drivers/net/ethernet/intel/i40e/i40e_xsk.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright(c) 2018 Intel Corporation. */
   3
   4 #include <linux/bpf_trace.h>
   5 #include <linux/stringify.h>
   6 #include <net/xdp_sock_drv.h>
   7 #include <net/xdp.h>
   8
   9 #include "i40e.h"
  10 #include "i40e_txrx_common.h"
  11 #include "i40e_xsk.h"
  12
  13 void i40e_clear_rx_bi_zc(struct i40e_ring *rx_ring)
  14 {
  15         memset(rx_ring->rx_bi_zc, 0,
  16                sizeof(*rx_ring->rx_bi_zc) * rx_ring->count);
  17 }
  18
  19 static struct xdp_buff **i40e_rx_bi(struct i40e_ring *rx_ring, u32 idx)
  20 {
  21         return &rx_ring->rx_bi_zc[idx];
  22 }
  23
  24 /**
  25  * i40e_realloc_rx_xdp_bi - reallocate SW ring for either XSK or normal buffer
  26  * @rx_ring: Current rx ring
  27  * @pool_present: is pool for XSK present
  28  *
  29  * Try allocating memory and return ENOMEM, if failed to allocate.
  30  * If allocation was successful, substitute buffer with allocated one.
  31  * Returns 0 on success, negative on failure
  32  */
  33 static int i40e_realloc_rx_xdp_bi(struct i40e_ring *rx_ring, bool pool_present)
  34 {
  35         size_t elem_size = pool_present ? sizeof(*rx_ring->rx_bi_zc) :
  36                                           sizeof(*rx_ring->rx_bi);
  37         void *sw_ring = kcalloc(rx_ring->count, elem_size, GFP_KERNEL);
  38
  39         if (!sw_ring)
  40                 return -ENOMEM;
  41
  42         if (pool_present) {
  43                 kfree(rx_ring->rx_bi);
  44                 rx_ring->rx_bi = NULL;
  45                 rx_ring->rx_bi_zc = sw_ring;
  46         } else {
  47                 kfree(rx_ring->rx_bi_zc);
  48                 rx_ring->rx_bi_zc = NULL;
  49                 rx_ring->rx_bi = sw_ring;
  50         }
  51         return 0;
  52 }
  53
  54 /**
  55  * i40e_realloc_rx_bi_zc - reallocate rx SW rings
  56  * @vsi: Current VSI
  57  * @zc: is zero copy set
  58  *
  59  * Reallocate buffer for rx_rings that might be used by XSK.
  60  * XDP requires more memory, than rx_buf provides.
  61  * Returns 0 on success, negative on failure
  62  */
  63 int i40e_realloc_rx_bi_zc(struct i40e_vsi *vsi, bool zc)
  64 {
  65         struct i40e_ring *rx_ring;
  66         unsigned long q;
  67
  68         for_each_set_bit(q, vsi->af_xdp_zc_qps, vsi->alloc_queue_pairs) {
  69                 rx_ring = vsi->rx_rings[q];
  70                 if (i40e_realloc_rx_xdp_bi(rx_ring, zc))
  71                         return -ENOMEM;
  72         }
  73         return 0;
  74 }
  75
  76 /**
  77  * i40e_xsk_pool_enable - Enable/associate an AF_XDP buffer pool to a
  78  * certain ring/qid
  79  * @vsi: Current VSI
  80  * @pool: buffer pool
  81  * @qid: Rx ring to associate buffer pool with
  82  *
  83  * Returns 0 on success, <0 on failure
  84  **/
  85 static int i40e_xsk_pool_enable(struct i40e_vsi *vsi,
  86                                 struct xsk_buff_pool *pool,
  87                                 u16 qid)
  88 {
  89         struct net_device *netdev = vsi->netdev;
