drivers/infiniband/hw/qib/qib_driver.c

   1 /*
   2  * Copyright (c) 2013 Intel Corporation. All rights reserved.
   3  * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved.
   4  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/spinlock.h>
  36 #include <linux/pci.h>
  37 #include <linux/io.h>
  38 #include <linux/delay.h>
  39 #include <linux/netdevice.h>
  40 #include <linux/vmalloc.h>
  41 #include <linux/module.h>
  42 #include <linux/prefetch.h>
  43
  44 #include "qib.h"
  45
  46 /*
  47  * The size has to be longer than this string, so we can append
  48  * board/chip information to it in the init code.
  49  */
  50 const char ib_qib_version[] = QIB_DRIVER_VERSION "\n";
  51
  52 DEFINE_SPINLOCK(qib_devs_lock);
  53 LIST_HEAD(qib_dev_list);
  54 DEFINE_MUTEX(qib_mutex);        /* general driver use */
  55
  56 unsigned qib_ibmtu;
  57 module_param_named(ibmtu, qib_ibmtu, uint, S_IRUGO);
  58 MODULE_PARM_DESC(ibmtu, "Set max IB MTU (0=2KB, 1=256, 2=512, ... 5=4096");
  59
  60 unsigned qib_compat_ddr_negotiate = 1;
  61 module_param_named(compat_ddr_negotiate, qib_compat_ddr_negotiate, uint,
  62                    S_IWUSR | S_IRUGO);
  63 MODULE_PARM_DESC(compat_ddr_negotiate,
  64                  "Attempt pre-IBTA 1.2 DDR speed negotiation");
  65
  66 MODULE_LICENSE("Dual BSD/GPL");
  67 MODULE_AUTHOR("Intel <[email protected]>");
  68 MODULE_DESCRIPTION("Intel IB driver");
  69
  70 /*
  71  * QIB_PIO_MAXIBHDR is the max IB header size allowed for in our
  72  * PIO send buffers.  This is well beyond anything currently
  73  * defined in the InfiniBand spec.
  74  */
  75 #define QIB_PIO_MAXIBHDR 128
  76
  77 /*
  78  * QIB_MAX_PKT_RCV is the max # if packets processed per receive interrupt.
  79  */
  80 #define QIB_MAX_PKT_RECV 64
  81
  82 struct qlogic_ib_stats qib_stats;
  83
  84 struct pci_dev *qib_get_pci_dev(struct rvt_dev_info *rdi)
  85 {
  86         struct qib_ibdev *ibdev = container_of(rdi, struct qib_ibdev, rdi);
  87         struct qib_devdata *dd = container_of(ibdev,
  88                                               struct qib_devdata, verbs_dev);
  89         return dd->pcidev;
  90 }
  91
  92 /*
  93  * Return count of units with at least one port ACTIVE.
  94  */
  95 int qib_count_active_units(void)
  96 {
  97         struct qib_devdata *dd;
  98         struct qib_pportdata *ppd;
  99         unsigned long flags;
 100         int pidx, nunits_active = 0;
 101
 102         spin_lock_irqsave(&qib_devs_lock, flags);
 103         list_for_each_entry(dd, &qib_dev_list, list) {
 104                 if (!(dd->flags & QIB_PRESENT) || !dd->kregbase)
 105                         continue;
 106                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 107                         ppd = dd->pport + pidx;
 108                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 109                                          QIBL_LINKARMED | QIBL_LINKACTIVE))) {
 110                                 nunits_active++;
 111                                 break;
 112                         }
 113                 }
 114         }
 115         spin_unlock_irqrestore(&qib_devs_lock, flags);
 116         return nunits_active;
 117 }
 118
 119 /*
 120  * Return count of all units, optionally return in arguments
 121  * the number of usable (present) units, and the number of
 122  * ports that are up.
