drivers/infiniband/hw/hfi1/driver.c

   1 /*
   2  * Copyright(c) 2015, 2016 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47
  48 #include <linux/spinlock.h>
  49 #include <linux/pci.h>
  50 #include <linux/io.h>
  51 #include <linux/delay.h>
  52 #include <linux/netdevice.h>
  53 #include <linux/vmalloc.h>
  54 #include <linux/module.h>
  55 #include <linux/prefetch.h>
  56 #include <rdma/ib_verbs.h>
  57
  58 #include "hfi.h"
  59 #include "trace.h"
  60 #include "qp.h"
  61 #include "sdma.h"
  62
  63 #undef pr_fmt
  64 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  65
  66 /*
  67  * The size has to be longer than this string, so we can append
  68  * board/chip information to it in the initialization code.
  69  */
  70 const char ib_hfi1_version[] = HFI1_DRIVER_VERSION "\n";
  71
  72 DEFINE_SPINLOCK(hfi1_devs_lock);
  73 LIST_HEAD(hfi1_dev_list);
  74 DEFINE_MUTEX(hfi1_mutex);       /* general driver use */
  75
  76 unsigned int hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
  77 module_param_named(max_mtu, hfi1_max_mtu, uint, S_IRUGO);
  78 MODULE_PARM_DESC(max_mtu, "Set max MTU bytes, default is " __stringify(
  79                  HFI1_DEFAULT_MAX_MTU));
  80
  81 unsigned int hfi1_cu = 1;
  82 module_param_named(cu, hfi1_cu, uint, S_IRUGO);
  83 MODULE_PARM_DESC(cu, "Credit return units");
  84
  85 unsigned long hfi1_cap_mask = HFI1_CAP_MASK_DEFAULT;
  86 static int hfi1_caps_set(const char *, const struct kernel_param *);
  87 static int hfi1_caps_get(char *, const struct kernel_param *);
  88 static const struct kernel_param_ops cap_ops = {
  89         .set = hfi1_caps_set,
  90         .get = hfi1_caps_get
  91 };
  92 module_param_cb(cap_mask, &cap_ops, &hfi1_cap_mask, S_IWUSR | S_IRUGO);
  93 MODULE_PARM_DESC(cap_mask, "Bit mask of enabled/disabled HW features");
  94
  95 MODULE_LICENSE("Dual BSD/GPL");
  96 MODULE_DESCRIPTION("Intel Omni-Path Architecture driver");
  97 MODULE_VERSION(HFI1_DRIVER_VERSION);
  98
  99 /*
 100  * MAX_PKT_RCV is the max # if packets processed per receive interrupt.
 101  */
 102 #define MAX_PKT_RECV 64
 103 /*
 104  * MAX_PKT_THREAD_RCV is the max # of packets processed before
 105  * the qp_wait_list queue is flushed.
 106  */
 107 #define MAX_PKT_RECV_THREAD (MAX_PKT_RECV * 4)
 108 #define EGR_HEAD_UPDATE_THRESHOLD 16
 109
 110 struct hfi1_ib_stats hfi1_stats;
 111
 112 static int hfi1_caps_set(const char *val, const struct kernel_param *kp)
 113 {
 114         int ret = 0;
 115         unsigned long *cap_mask_ptr = (unsigned long *)kp->arg,
 116                 cap_mask = *cap_mask_ptr, value, diff,
 117                 write_mask = ((HFI1_CAP_WRITABLE_MASK << HFI1_CAP_USER_SHIFT) |
 118                               HFI1_CAP_WRITABLE_MASK);
 119
 120         ret = kstrtoul(val, 0, &value);
 121         if (ret) {
 122                 pr_warn("Invalid module parameter value for 'cap_mask'\n");
 123                 goto done;
 124         }
 125         /* Get the changed bits (except the locked bit) */
 126         diff = value ^ (cap_mask & ~HFI1_CAP_LOCKED_SMASK);
 127
 128         /* Remove any bits that are not allowed to change after driver load */
 129         if (HFI1_CAP_LOCKED() && (diff & ~write_mask)) {
 130                 pr_warn("Ignoring non-writable capability bits %#lx\n",
 131                         diff & ~write_mask);
 132                 diff &= write_mask;
 133         }
 134
 135         /* Mask off any reserved bits */
 136         diff &= ~HFI1_CAP_RESERVED_MASK;
 137         /* Clear any previously set and changing bits */
 138         cap_mask &= ~diff;
 139         /* Update the bits with the new capability */
 140         cap_mask |= (value & diff);
 141         /* Check for any kernel/user restrictions */
 142         diff = (cap_mask & (HFI1_CAP_MUST_HAVE_KERN << HFI1_CAP_USER_SHIFT)) ^
 143                 ((cap_mask & HFI1_CAP_MUST_HAVE_KERN) << HFI1_CAP_USER_SHIFT);
 144         cap_mask &= ~diff;
 145         /* Set the bitmask to the final set */
 146         *cap_mask_ptr = cap_mask;
 147 done:
 148         return ret;
 149 }
 150
 151 static int hfi1_caps_get(char *buffer, const struct kernel_param *kp)
 152 {
 153         unsigned long cap_mask = *(unsigned long *)kp->arg;
 154
 155         cap_mask &= ~HFI1_CAP_LOCKED_SMASK;
 156         cap_mask |= ((cap_mask & HFI1_CAP_K2U) << HFI1_CAP_USER_SHIFT);
 157
 158         return scnprintf(buffer, PAGE_SIZE, "0x%lx", cap_mask);
 159 }
 160
 161 const char *get_unit_name(int unit)
 162 {
 163         static char iname[16];
 164
 165         snprintf(iname, sizeof(iname), DRIVER_NAME "_%u", unit);
 166         return iname;
 167 }
 168
 169 const char *get_card_name(struct rvt_dev_info *rdi)
 170 {
 171         struct hfi1_ibdev *ibdev = container_of(rdi, struct hfi1_ibdev, rdi);
 172         struct hfi1_devdata *dd = container_of(ibdev,
 173                                                struct hfi1_devdata, verbs_dev);
 174         return get_unit_name(dd->unit);
 175 }
 176
 177 struct pci_dev *get_pci_dev(struct rvt_dev_info *rdi)
 178 {
 179         struct hfi1_ibdev *ibdev = container_of(rdi, struct hfi1_ibdev, rdi);
 180         struct hfi1_devdata *dd = container_of(ibdev,
 181                                                struct hfi1_devdata, verbs_dev);
 182         return dd->pcidev;
 183 }
 184
 185 /*
 186  * Return count of units with at least one port ACTIVE.
