring->next_to_use = ntu;
1035}
1036
1037/**
1038 * ice_fill_tx_hw_ring - produce the number of Tx descriptors onto ring
1039 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
1040 * @xsk_pool: XSK buffer pool to pick buffers to be consumed by HW
1041 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
1042 * @nb_pkts: count of packets to be send
1043 * @total_bytes: bytes accumulator that will be used for stats update
1044 */
1045static void ice_fill_tx_hw_ring(struct ice_tx_ring *xdp_ring,
1046				struct xsk_buff_pool *xsk_pool,
1047				struct xdp_desc *descs, u32 nb_pkts,
1048				unsigned int *total_bytes)
1049{
1050	u32 batched, leftover, i;
1051
1052	batched = ALIGN_DOWN(nb_pkts, PKTS_PER_BATCH);
1053	leftover = nb_pkts & (PKTS_PER_BATCH - 1);
1054	for (i = 0; i < batched; i += PKTS_PER_BATCH)
1055		ice_xmit_pkt_batch(xdp_ring, xsk_pool, &descs[i], total_bytes);
1056	for (; i < batched + leftover; i++)
1057		ice_xmit_pkt(xdp_ring, xsk_pool, &descs[i], total_bytes);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1058}
1059
1060/**
1061 * ice_xmit_zc - take entries from XSK Tx ring and place them onto HW Tx ring
1062 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
1063 * @xsk_pool: AF_XDP buffer pool pointer
1064 *
1065 * Returns true if there is no more work that needs to be done, false otherwise
1066 */
1067bool ice_xmit_zc(struct ice_tx_ring *xdp_ring, struct xsk_buff_pool *xsk_pool)
1068{
1069	struct xdp_desc *descs = xsk_pool->tx_descs;
1070	u32 nb_pkts, nb_processed = 0;
1071	unsigned int total_bytes = 0;
1072	int budget;
1073
1074	ice_clean_xdp_irq_zc(xdp_ring, xsk_pool);
1075
1076	if (!netif_carrier_ok(xdp_ring->vsi->netdev) ||
1077	    !netif_running(xdp_ring->vsi->netdev))
1078		return true;
1079
1080	budget = ICE_DESC_UNUSED(xdp_ring);
1081	budget = min_t(u16, budget, ICE_RING_QUARTER(xdp_ring));
1082
1083	nb_pkts = xsk_tx_peek_release_desc_batch(xsk_pool, budget);
1084	if (!nb_pkts)
1085		return true;
1086
1087	if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
1088		nb_processed = xdp_ring->count - xdp_ring->next_to_use;
1089		ice_fill_tx_hw_ring(xdp_ring, xsk_pool, descs, nb_processed,
1090				    &total_bytes);
1091		xdp_ring->next_to_use = 0;
1092	}
1093
1094	ice_fill_tx_hw_ring(xdp_ring, xsk_pool, &descs[nb_processed],
1095			    nb_pkts - nb_processed, &total_bytes);
1096
1097	ice_set_rs_bit(xdp_ring);
1098	ice_xdp_ring_update_tail(xdp_ring);
1099	ice_update_tx_ring_stats(xdp_ring, nb_pkts, total_bytes);
1100
1101	if (xsk_uses_need_wakeup(xsk_pool))
1102		xsk_set_tx_need_wakeup(xsk_pool);
1103
1104	return nb_pkts < budget;
1105}
1106
1107/**
1108 * ice_xsk_wakeup - Implements ndo_xsk_wakeup
1109 * @netdev: net_device
1110 * @queue_id: queue to wake up
1111 * @flags: ignored in our case, since we have Rx and Tx in the same NAPI
1112 *
1113 * Returns negative on error, zero otherwise.
1114 */
1115int
1116ice_xsk_wakeup(struct net_device *netdev, u32 queue_id,
1117	       u32 __always_unused flags)
1118{
1119	struct ice_netdev_priv *np = netdev_priv(netdev);
1120	struct ice_q_vector *q_vector;
1121	struct ice_vsi *vsi = np->vsi;
1122	struct ice_tx_ring *ring;
1123
1124	if (test_bit(ICE_VSI_DOWN, vsi->state) || !netif_carrier_ok(netdev))
1125		return -ENETDOWN;
1126
1127	if (!ice_is_xdp_ena_vsi(vsi))
1128		return -EINVAL;
1129
1130	if (queue_id >= vsi->num_txq || queue_id >= vsi->num_rxq)
1131		return -EINVAL;
1132
1133	ring = vsi->rx_rings[queue_id]->xdp_ring;
1134
1135	if (!READ_ONCE(ring->xsk_pool))
1136		return -EINVAL;
1137
1138	/* The idea here is that if NAPI is running, mark a miss, so
1139	 * it will run again. If not, trigger an interrupt and
1140	 * schedule the NAPI from interrupt context. If NAPI would be
1141	 * scheduled here, the interrupt affinity would not be
1142	 * honored.
1143	 */
1144	q_vector = ring->q_vector;
1145	if (!napi_if_scheduled_mark_missed(&q_vector->napi))
1146		ice_trigger_sw_intr(&vsi->back->hw, q_vector);
1147
1148	return 0;
1149}
1150
1151/**
1152 * ice_xsk_any_rx_ring_ena - Checks if Rx rings have AF_XDP buff pool attached
1153 * @vsi: VSI to be checked
1154 *
1155 * Returns true if any of the Rx rings has an AF_XDP buff pool attached
1156 */
1157bool ice_xsk_any_rx_ring_ena(struct ice_vsi *vsi)
1158{
1159	int i;
1160
1161	ice_for_each_rxq(vsi, i) {
1162		if (xsk_get_pool_from_qid(vsi->netdev, i))
1163			return true;
1164	}
1165
1166	return false;
1167}
1168
1169/**
1170 * ice_xsk_clean_rx_ring - clean buffer pool queues connected to a given Rx ring
1171 * @rx_ring: ring to be cleaned
1172 */
1173void ice_xsk_clean_rx_ring(struct ice_rx_ring *rx_ring)
1174{
1175	u16 ntc = rx_ring->next_to_clean;
1176	u16 ntu = rx_ring->next_to_use;
1177
1178	while (ntc != ntu) {
1179		struct xdp_buff *xdp = *ice_xdp_buf(rx_ring, ntc);
1180
1181		xsk_buff_free(xdp);
1182		ntc++;
1183		if (ntc >= rx_ring->count)
1184			ntc = 0;
1185	}
1186}
1187
1188/**
1189 * ice_xsk_clean_xdp_ring - Clean the XDP Tx ring and its buffer pool queues
1190 * @xdp_ring: XDP_Tx ring
1191 */
1192void ice_xsk_clean_xdp_ring(struct ice_tx_ring *xdp_ring)
1193{
1194	u16 ntc = xdp_ring->next_to_clean, ntu = xdp_ring->next_to_use;
1195	u32 xsk_frames = 0;
1196
1197	while (ntc != ntu) {
1198		struct ice_tx_buf *tx_buf = &xdp_ring->tx_buf[ntc];
1199
1200		if (tx_buf->type == ICE_TX_BUF_XSK_TX) {
1201			tx_buf->type = ICE_TX_BUF_EMPTY;
1202			xsk_buff_free(tx_buf->xdp);
1203		} else {
1204			xsk_frames++;
1205		}
 
