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 <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 "cvmx-helper.h"
  56 #include "cvmx-wqe.h"
  57 #include "cvmx-fau.h"
  58 #include "cvmx-pow.h"
  59 #include "cvmx-pip.h"
  60 #include "cvmx-scratch.h"
  61
  62 #include "cvmx-gmxx-defs.h"
  63
  64 struct cvm_napi_wrapper {
  65         struct napi_struct napi;
  66 } ____cacheline_aligned_in_smp;
  67
  68 static struct cvm_napi_wrapper cvm_oct_napi[NR_CPUS] __cacheline_aligned_in_smp;
  69
  70 struct cvm_oct_core_state {
  71         int baseline_cores;
  72         /*
  73          * The number of additional cores that could be processing
  74          * input packtes.
  75          */
  76         atomic_t available_cores;
  77         cpumask_t cpu_state;
  78 } ____cacheline_aligned_in_smp;
  79
  80 static struct cvm_oct_core_state core_state __cacheline_aligned_in_smp;
  81
  82 static void cvm_oct_enable_napi(void *_)
  83 {
  84         int cpu = smp_processor_id();
  85         napi_schedule(&cvm_oct_napi[cpu].napi);
  86 }
  87
  88 static void cvm_oct_enable_one_cpu(void)
  89 {
  90         int v;
  91         int cpu;
  92
  93         /* Check to see if more CPUs are available for receive processing... */
  94         v = atomic_sub_if_positive(1, &core_state.available_cores);
  95         if (v < 0)
  96                 return;
  97
  98         /* ... if a CPU is available, Turn on NAPI polling for that CPU.  */
  99         for_each_online_cpu(cpu) {
 100                 if (!cpu_test_and_set(cpu, core_state.cpu_state)) {
 101                         v = smp_call_function_single(cpu, cvm_oct_enable_napi,
 102                                                      NULL, 0);
 103                         if (v)
 104                                 panic("Can't enable NAPI.");
 105                         break;
 106                 }
 107         }
 108 }
 109
 110 static void cvm_oct_no_more_work(void)
 111 {
 112         int cpu = smp_processor_id();
 113
 114         /*
 115          * CPU zero is special.  It always has the irq enabled when
 116          * waiting for incoming packets.
 117          */
 118         if (cpu == 0) {
 119                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 120                 return;
 121         }
 122
 123         cpu_clear(cpu, core_state.cpu_state);
 124         atomic_add(1, &core_state.available_cores);
 125 }
 126
 127 /**
 128  * cvm_oct_do_interrupt - interrupt handler.
 129  *
 130  * The interrupt occurs whenever the POW has packets in our group.
 131  *
 132  */
 133 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
 134 {
 135         /* Disable the IRQ and start napi_poll. */
 136         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 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
 306                 skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
 307                 if (likely(skb_in_hw)) {
 308                         skb = *pskb;
 309                         prefetch(&skb->head);
 310                         prefetch(&skb->len);
 311                 }
 312                 prefetch(cvm_oct_device[work->ipprt]);
 313
 314                 /* Immediately throw away all packets with receive errors */
 315                 if (unlikely(work->word2.snoip.rcv_error)) {
 316                         if (cvm_oct_check_rcv_error(work))
 317                                 continue;
 318                 }
 319
 320                 /*
 321                  * We can only use the zero copy path if skbuffs are
 322                  * in the FPA pool and the packet fits in a single
 323                  * buffer.
 324                  */
 325                 if (likely(skb_in_hw)) {
 326                         skb->data = skb->head + work->packet_ptr.s.addr - cvmx_ptr_to_phys(skb->head);
 327                         prefetch(skb->data);
 328                         skb->len = work->len;
 329                         skb_set_tail_pointer(skb, skb->len);
 330                         packet_not_copied = 1;
 331                 } else {
 332                         /*
 333                          * We have to copy the packet. First allocate
 334                          * an skbuff for it.