 187  */
 188 int hfi1_count_active_units(void)
 189 {
 190         struct hfi1_devdata *dd;
 191         struct hfi1_pportdata *ppd;
 192         unsigned long flags;
 193         int pidx, nunits_active = 0;
 194
 195         spin_lock_irqsave(&hfi1_devs_lock, flags);
 196         list_for_each_entry(dd, &hfi1_dev_list, list) {
 197                 if (!(dd->flags & HFI1_PRESENT) || !dd->kregbase)
 198                         continue;
 199                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 200                         ppd = dd->pport + pidx;
 201                         if (ppd->lid && ppd->linkup) {
 202                                 nunits_active++;
 203                                 break;
 204                         }
 205                 }
 206         }
 207         spin_unlock_irqrestore(&hfi1_devs_lock, flags);
 208         return nunits_active;
 209 }
 210
 211 /*
 212  * Return count of all units, optionally return in arguments
 213  * the number of usable (present) units, and the number of
 214  * ports that are up.
 215  */
 216 int hfi1_count_units(int *npresentp, int *nupp)
 217 {
 218         int nunits = 0, npresent = 0, nup = 0;
 219         struct hfi1_devdata *dd;
 220         unsigned long flags;
 221         int pidx;
 222         struct hfi1_pportdata *ppd;
 223
 224         spin_lock_irqsave(&hfi1_devs_lock, flags);
 225
 226         list_for_each_entry(dd, &hfi1_dev_list, list) {
 227                 nunits++;
 228                 if ((dd->flags & HFI1_PRESENT) && dd->kregbase)
 229                         npresent++;
 230                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 231                         ppd = dd->pport + pidx;
 232                         if (ppd->lid && ppd->linkup)
 233                                 nup++;
 234                 }
 235         }
 236
 237         spin_unlock_irqrestore(&hfi1_devs_lock, flags);
 238
 239         if (npresentp)
 240                 *npresentp = npresent;
 241         if (nupp)
 242                 *nupp = nup;
 243
 244         return nunits;
 245 }
 246
 247 /*
 248  * Get address of eager buffer from it's index (allocated in chunks, not
 249  * contiguous).
 250  */
 251 static inline void *get_egrbuf(const struct hfi1_ctxtdata *rcd, u64 rhf,
 252                                u8 *update)
 253 {
 254         u32 idx = rhf_egr_index(rhf), offset = rhf_egr_buf_offset(rhf);
 255
 256         *update |= !(idx & (rcd->egrbufs.threshold - 1)) && !offset;
 257         return (void *)(((u64)(rcd->egrbufs.rcvtids[idx].addr)) +
 258                         (offset * RCV_BUF_BLOCK_SIZE));
 259 }
 260
 261 /*
 262  * Validate and encode the a given RcvArray Buffer size.
 263  * The function will check whether the given size falls within
 264  * allowed size ranges for the respective type and, optionally,
 265  * return the proper encoding.
 266  */
 267 int hfi1_rcvbuf_validate(u32 size, u8 type, u16 *encoded)
 268 {
 269         if (unlikely(!PAGE_ALIGNED(size)))
 270                 return 0;
 271         if (unlikely(size < MIN_EAGER_BUFFER))
 272                 return 0;
 273         if (size >
 274             (type == PT_EAGER ? MAX_EAGER_BUFFER : MAX_EXPECTED_BUFFER))
 275                 return 0;
 276         if (encoded)
 277                 *encoded = ilog2(size / PAGE_SIZE) + 1;
 278         return 1;
 279 }
 280
 281 static void rcv_hdrerr(struct hfi1_ctxtdata *rcd, struct hfi1_pportdata *ppd,
 282                        struct hfi1_packet *packet)
 283 {
 284         struct ib_header *rhdr = packet->hdr;
 285         u32 rte = rhf_rcv_type_err(packet->rhf);
 286         int lnh = be16_to_cpu(rhdr->lrh[0]) & 3;
 287         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
 288         struct hfi1_devdata *dd = ppd->dd;
 289         struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
 290
 291         if (packet->rhf & (RHF_VCRC_ERR | RHF_ICRC_ERR))
 292                 return;
 293
 294         if (packet->rhf & RHF_TID_ERR) {
 295                 /* For TIDERR and RC QPs preemptively schedule a NAK */
 296                 struct ib_other_headers *ohdr = NULL;
 297                 u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */
 298                 u16 lid  = be16_to_cpu(rhdr->lrh[1]);
 299                 u32 qp_num;
 300                 u32 rcv_flags = 0;
 301
 302                 /* Sanity check packet */
 303                 if (tlen < 24)
 304                         goto drop;
 305
 306                 /* Check for GRH */
 307                 if (lnh == HFI1_LRH_BTH) {
 308                         ohdr = &rhdr->u.oth;
 309                 } else if (lnh == HFI1_LRH_GRH) {
 310                         u32 vtf;
 311
 312                         ohdr = &rhdr->u.l.oth;
 313                         if (rhdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
 314                                 goto drop;
 315                         vtf = be32_to_cpu(rhdr->u.l.grh.version_tclass_flow);
 316                         if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
 317                                 goto drop;
 318                         rcv_flags |= HFI1_HAS_GRH;
 319                 } else {
 320                         goto drop;
 321                 }
 322                 /* Get the destination QP number. */
 323                 qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
 324                 if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
 325                         struct rvt_qp *qp;
 326                         unsigned long flags;
 327
 328                         rcu_read_lock();
 329                         qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
 330                         if (!qp) {
 331                                 rcu_read_unlock();
 332                                 goto drop;
 333                         }
 334
 335                         /*
 336                          * Handle only RC QPs - for other QP types drop error
 337                          * packet.
 338                          */
 339                         spin_lock_irqsave(&qp->r_lock, flags);
 340
 341                         /* Check for valid receive state. */
 342                         if (!(ib_rvt_state_ops[qp->state] &
 343                               RVT_PROCESS_RECV_OK)) {
 344                                 ibp->rvp.n_pkt_drops++;
 345                         }
 346
 347                         switch (qp->ibqp.qp_type) {
 348                         case IB_QPT_RC:
 349                                 hfi1_rc_hdrerr(
 350                                         rcd,
 351                                         rhdr,
 352                                         rcv_flags,
 353                                         qp);
 354                                 break;
 355                         default:
 356                                 /* For now don't handle any other QP types */
 357                                 break;
 358                         }
 359
 360                         spin_unlock_irqrestore(&qp->r_lock, flags);
 361                         rcu_read_unlock();
 362                 } /* Unicast QP */
 363         } /* Valid packet with TIDErr */
 364
 365         /* handle "RcvTypeErr" flags */
 366         switch (rte) {
 367         case RHF_RTE_ERROR_OP_CODE_ERR:
 368         {
 369                 u32 opcode;
 370                 void *ebuf = NULL;
 371                 __be32 *bth = NULL;
 372
 373                 if (rhf_use_egr_bfr(packet->rhf))
 374                         ebuf = packet->ebuf;
 375
 376                 if (!ebuf)
 377                         goto drop; /* this should never happen */
 378
 379                 if (lnh == HFI1_LRH_BTH)
 380                         bth = (__be32 *)ebuf;
 381                 else if (lnh == HFI1_LRH_GRH)
 382                         bth = (__be32 *)((char *)ebuf + sizeof(struct ib_grh));
 383                 else
 384                         goto drop;
 385
 386                 opcode = be32_to_cpu(bth[0]) >> 24;
 387                 opcode &= 0xff;
 388
 389                 if (opcode == IB_OPCODE_CNP) {
 390                         /*
 391                          * Only in pre-B0 h/w is the CNP_OPCODE handled
 392                          * via this code path.
