drivers/staging/octeon/ethernet-rx.c

   1 /*
   2  * This file is based on code from OCTEON SDK by Cavium Networks.
   3  *
   4  * Copyright (c) 2003-2010 Cavium Networks
   5  *
   6  * This file is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License, Version 2, as
   8  * published by the Free Software Foundation.
   9  */
  10
  11 #include <linux/module.h>
  12 #include <linux/kernel.h>
  13 #include <linux/cache.h>
  14 #include <linux/cpumask.h>
  15 #include <linux/netdevice.h>
  16 #include <linux/etherdevice.h>
  17 #include <linux/ip.h>
  18 #include <linux/string.h>
  19 #include <linux/prefetch.h>
  20 #include <linux/ratelimit.h>
  21 #include <linux/smp.h>
  22 #include <linux/interrupt.h>
  23 #include <net/dst.h>
  24 #ifdef CONFIG_XFRM
  25 #include <linux/xfrm.h>
  26 #include <net/xfrm.h>
  27 #endif /* CONFIG_XFRM */
  28
  29 #include <linux/atomic.h>
  30
  31 #include <asm/octeon/octeon.h>
  32
  33 #include "ethernet-defines.h"
  34 #include "ethernet-mem.h"
  35 #include "ethernet-rx.h"
  36 #include "octeon-ethernet.h"
  37 #include "ethernet-util.h"
  38
  39 #include <asm/octeon/cvmx-helper.h>
  40 #include <asm/octeon/cvmx-wqe.h>
  41 #include <asm/octeon/cvmx-fau.h>
  42 #include <asm/octeon/cvmx-pow.h>
  43 #include <asm/octeon/cvmx-pip.h>
  44 #include <asm/octeon/cvmx-scratch.h>
  45
  46 #include <asm/octeon/cvmx-gmxx-defs.h>
  47
  48 static struct napi_struct cvm_oct_napi;
  49
  50 /**
  51  * cvm_oct_do_interrupt - interrupt handler.
  52  *
  53  * The interrupt occurs whenever the POW has packets in our group.
  54  *
  55  */
  56 static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
  57 {
  58         /* Disable the IRQ and start napi_poll. */
  59         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
  60         napi_schedule(&cvm_oct_napi);
  61
  62         return IRQ_HANDLED;
  63 }
  64
  65 /**
  66  * cvm_oct_check_rcv_error - process receive errors
  67  * @work: Work queue entry pointing to the packet.
  68  *
  69  * Returns Non-zero if the packet can be dropped, zero otherwise.
  70  */
  71 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
  72 {
  73         if ((work->word2.snoip.err_code == 10) && (work->len <= 64)) {
  74                 /*
  75                  * Ignore length errors on min size packets. Some
  76                  * equipment incorrectly pads packets to 64+4FCS
  77                  * instead of 60+4FCS.  Note these packets still get
  78                  * counted as frame errors.
  79                  */
  80         } else if (work->word2.snoip.err_code == 5 ||
  81                    work->word2.snoip.err_code == 7) {
  82                 /*
  83                  * We received a packet with either an alignment error
  84                  * or a FCS error. This may be signalling that we are
  85                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
  86                  * off. If this is the case we need to parse the
  87                  * packet to determine if we can remove a non spec
  88                  * preamble and generate a correct packet.
