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/init.h>
  33 #include <linux/etherdevice.h>
  34 #include <linux/ip.h>
  35 #include <linux/string.h>
  36 #include <linux/prefetch.h>
  37 #include <linux/ratelimit.h>
  38 #include <linux/smp.h>
  39 #include <linux/interrupt.h>
  40 #include <net/dst.h>
  41 #ifdef CONFIG_XFRM
  42 #include <linux/xfrm.h>
  43 #include <net/xfrm.h>
  44 #endif /* CONFIG_XFRM */
  45
  46 #include <linux/atomic.h>
  47
  48 #include <asm/octeon/octeon.h>
  49
  50 #include "ethernet-defines.h"
  51 #include "ethernet-mem.h"
  52 #include "ethernet-rx.h"
  53 #include "octeon-ethernet.h"
  54 #include "ethernet-util.h"
  55
  56 #include <asm/octeon/cvmx-helper.h>
  57 #include <asm/octeon/cvmx-wqe.h>
  58 #include <asm/octeon/cvmx-fau.h>
  59 #include <asm/octeon/cvmx-pow.h>
  60 #include <asm/octeon/cvmx-pip.h>
  61 #include <asm/octeon/cvmx-scratch.h>
  62
  63 #include <asm/octeon/cvmx-gmxx-defs.h>
  64
  65 struct cvm_napi_wrapper {
  66         struct napi_struct napi;
  67 } ____cacheline_aligned_in_smp;
  68
  69 static struct cvm_napi_wrapper cvm_oct_napi[NR_CPUS] __cacheline_aligned_in_smp;
  70
  71 struct cvm_oct_core_state {
  72         int baseline_cores;
  73         /*
  74          * The number of additional cores that could be processing
  75          * input packets.
  76          */
  77         atomic_t available_cores;
  78         cpumask_t cpu_state;
  79 } ____cacheline_aligned_in_smp;
  80
  81 static struct cvm_oct_core_state core_state __cacheline_aligned_in_smp;
  82
  83 static int cvm_irq_cpu;
  84
  85 static void cvm_oct_enable_napi(void *_)
  86 {
  87         int cpu = smp_processor_id();
  88         napi_schedule(&cvm_oct_napi[cpu].napi);
  89 }
  90
  91 static void cvm_oct_enable_one_cpu(void)
  92 {
  93         int v;
  94         int cpu;
  95
  96         /* Check to see if more CPUs are available for receive processing... */
  97         v = atomic_sub_if_positive(1, &core_state.available_cores);
  98         if (v < 0)
  99                 return;
 100
 101         /* ... if a CPU is available, Turn on NAPI polling for that CPU.  */
 102         for_each_online_cpu(cpu) {
 103                 if (!cpu_test_and_set(cpu, core_state.cpu_state)) {
 104                         v = smp_call_function_single(cpu, cvm_oct_enable_napi,
 105                                                      NULL, 0);
 106                         if (v)
 107                                 panic("Can't enable NAPI.");
 108                         break;
 109                 }
 110         }
 111 }
 112
 113 static void cvm_oct_no_more_work(void)
 114 {
 115         int cpu = smp_processor_id();
 116
 117         if (cpu == cvm_irq_cpu) {
 118                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 119                 return;
 120         }
 121
 122         cpu_clear(cpu, core_state.cpu_state);
 123         atomic_add(1, &core_state.available_cores);
 124 }
 125
 126 /**
 127  * cvm_oct_do_interrupt - interrupt handler.
 128  *
 129  * The interrupt occurs whenever the POW has packets in our group.
 130  *
 131  */
 132 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
 133 {
 134         /* Disable the IRQ and start napi_poll. */
 135         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 136         cvm_irq_cpu = smp_processor_id();
 137         cvm_oct_enable_napi(NULL);
 138
 139         return IRQ_HANDLED;
 140 }
 141
 142 /**
 143  * cvm_oct_check_rcv_error - process receive errors
 144  * @work: Work queue entry pointing to the packet.
 145  *
 146  * Returns Non-zero if the packet can be dropped, zero otherwise.
 147  */
 148 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
 149 {
 150         if ((work->word2.snoip.err_code == 10) && (work->len <= 64)) {
 151                 /*
 152                  * Ignore length errors on min size packets. Some
 153                  * equipment incorrectly pads packets to 64+4FCS
 154                  * instead of 60+4FCS.  Note these packets still get
 155                  * counted as frame errors.
 156                  */
 157         } else
 158             if (USE_10MBPS_PREAMBLE_WORKAROUND
 159                 && ((work->word2.snoip.err_code == 5)
 160                     || (work->word2.snoip.err_code == 7))) {
 161
 162                 /*
 163                  * We received a packet with either an alignment error
 164                  * or a FCS error. This may be signalling that we are
 165                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
 166                  * off. If this is the case we need to parse the
 167                  * packet to determine if we can remove a non spec
 168                  * preamble and generate a correct packet.