 393                          */
 394                         struct rvt_qp *qp = NULL;
 395                         u32 lqpn, rqpn;
 396                         u16 rlid;
 397                         u8 svc_type, sl, sc5;
 398
 399                         sc5 = hdr2sc(rhdr, packet->rhf);
 400                         sl = ibp->sc_to_sl[sc5];
 401
 402                         lqpn = be32_to_cpu(bth[1]) & RVT_QPN_MASK;
 403                         rcu_read_lock();
 404                         qp = rvt_lookup_qpn(rdi, &ibp->rvp, lqpn);
 405                         if (!qp) {
 406                                 rcu_read_unlock();
 407                                 goto drop;
 408                         }
 409
 410                         switch (qp->ibqp.qp_type) {
 411                         case IB_QPT_UD:
 412                                 rlid = 0;
 413                                 rqpn = 0;
 414                                 svc_type = IB_CC_SVCTYPE_UD;
 415                                 break;
 416                         case IB_QPT_UC:
 417                                 rlid = be16_to_cpu(rhdr->lrh[3]);
 418                                 rqpn = qp->remote_qpn;
 419                                 svc_type = IB_CC_SVCTYPE_UC;
 420                                 break;
 421                         default:
 422                                 goto drop;
 423                         }
 424
 425                         process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
 426                         rcu_read_unlock();
 427                 }
 428
 429                 packet->rhf &= ~RHF_RCV_TYPE_ERR_SMASK;
 430                 break;
 431         }
 432         default:
 433                 break;
 434         }
 435
 436 drop:
 437         return;
 438 }
 439
 440 static inline void init_packet(struct hfi1_ctxtdata *rcd,
 441                                struct hfi1_packet *packet)
 442 {
 443         packet->rsize = rcd->rcvhdrqentsize; /* words */
 444         packet->maxcnt = rcd->rcvhdrq_cnt * packet->rsize; /* words */
 445         packet->rcd = rcd;
 446         packet->updegr = 0;
 447         packet->etail = -1;
 448         packet->rhf_addr = get_rhf_addr(rcd);
 449         packet->rhf = rhf_to_cpu(packet->rhf_addr);
 450         packet->rhqoff = rcd->head;
 451         packet->numpkt = 0;
 452         packet->rcv_flags = 0;
 453 }
 454
 455 void hfi1_process_ecn_slowpath(struct rvt_qp *qp, struct hfi1_packet *pkt,
 456                                bool do_cnp)
 457 {
 458         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 459         struct ib_header *hdr = pkt->hdr;
 460         struct ib_other_headers *ohdr = pkt->ohdr;
 461         struct ib_grh *grh = NULL;
 462         u32 rqpn = 0, bth1;
 463         u16 rlid, dlid = be16_to_cpu(hdr->lrh[1]);
 464         u8 sc, svc_type;
 465         bool is_mcast = false;
 466
 467         if (pkt->rcv_flags & HFI1_HAS_GRH)
 468                 grh = &hdr->u.l.grh;
 469
 470         switch (qp->ibqp.qp_type) {
 471         case IB_QPT_SMI:
 472         case IB_QPT_GSI:
 473         case IB_QPT_UD:
 474                 rlid = be16_to_cpu(hdr->lrh[3]);
 475                 rqpn = be32_to_cpu(ohdr->u.ud.deth[1]) & RVT_QPN_MASK;
 476                 svc_type = IB_CC_SVCTYPE_UD;
 477                 is_mcast = (dlid > be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
 478                         (dlid != be16_to_cpu(IB_LID_PERMISSIVE));
 479                 break;
 480         case IB_QPT_UC:
 481                 rlid = qp->remote_ah_attr.dlid;
 482                 rqpn = qp->remote_qpn;
 483                 svc_type = IB_CC_SVCTYPE_UC;
 484                 break;
 485         case IB_QPT_RC:
 486                 rlid = qp->remote_ah_attr.dlid;
 487                 rqpn = qp->remote_qpn;
 488                 svc_type = IB_CC_SVCTYPE_RC;
 489                 break;
 490         default:
 491                 return;
 492         }
 493
 494         sc = hdr2sc(hdr, pkt->rhf);
 495
 496         bth1 = be32_to_cpu(ohdr->bth[1]);
 497         if (do_cnp && (bth1 & HFI1_FECN_SMASK)) {
 498                 u16 pkey = (u16)be32_to_cpu(ohdr->bth[0]);
 499
 500                 return_cnp(ibp, qp, rqpn, pkey, dlid, rlid, sc, grh);
 501         }
 502
 503         if (!is_mcast && (bth1 & HFI1_BECN_SMASK)) {
 504                 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 505                 u32 lqpn = bth1 & RVT_QPN_MASK;
 506                 u8 sl = ibp->sc_to_sl[sc];
 507
 508                 process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
 509         }
 510
 511 }
 512
 513 struct ps_mdata {
 514         struct hfi1_ctxtdata *rcd;
 515         u32 rsize;
 516         u32 maxcnt;
 517         u32 ps_head;
 518         u32 ps_tail;
 519         u32 ps_seq;
 520 };
 521
 522 static inline void init_ps_mdata(struct ps_mdata *mdata,
 523                                  struct hfi1_packet *packet)
 524 {
 525         struct hfi1_ctxtdata *rcd = packet->rcd;
 526
 527         mdata->rcd = rcd;
 528         mdata->rsize = packet->rsize;
 529         mdata->maxcnt = packet->maxcnt;
 530         mdata->ps_head = packet->rhqoff;
 531
 532         if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
 533                 mdata->ps_tail = get_rcvhdrtail(rcd);
 534                 if (rcd->ctxt == HFI1_CTRL_CTXT)
 535                         mdata->ps_seq = rcd->seq_cnt;
 536                 else
 537                         mdata->ps_seq = 0; /* not used with DMA_RTAIL */
 538         } else {
 539                 mdata->ps_tail = 0; /* used only with DMA_RTAIL*/
 540                 mdata->ps_seq = rcd->seq_cnt;
 541         }
 542 }
 543
 544 static inline int ps_done(struct ps_mdata *mdata, u64 rhf,
 545                           struct hfi1_ctxtdata *rcd)
 546 {
 547         if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL))
 548                 return mdata->ps_head == mdata->ps_tail;
 549         return mdata->ps_seq != rhf_rcv_seq(rhf);
 550 }
 551
 552 static inline int ps_skip(struct ps_mdata *mdata, u64 rhf,
 553                           struct hfi1_ctxtdata *rcd)
 554 {
 555         /*
 556          * Control context can potentially receive an invalid rhf.
