kernel/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  * @cpl: Interrupt number. Unused
  53  * @dev_id: Cookie to identify the device. Unused
  54  *
  55  * The interrupt occurs whenever the POW has packets in our group.
  56  *
  57  */
  58 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
  59 {
  60         /* Disable the IRQ and start napi_poll. */
  61         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
  62         napi_schedule(&cvm_oct_napi);
  63
  64         return IRQ_HANDLED;
  65 }
  66
  67 /**
  68  * cvm_oct_check_rcv_error - process receive errors
  69  * @work: Work queue entry pointing to the packet.
  70  *
  71  * Returns Non-zero if the packet can be dropped, zero otherwise.
  72  */
  73 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
  74 {
  75         int port;
  76
  77         if (octeon_has_feature(OCTEON_FEATURE_PKND))
  78                 port = work->word0.pip.cn68xx.pknd;
  79         else
  80                 port = work->word1.cn38xx.ipprt;
  81
  82         if ((work->word2.snoip.err_code == 10) && (work->word1.len <= 64)) {
  83                 /*
  84                  * Ignore length errors on min size packets. Some
  85                  * equipment incorrectly pads packets to 64+4FCS
  86                  * instead of 60+4FCS.  Note these packets still get
  87                  * counted as frame errors.
  88                  */
  89         } else if (work->word2.snoip.err_code == 5 ||
  90                    work->word2.snoip.err_code == 7) {
  91                 /*
  92                  * We received a packet with either an alignment error
  93                  * or a FCS error. This may be signalling that we are
  94                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
  95                  * off. If this is the case we need to parse the
  96                  * packet to determine if we can remove a non spec
  97                  * preamble and generate a correct packet.
  98                  */
  99                 int interface = cvmx_helper_get_interface_num(port);
 100                 int index = cvmx_helper_get_interface_index_num(port);
 101                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
 102
 103                 gmxx_rxx_frm_ctl.u64 =
 104                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 105                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 106
 107                         u8 *ptr =
 108                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 109                         int i = 0;
 110
 111                         while (i < work->word1.len - 1) {
 112                                 if (*ptr != 0x55)
 113                                         break;
 114                                 ptr++;
 115                                 i++;
 116                         }
 117
 118                         if (*ptr == 0xd5) {
 119                                 /*
 120                                   printk_ratelimited("Port %d received 0xd5 preamble\n",
 121                                           port);
 122                                  */
 123                                 work->packet_ptr.s.addr += i + 1;
 124                                 work->word1.len -= i + 5;
 125                         } else if ((*ptr & 0xf) == 0xd) {
 126                                 /*
 127                                   printk_ratelimited("Port %d received 0x?d preamble\n",
 128                                           port);
 129                                  */
 130                                 work->packet_ptr.s.addr += i;
 131                                 work->word1.len -= i + 4;
 132                                 for (i = 0; i < work->word1.len; i++) {
 133                                         *ptr =
 134                                             ((*ptr & 0xf0) >> 4) |
 135                                             ((*(ptr + 1) & 0xf) << 4);
 136                                         ptr++;
 137                                 }
 138                         } else {
 139                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 140                                                    port);
 141                                 /*
 142                                    cvmx_helper_dump_packet(work);
 143                                  */
 144                                 cvm_oct_free_work(work);
 145                                 return 1;
 146                         }
 147                 }
 148         } else {
 149                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 150                                    port, work->word2.snoip.err_code);
 151                 cvm_oct_free_work(work);
 152                 return 1;
 153         }
 154
 155         return 0;
 156 }
 157
 158 /**
 159  * cvm_oct_napi_poll - the NAPI poll function.
 160  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 161  * @budget: Maximum number of packets to receive.
 162  *
 163  * Returns the number of packets processed.