  90         bool if_running;
  91         int err;
  92
  93         if (vsi->type != I40E_VSI_MAIN)
  94                 return -EINVAL;
  95
  96         if (qid >= vsi->num_queue_pairs)
  97                 return -EINVAL;
  98
  99         if (qid >= netdev->real_num_rx_queues ||
 100             qid >= netdev->real_num_tx_queues)
 101                 return -EINVAL;
 102
 103         err = xsk_pool_dma_map(pool, &vsi->back->pdev->dev, I40E_RX_DMA_ATTR);
 104         if (err)
 105                 return err;
 106
 107         set_bit(qid, vsi->af_xdp_zc_qps);
 108
 109         if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
 110
 111         if (if_running) {
 112                 err = i40e_queue_pair_disable(vsi, qid);
 113                 if (err)
 114                         return err;
 115
 116                 err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], true);
 117                 if (err)
 118                         return err;
 119
 120                 err = i40e_queue_pair_enable(vsi, qid);
 121                 if (err)
 122                         return err;
 123
 124                 /* Kick start the NAPI context so that receiving will start */
 125                 err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
 126                 if (err)
 127                         return err;
 128         }
 129
 130         return 0;
 131 }
 132
 133 /**
 134  * i40e_xsk_pool_disable - Disassociate an AF_XDP buffer pool from a
 135  * certain ring/qid
 136  * @vsi: Current VSI
 137  * @qid: Rx ring to associate buffer pool with
 138  *
 139  * Returns 0 on success, <0 on failure
 140  **/
 141 static int i40e_xsk_pool_disable(struct i40e_vsi *vsi, u16 qid)
 142 {
 143         struct net_device *netdev = vsi->netdev;
 144         struct xsk_buff_pool *pool;
 145         bool if_running;
 146         int err;
 147
 148         pool = xsk_get_pool_from_qid(netdev, qid);
 149         if (!pool)
 150                 return -EINVAL;
 151
 152         if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
 153
 154         if (if_running) {
 155                 err = i40e_queue_pair_disable(vsi, qid);
 156                 if (err)
 157                         return err;
 158         }
 159
 160         clear_bit(qid, vsi->af_xdp_zc_qps);
 161         xsk_pool_dma_unmap(pool, I40E_RX_DMA_ATTR);
 162
 163         if (if_running) {
 164                 err = i40e_realloc_rx_xdp_bi(vsi->rx_rings[qid], false);
 165                 if (err)
 166                         return err;
 167                 err = i40e_queue_pair_enable(vsi, qid);
 168                 if (err)
 169                         return err;
 170         }
 171
 172         return 0;
 173 }
 174
 175 /**
 176  * i40e_xsk_pool_setup - Enable/disassociate an AF_XDP buffer pool to/from
 177  * a ring/qid
 178  * @vsi: Current VSI
 179  * @pool: Buffer pool to enable/associate to a ring, or NULL to disable
 180  * @qid: Rx ring to (dis)associate buffer pool (from)to
 181  *
 182  * This function enables or disables a buffer pool to a certain ring.