 335                          */
 336                         skb = dev_alloc_skb(work->len);
 337                         if (!skb) {
 338                                 printk_ratelimited("Port %d failed to allocate "
 339                                                    "skbuff, packet dropped\n",
 340                                                    work->ipprt);
 341                                 cvm_oct_free_work(work);
 342                                 continue;
 343                         }
 344
 345                         /*
 346                          * Check if we've received a packet that was
 347                          * entirely stored in the work entry.
 348                          */
 349                         if (unlikely(work->word2.s.bufs == 0)) {
 350                                 uint8_t *ptr = work->packet_data;
 351
 352                                 if (likely(!work->word2.s.not_IP)) {
 353                                         /*
 354                                          * The beginning of the packet
 355                                          * moves for IP packets.
 356                                          */
 357                                         if (work->word2.s.is_v6)
 358                                                 ptr += 2;
 359                                         else
 360                                                 ptr += 6;
 361                                 }
 362                                 memcpy(skb_put(skb, work->len), ptr, work->len);
 363                                 /* No packet buffers to free */
 364                         } else {
 365                                 int segments = work->word2.s.bufs;
 366                                 union cvmx_buf_ptr segment_ptr = work->packet_ptr;
 367                                 int len = work->len;
 368
 369                                 while (segments--) {
 370                                         union cvmx_buf_ptr next_ptr =
 371                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 372
 373                         /*
 374                          * Octeon Errata PKI-100: The segment size is
 375                          * wrong. Until it is fixed, calculate the
 376                          * segment size based on the packet pool
 377                          * buffer size. When it is fixed, the
 378                          * following line should be replaced with this
 379                          * one: int segment_size =
 380                          * segment_ptr.s.size;
 381                          */
 382                                         int segment_size = CVMX_FPA_PACKET_POOL_SIZE -
 383                                                 (segment_ptr.s.addr - (((segment_ptr.s.addr >> 7) - segment_ptr.s.back) << 7));
 384                                         /*
 385                                          * Don't copy more than what
 386                                          * is left in the packet.
 387                                          */
 388                                         if (segment_size > len)
 389                                                 segment_size = len;
 390                                         /* Copy the data into the packet */
 391                                         memcpy(skb_put(skb, segment_size),
 392                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 393                                                segment_size);
 394                                         len -= segment_size;
 395                                         segment_ptr = next_ptr;
 396                                 }
 397                         }
 398                         packet_not_copied = 0;
 399                 }
 400
 401                 if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
 402                            cvm_oct_device[work->ipprt])) {
 403                         struct net_device *dev = cvm_oct_device[work->ipprt];
 404                         struct octeon_ethernet *priv = netdev_priv(dev);
 405
 406                         /*
 407                          * Only accept packets for devices that are
 408                          * currently up.
 409                          */
 410                         if (likely(dev->flags & IFF_UP)) {
 411                                 skb->protocol = eth_type_trans(skb, dev);
 412                                 skb->dev = dev;
 413
 414                                 if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc || work->word2.s.L4_error))
 415                                         skb->ip_summed = CHECKSUM_NONE;
 416                                 else
 417                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 418
 419                                 /* Increment RX stats for virtual ports */
 420                                 if (work->ipprt >= CVMX_PIP_NUM_INPUT_PORTS) {
 421 #ifdef CONFIG_64BIT
 422                                         atomic64_add(1, (atomic64_t *)&priv->stats.rx_packets);
 423                                         atomic64_add(skb->len, (atomic64_t *)&priv->stats.rx_bytes);
 424 #else
 425                                         atomic_add(1, (atomic_t *)&priv->stats.rx_packets);
 426                                         atomic_add(skb->len, (atomic_t *)&priv->stats.rx_bytes);
 427 #endif
 428                                 }
 429                                 netif_receive_skb(skb);
 430                                 rx_count++;
 431                         } else {
 432                                 /* Drop any packet received for a device that isn't up */
 433                                 /*
 434                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 435                                            dev->name);
 436                                 */
 437 #ifdef CONFIG_64BIT
 438                                 atomic64_add(1, (atomic64_t *)&priv->stats.rx_dropped);
 439 #else
 440                                 atomic_add(1, (atomic_t *)&priv->stats.rx_dropped);
 441 #endif
 442                                 dev_kfree_skb_irq(skb);
 443                         }
 444                 } else {
 445                         /*
 446                          * Drop any packet received for a device that
 447                          * doesn't exist.
