drivers/staging/octeon/ethernet-rx.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * This file is based on code from OCTEON SDK by Cavium Networks.
   4  *
   5  * Copyright (c) 2003-2010 Cavium Networks
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/kernel.h>
  10 #include <linux/cache.h>
  11 #include <linux/cpumask.h>
  12 #include <linux/netdevice.h>
  13 #include <linux/etherdevice.h>
  14 #include <linux/ip.h>
  15 #include <linux/string.h>
  16 #include <linux/prefetch.h>
  17 #include <linux/ratelimit.h>
  18 #include <linux/smp.h>
  19 #include <linux/interrupt.h>
  20 #include <net/dst.h>
  21 #ifdef CONFIG_XFRM
  22 #include <linux/xfrm.h>
  23 #include <net/xfrm.h>
  24 #endif /* CONFIG_XFRM */
  25
  26 #include "octeon-ethernet.h"
  27 #include "ethernet-defines.h"
  28 #include "ethernet-mem.h"
  29 #include "ethernet-rx.h"
  30 #include "ethernet-util.h"
  31
  32 static atomic_t oct_rx_ready = ATOMIC_INIT(0);
  33
  34 static struct oct_rx_group {
  35         int irq;
  36         int group;
  37         struct napi_struct napi;
  38 } oct_rx_group[16];
  39
  40 /**
  41  * cvm_oct_do_interrupt - interrupt handler.
  42  * @irq: Interrupt number.
  43  * @napi_id: Cookie to identify the NAPI instance.
  44  *
  45  * The interrupt occurs whenever the POW has packets in our group.
  46  *
  47  */
  48 static irqreturn_t cvm_oct_do_interrupt(int irq, void *napi_id)
  49 {
  50         /* Disable the IRQ and start napi_poll. */
  51         disable_irq_nosync(irq);
  52         napi_schedule(napi_id);
  53
  54         return IRQ_HANDLED;
  55 }
  56
  57 /**
  58  * cvm_oct_check_rcv_error - process receive errors
  59  * @work: Work queue entry pointing to the packet.
  60  *
  61  * Returns Non-zero if the packet can be dropped, zero otherwise.
  62  */
  63 static inline int cvm_oct_check_rcv_error(struct cvmx_wqe *work)
  64 {
  65         int port;
  66
  67         if (octeon_has_feature(OCTEON_FEATURE_PKND))
  68                 port = work->word0.pip.cn68xx.pknd;
  69         else
  70                 port = work->word1.cn38xx.ipprt;
  71
  72         if ((work->word2.snoip.err_code == 10) && (work->word1.len <= 64)) {
  73                 /*
  74                  * Ignore length errors on min size packets. Some
  75                  * equipment incorrectly pads packets to 64+4FCS
  76                  * instead of 60+4FCS.  Note these packets still get
  77                  * counted as frame errors.
  78                  */
  79         } else if (work->word2.snoip.err_code == 5 ||
  80                    work->word2.snoip.err_code == 7) {
  81                 /*
  82                  * We received a packet with either an alignment error
  83                  * or a FCS error. This may be signalling that we are
  84                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
  85                  * off. If this is the case we need to parse the
  86                  * packet to determine if we can remove a non spec
  87                  * preamble and generate a correct packet.
  88                  */
  89                 int interface = cvmx_helper_get_interface_num(port);
  90                 int index = cvmx_helper_get_interface_index_num(port);
  91                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
  92
  93                 gmxx_rxx_frm_ctl.u64 =
  94                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
  95                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
  96                         u8 *ptr =
  97                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
  98                         int i = 0;
  99
 100                         while (i < work->word1.len - 1) {
 101                                 if (*ptr != 0x55)
 102                                         break;
 103                                 ptr++;
 104                                 i++;
 105                         }
 106
 107                         if (*ptr == 0xd5) {
 108                                 /* Port received 0xd5 preamble */
 109                                 work->packet_ptr.s.addr += i + 1;
 110                                 work->word1.len -= i + 5;
 111                         } else if ((*ptr & 0xf) == 0xd) {
 112                                 /* Port received 0xd preamble */
 113                                 work->packet_ptr.s.addr += i;
 114                                 work->word1.len -= i + 4;
 115                                 for (i = 0; i < work->word1.len; i++) {
 116                                         *ptr =
 117                                             ((*ptr & 0xf0) >> 4) |
 118                                             ((*(ptr + 1) & 0xf) << 4);
 119                                         ptr++;
 120                                 }
 121                         } else {
 122                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 123                                                    port);
 124                                 cvm_oct_free_work(work);
 125                                 return 1;
 126                         }
 127                 }
 128         } else {
 129                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 130                                    port, work->word2.snoip.err_code);
 131                 cvm_oct_free_work(work);
 132                 return 1;
 133         }
 134
 135         return 0;
 136 }
 137
 138 static void copy_segments_to_skb(struct cvmx_wqe *work, struct sk_buff *skb)
 139 {
 140         int segments = work->word2.s.bufs;
 141         union cvmx_buf_ptr segment_ptr = work->packet_ptr;
 142         int len = work->word1.len;
 143         int segment_size;
 144
 145         while (segments--) {
 146                 union cvmx_buf_ptr next_ptr;
 147
 148                 next_ptr = *(union cvmx_buf_ptr *)
 149                         cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 150
 151                 /*
 152                  * Octeon Errata PKI-100: The segment size is wrong.
