drivers/infiniband/hw/hfi1/driver.c

   1 /*
   2  * Copyright(c) 2015-2017 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 #include "debugfs.h"
  63 #include "vnic.h"
  64
  65 #undef pr_fmt
  66 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
  67
  68 /*
  69  * The size has to be longer than this string, so we can append
  70  * board/chip information to it in the initialization code.
  71  */
  72 const char ib_hfi1_version[] = HFI1_DRIVER_VERSION "\n";
  73
  74 DEFINE_SPINLOCK(hfi1_devs_lock);
  75 LIST_HEAD(hfi1_dev_list);
  76 DEFINE_MUTEX(hfi1_mutex);       /* general driver use */
  77
  78 unsigned int hfi1_max_mtu = HFI1_DEFAULT_MAX_MTU;
  79 module_param_named(max_mtu, hfi1_max_mtu, uint, S_IRUGO);
  80 MODULE_PARM_DESC(max_mtu, "Set max MTU bytes, default is " __stringify(
  81                  HFI1_DEFAULT_MAX_MTU));
  82
  83 unsigned int hfi1_cu = 1;
  84 module_param_named(cu, hfi1_cu, uint, S_IRUGO);
  85 MODULE_PARM_DESC(cu, "Credit return units");
  86
  87 unsigned long hfi1_cap_mask = HFI1_CAP_MASK_DEFAULT;
  88 static int hfi1_caps_set(const char *val, const struct kernel_param *kp);
  89 static int hfi1_caps_get(char *buffer, const struct kernel_param *kp);
  90 static const struct kernel_param_ops cap_ops = {
  91         .set = hfi1_caps_set,
  92         .get = hfi1_caps_get
  93 };
  94 module_param_cb(cap_mask, &cap_ops, &hfi1_cap_mask, S_IWUSR | S_IRUGO);
  95 MODULE_PARM_DESC(cap_mask, "Bit mask of enabled/disabled HW features");
  96
  97 MODULE_LICENSE("Dual BSD/GPL");
  98 MODULE_DESCRIPTION("Intel Omni-Path Architecture driver");
  99
 100 /*
 101  * MAX_PKT_RCV is the max # if packets processed per receive interrupt.
 102  */
 103 #define MAX_PKT_RECV 64
 104 /*
 105  * MAX_PKT_THREAD_RCV is the max # of packets processed before
 106  * the qp_wait_list queue is flushed.
 107  */
 108 #define MAX_PKT_RECV_THREAD (MAX_PKT_RECV * 4)
 109 #define EGR_HEAD_UPDATE_THRESHOLD 16
 110
 111 struct hfi1_ib_stats hfi1_stats;
 112
 113 static int hfi1_caps_set(const char *val, const struct kernel_param *kp)
 114 {
 115         int ret = 0;
 116         unsigned long *cap_mask_ptr = (unsigned long *)kp->arg,
 117                 cap_mask = *cap_mask_ptr, value, diff,
 118                 write_mask = ((HFI1_CAP_WRITABLE_MASK << HFI1_CAP_USER_SHIFT) |
 119                               HFI1_CAP_WRITABLE_MASK);
 120
 121         ret = kstrtoul(val, 0, &value);
 122         if (ret) {
 123                 pr_warn("Invalid module parameter value for 'cap_mask'\n");
 124                 goto done;
 125         }
 126         /* Get the changed bits (except the locked bit) */
 127         diff = value ^ (cap_mask & ~HFI1_CAP_LOCKED_SMASK);
 128
 129         /* Remove any bits that are not allowed to change after driver load */
 130         if (HFI1_CAP_LOCKED() && (diff & ~write_mask)) {
 131                 pr_warn("Ignoring non-writable capability bits %#lx\n",
 132                         diff & ~write_mask);
 133                 diff &= write_mask;
 134         }
 135
 136         /* Mask off any reserved bits */
 137         diff &= ~HFI1_CAP_RESERVED_MASK;
 138         /* Clear any previously set and changing bits */
 139         cap_mask &= ~diff;
 140         /* Update the bits with the new capability */
 141         cap_mask |= (value & diff);
 142         /* Check for any kernel/user restrictions */
 143         diff = (cap_mask & (HFI1_CAP_MUST_HAVE_KERN << HFI1_CAP_USER_SHIFT)) ^
 144                 ((cap_mask & HFI1_CAP_MUST_HAVE_KERN) << HFI1_CAP_USER_SHIFT);
 145         cap_mask &= ~diff;
 146         /* Set the bitmask to the final set */
 147         *cap_mask_ptr = cap_mask;
 148 done:
 149         return ret;
 150 }
 151
 152 static int hfi1_caps_get(char *buffer, const struct kernel_param *kp)
 153 {
 154         unsigned long cap_mask = *(unsigned long *)kp->arg;
 155
 156         cap_mask &= ~HFI1_CAP_LOCKED_SMASK;
 157         cap_mask |= ((cap_mask & HFI1_CAP_K2U) << HFI1_CAP_USER_SHIFT);
 158
 159         return scnprintf(buffer, PAGE_SIZE, "0x%lx", cap_mask);
 160 }
 161
 162 struct pci_dev *get_pci_dev(struct rvt_dev_info *rdi)
 163 {
 164         struct hfi1_ibdev *ibdev = container_of(rdi, struct hfi1_ibdev, rdi);
 165         struct hfi1_devdata *dd = container_of(ibdev,
 166                                                struct hfi1_devdata, verbs_dev);
 167         return dd->pcidev;
 168 }
 169
 170 /*
 171  * Return count of units with at least one port ACTIVE.
 172  */
 173 int hfi1_count_active_units(void)
 174 {
 175         struct hfi1_devdata *dd;
 176         struct hfi1_pportdata *ppd;
 177         unsigned long flags;
 178         int pidx, nunits_active = 0;
 179
 180         spin_lock_irqsave(&hfi1_devs_lock, flags);
 181         list_for_each_entry(dd, &hfi1_dev_list, list) {
 182                 if (!(dd->flags & HFI1_PRESENT) || !dd->kregbase1)
 183                         continue;
 184                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 185                         ppd = dd->pport + pidx;
 186                         if (ppd->lid && ppd->linkup) {
 187                                 nunits_active++;
 188                                 break;
 189                         }
 190                 }
 191         }
 192         spin_unlock_irqrestore(&hfi1_devs_lock, flags);
 193         return nunits_active;
 194 }
 195
 196 /*
 197  * Get address of eager buffer from it's index (allocated in chunks, not
 198  * contiguous).
 199  */
 200 static inline void *get_egrbuf(const struct hfi1_ctxtdata *rcd, u64 rhf,
 201                                u8 *update)
 202 {
 203         u32 idx = rhf_egr_index(rhf), offset = rhf_egr_buf_offset(rhf);
 204
 205         *update |= !(idx & (rcd->egrbufs.threshold - 1)) && !offset;
 206         return (void *)(((u64)(rcd->egrbufs.rcvtids[idx].addr)) +
 207                         (offset * RCV_BUF_BLOCK_SIZE));
 208 }
 209
 210 static inline void *hfi1_get_header(struct hfi1_devdata *dd,
 211                                     __le32 *rhf_addr)
 212 {
 213         u32 offset = rhf_hdrq_offset(rhf_to_cpu(rhf_addr));
 214
 215         return (void *)(rhf_addr - dd->rhf_offset + offset);
 216 }
 217
 218 static inline struct ib_header *hfi1_get_msgheader(struct hfi1_devdata *dd,
 219                                                    __le32 *rhf_addr)
 220 {
 221         return (struct ib_header *)hfi1_get_header(dd, rhf_addr);
 222 }
 223
 224 static inline struct hfi1_16b_header
 225                 *hfi1_get_16B_header(struct hfi1_devdata *dd,
 226                                      __le32 *rhf_addr)
 227 {
 228         return (struct hfi1_16b_header *)hfi1_get_header(dd, rhf_addr);
 229 }
 230
 231 /*
 232  * Validate and encode the a given RcvArray Buffer size.
 233  * The function will check whether the given size falls within
 234  * allowed size ranges for the respective type and, optionally,
 235  * return the proper encoding.
