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 <linux/atomic.h>
  30
  31 #include <asm/octeon/octeon.h>
  32
  33 #include "ethernet-defines.h"
  34 #include "ethernet-mem.h"
  35 #include "ethernet-rx.h"
  36 #include "octeon-ethernet.h"
  37 #include "ethernet-util.h"
  38
  39 #include <asm/octeon/cvmx-helper.h>
  40 #include <asm/octeon/cvmx-wqe.h>
  41 #include <asm/octeon/cvmx-fau.h>
  42 #include <asm/octeon/cvmx-pow.h>
  43 #include <asm/octeon/cvmx-pip.h>
  44 #include <asm/octeon/cvmx-scratch.h>
  45
  46 #include <asm/octeon/cvmx-gmxx-defs.h>
  47
  48 static struct napi_struct cvm_oct_napi;
  49
  50 /**
  51  * cvm_oct_do_interrupt - interrupt handler.
  52  *
  53  * The interrupt occurs whenever the POW has packets in our group.
  54  *
  55  */
  56 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
  57 {
  58         /* Disable the IRQ and start napi_poll. */
  59         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
  60         napi_schedule(&cvm_oct_napi);
  61
  62         return IRQ_HANDLED;
  63 }
  64
  65 /**
  66  * cvm_oct_check_rcv_error - process receive errors
  67  * @work: Work queue entry pointing to the packet.
  68  *
  69  * Returns Non-zero if the packet can be dropped, zero otherwise.
  70  */
  71 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
  72 {
  73         int port;
  74
  75         if (octeon_has_feature(OCTEON_FEATURE_PKND))
  76                 port = work->word0.pip.cn68xx.pknd;
  77         else
  78                 port = work->word1.cn38xx.ipprt;
  79
  80         if ((work->word2.snoip.err_code == 10) && (work->word1.len <= 64)) {
  81                 /*
  82                  * Ignore length errors on min size packets. Some
  83                  * equipment incorrectly pads packets to 64+4FCS
  84                  * instead of 60+4FCS.  Note these packets still get
  85                  * counted as frame errors.
  86                  */
  87         } else if (work->word2.snoip.err_code == 5 ||
  88                    work->word2.snoip.err_code == 7) {
  89                 /*
  90                  * We received a packet with either an alignment error
  91                  * or a FCS error. This may be signalling that we are
  92                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
  93                  * off. If this is the case we need to parse the
  94                  * packet to determine if we can remove a non spec
  95                  * preamble and generate a correct packet.
  96                  */
  97                 int interface = cvmx_helper_get_interface_num(port);
  98                 int index = cvmx_helper_get_interface_index_num(port);
  99                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
 100
 101                 gmxx_rxx_frm_ctl.u64 =
 102                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 103                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 104
 105                         uint8_t *ptr =
 106                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 107                         int i = 0;
 108
 109                         while (i < work->word1.len - 1) {
 110                                 if (*ptr != 0x55)
 111                                         break;
 112                                 ptr++;
 113                                 i++;
 114                         }
 115
 116                         if (*ptr == 0xd5) {
 117                                 /*
 118                                   printk_ratelimited("Port %d received 0xd5 preamble\n",
 119                                           port);
 120                                  */
 121                                 work->packet_ptr.s.addr += i + 1;
 122                                 work->word1.len -= i + 5;
 123                         } else if ((*ptr & 0xf) == 0xd) {
 124                                 /*
 125                                   printk_ratelimited("Port %d received 0x?d preamble\n",
 126                                           port);
 127                                  */
 128                                 work->packet_ptr.s.addr += i;
 129                                 work->word1.len -= i + 4;
 130                                 for (i = 0; i < work->word1.len; i++) {
 131                                         *ptr =
 132                                             ((*ptr & 0xf0) >> 4) |
 133                                             ((*(ptr + 1) & 0xf) << 4);
 134                                         ptr++;
 135                                 }
 136                         } else {
 137                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 138                                                    port);
 139                                 /*
 140                                    cvmx_helper_dump_packet(work);
 141                                  */
 142                                 cvm_oct_free_work(work);
 143                                 return 1;
 144                         }
 145                 }
 146         } else {
 147                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 148                                    port, work->word2.snoip.err_code);
 149                 cvm_oct_free_work(work);
 150                 return 1;
 151         }
 152
 153         return 0;
 154 }
 155
 156 /**
 157  * cvm_oct_napi_poll - the NAPI poll function.