 123  */
 124 int qib_count_units(int *npresentp, int *nupp)
 125 {
 126         int nunits = 0, npresent = 0, nup = 0;
 127         struct qib_devdata *dd;
 128         unsigned long flags;
 129         int pidx;
 130         struct qib_pportdata *ppd;
 131
 132         spin_lock_irqsave(&qib_devs_lock, flags);
 133
 134         list_for_each_entry(dd, &qib_dev_list, list) {
 135                 nunits++;
 136                 if ((dd->flags & QIB_PRESENT) && dd->kregbase)
 137                         npresent++;
 138                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 139                         ppd = dd->pport + pidx;
 140                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 141                                          QIBL_LINKARMED | QIBL_LINKACTIVE)))
 142                                 nup++;
 143                 }
 144         }
 145
 146         spin_unlock_irqrestore(&qib_devs_lock, flags);
 147
 148         if (npresentp)
 149                 *npresentp = npresent;
 150         if (nupp)
 151                 *nupp = nup;
 152
 153         return nunits;
 154 }
 155
 156 /**
 157  * qib_wait_linkstate - wait for an IB link state change to occur
 158  * @dd: the qlogic_ib device
 159  * @state: the state to wait for
 160  * @msecs: the number of milliseconds to wait
 161  *
 162  * wait up to msecs milliseconds for IB link state change to occur for
 163  * now, take the easy polling route.  Currently used only by
 164  * qib_set_linkstate.  Returns 0 if state reached, otherwise
 165  * -ETIMEDOUT state can have multiple states set, for any of several
 166  * transitions.
 167  */
 168 int qib_wait_linkstate(struct qib_pportdata *ppd, u32 state, int msecs)
 169 {
 170         int ret;
 171         unsigned long flags;
 172
 173         spin_lock_irqsave(&ppd->lflags_lock, flags);
 174         if (ppd->state_wanted) {
 175                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 176                 ret = -EBUSY;
 177                 goto bail;
 178         }
 179         ppd->state_wanted = state;
 180         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 181         wait_event_interruptible_timeout(ppd->state_wait,
 182                                          (ppd->lflags & state),
 183                                          msecs_to_jiffies(msecs));
 184         spin_lock_irqsave(&ppd->lflags_lock, flags);
 185         ppd->state_wanted = 0;
 186         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 187
 188         if (!(ppd->lflags & state))
 189                 ret = -ETIMEDOUT;
 190         else
 191                 ret = 0;
 192 bail:
 193         return ret;
 194 }
 195
 196 int qib_set_linkstate(struct qib_pportdata *ppd, u8 newstate)
 197 {
 198         u32 lstate;
 199         int ret;
 200         struct qib_devdata *dd = ppd->dd;
 201         unsigned long flags;
 202
 203         switch (newstate) {
 204         case QIB_IB_LINKDOWN_ONLY:
 205                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 206                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_NOP);
 207                 /* don't wait */
 208                 ret = 0;
 209                 goto bail;
 210
 211         case QIB_IB_LINKDOWN:
 212                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 213                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_POLL);
 214                 /* don't wait */
 215                 ret = 0;
 216                 goto bail;
 217
 218         case QIB_IB_LINKDOWN_SLEEP:
 219                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 220                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_SLEEP);
 221                 /* don't wait */
 222                 ret = 0;
 223                 goto bail;
 224
 225         case QIB_IB_LINKDOWN_DISABLE:
 226                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 227                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_DISABLE);
 228                 /* don't wait */
 229                 ret = 0;
 230                 goto bail;
 231
 232         case QIB_IB_LINKARM:
 233                 if (ppd->lflags & QIBL_LINKARMED) {
 234                         ret = 0;
 235                         goto bail;
 236                 }
 237                 if (!(ppd->lflags & (QIBL_LINKINIT | QIBL_LINKACTIVE))) {
 238                         ret = -EINVAL;
 239                         goto bail;
 240                 }
 241                 /*
 242                  * Since the port can be ACTIVE when we ask for ARMED,
 243                  * clear QIBL_LINKV so we can wait for a transition.
 244                  * If the link isn't ARMED, then something else happened
 245                  * and there is no point waiting for ARMED.