 236  */
 237 int hfi1_rcvbuf_validate(u32 size, u8 type, u16 *encoded)
 238 {
 239         if (unlikely(!PAGE_ALIGNED(size)))
 240                 return 0;
 241         if (unlikely(size < MIN_EAGER_BUFFER))
 242                 return 0;
 243         if (size >
 244             (type == PT_EAGER ? MAX_EAGER_BUFFER : MAX_EXPECTED_BUFFER))
 245                 return 0;
 246         if (encoded)
 247                 *encoded = ilog2(size / PAGE_SIZE) + 1;
 248         return 1;
 249 }
 250
 251 static void rcv_hdrerr(struct hfi1_ctxtdata *rcd, struct hfi1_pportdata *ppd,
 252                        struct hfi1_packet *packet)
 253 {
 254         struct ib_header *rhdr = packet->hdr;
 255         u32 rte = rhf_rcv_type_err(packet->rhf);
 256         u32 mlid_base;
 257         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
 258         struct hfi1_devdata *dd = ppd->dd;
 259         struct hfi1_ibdev *verbs_dev = &dd->verbs_dev;
 260         struct rvt_dev_info *rdi = &verbs_dev->rdi;
 261
 262         if ((packet->rhf & RHF_DC_ERR) &&
 263             hfi1_dbg_fault_suppress_err(verbs_dev))
 264                 return;
 265
 266         if (packet->rhf & (RHF_VCRC_ERR | RHF_ICRC_ERR))
 267                 return;
 268
 269         if (packet->etype == RHF_RCV_TYPE_BYPASS) {
 270                 goto drop;
 271         } else {
 272                 u8 lnh = ib_get_lnh(rhdr);
 273
 274                 mlid_base = be16_to_cpu(IB_MULTICAST_LID_BASE);
 275                 if (lnh == HFI1_LRH_BTH) {
 276                         packet->ohdr = &rhdr->u.oth;
 277                 } else if (lnh == HFI1_LRH_GRH) {
 278                         packet->ohdr = &rhdr->u.l.oth;
 279                         packet->grh = &rhdr->u.l.grh;
 280                 } else {
 281                         goto drop;
 282                 }
 283         }
 284
 285         if (packet->rhf & RHF_TID_ERR) {
 286                 /* For TIDERR and RC QPs preemptively schedule a NAK */
 287                 u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */
 288                 u32 dlid = ib_get_dlid(rhdr);
 289                 u32 qp_num;
 290
 291                 /* Sanity check packet */
 292                 if (tlen < 24)
 293                         goto drop;
 294
 295                 /* Check for GRH */
 296                 if (packet->grh) {
 297                         u32 vtf;
 298                         struct ib_grh *grh = packet->grh;
 299
 300                         if (grh->next_hdr != IB_GRH_NEXT_HDR)
 301                                 goto drop;
 302                         vtf = be32_to_cpu(grh->version_tclass_flow);
 303                         if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
 304                                 goto drop;
 305                 }
 306
 307                 /* Get the destination QP number. */
 308                 qp_num = ib_bth_get_qpn(packet->ohdr);
 309                 if (dlid < mlid_base) {
 310                         struct rvt_qp *qp;
 311                         unsigned long flags;
 312
 313                         rcu_read_lock();
 314                         qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
 315                         if (!qp) {
 316                                 rcu_read_unlock();
 317                                 goto drop;
 318                         }
 319
 320                         /*
 321                          * Handle only RC QPs - for other QP types drop error
 322                          * packet.
 323                          */
 324                         spin_lock_irqsave(&qp->r_lock, flags);
 325
 326                         /* Check for valid receive state. */
 327                         if (!(ib_rvt_state_ops[qp->state] &
 328                               RVT_PROCESS_RECV_OK)) {
 329                                 ibp->rvp.n_pkt_drops++;
 330                         }
 331
 332                         switch (qp->ibqp.qp_type) {
 333                         case IB_QPT_RC:
 334                                 hfi1_rc_hdrerr(rcd, packet, qp);
 335                                 break;
 336                         default:
 337                                 /* For now don't handle any other QP types */
 338                                 break;
 339                         }
 340
 341                         spin_unlock_irqrestore(&qp->r_lock, flags);
 342                         rcu_read_unlock();
 343                 } /* Unicast QP */
 344         } /* Valid packet with TIDErr */
 345
 346         /* handle "RcvTypeErr" flags */
 347         switch (rte) {
 348         case RHF_RTE_ERROR_OP_CODE_ERR:
 349         {
 350                 void *ebuf = NULL;
 351                 u8 opcode;
 352
 353                 if (rhf_use_egr_bfr(packet->rhf))
 354                         ebuf = packet->ebuf;
 355
 356                 if (!ebuf)
 357                         goto drop; /* this should never happen */
 358
 359                 opcode = ib_bth_get_opcode(packet->ohdr);
 360                 if (opcode == IB_OPCODE_CNP) {
 361                         /*
 362                          * Only in pre-B0 h/w is the CNP_OPCODE handled
 363                          * via this code path.
 364                          */
 365                         struct rvt_qp *qp = NULL;
 366                         u32 lqpn, rqpn;
 367                         u16 rlid;
 368                         u8 svc_type, sl, sc5;
 369
 370                         sc5 = hfi1_9B_get_sc5(rhdr, packet->rhf);
 371                         sl = ibp->sc_to_sl[sc5];
 372
 373                         lqpn = ib_bth_get_qpn(packet->ohdr);
 374                         rcu_read_lock();
 375                         qp = rvt_lookup_qpn(rdi, &ibp->rvp, lqpn);
 376                         if (!qp) {
 377                                 rcu_read_unlock();
 378                                 goto drop;
 379                         }
 380
 381                         switch (qp->ibqp.qp_type) {
 382                         case IB_QPT_UD:
 383                                 rlid = 0;
 384                                 rqpn = 0;
 385                                 svc_type = IB_CC_SVCTYPE_UD;
 386                                 break;
 387                         case IB_QPT_UC:
 388                                 rlid = ib_get_slid(rhdr);
 389                                 rqpn = qp->remote_qpn;
 390                                 svc_type = IB_CC_SVCTYPE_UC;
 391                                 break;
 392                         default:
 393                                 rcu_read_unlock();
 394                                 goto drop;
 395                         }
 396
 397                         process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
 398                         rcu_read_unlock();
 399                 }
 400
 401                 packet->rhf &= ~RHF_RCV_TYPE_ERR_SMASK;
 402                 break;
 403         }
 404         default:
 405                 break;
 406         }
 407
 408 drop:
 409         return;
 410 }
 411
 412 static inline void init_packet(struct hfi1_ctxtdata *rcd,
 413                                struct hfi1_packet *packet)
 414 {
 415         packet->rsize = rcd->rcvhdrqentsize; /* words */
 416         packet->maxcnt = rcd->rcvhdrq_cnt * packet->rsize; /* words */
 417         packet->rcd = rcd;
 418         packet->updegr = 0;
 419         packet->etail = -1;
 420         packet->rhf_addr = get_rhf_addr(rcd);
 421         packet->rhf = rhf_to_cpu(packet->rhf_addr);
 422         packet->rhqoff = rcd->head;
 423         packet->numpkt = 0;
 424 }
 425
 426 /* We support only two types - 9B and 16B for now */
 427 static const hfi1_handle_cnp hfi1_handle_cnp_tbl[2] = {
 428         [HFI1_PKT_TYPE_9B] = &return_cnp,
 429         [HFI1_PKT_TYPE_16B] = &return_cnp_16B
 430 };
 431
 432 void hfi1_process_ecn_slowpath(struct rvt_qp *qp, struct hfi1_packet *pkt,
 433                                bool do_cnp)
 434 {
 435         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 436         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 437         struct ib_other_headers *ohdr = pkt->ohdr;
 438         struct ib_grh *grh = pkt->grh;
 439         u32 rqpn = 0, bth1;
 440         u16 pkey;
 441         u32 rlid, slid, dlid = 0;
 442         u8 hdr_type, sc, svc_type;
 443         bool is_mcast = false;
