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 <ibsupport@intel.com>");
  68 MODULE_DESCRIPTION("Intel IB driver");
  69 MODULE_VERSION(QIB_DRIVER_VERSION);
  70
  71 /*
  72  * QIB_PIO_MAXIBHDR is the max IB header size allowed for in our
  73  * PIO send buffers.  This is well beyond anything currently
  74  * defined in the InfiniBand spec.
  75  */
  76 #define QIB_PIO_MAXIBHDR 128
  77
  78 /*
  79  * QIB_MAX_PKT_RCV is the max # if packets processed per receive interrupt.
  80  */
  81 #define QIB_MAX_PKT_RECV 64
  82
  83 struct qlogic_ib_stats qib_stats;
  84
  85 const char *qib_get_unit_name(int unit)
  86 {
  87         static char iname[16];
  88
  89         snprintf(iname, sizeof iname, "infinipath%u", unit);
  90         return iname;
  91 }
  92
  93 /*
  94  * Return count of units with at least one port ACTIVE.
  95  */
  96 int qib_count_active_units(void)
  97 {
  98         struct qib_devdata *dd;
  99         struct qib_pportdata *ppd;
 100         unsigned long flags;
 101         int pidx, nunits_active = 0;
 102
 103         spin_lock_irqsave(&qib_devs_lock, flags);
 104         list_for_each_entry(dd, &qib_dev_list, list) {
 105                 if (!(dd->flags & QIB_PRESENT) || !dd->kregbase)
 106                         continue;
 107                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 108                         ppd = dd->pport + pidx;
 109                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 110                                          QIBL_LINKARMED | QIBL_LINKACTIVE))) {
 111                                 nunits_active++;
 112                                 break;
 113                         }
 114                 }
 115         }
 116         spin_unlock_irqrestore(&qib_devs_lock, flags);
 117         return nunits_active;
 118 }
 119
 120 /*
 121  * Return count of all units, optionally return in arguments
 122  * the number of usable (present) units, and the number of
 123  * ports that are up.
 124  */
 125 int qib_count_units(int *npresentp, int *nupp)
 126 {
 127         int nunits = 0, npresent = 0, nup = 0;
 128         struct qib_devdata *dd;
 129         unsigned long flags;
 130         int pidx;
 131         struct qib_pportdata *ppd;
 132
 133         spin_lock_irqsave(&qib_devs_lock, flags);
 134
 135         list_for_each_entry(dd, &qib_dev_list, list) {
 136                 nunits++;
 137                 if ((dd->flags & QIB_PRESENT) && dd->kregbase)
 138                         npresent++;
 139                 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 140                         ppd = dd->pport + pidx;
 141                         if (ppd->lid && (ppd->lflags & (QIBL_LINKINIT |
 142                                          QIBL_LINKARMED | QIBL_LINKACTIVE)))
 143                                 nup++;
 144                 }
 145         }
 146
 147         spin_unlock_irqrestore(&qib_devs_lock, flags);
 148
 149         if (npresentp)
 150                 *npresentp = npresent;
 151         if (nupp)
 152                 *nupp = nup;
 153
 154         return nunits;
 155 }
 156
 157 /**
 158  * qib_wait_linkstate - wait for an IB link state change to occur
 159  * @dd: the qlogic_ib device
 160  * @state: the state to wait for
 161  * @msecs: the number of milliseconds to wait
 162  *
 163  * wait up to msecs milliseconds for IB link state change to occur for
 164  * now, take the easy polling route.  Currently used only by
 165  * qib_set_linkstate.  Returns 0 if state reached, otherwise
 166  * -ETIMEDOUT state can have multiple states set, for any of several
 167  * transitions.