 246                  */
 247                 spin_lock_irqsave(&ppd->lflags_lock, flags);
 248                 ppd->lflags &= ~QIBL_LINKV;
 249                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 250                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 251                                  IB_LINKCMD_ARMED | IB_LINKINITCMD_NOP);
 252                 lstate = QIBL_LINKV;
 253                 break;
 254
 255         case QIB_IB_LINKACTIVE:
 256                 if (ppd->lflags & QIBL_LINKACTIVE) {
 257                         ret = 0;
 258                         goto bail;
 259                 }
 260                 if (!(ppd->lflags & QIBL_LINKARMED)) {
 261                         ret = -EINVAL;
 262                         goto bail;
 263                 }
 264                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 265                                  IB_LINKCMD_ACTIVE | IB_LINKINITCMD_NOP);
 266                 lstate = QIBL_LINKACTIVE;
 267                 break;
 268
 269         default:
 270                 ret = -EINVAL;
 271                 goto bail;
 272         }
 273         ret = qib_wait_linkstate(ppd, lstate, 10);
 274
 275 bail:
 276         return ret;
 277 }
 278
 279 /*
 280  * Get address of eager buffer from it's index (allocated in chunks, not
 281  * contiguous).
 282  */
 283 static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
 284 {
 285         const u32 chunk = etail >> rcd->rcvegrbufs_perchunk_shift;
 286         const u32 idx =  etail & ((u32)rcd->rcvegrbufs_perchunk - 1);
 287
 288         return rcd->rcvegrbuf[chunk] + (idx << rcd->dd->rcvegrbufsize_shift);
 289 }
 290
 291 /*
 292  * Returns 1 if error was a CRC, else 0.
 293  * Needed for some chip's synthesized error counters.
 294  */
 295 static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
 296                           u32 ctxt, u32 eflags, u32 l, u32 etail,
 297                           __le32 *rhf_addr, struct qib_message_header *rhdr)
 298 {
 299         u32 ret = 0;
 300
 301         if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
 302                 ret = 1;
 303         else if (eflags == QLOGIC_IB_RHF_H_TIDERR) {
 304                 /* For TIDERR and RC QPs premptively schedule a NAK */
 305                 struct ib_header *hdr = (struct ib_header *)rhdr;
 306                 struct ib_other_headers *ohdr = NULL;
 307                 struct qib_ibport *ibp = &ppd->ibport_data;
 308                 struct qib_devdata *dd = ppd->dd;
 309                 struct rvt_dev_info *rdi = &dd->verbs_dev.rdi;
 310                 struct rvt_qp *qp = NULL;
 311                 u32 tlen = qib_hdrget_length_in_bytes(rhf_addr);
 312                 u16 lid  = be16_to_cpu(hdr->lrh[1]);
 313                 int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 314                 u32 qp_num;
 315                 u32 opcode;
 316                 u32 psn;
 317                 int diff;
 318
 319                 /* Sanity check packet */
 320                 if (tlen < 24)
 321                         goto drop;
 322
 323                 if (lid < be16_to_cpu(IB_MULTICAST_LID_BASE)) {
 324                         lid &= ~((1 << ppd->lmc) - 1);
 325                         if (unlikely(lid != ppd->lid))
 326                                 goto drop;
 327                 }
 328
 329                 /* Check for GRH */
 330                 if (lnh == QIB_LRH_BTH)
 331                         ohdr = &hdr->u.oth;
 332                 else if (lnh == QIB_LRH_GRH) {
 333                         u32 vtf;
 334
 335                         ohdr = &hdr->u.l.oth;
 336                         if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
 337                                 goto drop;
 338                         vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
 339                         if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
 340                                 goto drop;
 341                 } else
 342                         goto drop;
 343
 344                 /* Get opcode and PSN from packet */
 345                 opcode = be32_to_cpu(ohdr->bth[0]);
 346                 opcode >>= 24;
 347                 psn = be32_to_cpu(ohdr->bth[2]);
 348
 349                 /* Get the destination QP number. */
 350                 qp_num = be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
 351                 if (qp_num != QIB_MULTICAST_QPN) {
 352                         int ruc_res;
 353
 354                         rcu_read_lock();
 355                         qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num);
 356                         if (!qp) {
 357                                 rcu_read_unlock();
 358                                 goto drop;
 359                         }
 360
 361                         /*
 362                          * Handle only RC QPs - for other QP types drop error
 363                          * packet.