 444
 445         /* can be called from prescan */
 446         if (pkt->etype == RHF_RCV_TYPE_BYPASS) {
 447                 is_mcast = hfi1_is_16B_mcast(dlid);
 448                 pkey = hfi1_16B_get_pkey(pkt->hdr);
 449                 sc = hfi1_16B_get_sc(pkt->hdr);
 450                 dlid = hfi1_16B_get_dlid(pkt->hdr);
 451                 slid = hfi1_16B_get_slid(pkt->hdr);
 452                 hdr_type = HFI1_PKT_TYPE_16B;
 453         } else {
 454                 is_mcast = (dlid > be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
 455                            (dlid != be16_to_cpu(IB_LID_PERMISSIVE));
 456                 pkey = ib_bth_get_pkey(ohdr);
 457                 sc = hfi1_9B_get_sc5(pkt->hdr, pkt->rhf);
 458                 dlid = ib_get_dlid(pkt->hdr);
 459                 slid = ib_get_slid(pkt->hdr);
 460                 hdr_type = HFI1_PKT_TYPE_9B;
 461         }
 462
 463         switch (qp->ibqp.qp_type) {
 464         case IB_QPT_UD:
 465                 dlid = ppd->lid;
 466                 rlid = slid;
 467                 rqpn = ib_get_sqpn(pkt->ohdr);
 468                 svc_type = IB_CC_SVCTYPE_UD;
 469                 break;
 470         case IB_QPT_SMI:
 471         case IB_QPT_GSI:
 472                 rlid = slid;
 473                 rqpn = ib_get_sqpn(pkt->ohdr);
 474                 svc_type = IB_CC_SVCTYPE_UD;
 475                 break;
 476         case IB_QPT_UC:
 477                 rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
 478                 rqpn = qp->remote_qpn;
 479                 svc_type = IB_CC_SVCTYPE_UC;
 480                 break;
 481         case IB_QPT_RC:
 482                 rlid = rdma_ah_get_dlid(&qp->remote_ah_attr);
 483                 rqpn = qp->remote_qpn;
 484                 svc_type = IB_CC_SVCTYPE_RC;
 485                 break;
 486         default:
 487                 return;
 488         }
 489
 490         bth1 = be32_to_cpu(ohdr->bth[1]);
 491         /* Call appropriate CNP handler */
 492         if (do_cnp && (bth1 & IB_FECN_SMASK))
 493                 hfi1_handle_cnp_tbl[hdr_type](ibp, qp, rqpn, pkey,
 494                                               dlid, rlid, sc, grh);
 495
 496         if (!is_mcast && (bth1 & IB_BECN_SMASK)) {
 497                 u32 lqpn = bth1 & RVT_QPN_MASK;
 498                 u8 sl = ibp->sc_to_sl[sc];
 499
 500                 process_becn(ppd, sl, rlid, lqpn, rqpn, svc_type);
 501         }
 502
 503 }
 504
 505 struct ps_mdata {
 506         struct hfi1_ctxtdata *rcd;
 507         u32 rsize;
 508         u32 maxcnt;
 509         u32 ps_head;
 510         u32 ps_tail;
 511         u32 ps_seq;
 512 };
 513
 514 static inline void init_ps_mdata(struct ps_mdata *mdata,
 515                                  struct hfi1_packet *packet)
 516 {
 517         struct hfi1_ctxtdata *rcd = packet->rcd;
 518
 519         mdata->rcd = rcd;
 520         mdata->rsize = packet->rsize;
 521         mdata->maxcnt = packet->maxcnt;
 522         mdata->ps_head = packet->rhqoff;
 523
 524         if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
 525                 mdata->ps_tail = get_rcvhdrtail(rcd);
 526                 if (rcd->ctxt == HFI1_CTRL_CTXT)
 527                         mdata->ps_seq = rcd->seq_cnt;
 528                 else
 529                         mdata->ps_seq = 0; /* not used with DMA_RTAIL */
 530         } else {
 531                 mdata->ps_tail = 0; /* used only with DMA_RTAIL*/
 532                 mdata->ps_seq = rcd->seq_cnt;
 533         }
 534 }
 535
 536 static inline int ps_done(struct ps_mdata *mdata, u64 rhf,
 537                           struct hfi1_ctxtdata *rcd)
 538 {
 539         if (HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL))
 540                 return mdata->ps_head == mdata->ps_tail;
 541         return mdata->ps_seq != rhf_rcv_seq(rhf);
 542 }
 543
 544 static inline int ps_skip(struct ps_mdata *mdata, u64 rhf,
 545                           struct hfi1_ctxtdata *rcd)
 546 {
 547         /*
 548          * Control context can potentially receive an invalid rhf.
 549          * Drop such packets.
 550          */
 551         if ((rcd->ctxt == HFI1_CTRL_CTXT) && (mdata->ps_head != mdata->ps_tail))
 552                 return mdata->ps_seq != rhf_rcv_seq(rhf);
 553
 554         return 0;
 555 }
 556
 557 static inline void update_ps_mdata(struct ps_mdata *mdata,
 558                                    struct hfi1_ctxtdata *rcd)
 559 {
 560         mdata->ps_head += mdata->rsize;
 561         if (mdata->ps_head >= mdata->maxcnt)
 562                 mdata->ps_head = 0;
 563
 564         /* Control context must do seq counting */
 565         if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ||
 566             (rcd->ctxt == HFI1_CTRL_CTXT)) {
 567                 if (++mdata->ps_seq > 13)
 568                         mdata->ps_seq = 1;
 569         }
 570 }
 571
 572 /*
 573  * prescan_rxq - search through the receive queue looking for packets
 574  * containing Excplicit Congestion Notifications (FECNs, or BECNs).
 575  * When an ECN is found, process the Congestion Notification, and toggle
 576  * it off.
 577  * This is declared as a macro to allow quick checking of the port to avoid
 578  * the overhead of a function call if not enabled.
 579  */
 580 #define prescan_rxq(rcd, packet) \
 581         do { \
 582                 if (rcd->ppd->cc_prescan) \
 583                         __prescan_rxq(packet); \
 584         } while (0)
 585 static void __prescan_rxq(struct hfi1_packet *packet)
 586 {
 587         struct hfi1_ctxtdata *rcd = packet->rcd;
 588         struct ps_mdata mdata;
 589
 590         init_ps_mdata(&mdata, packet);
 591
 592         while (1) {
 593                 struct hfi1_devdata *dd = rcd->dd;
 594                 struct hfi1_ibport *ibp = rcd_to_iport(rcd);
 595                 __le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head +
 596                                          dd->rhf_offset;
 597                 struct rvt_qp *qp;
 598                 struct ib_header *hdr;
 599                 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
 600                 u64 rhf = rhf_to_cpu(rhf_addr);
 601                 u32 etype = rhf_rcv_type(rhf), qpn, bth1;
 602                 int is_ecn = 0;
 603                 u8 lnh;
 604
 605                 if (ps_done(&mdata, rhf, rcd))
 606                         break;
 607
 608                 if (ps_skip(&mdata, rhf, rcd))
 609                         goto next;
 610
 611                 if (etype != RHF_RCV_TYPE_IB)
 612                         goto next;
 613
 614                 packet->hdr = hfi1_get_msgheader(dd, rhf_addr);
 615                 hdr = packet->hdr;
 616                 lnh = ib_get_lnh(hdr);
 617
 618                 if (lnh == HFI1_LRH_BTH) {
 619                         packet->ohdr = &hdr->u.oth;
 620                         packet->grh = NULL;
 621                 } else if (lnh == HFI1_LRH_GRH) {
 622                         packet->ohdr = &hdr->u.l.oth;
 623                         packet->grh = &hdr->u.l.grh;
 624                 } else {
 625                         goto next; /* just in case */
 626                 }
 627
 628                 bth1 = be32_to_cpu(packet->ohdr->bth[1]);
 629                 is_ecn = !!(bth1 & (IB_FECN_SMASK | IB_BECN_SMASK));
 630
 631                 if (!is_ecn)
 632                         goto next;
 633
 634                 qpn = bth1 & RVT_QPN_MASK;
 635                 rcu_read_lock();
 636                 qp = rvt_lookup_qpn(rdi, &ibp->rvp, qpn);
 637
 638                 if (!qp) {
 639                         rcu_read_unlock();
 640                         goto next;
 641                 }
 642
 643                 process_ecn(qp, packet, true);
 644                 rcu_read_unlock();
 645
 646                 /* turn off BECN, FECN */
 647                 bth1 &= ~(IB_FECN_SMASK | IB_BECN_SMASK);
 648                 packet->ohdr->bth[1] = cpu_to_be32(bth1);
 649 next:
 650                 update_ps_mdata(&mdata, rcd);
 651         }
 652 }
 653
 654 static void process_rcv_qp_work(struct hfi1_packet *packet)
 655 {
 656         struct rvt_qp *qp, *nqp;
 657         struct hfi1_ctxtdata *rcd = packet->rcd;
 658
 659         /*
 660          * Iterate over all QPs waiting to respond.
 661          * The list won't change since the IRQ is only run on one CPU.