  89                  */
  90                 int interface = cvmx_helper_get_interface_num(work->ipprt);
  91                 int index = cvmx_helper_get_interface_index_num(work->ipprt);
  92                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
  93
  94                 gmxx_rxx_frm_ctl.u64 =
  95                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
  96                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
  97
  98                         uint8_t *ptr =
  99                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 100                         int i = 0;
 101
 102                         while (i < work->len - 1) {
 103                                 if (*ptr != 0x55)
 104                                         break;
 105                                 ptr++;
 106                                 i++;
 107                         }
 108
 109                         if (*ptr == 0xd5) {
 110                                 /*
 111                                   printk_ratelimited("Port %d received 0xd5 preamble\n",
 112                                           work->ipprt);
 113                                  */
 114                                 work->packet_ptr.s.addr += i + 1;
 115                                 work->len -= i + 5;
 116                         } else if ((*ptr & 0xf) == 0xd) {
 117                                 /*
 118                                   printk_ratelimited("Port %d received 0x?d preamble\n",
 119                                           work->ipprt);
 120                                  */
 121                                 work->packet_ptr.s.addr += i;
 122                                 work->len -= i + 4;
 123                                 for (i = 0; i < work->len; i++) {
 124                                         *ptr =
 125                                             ((*ptr & 0xf0) >> 4) |
 126                                             ((*(ptr + 1) & 0xf) << 4);
 127                                         ptr++;
 128                                 }
 129                         } else {
 130                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 131                                                    work->ipprt);
 132                                 /*
 133                                    cvmx_helper_dump_packet(work);
 134                                  */
 135                                 cvm_oct_free_work(work);
 136                                 return 1;
 137                         }
 138                 }
 139         } else {
 140                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 141                                    work->ipprt, work->word2.snoip.err_code);
 142                 cvm_oct_free_work(work);
 143                 return 1;
 144         }
 145
 146         return 0;
 147 }
 148
 149 /**
 150  * cvm_oct_napi_poll - the NAPI poll function.
 151  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 152  * @budget: Maximum number of packets to receive.
 153  *
 154  * Returns the number of packets processed.
 155  */
 156 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 157 {
 158         const int       coreid = cvmx_get_core_num();
 159         uint64_t        old_group_mask;
 160         uint64_t        old_scratch;
 161         int             rx_count = 0;
 162         int             did_work_request = 0;
 163         int             packet_not_copied;
 164
 165         /* Prefetch cvm_oct_device since we know we need it soon */
 166         prefetch(cvm_oct_device);
 167
 168         if (USE_ASYNC_IOBDMA) {
 169                 /* Save scratch in case userspace is using it */
 170                 CVMX_SYNCIOBDMA;
 171                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 172         }
 173
 174         /* Only allow work for our group (and preserve priorities) */
 175         old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 176         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 177                        (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 178
 179         if (USE_ASYNC_IOBDMA) {
 180                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 181                 did_work_request = 1;
 182         }
 183
 184         while (rx_count < budget) {
 185                 struct sk_buff *skb = NULL;
 186                 struct sk_buff **pskb = NULL;
 187                 int skb_in_hw;
 188                 cvmx_wqe_t *work;
 189
 190                 if (USE_ASYNC_IOBDMA && did_work_request)
 191                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 192                 else
 193                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 194
 195                 prefetch(work);
 196                 did_work_request = 0;
 197                 if (work == NULL) {
 198                         union cvmx_pow_wq_int wq_int;
 199
 200                         wq_int.u64 = 0;
 201                         wq_int.s.iq_dis = 1 << pow_receive_group;
 202                         wq_int.s.wq_int = 1 << pow_receive_group;
 203                         cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 204                         break;
 205                 }
 206                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) -
 207                         sizeof(void *));
 208                 prefetch(pskb);
 209
 210                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 211                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 212                                                             CVMX_POW_NO_WAIT);
 213                         did_work_request = 1;
 214                 }
 215                 rx_count++;
 216
 217                 skb_in_hw = work->word2.s.bufs == 1;
 218                 if (likely(skb_in_hw)) {
 219                         skb = *pskb;
 220                         prefetch(&skb->head);
 221                         prefetch(&skb->len);
 222                 }
 223                 prefetch(cvm_oct_device[work->ipprt]);
 224
 225                 /* Immediately throw away all packets with receive errors */
 226                 if (unlikely(work->word2.snoip.rcv_error)) {
 227                         if (cvm_oct_check_rcv_error(work))
 228                                 continue;
 229                 }
 230
 231                 /*
 232                  * We can only use the zero copy path if skbuffs are
 233                  * in the FPA pool and the packet fits in a single
 234                  * buffer.
