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 <net/xdp_sock.h>
   6 #include <net/xdp.h>
   7
   8 #include "i40e.h"
   9 #include "i40e_txrx_common.h"
  10 #include "i40e_xsk.h"
  11
  12 /**
  13  * i40e_xsk_umem_dma_map - DMA maps all UMEM memory for the netdev
  14  * @vsi: Current VSI
  15  * @umem: UMEM to DMA map
  16  *
  17  * Returns 0 on success, <0 on failure
  18  **/
  19 static int i40e_xsk_umem_dma_map(struct i40e_vsi *vsi, struct xdp_umem *umem)
  20 {
  21         struct i40e_pf *pf = vsi->back;
  22         struct device *dev;
  23         unsigned int i, j;
  24         dma_addr_t dma;
  25
  26         dev = &pf->pdev->dev;
  27         for (i = 0; i < umem->npgs; i++) {
  28                 dma = dma_map_page_attrs(dev, umem->pgs[i], 0, PAGE_SIZE,
  29                                          DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
  30                 if (dma_mapping_error(dev, dma))
  31                         goto out_unmap;
  32
  33                 umem->pages[i].dma = dma;
  34         }
  35
  36         return 0;
  37
  38 out_unmap:
  39         for (j = 0; j < i; j++) {
  40                 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
  41                                      DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
  42                 umem->pages[i].dma = 0;
  43         }
  44
  45         return -1;
  46 }
  47
  48 /**
  49  * i40e_xsk_umem_dma_unmap - DMA unmaps all UMEM memory for the netdev
  50  * @vsi: Current VSI
  51  * @umem: UMEM to DMA map
  52  **/
  53 static void i40e_xsk_umem_dma_unmap(struct i40e_vsi *vsi, struct xdp_umem *umem)
  54 {
  55         struct i40e_pf *pf = vsi->back;
  56         struct device *dev;
  57         unsigned int i;
  58
  59         dev = &pf->pdev->dev;
  60
  61         for (i = 0; i < umem->npgs; i++) {
  62                 dma_unmap_page_attrs(dev, umem->pages[i].dma, PAGE_SIZE,
  63                                      DMA_BIDIRECTIONAL, I40E_RX_DMA_ATTR);
  64
  65                 umem->pages[i].dma = 0;
  66         }
  67 }
  68
  69 /**
  70  * i40e_xsk_umem_enable - Enable/associate a UMEM to a certain ring/qid
  71  * @vsi: Current VSI
  72  * @umem: UMEM
  73  * @qid: Rx ring to associate UMEM to
  74  *
  75  * Returns 0 on success, <0 on failure
  76  **/
  77 static int i40e_xsk_umem_enable(struct i40e_vsi *vsi, struct xdp_umem *umem,
  78                                 u16 qid)
  79 {
  80         struct net_device *netdev = vsi->netdev;
  81         struct xdp_umem_fq_reuse *reuseq;
  82         bool if_running;
  83         int err;
  84
  85         if (vsi->type != I40E_VSI_MAIN)
  86                 return -EINVAL;
  87
  88         if (qid >= vsi->num_queue_pairs)
  89                 return -EINVAL;
  90
  91         if (qid >= netdev->real_num_rx_queues ||
  92             qid >= netdev->real_num_tx_queues)
  93                 return -EINVAL;
  94
  95         reuseq = xsk_reuseq_prepare(vsi->rx_rings[0]->count);
  96         if (!reuseq)
  97                 return -ENOMEM;
  98
  99         xsk_reuseq_free(xsk_reuseq_swap(umem, reuseq));
 100
 101         err = i40e_xsk_umem_dma_map(vsi, umem);
 102         if (err)
 103                 return err;
 104
 105         set_bit(qid, vsi->af_xdp_zc_qps);
 106
 107         if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
 108
 109         if (if_running) {
 110                 err = i40e_queue_pair_disable(vsi, qid);
 111                 if (err)
 112                         return err;
 113
 114                 err = i40e_queue_pair_enable(vsi, qid);
 115                 if (err)
 116                         return err;
 117
 118                 /* Kick start the NAPI context so that receiving will start */
 119                 err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
 120                 if (err)
 121                         return err;
 122         }
 123
 124         return 0;
 125 }
 126
 127 /**
 128  * i40e_xsk_umem_disable - Disassociate a UMEM from a certain ring/qid
 129  * @vsi: Current VSI
 130  * @qid: Rx ring to associate UMEM to
 131  *