 557          * Drop such packets.
 558          */
 559         if ((rcd->ctxt == HFI1_CTRL_CTXT) && (mdata->ps_head != mdata->ps_tail))
 560                 return mdata->ps_seq != rhf_rcv_seq(rhf);
 561
 562         return 0;
 563 }
 564
 565 static inline void update_ps_mdata(struct ps_mdata *mdata,
 566                                    struct hfi1_ctxtdata *rcd)
 567 {
 568         mdata->ps_head += mdata->rsize;
 569         if (mdata->ps_head >= mdata->maxcnt)
 570                 mdata->ps_head = 0;
 571
 572         /* Control context must do seq counting */
 573         if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ||
 574             (rcd->ctxt == HFI1_CTRL_CTXT)) {
 575                 if (++mdata->ps_seq > 13)
 576                         mdata->ps_seq = 1;
 577         }
 578 }
 579
 580 /*
 581  * prescan_rxq - search through the receive queue looking for packets
 582  * containing Excplicit Congestion Notifications (FECNs, or BECNs).
 583  * When an ECN is found, process the Congestion Notification, and toggle
 584  * it off.
 585  * This is declared as a macro to allow quick checking of the port to avoid
 586  * the overhead of a function call if not enabled.
 587  */
 588 #define prescan_rxq(rcd, packet) \
 589         do { \
 590                 if (rcd->ppd->cc_prescan) \
 591                         __prescan_rxq(packet); \
 592         } while (0)
 593 static void __prescan_rxq(struct hfi1_packet *packet)
 594 {
 595         struct hfi1_ctxtdata *rcd = packet->rcd;
 596         struct ps_mdata mdata;
 597
 598         init_ps_mdata(&mdata, packet);
 599
 600         while (1) {
 601                 struct hfi1_devdata *dd = rcd->dd;
 602                 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
 603                 __le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head +
 604                                          dd->rhf_offset;
 605                 struct rvt_qp *qp;
 606                 struct ib_header *hdr;
 607                 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
 608                 u64 rhf = rhf_to_cpu(rhf_addr);
 609                 u32 etype = rhf_rcv_type(rhf), qpn, bth1;
 610                 int is_ecn = 0;
 611                 u8 lnh;
 612
 613                 if (ps_done(&mdata, rhf, rcd))
 614                         break;
 615
 616                 if (ps_skip(&mdata, rhf, rcd))
 617                         goto next;
 618
 619                 if (etype != RHF_RCV_TYPE_IB)
 620                         goto next;
 621
 622                 packet->hdr = hfi1_get_msgheader(dd, rhf_addr);
 623                 hdr = packet->hdr;
 624
 625                 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 626
 627                 if (lnh == HFI1_LRH_BTH) {
 628                         packet->ohdr = &hdr->u.oth;
 629                 } else if (lnh == HFI1_LRH_GRH) {
 630                         packet->ohdr = &hdr->u.l.oth;
 631                         packet->rcv_flags |= HFI1_HAS_GRH;
 632                 } else {
 633                         goto next; /* just in case */
 634                 }
 635
 636                 bth1 = be32_to_cpu(packet->ohdr->bth[1]);
 637                 is_ecn = !!(bth1 & (HFI1_FECN_SMASK | HFI1_BECN_SMASK));
 638
 639                 if (!is_ecn)
 640                         goto next;
 641
 642                 qpn = bth1 & RVT_QPN_MASK;
 643                 rcu_read_lock();
 644                 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qpn);
 645
 646                 if (!qp) {
 647                         rcu_read_unlock();
 648                         goto next;
 649                 }
 650
 651                 process_ecn(qp, packet, true);
 652                 rcu_read_unlock();
 653
 654                 /* turn off BECN, FECN */
 655                 bth1 &= ~(HFI1_FECN_SMASK | HFI1_BECN_SMASK);
 656                 packet->ohdr->bth[1] = cpu_to_be32(bth1);
 657 next:
 658                 update_ps_mdata(&mdata, rcd);
 659         }
 660 }
 661
 662 static void process_rcv_qp_work(struct hfi1_ctxtdata *rcd)
 663 {
 664         struct rvt_qp *qp, *nqp;
 665
 666         /*
 667          * Iterate over all QPs waiting to respond.
 668          * The list won't change since the IRQ is only run on one CPU.
 669          */
 670         list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
 671                 list_del_init(&qp->rspwait);
 672                 if (qp->r_flags & RVT_R_RSP_NAK) {
 673                         qp->r_flags &= ~RVT_R_RSP_NAK;
 674                         hfi1_send_rc_ack(rcd, qp, 0);
 675                 }
 676                 if (qp->r_flags & RVT_R_RSP_SEND) {
 677                         unsigned long flags;
 678
 679                         qp->r_flags &= ~RVT_R_RSP_SEND;
 680                         spin_lock_irqsave(&qp->s_lock, flags);
 681                         if (ib_rvt_state_ops[qp->state] &
 682                                         RVT_PROCESS_OR_FLUSH_SEND)
 683                                 hfi1_schedule_send(qp);
 684                         spin_unlock_irqrestore(&qp->s_lock, flags);
 685                 }
 686                 rvt_put_qp(qp);
 687         }
 688 }
 689
 690 static noinline int max_packet_exceeded(struct hfi1_packet *packet, int thread)
 691 {
 692         if (thread) {
 693                 if ((packet->numpkt & (MAX_PKT_RECV_THREAD - 1)) == 0)
 694                         /* allow defered processing */
 695                         process_rcv_qp_work(packet->rcd);
 696                 cond_resched();
 697                 return RCV_PKT_OK;
 698         } else {
 699                 this_cpu_inc(*packet->rcd->dd->rcv_limit);
 700                 return RCV_PKT_LIMIT;
 701         }
 702 }
 703
 704 static inline int check_max_packet(struct hfi1_packet *packet, int thread)
 705 {
 706         int ret = RCV_PKT_OK;
 707
 708         if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0))
 709                 ret = max_packet_exceeded(packet, thread);
 710         return ret;
 711 }
 712
 713 static noinline int skip_rcv_packet(struct hfi1_packet *packet, int thread)
 714 {
 715         int ret;
 716
 717         /* Set up for the next packet */
 718         packet->rhqoff += packet->rsize;
 719         if (packet->rhqoff >= packet->maxcnt)
 720                 packet->rhqoff = 0;
 721
 722         packet->numpkt++;
 723         ret = check_max_packet(packet, thread);
 724
 725         packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
 726                                      packet->rcd->dd->rhf_offset;
 727         packet->rhf = rhf_to_cpu(packet->rhf_addr);
 728
 729         return ret;
 730 }
 731
 732 static inline int process_rcv_packet(struct hfi1_packet *packet, int thread)
 733 {
 734         int ret;
 735
 736         packet->hdr = hfi1_get_msgheader(packet->rcd->dd,
 737                                          packet->rhf_addr);
 738         packet->hlen = (u8 *)packet->rhf_addr - (u8 *)packet->hdr;
 739         packet->etype = rhf_rcv_type(packet->rhf);
 740         /* total length */
 741         packet->tlen = rhf_pkt_len(packet->rhf); /* in bytes */
 742         /* retrieve eager buffer details */
 743         packet->ebuf = NULL;
 744         if (rhf_use_egr_bfr(packet->rhf)) {
 745                 packet->etail = rhf_egr_index(packet->rhf);
 746                 packet->ebuf = get_egrbuf(packet->rcd, packet->rhf,
 747                                  &packet->updegr);
 748                 /*
 749                  * Prefetch the contents of the eager buffer.  It is
 750                  * OK to send a negative length to prefetch_range().
