err.no Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_file_ops.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/pci.h>
  35 #include <linux/poll.h>
  36 #include <linux/cdev.h>
  37 #include <linux/swap.h>
  38 #include <linux/vmalloc.h>
  39 #include <asm/pgtable.h>
  40
  41 #include "ipath_kernel.h"
  42 #include "ips_common.h"
  43 #include "ipath_layer.h"
  44
  45 static int ipath_open(struct inode *, struct file *);
  46 static int ipath_close(struct inode *, struct file *);
  47 static ssize_t ipath_write(struct file *, const char __user *, size_t,
  48                            loff_t *);
  49 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
  50 static int ipath_mmap(struct file *, struct vm_area_struct *);
  51
  52 static struct file_operations ipath_file_ops = {
  53         .owner = THIS_MODULE,
  54         .write = ipath_write,
  55         .open = ipath_open,
  56         .release = ipath_close,
  57         .poll = ipath_poll,
  58         .mmap = ipath_mmap
  59 };
  60
  61 static int ipath_get_base_info(struct ipath_portdata *pd,
  62                                void __user *ubase, size_t ubase_size)
  63 {
  64         int ret = 0;
  65         struct ipath_base_info *kinfo = NULL;
  66         struct ipath_devdata *dd = pd->port_dd;
  67
  68         if (ubase_size < sizeof(*kinfo)) {
  69                 ipath_cdbg(PROC,
  70                            "Base size %lu, need %lu (version mismatch?)\n",
  71                            (unsigned long) ubase_size,
  72                            (unsigned long) sizeof(*kinfo));
  73                 ret = -EINVAL;
  74                 goto bail;
  75         }
  76
  77         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
  78         if (kinfo == NULL) {
  79                 ret = -ENOMEM;
  80                 goto bail;
  81         }
  82
  83         ret = dd->ipath_f_get_base_info(pd, kinfo);
  84         if (ret < 0)
  85                 goto bail;
  86
  87         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
  88         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
  89         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
  90         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
  91         /*
  92          * have to mmap whole thing
  93          */
  94         kinfo->spi_rcv_egrbuftotlen =
  95                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
  96         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
  97         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
  98                 pd->port_rcvegrbuf_chunks;
  99         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt;
 100         /*
 101          * for this use, may be ipath_cfgports summed over all chips that
 102          * are are configured and present
 103          */
 104         kinfo->spi_nports = dd->ipath_cfgports;
 105         /* unit (chip/board) our port is on */
 106         kinfo->spi_unit = dd->ipath_unit;
 107         /* for now, only a single page */
 108         kinfo->spi_tid_maxsize = PAGE_SIZE;
 109
 110         /*
 111          * Doing this per port, and based on the skip value, etc.  This has
 112          * to be the actual buffer size, since the protocol code treats it
 113          * as an array.
 114          *
 115          * These have to be set to user addresses in the user code via mmap.
 116          * These values are used on return to user code for the mmap target
 117          * addresses only.  For 32 bit, same 44 bit address problem, so use
 118          * the physical address, not virtual.  Before 2.6.11, using the
 119          * page_address() macro worked, but in 2.6.11, even that returns the
 120          * full 64 bit address (upper bits all 1's).  So far, using the
 121          * physical addresses (or chip offsets, for chip mapping) works, but
 122          * no doubt some future kernel release will chang that, and we'll be
 123          * on to yet another method of dealing with this
 124          */
 125         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
 126         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
 127         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
 128         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
 129                 (void *) dd->ipath_statusp -
 130                 (void *) dd->ipath_pioavailregs_dma;
 131         kinfo->spi_piobufbase = (u64) pd->port_piobufs;
 132         kinfo->__spi_uregbase =
 133                 dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
 134
 135         kinfo->spi_pioindex = dd->ipath_pbufsport * (pd->port_port - 1);
 136         kinfo->spi_piocnt = dd->ipath_pbufsport;
 137         kinfo->spi_pioalign = dd->ipath_palign;
 138
 139         kinfo->spi_qpair = IPATH_KD_QP;
 140         kinfo->spi_piosize = dd->ipath_ibmaxlen;
 141         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
 142         kinfo->spi_port = pd->port_port;
 143         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
 144         kinfo->spi_hw_version = dd->ipath_revision;
 145
 146         if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
 147                 ret = -EFAULT;
 148
 149 bail:
 150         kfree(kinfo);
 151         return ret;
 152 }
 153
 154 /**
 155  * ipath_tid_update - update a port TID
 156  * @pd: the port
 157  * @ti: the TID information
 158  *
 159  * The new implementation as of Oct 2004 is that the driver assigns
 160  * the tid and returns it to the caller.   To make it easier to
 161  * catch bugs, and to reduce search time, we keep a cursor for
 162  * each port, walking the shadow tid array to find one that's not
 163  * in use.
