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         if ((work->word2.snoip.err_code == 10) && (work->len <= 64)) {
  74                 /*
  75                  * Ignore length errors on min size packets. Some
  76                  * equipment incorrectly pads packets to 64+4FCS
  77                  * instead of 60+4FCS.  Note these packets still get
  78                  * counted as frame errors.
  79                  */
  80         } else if (work->word2.snoip.err_code == 5 ||
  81                    work->word2.snoip.err_code == 7) {
  82                 /*
  83                  * We received a packet with either an alignment error
  84                  * or a FCS error. This may be signalling that we are
  85                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
  86                  * off. If this is the case we need to parse the
  87                  * packet to determine if we can remove a non spec
  88                  * preamble and generate a correct packet.
  89                  */
  90                 int interface = cvmx_helper_get_interface_num(work->ipprt);
  91                 int index = cvmx_helper_get_interface_index_num(work->ipprt);
  92                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
  93
  94                 gmxx_rxx_frm_ctl.u64 =
  95                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
  96                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
  97
  98                         uint8_t *ptr =
  99                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 100                         int i = 0;
 101
 102                         while (i < work->len - 1) {
 103                                 if (*ptr != 0x55)
 104                                         break;
 105                                 ptr++;
 106                                 i++;
 107                         }
 108
 109                         if (*ptr == 0xd5) {
 110                                 /*
 111                                   printk_ratelimited("Port %d received 0xd5 preamble\n",
 112                                           work->ipprt);
 113                                  */
 114                                 work->packet_ptr.s.addr += i + 1;
 115                                 work->len -= i + 5;
 116                         } else if ((*ptr & 0xf) == 0xd) {
 117                                 /*
 118                                   printk_ratelimited("Port %d received 0x?d preamble\n",
 119                                           work->ipprt);
 120                                  */
 121                                 work->packet_ptr.s.addr += i;
 122                                 work->len -= i + 4;
 123                                 for (i = 0; i < work->len; i++) {
 124                                         *ptr =
 125                                             ((*ptr & 0xf0) >> 4) |
 126                                             ((*(ptr + 1) & 0xf) << 4);
 127                                         ptr++;
 128                                 }
 129                         } else {
 130                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 131                                                    work->ipprt);
 132                                 /*
 133                                    cvmx_helper_dump_packet(work);
 134                                  */
 135                                 cvm_oct_free_work(work);
 136                                 return 1;
 137                         }
 138                 }
 139         } else {
 140                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 141                                    work->ipprt, work->word2.snoip.err_code);
 142                 cvm_oct_free_work(work);
 143                 return 1;
 144         }
 145
 146         return 0;
 147 }
 148
 149 /**
 150  * cvm_oct_napi_poll - the NAPI poll function.
 151  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 152  * @budget: Maximum number of packets to receive.
 153  *
 154  * Returns the number of packets processed.
 155  */
 156 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 157 {
 158         const int       coreid = cvmx_get_core_num();
 159         uint64_t        old_group_mask;
 160         uint64_t        old_scratch;
 161         int             rx_count = 0;
 162         int             did_work_request = 0;
 163         int             packet_not_copied;
 164
 165         /* Prefetch cvm_oct_device since we know we need it soon */
 166         prefetch(cvm_oct_device);
 167
 168         if (USE_ASYNC_IOBDMA) {
 169                 /* Save scratch in case userspace is using it */
 170                 CVMX_SYNCIOBDMA;
 171                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 172         }
 173
 174         /* Only allow work for our group (and preserve priorities) */
 175         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 176                 old_group_mask = cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid));
 177                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid),
 178                                 1ull << pow_receive_group);
 179                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 180         } else {
 181                 old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 182                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 183                         (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 184         }
 185
 186         if (USE_ASYNC_IOBDMA) {
 187                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 188                 did_work_request = 1;
 189         }
 190
 191         while (rx_count < budget) {
 192                 struct sk_buff *skb = NULL;
 193                 struct sk_buff **pskb = NULL;
 194                 int skb_in_hw;
 195                 cvmx_wqe_t *work;
 196
 197                 if (USE_ASYNC_IOBDMA && did_work_request)
 198                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 199                 else
 200                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 201
 202                 prefetch(work);
 203                 did_work_request = 0;
 204                 if (work == NULL) {
 205                         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 206                                 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS,
 207                                                1ull << pow_receive_group);
 208                                 cvmx_write_csr(CVMX_SSO_WQ_INT,
 209                                                1ull << pow_receive_group);
 210                         } else {
 211                                 union cvmx_pow_wq_int wq_int;
 212
 213                                 wq_int.u64 = 0;
 214                                 wq_int.s.iq_dis = 1 << pow_receive_group;
 215                                 wq_int.s.wq_int = 1 << pow_receive_group;
 216                                 cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 217                         }
 218                         break;
 219                 }
 220                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) -
 221                         sizeof(void *));
 222                 prefetch(pskb);
 223
 224                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 225                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 226                                                             CVMX_POW_NO_WAIT);
 227                         did_work_request = 1;
 228                 }
 229                 rx_count++;
 230
 231                 skb_in_hw = work->word2.s.bufs == 1;
 232                 if (likely(skb_in_hw)) {
 233                         skb = *pskb;
 234                         prefetch(&skb->head);
 235                         prefetch(&skb->len);
 236                 }
 237                 prefetch(cvm_oct_device[work->ipprt]);
 238
 239                 /* Immediately throw away all packets with receive errors */
 240                 if (unlikely(work->word2.snoip.rcv_error)) {
 241                         if (cvm_oct_check_rcv_error(work))
 242                                 continue;
 243                 }
 244
 245                 /*
 246                  * We can only use the zero copy path if skbuffs are
 247                  * in the FPA pool and the packet fits in a single
 248                  * buffer.