 662          */
 663         list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
 664                 list_del_init(&qp->rspwait);
 665                 if (qp->r_flags & RVT_R_RSP_NAK) {
 666                         qp->r_flags &= ~RVT_R_RSP_NAK;
 667                         packet->qp = qp;
 668                         hfi1_send_rc_ack(packet, 0);
 669                 }
 670                 if (qp->r_flags & RVT_R_RSP_SEND) {
 671                         unsigned long flags;
 672
 673                         qp->r_flags &= ~RVT_R_RSP_SEND;
 674                         spin_lock_irqsave(&qp->s_lock, flags);
 675                         if (ib_rvt_state_ops[qp->state] &
 676                                         RVT_PROCESS_OR_FLUSH_SEND)
 677                                 hfi1_schedule_send(qp);
 678                         spin_unlock_irqrestore(&qp->s_lock, flags);
 679                 }
 680                 rvt_put_qp(qp);
 681         }
 682 }
 683
 684 static noinline int max_packet_exceeded(struct hfi1_packet *packet, int thread)
 685 {
 686         if (thread) {
 687                 if ((packet->numpkt & (MAX_PKT_RECV_THREAD - 1)) == 0)
 688                         /* allow defered processing */
 689                         process_rcv_qp_work(packet);
 690                 cond_resched();
 691                 return RCV_PKT_OK;
 692         } else {
 693                 this_cpu_inc(*packet->rcd->dd->rcv_limit);
 694                 return RCV_PKT_LIMIT;
 695         }
 696 }
 697
 698 static inline int check_max_packet(struct hfi1_packet *packet, int thread)
 699 {
 700         int ret = RCV_PKT_OK;
 701
 702         if (unlikely((packet->numpkt & (MAX_PKT_RECV - 1)) == 0))
 703                 ret = max_packet_exceeded(packet, thread);
 704         return ret;
 705 }
 706
 707 static noinline int skip_rcv_packet(struct hfi1_packet *packet, int thread)
 708 {
 709         int ret;
 710
 711         /* Set up for the next packet */
 712         packet->rhqoff += packet->rsize;
 713         if (packet->rhqoff >= packet->maxcnt)
 714                 packet->rhqoff = 0;
 715
 716         packet->numpkt++;
 717         ret = check_max_packet(packet, thread);
 718
 719         packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
 720                                      packet->rcd->dd->rhf_offset;
 721         packet->rhf = rhf_to_cpu(packet->rhf_addr);
 722
 723         return ret;
 724 }
 725
 726 static inline int process_rcv_packet(struct hfi1_packet *packet, int thread)
 727 {
 728         int ret;
 729
 730         packet->etype = rhf_rcv_type(packet->rhf);
 731
 732         /* total length */
 733         packet->tlen = rhf_pkt_len(packet->rhf); /* in bytes */
 734         /* retrieve eager buffer details */
 735         packet->ebuf = NULL;
 736         if (rhf_use_egr_bfr(packet->rhf)) {
 737                 packet->etail = rhf_egr_index(packet->rhf);
 738                 packet->ebuf = get_egrbuf(packet->rcd, packet->rhf,
 739                                  &packet->updegr);
 740                 /*
 741                  * Prefetch the contents of the eager buffer.  It is
 742                  * OK to send a negative length to prefetch_range().
 743                  * The +2 is the size of the RHF.
 744                  */
 745                 prefetch_range(packet->ebuf,
 746                                packet->tlen - ((packet->rcd->rcvhdrqentsize -
 747                                                (rhf_hdrq_offset(packet->rhf)
 748                                                 + 2)) * 4));
 749         }
 750
 751         /*
 752          * Call a type specific handler for the packet. We
 753          * should be able to trust that etype won't be beyond
 754          * the range of valid indexes. If so something is really
 755          * wrong and we can probably just let things come
 756          * crashing down. There is no need to eat another
 757          * comparison in this performance critical code.
 758          */
 759         packet->rcd->dd->rhf_rcv_function_map[packet->etype](packet);
 760         packet->numpkt++;
 761
 762         /* Set up for the next packet */
 763         packet->rhqoff += packet->rsize;
 764         if (packet->rhqoff >= packet->maxcnt)
 765                 packet->rhqoff = 0;
 766
 767         ret = check_max_packet(packet, thread);
 768
 769         packet->rhf_addr = (__le32 *)packet->rcd->rcvhdrq + packet->rhqoff +
 770                                       packet->rcd->dd->rhf_offset;
 771         packet->rhf = rhf_to_cpu(packet->rhf_addr);
 772
 773         return ret;
 774 }
 775
 776 static inline void process_rcv_update(int last, struct hfi1_packet *packet)
 777 {
 778         /*
 779          * Update head regs etc., every 16 packets, if not last pkt,
 780          * to help prevent rcvhdrq overflows, when many packets
 781          * are processed and queue is nearly full.
 782          * Don't request an interrupt for intermediate updates.
 783          */
 784         if (!last && !(packet->numpkt & 0xf)) {
 785                 update_usrhead(packet->rcd, packet->rhqoff, packet->updegr,
 786                                packet->etail, 0, 0);
 787                 packet->updegr = 0;
 788         }
 789         packet->grh = NULL;
 790 }
 791
 792 static inline void finish_packet(struct hfi1_packet *packet)
 793 {
 794         /*
 795          * Nothing we need to free for the packet.
 796          *
 797          * The only thing we need to do is a final update and call for an
 798          * interrupt
 799          */
 800         update_usrhead(packet->rcd, packet->rcd->head, packet->updegr,
 801                        packet->etail, rcv_intr_dynamic, packet->numpkt);
 802 }
 803
 804 /*
 805  * Handle receive interrupts when using the no dma rtail option.
 806  */
 807 int handle_receive_interrupt_nodma_rtail(struct hfi1_ctxtdata *rcd, int thread)
 808 {
 809         u32 seq;
 810         int last = RCV_PKT_OK;
 811         struct hfi1_packet packet;
 812
 813         init_packet(rcd, &packet);
 814         seq = rhf_rcv_seq(packet.rhf);
 815         if (seq != rcd->seq_cnt) {
 816                 last = RCV_PKT_DONE;
 817                 goto bail;
 818         }
 819
 820         prescan_rxq(rcd, &packet);
 821
 822         while (last == RCV_PKT_OK) {
 823                 last = process_rcv_packet(&packet, thread);
 824                 seq = rhf_rcv_seq(packet.rhf);
 825                 if (++rcd->seq_cnt > 13)
 826                         rcd->seq_cnt = 1;
 827                 if (seq != rcd->seq_cnt)
 828                         last = RCV_PKT_DONE;
 829                 process_rcv_update(last, &packet);
 830         }
 831         process_rcv_qp_work(&packet);
 832         rcd->head = packet.rhqoff;
 833 bail:
 834         finish_packet(&packet);
 835         return last;
 836 }
 837
 838 int handle_receive_interrupt_dma_rtail(struct hfi1_ctxtdata *rcd, int thread)
 839 {
 840         u32 hdrqtail;
 841         int last = RCV_PKT_OK;
 842         struct hfi1_packet packet;
 843
 844         init_packet(rcd, &packet);
 845         hdrqtail = get_rcvhdrtail(rcd);
 846         if (packet.rhqoff == hdrqtail) {
 847                 last = RCV_PKT_DONE;
 848                 goto bail;
 849         }
 850         smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
 851
 852         prescan_rxq(rcd, &packet);
 853
 854         while (last == RCV_PKT_OK) {
 855                 last = process_rcv_packet(&packet, thread);
 856                 if (packet.rhqoff == hdrqtail)
 857                         last = RCV_PKT_DONE;
 858                 process_rcv_update(last, &packet);
 859         }
 860         process_rcv_qp_work(&packet);
 861         rcd->head = packet.rhqoff;
 862 bail:
 863         finish_packet(&packet);
 864         return last;
 865 }
 866
 867 static inline void set_nodma_rtail(struct hfi1_devdata *dd, u16 ctxt)
 868 {
 869         struct hfi1_ctxtdata *rcd;
 870         u16 i;
 871
 872         /*
 873          * For dynamically allocated kernel contexts (like vnic) switch
 874          * interrupt handler only for that context. Otherwise, switch
 875          * interrupt handler for all statically allocated kernel contexts.
 876          */
 877         if (ctxt >= dd->first_dyn_alloc_ctxt) {
 878                 rcd = hfi1_rcd_get_by_index_safe(dd, ctxt);
 879                 if (rcd) {
 880                         rcd->do_interrupt =
 881                                 &handle_receive_interrupt_nodma_rtail;
 882                         hfi1_rcd_put(rcd);
 883                 }
 884                 return;
 885         }
 886
 887         for (i = HFI1_CTRL_CTXT + 1; i < dd->first_dyn_alloc_ctxt; i++) {
 888                 rcd = hfi1_rcd_get_by_index(dd, i);
 889                 if (rcd)
 890                         rcd->do_interrupt =
 891                                 &handle_receive_interrupt_nodma_rtail;
 892                 hfi1_rcd_put(rcd);
 893         }
 894 }
 895
 896 static inline void set_dma_rtail(struct hfi1_devdata *dd, u16 ctxt)
 897 {
 898         struct hfi1_ctxtdata *rcd;
 899         u16 i;
 900
 901         /*
 902          * For dynamically allocated kernel contexts (like vnic) switch
 903          * interrupt handler only for that context. Otherwise, switch
 904          * interrupt handler for all statically allocated kernel contexts.