 164  *
 165  * For now, if we can't allocate the full list, we fail, although
 166  * in the long run, we'll allocate as many as we can, and the
 167  * caller will deal with that by trying the remaining pages later.
 168  * That means that when we fail, we have to mark the tids as not in
 169  * use again, in our shadow copy.
 170  *
 171  * It's up to the caller to free the tids when they are done.
 172  * We'll unlock the pages as they free them.
 173  *
 174  * Also, right now we are locking one page at a time, but since
 175  * the intended use of this routine is for a single group of
 176  * virtually contiguous pages, that should change to improve
 177  * performance.
 178  */
 179 static int ipath_tid_update(struct ipath_portdata *pd,
 180                             const struct ipath_tid_info *ti)
 181 {
 182         int ret = 0, ntids;
 183         u32 tid, porttid, cnt, i, tidcnt;
 184         u16 *tidlist;
 185         struct ipath_devdata *dd = pd->port_dd;
 186         u64 physaddr;
 187         unsigned long vaddr;
 188         u64 __iomem *tidbase;
 189         unsigned long tidmap[8];
 190         struct page **pagep = NULL;
 191
 192         if (!dd->ipath_pageshadow) {
 193                 ret = -ENOMEM;
 194                 goto done;
 195         }
 196
 197         cnt = ti->tidcnt;
 198         if (!cnt) {
 199                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
 200                           (unsigned long long) ti->tidlist);
 201                 /*
 202                  * Should we treat as success?  likely a bug
 203                  */
 204                 ret = -EFAULT;
 205                 goto done;
 206         }
 207         tidcnt = dd->ipath_rcvtidcnt;
 208         if (cnt >= tidcnt) {
 209                 /* make sure it all fits in port_tid_pg_list */
 210                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
 211                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
 212                 cnt = tidcnt;
 213         }
 214         pagep = (struct page **)pd->port_tid_pg_list;
 215         tidlist = (u16 *) (&pagep[cnt]);
 216
 217         memset(tidmap, 0, sizeof(tidmap));
 218         tid = pd->port_tidcursor;
 219         /* before decrement; chip actual # */
 220         porttid = pd->port_port * tidcnt;
 221         ntids = tidcnt;
 222         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
 223                                    dd->ipath_rcvtidbase +
 224                                    porttid * sizeof(*tidbase));
 225
 226         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
 227                    pd->port_port, cnt, tid, tidbase);
 228
 229         /* virtual address of first page in transfer */
 230         vaddr = ti->tidvaddr;
 231         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
 232                        cnt * PAGE_SIZE)) {
 233                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
 234                           (void *)vaddr, cnt);
 235                 ret = -EFAULT;
 236                 goto done;
 237         }
 238         ret = ipath_get_user_pages(vaddr, cnt, pagep);
 239         if (ret) {
 240                 if (ret == -EBUSY) {
 241                         ipath_dbg("Failed to lock addr %p, %u pages "
 242                                   "(already locked)\n",
 243                                   (void *) vaddr, cnt);
 244                         /*
 245                          * for now, continue, and see what happens but with
 246                          * the new implementation, this should never happen,
 247                          * unless perhaps the user has mpin'ed the pages
 248                          * themselves (something we need to test)
 249                          */
 250                         ret = 0;
 251                 } else {
 252                         dev_info(&dd->pcidev->dev,
 253                                  "Failed to lock addr %p, %u pages: "
 254                                  "errno %d\n", (void *) vaddr, cnt, -ret);
 255                         goto done;
 256                 }
 257         }
 258         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
 259                 for (; ntids--; tid++) {
 260                         if (tid == tidcnt)
 261                                 tid = 0;
 262                         if (!dd->ipath_pageshadow[porttid + tid])
 263                                 break;
 264                 }
 265                 if (ntids < 0) {
 266                         /*
 267                          * oops, wrapped all the way through their TIDs,
 268                          * and didn't have enough free; see comments at
 269                          * start of routine
 270                          */
 271                         ipath_dbg("Not enough free TIDs for %u pages "
 272                                   "(index %d), failing\n", cnt, i);
 273                         i--;    /* last tidlist[i] not filled in */
 274                         ret = -ENOMEM;
 275                         break;
 276                 }
 277                 tidlist[i] = tid;
 278                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
 279                            "vaddr %lx\n", i, tid, vaddr);
 280                 /* we "know" system pages and TID pages are same size */
 281                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
 282                 /*
 283                  * don't need atomic or it's overhead
 284                  */
 285                 __set_bit(tid, tidmap);
 286                 physaddr = page_to_phys(pagep[i]);
 287                 ipath_stats.sps_pagelocks++;
 288                 ipath_cdbg(VERBOSE,
 289                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
 290                            tid, vaddr, (unsigned long long) physaddr,
 291                            pagep[i]);
 292                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr);
 293                 /*
 294                  * don't check this tid in ipath_portshadow, since we
 295                  * just filled it in; start with the next one.