1206
1207		ntc++;
1208		if (ntc >= xdp_ring->count)
1209			ntc = 0;
1210	}
1211
1212	if (xsk_frames)
1213		xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
1214}
v6.2
   1// SPDX-License-Identifier: GPL-2.0
   2/* Copyright (c) 2019, Intel Corporation. */
   3
   4#include <linux/bpf_trace.h>
   5#include <net/xdp_sock_drv.h>
   6#include <net/xdp.h>
   7#include "ice.h"
   8#include "ice_base.h"
   9#include "ice_type.h"
  10#include "ice_xsk.h"
  11#include "ice_txrx.h"
  12#include "ice_txrx_lib.h"
  13#include "ice_lib.h"
  14
  15static struct xdp_buff **ice_xdp_buf(struct ice_rx_ring *rx_ring, u32 idx)
  16{
  17	return &rx_ring->xdp_buf[idx];
  18}
  19
  20/**
  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
  24 */
  25static void ice_qp_reset_stats(struct ice_vsi *vsi, u16 q_idx)
  26{
  27	struct ice_vsi_stats *vsi_stat;
  28	struct ice_pf *pf;
  29
  30	pf = vsi->back;
  31	if (!pf->vsi_stats)
  32		return;
  33
  34	vsi_stat = pf->vsi_stats[vsi->idx];
  35	if (!vsi_stat)
  36		return;
  37
  38	memset(&vsi_stat->rx_ring_stats[q_idx]->rx_stats, 0,
  39	       sizeof(vsi_stat->rx_ring_stats[q_idx]->rx_stats));
  40	memset(&vsi_stat->tx_ring_stats[q_idx]->stats, 0,
  41	       sizeof(vsi_stat->tx_ring_stats[q_idx]->stats));
  42	if (ice_is_xdp_ena_vsi(vsi))
  43		memset(&vsi->xdp_rings[q_idx]->ring_stats->stats, 0,
  44		       sizeof(vsi->xdp_rings[q_idx]->ring_stats->stats));
  45}
  46
  47/**
  48 * ice_qp_clean_rings - Cleans all the rings of a given index
  49 * @vsi: VSI that contains rings of interest
  50 * @q_idx: ring index in array
  51 */
  52static void ice_qp_clean_rings(struct ice_vsi *vsi, u16 q_idx)
  53{
  54	ice_clean_tx_ring(vsi->tx_rings[q_idx]);
  55	if (ice_is_xdp_ena_vsi(vsi)) {
  56		synchronize_rcu();
  57		ice_clean_tx_ring(vsi->xdp_rings[q_idx]);
  58	}
  59	ice_clean_rx_ring(vsi->rx_rings[q_idx]);
  60}
  61
  62/**
  63 * ice_qvec_toggle_napi - Enables/disables NAPI for a given q_vector
  64 * @vsi: VSI that has netdev
  65 * @q_vector: q_vector that has NAPI context
  66 * @enable: true for enable, false for disable
  67 */
  68static void
  69ice_qvec_toggle_napi(struct ice_vsi *vsi, struct ice_q_vector *q_vector,
  70		     bool enable)
  71{
  72	if (!vsi->netdev || !q_vector)
  73		return;
  74
  75	if (enable)
  76		napi_enable(&q_vector->napi);
  77	else
  78		napi_disable(&q_vector->napi);
  79}
  80
  81/**
  82 * ice_qvec_dis_irq - Mask off queue interrupt generation on given ring
  83 * @vsi: the VSI that contains queue vector being un-configured
  84 * @rx_ring: Rx ring that will have its IRQ disabled
  85 * @q_vector: queue vector
  86 */
  87static void
  88ice_qvec_dis_irq(struct ice_vsi *vsi, struct ice_rx_ring *rx_ring,
  89		 struct ice_q_vector *q_vector)
  90{
  91	struct ice_pf *pf = vsi->back;
  92	struct ice_hw *hw = &pf->hw;
  93	int base = vsi->base_vector;
  94	u16 reg;
  95	u32 val;
  96
  97	/* QINT_TQCTL is being cleared in ice_vsi_stop_tx_ring, so handle
  98	 * here only QINT_RQCTL
  99	 */
 100	reg = rx_ring->reg_idx;
 101	val = rd32(hw, QINT_RQCTL(reg));
 102	val &= ~QINT_RQCTL_CAUSE_ENA_M;
 103	wr32(hw, QINT_RQCTL(reg), val);
 104
 105	if (q_vector) {
 106		u16 v_idx = q_vector->v_idx;
 107
 108		wr32(hw, GLINT_DYN_CTL(q_vector->reg_idx), 0);
 109		ice_flush(hw);
 110		synchronize_irq(pf->msix_entries[v_idx + base].vector);
 111	}
 112}
 113
 114/**
 115 * ice_qvec_cfg_msix - Enable IRQ for given queue vector
 116 * @vsi: the VSI that contains queue vector
 117 * @q_vector: queue vector
 