 132  * Returns 0 on success, <0 on failure
 133  **/
 134 static int i40e_xsk_umem_disable(struct i40e_vsi *vsi, u16 qid)
 135 {
 136         struct net_device *netdev = vsi->netdev;
 137         struct xdp_umem *umem;
 138         bool if_running;
 139         int err;
 140
 141         umem = xdp_get_umem_from_qid(netdev, qid);
 142         if (!umem)
 143                 return -EINVAL;
 144
 145         if_running = netif_running(vsi->netdev) && i40e_enabled_xdp_vsi(vsi);
 146
 147         if (if_running) {
 148                 err = i40e_queue_pair_disable(vsi, qid);
 149                 if (err)
 150                         return err;
 151         }
 152
 153         clear_bit(qid, vsi->af_xdp_zc_qps);
 154         i40e_xsk_umem_dma_unmap(vsi, umem);
 155
 156         if (if_running) {
 157                 err = i40e_queue_pair_enable(vsi, qid);
 158                 if (err)
 159                         return err;
 160
 161                 /* Kick start the NAPI context so that receiving will start */
 162                 err = i40e_xsk_wakeup(vsi->netdev, qid, XDP_WAKEUP_RX);
 163                 if (err)
 164                         return err;
 165         }
 166
 167         return 0;
 168 }
 169
 170 /**
 171  * i40e_xsk_umem_setup - Enable/disassociate a UMEM to/from a ring/qid
 172  * @vsi: Current VSI
 173  * @umem: UMEM to enable/associate to a ring, or NULL to disable
 174  * @qid: Rx ring to (dis)associate UMEM (from)to
 175  *
 176  * This function enables or disables a UMEM to a certain ring.
 177  *
 178  * Returns 0 on success, <0 on failure
 179  **/
 180 int i40e_xsk_umem_setup(struct i40e_vsi *vsi, struct xdp_umem *umem,
 181                         u16 qid)
 182 {
 183         return umem ? i40e_xsk_umem_enable(vsi, umem, qid) :
 184                 i40e_xsk_umem_disable(vsi, qid);
 185 }
 186
 187 /**
 188  * i40e_run_xdp_zc - Executes an XDP program on an xdp_buff
 189  * @rx_ring: Rx ring
 190  * @xdp: xdp_buff used as input to the XDP program
 191  *
 192  * This function enables or disables a UMEM to a certain ring.
 193  *
 194  * Returns any of I40E_XDP_{PASS, CONSUMED, TX, REDIR}
 195  **/
 196 static int i40e_run_xdp_zc(struct i40e_ring *rx_ring, struct xdp_buff *xdp)
 197 {
 198         struct xdp_umem *umem = rx_ring->xsk_umem;
 199         int err, result = I40E_XDP_PASS;
 200         struct i40e_ring *xdp_ring;
 201         struct bpf_prog *xdp_prog;
 202         u64 offset;
 203         u32 act;
 204
 205         rcu_read_lock();
 206         /* NB! xdp_prog will always be !NULL, due to the fact that
 207          * this path is enabled by setting an XDP program.
 208          */
 209         xdp_prog = READ_ONCE(rx_ring->xdp_prog);
 210         act = bpf_prog_run_xdp(xdp_prog, xdp);
 211         offset = xdp->data - xdp->data_hard_start;
 212
 213         xdp->handle = xsk_umem_adjust_offset(umem, xdp->handle, offset);
 214
 215         switch (act) {
 216         case XDP_PASS:
 217                 break;
 218         case XDP_TX:
 219                 xdp_ring = rx_ring->vsi->xdp_rings[rx_ring->queue_index];
 220                 result = i40e_xmit_xdp_tx_ring(xdp, xdp_ring);
 221                 break;
 222         case XDP_REDIRECT:
 223                 err = xdp_do_redirect(rx_ring->netdev, xdp, xdp_prog);
 224                 result = !err ? I40E_XDP_REDIR : I40E_XDP_CONSUMED;
 225                 break;
 226         default:
 227                 bpf_warn_invalid_xdp_action(act);
 228                 /* fall through */
 229         case XDP_ABORTED:
 230                 trace_xdp_exception(rx_ring->netdev, xdp_prog, act);
 231                 /* fallthrough -- handle aborts by dropping packet */
 232         case XDP_DROP:
 233                 result = I40E_XDP_CONSUMED;
 234                 break;
 235         }
 236         rcu_read_unlock();
 237         return result;
 238 }
 239
 240 /**
 241  * i40e_alloc_buffer_zc - Allocates an i40e_rx_buffer
 242  * @rx_ring: Rx ring
 243  * @bi: Rx buffer to populate
 244  *
 245  * This function allocates an Rx buffer. The buffer can come from fill
 246  * queue, or via the recycle queue (next_to_alloc).