 183  *
 184  * Returns 0 on success, <0 on failure
 185  **/
 186 int i40e_xsk_pool_setup(struct i40e_vsi *vsi, struct xsk_buff_pool *pool,
 187                         u16 qid)
 188 {
 189         return pool ? i40e_xsk_pool_enable(vsi, pool, qid) :
 190                 i40e_xsk_pool_disable(vsi, qid);
 191 }
 192
 193 /**
 194  * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
 195  * @rx_ring: Rx ring
 196  * @xdp: xdp_buff used as input to the XDP program
 197  * @xdp_prog: XDP program to run
 198  *
 199  * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
 200  **/
 201 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp,
 202                            struct bpf_prog *xdp_prog)
 203 {
 204         int err, result = I40E_XDP_PASS;
 205         struct i40e_ring *xdp_ring;
 206         u32 act;
 207
 208         act = bpf_prog_run_xdp(xdp_prog, xdp);
 209
 210         if (likely(act == XDP_REDIRECT)) {
 211                 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
 212                 if (!err)
 213                         return I40E_XDP_REDIR;
 214                 if (xsk_uses_need_wakeup(rx_ring->xsk_pool) && err == -ENOBUFS)
 215                         result = I40E_XDP_EXIT;
 216                 else
 217                         result = I40E_XDP_CONSUMED;
 218                 goto out_failure;
 219         }
 220
 221         switch (act) {
 222         case XDP_PASS:
 223                 break;
 224         case XDP_TX:
 225                 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
 226                 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
 227                 if (result == I40E_XDP_CONSUMED)
 228                         goto out_failure;
 229                 break;
 230         case XDP_DROP:
 231                 result = I40E_XDP_CONSUMED;
 232                 break;
 233         default:
 234                 bpf_warn_invalid_xdp_action(rx_ring->netdev, xdp_prog, act);
 235                 fallthrough;
 236         case XDP_ABORTED:
 237                 result = I40E_XDP_CONSUMED;
 238 out_failure:
 239                 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
 240         }
 241         return result;
 242 }
 243
 244 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
 245 {
 246         u16 ntu = rx_ring->next_to_use;
 247         union i40e_rx_desc *rx_desc;
 248         struct xdp_buff **xdp;
 249         u32 nb_buffs, i;
 250         dma_addr_t dma;
 251
 252         rx_desc = I40E_RX_DESC(rx_ring, ntu);
 253         xdp = i40e_rx_bi(rx_ring, ntu);
 254
 255         nb_buffs = min_t(u16, count, rx_ring->count - ntu);
 256         nb_buffs = xsk_buff_alloc_batch(rx_ring->xsk_pool, xdp, nb_buffs);
 257         if (!nb_buffs)
 258                 return false;
 259
 260         i = nb_buffs;
 261         while (i--) {
 262                 dma = xsk_buff_xdp_get_dma(*xdp);
 263                 rx_desc->read.pkt_addr = cpu_to_le64(dma);
 264                 rx_desc->read.hdr_addr = 0;
 265
 266                 rx_desc++;
 267                 xdp++;
 268         }
 269
 270         ntu += nb_buffs;
 271         if (ntu == rx_ring->count) {
 272                 rx_desc = I40E_RX_DESC(rx_ring, 0);
 273                 ntu = 0;
 274         }
 275
 276         /* clear the status bits for the next_to_use descriptor */
 277         rx_desc->wb.qword1.status_error_len = 0;
 278         i40e_release_rx_desc(rx_ring, ntu);
 279
 280         return count == nb_buffs;
 281 }
 282
 283 /**
 284  * i40e_construct_skb_zc - Create skbuff from zero-copy Rx buffer
 285  * @rx_ring: Rx ring
 286  * @xdp: xdp_buff
 287  *
 288  * This functions allocates a new skb from a zero-copy Rx buffer.
 289  *
 290  * Returns the skb, or NULL on failure.