 905          */
 906         if (ctxt >= dd->first_dyn_alloc_ctxt) {
 907                 rcd = hfi1_rcd_get_by_index_safe(dd, ctxt);
 908                 if (rcd) {
 909                         rcd->do_interrupt =
 910                                 &handle_receive_interrupt_dma_rtail;
 911                         hfi1_rcd_put(rcd);
 912                 }
 913                 return;
 914         }
 915
 916         for (i = HFI1_CTRL_CTXT + 1; i < dd->first_dyn_alloc_ctxt; i++) {
 917                 rcd = hfi1_rcd_get_by_index(dd, i);
 918                 if (rcd)
 919                         rcd->do_interrupt =
 920                                 &handle_receive_interrupt_dma_rtail;
 921                 hfi1_rcd_put(rcd);
 922         }
 923 }
 924
 925 void set_all_slowpath(struct hfi1_devdata *dd)
 926 {
 927         struct hfi1_ctxtdata *rcd;
 928         u16 i;
 929
 930         /* HFI1_CTRL_CTXT must always use the slow path interrupt handler */
 931         for (i = HFI1_CTRL_CTXT + 1; i < dd->num_rcv_contexts; i++) {
 932                 rcd = hfi1_rcd_get_by_index(dd, i);
 933                 if (!rcd)
 934                         continue;
 935                 if (i < dd->first_dyn_alloc_ctxt || rcd->is_vnic)
 936                         rcd->do_interrupt = &handle_receive_interrupt;
 937
 938                 hfi1_rcd_put(rcd);
 939         }
 940 }
 941
 942 static inline int set_armed_to_active(struct hfi1_ctxtdata *rcd,
 943                                       struct hfi1_packet *packet,
 944                                       struct hfi1_devdata *dd)
 945 {
 946         struct work_struct *lsaw = &rcd->ppd->linkstate_active_work;
 947         u8 etype = rhf_rcv_type(packet->rhf);
 948         u8 sc = SC15_PACKET;
 949
 950         if (etype == RHF_RCV_TYPE_IB) {
 951                 struct ib_header *hdr = hfi1_get_msgheader(packet->rcd->dd,
 952                                                            packet->rhf_addr);
 953                 sc = hfi1_9B_get_sc5(hdr, packet->rhf);
 954         } else if (etype == RHF_RCV_TYPE_BYPASS) {
 955                 struct hfi1_16b_header *hdr = hfi1_get_16B_header(
 956                                                 packet->rcd->dd,
 957                                                 packet->rhf_addr);
 958                 sc = hfi1_16B_get_sc(hdr);
 959         }
 960         if (sc != SC15_PACKET) {
 961                 int hwstate = driver_lstate(rcd->ppd);
 962
 963                 if (hwstate != IB_PORT_ACTIVE) {
 964                         dd_dev_info(dd,
 965                                     "Unexpected link state %s\n",
 966                                     opa_lstate_name(hwstate));
 967                         return 0;
 968                 }
 969
 970                 queue_work(rcd->ppd->link_wq, lsaw);
 971                 return 1;
 972         }
 973         return 0;
 974 }
 975
 976 /*
 977  * handle_receive_interrupt - receive a packet
 978  * @rcd: the context
 979  *
 980  * Called from interrupt handler for errors or receive interrupt.
 981  * This is the slow path interrupt handler.
 982  */
 983 int handle_receive_interrupt(struct hfi1_ctxtdata *rcd, int thread)
 984 {
 985         struct hfi1_devdata *dd = rcd->dd;
 986         u32 hdrqtail;
 987         int needset, last = RCV_PKT_OK;
 988         struct hfi1_packet packet;
 989         int skip_pkt = 0;
 990
 991         /* Control context will always use the slow path interrupt handler */
 992         needset = (rcd->ctxt == HFI1_CTRL_CTXT) ? 0 : 1;
 993
 994         init_packet(rcd, &packet);
 995
 996         if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
 997                 u32 seq = rhf_rcv_seq(packet.rhf);
 998
 999                 if (seq != rcd->seq_cnt) {
1000                         last = RCV_PKT_DONE;
1001                         goto bail;
1002                 }
1003                 hdrqtail = 0;
1004         } else {
1005                 hdrqtail = get_rcvhdrtail(rcd);
1006                 if (packet.rhqoff == hdrqtail) {
1007                         last = RCV_PKT_DONE;
1008                         goto bail;
1009                 }
1010                 smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
1011
1012                 /*
1013                  * Control context can potentially receive an invalid
1014                  * rhf. Drop such packets.
1015                  */
1016                 if (rcd->ctxt == HFI1_CTRL_CTXT) {
1017                         u32 seq = rhf_rcv_seq(packet.rhf);
1018
1019                         if (seq != rcd->seq_cnt)
1020                                 skip_pkt = 1;
1021                 }
1022         }
1023
1024         prescan_rxq(rcd, &packet);
1025
1026         while (last == RCV_PKT_OK) {
1027                 if (unlikely(dd->do_drop &&
1028                              atomic_xchg(&dd->drop_packet, DROP_PACKET_OFF) ==
1029                              DROP_PACKET_ON)) {
1030                         dd->do_drop = 0;
1031
1032                         /* On to the next packet */
1033                         packet.rhqoff += packet.rsize;
1034                         packet.rhf_addr = (__le32 *)rcd->rcvhdrq +
1035                                           packet.rhqoff +
1036                                           dd->rhf_offset;
1037                         packet.rhf = rhf_to_cpu(packet.rhf_addr);
1038
1039                 } else if (skip_pkt) {
1040                         last = skip_rcv_packet(&packet, thread);
1041                         skip_pkt = 0;
1042                 } else {
1043                         /* Auto activate link on non-SC15 packet receive */
1044                         if (unlikely(rcd->ppd->host_link_state ==
1045                                      HLS_UP_ARMED) &&
1046                             set_armed_to_active(rcd, &packet, dd))
1047                                 goto bail;
1048                         last = process_rcv_packet(&packet, thread);
1049                 }
1050
1051                 if (!HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL)) {
1052                         u32 seq = rhf_rcv_seq(packet.rhf);
1053
1054                         if (++rcd->seq_cnt > 13)
1055                                 rcd->seq_cnt = 1;
1056                         if (seq != rcd->seq_cnt)
1057                                 last = RCV_PKT_DONE;
1058                         if (needset) {
1059                                 dd_dev_info(dd, "Switching to NO_DMA_RTAIL\n");
1060                                 set_nodma_rtail(dd, rcd->ctxt);
1061                                 needset = 0;
1062                         }
1063                 } else {
1064                         if (packet.rhqoff == hdrqtail)
1065                                 last = RCV_PKT_DONE;
1066                         /*
1067                          * Control context can potentially receive an invalid
1068                          * rhf. Drop such packets.
1069                          */
1070                         if (rcd->ctxt == HFI1_CTRL_CTXT) {
1071                                 u32 seq = rhf_rcv_seq(packet.rhf);
1072
1073                                 if (++rcd->seq_cnt > 13)
1074                                         rcd->seq_cnt = 1;
1075                                 if (!last && (seq != rcd->seq_cnt))
1076                                         skip_pkt = 1;
1077                         }
1078
1079                         if (needset) {
1080                                 dd_dev_info(dd,
1081                                             "Switching to DMA_RTAIL\n");
1082                                 set_dma_rtail(dd, rcd->ctxt);
1083                                 needset = 0;
1084                         }
1085                 }
1086
1087                 process_rcv_update(last, &packet);
1088         }
1089
1090         process_rcv_qp_work(&packet);
1091         rcd->head = packet.rhqoff;
1092
1093 bail:
1094         /*
1095          * Always write head at end, and setup rcv interrupt, even
1096          * if no packets were processed.
1097          */
1098         finish_packet(&packet);
1099         return last;
1100 }
1101
1102 /*
1103  * We may discover in the interrupt that the hardware link state has
1104  * changed from ARMED to ACTIVE (due to the arrival of a non-SC15 packet),
1105  * and we need to update the driver's notion of the link state.  We cannot
1106  * run set_link_state from interrupt context, so we queue this function on
1107  * a workqueue.
1108  *
1109  * We delay the regular interrupt processing until after the state changes
1110  * so that the link will be in the correct state by the time any application
1111  * we wake up attempts to send a reply to any message it received.
1112  * (Subsequent receive interrupts may possibly force the wakeup before we
1113  * update the link state.)
1114  *
1115  * The rcd is freed in hfi1_free_ctxtdata after hfi1_postinit_cleanup invokes
1116  * dd->f_cleanup(dd) to disable the interrupt handler and flush workqueues,
1117  * so we're safe from use-after-free of the rcd.
1118  */
1119 void receive_interrupt_work(struct work_struct *work)
1120 {
1121         struct hfi1_pportdata *ppd = container_of(work, struct hfi1_pportdata,
1122                                                   linkstate_active_work);
1123         struct hfi1_devdata *dd = ppd->dd;
1124         struct hfi1_ctxtdata *rcd;
1125         u16 i;
1126
1127         /* Received non-SC15 packet implies neighbor_normal */
1128         ppd->neighbor_normal = 1;
1129         set_link_state(ppd, HLS_UP_ACTIVE);
1130
1131         /*
1132          * Interrupt all statically allocated kernel contexts that could
1133          * have had an interrupt during auto activation.