 158  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 159  * @budget: Maximum number of packets to receive.
 160  *
 161  * Returns the number of packets processed.
 162  */
 163 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 164 {
 165         const int       coreid = cvmx_get_core_num();
 166         uint64_t        old_group_mask;
 167         uint64_t        old_scratch;
 168         int             rx_count = 0;
 169         int             did_work_request = 0;
 170         int             packet_not_copied;
 171
 172         /* Prefetch cvm_oct_device since we know we need it soon */
 173         prefetch(cvm_oct_device);
 174
 175         if (USE_ASYNC_IOBDMA) {
 176                 /* Save scratch in case userspace is using it */
 177                 CVMX_SYNCIOBDMA;
 178                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 179         }
 180
 181         /* Only allow work for our group (and preserve priorities) */
 182         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 183                 old_group_mask = cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid));
 184                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid),
 185                                 1ull << pow_receive_group);
 186                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 187         } else {
 188                 old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 189                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 190                         (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 191         }
 192
 193         if (USE_ASYNC_IOBDMA) {
 194                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 195                 did_work_request = 1;
 196         }
 197
 198         while (rx_count < budget) {
 199                 struct sk_buff *skb = NULL;
 200                 struct sk_buff **pskb = NULL;
 201                 int skb_in_hw;
 202                 cvmx_wqe_t *work;
 203                 int port;
 204
 205                 if (USE_ASYNC_IOBDMA && did_work_request)
 206                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 207                 else
 208                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 209
 210                 prefetch(work);
 211                 did_work_request = 0;
 212                 if (work == NULL) {
 213                         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 214                                 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS,
 215                                                1ull << pow_receive_group);
 216                                 cvmx_write_csr(CVMX_SSO_WQ_INT,
 217                                                1ull << pow_receive_group);
 218                         } else {
 219                                 union cvmx_pow_wq_int wq_int;
 220
 221                                 wq_int.u64 = 0;
 222                                 wq_int.s.iq_dis = 1 << pow_receive_group;
 223                                 wq_int.s.wq_int = 1 << pow_receive_group;
 224                                 cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 225                         }
 226                         break;
 227                 }
 228                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) -
 229                         sizeof(void *));
 230                 prefetch(pskb);
 231
 232                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 233                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 234                                                             CVMX_POW_NO_WAIT);
 235                         did_work_request = 1;
 236                 }
 237                 rx_count++;
 238
 239                 skb_in_hw = work->word2.s.bufs == 1;
 240                 if (likely(skb_in_hw)) {
 241                         skb = *pskb;
 242                         prefetch(&skb->head);
 243                         prefetch(&skb->len);
 244                 }
 245
 246                 if (octeon_has_feature(OCTEON_FEATURE_PKND))
 247                         port = work->word0.pip.cn68xx.pknd;
 248                 else
 249                         port = work->word1.cn38xx.ipprt;
 250
 251                 prefetch(cvm_oct_device[port]);
 252
 253                 /* Immediately throw away all packets with receive errors */
 254                 if (unlikely(work->word2.snoip.rcv_error)) {
 255                         if (cvm_oct_check_rcv_error(work))
 256                                 continue;
 257                 }
 258
 259                 /*
 260                  * We can only use the zero copy path if skbuffs are
 261                  * in the FPA pool and the packet fits in a single
 262                  * buffer.
 263                  */
 264                 if (likely(skb_in_hw)) {
 265                         skb->data = skb->head + work->packet_ptr.s.addr -
 266                                 cvmx_ptr_to_phys(skb->head);
 267                         prefetch(skb->data);
 268                         skb->len = work->word1.len;
 269                         skb_set_tail_pointer(skb, skb->len);
 270                         packet_not_copied = 1;
 271                 } else {
 272                         /*
 273                          * We have to copy the packet. First allocate
 274                          * an skbuff for it.