 247  *
 248  * Returns true for a successful allocation, false otherwise
 249  **/
 250 static bool i40e_alloc_buffer_zc(struct i40e_ring *rx_ring,
 251                                  struct i40e_rx_buffer *bi)
 252 {
 253         struct xdp_umem *umem = rx_ring->xsk_umem;
 254         void *addr = bi->addr;
 255         u64 handle, hr;
 256
 257         if (addr) {
 258                 rx_ring->rx_stats.page_reuse_count++;
 259                 return true;
 260         }
 261
 262         if (!xsk_umem_peek_addr(umem, &handle)) {
 263                 rx_ring->rx_stats.alloc_page_failed++;
 264                 return false;
 265         }
 266
 267         hr = umem->headroom + XDP_PACKET_HEADROOM;
 268
 269         bi->dma = xdp_umem_get_dma(umem, handle);
 270         bi->dma += hr;
 271
 272         bi->addr = xdp_umem_get_data(umem, handle);
 273         bi->addr += hr;
 274
 275         bi->handle = xsk_umem_adjust_offset(umem, handle, umem->headroom);
 276
 277         xsk_umem_discard_addr(umem);
 278         return true;
 279 }
 280
 281 /**
 282  * i40e_alloc_buffer_slow_zc - Allocates an i40e_rx_buffer
 283  * @rx_ring: Rx ring
 284  * @bi: Rx buffer to populate
 285  *
 286  * This function allocates an Rx buffer. The buffer can come from fill
 287  * queue, or via the reuse queue.
 288  *
 289  * Returns true for a successful allocation, false otherwise
 290  **/
 291 static bool i40e_alloc_buffer_slow_zc(struct i40e_ring *rx_ring,
 292                                       struct i40e_rx_buffer *bi)
 293 {
 294         struct xdp_umem *umem = rx_ring->xsk_umem;
 295         u64 handle, hr;
 296
 297         if (!xsk_umem_peek_addr_rq(umem, &handle)) {
 298                 rx_ring->rx_stats.alloc_page_failed++;
 299                 return false;
 300         }
 301
 302         handle &= rx_ring->xsk_umem->chunk_mask;
 303
 304         hr = umem->headroom + XDP_PACKET_HEADROOM;
 305
 306         bi->dma = xdp_umem_get_dma(umem, handle);
 307         bi->dma += hr;
 308
 309         bi->addr = xdp_umem_get_data(umem, handle);
 310         bi->addr += hr;
 311
 312         bi->handle = xsk_umem_adjust_offset(umem, handle, umem->headroom);
 313
 314         xsk_umem_discard_addr_rq(umem);
 315         return true;
 316 }
 317
 318 static __always_inline bool
 319 __i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count,
 320                            bool alloc(struct i40e_ring *rx_ring,
 321                                       struct i40e_rx_buffer *bi))
 322 {
 323         u16 ntu = rx_ring->next_to_use;
 324         union i40e_rx_desc *rx_desc;
 325         struct i40e_rx_buffer *bi;
 326         bool ok = true;
 327
 328         rx_desc = I40E_RX_DESC(rx_ring, ntu);
 329         bi = &rx_ring->rx_bi[ntu];
 330         do {
 331                 if (!alloc(rx_ring, bi)) {
 332                         ok = false;
 333                         goto no_buffers;
 334                 }
 335
 336                 dma_sync_single_range_for_device(rx_ring->dev, bi->dma, 0,
 337                                                  rx_ring->rx_buf_len,
 338                                                  DMA_BIDIRECTIONAL);
 339
 340                 rx_desc->read.pkt_addr = cpu_to_le64(bi->dma);
 341
 342                 rx_desc++;
 343                 bi++;
 344                 ntu++;
 345
 346                 if (unlikely(ntu == rx_ring->count)) {
 347                         rx_desc = I40E_RX_DESC(rx_ring, 0);
 348                         bi = rx_ring->rx_bi;
 349                         ntu = 0;
 350                 }
 351
 352                 rx_desc->wb.qword1.status_error_len = 0;
 353                 count--;
 354         } while (count);
 355
 356 no_buffers:
 357         if (rx_ring->next_to_use != ntu)
 358                 i40e_release_rx_desc(rx_ring, ntu);
 359
 360         return ok;
 361 }
 362
 363 /**
 364  * i40e_alloc_rx_buffers_zc - Allocates a number of Rx buffers
 365  * @rx_ring: Rx ring
 366  * @count: The number of buffers to allocate
 367  *
 368  * This function allocates a number of Rx buffers from the reuse queue
 369  * or fill ring and places them on the Rx ring.
