drivers/staging/octeon/ethernet-rx.c

   1 /**********************************************************************
   2  * Author: Cavium Networks
   3  *
   4  * Contact: support@caviumnetworks.com
   5  * This file is part of the OCTEON SDK
   6  *
   7  * Copyright (c) 2003-2010 Cavium Networks
   8  *
   9  * This file is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License, Version 2, as
  11  * published by the Free Software Foundation.
  12  *
  13  * This file is distributed in the hope that it will be useful, but
  14  * AS-IS and WITHOUT ANY WARRANTY; without even the implied warranty
  15  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE, TITLE, or
  16  * NONINFRINGEMENT.  See the GNU General Public License for more
  17  * details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this file; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
  22  * or visit http://www.gnu.org/licenses/.
  23  *
  24  * This file may also be available under a different license from Cavium.
  25  * Contact Cavium Networks for more information
  26 **********************************************************************/
  27 #include <linux/module.h>
  28 #include <linux/kernel.h>
  29 #include <linux/cache.h>
  30 #include <linux/cpumask.h>
  31 #include <linux/netdevice.h>
  32 #include <linux/etherdevice.h>
  33 #include <linux/ip.h>
  34 #include <linux/string.h>
  35 #include <linux/prefetch.h>
  36 #include <linux/ratelimit.h>
  37 #include <linux/smp.h>
  38 #include <linux/interrupt.h>
  39 #include <net/dst.h>
  40 #ifdef CONFIG_XFRM
  41 #include <linux/xfrm.h>
  42 #include <net/xfrm.h>
  43 #endif /* CONFIG_XFRM */
  44
  45 #include <linux/atomic.h>
  46
  47 #include <asm/octeon/octeon.h>
  48
  49 #include "ethernet-defines.h"
  50 #include "ethernet-mem.h"
  51 #include "ethernet-rx.h"
  52 #include "octeon-ethernet.h"
  53 #include "ethernet-util.h"
  54
  55 #include <asm/octeon/cvmx-helper.h>
  56 #include <asm/octeon/cvmx-wqe.h>
  57 #include <asm/octeon/cvmx-fau.h>
  58 #include <asm/octeon/cvmx-pow.h>
  59 #include <asm/octeon/cvmx-pip.h>
  60 #include <asm/octeon/cvmx-scratch.h>
  61
  62 #include <asm/octeon/cvmx-gmxx-defs.h>
  63
  64 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 packets.
  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 int cvm_irq_cpu;
  83
  84 static void cvm_oct_enable_napi(void *_)
  85 {
  86         int cpu = smp_processor_id();
  87         napi_schedule(&cvm_oct_napi[cpu].napi);
  88 }
  89
  90 static void cvm_oct_enable_one_cpu(void)
  91 {
  92         int v;
  93         int cpu;
  94
  95         /* Check to see if more CPUs are available for receive processing... */
  96         v = atomic_sub_if_positive(1, &core_state.available_cores);
  97         if (v < 0)
  98                 return;
  99
 100         /* ... if a CPU is available, Turn on NAPI polling for that CPU.  */
 101         for_each_online_cpu(cpu) {
 102                 if (!cpu_test_and_set(cpu, core_state.cpu_state)) {
 103                         v = smp_call_function_single(cpu, cvm_oct_enable_napi,
 104                                                      NULL, 0);
 105                         if (v)
 106                                 panic("Can't enable NAPI.");
 107                         break;
 108                 }
 109         }
 110 }
 111
 112 static void cvm_oct_no_more_work(void)
 113 {
 114         int cpu = smp_processor_id();
 115
 116         if (cpu == cvm_irq_cpu) {
 117                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 118                 return;
 119         }
 120
 121         cpu_clear(cpu, core_state.cpu_state);
 122         atomic_add(1, &core_state.available_cores);
 123 }
 124
 125 /**
 126  * cvm_oct_do_interrupt - interrupt handler.
 127  *
 128  * The interrupt occurs whenever the POW 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_irq_cpu = smp_processor_id();
 136         cvm_oct_enable_napi(NULL);
 137
 138         return IRQ_HANDLED;
 139 }
 140
 141 /**
 142  * cvm_oct_check_rcv_error - process receive errors
 143  * @work: Work queue entry pointing to the packet.