 296                  */
 297                 tid++;
 298         }
 299
 300         if (ret) {
 301                 u32 limit;
 302         cleanup:
 303                 /* jump here if copy out of updated info failed... */
 304                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
 305                           -ret, i, cnt);
 306                 /* same code that's in ipath_free_tid() */
 307                 limit = sizeof(tidmap) * BITS_PER_BYTE;
 308                 if (limit > tidcnt)
 309                         /* just in case size changes in future */
 310                         limit = tidcnt;
 311                 tid = find_first_bit((const unsigned long *)tidmap, limit);
 312                 for (; tid < limit; tid++) {
 313                         if (!test_bit(tid, tidmap))
 314                                 continue;
 315                         if (dd->ipath_pageshadow[porttid + tid]) {
 316                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
 317                                            tid);
 318                                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
 319                                                     dd->ipath_tidinvalid);
 320                                 dd->ipath_pageshadow[porttid + tid] = NULL;
 321                                 ipath_stats.sps_pageunlocks++;
 322                         }
 323                 }
 324                 ipath_release_user_pages(pagep, cnt);
 325         } else {
 326                 /*
 327                  * Copy the updated array, with ipath_tid's filled in, back
 328                  * to user.  Since we did the copy in already, this "should
 329                  * never fail" If it does, we have to clean up...
 330                  */
 331                 if (copy_to_user((void __user *)
 332                                  (unsigned long) ti->tidlist,
 333                                  tidlist, cnt * sizeof(*tidlist))) {
 334                         ret = -EFAULT;
 335                         goto cleanup;
 336                 }
 337                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
 338                                  tidmap, sizeof tidmap)) {
 339                         ret = -EFAULT;
 340                         goto cleanup;
 341                 }
 342                 if (tid == tidcnt)
 343                         tid = 0;
 344                 pd->port_tidcursor = tid;
 345         }
 346
 347 done:
 348         if (ret)
 349                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
 350                           ti->tidcnt, -ret);
 351         return ret;
 352 }
 353
 354 /**
 355  * ipath_tid_free - free a port TID
 356  * @pd: the port
 357  * @ti: the TID info
 358  *
 359  * right now we are unlocking one page at a time, but since
 360  * the intended use of this routine is for a single group of
 361  * virtually contiguous pages, that should change to improve
 362  * performance.  We check that the TID is in range for this port
 363  * but otherwise don't check validity; if user has an error and
 364  * frees the wrong tid, it's only their own data that can thereby
 365  * be corrupted.  We do check that the TID was in use, for sanity
 366  * We always use our idea of the saved address, not the address that
 367  * they pass in to us.
 368  */
 369
 370 static int ipath_tid_free(struct ipath_portdata *pd,
 371                           const struct ipath_tid_info *ti)
 372 {
 373         int ret = 0;
 374         u32 tid, porttid, cnt, limit, tidcnt;
 375         struct ipath_devdata *dd = pd->port_dd;
 376         u64 __iomem *tidbase;
 377         unsigned long tidmap[8];
 378
 379         if (!dd->ipath_pageshadow) {
 380                 ret = -ENOMEM;
 381                 goto done;
 382         }
 383
 384         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
 385                            sizeof tidmap)) {
 386                 ret = -EFAULT;
 387                 goto done;
 388         }
 389
 390         porttid = pd->port_port * dd->ipath_rcvtidcnt;
 391         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
 392                                    dd->ipath_rcvtidbase +
 393                                    porttid * sizeof(*tidbase));
 394
 395         tidcnt = dd->ipath_rcvtidcnt;
 396         limit = sizeof(tidmap) * BITS_PER_BYTE;
 397         if (limit > tidcnt)
 398                 /* just in case size changes in future */
 399                 limit = tidcnt;
 400         tid = find_first_bit(tidmap, limit);
 401         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
 402                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
 403                    limit, tid, porttid);
 404         for (cnt = 0; tid < limit; tid++) {
 405                 /*
 406                  * small optimization; if we detect a run of 3 or so without
 407                  * any set, use find_first_bit again.  That's mainly to
 408                  * accelerate the case where we wrapped, so we have some at
 409                  * the beginning, and some at the end, and a big gap
 410                  * in the middle.