 370  *
 371  * Returns true for a successful allocation, false otherwise
 372  **/
 373 bool i40e_alloc_rx_buffers_zc(struct i40e_ring *rx_ring, u16 count)
 374 {
 375         return __i40e_alloc_rx_buffers_zc(rx_ring, count,
 376                                           i40e_alloc_buffer_slow_zc);
 377 }
 378
 379 /**
 380  * i40e_alloc_rx_buffers_fast_zc - Allocates a number of Rx buffers
 381  * @rx_ring: Rx ring
 382  * @count: The number of buffers to allocate
 383  *
 384  * This function allocates a number of Rx buffers from the fill ring
 385  * or the internal recycle mechanism and places them on the Rx ring.
 386  *
 387  * Returns true for a successful allocation, false otherwise
 388  **/
 389 static bool i40e_alloc_rx_buffers_fast_zc(struct i40e_ring *rx_ring, u16 count)
 390 {
 391         return __i40e_alloc_rx_buffers_zc(rx_ring, count,
 392                                           i40e_alloc_buffer_zc);
 393 }
 394
 395 /**
 396  * i40e_get_rx_buffer_zc - Return the current Rx buffer
 397  * @rx_ring: Rx ring
 398  * @size: The size of the rx buffer (read from descriptor)
 399  *
 400  * This function returns the current, received Rx buffer, and also
 401  * does DMA synchronization.  the Rx ring.
 402  *
 403  * Returns the received Rx buffer
 404  **/
 405 static struct i40e_rx_buffer *i40e_get_rx_buffer_zc(struct i40e_ring *rx_ring,
 406                                                     const unsigned int size)
 407 {
 408         struct i40e_rx_buffer *bi;
 409
 410         bi = &rx_ring->rx_bi[rx_ring->next_to_clean];
 411
 412         /* we are reusing so sync this buffer for CPU use */
 413         dma_sync_single_range_for_cpu(rx_ring->dev,
 414                                       bi->dma, 0,
 415                                       size,
 416                                       DMA_BIDIRECTIONAL);
 417
 418         return bi;
 419 }
 420
 421 /**
 422  * i40e_reuse_rx_buffer_zc - Recycle an Rx buffer
 423  * @rx_ring: Rx ring
 424  * @old_bi: The Rx buffer to recycle
 425  *
 426  * This function recycles a finished Rx buffer, and places it on the
 427  * recycle queue (next_to_alloc).
 428  **/
 429 static void i40e_reuse_rx_buffer_zc(struct i40e_ring *rx_ring,
 430                                     struct i40e_rx_buffer *old_bi)
 431 {
 432         struct i40e_rx_buffer *new_bi = &rx_ring->rx_bi[rx_ring->next_to_alloc];
 433         u16 nta = rx_ring->next_to_alloc;
 434
 435         /* update, and store next to alloc */
 436         nta++;
 437         rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
 438
 439         /* transfer page from old buffer to new buffer */
 440         new_bi->dma = old_bi->dma;
 441         new_bi->addr = old_bi->addr;
 442         new_bi->handle = old_bi->handle;
 443
 444         old_bi->addr = NULL;
 445 }
 446
 447 /**
 448  * i40e_zca_free - Free callback for MEM_TYPE_ZERO_COPY allocations
 449  * @alloc: Zero-copy allocator
 450  * @handle: Buffer handle
 451  **/
 452 void i40e_zca_free(struct zero_copy_allocator *alloc, unsigned long handle)
 453 {
 454         struct i40e_rx_buffer *bi;
 455         struct i40e_ring *rx_ring;
 456         u64 hr, mask;
 457         u16 nta;
 458
 459         rx_ring = container_of(alloc, struct i40e_ring, zca);
 460         hr = rx_ring->xsk_umem->headroom + XDP_PACKET_HEADROOM;
 461         mask = rx_ring->xsk_umem->chunk_mask;
 462
 463         nta = rx_ring->next_to_alloc;
 464         bi = &rx_ring->rx_bi[nta];
 465
 466         nta++;
 467         rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
 468
 469         handle &= mask;
 470
 471         bi->dma = xdp_umem_get_dma(rx_ring->xsk_umem, handle);
 472         bi->dma += hr;
 473
 474         bi->addr = xdp_umem_get_data(rx_ring->xsk_umem, handle);
 475         bi->addr += hr;
 476
 477         bi->handle = xsk_umem_adjust_offset(rx_ring->xsk_umem, (u64)handle,
 478                                             rx_ring->xsk_umem->headroom);
 479 }
 480
 481 /**
 482  * i40e_construct_skb_zc - Create skbufff from zero-copy Rx buffer
 483  * @rx_ring: Rx ring
 484  * @bi: Rx buffer
 485  * @xdp: xdp_buff
 486  *
 487  * This functions allocates a new skb from a zero-copy Rx buffer.