 118 */
 119static void
 120ice_qvec_cfg_msix(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
 121{
 122	u16 reg_idx = q_vector->reg_idx;
 123	struct ice_pf *pf = vsi->back;
 124	struct ice_hw *hw = &pf->hw;
 125	struct ice_tx_ring *tx_ring;
 126	struct ice_rx_ring *rx_ring;
 127
 128	ice_cfg_itr(hw, q_vector);
 129
 130	ice_for_each_tx_ring(tx_ring, q_vector->tx)
 131		ice_cfg_txq_interrupt(vsi, tx_ring->reg_idx, reg_idx,
 132				      q_vector->tx.itr_idx);
 133
 134	ice_for_each_rx_ring(rx_ring, q_vector->rx)
 135		ice_cfg_rxq_interrupt(vsi, rx_ring->reg_idx, reg_idx,
 136				      q_vector->rx.itr_idx);
 
 
 
 
 137
 138	ice_flush(hw);
 139}
 140
 141/**
 142 * ice_qvec_ena_irq - Enable IRQ for given queue vector
 143 * @vsi: the VSI that contains queue vector
 144 * @q_vector: queue vector
 145 */
 146static void ice_qvec_ena_irq(struct ice_vsi *vsi, struct ice_q_vector *q_vector)
 147{
 148	struct ice_pf *pf = vsi->back;
 149	struct ice_hw *hw = &pf->hw;
 150
 151	ice_irq_dynamic_ena(hw, vsi, q_vector);
 152
 153	ice_flush(hw);
 154}
 155
 156/**
 157 * ice_qp_dis - Disables a queue pair
 158 * @vsi: VSI of interest
 159 * @q_idx: ring index in array
 160 *
 161 * Returns 0 on success, negative on failure.
 162 */
 163static int ice_qp_dis(struct ice_vsi *vsi, u16 q_idx)
 164{
 165	struct ice_txq_meta txq_meta = { };
 166	struct ice_q_vector *q_vector;
 167	struct ice_tx_ring *tx_ring;
 168	struct ice_rx_ring *rx_ring;
 169	int timeout = 50;
 170	int err;
 171
 172	if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
 173		return -EINVAL;
 174
 175	tx_ring = vsi->tx_rings[q_idx];
 176	rx_ring = vsi->rx_rings[q_idx];
 177	q_vector = rx_ring->q_vector;
 178
 179	while (test_and_set_bit(ICE_CFG_BUSY, vsi->state)) {
 180		timeout--;
 181		if (!timeout)
 182			return -EBUSY;
 183		usleep_range(1000, 2000);
 184	}
 185	netif_tx_stop_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
 186
 187	ice_qvec_dis_irq(vsi, rx_ring, q_vector);
 
 188
 189	ice_fill_txq_meta(vsi, tx_ring, &txq_meta);
 190	err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, tx_ring, &txq_meta);
 191	if (err)
 192		return err;
 193	if (ice_is_xdp_ena_vsi(vsi)) {
 194		struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
 195
 196		memset(&txq_meta, 0, sizeof(txq_meta));
 197		ice_fill_txq_meta(vsi, xdp_ring, &txq_meta);
 198		err = ice_vsi_stop_tx_ring(vsi, ICE_NO_RESET, 0, xdp_ring,
 199					   &txq_meta);
 200		if (err)
 201			return err;
 202	}
 203	err = ice_vsi_ctrl_one_rx_ring(vsi, false, q_idx, true);
 204	if (err)
 205		return err;
 206	ice_clean_rx_ring(rx_ring);
 207
 208	ice_qvec_toggle_napi(vsi, q_vector, false);
 209	ice_qp_clean_rings(vsi, q_idx);
 210	ice_qp_reset_stats(vsi, q_idx);
 211
 212	return 0;
 213}
 214
 215/**
 216 * ice_qp_ena - Enables a queue pair
 217 * @vsi: VSI of interest
 218 * @q_idx: ring index in array
 219 *
 220 * Returns 0 on success, negative on failure.
 221 */
 222static int ice_qp_ena(struct ice_vsi *vsi, u16 q_idx)
 223{
 224	struct ice_aqc_add_tx_qgrp *qg_buf;
 225	struct ice_q_vector *q_vector;
 226	struct ice_tx_ring *tx_ring;
 227	struct ice_rx_ring *rx_ring;
 228	u16 size;
 229	int err;
 230
 231	if (q_idx >= vsi->num_rxq || q_idx >= vsi->num_txq)
 232		return -EINVAL;
 233
 234	size = struct_size(qg_buf, txqs, 1);
 235	qg_buf = kzalloc(size, GFP_KERNEL);
 236	if (!qg_buf)
 237		return -ENOMEM;
 238
 239	qg_buf->num_txqs = 1;
 240
 241	tx_ring = vsi->tx_rings[q_idx];
 242	rx_ring = vsi->rx_rings[q_idx];
 243	q_vector = rx_ring->q_vector;
 244
 245	err = ice_vsi_cfg_txq(vsi, tx_ring, qg_buf);
 246	if (err)
 247		goto free_buf;
 248
 249	if (ice_is_xdp_ena_vsi(vsi)) {
 250		struct ice_tx_ring *xdp_ring = vsi->xdp_rings[q_idx];
 251
 252		memset(qg_buf, 0, size);
 253		qg_buf->num_txqs = 1;
 254		err = ice_vsi_cfg_txq(vsi, xdp_ring, qg_buf);
 255		if (err)
 256			goto free_buf;
 257		ice_set_ring_xdp(xdp_ring);
 258		ice_tx_xsk_pool(vsi, q_idx);
 259	}
 260
 261	err = ice_vsi_cfg_rxq(rx_ring);
 262	if (err)
 263		goto free_buf;
 264
 265	ice_qvec_cfg_msix(vsi, q_vector);
 
 266
 267	err = ice_vsi_ctrl_one_rx_ring(vsi, true, q_idx, true);
 268	if (err)
 269		goto free_buf;
 270
 271	clear_bit(ICE_CFG_BUSY, vsi->state);
 272	ice_qvec_toggle_napi(vsi, q_vector, true);
 273	ice_qvec_ena_irq(vsi, q_vector);
 274
 275	netif_tx_start_queue(netdev_get_tx_queue(vsi->netdev, q_idx));
 276free_buf:
 277	kfree(qg_buf);
 278	return err;
 
 
 
 
 