1134          */
1135         for (i = HFI1_CTRL_CTXT; i < dd->first_dyn_alloc_ctxt; i++) {
1136                 rcd = hfi1_rcd_get_by_index(dd, i);
1137                 if (rcd)
1138                         force_recv_intr(rcd);
1139                 hfi1_rcd_put(rcd);
1140         }
1141 }
1142
1143 /*
1144  * Convert a given MTU size to the on-wire MAD packet enumeration.
1145  * Return -1 if the size is invalid.
1146  */
1147 int mtu_to_enum(u32 mtu, int default_if_bad)
1148 {
1149         switch (mtu) {
1150         case     0: return OPA_MTU_0;
1151         case   256: return OPA_MTU_256;
1152         case   512: return OPA_MTU_512;
1153         case  1024: return OPA_MTU_1024;
1154         case  2048: return OPA_MTU_2048;
1155         case  4096: return OPA_MTU_4096;
1156         case  8192: return OPA_MTU_8192;
1157         case 10240: return OPA_MTU_10240;
1158         }
1159         return default_if_bad;
1160 }
1161
1162 u16 enum_to_mtu(int mtu)
1163 {
1164         switch (mtu) {
1165         case OPA_MTU_0:     return 0;
1166         case OPA_MTU_256:   return 256;
1167         case OPA_MTU_512:   return 512;
1168         case OPA_MTU_1024:  return 1024;
1169         case OPA_MTU_2048:  return 2048;
1170         case OPA_MTU_4096:  return 4096;
1171         case OPA_MTU_8192:  return 8192;
1172         case OPA_MTU_10240: return 10240;
1173         default: return 0xffff;
1174         }
1175 }
1176
1177 /*
1178  * set_mtu - set the MTU
1179  * @ppd: the per port data
1180  *
1181  * We can handle "any" incoming size, the issue here is whether we
1182  * need to restrict our outgoing size.  We do not deal with what happens
1183  * to programs that are already running when the size changes.
1184  */
1185 int set_mtu(struct hfi1_pportdata *ppd)
1186 {
1187         struct hfi1_devdata *dd = ppd->dd;
1188         int i, drain, ret = 0, is_up = 0;
1189
1190         ppd->ibmtu = 0;
1191         for (i = 0; i < ppd->vls_supported; i++)
1192                 if (ppd->ibmtu < dd->vld[i].mtu)
1193                         ppd->ibmtu = dd->vld[i].mtu;
1194         ppd->ibmaxlen = ppd->ibmtu + lrh_max_header_bytes(ppd->dd);
1195
1196         mutex_lock(&ppd->hls_lock);
1197         if (ppd->host_link_state == HLS_UP_INIT ||
1198             ppd->host_link_state == HLS_UP_ARMED ||
1199             ppd->host_link_state == HLS_UP_ACTIVE)
1200                 is_up = 1;
1201
1202         drain = !is_ax(dd) && is_up;
1203
1204         if (drain)
1205                 /*
1206                  * MTU is specified per-VL. To ensure that no packet gets
1207                  * stuck (due, e.g., to the MTU for the packet's VL being
1208                  * reduced), empty the per-VL FIFOs before adjusting MTU.
1209                  */
1210                 ret = stop_drain_data_vls(dd);
1211
1212         if (ret) {
1213                 dd_dev_err(dd, "%s: cannot stop/drain VLs - refusing to change per-VL MTUs\n",
1214                            __func__);
1215                 goto err;
1216         }
1217
1218         hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_MTU, 0);
1219
1220         if (drain)
1221                 open_fill_data_vls(dd); /* reopen all VLs */
1222
1223 err:
1224         mutex_unlock(&ppd->hls_lock);
1225
1226         return ret;
1227 }
1228
1229 int hfi1_set_lid(struct hfi1_pportdata *ppd, u32 lid, u8 lmc)
1230 {
1231         struct hfi1_devdata *dd = ppd->dd;
1232
1233         ppd->lid = lid;
1234         ppd->lmc = lmc;
1235         hfi1_set_ib_cfg(ppd, HFI1_IB_CFG_LIDLMC, 0);
1236
1237         dd_dev_info(dd, "port %u: got a lid: 0x%x\n", ppd->port, lid);
1238
1239         return 0;
1240 }
1241
1242 void shutdown_led_override(struct hfi1_pportdata *ppd)
1243 {
1244         struct hfi1_devdata *dd = ppd->dd;
1245
1246         /*
1247          * This pairs with the memory barrier in hfi1_start_led_override to
1248          * ensure that we read the correct state of LED beaconing represented
1249          * by led_override_timer_active
1250          */
1251         smp_rmb();
1252         if (atomic_read(&ppd->led_override_timer_active)) {
1253                 del_timer_sync(&ppd->led_override_timer);
1254                 atomic_set(&ppd->led_override_timer_active, 0);
1255                 /* Ensure the atomic_set is visible to all CPUs */
1256                 smp_wmb();
1257         }
1258
1259         /* Hand control of the LED to the DC for normal operation */
1260         write_csr(dd, DCC_CFG_LED_CNTRL, 0);
1261 }
1262
1263 static void run_led_override(struct timer_list *t)
1264 {
1265         struct hfi1_pportdata *ppd = from_timer(ppd, t, led_override_timer);
1266         struct hfi1_devdata *dd = ppd->dd;
1267         unsigned long timeout;
1268         int phase_idx;
1269
1270         if (!(dd->flags & HFI1_INITTED))
1271                 return;
1272
1273         phase_idx = ppd->led_override_phase & 1;
1274
1275         setextled(dd, phase_idx);
1276
1277         timeout = ppd->led_override_vals[phase_idx];
1278
1279         /* Set up for next phase */
1280         ppd->led_override_phase = !ppd->led_override_phase;
1281
1282         mod_timer(&ppd->led_override_timer, jiffies + timeout);
1283 }
1284
1285 /*
1286  * To have the LED blink in a particular pattern, provide timeon and timeoff
1287  * in milliseconds.
1288  * To turn off custom blinking and return to normal operation, use
1289  * shutdown_led_override()
1290  */
1291 void hfi1_start_led_override(struct hfi1_pportdata *ppd, unsigned int timeon,
1292                              unsigned int timeoff)
1293 {
1294         if (!(ppd->dd->flags & HFI1_INITTED))
1295                 return;
1296
1297         /* Convert to jiffies for direct use in timer */
1298         ppd->led_override_vals[0] = msecs_to_jiffies(timeoff);
1299         ppd->led_override_vals[1] = msecs_to_jiffies(timeon);
1300
1301         /* Arbitrarily start from LED on phase */
1302         ppd->led_override_phase = 1;
1303
1304         /*
1305          * If the timer has not already been started, do so. Use a "quick"
1306          * timeout so the handler will be called soon to look at our request.