 235                  */
 236                 if (likely(skb_in_hw)) {
 237                         skb->data = skb->head + work->packet_ptr.s.addr -
 238                                 cvmx_ptr_to_phys(skb->head);
 239                         prefetch(skb->data);
 240                         skb->len = work->len;
 241                         skb_set_tail_pointer(skb, skb->len);
 242                         packet_not_copied = 1;
 243                 } else {
 244                         /*
 245                          * We have to copy the packet. First allocate
 246                          * an skbuff for it.
 247                          */
 248                         skb = dev_alloc_skb(work->len);
 249                         if (!skb) {
 250                                 cvm_oct_free_work(work);
 251                                 continue;
 252                         }
 253
 254                         /*
 255                          * Check if we've received a packet that was
 256                          * entirely stored in the work entry.
 257                          */
 258                         if (unlikely(work->word2.s.bufs == 0)) {
 259                                 uint8_t *ptr = work->packet_data;
 260
 261                                 if (likely(!work->word2.s.not_IP)) {
 262                                         /*
 263                                          * The beginning of the packet
 264                                          * moves for IP packets.
 265                                          */
 266                                         if (work->word2.s.is_v6)
 267                                                 ptr += 2;
 268                                         else
 269                                                 ptr += 6;
 270                                 }
 271                                 memcpy(skb_put(skb, work->len), ptr, work->len);
 272                                 /* No packet buffers to free */
 273                         } else {
 274                                 int segments = work->word2.s.bufs;
 275                                 union cvmx_buf_ptr segment_ptr =
 276                                     work->packet_ptr;
 277                                 int len = work->len;
 278
 279                                 while (segments--) {
 280                                         union cvmx_buf_ptr next_ptr =
 281                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 282
 283                         /*
 284                          * Octeon Errata PKI-100: The segment size is
 285                          * wrong. Until it is fixed, calculate the
 286                          * segment size based on the packet pool
 287                          * buffer size. When it is fixed, the
 288                          * following line should be replaced with this
 289                          * one: int segment_size =
 290                          * segment_ptr.s.size;
 291                          */
 292                                         int segment_size =
 293                                             CVMX_FPA_PACKET_POOL_SIZE -
 294                                             (segment_ptr.s.addr -
 295                                              (((segment_ptr.s.addr >> 7) -
 296                                                segment_ptr.s.back) << 7));
 297                                         /*
 298                                          * Don't copy more than what
 299                                          * is left in the packet.
 300                                          */
 301                                         if (segment_size > len)
 302                                                 segment_size = len;
 303                                         /* Copy the data into the packet */
 304                                         memcpy(skb_put(skb, segment_size),
 305                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 306                                                segment_size);
 307                                         len -= segment_size;
 308                                         segment_ptr = next_ptr;
 309                                 }
 310                         }
 311                         packet_not_copied = 0;
 312                 }
 313
 314                 if (likely((work->ipprt < TOTAL_NUMBER_OF_PORTS) &&
 315                            cvm_oct_device[work->ipprt])) {
 316                         struct net_device *dev = cvm_oct_device[work->ipprt];
 317                         struct octeon_ethernet *priv = netdev_priv(dev);
 318
 319                         /*
 320                          * Only accept packets for devices that are
 321                          * currently up.