 168  */
 169 int qib_wait_linkstate(struct qib_pportdata *ppd, u32 state, int msecs)
 170 {
 171         int ret;
 172         unsigned long flags;
 173
 174         spin_lock_irqsave(&ppd->lflags_lock, flags);
 175         if (ppd->state_wanted) {
 176                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 177                 ret = -EBUSY;
 178                 goto bail;
 179         }
 180         ppd->state_wanted = state;
 181         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 182         wait_event_interruptible_timeout(ppd->state_wait,
 183                                          (ppd->lflags & state),
 184                                          msecs_to_jiffies(msecs));
 185         spin_lock_irqsave(&ppd->lflags_lock, flags);
 186         ppd->state_wanted = 0;
 187         spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 188
 189         if (!(ppd->lflags & state))
 190                 ret = -ETIMEDOUT;
 191         else
 192                 ret = 0;
 193 bail:
 194         return ret;
 195 }
 196
 197 int qib_set_linkstate(struct qib_pportdata *ppd, u8 newstate)
 198 {
 199         u32 lstate;
 200         int ret;
 201         struct qib_devdata *dd = ppd->dd;
 202         unsigned long flags;
 203
 204         switch (newstate) {
 205         case QIB_IB_LINKDOWN_ONLY:
 206                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 207                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_NOP);
 208                 /* don't wait */
 209                 ret = 0;
 210                 goto bail;
 211
 212         case QIB_IB_LINKDOWN:
 213                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 214                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_POLL);
 215                 /* don't wait */
 216                 ret = 0;
 217                 goto bail;
 218
 219         case QIB_IB_LINKDOWN_SLEEP:
 220                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 221                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_SLEEP);
 222                 /* don't wait */
 223                 ret = 0;
 224                 goto bail;
 225
 226         case QIB_IB_LINKDOWN_DISABLE:
 227                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 228                                  IB_LINKCMD_DOWN | IB_LINKINITCMD_DISABLE);
 229                 /* don't wait */
 230                 ret = 0;
 231                 goto bail;
 232
 233         case QIB_IB_LINKARM:
 234                 if (ppd->lflags & QIBL_LINKARMED) {
 235                         ret = 0;
 236                         goto bail;
 237                 }
 238                 if (!(ppd->lflags & (QIBL_LINKINIT | QIBL_LINKACTIVE))) {
 239                         ret = -EINVAL;
 240                         goto bail;
 241                 }
 242                 /*
 243                  * Since the port can be ACTIVE when we ask for ARMED,
 244                  * clear QIBL_LINKV so we can wait for a transition.
 245                  * If the link isn't ARMED, then something else happened
 246                  * and there is no point waiting for ARMED.
 247                  */
 248                 spin_lock_irqsave(&ppd->lflags_lock, flags);
 249                 ppd->lflags &= ~QIBL_LINKV;
 250                 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
 251                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 252                                  IB_LINKCMD_ARMED | IB_LINKINITCMD_NOP);
 253                 lstate = QIBL_LINKV;
 254                 break;
 255
 256         case QIB_IB_LINKACTIVE:
 257                 if (ppd->lflags & QIBL_LINKACTIVE) {
 258                         ret = 0;
 259                         goto bail;
 260                 }
 261                 if (!(ppd->lflags & QIBL_LINKARMED)) {
 262                         ret = -EINVAL;
 263                         goto bail;
 264                 }
 265                 dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LSTATE,
 266                                  IB_LINKCMD_ACTIVE | IB_LINKINITCMD_NOP);
 267                 lstate = QIBL_LINKACTIVE;
 268                 break;
 269
 270         default:
 271                 ret = -EINVAL;
 272                 goto bail;
 273         }
 274         ret = qib_wait_linkstate(ppd, lstate, 10);
 275
 276 bail:
 277         return ret;
 278 }
 279
 280 /*
 281  * Get address of eager buffer from it's index (allocated in chunks, not
 282  * contiguous).
 283  */
 284 static inline void *qib_get_egrbuf(const struct qib_ctxtdata *rcd, u32 etail)
 285 {
 286         const u32 chunk = etail >> rcd->rcvegrbufs_perchunk_shift;
 287         const u32 idx =  etail & ((u32)rcd->rcvegrbufs_perchunk - 1);
 288
 289         return rcd->rcvegrbuf[chunk] + (idx << rcd->dd->rcvegrbufsize_shift);
 290 }
 291
 292 /*
 293  * Returns 1 if error was a CRC, else 0.
 294  * Needed for some chip's synthesized error counters.
