drivers/staging/octeon/ethernet-rx.c

   1 /*
   2  * This file is based on code from OCTEON SDK by Cavium Networks.
   3  *
   4  * Copyright (c) 2003-2010 Cavium Networks
   5  *
   6  * This file is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License, Version 2, as
   8  * published by the Free Software Foundation.
   9  */
  10
  11 #include <linux/module.h>
  12 #include <linux/kernel.h>
  13 #include <linux/cache.h>
  14 #include <linux/cpumask.h>
  15 #include <linux/netdevice.h>
  16 #include <linux/etherdevice.h>
  17 #include <linux/ip.h>
  18 #include <linux/string.h>
  19 #include <linux/prefetch.h>
  20 #include <linux/ratelimit.h>
  21 #include <linux/smp.h>
  22 #include <linux/interrupt.h>
  23 #include <net/dst.h>
  24 #ifdef CONFIG_XFRM
  25 #include <linux/xfrm.h>
  26 #include <net/xfrm.h>
  27 #endif /* CONFIG_XFRM */
  28
  29 #include <asm/octeon/octeon.h>
  30
  31 #include "ethernet-defines.h"
  32 #include "ethernet-mem.h"
  33 #include "ethernet-rx.h"
  34 #include "octeon-ethernet.h"
  35 #include "ethernet-util.h"
  36
  37 #include <asm/octeon/cvmx-helper.h>
  38 #include <asm/octeon/cvmx-wqe.h>
  39 #include <asm/octeon/cvmx-fau.h>
  40 #include <asm/octeon/cvmx-pow.h>
  41 #include <asm/octeon/cvmx-pip.h>
  42 #include <asm/octeon/cvmx-scratch.h>
  43
  44 #include <asm/octeon/cvmx-gmxx-defs.h>
  45
  46 static struct oct_rx_group {
  47         int irq;
  48         int group;
  49         struct napi_struct napi;
  50 } oct_rx_group[16];
  51
  52 /**
  53  * cvm_oct_do_interrupt - interrupt handler.
  54  * @irq: Interrupt number.
  55  * @napi_id: Cookie to identify the NAPI instance.
  56  *
  57  * The interrupt occurs whenever the POW has packets in our group.
  58  *
  59  */
  60 static irqreturn_t cvm_oct_do_interrupt(int irq, void *napi_id)
  61 {
  62         /* Disable the IRQ and start napi_poll. */
  63         disable_irq_nosync(irq);
  64         napi_schedule(napi_id);
  65
  66         return IRQ_HANDLED;
  67 }
  68
  69 /**
  70  * cvm_oct_check_rcv_error - process receive errors
  71  * @work: Work queue entry pointing to the packet.
  72  *
  73  * Returns Non-zero if the packet can be dropped, zero otherwise.
  74  */
  75 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
  76 {
  77         int port;
  78
  79         if (octeon_has_feature(OCTEON_FEATURE_PKND))
  80                 port = work->word0.pip.cn68xx.pknd;
  81         else
  82                 port = work->word1.cn38xx.ipprt;
  83
  84         if ((work->word2.snoip.err_code == 10) && (work->word1.len <= 64)) {
  85                 /*
  86                  * Ignore length errors on min size packets. Some
  87                  * equipment incorrectly pads packets to 64+4FCS
  88                  * instead of 60+4FCS.  Note these packets still get
  89                  * counted as frame errors.
  90                  */
  91         } else if (work->word2.snoip.err_code == 5 ||
  92                    work->word2.snoip.err_code == 7) {
  93                 /*
  94                  * We received a packet with either an alignment error
  95                  * or a FCS error. This may be signalling that we are
  96                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
  97                  * off. If this is the case we need to parse the
  98                  * packet to determine if we can remove a non spec
  99                  * preamble and generate a correct packet.
