1215 lines · c
1// SPDX-License-Identifier: GPL-2.02/* 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 index22 * @vsi: VSI that contains rings of interest23 * @q_idx: ring index in array24 */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 index49 * @vsi: VSI that contains rings of interest50 * @q_idx: ring index in array51 */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_vector62 * @vsi: VSI that has netdev63 * @q_vector: q_vector that has NAPI context64 * @enable: true for enable, false for disable65 */66static void67ice_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 else76 napi_disable(&q_vector->napi);77}78 79/**80 * ice_qvec_dis_irq - Mask off queue interrupt generation on given ring81 * @vsi: the VSI that contains queue vector being un-configured82 * @rx_ring: Rx ring that will have its IRQ disabled83 * @q_vector: queue vector84 */85static void86ice_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 handle95 * here only QINT_RQCTL96 */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 vector111 * @vsi: the VSI that contains queue vector112 * @q_vector: queue vector113 * @qid: queue index114 */115static void116ice_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 vector142 * @vsi: the VSI that contains queue vector143 * @q_vector: queue vector144 */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 pair157 * @vsi: VSI of interest158 * @q_idx: ring index in array159 *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 pair209 * @vsi: VSI of interest210 * @q_idx: ring index in array211 *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 region262 * @vsi: Current VSI263 * @qid: queue ID264 *265 * Returns 0 on success, negative on failure266 */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 region281 * @vsi: Current VSI282 * @pool: pointer to a requested buffer pool region283 * @qid: queue ID284 *285 * Returns 0 on success, negative on failure286 */287static int288ice_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 buffer309 * @rx_ring: Rx ring310 * @pool_present: is pool for XSK present311 *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 failure315 */316static int317ice_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 pairs341 * @vsi: Current VSI342 * @zc: is zero copy set343 *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 failure347 */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 state367 * @vsi: Current VSI368 * @pool: buffer pool to enable/associate to a ring, NULL to disable369 * @qid: queue ID370 *371 * Returns 0 on success, negative on failure372 */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 it425 * @pool: XSK Buffer pool to pull the buffers from426 * @xdp: SW ring of xdp_buff that will hold the buffers427 * @rx_desc: Pointer to Rx descriptors that will be filled428 * @count: The number of buffers to allocate429 *430 * This function allocates a number of Rx buffers from the fill ring431 * 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 descriptors436 */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 basis451 * 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 buffers464 * @rx_ring: Rx ring465 * @xsk_pool: XSK buffer pool to pick buffers to be filled by HW466 * @count: The number of buffers to allocate467 *468 * Place the @count of descriptors onto Rx ring. Handle the ring wrap469 * for case where space from next_to_use up to the end of ring is less470 * 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 buffers515 * @rx_ring: Rx ring516 * @xsk_pool: XSK buffer pool to pick buffers to be filled by HW517 * @count: The number of buffers to allocate518 *519 * Wrapper for internal allocation routine; figure out how many tail520 * bumps should take place based on the given threshold521 *522 * Returns true if all calls to internal alloc routine succeeded523 */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 buffer541 * @rx_ring: Rx ring542 * @xdp: Pointer to XDP buffer543 *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 CQ604 * @xdp_ring: XDP Tx ring605 * @xsk_pool: AF_XDP buffer pool pointer606 */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 else626 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_TX666 * @xdp: XDP buffer to xmit667 * @xdp_ring: XDP ring to produce descriptor onto668 * @xsk_pool: AF_XDP buffer pool pointer669 *670 * note that this function works directly on xdp_buff, no need to convert671 * it to xdp_frame. xdp_buff pointer is stored to ice_tx_buf so that cleaning672 * side will be able to xsk_buff_free() it.673 *674 * Returns ICE_XDP_TX for successfully produced desc, ICE_XDP_CONSUMED if there675 * was not enough space on XDP ring676 */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 path751 * @rx_ring: Rx ring752 * @xdp: xdp_buff used as input to the XDP program753 * @xdp_prog: XDP program to run754 * @xdp_ring: ring to be used for XDP_TX action755 * @xsk_pool: AF_XDP buffer pool pointer756 *757 * Returns any of ICE_XDP_{PASS, CONSUMED, TX, REDIR}758 */759static int760ice_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 else776 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 int805ice_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 ring835 * @rx_ring: AF_XDP Rx ring836 * @xsk_pool: AF_XDP buffer pool pointer837 * @budget: NAPI budget838 *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 NULL858 */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 reading880 * any other fields out of the rx_desc until we have881 * 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, so966 * use rx_ring value instead of stack based one967 */968 if (failure || ntc == rx_ring->next_to_use)969 xsk_set_rx_need_wakeup(xsk_pool);970 else971 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 descriptor981 * @xdp_ring: XDP ring to produce the HW Tx descriptor on982 * @xsk_pool: XSK buffer pool to pick buffers to be consumed by HW983 * @desc: AF_XDP descriptor to pull the DMA address and length from984 * @total_bytes: bytes accumulator that will be used for stats update985 */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 descriptors1006 * @xdp_ring: XDP ring to produce the HW Tx descriptors on1007 * @xsk_pool: XSK buffer pool to pick buffers to be consumed by HW1008 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from1009 * @total_bytes: bytes accumulator that will be used for stats update1010 */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 ring1039 * @xdp_ring: XDP ring to produce the HW Tx descriptors on1040 * @xsk_pool: XSK buffer pool to pick buffers to be consumed by HW1041 * @descs: AF_XDP descriptors to pull the DMA addresses and lengths from1042 * @nb_pkts: count of packets to be send1043 * @total_bytes: bytes accumulator that will be used for stats update1044 */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 ring1062 * @xdp_ring: XDP ring to produce the HW Tx descriptors on1063 * @xsk_pool: AF_XDP buffer pool pointer1064 *1065 * Returns true if there is no more work that needs to be done, false otherwise1066 */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_wakeup1109 * @netdev: net_device1110 * @queue_id: queue to wake up1111 * @flags: ignored in our case, since we have Rx and Tx in the same NAPI1112 *1113 * Returns negative on error, zero otherwise.1114 */1115int1116ice_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, so1139 * it will run again. If not, trigger an interrupt and1140 * schedule the NAPI from interrupt context. If NAPI would be1141 * scheduled here, the interrupt affinity would not be1142 * 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 attached1153 * @vsi: VSI to be checked1154 *1155 * Returns true if any of the Rx rings has an AF_XDP buff pool attached1156 */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 ring1171 * @rx_ring: ring to be cleaned1172 */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 queues1190 * @xdp_ring: XDP_Tx ring1191 */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}1215