 295  */
 296 static u32 qib_rcv_hdrerr(struct qib_ctxtdata *rcd, struct qib_pportdata *ppd,
 297                           u32 ctxt, u32 eflags, u32 l, u32 etail,
 298                           __le32 *rhf_addr, struct qib_message_header *rhdr)
 299 {
 300         u32 ret = 0;
 301
 302         if (eflags & (QLOGIC_IB_RHF_H_ICRCERR | QLOGIC_IB_RHF_H_VCRCERR))
 303                 ret = 1;
 304         else if (eflags == QLOGIC_IB_RHF_H_TIDERR) {
 305                 /* For TIDERR and RC QPs premptively schedule a NAK */
 306                 struct qib_ib_header *hdr = (struct qib_ib_header *) rhdr;
 307                 struct qib_other_headers *ohdr = NULL;
 308                 struct qib_ibport *ibp = &ppd->ibport_data;
 309                 struct qib_qp *qp = NULL;
 310                 u32 tlen = qib_hdrget_length_in_bytes(rhf_addr);
 311                 u16 lid  = be16_to_cpu(hdr->lrh[1]);
 312                 int lnh = be16_to_cpu(hdr->lrh[0]) & 3;
 313                 u32 qp_num;
 314                 u32 opcode;
 315                 u32 psn;
 316                 int diff;
 317
 318                 /* Sanity check packet */
 319                 if (tlen < 24)
 320                         goto drop;
 321
 322                 if (lid < QIB_MULTICAST_LID_BASE) {
 323                         lid &= ~((1 << ppd->lmc) - 1);
 324                         if (unlikely(lid != ppd->lid))
 325                                 goto drop;
 326                 }
 327
 328                 /* Check for GRH */
 329                 if (lnh == QIB_LRH_BTH)
 330                         ohdr = &hdr->u.oth;
 331                 else if (lnh == QIB_LRH_GRH) {
 332                         u32 vtf;
 333
 334                         ohdr = &hdr->u.l.oth;
 335                         if (hdr->u.l.grh.next_hdr != IB_GRH_NEXT_HDR)
 336                                 goto drop;
 337                         vtf = be32_to_cpu(hdr->u.l.grh.version_tclass_flow);
 338                         if ((vtf >> IB_GRH_VERSION_SHIFT) != IB_GRH_VERSION)
 339                                 goto drop;
 340                 } else
 341                         goto drop;
 342
 343                 /* Get opcode and PSN from packet */
 344                 opcode = be32_to_cpu(ohdr->bth[0]);
 345                 opcode >>= 24;
 346                 psn = be32_to_cpu(ohdr->bth[2]);
 347
 348                 /* Get the destination QP number. */
 349                 qp_num = be32_to_cpu(ohdr->bth[1]) & QIB_QPN_MASK;
 350                 if (qp_num != QIB_MULTICAST_QPN) {
 351                         int ruc_res;
 352                         qp = qib_lookup_qpn(ibp, qp_num);
 353                         if (!qp)
 354                                 goto drop;
 355
 356                         /*
 357                          * Handle only RC QPs - for other QP types drop error
 358                          * packet.
 359                          */
 360                         spin_lock(&qp->r_lock);
 361
 362                         /* Check for valid receive state. */
 363                         if (!(ib_qib_state_ops[qp->state] &
 364                               QIB_PROCESS_RECV_OK)) {
 365                                 ibp->n_pkt_drops++;
 366                                 goto unlock;
 367                         }
 368
 369                         switch (qp->ibqp.qp_type) {
 370                         case IB_QPT_RC:
 371                                 ruc_res =
 372                                         qib_ruc_check_hdr(
 373                                                 ibp, hdr,
 374                                                 lnh == QIB_LRH_GRH,
 375                                                 qp,
 376                                                 be32_to_cpu(ohdr->bth[0]));
 377                                 if (ruc_res)
 378                                         goto unlock;
 379
 380                                 /* Only deal with RDMA Writes for now */
 381                                 if (opcode <
 382                                     IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
 383                                         diff = qib_cmp24(psn, qp->r_psn);
 384                                         if (!qp->r_nak_state && diff >= 0) {
 385                                                 ibp->n_rc_seqnak++;
 386                                                 qp->r_nak_state =
 387                                                         IB_NAK_PSN_ERROR;
 388                                                 /* Use the expected PSN. */
 389                                                 qp->r_ack_psn = qp->r_psn;
 390                                                 /*
 391                                                  * Wait to send the sequence
 392                                                  * NAK until all packets
 393                                                  * in the receive queue have
 394                                                  * been processed.
 395                                                  * Otherwise, we end up
 396                                                  * propagating congestion.
