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 packtes.
  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 void cvm_oct_enable_napi(void *_)
  84 {
  85         int cpu = smp_processor_id();
  86         napi_schedule(&cvm_oct_napi[cpu].napi);
  87 }
  88
  89 static void cvm_oct_enable_one_cpu(void)
  90 {
  91         int v;
  92         int cpu;
  93
  94         /* Check to see if more CPUs are available for receive processing... */
  95         v = atomic_sub_if_positive(1, &core_state.available_cores);
  96         if (v < 0)
  97                 return;
  98
  99         /* ... if a CPU is available, Turn on NAPI polling for that CPU.  */
 100         for_each_online_cpu(cpu) {
 101                 if (!cpu_test_and_set(cpu, core_state.cpu_state)) {
 102                         v = smp_call_function_single(cpu, cvm_oct_enable_napi,
 103                                                      NULL, 0);
 104                         if (v)
 105                                 panic("Can't enable NAPI.");
 106                         break;
 107                 }
 108         }
 109 }
 110
 111 static void cvm_oct_no_more_work(void)
 112 {
 113         int cpu = smp_processor_id();
 114
 115         /*
 116          * CPU zero is special.  It always has the irq enabled when
 117          * waiting for incoming packets.
 118          */
 119         if (cpu == 0) {
 120                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 121                 return;
 122         }
 123
 124         cpu_clear(cpu, core_state.cpu_state);
 125         atomic_add(1, &core_state.available_cores);
 126 }
 127
 128 /**
 129  * cvm_oct_do_interrupt - interrupt handler.
 130  *
 131  * The interrupt occurs whenever the POW has packets in our group.
 132  *
 133  */
 134 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
 135 {
 136         /* Disable the IRQ and start napi_poll. */
 137         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 138         cvm_oct_enable_napi(NULL);
 139
 140         return IRQ_HANDLED;
 141 }
 142
 143 /**
 144  * cvm_oct_check_rcv_error - process receive errors
 145  * @work: Work queue entry pointing to the packet.
 146  *
 147  * Returns Non-zero if the packet can be dropped, zero otherwise.
 148  */
 149 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
 150 {
 151         if ((work->word2.snoip.err_code == 10) && (work->len <= 64)) {
 152                 /*
 153                  * Ignore length errors on min size packets. Some
 154                  * equipment incorrectly pads packets to 64+4FCS
 155                  * instead of 60+4FCS.  Note these packets still get
 156                  * counted as frame errors.
 157                  */
 158         } else
 159             if (USE_10MBPS_PREAMBLE_WORKAROUND
 160                 && ((work->word2.snoip.err_code == 5)
 161                     || (work->word2.snoip.err_code == 7))) {
 162
 163                 /*
 164                  * We received a packet with either an alignment error
 165                  * or a FCS error. This may be signalling that we are
 166                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
 167                  * off. If this is the case we need to parse the
 168                  * packet to determine if we can remove a non spec
 169                  * preamble and generate a correct packet.
 170                  */
 171                 int interface = cvmx_helper_get_interface_num(work->ipprt);
 172                 int index = cvmx_helper_get_interface_index_num(work->ipprt);
 173                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
 174                 gmxx_rxx_frm_ctl.u64 =
 175                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 176                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 177
 178                         uint8_t *ptr =
 179                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 180                         int i = 0;
 181
 182                         while (i < work->len - 1) {
 183                                 if (*ptr != 0x55)
 184                                         break;
 185                                 ptr++;
 186                                 i++;
 187                         }
 188
 189                         if (*ptr == 0xd5) {
 190                                 /*
 191                                   printk_ratelimited("Port %d received 0xd5 preamble\n", work->ipprt);
 192                                  */
 193                                 work->packet_ptr.s.addr += i + 1;
 194                                 work->len -= i + 5;
 195                         } else if ((*ptr & 0xf) == 0xd) {
 196                                 /*
 197                                   printk_ratelimited("Port %d received 0x?d preamble\n", work->ipprt);
 198                                  */
 199                                 work->packet_ptr.s.addr += i;
 200                                 work->len -= i + 4;
 201                                 for (i = 0; i < work->len; i++) {
 202                                         *ptr =
 203                                             ((*ptr & 0xf0) >> 4) |
 204                                             ((*(ptr + 1) & 0xf) << 4);
 205                                         ptr++;
 206                                 }
 207                         } else {
 208                                 printk_ratelimited("Port %d unknown preamble, packet "
 209                                                    "dropped\n",
 210                                                    work->ipprt);
 211                                 /*
 212                                    cvmx_helper_dump_packet(work);
 213                                  */
 214                                 cvm_oct_free_work(work);
 215                                 return 1;
 216                         }
 217                 }
 218         } else {
 219                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 220                                    work->ipprt, work->word2.snoip.err_code);
 221                 cvm_oct_free_work(work);
 222                 return 1;
 223         }
 224
 225         return 0;
 226 }
 227
 228 /**
 229  * cvm_oct_napi_poll - the NAPI poll function.