 751                  * The +2 is the size of the RHF.
 752                  */
 753                 prefetch_range(packet->ebuf,
 754                                packet->tlen - ((packet->rcd->rcvhdrqentsize -
 755                                                (rhf_hdrq_offset(packet->rhf)
 756                                                 + 2)) * 4));
 757         }
 758
 759         /*
 760          * Call a type specific handler for the packet. We
 761          * should be able to trust that etype won't be beyond
 762          * the range of valid indexes. If so something is really
 763          * wrong and we can probably just let things come
 764          * crashing down. There is no need to eat another
 765          * comparison in this performance critical code.
 766          */
 767         packet->rcd->dd->rhf_rcv_function_map[packet->etype](packet);
 768         packet->numpkt++;
 769
 770         /* Set up for the next packet */
 771         packet->rhqoff += packet->rsize;
 772         if (packet->rhqoff >= packet->maxcnt)
 773                 packet->rhqoff = 0;
 774
 775         ret = check_max_packet(packet, thread);
 776
 777         packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
 778                                       packet->rcd->dd->rhf_offset;
 779         packet->rhf = rhf_to_cpu(packet->rhf_addr);
 780
 781         return ret;
 782 }
 783
 784 static inline void process_rcv_update(int last, struct hfi1_packet *packet)
 785 {
 786         /*
 787          * Update head regs etc., every 16 packets, if not last pkt,
 788          * to help prevent rcvhdrq overflows, when many packets
 789          * are processed and queue is nearly full.
 790          * Don't request an interrupt for intermediate updates.
 791          */
 792         if (!last && !(packet->numpkt & 0xf)) {
 793                 update_usrhead(packet->rcd, packet->rhqoff, packet->updegr,
 794                                packet->etail, 0, 0);
 795                 packet->updegr = 0;
 796         }
 797         packet->rcv_flags = 0;
 798 }
 799
 800 static inline void finish_packet(struct hfi1_packet *packet)
 801 {
 802         /*
 803          * Nothing we need to free for the packet.
 804          *
 805          * The only thing we need to do is a final update and call for an
 806          * interrupt
 807          */
 808         update_usrhead(packet->rcd, packet->rcd->head, packet->updegr,
 809                        packet->etail, rcv_intr_dynamic, packet->numpkt);
 810 }
 811
 812 /*
 813  * Handle receive interrupts when using the no dma rtail option.
 814  */
 815 int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata *rcd, int thread)
 816 {
 817         u32 seq;
 818         int last = RCV_PKT_OK;
 819         struct hfi1_packet packet;
 820
 821         init_packet(rcd, &packet);
 822         seq = rhf_rcv_seq(packet.rhf);
 823         if (seq != rcd->seq_cnt) {
 824                 last = RCV_PKT_DONE;
 825                 goto bail;
 826         }
 827
 828         prescan_rxq(rcd, &packet);
 829
 830         while (last == RCV_PKT_OK) {
 831                 last = process_rcv_packet(&packet, thread);
 832                 seq = rhf_rcv_seq(packet.rhf);
 833                 if (++rcd->seq_cnt > 13)
 834                         rcd->seq_cnt = 1;
 835                 if (seq != rcd->seq_cnt)
 836                         last = RCV_PKT_DONE;
 837                 process_rcv_update(last, &packet);
 838         }
 839         process_rcv_qp_work(rcd);
 840         rcd->head = packet.rhqoff;
 841 bail:
 842         finish_packet(&packet);
 843         return last;
 844 }
 845
 846 int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata *rcd, int thread)
 847 {
 848         u32 hdrqtail;
 849         int last = RCV_PKT_OK;
 850         struct hfi1_packet packet;
 851
 852         init_packet(rcd, &packet);
 853         hdrqtail = get_rcvhdrtail(rcd);
 854         if (packet.rhqoff == hdrqtail) {
 855                 last = RCV_PKT_DONE;
 856                 goto bail;
 857         }
 858         smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
 859
 860         prescan_rxq(rcd, &packet);
 861
 862         while (last == RCV_PKT_OK) {
 863                 last = process_rcv_packet(&packet, thread);
 864                 if (packet.rhqoff == hdrqtail)
 865                         last = RCV_PKT_DONE;
 866                 process_rcv_update(last, &packet);
 867         }
 868         process_rcv_qp_work(rcd);
 869         rcd->head = packet.rhqoff;
 870 bail:
 871         finish_packet(&packet);
 872         return last;
 873 }
 874
 875 static inline void set_all_nodma_rtail(struct hfi1_devdata *dd)
 876 {
 877         int i;
 878
 879         for (i = HFI1_CTRL_CTXT + 1; i < dd->first_user_ctxt; i++)
 880                 dd->rcd[i]->do_interrupt =
 881                         &handle_receive_interrupt_nodma_rtail;
 882 }
 883
 884 static inline void set_all_dma_rtail(struct hfi1_devdata *dd)
 885 {
 886         int i;
 887
 888         for (i = HFI1_CTRL_CTXT + 1; i < dd->first_user_ctxt; i++)
 889                 dd->rcd[i]->do_interrupt =
 890                         &handle_receive_interrupt_dma_rtail;
 891 }
 892
 893 void set_all_slowpath(struct hfi1_devdata *dd)
 894 {
 895         int i;
 896
 897         /* HFI1_CTRL_CTXT must always use the slow path interrupt handler */
 898         for (i = HFI1_CTRL_CTXT + 1; i < dd->first_user_ctxt; i++)
 899                 dd->rcd[i]->do_interrupt = &handle_receive_interrupt;
 900 }
 901
 902 static inline int set_armed_to_active(struct hfi1_ctxtdata *rcd,
 903                                       struct hfi1_packet *packet,
 904                                       struct hfi1_devdata *dd)
 905 {
 906         struct work_struct *lsaw = &rcd->ppd->linkstate_active_work;
 907         struct ib_header *hdr = hfi1_get_msgheader(packet->rcd->dd,
 908                                                    packet->rhf_addr);
 909         u8 etype = rhf_rcv_type(packet->rhf);
 910
 911         if (etype == RHF_RCV_TYPE_IB && hdr2sc(hdr, packet->rhf) != 0xf) {
 912                 int hwstate = read_logical_state(dd);
 913
 914                 if (hwstate != LSTATE_ACTIVE) {
 915                         dd_dev_info(dd, "Unexpected link state %d\n", hwstate);
 916                         return 0;
 917                 }
 918
 919                 queue_work(rcd->ppd->hfi1_wq, lsaw);
 920                 return 1;
 921         }
 922         return 0;
 923 }
 924
 925 /*
 926  * handle_receive_interrupt - receive a packet
 927  * @rcd: the context
 928  *
 929  * Called from interrupt handler for errors or receive interrupt.