 169                  */
 170                 int interface = cvmx_helper_get_interface_num(work->ipprt);
 171                 int index = cvmx_helper_get_interface_index_num(work->ipprt);
 172                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
 173                 gmxx_rxx_frm_ctl.u64 =
 174                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 175                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 176
 177                         uint8_t *ptr =
 178                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 179                         int i = 0;
 180
 181                         while (i < work->len - 1) {
 182                                 if (*ptr != 0x55)
 183                                         break;
 184                                 ptr++;
 185                                 i++;
 186                         }
 187
 188                         if (*ptr == 0xd5) {
 189                                 /*
 190                                   printk_ratelimited("Port %d received 0xd5 preamble\n", work->ipprt);
 191                                  */
 192                                 work->packet_ptr.s.addr += i + 1;
 193                                 work->len -= i + 5;
 194                         } else if ((*ptr & 0xf) == 0xd) {
 195                                 /*
 196                                   printk_ratelimited("Port %d received 0x?d preamble\n", work->ipprt);
 197                                  */
 198                                 work->packet_ptr.s.addr += i;
 199                                 work->len -= i + 4;
 200                                 for (i = 0; i < work->len; i++) {
 201                                         *ptr =
 202                                             ((*ptr & 0xf0) >> 4) |
 203                                             ((*(ptr + 1) & 0xf) << 4);
 204                                         ptr++;
 205                                 }
 206                         } else {
 207                                 printk_ratelimited("Port %d unknown preamble, packet "
 208                                                    "dropped\n",
 209                                                    work->ipprt);
 210                                 /*
 211                                    cvmx_helper_dump_packet(work);
 212                                  */
 213                                 cvm_oct_free_work(work);
 214                                 return 1;
 215                         }
 216                 }
 217         } else {
 218                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 219                                    work->ipprt, work->word2.snoip.err_code);
 220                 cvm_oct_free_work(work);
 221                 return 1;
 222         }
 223
 224         return 0;
 225 }
 226
 227 /**
 228  * cvm_oct_napi_poll - the NAPI poll function.
 229  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 230  * @budget: Maximum number of packets to receive.
 231  *
 232  * Returns the number of packets processed.
 233  */
 234 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 235 {
 236         const int       coreid = cvmx_get_core_num();
 237         uint64_t        old_group_mask;
 238         uint64_t        old_scratch;
 239         int             rx_count = 0;
 240         int             did_work_request = 0;
 241         int             packet_not_copied;
 242
 243         /* Prefetch cvm_oct_device since we know we need it soon */
 244         prefetch(cvm_oct_device);
 245
 246         if (USE_ASYNC_IOBDMA) {
 247                 /* Save scratch in case userspace is using it */
 248                 CVMX_SYNCIOBDMA;
 249                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 250         }
 251
 252         /* Only allow work for our group (and preserve priorities) */
 253         old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 254         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 255                        (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 256
 257         if (USE_ASYNC_IOBDMA) {
 258                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 259                 did_work_request = 1;
 260         }
 261
 262         while (rx_count < budget) {
 263                 struct sk_buff *skb = NULL;
 264                 struct sk_buff **pskb = NULL;
 265                 int skb_in_hw;
 266                 cvmx_wqe_t *work;
 267
 268                 if (USE_ASYNC_IOBDMA && did_work_request)
 269                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 270                 else
 271                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 272
 273                 prefetch(work);
 274                 did_work_request = 0;
 275                 if (work == NULL) {
 276                         union cvmx_pow_wq_int wq_int;
 277                         wq_int.u64 = 0;
 278                         wq_int.s.iq_dis = 1 << pow_receive_group;
 279                         wq_int.s.wq_int = 1 << pow_receive_group;
 280                         cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 281                         break;
 282                 }
 283                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) - sizeof(void *));
 284                 prefetch(pskb);
 285
 286                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 287                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 288                         did_work_request = 1;
 289                 }
 290
 291                 if (rx_count == 0) {
 292                         /*
 293                          * First time through, see if there is enough
 294                          * work waiting to merit waking another
 295                          * CPU.