 144  *
 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                                   printk_ratelimited("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                                   printk_ratelimited("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                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 207                                                    work->ipprt);
 208                                 /*
 209                                    cvmx_helper_dump_packet(work);
 210                                  */
 211                                 cvm_oct_free_work(work);
 212                                 return 1;
 213                         }
 214                 }
 215         } else {
 216                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 217                                    work->ipprt, work->word2.snoip.err_code);
 218                 cvm_oct_free_work(work);
 219                 return 1;
 220         }
 221
 222         return 0;
 223 }
 224
 225 /**
 226  * cvm_oct_napi_poll - the NAPI poll function.
 227  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 228  * @budget: Maximum number of packets to receive.
 229  *
 230  * Returns the number of packets processed.
 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                 rx_count++;
 304
 305                 skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
 306                 if (likely(skb_in_hw)) {
 307                         skb = *pskb;
 308                         prefetch(&skb->head);
 309                         prefetch(&skb->len);
 310                 }
 311                 prefetch(cvm_oct_device[work->ipprt]);
 312
 313                 /* Immediately throw away all packets with receive errors */
 314                 if (unlikely(work->word2.snoip.rcv_error)) {
 315                         if (cvm_oct_check_rcv_error(work))
 316                                 continue;
 317                 }
 318
 319                 /*
 320                  * We can only use the zero copy path if skbuffs are
 321                  * in the FPA pool and the packet fits in a single
 322                  * buffer.
 323                  */
 324                 if (likely(skb_in_hw)) {
 325                         skb->data = skb->head + work->packet_ptr.s.addr - cvmx_ptr_to_phys(skb->head);
 326                         prefetch(skb->data);
 327                         skb->len = work->len;
 328                         skb_set_tail_pointer(skb, skb->len);
 329                         packet_not_copied = 1;
 330                 } else {
 331                         /*
 332                          * We have to copy the packet. First allocate
 333                          * an skbuff for it.
 334                          */
 335                         skb = dev_alloc_skb(work->len);
 336                         if (!skb) {
 337                                 cvm_oct_free_work(work);
 338                                 continue;
 339                         }
 340
 341                         /*
 342                          * Check if we've received a packet that was
 343                          * entirely stored in the work entry.
 344                          */
 345                         if (unlikely(work->word2.s.bufs == 0)) {
 346                                 uint8_t *ptr = work->packet_data;
 347
 348                                 if (likely(!work->word2.s.not_IP)) {
 349                                         /*
 350                                          * The beginning of the packet
 351                                          * moves for IP packets.
 352                                          */
 353                                         if (work->word2.s.is_v6)
 354                                                 ptr += 2;
 355                                         else
 356                                                 ptr += 6;
 357                                 }
 358                                 memcpy(skb_put(skb, work->len), ptr, work->len);
 359                                 /* No packet buffers to free */
 360                         } else {
 361                                 int segments = work->word2.s.bufs;
 362                                 union cvmx_buf_ptr segment_ptr = work->packet_ptr;
 363                                 int len = work->len;
 364
 365                                 while (segments--) {
 366                                         union cvmx_buf_ptr next_ptr =
 367                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 368
 369                         /*
 370                          * Octeon Errata PKI-100: The segment size is
 371                          * wrong. Until it is fixed, calculate the
 372                          * segment size based on the packet pool
 373                          * buffer size. When it is fixed, the
 374                          * following line should be replaced with this
 375                          * one: int segment_size =
 376                          * segment_ptr.s.size;
 377                          */
 378                                         int segment_size = CVMX_FPA_PACKET_POOL_SIZE -
 379                                                 (segment_ptr.s.addr - (((segment_ptr.s.addr >> 7) - segment_ptr.s.back) << 7));
 380                                         /*
 381                                          * Don't copy more than what
 382                                          * is left in the packet.
 383                                          */
 384                                         if (segment_size > len)
 385                                                 segment_size = len;
 386                                         /* Copy the data into the packet */
 387                                         memcpy(skb_put(skb, segment_size),
 388                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 389                                                segment_size);
 390                                         len -= segment_size;
 391                                         segment_ptr = next_ptr;
 392                                 }
 393                         }
 394                         packet_not_copied = 0;
 395                 }
 396
 397                 if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
 398                            cvm_oct_device[work->ipprt])) {
 399                         struct net_device *dev = cvm_oct_device[work->ipprt];
 400                         struct octeon_ethernet *priv = netdev_priv(dev);
 401
 402                         /*
 403                          * Only accept packets for devices that are
 404                          * currently up.