 275                          */
 276                         skb = dev_alloc_skb(work->word1.len);
 277                         if (!skb) {
 278                                 cvm_oct_free_work(work);
 279                                 continue;
 280                         }
 281
 282                         /*
 283                          * Check if we've received a packet that was
 284                          * entirely stored in the work entry.
 285                          */
 286                         if (unlikely(work->word2.s.bufs == 0)) {
 287                                 uint8_t *ptr = work->packet_data;
 288
 289                                 if (likely(!work->word2.s.not_IP)) {
 290                                         /*
 291                                          * The beginning of the packet
 292                                          * moves for IP packets.
 293                                          */
 294                                         if (work->word2.s.is_v6)
 295                                                 ptr += 2;
 296                                         else
 297                                                 ptr += 6;
 298                                 }
 299                                 memcpy(skb_put(skb, work->word1.len), ptr,
 300                                        work->word1.len);
 301                                 /* No packet buffers to free */
 302                         } else {
 303                                 int segments = work->word2.s.bufs;
 304                                 union cvmx_buf_ptr segment_ptr =
 305                                     work->packet_ptr;
 306                                 int len = work->word1.len;
 307
 308                                 while (segments--) {
 309                                         union cvmx_buf_ptr next_ptr =
 310                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 311
 312                         /*
 313                          * Octeon Errata PKI-100: The segment size is
 314                          * wrong. Until it is fixed, calculate the
 315                          * segment size based on the packet pool
 316                          * buffer size. When it is fixed, the
 317                          * following line should be replaced with this
 318                          * one: int segment_size =
 319                          * segment_ptr.s.size;
 320                          */
 321                                         int segment_size =
 322                                             CVMX_FPA_PACKET_POOL_SIZE -
 323                                             (segment_ptr.s.addr -
 324                                              (((segment_ptr.s.addr >> 7) -
 325                                                segment_ptr.s.back) << 7));
 326                                         /*
 327                                          * Don't copy more than what
 328                                          * is left in the packet.
 329                                          */
 330                                         if (segment_size > len)
 331                                                 segment_size = len;
 332                                         /* Copy the data into the packet */
 333                                         memcpy(skb_put(skb, segment_size),
 334                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 335                                                segment_size);
 336                                         len -= segment_size;
 337                                         segment_ptr = next_ptr;
 338                                 }
 339                         }
 340                         packet_not_copied = 0;
 341                 }
 342                 if (likely((port < TOTAL_NUMBER_OF_PORTS) &&
 343                            cvm_oct_device[port])) {
 344                         struct net_device *dev = cvm_oct_device[port];
 345                         struct octeon_ethernet *priv = netdev_priv(dev);
 346
 347                         /*
 348                          * Only accept packets for devices that are
 349                          * currently up.
 350                          */
 351                         if (likely(dev->flags & IFF_UP)) {
 352                                 skb->protocol = eth_type_trans(skb, dev);
 353                                 skb->dev = dev;
 354
 355                                 if (unlikely(work->word2.s.not_IP ||
 356                                              work->word2.s.IP_exc ||
 357                                              work->word2.s.L4_error ||
 358                                              !work->word2.s.tcp_or_udp))
 359                                         skb->ip_summed = CHECKSUM_NONE;
 360                                 else
 361                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 362
 363                                 /* Increment RX stats for virtual ports */
 364                                 if (port >= CVMX_PIP_NUM_INPUT_PORTS) {
 365 #ifdef CONFIG_64BIT
 366                                         atomic64_add(1,
 367                                                      (atomic64_t *)&priv->stats.rx_packets);
 368                                         atomic64_add(skb->len,
 369                                                      (atomic64_t *)&priv->stats.rx_bytes);
 370 #else
 371                                         atomic_add(1,
 372                                                    (atomic_t *)&priv->stats.rx_packets);
 373                                         atomic_add(skb->len,
 374                                                    (atomic_t *)&priv->stats.rx_bytes);
 375 #endif
 376                                 }
 377                                 netif_receive_skb(skb);
 378                         } else {
 379                                 /* Drop any packet received for a device that isn't up */
 380                                 /*
 381                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 382                                            dev->name);
 383                                 */
 384 #ifdef CONFIG_64BIT
 385                                 atomic64_add(1,
 386                                              (atomic64_t *)&priv->stats.rx_dropped);
 387 #else
 388                                 atomic_add(1,
 389                                            (atomic_t *)&priv->stats.rx_dropped);
 390 #endif
 391                                 dev_kfree_skb_irq(skb);
 392                         }
 393                 } else {
 394                         /*
 395                          * Drop any packet received for a device that
 396                          * doesn't exist.
