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