 153                  *
 154                  * Until it is fixed, calculate the segment size based on
 155                  * the packet pool buffer size.
 156                  * When it is fixed, the following line should be replaced
 157                  * with this one:
 158                  * int segment_size = segment_ptr.s.size;
 159                  */
 160                 segment_size =
 161                         CVMX_FPA_PACKET_POOL_SIZE -
 162                         (segment_ptr.s.addr -
 163                          (((segment_ptr.s.addr >> 7) -
 164                            segment_ptr.s.back) << 7));
 165
 166                 /* Don't copy more than what is left in the packet */
 167                 if (segment_size > len)
 168                         segment_size = len;
 169
 170                 /* Copy the data into the packet */
 171                 skb_put_data(skb, cvmx_phys_to_ptr(segment_ptr.s.addr),
 172                              segment_size);
 173                 len -= segment_size;
 174                 segment_ptr = next_ptr;
 175         }
 176 }
 177
 178 static int cvm_oct_poll(struct oct_rx_group *rx_group, int budget)
 179 {
 180         const int       coreid = cvmx_get_core_num();
 181         u64     old_group_mask;
 182         u64     old_scratch;
 183         int             rx_count = 0;
 184         int             did_work_request = 0;
 185         int             packet_not_copied;
 186
 187         /* Prefetch cvm_oct_device since we know we need it soon */
 188         prefetch(cvm_oct_device);
 189
 190         if (USE_ASYNC_IOBDMA) {
 191                 /* Save scratch in case userspace is using it */
 192                 CVMX_SYNCIOBDMA;
 193                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 194         }
 195
 196         /* Only allow work for our group (and preserve priorities) */
 197         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 198                 old_group_mask = cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid));
 199                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid),
 200                                BIT(rx_group->group));
 201                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 202         } else {
 203                 old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 204                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 205                                (old_group_mask & ~0xFFFFull) |
 206                                BIT(rx_group->group));
 207         }
 208
 209         if (USE_ASYNC_IOBDMA) {
 210                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 211                 did_work_request = 1;
 212         }
 213
 214         while (rx_count < budget) {
 215                 struct sk_buff *skb = NULL;
 216                 struct sk_buff **pskb = NULL;
 217                 int skb_in_hw;
 218                 struct cvmx_wqe *work;
 219                 int port;
 220
 221                 if (USE_ASYNC_IOBDMA && did_work_request)
 222                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 223                 else
 224                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 225
 226                 prefetch(work);
 227                 did_work_request = 0;
 228                 if (!work) {
 229                         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 230                                 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS,
 231                                                BIT(rx_group->group));
 232                                 cvmx_write_csr(CVMX_SSO_WQ_INT,
 233                                                BIT(rx_group->group));
 234                         } else {
 235                                 union cvmx_pow_wq_int wq_int;
 236
 237                                 wq_int.u64 = 0;
 238                                 wq_int.s.iq_dis = BIT(rx_group->group);
 239                                 wq_int.s.wq_int = BIT(rx_group->group);
 240                                 cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 241                         }
 242                         break;
 243                 }
 244                 pskb = (struct sk_buff **)
 245                         (cvm_oct_get_buffer_ptr(work->packet_ptr) -
 246                         sizeof(void *));
 247                 prefetch(pskb);
 248
 249                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 250                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 251                                                             CVMX_POW_NO_WAIT);
 252                         did_work_request = 1;
 253                 }
 254                 rx_count++;
 255
 256                 skb_in_hw = work->word2.s.bufs == 1;
 257                 if (likely(skb_in_hw)) {
 258                         skb = *pskb;
 259                         prefetch(&skb->head);
 260                         prefetch(&skb->len);
 261                 }
 262
 263                 if (octeon_has_feature(OCTEON_FEATURE_PKND))
 264                         port = work->word0.pip.cn68xx.pknd;
 265                 else
 266                         port = work->word1.cn38xx.ipprt;
 267
 268                 prefetch(cvm_oct_device[port]);
 269
 270                 /* Immediately throw away all packets with receive errors */
 271                 if (unlikely(work->word2.snoip.rcv_error)) {
 272                         if (cvm_oct_check_rcv_error(work))
 273                                 continue;
 274                 }
 275
 276                 /*
 277                  * We can only use the zero copy path if skbuffs are
 278                  * in the FPA pool and the packet fits in a single
 279                  * buffer.