 230  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 231  * @budget: Maximum number of packets to receive.
 232  *
 233  * Returns the number of packets processed.
 234  */
 235 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 236 {
 237         const int       coreid = cvmx_get_core_num();
 238         uint64_t        old_group_mask;
 239         uint64_t        old_scratch;
 240         int             rx_count = 0;
 241         int             did_work_request = 0;
 242         int             packet_not_copied;
 243
 244         /* Prefetch cvm_oct_device since we know we need it soon */
 245         prefetch(cvm_oct_device);
 246
 247         if (USE_ASYNC_IOBDMA) {
 248                 /* Save scratch in case userspace is using it */
 249                 CVMX_SYNCIOBDMA;
 250                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 251         }
 252
 253         /* Only allow work for our group (and preserve priorities) */
 254         old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 255         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 256                        (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 257
 258         if (USE_ASYNC_IOBDMA) {
 259                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 260                 did_work_request = 1;
 261         }
 262
 263         while (rx_count < budget) {
 264                 struct sk_buff *skb = NULL;
 265                 struct sk_buff **pskb = NULL;
 266                 int skb_in_hw;
 267                 cvmx_wqe_t *work;
 268
 269                 if (USE_ASYNC_IOBDMA && did_work_request)
 270                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 271                 else
 272                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 273
 274                 prefetch(work);
 275                 did_work_request = 0;
 276                 if (work == NULL) {
 277                         union cvmx_pow_wq_int wq_int;
 278                         wq_int.u64 = 0;
 279                         wq_int.s.iq_dis = 1 << pow_receive_group;
 280                         wq_int.s.wq_int = 1 << pow_receive_group;
 281                         cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 282                         break;
 283                 }
 284                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) - sizeof(void *));
 285                 prefetch(pskb);
 286
 287                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 288                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 289                         did_work_request = 1;
 290                 }
 291
 292                 if (rx_count == 0) {
 293                         /*
 294                          * First time through, see if there is enough
 295                          * work waiting to merit waking another
 296                          * CPU.
 297                          */
 298                         union cvmx_pow_wq_int_cntx counts;
 299                         int backlog;
 300                         int cores_in_use = core_state.baseline_cores - atomic_read(&core_state.available_cores);
 301                         counts.u64 = cvmx_read_csr(CVMX_POW_WQ_INT_CNTX(pow_receive_group));
 302                         backlog = counts.s.iq_cnt + counts.s.ds_cnt;
 303                         if (backlog > budget * cores_in_use && napi != NULL)
 304                                 cvm_oct_enable_one_cpu();
 305                 }
 306                 rx_count++;
 307
 308                 skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
 309                 if (likely(skb_in_hw)) {
 310                         skb = *pskb;
 311                         prefetch(&skb->head);
 312                         prefetch(&skb->len);
 313                 }
 314                 prefetch(cvm_oct_device[work->ipprt]);
 315
 316                 /* Immediately throw away all packets with receive errors */
 317                 if (unlikely(work->word2.snoip.rcv_error)) {
 318                         if (cvm_oct_check_rcv_error(work))
 319                                 continue;
 320                 }
 321
 322                 /*
 323                  * We can only use the zero copy path if skbuffs are
 324                  * in the FPA pool and the packet fits in a single
 325                  * buffer.