 100                  */
 101                 int interface = cvmx_helper_get_interface_num(port);
 102                 int index = cvmx_helper_get_interface_index_num(port);
 103                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
 104
 105                 gmxx_rxx_frm_ctl.u64 =
 106                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 107                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 108                         u8 *ptr =
 109                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 110                         int i = 0;
 111
 112                         while (i < work->word1.len - 1) {
 113                                 if (*ptr != 0x55)
 114                                         break;
 115                                 ptr++;
 116                                 i++;
 117                         }
 118
 119                         if (*ptr == 0xd5) {
 120                                 /* Port received 0xd5 preamble */
 121                                 work->packet_ptr.s.addr += i + 1;
 122                                 work->word1.len -= i + 5;
 123                         } else if ((*ptr & 0xf) == 0xd) {
 124                                 /* Port received 0xd preamble */
 125                                 work->packet_ptr.s.addr += i;
 126                                 work->word1.len -= i + 4;
 127                                 for (i = 0; i < work->word1.len; i++) {
 128                                         *ptr =
 129                                             ((*ptr & 0xf0) >> 4) |
 130                                             ((*(ptr + 1) & 0xf) << 4);
 131                                         ptr++;
 132                                 }
 133                         } else {
 134                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 135                                                    port);
 136                                 cvm_oct_free_work(work);
 137                                 return 1;
 138                         }
 139                 }
 140         } else {
 141                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 142                                    port, work->word2.snoip.err_code);
 143                 cvm_oct_free_work(work);
 144                 return 1;
 145         }
 146
 147         return 0;
 148 }
 149
 150 static int cvm_oct_poll(struct oct_rx_group *rx_group, int budget)
 151 {
 152         const int       coreid = cvmx_get_core_num();
 153         u64     old_group_mask;
 154         u64     old_scratch;
 155         int             rx_count = 0;
 156         int             did_work_request = 0;
 157         int             packet_not_copied;
 158
 159         /* Prefetch cvm_oct_device since we know we need it soon */
 160         prefetch(cvm_oct_device);
 161
 162         if (USE_ASYNC_IOBDMA) {
 163                 /* Save scratch in case userspace is using it */
 164                 CVMX_SYNCIOBDMA;
 165                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 166         }
 167
 168         /* Only allow work for our group (and preserve priorities) */
 169         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 170                 old_group_mask = cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid));
 171                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid),
 172                                BIT(rx_group->group));
 173                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 174         } else {
 175                 old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 176                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 177                                (old_group_mask & ~0xFFFFull) |
 178                                BIT(rx_group->group));
 179         }
 180
 181         if (USE_ASYNC_IOBDMA) {
 182                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 183                 did_work_request = 1;
 184         }
 185
 186         while (rx_count < budget) {
 187                 struct sk_buff *skb = NULL;
 188                 struct sk_buff **pskb = NULL;
 189                 int skb_in_hw;
 190                 cvmx_wqe_t *work;
 191                 int port;
 192
 193                 if (USE_ASYNC_IOBDMA && did_work_request)
 194                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 195                 else
 196                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 197
 198                 prefetch(work);
 199                 did_work_request = 0;
 200                 if (!work) {
 201                         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 202                                 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS,
 203                                                BIT(rx_group->group));
 204                                 cvmx_write_csr(CVMX_SSO_WQ_INT,
 205                                                BIT(rx_group->group));
 206                         } else {
 207                                 union cvmx_pow_wq_int wq_int;
 208
 209                                 wq_int.u64 = 0;
 210                                 wq_int.s.iq_dis = BIT(rx_group->group);
 211                                 wq_int.s.wq_int = BIT(rx_group->group);
 212                                 cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 213                         }
 214                         break;
 215                 }
 216                 pskb = (struct sk_buff **)
 217                         (cvm_oct_get_buffer_ptr(work->packet_ptr) -
 218                         sizeof(void *));
 219                 prefetch(pskb);
 220
 221                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 222                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 223                                                             CVMX_POW_NO_WAIT);
 224                         did_work_request = 1;
 225                 }
 226                 rx_count++;
 227
 228                 skb_in_hw = work->word2.s.bufs == 1;
 229                 if (likely(skb_in_hw)) {
 230                         skb = *pskb;
 231                         prefetch(&skb->head);
 232                         prefetch(&skb->len);
 233                 }
 234
 235                 if (octeon_has_feature(OCTEON_FEATURE_PKND))
 236                         port = work->word0.pip.cn68xx.pknd;
 237                 else
 238                         port = work->word1.cn38xx.ipprt;
 239
 240                 prefetch(cvm_oct_device[port]);
 241
 242                 /* Immediately throw away all packets with receive errors */
 243                 if (unlikely(work->word2.snoip.rcv_error)) {
 244                         if (cvm_oct_check_rcv_error(work))
 245                                 continue;
 246                 }
 247
 248                 /*
 249                  * We can only use the zero copy path if skbuffs are
 250                  * in the FPA pool and the packet fits in a single
 251                  * buffer.
 252                  */
 253                 if (likely(skb_in_hw)) {
 254                         skb->data = skb->head + work->packet_ptr.s.addr -
 255                                 cvmx_ptr_to_phys(skb->head);
 256                         prefetch(skb->data);
 257                         skb->len = work->word1.len;
 258                         skb_set_tail_pointer(skb, skb->len);
 259                         packet_not_copied = 1;
 260                 } else {
 261                         /*
 262                          * We have to copy the packet. First allocate
 263                          * an skbuff for it.
 264                          */
 265                         skb = dev_alloc_skb(work->word1.len);
 266                         if (!skb) {
 267                                 cvm_oct_free_work(work);
 268                                 continue;
 269                         }
 270
 271                         /*
 272                          * Check if we've received a packet that was
 273                          * entirely stored in the work entry.