 488  *
 489  * Returns the skb, or NULL on failure.
 490  **/
 491 static struct sk_buff *i40e_construct_skb_zc(struct i40e_ring *rx_ring,
 492                                              struct i40e_rx_buffer *bi,
 493                                              struct xdp_buff *xdp)
 494 {
 495         unsigned int metasize = xdp->data - xdp->data_meta;
 496         unsigned int datasize = xdp->data_end - xdp->data;
 497         struct sk_buff *skb;
 498
 499         /* allocate a skb to store the frags */
 500         skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
 501                                xdp->data_end - xdp->data_hard_start,
 502                                GFP_ATOMIC | __GFP_NOWARN);
 503         if (unlikely(!skb))
 504                 return NULL;
 505
 506         skb_reserve(skb, xdp->data - xdp->data_hard_start);
 507         memcpy(__skb_put(skb, datasize), xdp->data, datasize);
 508         if (metasize)
 509                 skb_metadata_set(skb, metasize);
 510
 511         i40e_reuse_rx_buffer_zc(rx_ring, bi);
 512         return skb;
 513 }
 514
 515 /**
 516  * i40e_inc_ntc: Advance the next_to_clean index
 517  * @rx_ring: Rx ring
 518  **/
 519 static void i40e_inc_ntc(struct i40e_ring *rx_ring)
 520 {
 521         u32 ntc = rx_ring->next_to_clean + 1;
 522
 523         ntc = (ntc < rx_ring->count) ? ntc : 0;
 524         rx_ring->next_to_clean = ntc;
 525         prefetch(I40E_RX_DESC(rx_ring, ntc));
 526 }
 527
 528 /**
 529  * i40e_clean_rx_irq_zc - Consumes Rx packets from the hardware ring
 530  * @rx_ring: Rx ring
 531  * @budget: NAPI budget
 532  *
 533  * Returns amount of work completed
 534  **/
 535 int i40e_clean_rx_irq_zc(struct i40e_ring *rx_ring, int budget)
 536 {
 537         unsigned int total_rx_bytes = 0, total_rx_packets = 0;
 538         u16 cleaned_count = I40E_DESC_UNUSED(rx_ring);
 539         unsigned int xdp_res, xdp_xmit = 0;
 540         bool failure = false;
 541         struct sk_buff *skb;
 542         struct xdp_buff xdp;
 543
 544         xdp.rxq = &rx_ring->xdp_rxq;
 545
 546         while (likely(total_rx_packets < (unsigned int)budget)) {
 547                 struct i40e_rx_buffer *bi;
 548                 union i40e_rx_desc *rx_desc;
 549                 unsigned int size;
 550                 u64 qword;
 551
 552                 if (cleaned_count >= I40E_RX_BUFFER_WRITE) {
 553                         failure = failure ||
 554                                   !i40e_alloc_rx_buffers_fast_zc(rx_ring,
 555                                                                  cleaned_count);
 556                         cleaned_count = 0;
 557                 }
 558
 559                 rx_desc = I40E_RX_DESC(rx_ring, rx_ring->next_to_clean);
 560                 qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
 561
 562                 /* This memory barrier is needed to keep us from reading
 563                  * any other fields out of the rx_desc until we have
 564                  * verified the descriptor has been written back.