 405                          */
 406                         if (likely(dev->flags & IFF_UP)) {
 407                                 skb->protocol = eth_type_trans(skb, dev);
 408                                 skb->dev = dev;
 409
 410                                 if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc ||
 411                                         work->word2.s.L4_error || !work->word2.s.tcp_or_udp))
 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                         } else {
 428                                 /* Drop any packet received for a device that isn't up */
 429                                 /*
 430                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 431                                            dev->name);
 432                                 */
 433 #ifdef CONFIG_64BIT
 434                                 atomic64_add(1, (atomic64_t *)&priv->stats.rx_dropped);
 435 #else
 436                                 atomic_add(1, (atomic_t *)&priv->stats.rx_dropped);
 437 #endif
 438                                 dev_kfree_skb_irq(skb);
 439                         }
 440                 } else {
 441                         /*
 442                          * Drop any packet received for a device that
 443                          * doesn't exist.
 444                          */
 445                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 446                                    work->ipprt);
 447                         dev_kfree_skb_irq(skb);
 448                 }
 449                 /*
 450                  * Check to see if the skbuff and work share the same
 451                  * packet buffer.
 452                  */
 453                 if (USE_SKBUFFS_IN_HW && likely(packet_not_copied)) {
 454                         /*
 455                          * This buffer needs to be replaced, increment
 456                          * the number of buffers we need to free by
 457                          * one.
 458                          */
 459                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 460                                               1);
 461
 462                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL,
 463                                       DONT_WRITEBACK(1));
 464                 } else {
 465                         cvm_oct_free_work(work);
 466                 }
 467         }
 468         /* Restore the original POW group mask */
 469         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 470         if (USE_ASYNC_IOBDMA) {
 471                 /* Restore the scratch area */
 472                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 473         }
 474         cvm_oct_rx_refill_pool(0);
 475
 476         if (rx_count < budget && napi != NULL) {
 477                 /* No more work */
 478                 napi_complete(napi);
 479                 cvm_oct_no_more_work();
 480         }
 481         return rx_count;
 482 }
 483
 484 #ifdef CONFIG_NET_POLL_CONTROLLER
 485 /**
 486  * cvm_oct_poll_controller - poll for receive packets
 487  * device.
 488  *
 489  * @dev:    Device to poll. Unused
 490  */
 491 void cvm_oct_poll_controller(struct net_device *dev)
 492 {
 493         cvm_oct_napi_poll(NULL, 16);
 494 }
 495 #endif
 496
 497 void cvm_oct_rx_initialize(void)
 498 {
 499         int i;
 500         struct net_device *dev_for_napi = NULL;
 501         union cvmx_pow_wq_int_thrx int_thr;
 502         union cvmx_pow_wq_int_pc int_pc;
 503
 504         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 505                 if (cvm_oct_device[i]) {
 506                         dev_for_napi = cvm_oct_device[i];
 507                         break;
 508                 }
 509         }
 510
 511         if (NULL == dev_for_napi)
 512                 panic("No net_devices were allocated.");
 513
 514         if (max_rx_cpus >= 1 && max_rx_cpus < num_online_cpus())
 515                 atomic_set(&core_state.available_cores, max_rx_cpus);
 516         else
 517                 atomic_set(&core_state.available_cores, num_online_cpus());
 518         core_state.baseline_cores = atomic_read(&core_state.available_cores);
 519
 520         core_state.cpu_state = CPU_MASK_NONE;
 521         for_each_possible_cpu(i) {
 522                 netif_napi_add(dev_for_napi, &cvm_oct_napi[i].napi,
 523                                cvm_oct_napi_poll, rx_napi_weight);
 524                 napi_enable(&cvm_oct_napi[i].napi);
 525         }
 526         /* Register an IRQ handler to receive POW interrupts */
 527         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 528                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 529
 530         if (i)
 531                 panic("Could not acquire Ethernet IRQ %d\n",
 532                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 533
 534         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 535
 536         int_thr.u64 = 0;
 537         int_thr.s.tc_en = 1;
 538         int_thr.s.tc_thr = 1;
 539         /* Enable POW interrupt when our port has at least one packet */
 540         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
 541
 542         int_pc.u64 = 0;
 543         int_pc.s.pc_thr = 5;
 544         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 545
 546
 547         /* Scheduld NAPI now.  This will indirectly enable interrupts. */
 548         cvm_oct_enable_one_cpu();
 549 }
 550
 551 void cvm_oct_rx_shutdown(void)
 552 {
 553         int i;
 554         /* Shutdown all of the NAPIs */
 555         for_each_possible_cpu(i)
 556                 netif_napi_del(&cvm_oct_napi[i].napi);
 557 }