drivers/staging/octeon/ethernet-rx.c

   1 /**********************************************************************
   2  * Author: Cavium Networks
   3  *
   4  * Contact: support@caviumnetworks.com
   5  * This file is part of the OCTEON SDK
   6  *
   7  * Copyright (c) 2003-2010 Cavium Networks
   8  *
   9  * This file is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License, Version 2, as
  11  * published by the Free Software Foundation.
  12  *
  13  * This file is distributed in the hope that it will be useful, but
  14  * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty
  15  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or
  16  * NONINFRINGEMENT.  See the GNU General Public License for more
  17  * details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this file; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
  22  * or visit http://www.gnu.org/licenses/.
  23  *
  24  * This file may also be available under a different license from Cavium.
  25  * Contact Cavium Networks for more information
  26 **********************************************************************/
  27 #include <linux/module.h>
  28 #include <linux/kernel.h>
  29 #include <linux/cache.h>
  30 #include <linux/cpumask.h>
  31 #include <linux/netdevice.h>
  32 #include <linux/etherdevice.h>
  33 #include <linux/ip.h>
  34 #include <linux/string.h>
  35 #include <linux/prefetch.h>
  36 #include <linux/ratelimit.h>
  37 #include <linux/smp.h>
  38 #include <linux/interrupt.h>
  39 #include <net/dst.h>
  40 #ifdef CONFIG_XFRM
  41 #include <linux/xfrm.h>
  42 #include <net/xfrm.h>
  43 #endif /* CONFIG_XFRM */
  44
  45 #include <linux/atomic.h>
  46
  47 #include <asm/octeon/octeon.h>
  48
  49 #include "ethernet-defines.h"
  50 #include "ethernet-mem.h"
  51 #include "ethernet-rx.h"
  52 #include "octeon-ethernet.h"
  53 #include "ethernet-util.h"
  54
  55 #include <asm/octeon/cvmx-helper.h>
  56 #include <asm/octeon/cvmx-wqe.h>
  57 #include <asm/octeon/cvmx-fau.h>
  58 #include <asm/octeon/cvmx-pow.h>
  59 #include <asm/octeon/cvmx-pip.h>
  60 #include <asm/octeon/cvmx-scratch.h>
  61
  62 #include <asm/octeon/cvmx-gmxx-defs.h>
  63
  64 static struct napi_struct cvm_oct_napi;
  65
  66 /**
  67  * cvm_oct_do_interrupt - interrupt handler.
  68  *
  69  * The interrupt occurs whenever the POW has packets in our group.
  70  *
  71  */
  72 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
  73 {
  74         /* Disable the IRQ and start napi_poll. */
  75         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
  76         napi_schedule(&cvm_oct_napi);
  77
  78         return IRQ_HANDLED;
  79 }
  80
  81 /**
  82  * cvm_oct_check_rcv_error - process receive errors
  83  * @work: Work queue entry pointing to the packet.
  84  *
  85  * Returns Non-zero if the packet can be dropped, zero otherwise.
  86  */
  87 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
  88 {
  89         if ((work->word2.snoip.err_code == 10) && (work->len <= 64)) {
  90                 /*
  91                  * Ignore length errors on min size packets. Some
  92                  * equipment incorrectly pads packets to 64+4FCS
  93                  * instead of 60+4FCS.  Note these packets still get
  94                  * counted as frame errors.
  95                  */
  96         } else
  97             if (USE_10MBPS_PREAMBLE_WORKAROUND
  98                 && ((work->word2.snoip.err_code == 5)
  99                     || (work->word2.snoip.err_code == 7))) {
 100
 101                 /*
 102                  * We received a packet with either an alignment error
 103                  * or a FCS error. This may be signalling that we are
 104                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
 105                  * off. If this is the case we need to parse the
 106                  * packet to determine if we can remove a non spec
 107                  * preamble and generate a correct packet.