1307          */
1308         if (!timer_pending(&ppd->led_override_timer)) {
1309                 timer_setup(&ppd->led_override_timer, run_led_override, 0);
1310                 ppd->led_override_timer.expires = jiffies + 1;
1311                 add_timer(&ppd->led_override_timer);
1312                 atomic_set(&ppd->led_override_timer_active, 1);
1313                 /* Ensure the atomic_set is visible to all CPUs */
1314                 smp_wmb();
1315         }
1316 }
1317
1318 /**
1319  * hfi1_reset_device - reset the chip if possible
1320  * @unit: the device to reset
1321  *
1322  * Whether or not reset is successful, we attempt to re-initialize the chip
1323  * (that is, much like a driver unload/reload).  We clear the INITTED flag
1324  * so that the various entry points will fail until we reinitialize.  For
1325  * now, we only allow this if no user contexts are open that use chip resources
1326  */
1327 int hfi1_reset_device(int unit)
1328 {
1329         int ret;
1330         struct hfi1_devdata *dd = hfi1_lookup(unit);
1331         struct hfi1_pportdata *ppd;
1332         int pidx;
1333
1334         if (!dd) {
1335                 ret = -ENODEV;
1336                 goto bail;
1337         }
1338
1339         dd_dev_info(dd, "Reset on unit %u requested\n", unit);
1340
1341         if (!dd->kregbase1 || !(dd->flags & HFI1_PRESENT)) {
1342                 dd_dev_info(dd,
1343                             "Invalid unit number %u or not initialized or not present\n",
1344                             unit);
1345                 ret = -ENXIO;
1346                 goto bail;
1347         }
1348
1349         /* If there are any user/vnic contexts, we cannot reset */
1350         mutex_lock(&hfi1_mutex);
1351         if (dd->rcd)
1352                 if (hfi1_stats.sps_ctxts) {
1353                         mutex_unlock(&hfi1_mutex);
1354                         ret = -EBUSY;
1355                         goto bail;
1356                 }
1357         mutex_unlock(&hfi1_mutex);
1358
1359         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1360                 ppd = dd->pport + pidx;
1361
1362                 shutdown_led_override(ppd);
1363         }
1364         if (dd->flags & HFI1_HAS_SEND_DMA)
1365                 sdma_exit(dd);
1366
1367         hfi1_reset_cpu_counters(dd);
1368
1369         ret = hfi1_init(dd, 1);
1370
1371         if (ret)
1372                 dd_dev_err(dd,
1373                            "Reinitialize unit %u after reset failed with %d\n",
1374                            unit, ret);
1375         else
1376                 dd_dev_info(dd, "Reinitialized unit %u after resetting\n",
1377                             unit);
1378
1379 bail:
1380         return ret;
1381 }
1382
1383 static inline void hfi1_setup_ib_header(struct hfi1_packet *packet)
1384 {
1385         packet->hdr = (struct hfi1_ib_message_header *)
1386                         hfi1_get_msgheader(packet->rcd->dd,
1387                                            packet->rhf_addr);
1388         packet->hlen = (u8 *)packet->rhf_addr - (u8 *)packet->hdr;
1389 }
1390
1391 static int hfi1_bypass_ingress_pkt_check(struct hfi1_packet *packet)
1392 {
1393         struct hfi1_pportdata *ppd = packet->rcd->ppd;
1394
1395         /* slid and dlid cannot be 0 */
1396         if ((!packet->slid) || (!packet->dlid))
1397                 return -EINVAL;
1398
1399         /* Compare port lid with incoming packet dlid */
1400         if ((!(hfi1_is_16B_mcast(packet->dlid))) &&
1401             (packet->dlid !=
1402                 opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE), 16B))) {
1403                 if (packet->dlid != ppd->lid)
1404                         return -EINVAL;
1405         }
1406
1407         /* No multicast packets with SC15 */
1408         if ((hfi1_is_16B_mcast(packet->dlid)) && (packet->sc == 0xF))
1409                 return -EINVAL;
1410
1411         /* Packets with permissive DLID always on SC15 */
1412         if ((packet->dlid == opa_get_lid(be32_to_cpu(OPA_LID_PERMISSIVE),
1413                                          16B)) &&
1414             (packet->sc != 0xF))
1415                 return -EINVAL;
1416
1417         return 0;
1418 }
1419
1420 static int hfi1_setup_9B_packet(struct hfi1_packet *packet)
1421 {
1422         struct hfi1_ibport *ibp = rcd_to_iport(packet->rcd);
1423         struct ib_header *hdr;
1424         u8 lnh;
1425
1426         hfi1_setup_ib_header(packet);
1427         hdr = packet->hdr;
1428
1429         lnh = ib_get_lnh(hdr);
1430         if (lnh == HFI1_LRH_BTH) {
1431                 packet->ohdr = &hdr->u.oth;
1432                 packet->grh = NULL;
1433         } else if (lnh == HFI1_LRH_GRH) {
1434                 u32 vtf;
1435
1436                 packet->ohdr = &hdr->u.l.oth;
1437                 packet->grh = &hdr->u.l.grh;
1438                 if (packet->grh->next_hdr != IB_GRH_NEXT_HDR)
1439                         goto drop;
1440                 vtf = be32_to_cpu(packet->grh->version_tclass_flow);
1441                 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
1442                         goto drop;
1443         } else {
1444                 goto drop;
1445         }
1446
1447         /* Query commonly used fields from packet header */
1448         packet->payload = packet->ebuf;
1449         packet->opcode = ib_bth_get_opcode(packet->ohdr);
1450         packet->slid = ib_get_slid(hdr);
1451         packet->dlid = ib_get_dlid(hdr);
1452         if (unlikely((packet->dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
1453                      (packet->dlid != be16_to_cpu(IB_LID_PERMISSIVE))))
1454                 packet->dlid += opa_get_mcast_base(OPA_MCAST_NR) -
1455                                 be16_to_cpu(IB_MULTICAST_LID_BASE);
1456         packet->sl = ib_get_sl(hdr);
1457         packet->sc = hfi1_9B_get_sc5(hdr, packet->rhf);
1458         packet->pad = ib_bth_get_pad(packet->ohdr);
1459         packet->extra_byte = 0;
1460         packet->pkey = ib_bth_get_pkey(packet->ohdr);
1461         packet->migrated = ib_bth_is_migration(packet->ohdr);
1462
1463         return 0;
1464 drop:
1465         ibp->rvp.n_pkt_drops++;
1466         return -EINVAL;
1467 }
1468
1469 static int hfi1_setup_bypass_packet(struct hfi1_packet *packet)
1470 {
1471         /*
1472          * Bypass packets have a different header/payload split
1473          * compared to an IB packet.
1474          * Current split is set such that 16 bytes of the actual
1475          * header is in the header buffer and the remining is in
1476          * the eager buffer. We chose 16 since hfi1 driver only
1477          * supports 16B bypass packets and we will be able to
1478          * receive the entire LRH with such a split.
1479          */
1480
1481         struct hfi1_ctxtdata *rcd = packet->rcd;
1482         struct hfi1_pportdata *ppd = rcd->ppd;
1483         struct hfi1_ibport *ibp = &ppd->ibport_data;
1484         u8 l4;
1485         u8 grh_len;
1486
1487         packet->hdr = (struct hfi1_16b_header *)
1488                         hfi1_get_16B_header(packet->rcd->dd,
1489                                             packet->rhf_addr);
1490         packet->hlen = (u8 *)packet->rhf_addr - (u8 *)packet->hdr;
1491
1492         l4 = hfi1_16B_get_l4(packet->hdr);
1493         if (l4 == OPA_16B_L4_IB_LOCAL) {
1494                 grh_len = 0;
1495                 packet->ohdr = packet->ebuf;
1496                 packet->grh = NULL;
1497         } else if (l4 == OPA_16B_L4_IB_GLOBAL) {
1498                 u32 vtf;
1499
1500                 grh_len = sizeof(struct ib_grh);
1501                 packet->ohdr = packet->ebuf + grh_len;
1502                 packet->grh = packet->ebuf;
1503                 if (packet->grh->next_hdr != IB_GRH_NEXT_HDR)
1504                         goto drop;
1505                 vtf = be32_to_cpu(packet->grh->version_tclass_flow);
1506                 if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
1507                         goto drop;
1508         } else {
1509                 goto drop;
1510         }
1511
1512         /* Query commonly used fields from packet header */
1513         packet->opcode = ib_bth_get_opcode(packet->ohdr);
1514         /* hdr_len_by_opcode already has an IB LRH factored in */
1515         packet->hlen = hdr_len_by_opcode[packet->opcode] +
1516                 (LRH_16B_BYTES - LRH_9B_BYTES) + grh_len;
1517         packet->payload = packet->ebuf + packet->hlen - LRH_16B_BYTES;
1518         packet->slid = hfi1_16B_get_slid(packet->hdr);
1519         packet->dlid = hfi1_16B_get_dlid(packet->hdr);
1520         if (unlikely(hfi1_is_16B_mcast(packet->dlid)))
1521                 packet->dlid += opa_get_mcast_base(OPA_MCAST_NR) -
1522                                 opa_get_lid(opa_get_mcast_base(OPA_MCAST_NR),
1523                                             16B);
1524         packet->sc = hfi1_16B_get_sc(packet->hdr);
1525         packet->sl = ibp->sc_to_sl[packet->sc];
1526         packet->pad = hfi1_16B_bth_get_pad(packet->ohdr);
1527         packet->extra_byte = SIZE_OF_LT;
1528         packet->pkey = hfi1_16B_get_pkey(packet->hdr);
1529         packet->migrated = opa_bth_is_migration(packet->ohdr);
1530
1531         if (hfi1_bypass_ingress_pkt_check(packet))
1532                 goto drop;
1533
1534         return 0;
1535 drop:
1536         hfi1_cdbg(PKT, "%s: packet dropped\n", __func__);
1537         ibp->rvp.n_pkt_drops++;
1538         return -EINVAL;
1539 }
1540
1541 void handle_eflags(struct hfi1_packet *packet)
1542 {
1543         struct hfi1_ctxtdata *rcd = packet->rcd;
1544         u32 rte = rhf_rcv_type_err(packet->rhf);
1545
1546         rcv_hdrerr(rcd, rcd->ppd, packet);
1547         if (rhf_err_flags(packet->rhf))
1548                 dd_dev_err(rcd->dd,
1549                            "receive context %d: rhf 0x%016llx, errs [ %s%s%s%s%s%s%s%s] rte 0x%x\n",
1550                            rcd->ctxt, packet->rhf,
1551                            packet->rhf & RHF_K_HDR_LEN_ERR ? "k_hdr_len " : "",
1552                            packet->rhf & RHF_DC_UNC_ERR ? "dc_unc " : "",
1553                            packet->rhf & RHF_DC_ERR ? "dc " : "",
1554                            packet->rhf & RHF_TID_ERR ? "tid " : "",
1555                            packet->rhf & RHF_LEN_ERR ? "len " : "",
1556                            packet->rhf & RHF_ECC_ERR ? "ecc " : "",
1557                            packet->rhf & RHF_VCRC_ERR ? "vcrc " : "",
1558                            packet->rhf & RHF_ICRC_ERR ? "icrc " : "",
1559                            rte);
1560 }
1561
1562 /*
1563  * The following functions are called by the interrupt handler. They are type
1564  * specific handlers for each packet type.
