sys/mips/cavium/octe/ethernet-rx.c

   1 /*************************************************************************
   2 Copyright (c) 2003-2007  Cavium Networks (support@cavium.com). All rights
   3 reserved.
   4
   5
   6 Redistribution and use in source and binary forms, with or without
   7 modification, are permitted provided that the following conditions are
   8 met:
   9
  10     * Redistributions of source code must retain the above copyright
  11       notice, this list of conditions and the following disclaimer.
  12
  13     * Redistributions in binary form must reproduce the above
  14       copyright notice, this list of conditions and the following
  15       disclaimer in the documentation and/or other materials provided
  16       with the distribution.
  17
  18     * Neither the name of Cavium Networks nor the names of
  19       its contributors may be used to endorse or promote products
  20       derived from this software without specific prior written
  21       permission.
  22
  23 This Software, including technical data, may be subject to U.S. export  control laws, including the U.S. Export Administration Act and its  associated regulations, and may be subject to export or import  regulations in other countries.
  24
  25 TO THE MAXIMUM EXTENT PERMITTED BY LAW, THE SOFTWARE IS PROVIDED "AS IS"
  26 AND WITH ALL FAULTS AND CAVIUM  NETWORKS MAKES NO PROMISES, REPRESENTATIONS OR WARRANTIES, EITHER EXPRESS, IMPLIED, STATUTORY, OR OTHERWISE, WITH RESPECT TO THE SOFTWARE, INCLUDING ITS CONDITION, ITS CONFORMITY TO ANY REPRESENTATION OR DESCRIPTION, OR THE EXISTENCE OF ANY LATENT OR PATENT DEFECTS, AND CAVIUM SPECIFICALLY DISCLAIMS ALL IMPLIED (IF ANY) WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, FITNESS FOR A PARTICULAR PURPOSE, LACK OF VIRUSES, ACCURACY OR COMPLETENESS, QUIET ENJOYMENT, QUIET POSSESSION OR CORRESPONDENCE TO DESCRIPTION. THE ENTIRE  RISK ARISING OUT OF USE OR PERFORMANCE OF THE SOFTWARE LIES WITH YOU.
  27
  28 *************************************************************************/
  29
  30 #include <sys/cdefs.h>
  31 __FBSDID("$FreeBSD$");
  32
  33 #include <sys/param.h>
  34 #include <sys/systm.h>
  35 #include <sys/bus.h>
  36 #include <sys/endian.h>
  37 #include <sys/kernel.h>
  38 #include <sys/mbuf.h>
  39 #include <sys/socket.h>
  40 #include <sys/proc.h>
  41 #include <sys/sched.h>
  42 #include <sys/smp.h>
  43 #include <sys/taskqueue.h>
  44
  45 #include <net/ethernet.h>
  46 #include <net/if.h>
  47
  48 #include "wrapper-cvmx-includes.h"
  49 #include "ethernet-headers.h"
  50
  51 extern int pow_receive_group;
  52 extern struct ifnet *cvm_oct_device[];
  53
  54 static struct task cvm_oct_task;
  55 static struct taskqueue *cvm_oct_taskq;
  56
  57 static int cvm_oct_rx_active;
  58
  59 /**
  60  * Interrupt handler. The interrupt occurs whenever the POW
  61  * transitions from 0->1 packets in our group.
  62  *
  63  * @param cpl
  64  * @param dev_id
  65  * @param regs
  66  * @return
  67  */
  68 int cvm_oct_do_interrupt(void *dev_id)
  69 {
  70         /* Acknowledge the interrupt */
  71         if (INTERRUPT_LIMIT)
  72                 cvmx_write_csr(CVMX_POW_WQ_INT, 1<<pow_receive_group);
  73         else
  74                 cvmx_write_csr(CVMX_POW_WQ_INT, 0x10001<<pow_receive_group);
  75
  76         /*
  77          * Schedule task if there isn't one running.
  78          */
  79         if (atomic_cmpset_int(&cvm_oct_rx_active, 0, 1))
  80                 taskqueue_enqueue(cvm_oct_taskq, &cvm_oct_task);
  81
  82         return FILTER_HANDLED;
  83 }
  84
  85
  86 /**
  87  * This is called on receive errors, and determines if the packet
  88  * can be dropped early-on in cvm_oct_tasklet_rx().
