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 <asm/octeon/octeon.h>
  30
  31 #include "ethernet-defines.h"
  32 #include "ethernet-mem.h"
  33 #include "ethernet-rx.h"
  34 #include "octeon-ethernet.h"
  35 #include "ethernet-util.h"
  36
  37 #include <asm/octeon/cvmx-helper.h>
  38 #include <asm/octeon/cvmx-wqe.h>
  39 #include <asm/octeon/cvmx-fau.h>
  40 #include <asm/octeon/cvmx-pow.h>
  41 #include <asm/octeon/cvmx-pip.h>
  42 #include <asm/octeon/cvmx-scratch.h>
  43
  44 #include <asm/octeon/cvmx-gmxx-defs.h>
  45
  46 static struct napi_struct cvm_oct_napi;
  47
  48 /**
  49  * cvm_oct_do_interrupt - interrupt handler.
  50  * @cpl: Interrupt number. Unused
  51  * @dev_id: Cookie to identify the device. Unused
  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         int port;
  74
  75         if (octeon_has_feature(OCTEON_FEATURE_PKND))
  76                 port = work->word0.pip.cn68xx.pknd;
  77         else
  78                 port = work->word1.cn38xx.ipprt;
  79
  80         if ((work->word2.snoip.err_code == 10) && (work->word1.len <= 64)) {
  81                 /*
  82                  * Ignore length errors on min size packets. Some
  83                  * equipment incorrectly pads packets to 64+4FCS
  84                  * instead of 60+4FCS.  Note these packets still get
  85                  * counted as frame errors.
  86                  */
  87         } else if (work->word2.snoip.err_code == 5 ||
  88                    work->word2.snoip.err_code == 7) {
  89                 /*
  90                  * We received a packet with either an alignment error
  91                  * or a FCS error. This may be signalling that we are
  92                  * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
  93                  * off. If this is the case we need to parse the
  94                  * packet to determine if we can remove a non spec
  95                  * preamble and generate a correct packet.
  96                  */
  97                 int interface = cvmx_helper_get_interface_num(port);
  98                 int index = cvmx_helper_get_interface_index_num(port);
  99                 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl;
 100
 101                 gmxx_rxx_frm_ctl.u64 =
 102                     cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 103                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 104
 105                         u8 *ptr =
 106                             cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 107                         int i = 0;
 108
 109                         while (i < work->word1.len - 1) {
 110                                 if (*ptr != 0x55)
 111                                         break;
 112                                 ptr++;
 113                                 i++;
 114                         }
 115
 116                         if (*ptr == 0xd5) {
 117                                 /* Port received 0xd5 preamble */
 118                                 work->packet_ptr.s.addr += i + 1;
 119                                 work->word1.len -= i + 5;
 120                         } else if ((*ptr & 0xf) == 0xd) {
 121                                 /* Port received 0xd preamble */
 122                                 work->packet_ptr.s.addr += i;
 123                                 work->word1.len -= i + 4;
 124                                 for (i = 0; i < work->word1.len; i++) {
 125                                         *ptr =
 126                                             ((*ptr & 0xf0) >> 4) |
 127                                             ((*(ptr + 1) & 0xf) << 4);
 128                                         ptr++;
 129                                 }
 130                         } else {
 131                                 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
 132                                                    port);
 133                                 cvm_oct_free_work(work);
 134                                 return 1;
 135                         }
 136                 }
 137         } else {
 138                 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
 139                                    port, work->word2.snoip.err_code);
 140                 cvm_oct_free_work(work);
 141                 return 1;
 142         }
 143
 144         return 0;
 145 }
 146
 147 /**
 148  * cvm_oct_napi_poll - the NAPI poll function.
 149  * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
 150  * @budget: Maximum number of packets to receive.
 151  *
 152  * Returns the number of packets processed.