 291  **/
 292 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
 293                                              struct xdp_buff *xdp)
 294 {
 295         unsigned int totalsize = xdp->data_end - xdp->data_meta;
 296         unsigned int metasize = xdp->data - xdp->data_meta;
 297         struct skb_shared_info *sinfo = NULL;
 298         struct sk_buff *skb;
 299         u32 nr_frags = 0;
 300
 301         if (unlikely(xdp_buff_has_frags(xdp))) {
 302                 sinfo = xdp_get_shared_info_from_buff(xdp);
 303                 nr_frags = sinfo->nr_frags;
 304         }
 305         net_prefetch(xdp->data_meta);
 306
 307         /* allocate a skb to store the frags */
 308         skb = __napi_alloc_skb(&rx_ring->q_vector->napi, totalsize,
 309                                GFP_ATOMIC | __GFP_NOWARN);
 310         if (unlikely(!skb))
 311                 goto out;
 312
 313         memcpy(__skb_put(skb, totalsize), xdp->data_meta,
 314                ALIGN(totalsize, sizeof(long)));
 315
 316         if (metasize) {
 317                 skb_metadata_set(skb, metasize);
 318                 __skb_pull(skb, metasize);
 319         }
 320
 321         if (likely(!xdp_buff_has_frags(xdp)))
 322                 goto out;
 323
 324         for (int i = 0; i < nr_frags; i++) {
 325                 struct skb_shared_info *skinfo = skb_shinfo(skb);
 326                 skb_frag_t *frag = &sinfo->frags[i];
 327                 struct page *page;
 328                 void *addr;
 329
 330                 page = dev_alloc_page();
 331                 if (!page) {
 332                         dev_kfree_skb(skb);
 333                         return NULL;
 334                 }
 335                 addr = page_to_virt(page);
 336
 337                 memcpy(addr, skb_frag_page(frag), skb_frag_size(frag));
 338
 339                 __skb_fill_page_desc_noacc(skinfo, skinfo->nr_frags++,
 340                                            addr, 0, skb_frag_size(frag));
 341         }
 342
 343 out:
 344         xsk_buff_free(xdp);
 345         return skb;
 346 }
 347
 348 static void i40e_handle_xdp_result_zc(struct i40e_ring *rx_ring,
 349                                       struct xdp_buff *xdp_buff,
 350                                       union i40e_rx_desc *rx_desc,
 351                                       unsigned int *rx_packets,
 352                                       unsigned int *rx_bytes,
 353                                       unsigned int xdp_res,
 354                                       bool *failure)
 355 {
 356         struct sk_buff *skb;
 357
 358         *rx_packets = 1;
 359         *rx_bytes = xdp_get_buff_len(xdp_buff);
 360
 361         if (likely(xdp_res == I40E_XDP_REDIR) || xdp_res == I40E_XDP_TX)
 362                 return;
 363
 364         if (xdp_res == I40E_XDP_EXIT) {
 365                 *failure = true;
 366                 return;
 367         }
 368
 369         if (xdp_res == I40E_XDP_CONSUMED) {
 370                 xsk_buff_free(xdp_buff);
 371                 return;
 372         }
 373         if (xdp_res == I40E_XDP_PASS) {
 374                 /* NB! We are not checking for errors using
 375                  * i40e_test_staterr with
 376                  * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
 377                  * SBP is *not* set in PRT_SBPVSI (default not set).
 378                  */
 379                 skb = i40e_construct_skb_zc(rx_ring, xdp_buff);
 380                 if (!skb) {
 381                         rx_ring->rx_stats.alloc_buff_failed++;
 382                         *rx_packets = 0;
 383                         *rx_bytes = 0;
 384                         return;
 385                 }
 386
 387                 if (eth_skb_pad(skb)) {
 388                         *rx_packets = 0;
 389                         *rx_bytes = 0;
 390                         return;
 391                 }
 392
 393                 i40e_process_skb_fields(rx_ring, rx_desc, skb);
 394                 napi_gro_receive(&rx_ring->q_vector->napi, skb);
 395                 return;
 396         }
 397
 398         /* Should never get here, as all valid cases have been handled already.