 411                  */
 412                 if (!test_bit(tid, tidmap))
 413                         continue;
 414                 cnt++;
 415                 if (dd->ipath_pageshadow[porttid + tid]) {
 416                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
 417                                    pd->port_pid, tid);
 418                         dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
 419                                             dd->ipath_tidinvalid);
 420                         ipath_release_user_pages(
 421                                 &dd->ipath_pageshadow[porttid + tid], 1);
 422                         dd->ipath_pageshadow[porttid + tid] = NULL;
 423                         ipath_stats.sps_pageunlocks++;
 424                 } else
 425                         ipath_dbg("Unused tid %u, ignoring\n", tid);
 426         }
 427         if (cnt != ti->tidcnt)
 428                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
 429                           ti->tidcnt, cnt);
 430 done:
 431         if (ret)
 432                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
 433                           ti->tidcnt, -ret);
 434         return ret;
 435 }
 436
 437 /**
 438  * ipath_set_part_key - set a partition key
 439  * @pd: the port
 440  * @key: the key
 441  *
 442  * We can have up to 4 active at a time (other than the default, which is
 443  * always allowed).  This is somewhat tricky, since multiple ports may set
 444  * the same key, so we reference count them, and clean up at exit.  All 4
 445  * partition keys are packed into a single infinipath register.  It's an
 446  * error for a process to set the same pkey multiple times.  We provide no
 447  * mechanism to de-allocate a pkey at this time, we may eventually need to
 448  * do that.  I've used the atomic operations, and no locking, and only make
 449  * a single pass through what's available.  This should be more than
 450  * adequate for some time. I'll think about spinlocks or the like if and as
 451  * it's necessary.
 452  */
 453 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
 454 {
 455         struct ipath_devdata *dd = pd->port_dd;
 456         int i, any = 0, pidx = -1;
 457         u16 lkey = key & 0x7FFF;
 458         int ret;
 459
 460         if (lkey == (IPS_DEFAULT_P_KEY & 0x7FFF)) {
 461                 /* nothing to do; this key always valid */
 462                 ret = 0;
 463                 goto bail;
 464         }
 465
 466         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
 467                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
 468                    pd->port_port, key, dd->ipath_pkeys[0],
 469                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
 470                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
 471                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
 472                    atomic_read(&dd->ipath_pkeyrefs[3]));
 473
 474         if (!lkey) {
 475                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
 476                            pd->port_port);
 477                 ret = -EINVAL;
 478                 goto bail;
 479         }
 480
 481         /*
 482          * Set the full membership bit, because it has to be
 483          * set in the register or the packet, and it seems
 484          * cleaner to set in the register than to force all
 485          * callers to set it. (see bug 4331)
 486          */
 487         key |= 0x8000;
 488
 489         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 490                 if (!pd->port_pkeys[i] && pidx == -1)
 491                         pidx = i;
 492                 if (pd->port_pkeys[i] == key) {
 493                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
 494                                    "(%x) more than once\n",
 495                                    pd->port_port, key);
 496                         ret = -EEXIST;
 497                         goto bail;
 498                 }
 499         }
 500         if (pidx == -1) {
 501                 ipath_dbg("All pkeys for port %u already in use, "
 502                           "can't set %x\n", pd->port_port, key);
 503                 ret = -EBUSY;
 504                 goto bail;
 505         }
 506         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 507                 if (!dd->ipath_pkeys[i]) {
 508                         any++;
 509                         continue;
 510                 }
 511                 if (dd->ipath_pkeys[i] == key) {
 512                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
 513
 514                         if (atomic_inc_return(pkrefs) > 1) {
 515                                 pd->port_pkeys[pidx] = key;
 516                                 ipath_cdbg(VERBOSE, "p%u set key %x "
 517                                            "matches #%d, count now %d\n",
 518                                            pd->port_port, key, i,
 519                                            atomic_read(pkrefs));
 520                                 ret = 0;
 521                                 goto bail;
 522                         } else {
 523                                 /*
 524                                  * lost race, decrement count, catch below
 525                                  */
 526                                 atomic_dec(pkrefs);
 527                                 ipath_cdbg(VERBOSE, "Lost race, count was "
 528                                            "0, after dec, it's %d\n",
 529                                            atomic_read(pkrefs));
 530                                 any++;
 531                         }
 532                 }
 533                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
 534                         /*
 535                          * It makes no sense to have both the limited and
 536                          * full membership PKEY set at the same time since
 537                          * the unlimited one will disable the limited one.