 279}
 280
 281/**
 282 * ice_xsk_pool_disable - disable a buffer pool region
 283 * @vsi: Current VSI
 284 * @qid: queue ID
 285 *
 286 * Returns 0 on success, negative on failure
 287 */
 288static int ice_xsk_pool_disable(struct ice_vsi *vsi, u16 qid)
 289{
 290	struct xsk_buff_pool *pool = xsk_get_pool_from_qid(vsi->netdev, qid);
 291
 292	if (!pool)
 293		return -EINVAL;
 294
 295	clear_bit(qid, vsi->af_xdp_zc_qps);
 296	xsk_pool_dma_unmap(pool, ICE_RX_DMA_ATTR);
 297
 298	return 0;
 299}
 300
 301/**
 302 * ice_xsk_pool_enable - enable a buffer pool region
 303 * @vsi: Current VSI
 304 * @pool: pointer to a requested buffer pool region
 305 * @qid: queue ID
 306 *
 307 * Returns 0 on success, negative on failure
 308 */
 309static int
 310ice_xsk_pool_enable(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
 311{
 312	int err;
 313
 314	if (vsi->type != ICE_VSI_PF)
 315		return -EINVAL;
 316
 317	if (qid >= vsi->netdev->real_num_rx_queues ||
 318	    qid >= vsi->netdev->real_num_tx_queues)
 319		return -EINVAL;
 320
 321	err = xsk_pool_dma_map(pool, ice_pf_to_dev(vsi->back),
 322			       ICE_RX_DMA_ATTR);
 323	if (err)
 324		return err;
 325
 326	set_bit(qid, vsi->af_xdp_zc_qps);
 327
 328	return 0;
 329}
 330
 331/**
 332 * ice_realloc_rx_xdp_bufs - reallocate for either XSK or normal buffer
 333 * @rx_ring: Rx ring
 334 * @pool_present: is pool for XSK present
 335 *
 336 * Try allocating memory and return ENOMEM, if failed to allocate.
 337 * If allocation was successful, substitute buffer with allocated one.
 338 * Returns 0 on success, negative on failure
 339 */
 340static int
 341ice_realloc_rx_xdp_bufs(struct ice_rx_ring *rx_ring, bool pool_present)
 342{
 343	size_t elem_size = pool_present ? sizeof(*rx_ring->xdp_buf) :
 344					  sizeof(*rx_ring->rx_buf);
 345	void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
 346
 347	if (!sw_ring)
 348		return -ENOMEM;
 349
 350	if (pool_present) {
 351		kfree(rx_ring->rx_buf);
 352		rx_ring->rx_buf = NULL;
 353		rx_ring->xdp_buf = sw_ring;
 354	} else {
 355		kfree(rx_ring->xdp_buf);
 356		rx_ring->xdp_buf = NULL;
 357		rx_ring->rx_buf = sw_ring;
 358	}
 359
 360	return 0;
 361}
 362
 363/**
 364 * ice_realloc_zc_buf - reallocate XDP ZC queue pairs
 365 * @vsi: Current VSI
 366 * @zc: is zero copy set
 367 *
 368 * Reallocate buffer for rx_rings that might be used by XSK.
 369 * XDP requires more memory, than rx_buf provides.
 370 * Returns 0 on success, negative on failure
 371 */
 372int ice_realloc_zc_buf(struct ice_vsi *vsi, bool zc)
 373{
 374	struct ice_rx_ring *rx_ring;
 375	unsigned long q;
 
 
 
 
 
 376
 377	for_each_set_bit(q, vsi->af_xdp_zc_qps,
 378			 max_t(int, vsi->alloc_txq, vsi->alloc_rxq)) {
 379		rx_ring = vsi->rx_rings[q];
 380		if (ice_realloc_rx_xdp_bufs(rx_ring, zc))
 381			return -ENOMEM;
 382	}
 383
 384	return 0;
 385}
 386
 387/**
 388 * ice_xsk_pool_setup - enable/disable a buffer pool region depending on its state
 389 * @vsi: Current VSI
 390 * @pool: buffer pool to enable/associate to a ring, NULL to disable
 391 * @qid: queue ID
 392 *
 393 * Returns 0 on success, negative on failure
 394 */
 395int ice_xsk_pool_setup(struct ice_vsi *vsi, struct xsk_buff_pool *pool, u16 qid)
 396{
 397	bool if_running, pool_present = !!pool;
 398	int ret = 0, pool_failure = 0;
 399
 400	if (qid >= vsi->num_rxq || qid >= vsi->num_txq) {
 401		netdev_err(vsi->netdev, "Please use queue id in scope of combined queues count\n");
 402		pool_failure = -EINVAL;
 403		goto failure;
 404	}
 405
 406	if_running = netif_running(vsi->netdev) && ice_is_xdp_ena_vsi(vsi);
 
 407
 408	if (if_running) {
 409		struct ice_rx_ring *rx_ring = vsi->rx_rings[qid];
 410
 411		ret = ice_qp_dis(vsi, qid);
 412		if (ret) {
 413			netdev_err(vsi->netdev, "ice_qp_dis error = %d\n", ret);
 414			goto xsk_pool_if_up;
 415		}
 416
 417		ret = ice_realloc_rx_xdp_bufs(rx_ring, pool_present);
 418		if (ret)
 419			goto xsk_pool_if_up;
 420	}
 421
 422	pool_failure = pool_present ? ice_xsk_pool_enable(vsi, pool, qid) :
 423				      ice_xsk_pool_disable(vsi, qid);
 424
 425xsk_pool_if_up:
 426	if (if_running) {
 427		ret = ice_qp_ena(vsi, qid);
 428		if (!ret && pool_present)
 429			napi_schedule(&vsi->rx_rings[qid]->xdp_ring->q_vector->napi);
 430		else if (ret)
 431			netdev_err(vsi->netdev, "ice_qp_ena error = %d\n", ret);
 432	}
 433
 434failure:
 435	if (pool_failure) {
 436		netdev_err(vsi->netdev, "Could not %sable buffer pool, error = %d\n",
 437			   pool_present ? "en" : "dis", pool_failure);
 438		return pool_failure;
 439	}
 440
 441	return ret;
 442}
 443
 444/**
 445 * ice_fill_rx_descs - pick buffers from XSK buffer pool and use it
 446 * @pool: XSK Buffer pool to pull the buffers from
 447 * @xdp: SW ring of xdp_buff that will hold the buffers
 448 * @rx_desc: Pointer to Rx descriptors that will be filled
 449 * @count: The number of buffers to allocate
 450 *
 451 * This function allocates a number of Rx buffers from the fill ring
 452 * or the internal recycle mechanism and places them on the Rx ring.
 453 *
 454 * Note that ring wrap should be handled by caller of this function.
 455 *
 456 * Returns the amount of allocated Rx descriptors
 457 */
 458static u16 ice_fill_rx_descs(struct xsk_buff_pool *pool, struct xdp_buff **xdp,
 459			     union ice_32b_rx_flex_desc *rx_desc, u16 count)
 460{
 461	dma_addr_t dma;
 462	u16 buffs;
 463	int i;
 464
 465	buffs = xsk_buff_alloc_batch(pool, xdp, count);
 466	for (i = 0; i < buffs; i++) {
 467		dma = xsk_buff_xdp_get_dma(*xdp);
 468		rx_desc->read.pkt_addr = cpu_to_le64(dma);
 469		rx_desc->wb.status_error0 = 0;
 470
 