  89  *
  90  * @param work Work queue entry pointing to the packet.
  91  * @return Non-zero if the packet can be dropped, zero otherwise.
  92  */
  93 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
  94 {
  95         if ((work->word2.snoip.err_code == 10) && (work->word1.s.len <= 64)) {
  96                 /* Ignore length errors on min size packets. Some equipment
  97                    incorrectly pads packets to 64+4FCS instead of 60+4FCS.
  98                    Note these packets still get counted as frame errors. */
  99         } else
 100         if (USE_10MBPS_PREAMBLE_WORKAROUND && ((work->word2.snoip.err_code == 5) || (work->word2.snoip.err_code == 7))) {
 101
 102                 /* We received a packet with either an alignment error or a
 103                    FCS error. This may be signalling that we are running
 104                    10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK} off. If this is the
 105                    case we need to parse the packet to determine if we can
 106                    remove a non spec preamble and generate a correct packet */
 107                 int interface = cvmx_helper_get_interface_num(work->word1.cn38xx.ipprt);
 108                 int index = cvmx_helper_get_interface_index_num(work->word1.cn38xx.ipprt);
 109                 cvmx_gmxx_rxx_frm_ctl_t gmxx_rxx_frm_ctl;
 110                 gmxx_rxx_frm_ctl.u64 = cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index, interface));
 111                 if (gmxx_rxx_frm_ctl.s.pre_chk == 0) {
 112
 113                         uint8_t *ptr = cvmx_phys_to_ptr(work->packet_ptr.s.addr);
 114                         int i = 0;
 115
 116                         while (i < work->word1.s.len-1) {
 117                                 if (*ptr != 0x55)
 118                                         break;
 119                                 ptr++;
 120                                 i++;
 121                         }
 122
 123                         if (*ptr == 0xd5) {
 124                                 /*
 125                                 DEBUGPRINT("Port %d received 0xd5 preamble\n", work->word1.cn38xx.ipprt);
 126                                 */
 127                                 work->packet_ptr.s.addr += i+1;
 128                                 work->word1.s.len -= i+5;
 129                         } else
 130                         if ((*ptr & 0xf) == 0xd) {
 131                                 /*
 132                                 DEBUGPRINT("Port %d received 0x?d preamble\n", work->word1.cn38xx.ipprt);
 133                                 */
 134                                 work->packet_ptr.s.addr += i;
 135                                 work->word1.s.len -= i+4;
 136                                 for (i = 0; i < work->word1.s.len; i++) {
 137                                         *ptr = ((*ptr&0xf0)>>4) | ((*(ptr+1)&0xf)<<4);
 138                                         ptr++;
 139                                 }
 140                         } else {
 141                                 DEBUGPRINT("Port %d unknown preamble, packet dropped\n", work->word1.cn38xx.ipprt);
 142                                 /*
 143                                 cvmx_helper_dump_packet(work);
 144                                 */
 145                                 cvm_oct_free_work(work);
 146                                 return 1;
 147                         }
 148                 }
 149         } else {
 150                 DEBUGPRINT("Port %d receive error code %d, packet dropped\n", work->word1.cn38xx.ipprt, work->word2.snoip.err_code);
 151                 cvm_oct_free_work(work);
 152                 return 1;
 153         }
 154
 155         return 0;
 156 }
 157
 158 /**
 159  * Tasklet function that is scheduled on a core when an interrupt occurs.