 274                          */
 275                         if (unlikely(work->word2.s.bufs == 0)) {
 276                                 u8 *ptr = work->packet_data;
 277
 278                                 if (likely(!work->word2.s.not_IP)) {
 279                                         /*
 280                                          * The beginning of the packet
 281                                          * moves for IP packets.
 282                                          */
 283                                         if (work->word2.s.is_v6)
 284                                                 ptr += 2;
 285                                         else
 286                                                 ptr += 6;
 287                                 }
 288                                 memcpy(skb_put(skb, work->word1.len), ptr,
 289                                        work->word1.len);
 290                                 /* No packet buffers to free */
 291                         } else {
 292                                 int segments = work->word2.s.bufs;
 293                                 union cvmx_buf_ptr segment_ptr =
 294                                     work->packet_ptr;
 295                                 int len = work->word1.len;
 296
 297                                 while (segments--) {
 298                                         union cvmx_buf_ptr next_ptr =
 299                                             *(union cvmx_buf_ptr *)
 300                                               cvmx_phys_to_ptr(
 301                                               segment_ptr.s.addr - 8);
 302
 303                         /*
 304                          * Octeon Errata PKI-100: The segment size is
 305                          * wrong. Until it is fixed, calculate the
 306                          * segment size based on the packet pool
 307                          * buffer size. When it is fixed, the
 308                          * following line should be replaced with this
 309                          * one: int segment_size =
 310                          * segment_ptr.s.size;
 311                          */
 312                                         int segment_size =
 313                                             CVMX_FPA_PACKET_POOL_SIZE -
 314                                             (segment_ptr.s.addr -
 315                                              (((segment_ptr.s.addr >> 7) -
 316                                                segment_ptr.s.back) << 7));
 317                                         /*
 318                                          * Don't copy more than what
 319                                          * is left in the packet.
 320                                          */
 321                                         if (segment_size > len)
 322                                                 segment_size = len;
 323                                         /* Copy the data into the packet */
 324                                         memcpy(skb_put(skb, segment_size),
 325                                                cvmx_phys_to_ptr(
 326                                                segment_ptr.s.addr),
 327                                                segment_size);
 328                                         len -= segment_size;
 329                                         segment_ptr = next_ptr;
 330                                 }
 331                         }
 332                         packet_not_copied = 0;
 333                 }
 334                 if (likely((port < TOTAL_NUMBER_OF_PORTS) &&
 335                            cvm_oct_device[port])) {
 336                         struct net_device *dev = cvm_oct_device[port];
 337                         struct octeon_ethernet *priv = netdev_priv(dev);
 338
 339                         /*
 340                          * Only accept packets for devices that are
 341                          * currently up.
 342                          */
 343                         if (likely(dev->flags & IFF_UP)) {
 344                                 skb->protocol = eth_type_trans(skb, dev);
 345                                 skb->dev = dev;
 346
 347                                 if (unlikely(work->word2.s.not_IP ||
 348                                              work->word2.s.IP_exc ||
 349                                              work->word2.s.L4_error ||
 350                                              !work->word2.s.tcp_or_udp))
 351                                         skb->ip_summed = CHECKSUM_NONE;
 352                                 else
 353                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 354
 355                                 /* Increment RX stats for virtual ports */
 356                                 if (port >= CVMX_PIP_NUM_INPUT_PORTS) {
 357                                         priv->stats.rx_packets++;
 358                                         priv->stats.rx_bytes += skb->len;
 359                                 }
 360                                 netif_receive_skb(skb);
 361                         } else {
 362                                 /*
 363                                  * Drop any packet received for a device that
 364                                  * isn't up.
 365                                  */
 366                                 priv->stats.rx_dropped++;
 367                                 dev_kfree_skb_irq(skb);
 368                         }
 369                 } else {
 370                         /*
 371                          * Drop any packet received for a device that
 372                          * doesn't exist.
 373                          */
 374                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 375                                            port);
 376                         dev_kfree_skb_irq(skb);
 377                 }
 378                 /*
 379                  * Check to see if the skbuff and work share the same
 380                  * packet buffer.
 381                  */
 382                 if (likely(packet_not_copied)) {
 383                         /*
 384                          * This buffer needs to be replaced, increment
 385                          * the number of buffers we need to free by
 386                          * one.
 387                          */
 388                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 389                                               1);
 390
 391                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 392                 } else {
 393                         cvm_oct_free_work(work);
 394                 }
 395         }
 396         /* Restore the original POW group mask */
 397         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 398                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid), old_group_mask);
 399                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 400         } else {
 401                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 402         }
 403
 404         if (USE_ASYNC_IOBDMA) {
 405                 /* Restore the scratch area */
 406                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 407         }
 408         cvm_oct_rx_refill_pool(0);
 409
 410         return rx_count;
 411 }
 412
 413 /**
 414  * cvm_oct_napi_poll - the NAPI poll function.