 
 
 
 
 471		rx_desc++;
 472		xdp++;
 473	}
 474
 475	return buffs;
 476}
 477
 478/**
 479 * __ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
 480 * @rx_ring: Rx ring
 
 481 * @count: The number of buffers to allocate
 482 *
 483 * Place the @count of descriptors onto Rx ring. Handle the ring wrap
 484 * for case where space from next_to_use up to the end of ring is less
 485 * than @count. Finally do a tail bump.
 486 *
 487 * Returns true if all allocations were successful, false if any fail.
 488 */
 489static bool __ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
 
 490{
 491	u32 nb_buffs_extra = 0, nb_buffs = 0;
 492	union ice_32b_rx_flex_desc *rx_desc;
 493	u16 ntu = rx_ring->next_to_use;
 494	u16 total_count = count;
 495	struct xdp_buff **xdp;
 496
 497	rx_desc = ICE_RX_DESC(rx_ring, ntu);
 498	xdp = ice_xdp_buf(rx_ring, ntu);
 499
 500	if (ntu + count >= rx_ring->count) {
 501		nb_buffs_extra = ice_fill_rx_descs(rx_ring->xsk_pool, xdp,
 502						   rx_desc,
 503						   rx_ring->count - ntu);
 504		if (nb_buffs_extra != rx_ring->count - ntu) {
 505			ntu += nb_buffs_extra;
 506			goto exit;
 507		}
 508		rx_desc = ICE_RX_DESC(rx_ring, 0);
 509		xdp = ice_xdp_buf(rx_ring, 0);
 510		ntu = 0;
 511		count -= nb_buffs_extra;
 512		ice_release_rx_desc(rx_ring, 0);
 513	}
 514
 515	nb_buffs = ice_fill_rx_descs(rx_ring->xsk_pool, xdp, rx_desc, count);
 516
 517	ntu += nb_buffs;
 518	if (ntu == rx_ring->count)
 519		ntu = 0;
 520
 521exit:
 522	if (rx_ring->next_to_use != ntu)
 523		ice_release_rx_desc(rx_ring, ntu);
 524
 525	return total_count == (nb_buffs_extra + nb_buffs);
 526}
 527
 528/**
 529 * ice_alloc_rx_bufs_zc - allocate a number of Rx buffers
 530 * @rx_ring: Rx ring
 
 531 * @count: The number of buffers to allocate
 532 *
 533 * Wrapper for internal allocation routine; figure out how many tail
 534 * bumps should take place based on the given threshold
 535 *
 536 * Returns true if all calls to internal alloc routine succeeded
 537 */
 538bool ice_alloc_rx_bufs_zc(struct ice_rx_ring *rx_ring, u16 count)
 
 539{
 540	u16 rx_thresh = ICE_RING_QUARTER(rx_ring);
 541	u16 leftover, i, tail_bumps;
 542
 543	tail_bumps = count / rx_thresh;
 544	leftover = count - (tail_bumps * rx_thresh);
 545
 546	for (i = 0; i < tail_bumps; i++)
 547		if (!__ice_alloc_rx_bufs_zc(rx_ring, rx_thresh))
 548			return false;
 549	return __ice_alloc_rx_bufs_zc(rx_ring, leftover);
 550}
 551
 552/**
 553 * ice_bump_ntc - Bump the next_to_clean counter of an Rx ring
 554 * @rx_ring: Rx ring
 555 */
 556static void ice_bump_ntc(struct ice_rx_ring *rx_ring)
 557{
 558	int ntc = rx_ring->next_to_clean + 1;
 559
 560	ntc = (ntc < rx_ring->count) ? ntc : 0;
 561	rx_ring->next_to_clean = ntc;
 562	prefetch(ICE_RX_DESC(rx_ring, ntc));
 563}
 564
 565/**
 566 * ice_construct_skb_zc - Create an sk_buff from zero-copy buffer
 567 * @rx_ring: Rx ring
 568 * @xdp: Pointer to XDP buffer
 569 *
 570 * This function allocates a new skb from a zero-copy Rx buffer.
 571 *
 572 * Returns the skb on success, NULL on failure.
 573 */
 574static struct sk_buff *
 575ice_construct_skb_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp)
 576{
 577	unsigned int totalsize = xdp->data_end - xdp->data_meta;
 578	unsigned int metasize = xdp->data - xdp->data_meta;
 
 579	struct sk_buff *skb;
 
 580
 
 
 
 
 581	net_prefetch(xdp->data_meta);
 582
 583	skb = __napi_alloc_skb(&rx_ring->q_vector->napi, totalsize,
 584			       GFP_ATOMIC | __GFP_NOWARN);
 585	if (unlikely(!skb))
 586		return NULL;
 587
 588	memcpy(__skb_put(skb, totalsize), xdp->data_meta,
 589	       ALIGN(totalsize, sizeof(long)));
 590
 591	if (metasize) {
 592		skb_metadata_set(skb, metasize);
 593		__skb_pull(skb, metasize);
 594	}
 595
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 596	xsk_buff_free(xdp);
 597	return skb;
 598}
 599
 600/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 601 * ice_run_xdp_zc - Executes an XDP program in zero-copy path
 602 * @rx_ring: Rx ring
 603 * @xdp: xdp_buff used as input to the XDP program
 604 * @xdp_prog: XDP program to run
 605 * @xdp_ring: ring to be used for XDP_TX action
 