 538                          */
 539                         ret = -EEXIST;
 540                         goto bail;
 541                 }
 542         }
 543         if (!any) {
 544                 ipath_dbg("port %u, all pkeys already in use, "
 545                           "can't set %x\n", pd->port_port, key);
 546                 ret = -EBUSY;
 547                 goto bail;
 548         }
 549         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
 550                 if (!dd->ipath_pkeys[i] &&
 551                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
 552                         u64 pkey;
 553
 554                         /* for ipathstats, etc. */
 555                         ipath_stats.sps_pkeys[i] = lkey;
 556                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
 557                         pkey =
 558                                 (u64) dd->ipath_pkeys[0] |
 559                                 ((u64) dd->ipath_pkeys[1] << 16) |
 560                                 ((u64) dd->ipath_pkeys[2] << 32) |
 561                                 ((u64) dd->ipath_pkeys[3] << 48);
 562                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
 563                                    "portidx %d, new pkey reg %llx\n",
 564                                    pd->port_port, key, i, pidx,
 565                                    (unsigned long long) pkey);
 566                         ipath_write_kreg(
 567                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
 568
 569                         ret = 0;
 570                         goto bail;
 571                 }
 572         }
 573         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
 574                   "can't set %x\n", pd->port_port, key);
 575         ret = -EBUSY;
 576
 577 bail:
 578         return ret;
 579 }
 580
 581 /**
 582  * ipath_manage_rcvq - manage a port's receive queue
 583  * @pd: the port
 584  * @start_stop: action to carry out
 585  *
 586  * start_stop == 0 disables receive on the port, for use in queue
 587  * overflow conditions.  start_stop==1 re-enables, to be used to
 588  * re-init the software copy of the head register
 589  */
 590 static int ipath_manage_rcvq(struct ipath_portdata *pd, int start_stop)
 591 {
 592         struct ipath_devdata *dd = pd->port_dd;
 593         u64 tval;
 594
 595         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u\n",
 596                    start_stop ? "en" : "dis", dd->ipath_unit,
 597                    pd->port_port);
 598         /* atomically clear receive enable port. */
 599         if (start_stop) {
 600                 /*
 601                  * On enable, force in-memory copy of the tail register to
 602                  * 0, so that protocol code doesn't have to worry about
 603                  * whether or not the chip has yet updated the in-memory
 604                  * copy or not on return from the system call. The chip
 605                  * always resets it's tail register back to 0 on a
 606                  * transition from disabled to enabled.  This could cause a
 607                  * problem if software was broken, and did the enable w/o
 608                  * the disable, but eventually the in-memory copy will be
 609                  * updated and correct itself, even in the face of software
 610                  * bugs.
 611                  */
 612                 *pd->port_rcvhdrtail_kvaddr = 0;
 613                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 614                         &dd->ipath_rcvctrl);
 615         } else
 616                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 617                           &dd->ipath_rcvctrl);
 618         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 619                          dd->ipath_rcvctrl);
 620         /* now be sure chip saw it before we return */
 621         tval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 622         if (start_stop) {
 623                 /*
 624                  * And try to be sure that tail reg update has happened too.
 625                  * This should in theory interlock with the RXE changes to
 626                  * the tail register.  Don't assign it to the tail register
 627                  * in memory copy, since we could overwrite an update by the
 628                  * chip if we did.
 629                  */
 630                 tval = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 631         }
 632         /* always; new head should be equal to new tail; see above */
 633         return 0;
 634 }
 635
 636 static void ipath_clean_part_key(struct ipath_portdata *pd,
 637                                  struct ipath_devdata *dd)
 638 {
 639         int i, j, pchanged = 0;
 640         u64 oldpkey;
 641
 642         /* for debugging only */
 643         oldpkey = (u64) dd->ipath_pkeys[0] |
 644                 ((u64) dd->ipath_pkeys[1] << 16) |
 645                 ((u64) dd->ipath_pkeys[2] << 32) |
 646                 ((u64) dd->ipath_pkeys[3] << 48);
 647
 648         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
 649                 if (!pd->port_pkeys[i])
 650                         continue;
 651                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
 652                            pd->port_pkeys[i]);
 653                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
 654                         /* check for match independent of the global bit */
 655                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
 656                             (pd->port_pkeys[i] & 0x7fff))
 657                                 continue;
 658                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
 659                                 ipath_cdbg(VERBOSE, "p%u clear key "
 660                                            "%x matches #%d\n",
 661                                            pd->port_port,
 662                                            pd->port_pkeys[i], j);
 663                                 ipath_stats.sps_pkeys[j] =
 664                                         dd->ipath_pkeys[j] = 0;
 665                                 pchanged++;
 666                         }
 667                         else ipath_cdbg(
 668                                 VERBOSE, "p%u key %x matches #%d, "
 669                                 "but ref still %d\n", pd->port_port,
 670                                 pd->port_pkeys[i], j,
 671                                 atomic_read(&dd->ipath_pkeyrefs[j]));
 672                         break;
 673                 }
 674                 pd->port_pkeys[i] = 0;
 675         }
 676         if (pchanged) {
 677                 u64 pkey = (u64) dd->ipath_pkeys[0] |
 678                         ((u64) dd->ipath_pkeys[1] << 16) |
 679                         ((u64) dd->ipath_pkeys[2] << 32) |
 680                         ((u64) dd->ipath_pkeys[3] << 48);
 681                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
 682                            "new pkey reg %llx\n", pd->port_port,
 683                            (unsigned long long) oldpkey,
 684                            (unsigned long long) pkey);
 685                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
 686                                  pkey);
 687         }
 688 }
 689
 690 /**
 691  * ipath_create_user_egr - allocate eager TID buffers
 692  * @pd: the port to allocate TID buffers for
 693  *
 694  * This routine is now quite different for user and kernel, because
 695  * the kernel uses skb's, for the accelerated network performance
 696  * This is the user port version
 697  *
 698  * Allocate the eager TID buffers and program them into infinipath
 699  * They are no longer completely contiguous, we do multiple allocation
 700  * calls.