 399          */
 400         WARN_ON_ONCE(1);
 401 }
 402
 403 static int
 404 i40e_add_xsk_frag(struct i40e_ring *rx_ring, struct xdp_buff *first,
 405                   struct xdp_buff *xdp, const unsigned int size)
 406 {
 407         struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(first);
 408
 409         if (!xdp_buff_has_frags(first)) {
 410                 sinfo->nr_frags = 0;
 411                 sinfo->xdp_frags_size = 0;
 412                 xdp_buff_set_frags_flag(first);
 413         }
 414
 415         if (unlikely(sinfo->nr_frags == MAX_SKB_FRAGS)) {
 416                 xsk_buff_free(first);
 417                 return -ENOMEM;
 418         }
 419
 420         __skb_fill_page_desc_noacc(sinfo, sinfo->nr_frags++,
 421                                    virt_to_page(xdp->data_hard_start), 0, size);
 422         sinfo->xdp_frags_size += size;
 423         xsk_buff_add_frag(xdp);
 424
 425         return 0;
 426 }
 427
 428 /**
 429  * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
 430  * @rx_ring: Rx ring
 431  * @budget: NAPI budget
 432  *
 433  * Returns amount of work completed
 434  **/
 435 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
 436 {
 437         unsigned int total_rx_bytes = 0, total_rx_packets = 0;
 438         u16 next_to_process = rx_ring->next_to_process;
 439         u16 next_to_clean = rx_ring->next_to_clean;
 440         u16 count_mask = rx_ring->count - 1;
 441         unsigned int xdp_res, xdp_xmit = 0;
 442         struct xdp_buff *first = NULL;
 443         struct bpf_prog *xdp_prog;
 444         bool failure = false;
 445         u16 cleaned_count;
 446
 447         if (next_to_process != next_to_clean)
 448                 first = *i40e_rx_bi(rx_ring, next_to_clean);
 449
 450         /* NB! xdp_prog will always be !NULL, due to the fact that
 451          * this path is enabled by setting an XDP program.
 452          */
 453         xdp_prog = READ_ONCE(rx_ring->xdp_prog);
 454
 455         while (likely(total_rx_packets < (unsigned int)budget)) {
 456                 union i40e_rx_desc *rx_desc;
 457                 unsigned int rx_packets;
 458                 unsigned int rx_bytes;
 459                 struct xdp_buff *bi;
 460                 unsigned int size;
 461                 u64 qword;
 462
 463                 rx_desc = I40E_RX_DESC(rx_ring, next_to_process);
 464                 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
 465
 466                 /* This memory barrier is needed to keep us from reading
 467                  * any other fields out of the rx_desc until we have
 468                  * verified the descriptor has been written back.
 469                  */
 470                 dma_rmb();
 471
 472                 if (i40e_rx_is_programming_status(qword)) {
 473                         i40e_clean_programming_status(rx_ring,
 474                                                       rx_desc->raw.qword[0],
 475                                                       qword);
 476                         bi = *i40e_rx_bi(rx_ring, next_to_process);
 477                         xsk_buff_free(bi);
 478                         next_to_process = (next_to_process + 1) & count_mask;
 479                         continue;
 480                 }
 481
 482                 size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 483                        I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
 484                 if (!size)
 485                         break;
 486
 487                 bi = *i40e_rx_bi(rx_ring, next_to_process);
 488                 xsk_buff_set_size(bi, size);
 489                 xsk_buff_dma_sync_for_cpu(bi, rx_ring->xsk_pool);
 490
 491                 if (!first)
 492                         first = bi;
 493                 else if (i40e_add_xsk_frag(rx_ring, first, bi, size))
 494                         break;
 495
 496                 next_to_process = (next_to_process + 1) & count_mask;
 497
 498                 if (i40e_is_non_eop(rx_ring, rx_desc))
 499                         continue;
 500
 501                 xdp_res = i40e_run_xdp_zc(rx_ring, first, xdp_prog);
 502                 i40e_handle_xdp_result_zc(rx_ring, first, rx_desc, &rx_packets,