 606 *
 607 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}
 608 */
 609static int
 610ice_run_xdp_zc(struct ice_rx_ring *rx_ring, struct xdp_buff *xdp,
 611	       struct bpf_prog *xdp_prog, struct ice_tx_ring *xdp_ring)
 
 612{
 613	int err, result = ICE_XDP_PASS;
 614	u32 act;
 615
 616	act = bpf_prog_run_xdp(xdp_prog, xdp);
 617
 618	if (likely(act == XDP_REDIRECT)) {
 619		err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
 620		if (!err)
 621			return ICE_XDP_REDIR;
 622		if (xsk_uses_need_wakeup(rx_ring->xsk_pool) && err == -ENOBUFS)
 623			result = ICE_XDP_EXIT;
 624		else
 625			result = ICE_XDP_CONSUMED;
 626		goto out_failure;
 627	}
 628
 629	switch (act) {
 630	case XDP_PASS:
 631		break;
 632	case XDP_TX:
 633		result = ice_xmit_xdp_buff(xdp, xdp_ring);
 634		if (result == ICE_XDP_CONSUMED)
 635			goto out_failure;
 636		break;
 637	case XDP_DROP:
 638		result = ICE_XDP_CONSUMED;
 639		break;
 640	default:
 641		bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
 642		fallthrough;
 643	case XDP_ABORTED:
 644		result = ICE_XDP_CONSUMED;
 645out_failure:
 646		trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
 647		break;
 648	}
 649
 650	return result;
 651}
 652
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 653/**
 654 * ice_clean_rx_irq_zc - consumes packets from the hardware ring
 655 * @rx_ring: AF_XDP Rx ring
 
 656 * @budget: NAPI budget
 657 *
 658 * Returns number of processed packets on success, remaining budget on failure.
 659 */
 660int ice_clean_rx_irq_zc(struct ice_rx_ring *rx_ring, int budget)
 
 
 661{
 662	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
 
 
 
 663	struct ice_tx_ring *xdp_ring;
 664	unsigned int xdp_xmit = 0;
 665	struct bpf_prog *xdp_prog;
 
 666	bool failure = false;
 667	int entries_to_alloc;
 668
 669	/* ZC patch is enabled only when XDP program is set,
 670	 * so here it can not be NULL
 671	 */
 672	xdp_prog = READ_ONCE(rx_ring->xdp_prog);
 673	xdp_ring = rx_ring->xdp_ring;
 674
 
 
 
 675	while (likely(total_rx_packets < (unsigned int)budget)) {
 676		union ice_32b_rx_flex_desc *rx_desc;
 677		unsigned int size, xdp_res = 0;
 678		struct xdp_buff *xdp;
 679		struct sk_buff *skb;
 680		u16 stat_err_bits;
 681		u16 vlan_tag = 0;
 682		u16 rx_ptype;
 683
 684		rx_desc = ICE_RX_DESC(rx_ring, rx_ring->next_to_clean);
 685
 686		stat_err_bits = BIT(ICE_RX_FLEX_DESC_STATUS0_DD_S);
 687		if (!ice_test_staterr(rx_desc->wb.status_error0, stat_err_bits))
 688			break;
 689
 690		/* This memory barrier is needed to keep us from reading
 691		 * any other fields out of the rx_desc until we have
 692		 * verified the descriptor has been written back.
 693		 */
 694		dma_rmb();
 695
 696		if (unlikely(rx_ring->next_to_clean == rx_ring->next_to_use))
 697			break;
 698
 699		xdp = *ice_xdp_buf(rx_ring, rx_ring->next_to_clean);
 700
 701		size = le16_to_cpu(rx_desc->wb.pkt_len) &
 702				   ICE_RX_FLX_DESC_PKT_LEN_M;
 703		if (!size) {
 704			xdp->data = NULL;
 705			xdp->data_end = NULL;
 706			xdp->data_hard_start = NULL;
 707			xdp->data_meta = NULL;
 708			goto construct_skb;
 
 
 709		}
 710
 711		xsk_buff_set_size(xdp, size);
 712		xsk_buff_dma_sync_for_cpu(xdp, rx_ring->xsk_pool);
 
 
 
 713
 714		xdp_res = ice_run_xdp_zc(rx_ring, xdp, xdp_prog, xdp_ring);
 
 715		if (likely(xdp_res & (ICE_XDP_TX | ICE_XDP_REDIR))) {
 716			xdp_xmit |= xdp_res;
 717		} else if (xdp_res == ICE_XDP_EXIT) {
 718			failure = true;
 
 
 719			break;
 720		} else if (xdp_res == ICE_XDP_CONSUMED) {
 721			xsk_buff_free(xdp);
 722		} else if (xdp_res == ICE_XDP_PASS) {
 723			goto construct_skb;
 724		}
 725
 726		total_rx_bytes += size;
 727		total_rx_packets++;
 728
 729		ice_bump_ntc(rx_ring);
 
 730		continue;
 731
 732construct_skb:
 733		/* XDP_PASS path */
 734		skb = ice_construct_skb_zc(rx_ring, xdp);
 735		if (!skb) {
 736			rx_ring->ring_stats->rx_stats.alloc_buf_failed++;
 737			break;
 738		}
 739
 740		ice_bump_ntc(rx_ring);
 
 741
 742		if (eth_skb_pad(skb)) {
 743			skb = NULL;
 744			continue;
 745		}
 746
 747		total_rx_bytes += skb->len;
 748		total_rx_packets++;
 749
 750		vlan_tag = ice_get_vlan_tag_from_rx_desc(rx_desc);
 751
 752		rx_ptype = le16_to_cpu(rx_desc->wb.ptype_flex_flags0) &
 753				       ICE_RX_FLEX_DESC_PTYPE_M;
 754
 755		ice_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
 756		ice_receive_skb(rx_ring, skb, vlan_tag);
 757	}
 758
 759	entries_to_alloc = ICE_DESC_UNUSED(rx_ring);
 
 760	if (entries_to_alloc > ICE_RING_QUARTER(rx_ring))
 761		failure |= !ice_alloc_rx_bufs_zc(rx_ring, entries_to_alloc);
 