 326                  */
 327                 if (likely(skb_in_hw)) {
 328                         skb->data = skb->head + work->packet_ptr.s.addr - cvmx_ptr_to_phys(skb->head);
 329                         prefetch(skb->data);
 330                         skb->len = work->len;
 331                         skb_set_tail_pointer(skb, skb->len);
 332                         packet_not_copied = 1;
 333                 } else {
 334                         /*
 335                          * We have to copy the packet. First allocate
 336                          * an skbuff for it.
 337                          */
 338                         skb = dev_alloc_skb(work->len);
 339                         if (!skb) {
 340                                 cvm_oct_free_work(work);
 341                                 continue;
 342                         }
 343
 344                         /*
 345                          * Check if we've received a packet that was
 346                          * entirely stored in the work entry.
 347                          */
 348                         if (unlikely(work->word2.s.bufs == 0)) {
 349                                 uint8_t *ptr = work->packet_data;
 350
 351                                 if (likely(!work->word2.s.not_IP)) {
 352                                         /*
 353                                          * The beginning of the packet
 354                                          * moves for IP packets.
 355                                          */
 356                                         if (work->word2.s.is_v6)
 357                                                 ptr += 2;
 358                                         else
 359                                                 ptr += 6;
 360                                 }
 361                                 memcpy(skb_put(skb, work->len), ptr, work->len);
 362                                 /* No packet buffers to free */
 363                         } else {
 364                                 int segments = work->word2.s.bufs;
 365                                 union cvmx_buf_ptr segment_ptr = work->packet_ptr;
 366                                 int len = work->len;
 367
 368                                 while (segments--) {
 369                                         union cvmx_buf_ptr next_ptr =
 370                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 371
 372                         /*
 373                          * Octeon Errata PKI-100: The segment size is
 374                          * wrong. Until it is fixed, calculate the
 375                          * segment size based on the packet pool
 376                          * buffer size. When it is fixed, the
 377                          * following line should be replaced with this
 378                          * one: int segment_size =
 379                          * segment_ptr.s.size;
 380                          */
 381                                         int segment_size = CVMX_FPA_PACKET_POOL_SIZE -
 382                                                 (segment_ptr.s.addr - (((segment_ptr.s.addr >> 7) - segment_ptr.s.back) << 7));
 383                                         /*
 384                                          * Don't copy more than what
 385                                          * is left in the packet.
 386                                          */
 387                                         if (segment_size > len)
 388                                                 segment_size = len;
 389                                         /* Copy the data into the packet */
 390                                         memcpy(skb_put(skb, segment_size),
 391                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 392                                                segment_size);
 393                                         len -= segment_size;
 394                                         segment_ptr = next_ptr;
 395                                 }
 396                         }
 397                         packet_not_copied = 0;
 398                 }
 399
 400                 if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
 401                            cvm_oct_device[work->ipprt])) {
 402                         struct net_device *dev = cvm_oct_device[work->ipprt];
 403                         struct octeon_ethernet *priv = netdev_priv(dev);
 404
 405                         /*
 406                          * Only accept packets for devices that are
 407                          * currently up.
 408                          */
 409                         if (likely(dev->flags & IFF_UP)) {
 410                                 skb->protocol = eth_type_trans(skb, dev);
 411                                 skb->dev = dev;
 412
 413                                 if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc ||
 414                                         work->word2.s.L4_error || !work->word2.s.tcp_or_udp))
 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                         } else {
 431                                 /* Drop any packet received for a device that isn't up */
 432                                 /*
 433                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 434                                            dev->name);
 435                                 */
 436 #ifdef CONFIG_64BIT
 437                                 atomic64_add(1, (atomic64_t *)&priv->stats.rx_dropped);
 438 #else
 439                                 atomic_add(1, (atomic_t *)&priv->stats.rx_dropped);
 440 #endif
 441                                 dev_kfree_skb_irq(skb);
 442                         }
 443                 } else {
 444                         /*
 445                          * Drop any packet received for a device that
 446                          * doesn't exist.