 364                          */
 365                         spin_lock(&qp->r_lock);
 366
 367                         /* Check for valid receive state. */
 368                         if (!(ib_rvt_state_ops[qp->state] &
 369                               RVT_PROCESS_RECV_OK)) {
 370                                 ibp->rvp.n_pkt_drops++;
 371                                 goto unlock;
 372                         }
 373
 374                         switch (qp->ibqp.qp_type) {
 375                         case IB_QPT_RC:
 376                                 ruc_res =
 377                                         qib_ruc_check_hdr(
 378                                                 ibp, hdr,
 379                                                 lnh == QIB_LRH_GRH,
 380                                                 qp,
 381                                                 be32_to_cpu(ohdr->bth[0]));
 382                                 if (ruc_res)
 383                                         goto unlock;
 384
 385                                 /* Only deal with RDMA Writes for now */
 386                                 if (opcode <
 387                                     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
 388                                         diff = qib_cmp24(psn, qp->r_psn);
 389                                         if (!qp->r_nak_state && diff >= 0) {
 390                                                 ibp->rvp.n_rc_seqnak++;
 391                                                 qp->r_nak_state =
 392                                                         IB_NAK_PSN_ERROR;
 393                                                 /* Use the expected PSN. */
 394                                                 qp->r_ack_psn = qp->r_psn;
 395                                                 /*
 396                                                  * Wait to send the sequence
 397                                                  * NAK until all packets
 398                                                  * in the receive queue have
 399                                                  * been processed.
 400                                                  * Otherwise, we end up
 401                                                  * propagating congestion.
 402                                                  */
 403                                                 if (list_empty(&qp->rspwait)) {
 404                                                         qp->r_flags |=
 405                                                                 RVT_R_RSP_NAK;
 406                                                         rvt_get_qp(qp);
 407                                                         list_add_tail(
 408                                                          &qp->rspwait,
 409                                                          &rcd->qp_wait_list);
 410                                                 }
 411                                         } /* Out of sequence NAK */
 412                                 } /* QP Request NAKs */
 413                                 break;
 414                         case IB_QPT_SMI:
 415                         case IB_QPT_GSI:
 416                         case IB_QPT_UD:
 417                         case IB_QPT_UC:
 418                         default:
 419                                 /* For now don't handle any other QP types */
 420                                 break;
 421                         }
 422
 423 unlock:
 424                         spin_unlock(&qp->r_lock);
 425                         rcu_read_unlock();
 426                 } /* Unicast QP */
 427         } /* Valid packet with TIDErr */
 428
 429 drop:
 430         return ret;
 431 }
 432
 433 /*
 434  * qib_kreceive - receive a packet
 435  * @rcd: the qlogic_ib context
 436  * @llic: gets count of good packets needed to clear lli,
 437  *          (used with chips that need need to track crcs for lli)
 438  *
 439  * called from interrupt handler for errors or receive interrupt
 440  * Returns number of CRC error packets, needed by some chips for
 441  * local link integrity tracking.   crcs are adjusted down by following
 442  * good packets, if any, and count of good packets is also tracked.
 443  */
 444 u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
 445 {
 446         struct qib_devdata *dd = rcd->dd;
 447         struct qib_pportdata *ppd = rcd->ppd;
 448         __le32 *rhf_addr;
 449         void *ebuf;
 450         const u32 rsize = dd->rcvhdrentsize;        /* words */
 451         const u32 maxcnt = dd->rcvhdrcnt * rsize;   /* words */
 452         u32 etail = -1, l, hdrqtail;
 453         struct qib_message_header *hdr;
 454         u32 eflags, etype, tlen, i = 0, updegr = 0, crcs = 0;
 455         int last;
 456         u64 lval;
 457         struct rvt_qp *qp, *nqp;
 458
 459         l = rcd->head;
 460         rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
 461         if (dd->flags & QIB_NODMA_RTAIL) {
 462                 u32 seq = qib_hdrget_seq(rhf_addr);
 463
 464                 if (seq != rcd->seq_cnt)
 465                         goto bail;
 466                 hdrqtail = 0;
 467         } else {
 468                 hdrqtail = qib_get_rcvhdrtail(rcd);
 469                 if (l == hdrqtail)
 470                         goto bail;
 471                 smp_rmb();  /* prevent speculative reads of dma'ed hdrq */
 472         }
 473
 474         for (last = 0, i = 1; !last; i += !last) {
 475                 hdr = dd->f_get_msgheader(dd, rhf_addr);
 476                 eflags = qib_hdrget_err_flags(rhf_addr);
 477                 etype = qib_hdrget_rcv_type(rhf_addr);
 478                 /* total length */
 479                 tlen = qib_hdrget_length_in_bytes(rhf_addr);
 480                 ebuf = NULL;
 481                 if ((dd->flags & QIB_NODMA_RTAIL) ?