 930  * This is the slow path interrupt handler.
 931  */
 932 int handle_receive_interrupt(struct hfi1_ctxtdata *rcd, int thread)
 933 {
 934         struct hfi1_devdata *dd = rcd->dd;
 935         u32 hdrqtail;
 936         int needset, last = RCV_PKT_OK;
 937         struct hfi1_packet packet;
 938         int skip_pkt = 0;
 939
 940         /* Control context will always use the slow path interrupt handler */
 941         needset = (rcd->ctxt == HFI1_CTRL_CTXT) ? 0 : 1;
 942
 943         init_packet(rcd, &packet);
 944
 945         if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
 946                 u32 seq = rhf_rcv_seq(packet.rhf);
 947
 948                 if (seq != rcd->seq_cnt) {
 949                         last = RCV_PKT_DONE;
 950                         goto bail;
 951                 }
 952                 hdrqtail = 0;
 953         } else {
 954                 hdrqtail = get_rcvhdrtail(rcd);
 955                 if (packet.rhqoff == hdrqtail) {
 956                         last = RCV_PKT_DONE;
 957                         goto bail;
 958                 }
 959                 smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
 960
 961                 /*
 962                  * Control context can potentially receive an invalid
 963                  * rhf. Drop such packets.
 964                  */
 965                 if (rcd->ctxt == HFI1_CTRL_CTXT) {
 966                         u32 seq = rhf_rcv_seq(packet.rhf);
 967
 968                         if (seq != rcd->seq_cnt)
 969                                 skip_pkt = 1;
 970                 }
 971         }
 972
 973         prescan_rxq(rcd, &packet);
 974
 975         while (last == RCV_PKT_OK) {
 976                 if (unlikely(dd->do_drop &&
 977                              atomic_xchg(&dd->drop_packet, DROP_PACKET_OFF) ==
 978                              DROP_PACKET_ON)) {
 979                         dd->do_drop = 0;
 980
 981                         /* On to the next packet */
 982                         packet.rhqoff += packet.rsize;
 983                         packet.rhf_addr = (__le32 *)rcd->rcvhdrq +
 984                                           packet.rhqoff +
 985                                           dd->rhf_offset;
 986                         packet.rhf = rhf_to_cpu(packet.rhf_addr);
 987
 988                 } else if (skip_pkt) {
 989                         last = skip_rcv_packet(&packet, thread);
 990                         skip_pkt = 0;
 991                 } else {
 992                         /* Auto activate link on non-SC15 packet receive */
 993                         if (unlikely(rcd->ppd->host_link_state ==
 994                                      HLS_UP_ARMED) &&
 995                             set_armed_to_active(rcd, &packet, dd))
 996                                 goto bail;
 997                         last = process_rcv_packet(&packet, thread);
 998                 }
 999
1000                 if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
1001                         u32 seq = rhf_rcv_seq(packet.rhf);
1002
1003                         if (++rcd->seq_cnt > 13)
1004                                 rcd->seq_cnt = 1;
1005                         if (seq != rcd->seq_cnt)
1006                                 last = RCV_PKT_DONE;
1007                         if (needset) {
1008                                 dd_dev_info(dd, "Switching to NO_DMA_RTAIL\n");
1009                                 set_all_nodma_rtail(dd);
1010                                 needset = 0;
1011                         }
1012                 } else {
1013                         if (packet.rhqoff == hdrqtail)
1014                                 last = RCV_PKT_DONE;
1015                         /*
1016                          * Control context can potentially receive an invalid
1017                          * rhf. Drop such packets.
1018                          */
1019                         if (rcd->ctxt == HFI1_CTRL_CTXT) {
1020                                 u32 seq = rhf_rcv_seq(packet.rhf);
1021
1022                                 if (++rcd->seq_cnt > 13)
1023                                         rcd->seq_cnt = 1;
1024                                 if (!last && (seq != rcd->seq_cnt))
1025                                         skip_pkt = 1;
1026                         }
1027
1028                         if (needset) {
1029                                 dd_dev_info(dd,
1030                                             "Switching to DMA_RTAIL\n");
1031                                 set_all_dma_rtail(dd);
1032                                 needset = 0;
1033                         }
1034                 }
1035
1036                 process_rcv_update(last, &packet);
1037         }
1038
1039         process_rcv_qp_work(rcd);
1040         rcd->head = packet.rhqoff;
1041
1042 bail:
1043         /*
1044          * Always write head at end, and setup rcv interrupt, even
1045          * if no packets were processed.
1046          */
1047         finish_packet(&packet);
1048         return last;
1049 }
1050
1051 /*
1052  * We may discover in the interrupt that the hardware link state has
1053  * changed from ARMED to ACTIVE (due to the arrival of a non-SC15 packet),
1054  * and we need to update the driver's notion of the link state.  We cannot
1055  * run set_link_state from interrupt context, so we queue this function on
1056  * a workqueue.
1057  *
1058  * We delay the regular interrupt processing until after the state changes
1059  * so that the link will be in the correct state by the time any application
1060  * we wake up attempts to send a reply to any message it received.
1061  * (Subsequent receive interrupts may possibly force the wakeup before we
1062  * update the link state.)
1063  *
1064  * The rcd is freed in hfi1_free_ctxtdata after hfi1_postinit_cleanup invokes
1065  * dd->f_cleanup(dd) to disable the interrupt handler and flush workqueues,
1066  * so we're safe from use-after-free of the rcd.
1067  */
1068 void receive_interrupt_work(struct work_struct *work)
1069 {
1070         struct hfi1_pportdata *ppd = container_of(work, struct hfi1_pportdata,
1071                                                   linkstate_active_work);
1072         struct hfi1_devdata *dd = ppd->dd;
1073         int i;
1074
1075         /* Received non-SC15 packet implies neighbor_normal */
1076         ppd->neighbor_normal = 1;
1077         set_link_state(ppd, HLS_UP_ACTIVE);
1078
1079         /*
1080          * Interrupt all kernel contexts that could have had an
1081          * interrupt during auto activation.
1082          */
1083         for (i = HFI1_CTRL_CTXT; i < dd->first_user_ctxt; i++)
1084                 force_recv_intr(dd->rcd[i]);
1085 }
1086
1087 /*
1088  * Convert a given MTU size to the on-wire MAD packet enumeration.