 503                                           &rx_bytes, xdp_res, &failure);
 504                 first->flags = 0;
 505                 next_to_clean = next_to_process;
 506                 if (failure)
 507                         break;
 508                 total_rx_packets += rx_packets;
 509                 total_rx_bytes += rx_bytes;
 510                 xdp_xmit |= xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR);
 511                 first = NULL;
 512         }
 513
 514         rx_ring->next_to_clean = next_to_clean;
 515         rx_ring->next_to_process = next_to_process;
 516         cleaned_count = (next_to_clean - rx_ring->next_to_use - 1) & count_mask;
 517
 518         if (cleaned_count >= I40E_RX_BUFFER_WRITE)
 519                 failure |= !i40e_alloc_rx_buffers_zc(rx_ring, cleaned_count);
 520
 521         i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
 522         i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
 523
 524         if (xsk_uses_need_wakeup(rx_ring->xsk_pool)) {
 525                 if (failure || next_to_clean == rx_ring->next_to_use)
 526                         xsk_set_rx_need_wakeup(rx_ring->xsk_pool);
 527                 else
 528                         xsk_clear_rx_need_wakeup(rx_ring->xsk_pool);
 529
 530                 return (int)total_rx_packets;
 531         }
 532         return failure ? budget : (int)total_rx_packets;
 533 }
 534
 535 static void i40e_xmit_pkt(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
 536                           unsigned int *total_bytes)
 537 {
 538         u32 cmd = I40E_TX_DESC_CMD_ICRC | xsk_is_eop_desc(desc);
 539         struct i40e_tx_desc *tx_desc;
 540         dma_addr_t dma;
 541
 542         dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc->addr);
 543         xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc->len);
 544
 545         tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use++);
 546         tx_desc->buffer_addr = cpu_to_le64(dma);
 547         tx_desc->cmd_type_offset_bsz = build_ctob(cmd, 0, desc->len, 0);
 548
 549         *total_bytes += desc->len;
 550 }
 551
 552 static void i40e_xmit_pkt_batch(struct i40e_ring *xdp_ring, struct xdp_desc *desc,
 553                                 unsigned int *total_bytes)
 554 {
 555         u16 ntu = xdp_ring->next_to_use;
 556         struct i40e_tx_desc *tx_desc;
 557         dma_addr_t dma;
 558         u32 i;
 559
 560         loop_unrolled_for(i = 0; i < PKTS_PER_BATCH; i++) {
 561                 u32 cmd = I40E_TX_DESC_CMD_ICRC | xsk_is_eop_desc(&desc[i]);
 562
 563                 dma = xsk_buff_raw_get_dma(xdp_ring->xsk_pool, desc[i].addr);
 564                 xsk_buff_raw_dma_sync_for_device(xdp_ring->xsk_pool, dma, desc[i].len);
 565
 566                 tx_desc = I40E_TX_DESC(xdp_ring, ntu++);
 567                 tx_desc->buffer_addr = cpu_to_le64(dma);
 568                 tx_desc->cmd_type_offset_bsz = build_ctob(cmd, 0, desc[i].len, 0);
 569
 570                 *total_bytes += desc[i].len;
 571         }
 572
 573         xdp_ring->next_to_use = ntu;
 574 }
 575
 576 static void i40e_fill_tx_hw_ring(struct i40e_ring *xdp_ring, struct xdp_desc *descs, u32 nb_pkts,
 577                                  unsigned int *total_bytes)
 578 {
 579         u32 batched, leftover, i;
 580
 581         batched = nb_pkts & ~(PKTS_PER_BATCH - 1);
 582         leftover = nb_pkts & (PKTS_PER_BATCH - 1);
 583         for (i = 0; i < batched; i += PKTS_PER_BATCH)
 584                 i40e_xmit_pkt_batch(xdp_ring, &descs[i], total_bytes);
 585         for (i = batched; i < batched + leftover; i++)
 586                 i40e_xmit_pkt(xdp_ring, &descs[i], total_bytes);
 587 }
 588
 589 static void i40e_set_rs_bit(struct i40e_ring *xdp_ring)
 590 {
 591         u16 ntu = xdp_ring->next_to_use ? xdp_ring->next_to_use - 1 : xdp_ring->count - 1;
 592         struct i40e_tx_desc *tx_desc;
 593
 594         tx_desc = I40E_TX_DESC(xdp_ring, ntu);
 595         tx_desc->cmd_type_offset_bsz |= cpu_to_le64(I40E_TX_DESC_CMD_RS << I40E_TXD_QW1_CMD_SHIFT);
 596 }
 597
 598 /**
 599  * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
 600  * @xdp_ring: XDP Tx ring
 601  * @budget: NAPI budget
 602  *
 603  * Returns true if the work is finished.