 701  */
 702 static int ipath_create_user_egr(struct ipath_portdata *pd)
 703 {
 704         struct ipath_devdata *dd = pd->port_dd;
 705         unsigned e, egrcnt, alloced, egrperchunk, chunk, egrsize, egroff;
 706         size_t size;
 707         int ret;
 708         gfp_t gfp_flags;
 709
 710         /*
 711          * GFP_USER, but without GFP_FS, so buffer cache can be
 712          * coalesced (we hope); otherwise, even at order 4,
 713          * heavy filesystem activity makes these fail, and we can
 714          * use compound pages.
 715          */
 716         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
 717
 718         egrcnt = dd->ipath_rcvegrcnt;
 719         /* TID number offset for this port */
 720         egroff = pd->port_port * egrcnt;
 721         egrsize = dd->ipath_rcvegrbufsize;
 722         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
 723                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
 724
 725         /*
 726          * to avoid wasting a lot of memory, we allocate 32KB chunks of
 727          * physically contiguous memory, advance through it until used up
 728          * and then allocate more.  Of course, we need memory to store those
 729          * extra pointers, now.  Started out with 256KB, but under heavy
 730          * memory pressure (creating large files and then copying them over
 731          * NFS while doing lots of MPI jobs), we hit some allocation
 732          * failures, even though we can sleep...  (2.6.10) Still get
 733          * failures at 64K.  32K is the lowest we can go without wasting
 734          * additional memory.
 735          */
 736         size = 0x8000;
 737         alloced = ALIGN(egrsize * egrcnt, size);
 738         egrperchunk = size / egrsize;
 739         chunk = (egrcnt + egrperchunk - 1) / egrperchunk;
 740         pd->port_rcvegrbuf_chunks = chunk;
 741         pd->port_rcvegrbufs_perchunk = egrperchunk;
 742         pd->port_rcvegrbuf_size = size;
 743         pd->port_rcvegrbuf = vmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]));
 744         if (!pd->port_rcvegrbuf) {
 745                 ret = -ENOMEM;
 746                 goto bail;
 747         }
 748         pd->port_rcvegrbuf_phys =
 749                 vmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]));
 750         if (!pd->port_rcvegrbuf_phys) {
 751                 ret = -ENOMEM;
 752                 goto bail_rcvegrbuf;
 753         }
 754         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
 755
 756                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
 757                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
 758                         gfp_flags);
 759
 760                 if (!pd->port_rcvegrbuf[e]) {
 761                         ret = -ENOMEM;
 762                         goto bail_rcvegrbuf_phys;
 763                 }
 764         }
 765
 766         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
 767
 768         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
 769                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
 770                 unsigned i;
 771
 772                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
 773                         dd->ipath_f_put_tid(dd, e + egroff +
 774                                             (u64 __iomem *)
 775                                             ((char __iomem *)
 776                                              dd->ipath_kregbase +
 777                                              dd->ipath_rcvegrbase), 0, pa);
 778                         pa += egrsize;
 779                 }
 780                 cond_resched(); /* don't hog the cpu */
 781         }
 782
 783         ret = 0;
 784         goto bail;
 785
 786 bail_rcvegrbuf_phys:
 787         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
 788                      pd->port_rcvegrbuf[e]; e++)
 789                 dma_free_coherent(&dd->pcidev->dev, size,
 790                                   pd->port_rcvegrbuf[e],
 791                                   pd->port_rcvegrbuf_phys[e]);
 792
 793         vfree(pd->port_rcvegrbuf_phys);
 794         pd->port_rcvegrbuf_phys = NULL;
 795 bail_rcvegrbuf:
 796         vfree(pd->port_rcvegrbuf);
 797         pd->port_rcvegrbuf = NULL;
 798 bail:
 799         return ret;
 800 }
 801
 802 static int ipath_do_user_init(struct ipath_portdata *pd,
 803                               const struct ipath_user_info *uinfo)
 804 {
 805         int ret = 0;
 806         struct ipath_devdata *dd = pd->port_dd;
 807         u64 physaddr, uaddr, off, atmp;
 808         struct page *pagep;
 809         u32 head32;
 810         u64 head;
 811