 482                     qib_hdrget_use_egr_buf(rhf_addr) :
 483                     (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
 484                         etail = qib_hdrget_index(rhf_addr);
 485                         updegr = 1;
 486                         if (tlen > sizeof(*hdr) ||
 487                             etype >= RCVHQ_RCV_TYPE_NON_KD) {
 488                                 ebuf = qib_get_egrbuf(rcd, etail);
 489                                 prefetch_range(ebuf, tlen - sizeof(*hdr));
 490                         }
 491                 }
 492                 if (!eflags) {
 493                         u16 lrh_len = be16_to_cpu(hdr->lrh[2]) << 2;
 494
 495                         if (lrh_len != tlen) {
 496                                 qib_stats.sps_lenerrs++;
 497                                 goto move_along;
 498                         }
 499                 }
 500                 if (etype == RCVHQ_RCV_TYPE_NON_KD && !eflags &&
 501                     ebuf == NULL &&
 502                     tlen > (dd->rcvhdrentsize - 2 + 1 -
 503                                 qib_hdrget_offset(rhf_addr)) << 2) {
 504                         goto move_along;
 505                 }
 506
 507                 /*
 508                  * Both tiderr and qibhdrerr are set for all plain IB
 509                  * packets; only qibhdrerr should be set.
 510                  */
 511                 if (unlikely(eflags))
 512                         crcs += qib_rcv_hdrerr(rcd, ppd, rcd->ctxt, eflags, l,
 513                                                etail, rhf_addr, hdr);
 514                 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
 515                         qib_ib_rcv(rcd, hdr, ebuf, tlen);
 516                         if (crcs)
 517                                 crcs--;
 518                         else if (llic && *llic)
 519                                 --*llic;
 520                 }
 521 move_along:
 522                 l += rsize;
 523                 if (l >= maxcnt)
 524                         l = 0;
 525                 if (i == QIB_MAX_PKT_RECV)
 526                         last = 1;
 527
 528                 rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
 529                 if (dd->flags & QIB_NODMA_RTAIL) {
 530                         u32 seq = qib_hdrget_seq(rhf_addr);
 531
 532                         if (++rcd->seq_cnt > 13)
 533                                 rcd->seq_cnt = 1;
 534                         if (seq != rcd->seq_cnt)
 535                                 last = 1;
 536                 } else if (l == hdrqtail)
 537                         last = 1;
 538                 /*
 539                  * Update head regs etc., every 16 packets, if not last pkt,
 540                  * to help prevent rcvhdrq overflows, when many packets
 541                  * are processed and queue is nearly full.
 542                  * Don't request an interrupt for intermediate updates.
 543                  */
 544                 lval = l;
 545                 if (!last && !(i & 0xf)) {
 546                         dd->f_update_usrhead(rcd, lval, updegr, etail, i);
 547                         updegr = 0;
 548                 }
 549         }
 550
 551         rcd->head = l;
 552
 553         /*
 554          * Iterate over all QPs waiting to respond.
 555          * The list won't change since the IRQ is only run on one CPU.
 556          */
 557         list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
 558                 list_del_init(&qp->rspwait);
 559                 if (qp->r_flags & RVT_R_RSP_NAK) {
 560                         qp->r_flags &= ~RVT_R_RSP_NAK;
 561                         qib_send_rc_ack(qp);
 562                 }
 563                 if (qp->r_flags & RVT_R_RSP_SEND) {
 564                         unsigned long flags;
 565
 566                         qp->r_flags &= ~RVT_R_RSP_SEND;
 567                         spin_lock_irqsave(&qp->s_lock, flags);
 568                         if (ib_rvt_state_ops[qp->state] &
 569                                         RVT_PROCESS_OR_FLUSH_SEND)
 570                                 qib_schedule_send(qp);
 571                         spin_unlock_irqrestore(&qp->s_lock, flags);
 572                 }
 573                 rvt_put_qp(qp);
 574         }
 575
 576 bail:
 577         /* Report number of packets consumed */
 578         if (npkts)
 579                 *npkts = i;
 580
 581         /*
 582          * Always write head at end, and setup rcv interrupt, even
 583          * if no packets were processed.