 604  **/
 605 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
 606 {
 607         struct xdp_desc *descs = xdp_ring->xsk_pool->tx_descs;
 608         u32 nb_pkts, nb_processed = 0;
 609         unsigned int total_bytes = 0;
 610
 611         nb_pkts = xsk_tx_peek_release_desc_batch(xdp_ring->xsk_pool, budget);
 612         if (!nb_pkts)
 613                 return true;
 614
 615         if (xdp_ring->next_to_use + nb_pkts >= xdp_ring->count) {
 616                 nb_processed = xdp_ring->count - xdp_ring->next_to_use;
 617                 i40e_fill_tx_hw_ring(xdp_ring, descs, nb_processed, &total_bytes);
 618                 xdp_ring->next_to_use = 0;
 619         }
 620
 621         i40e_fill_tx_hw_ring(xdp_ring, &descs[nb_processed], nb_pkts - nb_processed,
 622                              &total_bytes);
 623
 624         /* Request an interrupt for the last frame and bump tail ptr. */
 625         i40e_set_rs_bit(xdp_ring);
 626         i40e_xdp_ring_update_tail(xdp_ring);
 627
 628         i40e_update_tx_stats(xdp_ring, nb_pkts, total_bytes);
 629
 630         return nb_pkts < budget;
 631 }
 632
 633 /**
 634  * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
 635  * @tx_ring: XDP Tx ring
 636  * @tx_bi: Tx buffer info to clean
 637  **/
 638 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
 639                                      struct i40e_tx_buffer *tx_bi)
 640 {
 641         xdp_return_frame(tx_bi->xdpf);
 642         tx_ring->xdp_tx_active--;
 643         dma_unmap_single(tx_ring->dev,
 644                          dma_unmap_addr(tx_bi, dma),
 645                          dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
 646         dma_unmap_len_set(tx_bi, len, 0);
 647 }
 648
 649 /**
 650  * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
 651  * @vsi: Current VSI
 652  * @tx_ring: XDP Tx ring
 653  *
 654  * Returns true if cleanup/transmission is done.
 655  **/
 656 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi, struct i40e_ring *tx_ring)
 657 {
 658         struct xsk_buff_pool *bp = tx_ring->xsk_pool;
 659         u32 i, completed_frames, xsk_frames = 0;
 660         u32 head_idx = i40e_get_head(tx_ring);
 661         struct i40e_tx_buffer *tx_bi;
 662         unsigned int ntc;
 663
 664         if (head_idx < tx_ring->next_to_clean)
 665                 head_idx += tx_ring->count;
 666         completed_frames = head_idx - tx_ring->next_to_clean;
 667
 668         if (completed_frames == 0)
 669                 goto out_xmit;
 670
 671         if (likely(!tx_ring->xdp_tx_active)) {
 672                 xsk_frames = completed_frames;
 673                 goto skip;
 674         }
 675
 676         ntc = tx_ring->next_to_clean;
 677
 678         for (i = 0; i < completed_frames; i++) {
 679                 tx_bi = &tx_ring->tx_bi[ntc];
 680
 681                 if (tx_bi->xdpf) {
 682                         i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
 683                         tx_bi->xdpf = NULL;
 684                 } else {
 685                         xsk_frames++;
 686                 }
 687
 688                 if (++ntc >= tx_ring->count)
 689                         ntc = 0;
 690         }
 691
 692 skip:
 693         tx_ring->next_to_clean += completed_frames;
 694         if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
 695                 tx_ring->next_to_clean -= tx_ring->count;
 696
 697         if (xsk_frames)
 698                 xsk_tx_completed(bp, xsk_frames);
 699
 700         i40e_arm_wb(tx_ring, vsi, completed_frames);
 701
 702 out_xmit:
 703         if (xsk_uses_need_wakeup(tx_ring->xsk_pool))
 704                 xsk_set_tx_need_wakeup(tx_ring->xsk_pool);
 705
 706         return i40e_xmit_zc(tx_ring, I40E_DESC_UNUSED(tx_ring));
 707 }
 708
 709 /**
 710  * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup
 711  * @dev: the netdevice
 712  * @queue_id: queue id to wake up
 713  * @flags: ignored in our case since we have Rx and Tx in the same NAPI.