 322                          */
 323                         if (likely(dev->flags & IFF_UP)) {
 324                                 skb->protocol = eth_type_trans(skb, dev);
 325                                 skb->dev = dev;
 326
 327                                 if (unlikely(work->word2.s.not_IP ||
 328                                              work->word2.s.IP_exc ||
 329                                              work->word2.s.L4_error ||
 330                                              !work->word2.s.tcp_or_udp))
 331                                         skb->ip_summed = CHECKSUM_NONE;
 332                                 else
 333                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 334
 335                                 /* Increment RX stats for virtual ports */
 336                                 if (work->ipprt >= CVMX_PIP_NUM_INPUT_PORTS) {
 337 #ifdef CONFIG_64BIT
 338                                         atomic64_add(1,
 339                                                      (atomic64_t *)&priv->stats.rx_packets);
 340                                         atomic64_add(skb->len,
 341                                                      (atomic64_t *)&priv->stats.rx_bytes);
 342 #else
 343                                         atomic_add(1,
 344                                                    (atomic_t *)&priv->stats.rx_packets);
 345                                         atomic_add(skb->len,
 346                                                    (atomic_t *)&priv->stats.rx_bytes);
 347 #endif
 348                                 }
 349                                 netif_receive_skb(skb);
 350                         } else {
 351                                 /* Drop any packet received for a device that isn't up */
 352                                 /*
 353                                   printk_ratelimited("%s: Device not up, packet dropped\n",
 354                                            dev->name);
 355                                 */
 356 #ifdef CONFIG_64BIT
 357                                 atomic64_add(1,
 358                                              (atomic64_t *)&priv->stats.rx_dropped);
 359 #else
 360                                 atomic_add(1,
 361                                            (atomic_t *)&priv->stats.rx_dropped);
 362 #endif
 363                                 dev_kfree_skb_irq(skb);
 364                         }
 365                 } else {
 366                         /*
 367                          * Drop any packet received for a device that
 368                          * doesn't exist.
 369                          */
 370                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 371                                    work->ipprt);
 372                         dev_kfree_skb_irq(skb);
 373                 }
 374                 /*
 375                  * Check to see if the skbuff and work share the same
 376                  * packet buffer.
 377                  */
 378                 if (likely(packet_not_copied)) {
 379                         /*
 380                          * This buffer needs to be replaced, increment
 381                          * the number of buffers we need to free by
 382                          * one.
 383                          */
 384                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 385                                               1);
 386
 387                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 388                 } else {
 389                         cvm_oct_free_work(work);
 390                 }
 391         }
 392         /* Restore the original POW group mask */
 393         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 394         if (USE_ASYNC_IOBDMA) {
 395                 /* Restore the scratch area */
 396                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 397         }
 398         cvm_oct_rx_refill_pool(0);
 399
 400         if (rx_count < budget && napi != NULL) {
 401                 /* No more work */
 402                 napi_complete(napi);
 403                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 404         }
 405         return rx_count;
 406 }
 407
 408 #ifdef CONFIG_NET_POLL_CONTROLLER
 409 /**
 410  * cvm_oct_poll_controller - poll for receive packets
 411  * device.
 412  *
 413  * @dev:    Device to poll. Unused
 414  */
 415 void cvm_oct_poll_controller(struct net_device *dev)
 416 {
 417         cvm_oct_napi_poll(NULL, 16);
 418 }
 419 #endif
 420
 421 void cvm_oct_rx_initialize(void)
 422 {
 423         int i;
 424         struct net_device *dev_for_napi = NULL;
 425         union cvmx_pow_wq_int_thrx int_thr;
 426         union cvmx_pow_wq_int_pc int_pc;
 427
 428         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 429                 if (cvm_oct_device[i]) {
 430                         dev_for_napi = cvm_oct_device[i];
 431                         break;
 432                 }
 433         }
 434
 435         if (NULL == dev_for_napi)
 436                 panic("No net_devices were allocated.");
 437
 438         netif_napi_add(dev_for_napi, &cvm_oct_napi, cvm_oct_napi_poll,
 439                        rx_napi_weight);
 440         napi_enable(&cvm_oct_napi);
 441
 442         /* Register an IRQ handler to receive POW interrupts */
 443         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 444                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 445
 446         if (i)
 447                 panic("Could not acquire Ethernet IRQ %d\n",
 448                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 449
 450         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 451
 452         int_thr.u64 = 0;
 453         int_thr.s.tc_en = 1;
 454         int_thr.s.tc_thr = 1;
 455         /* Enable POW interrupt when our port has at least one packet */
 456         cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
 457
 458         int_pc.u64 = 0;
 459         int_pc.s.pc_thr = 5;
 460         cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 461
 462         /* Schedule NAPI now. This will indirectly enable the interrupt. */
 463         napi_schedule(&cvm_oct_napi);
 464 }
 465
 466 void cvm_oct_rx_shutdown(void)
 467 {
 468         netif_napi_del(&cvm_oct_napi);
 469 }