 762
 763	ice_finalize_xdp_rx(xdp_ring, xdp_xmit);
 764	ice_update_rx_ring_stats(rx_ring, total_rx_packets, total_rx_bytes);
 765
 766	if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
 767		if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
 768			xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
 
 
 
 769		else
 770			xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
 771
 772		return (int)total_rx_packets;
 773	}
 774
 775	return failure ? budget : (int)total_rx_packets;
 776}
 777
 778/**
 779 * ice_clean_xdp_tx_buf - Free and unmap XDP Tx buffer
 780 * @xdp_ring: XDP Tx ring
 781 * @tx_buf: Tx buffer to clean
 782 */
 783static void
 784ice_clean_xdp_tx_buf(struct ice_tx_ring *xdp_ring, struct ice_tx_buf *tx_buf)
 785{
 786	page_frag_free(tx_buf->raw_buf);
 787	xdp_ring->xdp_tx_active--;
 788	dma_unmap_single(xdp_ring->dev, dma_unmap_addr(tx_buf, dma),
 789			 dma_unmap_len(tx_buf, len), DMA_TO_DEVICE);
 790	dma_unmap_len_set(tx_buf, len, 0);
 791}
 792
 793/**
 794 * ice_clean_xdp_irq_zc - produce AF_XDP descriptors to CQ
 795 * @xdp_ring: XDP Tx ring
 796 */
 797static void ice_clean_xdp_irq_zc(struct ice_tx_ring *xdp_ring)
 798{
 799	u16 ntc = xdp_ring->next_to_clean;
 800	struct ice_tx_desc *tx_desc;
 801	u16 cnt = xdp_ring->count;
 802	struct ice_tx_buf *tx_buf;
 803	u16 completed_frames = 0;
 804	u16 xsk_frames = 0;
 805	u16 last_rs;
 806	int i;
 807
 808	last_rs = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : cnt - 1;
 809	tx_desc = ICE_TX_DESC(xdp_ring, last_rs);
 810	if ((tx_desc->cmd_type_offset_bsz &
 811	    cpu_to_le64(ICE_TX_DESC_DTYPE_DESC_DONE))) {
 812		if (last_rs >= ntc)
 813			completed_frames = last_rs - ntc + 1;
 814		else
 815			completed_frames = last_rs + cnt - ntc + 1;
 816	}
 817
 818	if (!completed_frames)
 819		return;
 820
 821	if (likely(!xdp_ring->xdp_tx_active)) {
 822		xsk_frames = completed_frames;
 823		goto skip;
 824	}
 825
 826	ntc = xdp_ring->next_to_clean;
 827	for (i = 0; i < completed_frames; i++) {
 828		tx_buf = &xdp_ring->tx_buf[ntc];
 829
 830		if (tx_buf->raw_buf) {
 831			ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
 832			tx_buf->raw_buf = NULL;
 833		} else {
 834			xsk_frames++;
 835		}
 836
 837		ntc++;
 838		if (ntc >= xdp_ring->count)
 839			ntc = 0;
 840	}
 841skip:
 842	tx_desc->cmd_type_offset_bsz = 0;
 843	xdp_ring->next_to_clean += completed_frames;
 844	if (xdp_ring->next_to_clean >= cnt)
 845		xdp_ring->next_to_clean -= cnt;
 846	if (xsk_frames)
 847		xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
 848}
 849
 850/**
 851 * ice_xmit_pkt - produce a single HW Tx descriptor out of AF_XDP descriptor
 852 * @xdp_ring: XDP ring to produce the HW Tx descriptor on
 
 853 * @desc: AF_XDP descriptor to pull the DMA address and length from
 854 * @total_bytes: bytes accumulator that will be used for stats update
 855 */
 856static void ice_xmit_pkt(struct ice_tx_ring *xdp_ring, struct xdp_desc *desc,
 
 857			 unsigned int *total_bytes)
 858{
 859	struct ice_tx_desc *tx_desc;
 860	dma_addr_t dma;
 861
 862	dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
 863	xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
 864
 865	tx_desc = ICE_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
 866	tx_desc->buf_addr = cpu_to_le64(dma);
 867	tx_desc->cmd_type_offset_bsz = ice_build_ctob(ICE_TX_DESC_CMD_EOP,
 868						      0, desc->len, 0);
 869
 870	*total_bytes += desc->len;
 871}
 872
 873/**
 874 * ice_xmit_pkt_batch - produce a batch of HW Tx descriptors out of AF_XDP descriptors
 875 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
 
 876 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
 877 * @total_bytes: bytes accumulator that will be used for stats update
 878 */
 879static void ice_xmit_pkt_batch(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
 
 
 880			       unsigned int *total_bytes)
 881{
 882	u16 ntu = xdp_ring->next_to_use;
 883	struct ice_tx_desc *tx_desc;
 884	u32 i;
 885
 886	loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
 887		dma_addr_t dma;
 888
 889		dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, descs[i].addr);
 890		xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, descs[i].len);
 891
 892		tx_desc = ICE_TX_DESC(xdp_ring, ntu++);
 893		tx_desc->buf_addr = cpu_to_le64(dma);
 894		tx_desc->cmd_type_offset_bsz = ice_build_ctob(ICE_TX_DESC_CMD_EOP,
 895							      0, descs[i].len, 0);
 896
 897		*total_bytes += descs[i].len;
 898	}
 899
 900	xdp_ring->next_to_use = ntu;
 901}
 902
 903/**
 904 * ice_fill_tx_hw_ring - produce the number of Tx descriptors onto ring
 905 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
 
 906 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from
 907 * @nb_pkts: count of packets to be send
 908 * @total_bytes: bytes accumulator that will be used for stats update
 909 */
 910static void ice_fill_tx_hw_ring(struct ice_tx_ring *xdp_ring, struct xdp_desc *descs,
 911				u32 nb_pkts, unsigned int *total_bytes)
 