 415  * @napi: The NAPI instance.
 416  * @budget: Maximum number of packets to receive.
 417  *
 418  * Returns the number of packets processed.
 419  */
 420 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 421 {
 422         struct oct_rx_group *rx_group = container_of(napi, struct oct_rx_group,
 423                                                      napi);
 424         int rx_count;
 425
 426         rx_count = cvm_oct_poll(rx_group, budget);
 427
 428         if (rx_count < budget) {
 429                 /* No more work */
 430                 napi_complete(napi);
 431                 enable_irq(rx_group->irq);
 432         }
 433         return rx_count;
 434 }
 435
 436 #ifdef CONFIG_NET_POLL_CONTROLLER
 437 /**
 438  * cvm_oct_poll_controller - poll for receive packets
 439  * device.
 440  *
 441  * @dev:    Device to poll. Unused
 442  */
 443 void cvm_oct_poll_controller(struct net_device *dev)
 444 {
 445         int i;
 446
 447         for (i = 0; i < ARRAY_SIZE(oct_rx_group); i++) {
 448
 449                 if (!(pow_receive_groups & BIT(i)))
 450                         continue;
 451
 452                 cvm_oct_poll(&oct_rx_group[i], 16);
 453
 454         }
 455 }
 456 #endif
 457
 458 void cvm_oct_rx_initialize(void)
 459 {
 460         int i;
 461         struct net_device *dev_for_napi = NULL;
 462
 463         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 464                 if (cvm_oct_device[i]) {
 465                         dev_for_napi = cvm_oct_device[i];
 466                         break;
 467                 }
 468         }
 469
 470         if (!dev_for_napi)
 471                 panic("No net_devices were allocated.");
 472
 473         for (i = 0; i < ARRAY_SIZE(oct_rx_group); i++) {
 474                 int ret;
 475
 476                 if (!(pow_receive_groups & BIT(i)))
 477                         continue;
 478
 479                 netif_napi_add(dev_for_napi, &oct_rx_group[i].napi,
 480                                cvm_oct_napi_poll, rx_napi_weight);
 481                 napi_enable(&oct_rx_group[i].napi);
 482
 483                 oct_rx_group[i].irq = OCTEON_IRQ_WORKQ0 + i;
 484                 oct_rx_group[i].group = i;
 485
 486                 /* Register an IRQ handler to receive POW interrupts */
 487                 ret = request_irq(oct_rx_group[i].irq, cvm_oct_do_interrupt, 0,
 488                                   "Ethernet", &oct_rx_group[i].napi);
 489                 if (ret)
 490                         panic("Could not acquire Ethernet IRQ %d\n",
 491                               oct_rx_group[i].irq);
 492
 493                 disable_irq_nosync(oct_rx_group[i].irq);
 494
 495                 /* Enable POW interrupt when our port has at least one packet */
 496                 if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 497                         union cvmx_sso_wq_int_thrx int_thr;
 498                         union cvmx_pow_wq_int_pc int_pc;
 499
 500                         int_thr.u64 = 0;
 501                         int_thr.s.tc_en = 1;
 502                         int_thr.s.tc_thr = 1;
 503                         cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(i), int_thr.u64);
 504
 505                         int_pc.u64 = 0;
 506                         int_pc.s.pc_thr = 5;
 507                         cvmx_write_csr(CVMX_SSO_WQ_INT_PC, int_pc.u64);
 508                 } else {
 509                         union cvmx_pow_wq_int_thrx int_thr;
 510                         union cvmx_pow_wq_int_pc int_pc;
 511
 512                         int_thr.u64 = 0;
 513                         int_thr.s.tc_en = 1;
 514                         int_thr.s.tc_thr = 1;
 515                         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(i), int_thr.u64);
 516
 517                         int_pc.u64 = 0;
 518                         int_pc.s.pc_thr = 5;
 519                         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 520                 }
 521
 522                 /* Schedule NAPI now. This will indirectly enable the
 523                  * interrupt.
 524                  */
 525                 napi_schedule(&oct_rx_group[i].napi);
 526         }
 527 }
 528
 529 void cvm_oct_rx_shutdown(void)
 530 {
 531         int i;
 532
 533         for (i = 0; i < ARRAY_SIZE(oct_rx_group); i++) {
 534
 535                 if (!(pow_receive_groups & BIT(i)))
 536                         continue;
 537
 538                 /* Disable POW interrupt */
 539                 if (OCTEON_IS_MODEL(OCTEON_CN68XX))
 540                         cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(i), 0);
 541                 else
 542                         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(i), 0);
 543
 544                 /* Free the interrupt handler */
 545                 free_irq(oct_rx_group[i].irq, cvm_oct_device);
 546
 547                 netif_napi_del(&oct_rx_group[i].napi);
 548         }
 549 }