 108                  */
 109                 int interface = cvmx_helper_get_interface_num(work->ipprt);
 110                 int index = cvmx_helper_get_interface_index_num(work->ipprt);
 111                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
 112                 gmxx_rxx_frm_ctl.u64 =
 113                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 114                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 115
 116                         uint8_t *ptr =
 117                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 118                         int i = 0;
 119
 120                         while (i < work->len - 1) {
 121                                 if (*ptr != 0x55)
 122                                         break;
 123                                 ptr++;
 124                                 i++;
 125                         }
 126
 127                         if (*ptr == 0xd5) {
 128                                 /*
 129                                   printk_ratelimited("Port %d received 0xd5 preamble\n",
 130                                           work->ipprt);
 131                                  */
 132                                 work->packet_ptr.s.addr += i + 1;
 133                                 work->len -= i + 5;
 134                         } else if ((*ptr & 0xf) == 0xd) {
 135                                 /*
 136                                   printk_ratelimited("Port %d received 0x?d preamble\n",
 137                                           work->ipprt);
 138                                  */
 139                                 work->packet_ptr.s.addr += i;
 140                                 work->len -= i + 4;
 141                                 for (i = 0; i < work->len; i++) {
 142                                         *ptr =
 143                                             ((*ptr & 0xf0) >> 4) |
 144                                             ((*(ptr + 1) & 0xf) << 4);
 145                                         ptr++;
 146                                 }
 147                         } else {
 148                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 149                                                    work->ipprt);
 150                                 /*
 151                                    cvmx_helper_dump_packet(work);
 152                                  */
 153                                 cvm_oct_free_work(work);
 154                                 return 1;
 155                         }
 156                 }
 157         } else {
 158                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 159                                    work->ipprt, work->word2.snoip.err_code);
 160                 cvm_oct_free_work(work);
 161                 return 1;
 162         }
 163
 164         return 0;
 165 }
 166
 167 /**
 168  * cvm_oct_napi_poll - the NAPI poll function.
 169  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 170  * @budget: Maximum number of packets to receive.
 171  *
 172  * Returns the number of packets processed.
 173  */
 174 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 175 {
 176         const int       coreid = cvmx_get_core_num();
 177         uint64_t        old_group_mask;
 178         uint64_t        old_scratch;
 179         int             rx_count = 0;
 180         int             did_work_request = 0;
 181         int             packet_not_copied;
 182
 183         /* Prefetch cvm_oct_device since we know we need it soon */
 184         prefetch(cvm_oct_device);
 185
 186         if (USE_ASYNC_IOBDMA) {
 187                 /* Save scratch in case userspace is using it */
 188                 CVMX_SYNCIOBDMA;
 189                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 190         }
 191
 192         /* Only allow work for our group (and preserve priorities) */
 193         old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 194         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 195                        (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 196
 197         if (USE_ASYNC_IOBDMA) {
 198                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 199                 did_work_request = 1;
 200         }
 201
 202         while (rx_count < budget) {
 203                 struct sk_buff *skb = NULL;
 204                 struct sk_buff **pskb = NULL;
 205                 int skb_in_hw;
 206                 cvmx_wqe_t *work;
 207
 208                 if (USE_ASYNC_IOBDMA && did_work_request)
 209                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 210                 else
 211                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 212
 213                 prefetch(work);
 214                 did_work_request = 0;
 215                 if (work == NULL) {
 216                         union cvmx_pow_wq_int wq_int;
 217                         wq_int.u64 = 0;
 218                         wq_int.s.iq_dis = 1 << pow_receive_group;
 219                         wq_int.s.wq_int = 1 << pow_receive_group;
 220                         cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 221                         break;
 222                 }
 223                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) -
 224                         sizeof(void *));
 225                 prefetch(pskb);
 226
 227                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 228                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 229                                                             CVMX_POW_NO_WAIT);
 230                         did_work_request = 1;
 231                 }
 232                 rx_count++;
 233
 234                 skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
 235                 if (likely(skb_in_hw)) {
 236                         skb = *pskb;
 237                         prefetch(&skb->head);
 238                         prefetch(&skb->len);
 239                 }
 240                 prefetch(cvm_oct_device[work->ipprt]);
 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->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->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                                 uint8_t *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->len), ptr, work->len);
 289                                 /* No packet buffers to free */
 290                         } else {
 291                                 int segments = work->word2.s.bufs;
 292                                 union cvmx_buf_ptr segment_ptr =
 293                                     work->packet_ptr;
 294                                 int len = work->len;
 295
 296                                 while (segments--) {
 297                                         union cvmx_buf_ptr next_ptr =
 298                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 299
 300                         /*
 301                          * Octeon Errata PKI-100: The segment size is
 302                          * wrong. Until it is fixed, calculate the
 303                          * segment size based on the packet pool
 304                          * buffer size. When it is fixed, the
 305                          * following line should be replaced with this
 306                          * one: int segment_size =
 307                          * segment_ptr.s.size;
 308                          */
 309                                         int segment_size =
 310                                             CVMX_FPA_PACKET_POOL_SIZE -
 311                                             (segment_ptr.s.addr -
 312                                              (((segment_ptr.s.addr >> 7) -
 313                                                segment_ptr.s.back) << 7));
 314                                         /*
 315                                          * Don't copy more than what
 316                                          * is left in the packet.