 812         /* for now, if major version is different, bail */
 813         if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
 814                 dev_info(&dd->pcidev->dev,
 815                          "User major version %d not same as driver "
 816                          "major %d\n", uinfo->spu_userversion >> 16,
 817                          IPATH_USER_SWMAJOR);
 818                 ret = -ENODEV;
 819                 goto done;
 820         }
 821
 822         if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
 823                 ipath_dbg("User minor version %d not same as driver "
 824                           "minor %d\n", uinfo->spu_userversion & 0xffff,
 825                           IPATH_USER_SWMINOR);
 826
 827         if (uinfo->spu_rcvhdrsize) {
 828                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
 829                 if (ret)
 830                         goto done;
 831         }
 832
 833         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
 834
 835         /* set up for the rcvhdr Q tail register writeback to user memory */
 836         if (!uinfo->spu_rcvhdraddr ||
 837             !access_ok(VERIFY_WRITE, (u64 __user *) (unsigned long)
 838                        uinfo->spu_rcvhdraddr, sizeof(u64))) {
 839                 ipath_dbg("Port %d rcvhdrtail addr %llx not valid\n",
 840                           pd->port_port,
 841                           (unsigned long long) uinfo->spu_rcvhdraddr);
 842                 ret = -EINVAL;
 843                 goto done;
 844         }
 845
 846         off = offset_in_page(uinfo->spu_rcvhdraddr);
 847         uaddr = PAGE_MASK & (unsigned long) uinfo->spu_rcvhdraddr;
 848         ret = ipath_get_user_pages_nocopy(uaddr, &pagep);
 849         if (ret) {
 850                 dev_info(&dd->pcidev->dev, "Failed to lookup and lock "
 851                          "address %llx for rcvhdrtail: errno %d\n",
 852                          (unsigned long long) uinfo->spu_rcvhdraddr, -ret);
 853                 goto done;
 854         }
 855         ipath_stats.sps_pagelocks++;
 856         pd->port_rcvhdrtail_uaddr = uaddr;
 857         pd->port_rcvhdrtail_pagep = pagep;
 858         pd->port_rcvhdrtail_kvaddr =
 859                 page_address(pagep);
 860         pd->port_rcvhdrtail_kvaddr += off;
 861         physaddr = page_to_phys(pagep) + off;
 862         ipath_cdbg(VERBOSE, "port %d user addr %llx hdrtailaddr, %llx "
 863                    "physical (off=%llx)\n",
 864                    pd->port_port,
 865                    (unsigned long long) uinfo->spu_rcvhdraddr,
 866                    (unsigned long long) physaddr, (unsigned long long) off);
 867         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
 868                               pd->port_port, physaddr);
 869         atmp = ipath_read_kreg64_port(dd,
 870                                       dd->ipath_kregs->kr_rcvhdrtailaddr,
 871                                       pd->port_port);
 872         if (physaddr != atmp) {
 873                 ipath_dev_err(dd,
 874                               "Catastrophic software error, "
 875                               "RcvHdrTailAddr%u written as %llx, "
 876                               "read back as %llx\n", pd->port_port,
 877                               (unsigned long long) physaddr,
 878                               (unsigned long long) atmp);
 879                 ret = -EINVAL;
 880                 goto done;
 881         }
 882
 883         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
 884         pd->port_piobufs = dd->ipath_piobufbase +
 885                 dd->ipath_pbufsport * (pd->port_port -
 886                                        1) * dd->ipath_palign;
 887         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
 888                    pd->port_port, pd->port_piobufs);
 889
 890         /*
 891          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
 892          * array for time being.  If pd->port_port > chip-supported,
 893          * we need to do extra stuff here to handle by handling overflow
 894          * through port 0, someday
 895          */
 896         ret = ipath_create_rcvhdrq(dd, pd);
 897         if (!ret)
 898                 ret = ipath_create_user_egr(pd);
 899         if (ret)
 900                 goto done;
 901         /* enable receives now */
 902         /* atomically set enable bit for this port */
 903         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
 904                 &dd->ipath_rcvctrl);
 905
 906         /*
 907          * set the head registers for this port to the current values
 908          * of the tail pointers, since we don't know if they were
 909          * updated on last use of the port.