 397                                                  */
 398                                                 if (list_empty(&qp->rspwait)) {
 399                                                         qp->r_flags |=
 400                                                                 QIB_R_RSP_NAK;
 401                                                         atomic_inc(
 402                                                                 &qp->refcount);
 403                                                         list_add_tail(
 404                                                          &qp->rspwait,
 405                                                          &rcd->qp_wait_list);
 406                                                 }
 407                                         } /* Out of sequence NAK */
 408                                 } /* QP Request NAKs */
 409                                 break;
 410                         case IB_QPT_SMI:
 411                         case IB_QPT_GSI:
 412                         case IB_QPT_UD:
 413                         case IB_QPT_UC:
 414                         default:
 415                                 /* For now don't handle any other QP types */
 416                                 break;
 417                         }
 418
 419 unlock:
 420                         spin_unlock(&qp->r_lock);
 421                         /*
 422                          * Notify qib_destroy_qp() if it is waiting
 423                          * for us to finish.
 424                          */
 425                         if (atomic_dec_and_test(&qp->refcount))
 426                                 wake_up(&qp->wait);
 427                 } /* Unicast QP */
 428         } /* Valid packet with TIDErr */
 429
 430 drop:
 431         return ret;
 432 }
 433
 434 /*
 435  * qib_kreceive - receive a packet
 436  * @rcd: the qlogic_ib context
 437  * @llic: gets count of good packets needed to clear lli,
 438  *          (used with chips that need need to track crcs for lli)
 439  *
 440  * called from interrupt handler for errors or receive interrupt
 441  * Returns number of CRC error packets, needed by some chips for
 442  * local link integrity tracking.   crcs are adjusted down by following
 443  * good packets, if any, and count of good packets is also tracked.
 444  */
 445 u32 qib_kreceive(struct qib_ctxtdata *rcd, u32 *llic, u32 *npkts)
 446 {
 447         struct qib_devdata *dd = rcd->dd;
 448         struct qib_pportdata *ppd = rcd->ppd;
 449         __le32 *rhf_addr;
 450         void *ebuf;
 451         const u32 rsize = dd->rcvhdrentsize;        /* words */
 452         const u32 maxcnt = dd->rcvhdrcnt * rsize;   /* words */
 453         u32 etail = -1, l, hdrqtail;
 454         struct qib_message_header *hdr;
 455         u32 eflags, etype, tlen, i = 0, updegr = 0, crcs = 0;
 456         int last;
 457         u64 lval;
 458         struct qib_qp *qp, *nqp;
 459
 460         l = rcd->head;
 461         rhf_addr = (__le32 *) rcd->rcvhdrq + l + dd->rhf_offset;
 462         if (dd->flags & QIB_NODMA_RTAIL) {
 463                 u32 seq = qib_hdrget_seq(rhf_addr);
 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          * Notify qib_destroy_qp() if it is waiting
 552          * for lookaside_qp to finish.
 553          */
 554         if (rcd->lookaside_qp) {
 555                 if (atomic_dec_and_test(&rcd->lookaside_qp->refcount))
 556                         wake_up(&rcd->lookaside_qp->wait);
 557                 rcd->lookaside_qp = NULL;
 558         }
 559
 560         rcd->head = l;
 561
 562         /*
 563          * Iterate over all QPs waiting to respond.
 564          * The list won't change since the IRQ is only run on one CPU.
 565          */
 566         list_for_each_entry_safe(qp, nqp, &rcd->qp_wait_list, rspwait) {
 567                 list_del_init(&qp->rspwait);
 568                 if (qp->r_flags & QIB_R_RSP_NAK) {
 569                         qp->r_flags &= ~QIB_R_RSP_NAK;
 570                         qib_send_rc_ack(qp);
 571                 }
 572                 if (qp->r_flags & QIB_R_RSP_SEND) {
 573                         unsigned long flags;
 574
 575                         qp->r_flags &= ~QIB_R_RSP_SEND;
 576                         spin_lock_irqsave(&qp->s_lock, flags);
 577                         if (ib_qib_state_ops[qp->state] &
 578                                         QIB_PROCESS_OR_FLUSH_SEND)
 579                                 qib_schedule_send(qp);
 580                         spin_unlock_irqrestore(&qp->s_lock, flags);
 581                 }
 582                 if (atomic_dec_and_test(&qp->refcount))
 583                         wake_up(&qp->wait);
 584         }
 585
 586 bail:
 587         /* Report number of packets consumed */
 588         if (npkts)
 589                 *npkts = i;
 590
 591         /*
 592          * Always write head at end, and setup rcv interrupt, even
 593          * if no packets were processed.