 249                  */
 250                 if (likely(skb_in_hw)) {
 251                         skb->data = skb->head + work->packet_ptr.s.addr -
 252                                 cvmx_ptr_to_phys(skb->head);
 253                         prefetch(skb->data);
 254                         skb->len = work->len;
 255                         skb_set_tail_pointer(skb, skb->len);
 256                         packet_not_copied = 1;
 257                 } else {
 258                         /*
 259                          * We have to copy the packet. First allocate
 260                          * an skbuff for it.
 261                          */
 262                         skb = dev_alloc_skb(work->len);
 263                         if (!skb) {
 264                                 cvm_oct_free_work(work);
 265                                 continue;
 266                         }
 267
 268                         /*
 269                          * Check if we've received a packet that was
 270                          * entirely stored in the work entry.
 271                          */
 272                         if (unlikely(work->word2.s.bufs == 0)) {
 273                                 uint8_t *ptr = work->packet_data;
 274
 275                                 if (likely(!work->word2.s.not_IP)) {
 276                                         /*
 277                                          * The beginning of the packet
 278                                          * moves for IP packets.
 279                                          */
 280                                         if (work->word2.s.is_v6)
 281                                                 ptr += 2;
 282                                         else
 283                                                 ptr += 6;
 284                                 }
 285                                 memcpy(skb_put(skb, work->len), ptr, work->len);
 286                                 /* No packet buffers to free */
 287                         } else {
 288                                 int segments = work->word2.s.bufs;
 289                                 union cvmx_buf_ptr segment_ptr =
 290                                     work->packet_ptr;
 291                                 int len = work->len;
 292
 293                                 while (segments--) {
 294                                         union cvmx_buf_ptr next_ptr =
 295                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 296
 297                         /*
 298                          * Octeon Errata PKI-100: The segment size is
 299                          * wrong. Until it is fixed, calculate the
 300                          * segment size based on the packet pool
 301                          * buffer size. When it is fixed, the
 302                          * following line should be replaced with this
 303                          * one: int segment_size =
 304                          * segment_ptr.s.size;
 305                          */
 306                                         int segment_size =
 307                                             CVMX_FPA_PACKET_POOL_SIZE -
 308                                             (segment_ptr.s.addr -
 309                                              (((segment_ptr.s.addr >> 7) -
 310                                                segment_ptr.s.back) << 7));
 311                                         /*
 312                                          * Don't copy more than what
 313                                          * is left in the packet.
 314                                          */
 315                                         if (segment_size > len)
 316                                                 segment_size = len;
 317                                         /* Copy the data into the packet */
 318                                         memcpy(skb_put(skb, segment_size),
 319                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 320                                                segment_size);
 321                                         len -= segment_size;
 322                                         segment_ptr = next_ptr;
 323                                 }
 324                         }
 325                         packet_not_copied = 0;
 326                 }
 327
 328                 if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
 329                            cvm_oct_device[work->ipprt])) {
 330                         struct net_device *dev = cvm_oct_device[work->ipprt];
 331                         struct octeon_ethernet *priv = netdev_priv(dev);
 332
 333                         /*
 334                          * Only accept packets for devices that are
 335                          * currently up.