 160  *
 161  * @param unused
 162  */
 163 void cvm_oct_tasklet_rx(void *context, int pending)
 164 {
 165         int                 coreid;
 166         uint64_t            old_group_mask;
 167         int                 rx_count = 0;
 168         int                 number_to_free;
 169         int                 num_freed;
 170         int                 packet_not_copied;
 171
 172         sched_pin();
 173         coreid = cvmx_get_core_num();
 174
 175         /* Prefetch cvm_oct_device since we know we need it soon */
 176         CVMX_PREFETCH(cvm_oct_device, 0);
 177
 178         /* Only allow work for our group (and preserve priorities) */
 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         while (1) {
 184                 struct mbuf *m = NULL;
 185                 int mbuf_in_hw;
 186                 cvmx_wqe_t *work;
 187
 188                 if ((INTERRUPT_LIMIT == 0) || (rx_count < MAX_RX_PACKETS))
 189                         work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
 190                 else
 191                         work = NULL;
 192                 CVMX_PREFETCH(work, 0);
 193                 if (work == NULL)
 194                         break;
 195
 196                 mbuf_in_hw = work->word2.s.bufs == 1;
 197                 if ((mbuf_in_hw)) {
 198                         m = *(struct mbuf **)(cvm_oct_get_buffer_ptr(work->packet_ptr) - sizeof(void *));
 199                         CVMX_PREFETCH(m, offsetof(struct mbuf, m_data));
 200                         CVMX_PREFETCH(m, offsetof(struct mbuf, m_pkthdr));
 201                 }
 202                 CVMX_PREFETCH(cvm_oct_device[work->word1.cn38xx.ipprt], 0);
 203                 //CVMX_PREFETCH(m, 0);
 204
 205
 206                 rx_count++;
 207                 /* Immediately throw away all packets with receive errors */
 208                 if ((work->word2.snoip.rcv_error)) {
 209                         if (cvm_oct_check_rcv_error(work))
 210                                 continue;
 211                 }
 212
 213                 /* We can only use the zero copy path if mbufs are in the FPA pool
 214                    and the packet fits in a single buffer */
 215                 if ((mbuf_in_hw)) {
 216                         CVMX_PREFETCH(m->m_data, 0);
 217
 218                         m->m_pkthdr.len = m->m_len = work->word1.s.len;
 219
 220                         packet_not_copied = 1;
 221
 222                         /*
 223                          * Adjust the data pointer based on the offset
 224                          * of the packet within the buffer.
 225                          */
 226                         m->m_data += (work->packet_ptr.s.back << 7) + (work->packet_ptr.s.addr & 0x7f);
 227                 } else {
 228
 229                         /* We have to copy the packet. First allocate an
 230                            mbuf for it */
 231                         MGETHDR(m, M_NOWAIT, MT_DATA);
 232                         if (m == NULL) {
 233                                 DEBUGPRINT("Port %d failed to allocate mbuf, packet dropped\n", work->word1.cn38xx.ipprt);
 234                                 cvm_oct_free_work(work);
 235                                 continue;
 236                         }
 237
 238                         /* Check if we've received a packet that was entirely
 239                            stored in the work entry. This is untested */
 240                         if ((work->word2.s.bufs == 0)) {
 241                                 uint8_t *ptr = work->packet_data;
 242
 243                                 if (cvmx_likely(!work->word2.s.not_IP)) {
 244                                         /* The beginning of the packet moves
 245                                            for IP packets */
 246                                         if (work->word2.s.is_v6)
 247                                                 ptr += 2;
 248                                         else
 249                                                 ptr += 6;
 250                                 }
 251                                 panic("%s: not yet implemented; copy in small packet.", __func__);
 252                                 /* No packet buffers to free */
 253                         } else {
 254                                 int segments = work->word2.s.bufs;
 255                                 cvmx_buf_ptr_t segment_ptr = work->packet_ptr;
 256                                 int len = work->word1.s.len;
 257
 258                                 while (segments--) {
 259                                         cvmx_buf_ptr_t next_ptr = *(cvmx_buf_ptr_t *)cvmx_phys_to_ptr(segment_ptr.s.addr-8);
 260                                         /* Octeon Errata PKI-100: The segment
 261                                            size is wrong. Until it is fixed,
 262                                            calculate the segment size based on
 263                                            the packet pool buffer size. When
 264                                            it is fixed, the following line
 265                                            should be replaced with this one:
 266                                         int segment_size = segment_ptr.s.size; */
 267                                         int segment_size = CVMX_FPA_PACKET_POOL_SIZE - (segment_ptr.s.addr - (((segment_ptr.s.addr >> 7) - segment_ptr.s.back) << 7));
 268                                         /* Don't copy more than what is left
 269                                            in the packet */
 270                                         if (segment_size > len)
 271                                                 segment_size = len;