 584          */
 585         lval = (u64)rcd->head | dd->rhdrhead_intr_off;
 586         dd->f_update_usrhead(rcd, lval, updegr, etail, i);
 587         return crcs;
 588 }
 589
 590 /**
 591  * qib_set_mtu - set the MTU
 592  * @ppd: the perport data
 593  * @arg: the new MTU
 594  *
 595  * We can handle "any" incoming size, the issue here is whether we
 596  * need to restrict our outgoing size.   For now, we don't do any
 597  * sanity checking on this, and we don't deal with what happens to
 598  * programs that are already running when the size changes.
 599  * NOTE: changing the MTU will usually cause the IBC to go back to
 600  * link INIT state...
 601  */
 602 int qib_set_mtu(struct qib_pportdata *ppd, u16 arg)
 603 {
 604         u32 piosize;
 605         int ret, chk;
 606
 607         if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
 608             arg != 4096) {
 609                 ret = -EINVAL;
 610                 goto bail;
 611         }
 612         chk = ib_mtu_enum_to_int(qib_ibmtu);
 613         if (chk > 0 && arg > chk) {
 614                 ret = -EINVAL;
 615                 goto bail;
 616         }
 617
 618         piosize = ppd->ibmaxlen;
 619         ppd->ibmtu = arg;
 620
 621         if (arg >= (piosize - QIB_PIO_MAXIBHDR)) {
 622                 /* Only if it's not the initial value (or reset to it) */
 623                 if (piosize != ppd->init_ibmaxlen) {
 624                         if (arg > piosize && arg <= ppd->init_ibmaxlen)
 625                                 piosize = ppd->init_ibmaxlen - 2 * sizeof(u32);
 626                         ppd->ibmaxlen = piosize;
 627                 }
 628         } else if ((arg + QIB_PIO_MAXIBHDR) != ppd->ibmaxlen) {
 629                 piosize = arg + QIB_PIO_MAXIBHDR - 2 * sizeof(u32);
 630                 ppd->ibmaxlen = piosize;
 631         }
 632
 633         ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_MTU, 0);
 634
 635         ret = 0;
 636
 637 bail:
 638         return ret;
 639 }
 640
 641 int qib_set_lid(struct qib_pportdata *ppd, u32 lid, u8 lmc)
 642 {
 643         struct qib_devdata *dd = ppd->dd;
 644
 645         ppd->lid = lid;
 646         ppd->lmc = lmc;
 647
 648         dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LIDLMC,
 649                          lid | (~((1U << lmc) - 1)) << 16);
 650
 651         qib_devinfo(dd->pcidev, "IB%u:%u got a lid: 0x%x\n",
 652                     dd->unit, ppd->port, lid);
 653
 654         return 0;
 655 }
 656
 657 /*
 658  * Following deal with the "obviously simple" task of overriding the state
 659  * of the LEDS, which normally indicate link physical and logical status.
 660  * The complications arise in dealing with different hardware mappings
 661  * and the board-dependent routine being called from interrupts.
 662  * and then there's the requirement to _flash_ them.
 663  */
 664 #define LED_OVER_FREQ_SHIFT 8
 665 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
 666 /* Below is "non-zero" to force override, but both actual LEDs are off */
 667 #define LED_OVER_BOTH_OFF (8)
 668
 669 static void qib_run_led_override(struct timer_list *t)
 670 {
 671         struct qib_pportdata *ppd = from_timer(ppd, t,
 672                                                     led_override_timer);
 673         struct qib_devdata *dd = ppd->dd;
 674         int timeoff;
 675         int ph_idx;
 676
 677         if (!(dd->flags & QIB_INITTED))
 678                 return;
 679
 680         ph_idx = ppd->led_override_phase++ & 1;
 681         ppd->led_override = ppd->led_override_vals[ph_idx];
 682         timeoff = ppd->led_override_timeoff;
 683
 684         dd->f_setextled(ppd, 1);
 685         /*
 686          * don't re-fire the timer if user asked for it to be off; we let
 687          * it fire one more time after they turn it off to simplify
 688          */
 689         if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 690                 mod_timer(&ppd->led_override_timer, jiffies + timeoff);
 691 }
 692
 693 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val)
 694 {
 695         struct qib_devdata *dd = ppd->dd;
 696         int timeoff, freq;
 697
 698         if (!(dd->flags & QIB_INITTED))
 699                 return;
 700
 701         /* First check if we are blinking. If not, use 1HZ polling */
 702         timeoff = HZ;
 703         freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
 704
 705         if (freq) {
 706                 /* For blink, set each phase from one nybble of val */
 707                 ppd->led_override_vals[0] = val & 0xF;
 708                 ppd->led_override_vals[1] = (val >> 4) & 0xF;
 709                 timeoff = (HZ << 4)/freq;
 710         } else {
 711                 /* Non-blink set both phases the same. */
 712                 ppd->led_override_vals[0] = val & 0xF;
 713                 ppd->led_override_vals[1] = val & 0xF;
 714         }
 715         ppd->led_override_timeoff = timeoff;
 716
 717         /*
 718          * If the timer has not already been started, do so. Use a "quick"
 719          * timeout so the function will be called soon, to look at our request.