 336                          */
 337                         if (likely(dev->flags & IFF_UP)) {
 338                                 skb->protocol = eth_type_trans(skb, dev);
 339                                 skb->dev = dev;
 340
 341                                 if (unlikely(work->word2.s.not_IP ||
 342                                              work->word2.s.IP_exc ||
 343                                              work->word2.s.L4_error ||
 344                                              !work->word2.s.tcp_or_udp))
 345                                         skb->ip_summed = CHECKSUM_NONE;
 346                                 else
 347                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 348
 349                                 /* Increment RX stats for virtual ports */
 350                                 if (work->ipprt >= CVMX_PIP_NUM_INPUT_PORTS) {
 351 #ifdef CONFIG_64BIT
 352                                         atomic64_add(1,
 353                                                      (atomic64_t *)&priv->stats.rx_packets);
 354                                         atomic64_add(skb->len,
 355                                                      (atomic64_t *)&priv->stats.rx_bytes);
 356 #else
 357                                         atomic_add(1,
 358                                                    (atomic_t *)&priv->stats.rx_packets);
 359                                         atomic_add(skb->len,
 360                                                    (atomic_t *)&priv->stats.rx_bytes);
 361 #endif
 362                                 }
 363                                 netif_receive_skb(skb);
 364                         } else {
 365                                 /* Drop any packet received for a device that isn't up */
 366                                 /*
 367                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 368                                            dev->name);
 369                                 */
 370 #ifdef CONFIG_64BIT
 371                                 atomic64_add(1,
 372                                              (atomic64_t *)&priv->stats.rx_dropped);
 373 #else
 374                                 atomic_add(1,
 375                                            (atomic_t *)&priv->stats.rx_dropped);
 376 #endif
 377                                 dev_kfree_skb_irq(skb);
 378                         }
 379                 } else {
 380                         /*
 381                          * Drop any packet received for a device that
 382                          * doesn't exist.
 383                          */
 384                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 385                                    work->ipprt);
 386                         dev_kfree_skb_irq(skb);
 387                 }
 388                 /*
 389                  * Check to see if the skbuff and work share the same
 390                  * packet buffer.
 391                  */
 392                 if (likely(packet_not_copied)) {
 393                         /*
 394                          * This buffer needs to be replaced, increment
 395                          * the number of buffers we need to free by
 396                          * one.
 397                          */
 398                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 399                                               1);
 400
 401                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 402                 } else {
 403                         cvm_oct_free_work(work);
 404                 }
 405         }
 406         /* Restore the original POW group mask */
 407         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 408                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid), old_group_mask);
 409                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 410         } else {
 411                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 412         }
 413
 414         if (USE_ASYNC_IOBDMA) {
 415                 /* Restore the scratch area */
 416                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 417         }
 418         cvm_oct_rx_refill_pool(0);
 419
 420         if (rx_count < budget && napi != NULL) {
 421                 /* No more work */
 422                 napi_complete(napi);
 423                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 424         }
 425         return rx_count;
 426 }
 427
 428 #ifdef CONFIG_NET_POLL_CONTROLLER
 429 /**
 430  * cvm_oct_poll_controller - poll for receive packets
 431  * device.
 432  *
 433  * @dev:    Device to poll. Unused
 434  */
 435 void cvm_oct_poll_controller(struct net_device *dev)
 436 {
 437         cvm_oct_napi_poll(NULL, 16);
 438 }
 439 #endif
 440
 441 void cvm_oct_rx_initialize(void)
 442 {
 443         int i;
 444         struct net_device *dev_for_napi = NULL;
 445
 446         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 447                 if (cvm_oct_device[i]) {
 448                         dev_for_napi = cvm_oct_device[i];
 449                         break;
 450                 }
 451         }
 452
 453         if (NULL == dev_for_napi)
 454                 panic("No net_devices were allocated.");
 455
 456         netif_napi_add(dev_for_napi, &cvm_oct_napi, cvm_oct_napi_poll,
 457                        rx_napi_weight);
 458         napi_enable(&cvm_oct_napi);
 459
 460         /* Register an IRQ handler to receive POW interrupts */
 461         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 462                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 463
 464         if (i)
 465                 panic("Could not acquire Ethernet IRQ %d\n",
 466                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 467
 468         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 469
 470         /* Enable POW interrupt when our port has at least one packet */
 471         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 472                 union cvmx_sso_wq_int_thrx int_thr;
 473                 union cvmx_pow_wq_int_pc int_pc;
 474
 475                 int_thr.u64 = 0;
 476                 int_thr.s.tc_en = 1;
 477                 int_thr.s.tc_thr = 1;
 478                 cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(pow_receive_group),
 479                                int_thr.u64);
 480
 481                 int_pc.u64 = 0;
 482                 int_pc.s.pc_thr = 5;
 483                 cvmx_write_csr(CVMX_SSO_WQ_INT_PC, int_pc.u64);
 484         } else {
 485                 union cvmx_pow_wq_int_thrx int_thr;
 486                 union cvmx_pow_wq_int_pc int_pc;
 487
 488                 int_thr.u64 = 0;
 489                 int_thr.s.tc_en = 1;
 490                 int_thr.s.tc_thr = 1;
 491                 cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group),
 492                                int_thr.u64);
 493
 494                 int_pc.u64 = 0;
 495                 int_pc.s.pc_thr = 5;
 496                 cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 497         }
 498
 499         /* Schedule NAPI now. This will indirectly enable the interrupt. */
 500         napi_schedule(&cvm_oct_napi);
 501 }
 502
 503 void cvm_oct_rx_shutdown(void)
 504 {
 505         netif_napi_del(&cvm_oct_napi);
 506 }