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