 164  */
 165 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 166 {
 167         const int       coreid = cvmx_get_core_num();
 168         u64     old_group_mask;
 169         u64     old_scratch;
 170         int             rx_count = 0;
 171         int             did_work_request = 0;
 172         int             packet_not_copied;
 173
 174         /* Prefetch cvm_oct_device since we know we need it soon */
 175         prefetch(cvm_oct_device);
 176
 177         if (USE_ASYNC_IOBDMA) {
 178                 /* Save scratch in case userspace is using it */
 179                 CVMX_SYNCIOBDMA;
 180                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 181         }
 182
 183         /* Only allow work for our group (and preserve priorities) */
 184         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 185                 old_group_mask = cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid));
 186                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid),
 187                                 1ull << pow_receive_group);
 188                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 189         } else {
 190                 old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 191                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 192                         (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 193         }
 194
 195         if (USE_ASYNC_IOBDMA) {
 196                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 197                 did_work_request = 1;
 198         }
 199
 200         while (rx_count < budget) {
 201                 struct sk_buff *skb = NULL;
 202                 struct sk_buff **pskb = NULL;
 203                 int skb_in_hw;
 204                 cvmx_wqe_t *work;
 205                 int port;
 206
 207                 if (USE_ASYNC_IOBDMA && did_work_request)
 208                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 209                 else
 210                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 211
 212                 prefetch(work);
 213                 did_work_request = 0;
 214                 if (work == NULL) {
 215                         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 216                                 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS,
 217                                                1ull << pow_receive_group);
 218                                 cvmx_write_csr(CVMX_SSO_WQ_INT,
 219                                                1ull << pow_receive_group);
 220                         } else {
 221                                 union cvmx_pow_wq_int wq_int;
 222
 223                                 wq_int.u64 = 0;
 224                                 wq_int.s.iq_dis = 1 << pow_receive_group;
 225                                 wq_int.s.wq_int = 1 << pow_receive_group;
 226                                 cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 227                         }
 228                         break;
 229                 }
 230                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) -
 231                         sizeof(void *));
 232                 prefetch(pskb);
 233
 234                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 235                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 236                                                             CVMX_POW_NO_WAIT);
 237                         did_work_request = 1;
 238                 }
 239                 rx_count++;
 240
 241                 skb_in_hw = work->word2.s.bufs == 1;
 242                 if (likely(skb_in_hw)) {
 243                         skb = *pskb;
 244                         prefetch(&skb->head);
 245                         prefetch(&skb->len);
 246                 }
 247
 248                 if (octeon_has_feature(OCTEON_FEATURE_PKND))
 249                         port = work->word0.pip.cn68xx.pknd;
 250                 else
 251                         port = work->word1.cn38xx.ipprt;
 252
 253                 prefetch(cvm_oct_device[port]);
 254
 255                 /* Immediately throw away all packets with receive errors */
 256                 if (unlikely(work->word2.snoip.rcv_error)) {
 257                         if (cvm_oct_check_rcv_error(work))
 258                                 continue;
 259                 }
 260
 261                 /*
 262                  * We can only use the zero copy path if skbuffs are
 263                  * in the FPA pool and the packet fits in a single
 264                  * buffer.
 265                  */
 266                 if (likely(skb_in_hw)) {
 267                         skb->data = skb->head + work->packet_ptr.s.addr -
 268                                 cvmx_ptr_to_phys(skb->head);
 269                         prefetch(skb->data);
 270                         skb->len = work->word1.len;
 271                         skb_set_tail_pointer(skb, skb->len);
 272                         packet_not_copied = 1;
 273                 } else {
 274                         /*
 275                          * We have to copy the packet. First allocate
 276                          * an skbuff for it.
 277                          */
 278                         skb = dev_alloc_skb(work->word1.len);
 279                         if (!skb) {
 280                                 cvm_oct_free_work(work);
 281                                 continue;
 282                         }
 283
 284                         /*
 285                          * Check if we've received a packet that was
 286                          * entirely stored in the work entry.
 287                          */
 288                         if (unlikely(work->word2.s.bufs == 0)) {
 289                                 u8 *ptr = work->packet_data;
 290
 291                                 if (likely(!work->word2.s.not_IP)) {
 292                                         /*
 293                                          * The beginning of the packet
 294                                          * moves for IP packets.