 317                                          */
 318                                         if (segment_size > len)
 319                                                 segment_size = len;
 320                                         /* Copy the data into the packet */
 321                                         memcpy(skb_put(skb, segment_size),
 322                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 323                                                segment_size);
 324                                         len -= segment_size;
 325                                         segment_ptr = next_ptr;
 326                                 }
 327                         }
 328                         packet_not_copied = 0;
 329                 }
 330
 331                 if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
 332                            cvm_oct_device[work->ipprt])) {
 333                         struct net_device *dev = cvm_oct_device[work->ipprt];
 334                         struct octeon_ethernet *priv = netdev_priv(dev);
 335
 336                         /*
 337                          * Only accept packets for devices that are
 338                          * currently up.
 339                          */
 340                         if (likely(dev->flags & IFF_UP)) {
 341                                 skb->protocol = eth_type_trans(skb, dev);
 342                                 skb->dev = dev;
 343
 344                                 if (unlikely(work->word2.s.not_IP ||
 345                                              work->word2.s.IP_exc ||
 346                                              work->word2.s.L4_error ||
 347                                              !work->word2.s.tcp_or_udp))
 348                                         skb->ip_summed = CHECKSUM_NONE;
 349                                 else
 350                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 351
 352                                 /* Increment RX stats for virtual ports */
 353                                 if (work->ipprt >= CVMX_PIP_NUM_INPUT_PORTS) {
 354 #ifdef CONFIG_64BIT
 355                                         atomic64_add(1,
 356                                                      (atomic64_t *)&priv->stats.rx_packets);
 357                                         atomic64_add(skb->len,
 358                                                      (atomic64_t *)&priv->stats.rx_bytes);
 359 #else
 360                                         atomic_add(1,
 361                                                    (atomic_t *)&priv->stats.rx_packets);
 362                                         atomic_add(skb->len,
 363                                                    (atomic_t *)&priv->stats.rx_bytes);
 364 #endif
 365                                 }
 366                                 netif_receive_skb(skb);
 367                         } else {
 368                                 /* Drop any packet received for a device that isn't up */
 369                                 /*
 370                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 371                                            dev->name);
 372                                 */
 373 #ifdef CONFIG_64BIT
 374                                 atomic64_add(1,
 375                                              (atomic64_t *)&priv->stats.rx_dropped);
 376 #else
 377                                 atomic_add(1,
 378                                            (atomic_t *)&priv->stats.rx_dropped);
 379 #endif
 380                                 dev_kfree_skb_irq(skb);
 381                         }
 382                 } else {
 383                         /*
 384                          * Drop any packet received for a device that
 385                          * doesn't exist.
 386                          */
 387                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 388                                    work->ipprt);
 389                         dev_kfree_skb_irq(skb);
 390                 }
 391                 /*
 392                  * Check to see if the skbuff and work share the same
 393                  * packet buffer.
 394                  */
 395                 if (USE_SKBUFFS_IN_HW && likely(packet_not_copied)) {
 396                         /*
 397                          * This buffer needs to be replaced, increment
 398                          * the number of buffers we need to free by
 399                          * one.
 400                          */
 401                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 402                                               1);
 403
 404                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL,
 405                                       DONT_WRITEBACK(1));
 406                 } else {
 407                         cvm_oct_free_work(work);
 408                 }
 409         }
 410         /* Restore the original POW group mask */
 411         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 412         if (USE_ASYNC_IOBDMA) {
 413                 /* Restore the scratch area */
 414                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 415         }
 416         cvm_oct_rx_refill_pool(0);
 417
 418         if (rx_count < budget && napi != NULL) {
 419                 /* No more work */
 420                 napi_complete(napi);
 421                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 422         }
 423         return rx_count;
 424 }
 425
 426 #ifdef CONFIG_NET_POLL_CONTROLLER
 427 /**
 428  * cvm_oct_poll_controller - poll for receive packets
 429  * device.
 430  *
 431  * @dev:    Device to poll. Unused
 432  */
 433 void cvm_oct_poll_controller(struct net_device *dev)
 434 {
 435         cvm_oct_napi_poll(NULL, 16);
 436 }
 437 #endif
 438
 439 void cvm_oct_rx_initialize(void)
 440 {
 441         int i;
 442         struct net_device *dev_for_napi = NULL;
 443         union cvmx_pow_wq_int_thrx int_thr;
 444         union cvmx_pow_wq_int_pc int_pc;
 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         int_thr.u64 = 0;
 471         int_thr.s.tc_en = 1;
 472         int_thr.s.tc_thr = 1;
 473         /* Enable POW interrupt when our port has at least one packet */
 474         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
 475
 476         int_pc.u64 = 0;
 477         int_pc.s.pc_thr = 5;
 478         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 479
 480         /* Schedule NAPI now. This will indirectly enable the interrupt. */
 481         napi_schedule(&cvm_oct_napi);
 482 }
 483
 484 void cvm_oct_rx_shutdown(void)
 485 {
 486         netif_napi_del(&cvm_oct_napi);
 487 }