 153  */
 154 static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 155 {
 156         const int       coreid = cvmx_get_core_num();
 157         u64     old_group_mask;
 158         u64     old_scratch;
 159         int             rx_count = 0;
 160         int             did_work_request = 0;
 161         int             packet_not_copied;
 162
 163         /* Prefetch cvm_oct_device since we know we need it soon */
 164         prefetch(cvm_oct_device);
 165
 166         if (USE_ASYNC_IOBDMA) {
 167                 /* Save scratch in case userspace is using it */
 168                 CVMX_SYNCIOBDMA;
 169                 old_scratch = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 170         }
 171
 172         /* Only allow work for our group (and preserve priorities) */
 173         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 174                 old_group_mask = cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid));
 175                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid),
 176                                 1ull << pow_receive_group);
 177                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 178         } else {
 179                 old_group_mask = cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid));
 180                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 181                         (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 182         }
 183
 184         if (USE_ASYNC_IOBDMA) {
 185                 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
 186                 did_work_request = 1;
 187         }
 188
 189         while (rx_count < budget) {
 190                 struct sk_buff *skb = NULL;
 191                 struct sk_buff **pskb = NULL;
 192                 int skb_in_hw;
 193                 cvmx_wqe_t *work;
 194                 int port;
 195
 196                 if (USE_ASYNC_IOBDMA && did_work_request)
 197                         work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
 198                 else
 199                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 200
 201                 prefetch(work);
 202                 did_work_request = 0;
 203                 if (!work) {
 204                         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 205                                 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS,
 206                                                1ull << pow_receive_group);
 207                                 cvmx_write_csr(CVMX_SSO_WQ_INT,
 208                                                1ull << pow_receive_group);
 209                         } else {
 210                                 union cvmx_pow_wq_int wq_int;
 211
 212                                 wq_int.u64 = 0;
 213                                 wq_int.s.iq_dis = 1 << pow_receive_group;
 214                                 wq_int.s.wq_int = 1 << pow_receive_group;
 215                                 cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 216                         }
 217                         break;
 218                 }
 219                 pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) -
 220                         sizeof(void *));
 221                 prefetch(pskb);
 222
 223                 if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
 224                         cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH,
 225                                                             CVMX_POW_NO_WAIT);
 226                         did_work_request = 1;
 227                 }
 228                 rx_count++;
 229
 230                 skb_in_hw = work->word2.s.bufs == 1;
 231                 if (likely(skb_in_hw)) {
 232                         skb = *pskb;
 233                         prefetch(&skb->head);
 234                         prefetch(&skb->len);
 235                 }
 236
 237                 if (octeon_has_feature(OCTEON_FEATURE_PKND))
 238                         port = work->word0.pip.cn68xx.pknd;
 239                 else
 240                         port = work->word1.cn38xx.ipprt;
 241
 242                 prefetch(cvm_oct_device[port]);
 243
 244                 /* Immediately throw away all packets with receive errors */
 245                 if (unlikely(work->word2.snoip.rcv_error)) {
 246                         if (cvm_oct_check_rcv_error(work))
 247                                 continue;
 248                 }
 249
 250                 /*
 251                  * We can only use the zero copy path if skbuffs are
 252                  * in the FPA pool and the packet fits in a single
 253                  * buffer.
 254                  */
 255                 if (likely(skb_in_hw)) {
 256                         skb->data = skb->head + work->packet_ptr.s.addr -
 257                                 cvmx_ptr_to_phys(skb->head);
 258                         prefetch(skb->data);
 259                         skb->len = work->word1.len;
 260                         skb_set_tail_pointer(skb, skb->len);
 261                         packet_not_copied = 1;
 262                 } else {
 263                         /*
 264                          * We have to copy the packet. First allocate
 265                          * an skbuff for it.
 266                          */
 267                         skb = dev_alloc_skb(work->word1.len);
 268                         if (!skb) {
 269                                 cvm_oct_free_work(work);
 270                                 continue;
 271                         }
 272
 273                         /*
 274                          * Check if we've received a packet that was
 275                          * entirely stored in the work entry.