1089  * Return -1 if the size is invalid.
1090  */
1091 int mtu_to_enum(u32 mtu, int default_if_bad)
1092 {
1093         switch (mtu) {
1094         case     0: return OPA_MTU_0;
1095         case   256: return OPA_MTU_256;
1096         case   512: return OPA_MTU_512;
1097         case  1024: return OPA_MTU_1024;
1098         case  2048: return OPA_MTU_2048;
1099         case  4096: return OPA_MTU_4096;
1100         case  8192: return OPA_MTU_8192;
1101         case 10240: return OPA_MTU_10240;
1102         }
1103         return default_if_bad;
1104 }
1105
1106 u16 enum_to_mtu(int mtu)
1107 {
1108         switch (mtu) {
1109         case OPA_MTU_0:     return 0;
1110         case OPA_MTU_256:   return 256;
1111         case OPA_MTU_512:   return 512;
1112         case OPA_MTU_1024:  return 1024;
1113         case OPA_MTU_2048:  return 2048;
1114         case OPA_MTU_4096:  return 4096;
1115         case OPA_MTU_8192:  return 8192;
1116         case OPA_MTU_10240: return 10240;
1117         default: return 0xffff;
1118         }
1119 }
1120
1121 /*
1122  * set_mtu - set the MTU
1123  * @ppd: the per port data
1124  *
1125  * We can handle "any" incoming size, the issue here is whether we
1126  * need to restrict our outgoing size.  We do not deal with what happens
1127  * to programs that are already running when the size changes.
1128  */
1129 int set_mtu(struct hfi1_pportdata *ppd)
1130 {
1131         struct hfi1_devdata *dd = ppd->dd;
1132         int i, drain, ret = 0, is_up = 0;
1133
1134         ppd->ibmtu = 0;
1135         for (i = 0; i < ppd->vls_supported; i++)
1136                 if (ppd->ibmtu < dd->vld[i].mtu)
1137                         ppd->ibmtu = dd->vld[i].mtu;
1138         ppd->ibmaxlen = ppd->ibmtu + lrh_max_header_bytes(ppd->dd);
1139
1140         mutex_lock(&ppd->hls_lock);
1141         if (ppd->host_link_state == HLS_UP_INIT ||
1142             ppd->host_link_state == HLS_UP_ARMED ||
1143             ppd->host_link_state == HLS_UP_ACTIVE)
1144                 is_up = 1;
1145
1146         drain = !is_ax(dd) && is_up;
1147
1148         if (drain)
1149                 /*
1150                  * MTU is specified per-VL. To ensure that no packet gets
1151                  * stuck (due, e.g., to the MTU for the packet's VL being
1152                  * reduced), empty the per-VL FIFOs before adjusting MTU.
1153                  */
1154                 ret = stop_drain_data_vls(dd);
1155
1156         if (ret) {
1157                 dd_dev_err(dd, "%s: cannot stop/drain VLs - refusing to change per-VL MTUs\n",
1158                            __func__);
1159                 goto err;
1160         }
1161
1162         hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_MTU, 0);
1163
1164         if (drain)
1165                 open_fill_data_vls(dd); /* reopen all VLs */
1166
1167 err:
1168         mutex_unlock(&ppd->hls_lock);
1169
1170         return ret;
1171 }
1172
1173 int hfi1_set_lid(struct hfi1_pportdata *ppd, u32 lid, u8 lmc)
1174 {
1175         struct hfi1_devdata *dd = ppd->dd;
1176
1177         ppd->lid = lid;
1178         ppd->lmc = lmc;
1179         hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_LIDLMC, 0);
1180
1181         dd_dev_info(dd, "port %u: got a lid: 0x%x\n", ppd->port, lid);
1182
1183         return 0;
1184 }
1185
1186 void shutdown_led_override(struct hfi1_pportdata *ppd)
1187 {
1188         struct hfi1_devdata *dd = ppd->dd;
1189
1190         /*
1191          * This pairs with the memory barrier in hfi1_start_led_override to
1192          * ensure that we read the correct state of LED beaconing represented
1193          * by led_override_timer_active
1194          */
1195         smp_rmb();
1196         if (atomic_read(&ppd->led_override_timer_active)) {
1197                 del_timer_sync(&ppd->led_override_timer);
1198                 atomic_set(&ppd->led_override_timer_active, 0);
1199                 /* Ensure the atomic_set is visible to all CPUs */
1200                 smp_wmb();
1201         }
1202
1203         /* Hand control of the LED to the DC for normal operation */
1204         write_csr(dd, DCC_CFG_LED_CNTRL, 0);
1205 }
1206
1207 static void run_led_override(unsigned long opaque)
1208 {
1209         struct hfi1_pportdata *ppd = (struct hfi1_pportdata *)opaque;
1210         struct hfi1_devdata *dd = ppd->dd;
1211         unsigned long timeout;
1212         int phase_idx;
1213
1214         if (!(dd->flags & HFI1_INITTED))
1215                 return;
1216
1217         phase_idx = ppd->led_override_phase & 1;
1218
1219         setextled(dd, phase_idx);
1220
1221         timeout = ppd->led_override_vals[phase_idx];
1222
1223         /* Set up for next phase */
1224         ppd->led_override_phase = !ppd->led_override_phase;
1225
1226         mod_timer(&ppd->led_override_timer, jiffies + timeout);
1227 }
1228
1229 /*
1230  * To have the LED blink in a particular pattern, provide timeon and timeoff
1231  * in milliseconds.
1232  * To turn off custom blinking and return to normal operation, use
1233  * shutdown_led_override()
1234  */
1235 void hfi1_start_led_override(struct hfi1_pportdata *ppd, unsigned int timeon,
1236                              unsigned int timeoff)
1237 {
1238         if (!(ppd->dd->flags & HFI1_INITTED))
1239                 return;
1240
1241         /* Convert to jiffies for direct use in timer */
1242         ppd->led_override_vals[0] = msecs_to_jiffies(timeoff);
1243         ppd->led_override_vals[1] = msecs_to_jiffies(timeon);
1244
1245         /* Arbitrarily start from LED on phase */
1246         ppd->led_override_phase = 1;
1247
1248         /*
1249          * If the timer has not already been started, do so. Use a "quick"
1250          * timeout so the handler will be called soon to look at our request.