 720          */
 721         if (atomic_inc_return(&ppd->led_override_timer_active) == 1) {
 722                 /* Need to start timer */
 723                 timer_setup(&ppd->led_override_timer, qib_run_led_override, 0);
 724                 ppd->led_override_timer.expires = jiffies + 1;
 725                 add_timer(&ppd->led_override_timer);
 726         } else {
 727                 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 728                         mod_timer(&ppd->led_override_timer, jiffies + 1);
 729                 atomic_dec(&ppd->led_override_timer_active);
 730         }
 731 }
 732
 733 /**
 734  * qib_reset_device - reset the chip if possible
 735  * @unit: the device to reset
 736  *
 737  * Whether or not reset is successful, we attempt to re-initialize the chip
 738  * (that is, much like a driver unload/reload).  We clear the INITTED flag
 739  * so that the various entry points will fail until we reinitialize.  For
 740  * now, we only allow this if no user contexts are open that use chip resources
 741  */
 742 int qib_reset_device(int unit)
 743 {
 744         int ret, i;
 745         struct qib_devdata *dd = qib_lookup(unit);
 746         struct qib_pportdata *ppd;
 747         unsigned long flags;
 748         int pidx;
 749
 750         if (!dd) {
 751                 ret = -ENODEV;
 752                 goto bail;
 753         }
 754
 755         qib_devinfo(dd->pcidev, "Reset on unit %u requested\n", unit);
 756
 757         if (!dd->kregbase || !(dd->flags & QIB_PRESENT)) {
 758                 qib_devinfo(dd->pcidev,
 759                         "Invalid unit number %u or not initialized or not present\n",
 760                         unit);
 761                 ret = -ENXIO;
 762                 goto bail;
 763         }
 764
 765         spin_lock_irqsave(&dd->uctxt_lock, flags);
 766         if (dd->rcd)
 767                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
 768                         if (!dd->rcd[i] || !dd->rcd[i]->cnt)
 769                                 continue;
 770                         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 771                         ret = -EBUSY;
 772                         goto bail;
 773                 }
 774         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 775
 776         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 777                 ppd = dd->pport + pidx;
 778                 if (atomic_read(&ppd->led_override_timer_active)) {
 779                         /* Need to stop LED timer, _then_ shut off LEDs */
 780                         del_timer_sync(&ppd->led_override_timer);
 781                         atomic_set(&ppd->led_override_timer_active, 0);
 782                 }
 783
 784                 /* Shut off LEDs after we are sure timer is not running */
 785                 ppd->led_override = LED_OVER_BOTH_OFF;
 786                 dd->f_setextled(ppd, 0);
 787                 if (dd->flags & QIB_HAS_SEND_DMA)
 788                         qib_teardown_sdma(ppd);
 789         }
 790
 791         ret = dd->f_reset(dd);
 792         if (ret == 1)
 793                 ret = qib_init(dd, 1);
 794         else
 795                 ret = -EAGAIN;
 796         if (ret)
 797                 qib_dev_err(dd,
 798                         "Reinitialize unit %u after reset failed with %d\n",
 799                         unit, ret);
 800         else
 801                 qib_devinfo(dd->pcidev,
 802                         "Reinitialized unit %u after resetting\n",
 803                         unit);
 804
 805 bail:
 806         return ret;
 807 }