 714  *
 715  * Returns <0 for errors, 0 otherwise.
 716  **/
 717 int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 718 {
 719         struct i40e_netdev_priv *np = netdev_priv(dev);
 720         struct i40e_vsi *vsi = np->vsi;
 721         struct i40e_pf *pf = vsi->back;
 722         struct i40e_ring *ring;
 723
 724         if (test_bit(__I40E_CONFIG_BUSY, pf->state))
 725                 return -EAGAIN;
 726
 727         if (test_bit(__I40E_VSI_DOWN, vsi->state))
 728                 return -ENETDOWN;
 729
 730         if (!i40e_enabled_xdp_vsi(vsi))
 731                 return -EINVAL;
 732
 733         if (queue_id >= vsi->num_queue_pairs)
 734                 return -EINVAL;
 735
 736         if (!vsi->xdp_rings[queue_id]->xsk_pool)
 737                 return -EINVAL;
 738
 739         ring = vsi->xdp_rings[queue_id];
 740
 741         /* The idea here is that if NAPI is running, mark a miss, so
 742          * it will run again. If not, trigger an interrupt and
 743          * schedule the NAPI from interrupt context. If NAPI would be
 744          * scheduled here, the interrupt affinity would not be
 745          * honored.
 746          */
 747         if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
 748                 i40e_force_wb(vsi, ring->q_vector);
 749
 750         return 0;
 751 }
 752
 753 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
 754 {
 755         u16 count_mask = rx_ring->count - 1;
 756         u16 ntc = rx_ring->next_to_clean;
 757         u16 ntu = rx_ring->next_to_use;
 758
 759         for ( ; ntc != ntu; ntc = (ntc + 1)  & count_mask) {
 760                 struct xdp_buff *rx_bi = *i40e_rx_bi(rx_ring, ntc);
 761
 762                 xsk_buff_free(rx_bi);
 763         }
 764 }
 765
 766 /**
 767  * i40e_xsk_clean_tx_ring - Clean the XDP Tx ring on shutdown
 768  * @tx_ring: XDP Tx ring
 769  **/
 770 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
 771 {
 772         u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
 773         struct xsk_buff_pool *bp = tx_ring->xsk_pool;
 774         struct i40e_tx_buffer *tx_bi;
 775         u32 xsk_frames = 0;
 776
 777         while (ntc != ntu) {
 778                 tx_bi = &tx_ring->tx_bi[ntc];
 779
 780                 if (tx_bi->xdpf)
 781                         i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
 782                 else
 783                         xsk_frames++;
 784
 785                 tx_bi->xdpf = NULL;
 786
 787                 ntc++;
 788                 if (ntc >= tx_ring->count)
 789                         ntc = 0;
 790         }
 791
 792         if (xsk_frames)
 793                 xsk_tx_completed(bp, xsk_frames);
 794 }
 795
 796 /**
 797  * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have an AF_XDP
 798  * buffer pool attached
 799  * @vsi: vsi
 800  *
 801  * Returns true if any of the Rx rings has an AF_XDP buffer pool attached
 802  **/
 803 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
 804 {
 805         struct net_device *netdev = vsi->netdev;
 806         int i;
 807
 808         for (i = 0; i < vsi->num_queue_pairs; i++) {
 809                 if (xsk_get_pool_from_qid(netdev, i))
 810                         return true;
 811         }
 812
 813         return false;
 814 }