 296                          */
 297                         union cvmx_pow_wq_int_cntx counts;
 298                         int backlog;
 299                         int cores_in_use = core_state.baseline_cores - atomic_read(&core_state.available_cores);
 300                         counts.u64 = cvmx_read_csr(CVMX_POW_WQ_INT_CNTX(pow_receive_group));
 301                         backlog = counts.s.iq_cnt + counts.s.ds_cnt;
 302                         if (backlog > budget * cores_in_use && napi != NULL)
 303                                 cvm_oct_enable_one_cpu();
 304                 }
 305                 rx_count++;
 306
 307                 skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
 308                 if (likely(skb_in_hw)) {
 309                         skb = *pskb;
 310                         prefetch(&skb->head);
 311                         prefetch(&skb->len);
 312                 }
 313                 prefetch(cvm_oct_device[work->ipprt]);
 314
 315                 /* Immediately throw away all packets with receive errors */
 316                 if (unlikely(work->word2.snoip.rcv_error)) {
 317                         if (cvm_oct_check_rcv_error(work))
 318                                 continue;
 319                 }
 320
 321                 /*
 322                  * We can only use the zero copy path if skbuffs are
 323                  * in the FPA pool and the packet fits in a single
 324                  * buffer.
 325                  */
 326                 if (likely(skb_in_hw)) {
 327                         skb->data = skb->head + work->packet_ptr.s.addr - cvmx_ptr_to_phys(skb->head);
 328                         prefetch(skb->data);
 329                         skb->len = work->len;
 330                         skb_set_tail_pointer(skb, skb->len);
 331                         packet_not_copied = 1;
 332                 } else {
 333                         /*
 334                          * We have to copy the packet. First allocate
 335                          * an skbuff for it.
 336                          */
 337                         skb = dev_alloc_skb(work->len);
 338                         if (!skb) {
 339                                 cvm_oct_free_work(work);
 340                                 continue;
 341                         }
 342
 343                         /*
 344                          * Check if we've received a packet that was
 345                          * entirely stored in the work entry.
 346                          */
 347                         if (unlikely(work->word2.s.bufs == 0)) {
 348                                 uint8_t *ptr = work->packet_data;
 349
 350                                 if (likely(!work->word2.s.not_IP)) {
 351                                         /*
 352                                          * The beginning of the packet
 353                                          * moves for IP packets.
 354                                          */
 355                                         if (work->word2.s.is_v6)
 356                                                 ptr += 2;
 357                                         else
 358                                                 ptr += 6;
 359                                 }
 360                                 memcpy(skb_put(skb, work->len), ptr, work->len);
 361                                 /* No packet buffers to free */
 362                         } else {
 363                                 int segments = work->word2.s.bufs;
 364                                 union cvmx_buf_ptr segment_ptr = work->packet_ptr;
 365                                 int len = work->len;
 366
 367                                 while (segments--) {
 368                                         union cvmx_buf_ptr next_ptr =
 369                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 370
 371                         /*
 372                          * Octeon Errata PKI-100: The segment size is
 373                          * wrong. Until it is fixed, calculate the
 374                          * segment size based on the packet pool
 375                          * buffer size. When it is fixed, the
 376                          * following line should be replaced with this
 377                          * one: int segment_size =
 378                          * segment_ptr.s.size;
 379                          */
 380                                         int segment_size = CVMX_FPA_PACKET_POOL_SIZE -
 381                                                 (segment_ptr.s.addr - (((segment_ptr.s.addr >> 7) - segment_ptr.s.back) << 7));
 382                                         /*
 383                                          * Don't copy more than what
 384                                          * is left in the packet.
 385                                          */
 386                                         if (segment_size > len)
 387                                                 segment_size = len;
 388                                         /* Copy the data into the packet */
 389                                         memcpy(skb_put(skb, segment_size),
 390                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 391                                                segment_size);
 392                                         len -= segment_size;
 393                                         segment_ptr = next_ptr;
 394                                 }
 395                         }
 396                         packet_not_copied = 0;
 397                 }
 398
 399                 if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
 400                            cvm_oct_device[work->ipprt])) {
 401                         struct net_device *dev = cvm_oct_device[work->ipprt];
 402                         struct octeon_ethernet *priv = netdev_priv(dev);
 403
 404                         /*
 405                          * Only accept packets for devices that are
 406                          * currently up.