1565  */
1566 int process_receive_ib(struct hfi1_packet *packet)
1567 {
1568         if (unlikely(hfi1_dbg_fault_packet(packet)))
1569                 return RHF_RCV_CONTINUE;
1570
1571         if (hfi1_setup_9B_packet(packet))
1572                 return RHF_RCV_CONTINUE;
1573
1574         trace_hfi1_rcvhdr(packet);
1575
1576         if (unlikely(rhf_err_flags(packet->rhf))) {
1577                 handle_eflags(packet);
1578                 return RHF_RCV_CONTINUE;
1579         }
1580
1581         hfi1_ib_rcv(packet);
1582         return RHF_RCV_CONTINUE;
1583 }
1584
1585 static inline bool hfi1_is_vnic_packet(struct hfi1_packet *packet)
1586 {
1587         /* Packet received in VNIC context via RSM */
1588         if (packet->rcd->is_vnic)
1589                 return true;
1590
1591         if ((hfi1_16B_get_l2(packet->ebuf) == OPA_16B_L2_TYPE) &&
1592             (hfi1_16B_get_l4(packet->ebuf) == OPA_16B_L4_ETHR))
1593                 return true;
1594
1595         return false;
1596 }
1597
1598 int process_receive_bypass(struct hfi1_packet *packet)
1599 {
1600         struct hfi1_devdata *dd = packet->rcd->dd;
1601
1602         if (hfi1_is_vnic_packet(packet)) {
1603                 hfi1_vnic_bypass_rcv(packet);
1604                 return RHF_RCV_CONTINUE;
1605         }
1606
1607         if (hfi1_setup_bypass_packet(packet))
1608                 return RHF_RCV_CONTINUE;
1609
1610         trace_hfi1_rcvhdr(packet);
1611
1612         if (unlikely(rhf_err_flags(packet->rhf))) {
1613                 handle_eflags(packet);
1614                 return RHF_RCV_CONTINUE;
1615         }
1616
1617         if (hfi1_16B_get_l2(packet->hdr) == 0x2) {
1618                 hfi1_16B_rcv(packet);
1619         } else {
1620                 dd_dev_err(dd,
1621                            "Bypass packets other than 16B are not supported in normal operation. Dropping\n");
1622                 incr_cntr64(&dd->sw_rcv_bypass_packet_errors);
1623                 if (!(dd->err_info_rcvport.status_and_code &
1624                       OPA_EI_STATUS_SMASK)) {
1625                         u64 *flits = packet->ebuf;
1626
1627                         if (flits && !(packet->rhf & RHF_LEN_ERR)) {
1628                                 dd->err_info_rcvport.packet_flit1 = flits[0];
1629                                 dd->err_info_rcvport.packet_flit2 =
1630                                         packet->tlen > sizeof(flits[0]) ?
1631                                         flits[1] : 0;
1632                         }
1633                         dd->err_info_rcvport.status_and_code |=
1634                                 (OPA_EI_STATUS_SMASK | BAD_L2_ERR);
1635                 }
1636         }
1637         return RHF_RCV_CONTINUE;
1638 }
1639
1640 int process_receive_error(struct hfi1_packet *packet)
1641 {
1642         /* KHdrHCRCErr -- KDETH packet with a bad HCRC */
1643         if (unlikely(
1644                  hfi1_dbg_fault_suppress_err(&packet->rcd->dd->verbs_dev) &&
1645                  rhf_rcv_type_err(packet->rhf) == 3))
1646                 return RHF_RCV_CONTINUE;
1647
1648         hfi1_setup_ib_header(packet);
1649         handle_eflags(packet);
1650
1651         if (unlikely(rhf_err_flags(packet->rhf)))
1652                 dd_dev_err(packet->rcd->dd,
1653                            "Unhandled error packet received. Dropping.\n");
1654
1655         return RHF_RCV_CONTINUE;
1656 }
1657
1658 int kdeth_process_expected(struct hfi1_packet *packet)
1659 {
1660         if (unlikely(hfi1_dbg_fault_packet(packet)))
1661                 return RHF_RCV_CONTINUE;
1662
1663         hfi1_setup_ib_header(packet);
1664         if (unlikely(rhf_err_flags(packet->rhf)))
1665                 handle_eflags(packet);
1666
1667         dd_dev_err(packet->rcd->dd,
1668                    "Unhandled expected packet received. Dropping.\n");
1669         return RHF_RCV_CONTINUE;
1670 }
1671
1672 int kdeth_process_eager(struct hfi1_packet *packet)
1673 {
1674         hfi1_setup_ib_header(packet);
1675         if (unlikely(rhf_err_flags(packet->rhf)))
1676                 handle_eflags(packet);
1677         if (unlikely(hfi1_dbg_fault_packet(packet)))
1678                 return RHF_RCV_CONTINUE;
1679
1680         dd_dev_err(packet->rcd->dd,
1681                    "Unhandled eager packet received. Dropping.\n");
1682         return RHF_RCV_CONTINUE;
1683 }
1684
1685 int process_receive_invalid(struct hfi1_packet *packet)
1686 {
1687         dd_dev_err(packet->rcd->dd, "Invalid packet type %d. Dropping\n",
1688                    rhf_rcv_type(packet->rhf));
1689         return RHF_RCV_CONTINUE;
1690 }
1691
1692 void seqfile_dump_rcd(struct seq_file *s, struct hfi1_ctxtdata *rcd)
1693 {
1694         struct hfi1_packet packet;
1695         struct ps_mdata mdata;
1696
1697         seq_printf(s, "Rcd %u: RcvHdr cnt %u entsize %u %s head %llu tail %llu\n",
1698                    rcd->ctxt, rcd->rcvhdrq_cnt, rcd->rcvhdrqentsize,
1699                    HFI1_CAP_KGET_MASK(rcd->flags, DMA_RTAIL) ?
1700                    "dma_rtail" : "nodma_rtail",
1701                    read_uctxt_csr(rcd->dd, rcd->ctxt, RCV_HDR_HEAD) &
1702                    RCV_HDR_HEAD_HEAD_MASK,
1703                    read_uctxt_csr(rcd->dd, rcd->ctxt, RCV_HDR_TAIL));
1704
1705         init_packet(rcd, &packet);
1706         init_ps_mdata(&mdata, &packet);
1707
1708         while (1) {
1709                 struct hfi1_devdata *dd = rcd->dd;
1710                 __le32 *rhf_addr = (__le32 *)rcd->rcvhdrq + mdata.ps_head +
1711                                          dd->rhf_offset;
1712                 struct ib_header *hdr;
1713                 u64 rhf = rhf_to_cpu(rhf_addr);
1714                 u32 etype = rhf_rcv_type(rhf), qpn;
1715                 u8 opcode;
1716                 u32 psn;
1717                 u8 lnh;
1718
1719                 if (ps_done(&mdata, rhf, rcd))
1720                         break;
1721
1722                 if (ps_skip(&mdata, rhf, rcd))
1723                         goto next;
1724
1725                 if (etype > RHF_RCV_TYPE_IB)
1726                         goto next;
1727
1728                 packet.hdr = hfi1_get_msgheader(dd, rhf_addr);
1729                 hdr = packet.hdr;
1730
1731                 lnh = be16_to_cpu(hdr->lrh[0]) & 3;
1732
1733                 if (lnh == HFI1_LRH_BTH)
1734                         packet.ohdr = &hdr->u.oth;
1735                 else if (lnh == HFI1_LRH_GRH)
1736                         packet.ohdr = &hdr->u.l.oth;
1737                 else
1738                         goto next; /* just in case */
1739
1740                 opcode = (be32_to_cpu(packet.ohdr->bth[0]) >> 24);
1741                 qpn = be32_to_cpu(packet.ohdr->bth[1]) & RVT_QPN_MASK;
1742                 psn = mask_psn(be32_to_cpu(packet.ohdr->bth[2]));
1743
1744                 seq_printf(s, "\tEnt %u: opcode 0x%x, qpn 0x%x, psn 0x%x\n",
1745                            mdata.ps_head, opcode, qpn, psn);
1746 next:
1747                 update_ps_mdata(&mdata, rcd);
1748         }
1749 }