1251          */
1252         if (!timer_pending(&ppd->led_override_timer)) {
1253                 setup_timer(&ppd->led_override_timer, run_led_override,
1254                             (unsigned long)ppd);
1255                 ppd->led_override_timer.expires = jiffies + 1;
1256                 add_timer(&ppd->led_override_timer);
1257                 atomic_set(&ppd->led_override_timer_active, 1);
1258                 /* Ensure the atomic_set is visible to all CPUs */
1259                 smp_wmb();
1260         }
1261 }
1262
1263 /**
1264  * hfi1_reset_device - reset the chip if possible
1265  * @unit: the device to reset
1266  *
1267  * Whether or not reset is successful, we attempt to re-initialize the chip
1268  * (that is, much like a driver unload/reload).  We clear the INITTED flag
1269  * so that the various entry points will fail until we reinitialize.  For
1270  * now, we only allow this if no user contexts are open that use chip resources
1271  */
1272 int hfi1_reset_device(int unit)
1273 {
1274         int ret, i;
1275         struct hfi1_devdata *dd = hfi1_lookup(unit);
1276         struct hfi1_pportdata *ppd;
1277         unsigned long flags;
1278         int pidx;
1279
1280         if (!dd) {
1281                 ret = -ENODEV;
1282                 goto bail;
1283         }
1284
1285         dd_dev_info(dd, "Reset on unit %u requested\n", unit);
1286
1287         if (!dd->kregbase || !(dd->flags & HFI1_PRESENT)) {
1288                 dd_dev_info(dd,
1289                             "Invalid unit number %u or not initialized or not present\n",
1290                             unit);
1291                 ret = -ENXIO;
1292                 goto bail;
1293         }
1294
1295         spin_lock_irqsave(&dd->uctxt_lock, flags);
1296         if (dd->rcd)
1297                 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
1298                         if (!dd->rcd[i] || !dd->rcd[i]->cnt)
1299                                 continue;
1300                         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1301                         ret = -EBUSY;
1302                         goto bail;
1303                 }
1304         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1305
1306         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1307                 ppd = dd->pport + pidx;
1308
1309                 shutdown_led_override(ppd);
1310         }
1311         if (dd->flags & HFI1_HAS_SEND_DMA)
1312                 sdma_exit(dd);
1313
1314         hfi1_reset_cpu_counters(dd);
1315
1316         ret = hfi1_init(dd, 1);
1317
1318         if (ret)
1319                 dd_dev_err(dd,
1320                            "Reinitialize unit %u after reset failed with %d\n",
1321                            unit, ret);
1322         else
1323                 dd_dev_info(dd, "Reinitialized unit %u after resetting\n",
1324                             unit);
1325
1326 bail:
1327         return ret;
1328 }
1329
1330 void handle_eflags(struct hfi1_packet *packet)
1331 {
1332         struct hfi1_ctxtdata *rcd = packet->rcd;
1333         u32 rte = rhf_rcv_type_err(packet->rhf);
1334
1335         rcv_hdrerr(rcd, rcd->ppd, packet);
1336         if (rhf_err_flags(packet->rhf))
1337                 dd_dev_err(rcd->dd,
1338                            "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s%s] rte 0x%x\n",
1339                            rcd->ctxt, packet->rhf,
1340                            packet->rhf & RHF_K_HDR_LEN_ERR ? "k_hdr_len " : "",
1341                            packet->rhf & RHF_DC_UNC_ERR ? "dc_unc " : "",
1342                            packet->rhf & RHF_DC_ERR ? "dc " : "",
1343                            packet->rhf & RHF_TID_ERR ? "tid " : "",
1344                            packet->rhf & RHF_LEN_ERR ? "len " : "",
1345                            packet->rhf & RHF_ECC_ERR ? "ecc " : "",
1346                            packet->rhf & RHF_VCRC_ERR ? "vcrc " : "",
1347                            packet->rhf & RHF_ICRC_ERR ? "icrc " : "",
1348                            rte);
1349 }
1350
1351 /*
1352  * The following functions are called by the interrupt handler. They are type
1353  * specific handlers for each packet type.
1354  */
1355 int process_receive_ib(struct hfi1_packet *packet)
1356 {
1357         trace_hfi1_rcvhdr(packet->rcd->ppd->dd,
1358                           packet->rcd->ctxt,
1359                           rhf_err_flags(packet->rhf),
1360                           RHF_RCV_TYPE_IB,
1361                           packet->hlen,
1362                           packet->tlen,
1363                           packet->updegr,
1364                           rhf_egr_index(packet->rhf));
1365
1366         if (unlikely(rhf_err_flags(packet->rhf))) {
1367                 handle_eflags(packet);
1368                 return RHF_RCV_CONTINUE;
1369         }
1370
1371         hfi1_ib_rcv(packet);
1372         return RHF_RCV_CONTINUE;
1373 }
1374
1375 int process_receive_bypass(struct hfi1_packet *packet)
1376 {
1377         struct hfi1_devdata *dd = packet->rcd->dd;
1378
1379         if (unlikely(rhf_err_flags(packet->rhf)))
1380                 handle_eflags(packet);
1381
1382         dd_dev_err(dd,
1383                    "Bypass packets are not supported in normal operation. Dropping\n");
1384         incr_cntr64(&dd->sw_rcv_bypass_packet_errors);
1385         if (!(dd->err_info_rcvport.status_and_code & OPA_EI_STATUS_SMASK)) {
1386                 u64 *flits = packet->ebuf;
1387
1388                 if (flits && !(packet->rhf & RHF_LEN_ERR)) {
1389                         dd->err_info_rcvport.packet_flit1 = flits[0];
1390                         dd->err_info_rcvport.packet_flit2 =
1391                                 packet->tlen > sizeof(flits[0]) ? flits[1] : 0;
1392                 }
1393                 dd->err_info_rcvport.status_and_code |=
1394                         (OPA_EI_STATUS_SMASK | BAD_L2_ERR);
1395         }
1396         return RHF_RCV_CONTINUE;
1397 }
1398
1399 int process_receive_error(struct hfi1_packet *packet)
1400 {
1401         handle_eflags(packet);
1402
1403         if (unlikely(rhf_err_flags(packet->rhf)))
1404                 dd_dev_err(packet->rcd->dd,
1405                            "Unhandled error packet received. Dropping.\n");
1406
1407         return RHF_RCV_CONTINUE;
1408 }
1409
1410 int kdeth_process_expected(struct hfi1_packet *packet)
1411 {
1412         if (unlikely(rhf_err_flags(packet->rhf)))
1413                 handle_eflags(packet);
1414
1415         dd_dev_err(packet->rcd->dd,
1416                    "Unhandled expected packet received. Dropping.\n");
1417         return RHF_RCV_CONTINUE;
1418 }
1419
1420 int kdeth_process_eager(struct hfi1_packet *packet)
1421 {
1422         if (unlikely(rhf_err_flags(packet->rhf)))
1423                 handle_eflags(packet);
1424
1425         dd_dev_err(packet->rcd->dd,
1426                    "Unhandled eager packet received. Dropping.\n");
1427         return RHF_RCV_CONTINUE;
1428 }
1429
1430 int process_receive_invalid(struct hfi1_packet *packet)
1431 {
1432         dd_dev_err(packet->rcd->dd, "Invalid packet type %d. Dropping\n",
1433                    rhf_rcv_type(packet->rhf));
1434         return RHF_RCV_CONTINUE;
1435 }