 594          */
 595         lval = (u64)rcd->head | dd->rhdrhead_intr_off;
 596         dd->f_update_usrhead(rcd, lval, updegr, etail, i);
 597         return crcs;
 598 }
 599
 600 /**
 601  * qib_set_mtu - set the MTU
 602  * @ppd: the perport data
 603  * @arg: the new MTU
 604  *
 605  * We can handle "any" incoming size, the issue here is whether we
 606  * need to restrict our outgoing size.   For now, we don't do any
 607  * sanity checking on this, and we don't deal with what happens to
 608  * programs that are already running when the size changes.
 609  * NOTE: changing the MTU will usually cause the IBC to go back to
 610  * link INIT state...
 611  */
 612 int qib_set_mtu(struct qib_pportdata *ppd, u16 arg)
 613 {
 614         u32 piosize;
 615         int ret, chk;
 616
 617         if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
 618             arg != 4096) {
 619                 ret = -EINVAL;
 620                 goto bail;
 621         }
 622         chk = ib_mtu_enum_to_int(qib_ibmtu);
 623         if (chk > 0 && arg > chk) {
 624                 ret = -EINVAL;
 625                 goto bail;
 626         }
 627
 628         piosize = ppd->ibmaxlen;
 629         ppd->ibmtu = arg;
 630
 631         if (arg >= (piosize - QIB_PIO_MAXIBHDR)) {
 632                 /* Only if it's not the initial value (or reset to it) */
 633                 if (piosize != ppd->init_ibmaxlen) {
 634                         if (arg > piosize && arg <= ppd->init_ibmaxlen)
 635                                 piosize = ppd->init_ibmaxlen - 2 * sizeof(u32);
 636                         ppd->ibmaxlen = piosize;
 637                 }
 638         } else if ((arg + QIB_PIO_MAXIBHDR) != ppd->ibmaxlen) {
 639                 piosize = arg + QIB_PIO_MAXIBHDR - 2 * sizeof(u32);
 640                 ppd->ibmaxlen = piosize;
 641         }
 642
 643         ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_MTU, 0);
 644
 645         ret = 0;
 646
 647 bail:
 648         return ret;
 649 }
 650
 651 int qib_set_lid(struct qib_pportdata *ppd, u32 lid, u8 lmc)
 652 {
 653         struct qib_devdata *dd = ppd->dd;
 654         ppd->lid = lid;
 655         ppd->lmc = lmc;
 656
 657         dd->f_set_ib_cfg(ppd, QIB_IB_CFG_LIDLMC,
 658                          lid | (~((1U << lmc) - 1)) << 16);
 659
 660         qib_devinfo(dd->pcidev, "IB%u:%u got a lid: 0x%x\n",
 661                     dd->unit, ppd->port, lid);
 662
 663         return 0;
 664 }
 665
 666 /*
 667  * Following deal with the "obviously simple" task of overriding the state
 668  * of the LEDS, which normally indicate link physical and logical status.
 669  * The complications arise in dealing with different hardware mappings
 670  * and the board-dependent routine being called from interrupts.
 671  * and then there's the requirement to _flash_ them.
 672  */
 673 #define LED_OVER_FREQ_SHIFT 8
 674 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
 675 /* Below is "non-zero" to force override, but both actual LEDs are off */
 676 #define LED_OVER_BOTH_OFF (8)
 677
 678 static void qib_run_led_override(unsigned long opaque)
 679 {
 680         struct qib_pportdata *ppd = (struct qib_pportdata *)opaque;
 681         struct qib_devdata *dd = ppd->dd;
 682         int timeoff;
 683         int ph_idx;
 684
 685         if (!(dd->flags & QIB_INITTED))
 686                 return;
 687
 688         ph_idx = ppd->led_override_phase++ & 1;
 689         ppd->led_override = ppd->led_override_vals[ph_idx];
 690         timeoff = ppd->led_override_timeoff;
 691
 692         dd->f_setextled(ppd, 1);
 693         /*
 694          * don't re-fire the timer if user asked for it to be off; we let
 695          * it fire one more time after they turn it off to simplify
 696          */
 697         if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 698                 mod_timer(&ppd->led_override_timer, jiffies + timeoff);
 699 }
 700
 701 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val)
 702 {
 703         struct qib_devdata *dd = ppd->dd;
 704         int timeoff, freq;
 705
 706         if (!(dd->flags & QIB_INITTED))
 707                 return;
 708
 709         /* First check if we are blinking. If not, use 1HZ polling */
 710         timeoff = HZ;
 711         freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
 712
 713         if (freq) {
 714                 /* For blink, set each phase from one nybble of val */
 715                 ppd->led_override_vals[0] = val & 0xF;
 716                 ppd->led_override_vals[1] = (val >> 4) & 0xF;
 717                 timeoff = (HZ << 4)/freq;
 718         } else {
 719                 /* Non-blink set both phases the same. */
 720                 ppd->led_override_vals[0] = val & 0xF;
 721                 ppd->led_override_vals[1] = val & 0xF;
 722         }
 723         ppd->led_override_timeoff = timeoff;
 724
 725         /*
 726          * If the timer has not already been started, do so. Use a "quick"
 727          * timeout so the function will be called soon, to look at our request.