 397                          */
 398                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 399                                    port);
 400                         dev_kfree_skb_irq(skb);
 401                 }
 402                 /*
 403                  * Check to see if the skbuff and work share the same
 404                  * packet buffer.
 405                  */
 406                 if (likely(packet_not_copied)) {
 407                         /*
 408                          * This buffer needs to be replaced, increment
 409                          * the number of buffers we need to free by
 410                          * one.
 411                          */
 412                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 413                                               1);
 414
 415                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 416                 } else {
 417                         cvm_oct_free_work(work);
 418                 }
 419         }
 420         /* Restore the original POW group mask */
 421         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 422                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid), old_group_mask);
 423                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 424         } else {
 425                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 426         }
 427
 428         if (USE_ASYNC_IOBDMA) {
 429                 /* Restore the scratch area */
 430                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 431         }
 432         cvm_oct_rx_refill_pool(0);
 433
 434         if (rx_count < budget && napi != NULL) {
 435                 /* No more work */
 436                 napi_complete(napi);
 437                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 438         }
 439         return rx_count;
 440 }
 441
 442 #ifdef CONFIG_NET_POLL_CONTROLLER
 443 /**
 444  * cvm_oct_poll_controller - poll for receive packets
 445  * device.
 446  *
 447  * @dev:    Device to poll. Unused
 448  */
 449 void cvm_oct_poll_controller(struct net_device *dev)
 450 {
 451         cvm_oct_napi_poll(NULL, 16);
 452 }
 453 #endif
 454
 455 void cvm_oct_rx_initialize(void)
 456 {
 457         int i;
 458         struct net_device *dev_for_napi = NULL;
 459
 460         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 461                 if (cvm_oct_device[i]) {
 462                         dev_for_napi = cvm_oct_device[i];
 463                         break;
 464                 }
 465         }
 466
 467         if (NULL == dev_for_napi)
 468                 panic("No net_devices were allocated.");
 469
 470         netif_napi_add(dev_for_napi, &cvm_oct_napi, cvm_oct_napi_poll,
 471                        rx_napi_weight);
 472         napi_enable(&cvm_oct_napi);
 473
 474         /* Register an IRQ handler to receive POW interrupts */
 475         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 476                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 477
 478         if (i)
 479                 panic("Could not acquire Ethernet IRQ %d\n",
 480                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 481
 482         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 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(pow_receive_group),
 493                                int_thr.u64);
 494
 495                 int_pc.u64 = 0;
 496                 int_pc.s.pc_thr = 5;
 497                 cvmx_write_csr(CVMX_SSO_WQ_INT_PC, int_pc.u64);
 498         } else {
 499                 union cvmx_pow_wq_int_thrx int_thr;
 500                 union cvmx_pow_wq_int_pc int_pc;
 501
 502                 int_thr.u64 = 0;
 503                 int_thr.s.tc_en = 1;
 504                 int_thr.s.tc_thr = 1;
 505                 cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group),
 506                                int_thr.u64);
 507
 508                 int_pc.u64 = 0;
 509                 int_pc.s.pc_thr = 5;
 510                 cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 511         }
 512
 513         /* Schedule NAPI now. This will indirectly enable the interrupt. */
 514         napi_schedule(&cvm_oct_napi);
 515 }
 516
 517 void cvm_oct_rx_shutdown(void)
 518 {
 519         netif_napi_del(&cvm_oct_napi);
 520 }