 295                                          */
 296                                         if (work->word2.s.is_v6)
 297                                                 ptr += 2;
 298                                         else
 299                                                 ptr += 6;
 300                                 }
 301                                 memcpy(skb_put(skb, work->word1.len), ptr,
 302                                        work->word1.len);
 303                                 /* No packet buffers to free */
 304                         } else {
 305                                 int segments = work->word2.s.bufs;
 306                                 union cvmx_buf_ptr segment_ptr =
 307                                     work->packet_ptr;
 308                                 int len = work->word1.len;
 309
 310                                 while (segments--) {
 311                                         union cvmx_buf_ptr next_ptr =
 312                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 313
 314                         /*
 315                          * Octeon Errata PKI-100: The segment size is
 316                          * wrong. Until it is fixed, calculate the
 317                          * segment size based on the packet pool
 318                          * buffer size. When it is fixed, the
 319                          * following line should be replaced with this
 320                          * one: int segment_size =
 321                          * segment_ptr.s.size;
 322                          */
 323                                         int segment_size =
 324                                             CVMX_FPA_PACKET_POOL_SIZE -
 325                                             (segment_ptr.s.addr -
 326                                              (((segment_ptr.s.addr >> 7) -
 327                                                segment_ptr.s.back) << 7));
 328                                         /*
 329                                          * Don't copy more than what
 330                                          * is left in the packet.
 331                                          */
 332                                         if (segment_size > len)
 333                                                 segment_size = len;
 334                                         /* Copy the data into the packet */
 335                                         memcpy(skb_put(skb, segment_size),
 336                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 337                                                segment_size);
 338                                         len -= segment_size;
 339                                         segment_ptr = next_ptr;
 340                                 }
 341                         }
 342                         packet_not_copied = 0;
 343                 }
 344                 if (likely((port < TOTAL_NUMBER_OF_PORTS) &&
 345                            cvm_oct_device[port])) {
 346                         struct net_device *dev = cvm_oct_device[port];
 347                         struct octeon_ethernet *priv = netdev_priv(dev);
 348
 349                         /*
 350                          * Only accept packets for devices that are
 351                          * currently up.
 352                          */
 353                         if (likely(dev->flags & IFF_UP)) {
 354                                 skb->protocol = eth_type_trans(skb, dev);
 355                                 skb->dev = dev;
 356
 357                                 if (unlikely(work->word2.s.not_IP ||
 358                                              work->word2.s.IP_exc ||
 359                                              work->word2.s.L4_error ||
 360                                              !work->word2.s.tcp_or_udp))
 361                                         skb->ip_summed = CHECKSUM_NONE;
 362                                 else
 363                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 364
 365                                 /* Increment RX stats for virtual ports */
 366                                 if (port >= CVMX_PIP_NUM_INPUT_PORTS) {
 367 #ifdef CONFIG_64BIT
 368                                         atomic64_add(1,
 369                                                      (atomic64_t *)&priv->stats.rx_packets);
 370                                         atomic64_add(skb->len,
 371                                                      (atomic64_t *)&priv->stats.rx_bytes);
 372 #else
 373                                         atomic_add(1,
 374                                                    (atomic_t *)&priv->stats.rx_packets);
 375                                         atomic_add(skb->len,
 376                                                    (atomic_t *)&priv->stats.rx_bytes);
 377 #endif
 378                                 }
 379                                 netif_receive_skb(skb);
 380                         } else {
 381                                 /* Drop any packet received for a device that isn't up */
 382                                 /*
 383                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 384                                            dev->name);
 385                                 */
 386 #ifdef CONFIG_64BIT
 387                                 atomic64_add(1,
 388                                              (atomic64_t *)&priv->stats.rx_dropped);
 389 #else
 390                                 atomic_add(1,
 391                                            (atomic_t *)&priv->stats.rx_dropped);
 392 #endif
 393                                 dev_kfree_skb_irq(skb);
 394                         }
 395                 } else {
 396                         /*
 397                          * Drop any packet received for a device that
 398                          * doesn't exist.