 
 912{
 913	u32 batched, leftover, i;
 914
 915	batched = ALIGN_DOWN(nb_pkts, PKTS_PER_BATCH);
 916	leftover = nb_pkts & (PKTS_PER_BATCH - 1);
 917	for (i = 0; i < batched; i += PKTS_PER_BATCH)
 918		ice_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
 919	for (; i < batched + leftover; i++)
 920		ice_xmit_pkt(xdp_ring, &descs[i], total_bytes);
 921}
 922
 923/**
 924 * ice_set_rs_bit - set RS bit on last produced descriptor (one behind current NTU)
 925 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
 926 */
 927static void ice_set_rs_bit(struct ice_tx_ring *xdp_ring)
 928{
 929	u16 ntu = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : xdp_ring->count - 1;
 930	struct ice_tx_desc *tx_desc;
 931
 932	tx_desc = ICE_TX_DESC(xdp_ring, ntu);
 933	tx_desc->cmd_type_offset_bsz |=
 934		cpu_to_le64(ICE_TX_DESC_CMD_RS << ICE_TXD_QW1_CMD_S);
 935}
 936
 937/**
 938 * ice_xmit_zc - take entries from XSK Tx ring and place them onto HW Tx ring
 939 * @xdp_ring: XDP ring to produce the HW Tx descriptors on
 
 940 *
 941 * Returns true if there is no more work that needs to be done, false otherwise
 942 */
 943bool ice_xmit_zc(struct ice_tx_ring *xdp_ring)
 944{
 945	struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
 946	u32 nb_pkts, nb_processed = 0;
 947	unsigned int total_bytes = 0;
 948	int budget;
 949
 950	ice_clean_xdp_irq_zc(xdp_ring);
 
 
 
 
 951
 952	budget = ICE_DESC_UNUSED(xdp_ring);
 953	budget = min_t(u16, budget, ICE_RING_QUARTER(xdp_ring));
 954
 955	nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
 956	if (!nb_pkts)
 957		return true;
 958
 959	if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
 960		nb_processed = xdp_ring->count - xdp_ring->next_to_use;
 961		ice_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
 
 962		xdp_ring->next_to_use = 0;
 963	}
 964
 965	ice_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
 966			    &total_bytes);
 967
 968	ice_set_rs_bit(xdp_ring);
 969	ice_xdp_ring_update_tail(xdp_ring);
 970	ice_update_tx_ring_stats(xdp_ring, nb_pkts, total_bytes);
 971
 972	if (xsk_uses_need_wakeup(xdp_ring->xsk_pool))
 973		xsk_set_tx_need_wakeup(xdp_ring->xsk_pool);
 974
 975	return nb_pkts < budget;
 976}
 977
 978/**
 979 * ice_xsk_wakeup - Implements ndo_xsk_wakeup
 980 * @netdev: net_device
 981 * @queue_id: queue to wake up
 982 * @flags: ignored in our case, since we have Rx and Tx in the same NAPI
 983 *
 984 * Returns negative on error, zero otherwise.
 985 */
 986int
 987ice_xsk_wakeup(struct net_device *netdev, u32 queue_id,
 988	       u32 __always_unused flags)
 989{
 990	struct ice_netdev_priv *np = netdev_priv(netdev);
 991	struct ice_q_vector *q_vector;
 992	struct ice_vsi *vsi = np->vsi;
 993	struct ice_tx_ring *ring;
 994
 995	if (test_bit(ICE_VSI_DOWN, vsi->state))
 996		return -ENETDOWN;
 997
 998	if (!ice_is_xdp_ena_vsi(vsi))
 999		return -EINVAL;
1000
1001	if (queue_id >= vsi->num_txq || queue_id >= vsi->num_rxq)
1002		return -EINVAL;
1003
1004	ring = vsi->rx_rings[queue_id]->xdp_ring;
1005
1006	if (!ring->xsk_pool)
1007		return -EINVAL;
1008
1009	/* The idea here is that if NAPI is running, mark a miss, so
1010	 * it will run again. If not, trigger an interrupt and
1011	 * schedule the NAPI from interrupt context. If NAPI would be
1012	 * scheduled here, the interrupt affinity would not be
1013	 * honored.
1014	 */
1015	q_vector = ring->q_vector;
1016	if (!napi_if_scheduled_mark_missed(&q_vector->napi))
1017		ice_trigger_sw_intr(&vsi->back->hw, q_vector);
1018
1019	return 0;
1020}
1021
1022/**
1023 * ice_xsk_any_rx_ring_ena - Checks if Rx rings have AF_XDP buff pool attached
1024 * @vsi: VSI to be checked
1025 *
1026 * Returns true if any of the Rx rings has an AF_XDP buff pool attached
1027 */
1028bool ice_xsk_any_rx_ring_ena(struct ice_vsi *vsi)
1029{
1030	int i;
1031
1032	ice_for_each_rxq(vsi, i) {
1033		if (xsk_get_pool_from_qid(vsi->netdev, i))
1034			return true;
1035	}
1036
1037	return false;
1038}
1039
1040/**
1041 * ice_xsk_clean_rx_ring - clean buffer pool queues connected to a given Rx ring
1042 * @rx_ring: ring to be cleaned
1043 */
1044void ice_xsk_clean_rx_ring(struct ice_rx_ring *rx_ring)
1045{
1046	u16 ntc = rx_ring->next_to_clean;
1047	u16 ntu = rx_ring->next_to_use;
1048
1049	while (ntc != ntu) {
1050		struct xdp_buff *xdp = *ice_xdp_buf(rx_ring, ntc);
1051
1052		xsk_buff_free(xdp);
1053		ntc++;
1054		if (ntc >= rx_ring->count)
1055			ntc = 0;
1056	}
1057}
1058
1059/**
1060 * ice_xsk_clean_xdp_ring - Clean the XDP Tx ring and its buffer pool queues
1061 * @xdp_ring: XDP_Tx ring
1062 */
1063void ice_xsk_clean_xdp_ring(struct ice_tx_ring *xdp_ring)
1064{
1065	u16 ntc = xdp_ring->next_to_clean, ntu = xdp_ring->next_to_use;
1066	u32 xsk_frames = 0;
1067
1068	while (ntc != ntu) {
1069		struct ice_tx_buf *tx_buf = &xdp_ring->tx_buf[ntc];
1070
1071		if (tx_buf->raw_buf)
1072			ice_clean_xdp_tx_buf(xdp_ring, tx_buf);
1073		else
 
1074			xsk_frames++;
1075
1076		tx_buf->raw_buf = NULL;
1077
1078		ntc++;
1079		if (ntc >= xdp_ring->count)
1080			ntc = 0;
1081	}
1082
1083	if (xsk_frames)
1084		xsk_tx_completed(xdp_ring->xsk_pool, xsk_frames);
1085}