 280                  */
 281                 if (likely(skb_in_hw)) {
 282                         skb->data = skb->head + work->packet_ptr.s.addr -
 283                                 cvmx_ptr_to_phys(skb->head);
 284                         prefetch(skb->data);
 285                         skb->len = work->word1.len;
 286                         skb_set_tail_pointer(skb, skb->len);
 287                         packet_not_copied = 1;
 288                 } else {
 289                         /*
 290                          * We have to copy the packet. First allocate
 291                          * an skbuff for it.
 292                          */
 293                         skb = dev_alloc_skb(work->word1.len);
 294                         if (!skb) {
 295                                 cvm_oct_free_work(work);
 296                                 continue;
 297                         }
 298
 299                         /*
 300                          * Check if we've received a packet that was
 301                          * entirely stored in the work entry.
 302                          */
 303                         if (unlikely(work->word2.s.bufs == 0)) {
 304                                 u8 *ptr = work->packet_data;
 305
 306                                 if (likely(!work->word2.s.not_IP)) {
 307                                         /*
 308                                          * The beginning of the packet
 309                                          * moves for IP packets.
 310                                          */
 311                                         if (work->word2.s.is_v6)
 312                                                 ptr += 2;
 313                                         else
 314                                                 ptr += 6;
 315                                 }
 316                                 skb_put_data(skb, ptr, work->word1.len);
 317                                 /* No packet buffers to free */
 318                         } else {
 319                                 copy_segments_to_skb(work, skb);
 320                         }
 321                         packet_not_copied = 0;
 322                 }
 323                 if (likely((port < TOTAL_NUMBER_OF_PORTS) &&
 324                            cvm_oct_device[port])) {
 325                         struct net_device *dev = cvm_oct_device[port];
 326
 327                         /*
 328                          * Only accept packets for devices that are
 329                          * currently up.
 330                          */
 331                         if (likely(dev->flags & IFF_UP)) {
 332                                 skb->protocol = eth_type_trans(skb, dev);
 333                                 skb->dev = dev;
 334
 335                                 if (unlikely(work->word2.s.not_IP ||
 336                                              work->word2.s.IP_exc ||
 337                                              work->word2.s.L4_error ||
 338                                              !work->word2.s.tcp_or_udp))
 339                                         skb->ip_summed = CHECKSUM_NONE;
 340                                 else
 341                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 342
 343                                 /* Increment RX stats for virtual ports */
 344                                 if (port >= CVMX_PIP_NUM_INPUT_PORTS) {
 345                                         dev->stats.rx_packets++;
 346                                         dev->stats.rx_bytes += skb->len;
 347                                 }
 348                                 netif_receive_skb(skb);
 349                         } else {
 350                                 /*
 351                                  * Drop any packet received for a device that
 352                                  * isn't up.
 353                                  */
 354                                 dev->stats.rx_dropped++;
 355                                 dev_kfree_skb_irq(skb);
 356                         }
 357                 } else {
 358                         /*
 359                          * Drop any packet received for a device that
 360                          * doesn't exist.
 361                          */
 362                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 363                                            port);
 364                         dev_kfree_skb_irq(skb);
 365                 }
 366                 /*
 367                  * Check to see if the skbuff and work share the same
 368                  * packet buffer.
 369                  */
 370                 if (likely(packet_not_copied)) {
 371                         /*
 372                          * This buffer needs to be replaced, increment
 373                          * the number of buffers we need to free by
 374                          * one.
 375                          */
 376                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 377                                               1);
 378
 379                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 380                 } else {
 381                         cvm_oct_free_work(work);
 382                 }
 383         }
 384         /* Restore the original POW group mask */
 385         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 386                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid), old_group_mask);
 387                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 388         } else {
 389                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 390         }
 391
 392         if (USE_ASYNC_IOBDMA) {
 393                 /* Restore the scratch area */
 394                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 395         }
 396         cvm_oct_rx_refill_pool(0);
 397
 398         return rx_count;
 399 }
 400
 401 /**
 402  * cvm_oct_napi_poll - the NAPI poll function.
 403  * @napi: The NAPI instance.
 404  * @budget: Maximum number of packets to receive.
 405  *
 406  * Returns the number of packets processed.