 447                          */
 448                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 449                                    work->ipprt);
 450                         dev_kfree_skb_irq(skb);
 451                 }
 452                 /*
 453                  * Check to see if the skbuff and work share the same
 454                  * packet buffer.
 455                  */
 456                 if (USE_SKBUFFS_IN_HW && likely(packet_not_copied)) {
 457                         /*
 458                          * This buffer needs to be replaced, increment
 459                          * the number of buffers we need to free by
 460                          * one.
 461                          */
 462                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 463                                               1);
 464
 465                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL,
 466                                       DONT_WRITEBACK(1));
 467                 } else {
 468                         cvm_oct_free_work(work);
 469                 }
 470         }
 471         /* Restore the original POW group mask */
 472         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 473         if (USE_ASYNC_IOBDMA) {
 474                 /* Restore the scratch area */
 475                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 476         }
 477         cvm_oct_rx_refill_pool(0);
 478
 479         if (rx_count < budget && napi != NULL) {
 480                 /* No more work */
 481                 napi_complete(napi);
 482                 cvm_oct_no_more_work();
 483         }
 484         return rx_count;
 485 }
 486
 487 #ifdef CONFIG_NET_POLL_CONTROLLER
 488 /**
 489  * cvm_oct_poll_controller - poll for receive packets
 490  * device.
 491  *
 492  * @dev:    Device to poll. Unused
 493  */
 494 void cvm_oct_poll_controller(struct net_device *dev)
 495 {
 496         cvm_oct_napi_poll(NULL, 16);
 497 }
 498 #endif
 499
 500 void cvm_oct_rx_initialize(void)
 501 {
 502         int i;
 503         struct net_device *dev_for_napi = NULL;
 504         union cvmx_pow_wq_int_thrx int_thr;
 505         union cvmx_pow_wq_int_pc int_pc;
 506
 507         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 508                 if (cvm_oct_device[i]) {
 509                         dev_for_napi = cvm_oct_device[i];
 510                         break;
 511                 }
 512         }
 513
 514         if (NULL == dev_for_napi)
 515                 panic("No net_devices were allocated.");
 516
 517         if (max_rx_cpus > 1  && max_rx_cpus < num_online_cpus())
 518                 atomic_set(&core_state.available_cores, max_rx_cpus);
 519         else
 520                 atomic_set(&core_state.available_cores, num_online_cpus());
 521         core_state.baseline_cores = atomic_read(&core_state.available_cores);
 522
 523         core_state.cpu_state = CPU_MASK_NONE;
 524         for_each_possible_cpu(i) {
 525                 netif_napi_add(dev_for_napi, &cvm_oct_napi[i].napi,
 526                                cvm_oct_napi_poll, rx_napi_weight);
 527                 napi_enable(&cvm_oct_napi[i].napi);
 528         }
 529         /* Register an IRQ hander for to receive POW interrupts */
 530         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 531                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 532
 533         if (i)
 534                 panic("Could not acquire Ethernet IRQ %d\n",
 535                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 536
 537         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 538
 539         int_thr.u64 = 0;
 540         int_thr.s.tc_en = 1;
 541         int_thr.s.tc_thr = 1;
 542         /* Enable POW interrupt when our port has at least one packet */
 543         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
 544
 545         int_pc.u64 = 0;
 546         int_pc.s.pc_thr = 5;
 547         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 548
 549
 550         /* Scheduld NAPI now.  This will indirectly enable interrupts. */
 551         cvm_oct_enable_one_cpu();
 552 }
 553
 554 void cvm_oct_rx_shutdown(void)
 555 {
 556         int i;
 557         /* Shutdown all of the NAPIs */
 558         for_each_possible_cpu(i)
 559                 netif_napi_del(&cvm_oct_napi[i].napi);
 560 }