 276                          */
 277                         if (unlikely(work->word2.s.bufs == 0)) {
 278                                 u8 *ptr = work->packet_data;
 279
 280                                 if (likely(!work->word2.s.not_IP)) {
 281                                         /*
 282                                          * The beginning of the packet
 283                                          * moves for IP packets.
 284                                          */
 285                                         if (work->word2.s.is_v6)
 286                                                 ptr += 2;
 287                                         else
 288                                                 ptr += 6;
 289                                 }
 290                                 memcpy(skb_put(skb, work->word1.len), ptr,
 291                                        work->word1.len);
 292                                 /* No packet buffers to free */
 293                         } else {
 294                                 int segments = work->word2.s.bufs;
 295                                 union cvmx_buf_ptr segment_ptr =
 296                                     work->packet_ptr;
 297                                 int len = work->word1.len;
 298
 299                                 while (segments--) {
 300                                         union cvmx_buf_ptr next_ptr =
 301                                             *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
 302
 303                         /*
 304                          * Octeon Errata PKI-100: The segment size is
 305                          * wrong. Until it is fixed, calculate the
 306                          * segment size based on the packet pool
 307                          * buffer size. When it is fixed, the
 308                          * following line should be replaced with this
 309                          * one: int segment_size =
 310                          * segment_ptr.s.size;
 311                          */
 312                                         int segment_size =
 313                                             CVMX_FPA_PACKET_POOL_SIZE -
 314                                             (segment_ptr.s.addr -
 315                                              (((segment_ptr.s.addr >> 7) -
 316                                                segment_ptr.s.back) << 7));
 317                                         /*
 318                                          * Don't copy more than what
 319                                          * is left in the packet.
 320                                          */
 321                                         if (segment_size > len)
 322                                                 segment_size = len;
 323                                         /* Copy the data into the packet */
 324                                         memcpy(skb_put(skb, segment_size),
 325                                                cvmx_phys_to_ptr(segment_ptr.s.addr),
 326                                                segment_size);
 327                                         len -= segment_size;
 328                                         segment_ptr = next_ptr;
 329                                 }
 330                         }
 331                         packet_not_copied = 0;
 332                 }
 333                 if (likely((port < TOTAL_NUMBER_OF_PORTS) &&
 334                            cvm_oct_device[port])) {
 335                         struct net_device *dev = cvm_oct_device[port];
 336                         struct octeon_ethernet *priv = netdev_priv(dev);
 337
 338                         /*
 339                          * Only accept packets for devices that are
 340                          * currently up.
 341                          */
 342                         if (likely(dev->flags & IFF_UP)) {
 343                                 skb->protocol = eth_type_trans(skb, dev);
 344                                 skb->dev = dev;
 345
 346                                 if (unlikely(work->word2.s.not_IP ||
 347                                              work->word2.s.IP_exc ||
 348                                              work->word2.s.L4_error ||
 349                                              !work->word2.s.tcp_or_udp))
 350                                         skb->ip_summed = CHECKSUM_NONE;
 351                                 else
 352                                         skb->ip_summed = CHECKSUM_UNNECESSARY;
 353
 354                                 /* Increment RX stats for virtual ports */
 355                                 if (port >= CVMX_PIP_NUM_INPUT_PORTS) {
 356                                         priv->stats.rx_packets++;
 357                                         priv->stats.rx_bytes += skb->len;
 358                                 }
 359                                 netif_receive_skb(skb);
 360                         } else {
 361                                 /*
 362                                  * Drop any packet received for a device that
 363                                  * isn't up.
 364                                  */
 365                                 priv->stats.rx_dropped++;
 366                                 dev_kfree_skb_irq(skb);
 367                         }
 368                 } else {
 369                         /*
 370                          * Drop any packet received for a device that
 371                          * doesn't exist.