 407                          */
 408                         if (likely(dev->flags & IFF_UP)) {
 409                                 skb->protocol = eth_type_trans(skb, dev);
 410                                 skb->dev = dev;
 411
 412                                 if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc ||
 413                                         work->word2.s.L4_error || !work->word2.s.tcp_or_udp))
 414                                         skb->ip_summed = CHECKSUM_NONE;
 415                                 else
 416                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 417
 418                                 /* Increment RX stats for virtual ports */
 419                                 if (work->ipprt >= CVMX_PIP_NUM_INPUT_PORTS) {
 420 #ifdef CONFIG_64BIT
 421                                         atomic64_add(1, (atomic64_t *)&priv->stats.rx_packets);
 422                                         atomic64_add(skb->len, (atomic64_t *)&priv->stats.rx_bytes);
 423 #else
 424                                         atomic_add(1, (atomic_t *)&priv->stats.rx_packets);
 425                                         atomic_add(skb->len, (atomic_t *)&priv->stats.rx_bytes);
 426 #endif
 427                                 }
 428                                 netif_receive_skb(skb);
 429                         } else {
 430                                 /* Drop any packet received for a device that isn't up */
 431                                 /*
 432                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 433                                            dev->name);
 434                                 */
 435 #ifdef CONFIG_64BIT
 436                                 atomic64_add(1, (atomic64_t *)&priv->stats.rx_dropped);
 437 #else
 438                                 atomic_add(1, (atomic_t *)&priv->stats.rx_dropped);
 439 #endif
 440                                 dev_kfree_skb_irq(skb);
 441                         }
 442                 } else {
 443                         /*
 444                          * Drop any packet received for a device that
 445                          * doesn't exist.
 446                          */
 447                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 448                                    work->ipprt);
 449                         dev_kfree_skb_irq(skb);
 450                 }
 451                 /*
 452                  * Check to see if the skbuff and work share the same
 453                  * packet buffer.
 454                  */
 455                 if (USE_SKBUFFS_IN_HW && likely(packet_not_copied)) {
 456                         /*
 457                          * This buffer needs to be replaced, increment
 458                          * the number of buffers we need to free by
 459                          * one.
 460                          */
 461                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 462                                               1);
 463
 464                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL,
 465                                       DONT_WRITEBACK(1));
 466                 } else {
 467                         cvm_oct_free_work(work);
 468                 }
 469         }
 470         /* Restore the original POW group mask */
 471         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 472         if (USE_ASYNC_IOBDMA) {
 473                 /* Restore the scratch area */
 474                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 475         }
 476         cvm_oct_rx_refill_pool(0);
 477
 478         if (rx_count < budget && napi != NULL) {
 479                 /* No more work */
 480                 napi_complete(napi);
 481                 cvm_oct_no_more_work();
 482         }
 483         return rx_count;
 484 }
 485
 486 #ifdef CONFIG_NET_POLL_CONTROLLER
 487 /**
 488  * cvm_oct_poll_controller - poll for receive packets
 489  * device.
 490  *
 491  * @dev:    Device to poll. Unused
 492  */
 493 void cvm_oct_poll_controller(struct net_device *dev)
 494 {
 495         cvm_oct_napi_poll(NULL, 16);
 496 }
 497 #endif
 498
 499 void cvm_oct_rx_initialize(void)
 500 {
 501         int i;
 502         struct net_device *dev_for_napi = NULL;
 503         union cvmx_pow_wq_int_thrx int_thr;
 504         union cvmx_pow_wq_int_pc int_pc;
 505
 506         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 507                 if (cvm_oct_device[i]) {
 508                         dev_for_napi = cvm_oct_device[i];
 509                         break;
 510                 }
 511         }
 512
 513         if (NULL == dev_for_napi)
 514                 panic("No net_devices were allocated.");
 515
 516         if (max_rx_cpus >= 1 && max_rx_cpus < num_online_cpus())
 517                 atomic_set(&core_state.available_cores, max_rx_cpus);
 518         else
 519                 atomic_set(&core_state.available_cores, num_online_cpus());
 520         core_state.baseline_cores = atomic_read(&core_state.available_cores);
 521
 522         core_state.cpu_state = CPU_MASK_NONE;
 523         for_each_possible_cpu(i) {
 524                 netif_napi_add(dev_for_napi, &cvm_oct_napi[i].napi,
 525                                cvm_oct_napi_poll, rx_napi_weight);
 526                 napi_enable(&cvm_oct_napi[i].napi);
 527         }
 528         /* Register an IRQ handler to receive POW interrupts */
 529         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 530                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 531
 532         if (i)
 533                 panic("Could not acquire Ethernet IRQ %d\n",
 534                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 535
 536         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 537
 538         int_thr.u64 = 0;
 539         int_thr.s.tc_en = 1;
 540         int_thr.s.tc_thr = 1;
 541         /* Enable POW interrupt when our port has at least one packet */
 542         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
 543
 544         int_pc.u64 = 0;
 545         int_pc.s.pc_thr = 5;
 546         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 547
 548
 549         /* Scheduld NAPI now.  This will indirectly enable interrupts. */
 550         cvm_oct_enable_one_cpu();
 551 }
 552
 553 void cvm_oct_rx_shutdown(void)
 554 {
 555         int i;
 556         /* Shutdown all of the NAPIs */
 557         for_each_possible_cpu(i)
 558                 netif_napi_del(&cvm_oct_napi[i].napi);
 559 }