 565                  */
 566                 dma_rmb();
 567
 568                 bi = i40e_clean_programming_status(rx_ring, rx_desc,
 569                                                    qword);
 570                 if (unlikely(bi)) {
 571                         i40e_reuse_rx_buffer_zc(rx_ring, bi);
 572                         cleaned_count++;
 573                         continue;
 574                 }
 575
 576                 size = (qword & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
 577                        I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
 578                 if (!size)
 579                         break;
 580
 581                 bi = i40e_get_rx_buffer_zc(rx_ring, size);
 582                 xdp.data = bi->addr;
 583                 xdp.data_meta = xdp.data;
 584                 xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
 585                 xdp.data_end = xdp.data + size;
 586                 xdp.handle = bi->handle;
 587
 588                 xdp_res = i40e_run_xdp_zc(rx_ring, &xdp);
 589                 if (xdp_res) {
 590                         if (xdp_res & (I40E_XDP_TX | I40E_XDP_REDIR)) {
 591                                 xdp_xmit |= xdp_res;
 592                                 bi->addr = NULL;
 593                         } else {
 594                                 i40e_reuse_rx_buffer_zc(rx_ring, bi);
 595                         }
 596
 597                         total_rx_bytes += size;
 598                         total_rx_packets++;
 599
 600                         cleaned_count++;
 601                         i40e_inc_ntc(rx_ring);
 602                         continue;
 603                 }
 604
 605                 /* XDP_PASS path */
 606
 607                 /* NB! We are not checking for errors using
 608                  * i40e_test_staterr with
 609                  * BIT(I40E_RXD_QW1_ERROR_SHIFT). This is due to that
 610                  * SBP is *not* set in PRT_SBPVSI (default not set).
 611                  */
 612                 skb = i40e_construct_skb_zc(rx_ring, bi, &xdp);
 613                 if (!skb) {
 614                         rx_ring->rx_stats.alloc_buff_failed++;
 615                         break;
 616                 }
 617
 618                 cleaned_count++;
 619                 i40e_inc_ntc(rx_ring);
 620
 621                 if (eth_skb_pad(skb))
 622                         continue;
 623
 624                 total_rx_bytes += skb->len;
 625                 total_rx_packets++;
 626
 627                 i40e_process_skb_fields(rx_ring, rx_desc, skb);
 628                 napi_gro_receive(&rx_ring->q_vector->napi, skb);
 629         }
 630
 631         i40e_finalize_xdp_rx(rx_ring, xdp_xmit);
 632         i40e_update_rx_stats(rx_ring, total_rx_bytes, total_rx_packets);
 633
 634         if (xsk_umem_uses_need_wakeup(rx_ring->xsk_umem)) {
 635                 if (failure || rx_ring->next_to_clean == rx_ring->next_to_use)
 636                         xsk_set_rx_need_wakeup(rx_ring->xsk_umem);
 637                 else
 638                         xsk_clear_rx_need_wakeup(rx_ring->xsk_umem);
 639
 640                 return (int)total_rx_packets;
 641         }
 642         return failure ? budget : (int)total_rx_packets;
 643 }
 644
 645 /**
 646  * i40e_xmit_zc - Performs zero-copy Tx AF_XDP
 647  * @xdp_ring: XDP Tx ring
 648  * @budget: NAPI budget
 649  *
 650  * Returns true if the work is finished.