 399                          */
 400                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 401                                    port);
 402                         dev_kfree_skb_irq(skb);
 403                 }
 404                 /*
 405                  * Check to see if the skbuff and work share the same
 406                  * packet buffer.
 407                  */
 408                 if (likely(packet_not_copied)) {
 409                         /*
 410                          * This buffer needs to be replaced, increment
 411                          * the number of buffers we need to free by
 412                          * one.
 413                          */
 414                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 415                                               1);
 416
 417                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 418                 } else {
 419                         cvm_oct_free_work(work);
 420                 }
 421         }
 422         /* Restore the original POW group mask */
 423         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 424                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid), old_group_mask);
 425                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 426         } else {
 427                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 428         }
 429
 430         if (USE_ASYNC_IOBDMA) {
 431                 /* Restore the scratch area */
 432                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 433         }
 434         cvm_oct_rx_refill_pool(0);
 435
 436         if (rx_count < budget && napi != NULL) {
 437                 /* No more work */
 438                 napi_complete(napi);
 439                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 440         }
 441         return rx_count;
 442 }
 443
 444 #ifdef CONFIG_NET_POLL_CONTROLLER
 445 /**
 446  * cvm_oct_poll_controller - poll for receive packets
 447  * device.
 448  *
 449  * @dev:    Device to poll. Unused
 450  */
 451 void cvm_oct_poll_controller(struct net_device *dev)
 452 {
 453         cvm_oct_napi_poll(NULL, 16);
 454 }
 455 #endif
 456
 457 void cvm_oct_rx_initialize(void)
 458 {
 459         int i;
 460         struct net_device *dev_for_napi = NULL;
 461
 462         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 463                 if (cvm_oct_device[i]) {
 464                         dev_for_napi = cvm_oct_device[i];
 465                         break;
 466                 }
 467         }
 468
 469         if (NULL == dev_for_napi)
 470                 panic("No net_devices were allocated.");
 471
 472         netif_napi_add(dev_for_napi, &cvm_oct_napi, cvm_oct_napi_poll,
 473                        rx_napi_weight);
 474         napi_enable(&cvm_oct_napi);
 475
 476         /* Register an IRQ handler to receive POW interrupts */
 477         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 478                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 479
 480         if (i)
 481                 panic("Could not acquire Ethernet IRQ %d\n",
 482                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 483
 484         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 485
 486         /* Enable POW interrupt when our port has at least one packet */
 487         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 488                 union cvmx_sso_wq_int_thrx int_thr;
 489                 union cvmx_pow_wq_int_pc int_pc;
 490
 491                 int_thr.u64 = 0;
 492                 int_thr.s.tc_en = 1;
 493                 int_thr.s.tc_thr = 1;
 494                 cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(pow_receive_group),
 495                                int_thr.u64);
 496
 497                 int_pc.u64 = 0;
 498                 int_pc.s.pc_thr = 5;
 499                 cvmx_write_csr(CVMX_SSO_WQ_INT_PC, int_pc.u64);
 500         } else {
 501                 union cvmx_pow_wq_int_thrx int_thr;
 502                 union cvmx_pow_wq_int_pc int_pc;
 503
 504                 int_thr.u64 = 0;
 505                 int_thr.s.tc_en = 1;
 506                 int_thr.s.tc_thr = 1;
 507                 cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group),
 508                                int_thr.u64);
 509
 510                 int_pc.u64 = 0;
 511                 int_pc.s.pc_thr = 5;
 512                 cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 513         }
 514
 515         /* Schedule NAPI now. This will indirectly enable the interrupt. */
 516         napi_schedule(&cvm_oct_napi);
 517 }
 518
 519 void cvm_oct_rx_shutdown(void)
 520 {
 521         netif_napi_del(&cvm_oct_napi);
 522 }