 272                                         /* Copy the data into the packet */
 273                                         panic("%s: not yet implemented; copy in packet segments.", __func__);
 274 #if 0
 275                                         memcpy(m_put(m, segment_size), cvmx_phys_to_ptr(segment_ptr.s.addr), segment_size);
 276 #endif
 277                                         /* Reduce the amount of bytes left
 278                                            to copy */
 279                                         len -= segment_size;
 280                                         segment_ptr = next_ptr;
 281                                 }
 282                         }
 283                         packet_not_copied = 0;
 284                 }
 285
 286                 if (((work->word1.cn38xx.ipprt < TOTAL_NUMBER_OF_PORTS) &&
 287                     cvm_oct_device[work->word1.cn38xx.ipprt])) {
 288                         struct ifnet *ifp = cvm_oct_device[work->word1.cn38xx.ipprt];
 289
 290                         /* Only accept packets for devices
 291                            that are currently up */
 292                         if ((ifp->if_flags & IFF_UP)) {
 293                                 m->m_pkthdr.rcvif = ifp;
 294
 295                                 if ((ifp->if_capenable & IFCAP_RXCSUM) != 0) {
 296                                         if ((work->word2.s.not_IP || work->word2.s.IP_exc || work->word2.s.L4_error))
 297                                                 m->m_pkthdr.csum_flags = 0; /* XXX */
 298                                         else {
 299                                                 m->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
 300                                                 m->m_pkthdr.csum_data = 0xffff;
 301                                         }
 302                                 } else {
 303                                         m->m_pkthdr.csum_flags = 0; /* XXX */
 304                                 }
 305
 306                                 ifp->if_ipackets++;
 307
 308                                 (*ifp->if_input)(ifp, m);
 309                         } else {
 310                                 /* Drop any packet received for a device that isn't up */
 311                                 /*
 312                                 DEBUGPRINT("%s: Device not up, packet dropped\n",
 313                                            if_name(ifp));
 314                                 */
 315                                 m_freem(m);
 316                         }
 317                 } else {
 318                         /* Drop any packet received for a device that
 319                            doesn't exist */
 320                         DEBUGPRINT("Port %d not controlled by FreeBSD, packet dropped\n", work->word1.cn38xx.ipprt);
 321                         m_freem(m);
 322                 }
 323
 324                 /* Check to see if the mbuf and work share
 325                    the same packet buffer */
 326                 if ((packet_not_copied)) {
 327                         /* This buffer needs to be replaced, increment
 328                         the number of buffers we need to free by one */
 329                         cvmx_fau_atomic_add32(
 330                                 FAU_NUM_PACKET_BUFFERS_TO_FREE, 1);
 331
 332                         cvmx_fpa_free(work, CVMX_FPA_WQE_POOL,
 333                                       DONT_WRITEBACK(1));
 334                 } else
 335                         cvm_oct_free_work(work);
 336         }
 337
 338         /*
 339          * If we hit our limit, schedule another task while we clean up.
 340          */
 341         if (INTERRUPT_LIMIT != 0 && rx_count == MAX_RX_PACKETS) {
 342                 taskqueue_enqueue(cvm_oct_taskq, &cvm_oct_task);
 343         } else {
 344                 /*
 345                  * No more packets, all done.
 346                  */
 347                 if (!atomic_cmpset_int(&cvm_oct_rx_active, 1, 0))
 348                         panic("%s: inconsistent rx active state.", __func__);
 349         }
 350
 351         /* Restore the original POW group mask */
 352         cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 353
 354         /* Refill the packet buffer pool */
 355         number_to_free =
 356           cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 357
 358         if (number_to_free > 0) {
 359                 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 360                                       -number_to_free);
 361                 num_freed =
 362                         cvm_oct_mem_fill_fpa(CVMX_FPA_PACKET_POOL,
 363                                              CVMX_FPA_PACKET_POOL_SIZE,
 364                                              number_to_free);
 365                 if (num_freed != number_to_free) {
 366                         cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
 367                                               number_to_free - num_freed);
 368                 }
 369         }
 370         sched_unpin();
 371 }
 372
 373
 374
 375 void cvm_oct_rx_initialize(void)
 376 {
 377         TASK_INIT(&cvm_oct_task, 0, cvm_oct_tasklet_rx, NULL);
 378
 379         cvm_oct_taskq = taskqueue_create_fast("oct_rx", M_NOWAIT,
 380                                               taskqueue_thread_enqueue,
 381                                               &cvm_oct_taskq);
 382         taskqueue_start_threads(&cvm_oct_taskq, min(mp_ncpus, MAXCPU),
 383                                 PI_NET, "octe taskq");
 384 }
 385
 386 void cvm_oct_rx_shutdown(void)
 387 {
 388         panic("%s: not yet implemented.", __func__);
 389 }
 390