 651  **/
 652 static bool i40e_xmit_zc(struct i40e_ring *xdp_ring, unsigned int budget)
 653 {
 654         struct i40e_tx_desc *tx_desc = NULL;
 655         struct i40e_tx_buffer *tx_bi;
 656         bool work_done = true;
 657         struct xdp_desc desc;
 658         dma_addr_t dma;
 659
 660         while (budget-- > 0) {
 661                 if (!unlikely(I40E_DESC_UNUSED(xdp_ring))) {
 662                         xdp_ring->tx_stats.tx_busy++;
 663                         work_done = false;
 664                         break;
 665                 }
 666
 667                 if (!xsk_umem_consume_tx(xdp_ring->xsk_umem, &desc))
 668                         break;
 669
 670                 dma = xdp_umem_get_dma(xdp_ring->xsk_umem, desc.addr);
 671
 672                 dma_sync_single_for_device(xdp_ring->dev, dma, desc.len,
 673                                            DMA_BIDIRECTIONAL);
 674
 675                 tx_bi = &xdp_ring->tx_bi[xdp_ring->next_to_use];
 676                 tx_bi->bytecount = desc.len;
 677
 678                 tx_desc = I40E_TX_DESC(xdp_ring, xdp_ring->next_to_use);
 679                 tx_desc->buffer_addr = cpu_to_le64(dma);
 680                 tx_desc->cmd_type_offset_bsz =
 681                         build_ctob(I40E_TX_DESC_CMD_ICRC
 682                                    | I40E_TX_DESC_CMD_EOP,
 683                                    0, desc.len, 0);
 684
 685                 xdp_ring->next_to_use++;
 686                 if (xdp_ring->next_to_use == xdp_ring->count)
 687                         xdp_ring->next_to_use = 0;
 688         }
 689
 690         if (tx_desc) {
 691                 /* Request an interrupt for the last frame and bump tail ptr. */
 692                 tx_desc->cmd_type_offset_bsz |= (I40E_TX_DESC_CMD_RS <<
 693                                                  I40E_TXD_QW1_CMD_SHIFT);
 694                 i40e_xdp_ring_update_tail(xdp_ring);
 695
 696                 xsk_umem_consume_tx_done(xdp_ring->xsk_umem);
 697                 if (xsk_umem_uses_need_wakeup(xdp_ring->xsk_umem))
 698                         xsk_clear_tx_need_wakeup(xdp_ring->xsk_umem);
 699         }
 700
 701         return !!budget && work_done;
 702 }
 703
 704 /**
 705  * i40e_clean_xdp_tx_buffer - Frees and unmaps an XDP Tx entry
 706  * @tx_ring: XDP Tx ring
 707  * @tx_bi: Tx buffer info to clean
 708  **/
 709 static void i40e_clean_xdp_tx_buffer(struct i40e_ring *tx_ring,
 710                                      struct i40e_tx_buffer *tx_bi)
 711 {
 712         xdp_return_frame(tx_bi->xdpf);
 713         dma_unmap_single(tx_ring->dev,
 714                          dma_unmap_addr(tx_bi, dma),
 715                          dma_unmap_len(tx_bi, len), DMA_TO_DEVICE);
 716         dma_unmap_len_set(tx_bi, len, 0);
 717 }
 718
 719 /**
 720  * i40e_clean_xdp_tx_irq - Completes AF_XDP entries, and cleans XDP entries
 721  * @tx_ring: XDP Tx ring
 722  * @tx_bi: Tx buffer info to clean
 723  *
 724  * Returns true if cleanup/tranmission is done.
 725  **/
 726 bool i40e_clean_xdp_tx_irq(struct i40e_vsi *vsi,
 727                            struct i40e_ring *tx_ring, int napi_budget)
 728 {
 729         unsigned int ntc, total_bytes = 0, budget = vsi->work_limit;
 730         u32 i, completed_frames, frames_ready, xsk_frames = 0;
 731         struct xdp_umem *umem = tx_ring->xsk_umem;
 732         u32 head_idx = i40e_get_head(tx_ring);
 733         bool work_done = true, xmit_done;
 734         struct i40e_tx_buffer *tx_bi;
 735
 736         if (head_idx < tx_ring->next_to_clean)
 737                 head_idx += tx_ring->count;
 738         frames_ready = head_idx - tx_ring->next_to_clean;
 739
 740         if (frames_ready == 0) {
 741                 goto out_xmit;
 742         } else if (frames_ready > budget) {
 743                 completed_frames = budget;
 744                 work_done = false;
 745         } else {
 746                 completed_frames = frames_ready;
 747         }
 748
 749         ntc = tx_ring->next_to_clean;
 750
 751         for (i = 0; i < completed_frames; i++) {
 752                 tx_bi = &tx_ring->tx_bi[ntc];
 753
 754                 if (tx_bi->xdpf)
 755                         i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
 756                 else
 757                         xsk_frames++;
 758
 759                 tx_bi->xdpf = NULL;
 760                 total_bytes += tx_bi->bytecount;
 761
 762                 if (++ntc >= tx_ring->count)
 763                         ntc = 0;
 764         }
 765
 766         tx_ring->next_to_clean += completed_frames;
 767         if (unlikely(tx_ring->next_to_clean >= tx_ring->count))
 768                 tx_ring->next_to_clean -= tx_ring->count;
 769
 770         if (xsk_frames)
 771                 xsk_umem_complete_tx(umem, xsk_frames);
 772
 773         i40e_arm_wb(tx_ring, vsi, budget);
 774         i40e_update_tx_stats(tx_ring, completed_frames, total_bytes);
 775
 776 out_xmit:
 777         if (xsk_umem_uses_need_wakeup(tx_ring->xsk_umem)) {
 778                 if (tx_ring->next_to_clean == tx_ring->next_to_use)
 779                         xsk_set_tx_need_wakeup(tx_ring->xsk_umem);
 780                 else
 781                         xsk_clear_tx_need_wakeup(tx_ring->xsk_umem);
 782         }
 783
 784         xmit_done = i40e_xmit_zc(tx_ring, budget);
 785
 786         return work_done && xmit_done;
 787 }
 788
 789 /**
 790  * i40e_xsk_wakeup - Implements the ndo_xsk_wakeup
 791  * @dev: the netdevice
 792  * @queue_id: queue id to wake up
 793  * @flags: ignored in our case since we have Rx and Tx in the same NAPI.