 407  */
 408 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 409 {
 410         struct oct_rx_group *rx_group = container_of(napi, struct oct_rx_group,
 411                                                      napi);
 412         int rx_count;
 413
 414         rx_count = cvm_oct_poll(rx_group, budget);
 415
 416         if (rx_count < budget) {
 417                 /* No more work */
 418                 napi_complete_done(napi, rx_count);
 419                 enable_irq(rx_group->irq);
 420         }
 421         return rx_count;
 422 }
 423
 424 #ifdef CONFIG_NET_POLL_CONTROLLER
 425 /**
 426  * cvm_oct_poll_controller - poll for receive packets
 427  * device.
 428  *
 429  * @dev:    Device to poll. Unused
 430  */
 431 void cvm_oct_poll_controller(struct net_device *dev)
 432 {
 433         int i;
 434
 435         if (!atomic_read(&oct_rx_ready))
 436                 return;
 437
 438         for (i = 0; i < ARRAY_SIZE(oct_rx_group); i++) {
 439                 if (!(pow_receive_groups & BIT(i)))
 440                         continue;
 441
 442                 cvm_oct_poll(&oct_rx_group[i], 16);
 443         }
 444 }
 445 #endif
 446
 447 void cvm_oct_rx_initialize(void)
 448 {
 449         int i;
 450         struct net_device *dev_for_napi = NULL;
 451
 452         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 453                 if (cvm_oct_device[i]) {
 454                         dev_for_napi = cvm_oct_device[i];
 455                         break;
 456                 }
 457         }
 458
 459         if (!dev_for_napi)
 460                 panic("No net_devices were allocated.");
 461
 462         for (i = 0; i < ARRAY_SIZE(oct_rx_group); i++) {
 463                 int ret;
 464
 465                 if (!(pow_receive_groups & BIT(i)))
 466                         continue;
 467
 468                 netif_napi_add(dev_for_napi, &oct_rx_group[i].napi,
 469                                cvm_oct_napi_poll, rx_napi_weight);
 470                 napi_enable(&oct_rx_group[i].napi);
 471
 472                 oct_rx_group[i].irq = OCTEON_IRQ_WORKQ0 + i;
 473                 oct_rx_group[i].group = i;
 474
 475                 /* Register an IRQ handler to receive POW interrupts */
 476                 ret = request_irq(oct_rx_group[i].irq, cvm_oct_do_interrupt, 0,
 477                                   "Ethernet", &oct_rx_group[i].napi);
 478                 if (ret)
 479                         panic("Could not acquire Ethernet IRQ %d\n",
 480                               oct_rx_group[i].irq);
 481
 482                 disable_irq_nosync(oct_rx_group[i].irq);
 483
 484                 /* Enable POW interrupt when our port has at least one packet */
 485                 if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 486                         union cvmx_sso_wq_int_thrx int_thr;
 487                         union cvmx_pow_wq_int_pc int_pc;
 488
 489                         int_thr.u64 = 0;
 490                         int_thr.s.tc_en = 1;
 491                         int_thr.s.tc_thr = 1;
 492                         cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(i), int_thr.u64);
 493
 494                         int_pc.u64 = 0;
 495                         int_pc.s.pc_thr = 5;
 496                         cvmx_write_csr(CVMX_SSO_WQ_INT_PC, int_pc.u64);
 497                 } else {
 498                         union cvmx_pow_wq_int_thrx int_thr;
 499                         union cvmx_pow_wq_int_pc int_pc;
 500
 501                         int_thr.u64 = 0;
 502                         int_thr.s.tc_en = 1;
 503                         int_thr.s.tc_thr = 1;
 504                         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(i), int_thr.u64);
 505
 506                         int_pc.u64 = 0;
 507                         int_pc.s.pc_thr = 5;
 508                         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 509                 }
 510
 511                 /* Schedule NAPI now. This will indirectly enable the
 512                  * interrupt.
 513                  */
 514                 napi_schedule(&oct_rx_group[i].napi);
 515         }
 516         atomic_inc(&oct_rx_ready);
 517 }
 518
 519 void cvm_oct_rx_shutdown(void)
 520 {
 521         int i;
 522
 523         for (i = 0; i < ARRAY_SIZE(oct_rx_group); i++) {
 524                 if (!(pow_receive_groups & BIT(i)))
 525                         continue;
 526
 527                 /* Disable POW interrupt */
 528                 if (OCTEON_IS_MODEL(OCTEON_CN68XX))
 529                         cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(i), 0);
 530                 else
 531                         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(i), 0);
 532
 533                 /* Free the interrupt handler */
 534                 free_irq(oct_rx_group[i].irq, cvm_oct_device);
 535
 536                 netif_napi_del(&oct_rx_group[i].napi);
 537         }
 538 }