 728          */
 729         if (atomic_inc_return(&ppd->led_override_timer_active) == 1) {
 730                 /* Need to start timer */
 731                 init_timer(&ppd->led_override_timer);
 732                 ppd->led_override_timer.function = qib_run_led_override;
 733                 ppd->led_override_timer.data = (unsigned long) ppd;
 734                 ppd->led_override_timer.expires = jiffies + 1;
 735                 add_timer(&ppd->led_override_timer);
 736         } else {
 737                 if (ppd->led_override_vals[0] || ppd->led_override_vals[1])
 738                         mod_timer(&ppd->led_override_timer, jiffies + 1);
 739                 atomic_dec(&ppd->led_override_timer_active);
 740         }
 741 }
 742
 743 /**
 744  * qib_reset_device - reset the chip if possible
 745  * @unit: the device to reset
 746  *
 747  * Whether or not reset is successful, we attempt to re-initialize the chip
 748  * (that is, much like a driver unload/reload).  We clear the INITTED flag
 749  * so that the various entry points will fail until we reinitialize.  For
 750  * now, we only allow this if no user contexts are open that use chip resources
 751  */
 752 int qib_reset_device(int unit)
 753 {
 754         int ret, i;
 755         struct qib_devdata *dd = qib_lookup(unit);
 756         struct qib_pportdata *ppd;
 757         unsigned long flags;
 758         int pidx;
 759
 760         if (!dd) {
 761                 ret = -ENODEV;
 762                 goto bail;
 763         }
 764
 765         qib_devinfo(dd->pcidev, "Reset on unit %u requested\n", unit);
 766
 767         if (!dd->kregbase || !(dd->flags & QIB_PRESENT)) {
 768                 qib_devinfo(dd->pcidev,
 769                         "Invalid unit number %u or not initialized or not present\n",
 770                         unit);
 771                 ret = -ENXIO;
 772                 goto bail;
 773         }
 774
 775         spin_lock_irqsave(&dd->uctxt_lock, flags);
 776         if (dd->rcd)
 777                 for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
 778                         if (!dd->rcd[i] || !dd->rcd[i]->cnt)
 779                                 continue;
 780                         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 781                         ret = -EBUSY;
 782                         goto bail;
 783                 }
 784         spin_unlock_irqrestore(&dd->uctxt_lock, flags);
 785
 786         for (pidx = 0; pidx < dd->num_pports; ++pidx) {
 787                 ppd = dd->pport + pidx;
 788                 if (atomic_read(&ppd->led_override_timer_active)) {
 789                         /* Need to stop LED timer, _then_ shut off LEDs */
 790                         del_timer_sync(&ppd->led_override_timer);
 791                         atomic_set(&ppd->led_override_timer_active, 0);
 792                 }
 793
 794                 /* Shut off LEDs after we are sure timer is not running */
 795                 ppd->led_override = LED_OVER_BOTH_OFF;
 796                 dd->f_setextled(ppd, 0);
 797                 if (dd->flags & QIB_HAS_SEND_DMA)
 798                         qib_teardown_sdma(ppd);
 799         }
 800
 801         ret = dd->f_reset(dd);
 802         if (ret == 1)
 803                 ret = qib_init(dd, 1);
 804         else
 805                 ret = -EAGAIN;
 806         if (ret)
 807                 qib_dev_err(dd,
 808                         "Reinitialize unit %u after reset failed with %d\n",
 809                         unit, ret);
 810         else
 811                 qib_devinfo(dd->pcidev,
 812                         "Reinitialized unit %u after resetting\n",
 813                         unit);
 814
 815 bail:
 816         return ret;
 817 }