 794  *
 795  * Returns <0 for errors, 0 otherwise.
 796  **/
 797 int i40e_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 798 {
 799         struct i40e_netdev_priv *np = netdev_priv(dev);
 800         struct i40e_vsi *vsi = np->vsi;
 801         struct i40e_ring *ring;
 802
 803         if (test_bit(__I40E_VSI_DOWN, vsi->state))
 804                 return -ENETDOWN;
 805
 806         if (!i40e_enabled_xdp_vsi(vsi))
 807                 return -ENXIO;
 808
 809         if (queue_id >= vsi->num_queue_pairs)
 810                 return -ENXIO;
 811
 812         if (!vsi->xdp_rings[queue_id]->xsk_umem)
 813                 return -ENXIO;
 814
 815         ring = vsi->xdp_rings[queue_id];
 816
 817         /* The idea here is that if NAPI is running, mark a miss, so
 818          * it will run again. If not, trigger an interrupt and
 819          * schedule the NAPI from interrupt context. If NAPI would be
 820          * scheduled here, the interrupt affinity would not be
 821          * honored.
 822          */
 823         if (!napi_if_scheduled_mark_missed(&ring->q_vector->napi))
 824                 i40e_force_wb(vsi, ring->q_vector);
 825
 826         return 0;
 827 }
 828
 829 void i40e_xsk_clean_rx_ring(struct i40e_ring *rx_ring)
 830 {
 831         u16 i;
 832
 833         for (i = 0; i < rx_ring->count; i++) {
 834                 struct i40e_rx_buffer *rx_bi = &rx_ring->rx_bi[i];
 835
 836                 if (!rx_bi->addr)
 837                         continue;
 838
 839                 xsk_umem_fq_reuse(rx_ring->xsk_umem, rx_bi->handle);
 840                 rx_bi->addr = NULL;
 841         }
 842 }
 843
 844 /**
 845  * i40e_xsk_clean_xdp_ring - Clean the XDP Tx ring on shutdown
 846  * @xdp_ring: XDP Tx ring
 847  **/
 848 void i40e_xsk_clean_tx_ring(struct i40e_ring *tx_ring)
 849 {
 850         u16 ntc = tx_ring->next_to_clean, ntu = tx_ring->next_to_use;
 851         struct xdp_umem *umem = tx_ring->xsk_umem;
 852         struct i40e_tx_buffer *tx_bi;
 853         u32 xsk_frames = 0;
 854
 855         while (ntc != ntu) {
 856                 tx_bi = &tx_ring->tx_bi[ntc];
 857
 858                 if (tx_bi->xdpf)
 859                         i40e_clean_xdp_tx_buffer(tx_ring, tx_bi);
 860                 else
 861                         xsk_frames++;
 862
 863                 tx_bi->xdpf = NULL;
 864
 865                 ntc++;
 866                 if (ntc >= tx_ring->count)
 867                         ntc = 0;
 868         }
 869
 870         if (xsk_frames)
 871                 xsk_umem_complete_tx(umem, xsk_frames);
 872 }
 873
 874 /**
 875  * i40e_xsk_any_rx_ring_enabled - Checks if Rx rings have AF_XDP UMEM attached
 876  * @vsi: vsi
 877  *
 878  * Returns true if any of the Rx rings has an AF_XDP UMEM attached
 879  **/
 880 bool i40e_xsk_any_rx_ring_enabled(struct i40e_vsi *vsi)
 881 {
 882         struct net_device *netdev = vsi->netdev;
 883         int i;
 884
 885         for (i = 0; i < vsi->num_queue_pairs; i++) {
 886                 if (xdp_get_umem_from_qid(netdev, i))
 887                         return true;
 888         }
 889
 890         return false;
 891 }