 372                          */
 373                         printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
 374                                    port);
 375                         dev_kfree_skb_irq(skb);
 376                 }
 377                 /*
 378                  * Check to see if the skbuff and work share the same
 379                  * packet buffer.
 380                  */
 381                 if (likely(packet_not_copied)) {
 382                         /*
 383                          * This buffer needs to be replaced, increment
 384                          * the number of buffers we need to free by
 385                          * one.
 386                          */
 387                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 388                                               1);
 389
 390                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL, 1);
 391                 } else {
 392                         cvm_oct_free_work(work);
 393                 }
 394         }
 395         /* Restore the original POW group mask */
 396         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 397                 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid), old_group_mask);
 398                 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid)); /* Flush */
 399         } else {
 400                 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 401         }
 402
 403         if (USE_ASYNC_IOBDMA) {
 404                 /* Restore the scratch area */
 405                 cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 406         }
 407         cvm_oct_rx_refill_pool(0);
 408
 409         if (rx_count < budget && napi) {
 410                 /* No more work */
 411                 napi_complete(napi);
 412                 enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 413         }
 414         return rx_count;
 415 }
 416
 417 #ifdef CONFIG_NET_POLL_CONTROLLER
 418 /**
 419  * cvm_oct_poll_controller - poll for receive packets
 420  * device.
 421  *
 422  * @dev:    Device to poll. Unused
 423  */
 424 void cvm_oct_poll_controller(struct net_device *dev)
 425 {
 426         cvm_oct_napi_poll(NULL, 16);
 427 }
 428 #endif
 429
 430 void cvm_oct_rx_initialize(void)
 431 {
 432         int i;
 433         struct net_device *dev_for_napi = NULL;
 434
 435         for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
 436                 if (cvm_oct_device[i]) {
 437                         dev_for_napi = cvm_oct_device[i];
 438                         break;
 439                 }
 440         }
 441
 442         if (!dev_for_napi)
 443                 panic("No net_devices were allocated.");
 444
 445         netif_napi_add(dev_for_napi, &cvm_oct_napi, cvm_oct_napi_poll,
 446                        rx_napi_weight);
 447         napi_enable(&cvm_oct_napi);
 448
 449         /* Register an IRQ handler to receive POW interrupts */
 450         i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
 451                         cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
 452
 453         if (i)
 454                 panic("Could not acquire Ethernet IRQ %d\n",
 455                       OCTEON_IRQ_WORKQ0 + pow_receive_group);
 456
 457         disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
 458
 459         /* Enable POW interrupt when our port has at least one packet */
 460         if (OCTEON_IS_MODEL(OCTEON_CN68XX)) {
 461                 union cvmx_sso_wq_int_thrx int_thr;
 462                 union cvmx_pow_wq_int_pc int_pc;
 463
 464                 int_thr.u64 = 0;
 465                 int_thr.s.tc_en = 1;
 466                 int_thr.s.tc_thr = 1;
 467                 cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(pow_receive_group),
 468                                int_thr.u64);
 469
 470                 int_pc.u64 = 0;
 471                 int_pc.s.pc_thr = 5;
 472                 cvmx_write_csr(CVMX_SSO_WQ_INT_PC, int_pc.u64);
 473         } else {
 474                 union cvmx_pow_wq_int_thrx int_thr;
 475                 union cvmx_pow_wq_int_pc int_pc;
 476
 477                 int_thr.u64 = 0;
 478                 int_thr.s.tc_en = 1;
 479                 int_thr.s.tc_thr = 1;
 480                 cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group),
 481                                int_thr.u64);
 482
 483                 int_pc.u64 = 0;
 484                 int_pc.s.pc_thr = 5;
 485                 cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
 486         }
 487
 488         /* Schedule NAPI now. This will indirectly enable the interrupt. */
 489         napi_schedule(&cvm_oct_napi);
 490 }
 491
 492 void cvm_oct_rx_shutdown(void)
 493 {
 494         netif_napi_del(&cvm_oct_napi);
 495 }