 448                          */
 449                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 450                                    work->ipprt);
 451                         dev_kfree_skb_irq(skb);
 452                 }
 453                 /*
 454                  * Check to see if the skbuff and work share the same
 455                  * packet buffer.
 456                  */
 457                 if (USE_SKBUFFS_IN_HW && likely(packet_not_copied)) {
 458                         /*
 459                          * This buffer needs to be replaced, increment
 460                          * the number of buffers we need to free by
 461                          * one.
 462                          */
 463                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 464                                               1);
 465
 466                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL,
 467                                       DONT_WRITEBACK(1));
 468                 } else {
 469                         cvm_oct_free_work(work);
 470                 }
 471         }
 472         /* Restore the original POW group mask */
 473         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 474         if (USE_ASYNC_IOBDMA) {
 475                 /* Restore the scratch area */
 476                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 477         }
 478         cvm_oct_rx_refill_pool(0);
 479
 480         if (rx_count < budget && napi != NULL) {
 481                 /* No more work */
 482                 napi_complete(napi);
 483                 cvm_oct_no_more_work();
 484         }
 485         return rx_count;
 486 }
 487
 488 #ifdef CONFIG_NET_POLL_CONTROLLER
 489 /**
 490  * cvm_oct_poll_controller - poll for receive packets
 491  * device.
 492  *
 493  * @dev:    Device to poll. Unused
 494  */
 495 void cvm_oct_poll_controller(struct net_device *dev)
 496 {
 497         cvm_oct_napi_poll(NULL, 16);
 498 }
 499 #endif
 500
 501 void cvm_oct_rx_initialize(void)
 502 {
 503         int i;
 504         struct net_device *dev_for_napi = NULL;
 505         union cvmx_pow_wq_int_thrx int_thr;
 506         union cvmx_pow_wq_int_pc int_pc;
 507
 508         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 509                 if (cvm_oct_device[i]) {
 510                         dev_for_napi = cvm_oct_device[i];
 511                         break;
 512                 }
 513         }
 514
 515         if (NULL == dev_for_napi)
 516                 panic("No net_devices were allocated.");
 517
 518         if (max_rx_cpus > 1  && max_rx_cpus < num_online_cpus())
 519                 atomic_set(&core_state.available_cores, max_rx_cpus);
 520         else
 521                 atomic_set(&core_state.available_cores, num_online_cpus());
 522         core_state.baseline_cores = atomic_read(&core_state.available_cores);
 523
 524         core_state.cpu_state = CPU_MASK_NONE;
 525         for_each_possible_cpu(i) {
 526                 netif_napi_add(dev_for_napi, &cvm_oct_napi[i].napi,
 527                                cvm_oct_napi_poll, rx_napi_weight);
 528                 napi_enable(&cvm_oct_napi[i].napi);
 529         }
 530         /* Register an IRQ hander for to receive POW interrupts */
 531         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 532                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 533
 534         if (i)
 535                 panic("Could not acquire Ethernet IRQ %d\n",
 536                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 537
 538         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 539
 540         int_thr.u64 = 0;
 541         int_thr.s.tc_en = 1;
 542         int_thr.s.tc_thr = 1;
 543         /* Enable POW interrupt when our port has at least one packet */
 544         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
 545
 546         int_pc.u64 = 0;
 547         int_pc.s.pc_thr = 5;
 548         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 549
 550
 551         /* Scheduld NAPI now.  This will indirectly enable interrupts. */
 552         cvm_oct_enable_one_cpu();
 553 }
 554
 555 void cvm_oct_rx_shutdown(void)
 556 {
 557         int i;
 558         /* Shutdown all of the NAPIs */
 559         for_each_possible_cpu(i)
 560                 netif_napi_del(&cvm_oct_napi[i].napi);
 561 }