 910          */
 911         head32 = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
 912         head = (u64) head32;
 913         ipath_write_ureg(dd, ur_rcvhdrhead, head, pd->port_port);
 914         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
 915         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
 916         dd->ipath_lastegrheads[pd->port_port] = -1;
 917         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
 918         ipath_cdbg(VERBOSE, "Wrote port%d head %llx, egrhead %x from "
 919                    "tail regs\n", pd->port_port,
 920                    (unsigned long long) head, head32);
 921         pd->port_tidcursor = 0; /* start at beginning after open */
 922         /*
 923          * now enable the port; the tail registers will be written to memory
 924          * by the chip as soon as it sees the write to
 925          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
 926          * transition from 0 to 1, so clear it first, then set it as part of
 927          * enabling the port.  This will (very briefly) affect any other
 928          * open ports, but it shouldn't be long enough to be an issue.
 929          */
 930         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 931                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
 932         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 933                          dd->ipath_rcvctrl);
 934
 935 done:
 936         return ret;
 937 }
 938
 939 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
 940                      u64 ureg)
 941 {
 942         unsigned long phys;
 943         int ret;
 944
 945         /* it's the real hardware, so io_remap works */
 946
 947         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
 948                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
 949                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
 950                 ret = -EFAULT;
 951         } else {
 952                 phys = dd->ipath_physaddr + ureg;
 953                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 954
 955                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 956                 ret = io_remap_pfn_range(vma, vma->vm_start,
 957                                          phys >> PAGE_SHIFT,
 958                                          vma->vm_end - vma->vm_start,
 959                                          vma->vm_page_prot);
 960         }
 961         return ret;
 962 }
 963
 964 static int mmap_piobufs(struct vm_area_struct *vma,
 965                         struct ipath_devdata *dd,
 966                         struct ipath_portdata *pd)
 967 {
 968         unsigned long phys;
 969         int ret;
 970
 971         /*
 972          * When we map the PIO buffers, we want to map them as writeonly, no
 973          * read possible.
 974          */
 975
 976         if ((vma->vm_end - vma->vm_start) >
 977             (dd->ipath_pbufsport * dd->ipath_palign)) {
 978                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
 979                          "reqlen %lx > PAGE\n",
 980                          vma->vm_end - vma->vm_start);
 981                 ret = -EFAULT;
 982                 goto bail;
 983         }
 984
 985         phys = dd->ipath_physaddr + pd->port_piobufs;
 986         /*
 987          * Do *NOT* mark this as non-cached (PWT bit), or we don't get the
 988          * write combining behavior we want on the PIO buffers!
 989          * vma->vm_page_prot =
 990          *        pgprot_noncached(vma->vm_page_prot);
 991          */
 992
 993         if (vma->vm_flags & VM_READ) {
 994                 dev_info(&dd->pcidev->dev,
 995                          "Can't map piobufs as readable (flags=%lx)\n",
 996                          vma->vm_flags);
 997                 ret = -EPERM;
 998                 goto bail;
 999         }
1000
1001         /* don't allow them to later change to readable with mprotect */
1002
1003         vma->vm_flags &= ~VM_MAYWRITE;
1004         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1005
1006         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1007                                  vma->vm_end - vma->vm_start,
1008                                  vma->vm_page_prot);
1009 bail:
1010         return ret;
1011 }
1012
1013 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1014                            struct ipath_portdata *pd)
1015 {
1016         struct ipath_devdata *dd = pd->port_dd;
1017         unsigned long start, size;
1018         size_t total_size, i;
1019         dma_addr_t *phys;
1020         int ret;
1021
1022         if (!pd->port_rcvegrbuf) {
1023                 ret = -EFAULT;
1024                 goto bail;
1025         }
1026
1027         size = pd->port_rcvegrbuf_size;
1028         total_size = pd->port_rcvegrbuf_chunks * size;
1029         if ((vma->vm_end - vma->vm_start) > total_size) {
1030                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1031                          "reqlen %lx > actual %lx\n",
1032                          vma->vm_end - vma->vm_start,
1033                          (unsigned long) total_size);
1034                 ret = -EFAULT;
1035                 goto bail;
1036         }
1037
1038         if (vma->vm_flags & VM_WRITE) {
1039                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1040                          "writable (flags=%lx)\n", vma->vm_flags);
1041                 ret = -EPERM;
1042                 goto bail;
1043         }
1044
1045         start = vma->vm_start;
1046         phys = pd->port_rcvegrbuf_phys;
1047
1048         /* don't allow them to later change to writeable with mprotect */
1049         vma->vm_flags &= ~VM_MAYWRITE;
1050
1051         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1052                 ret = remap_pfn_range(vma, start, phys[i] >> PAGE_SHIFT,
1053                                       size, vma->vm_page_prot);
1054                 if (ret < 0)
1055                         goto bail;
1056         }
1057         ret = 0;
1058
1059 bail:
1060         return ret;
1061 }
1062
1063 static int mmap_rcvhdrq(struct vm_area_struct *vma,
1064                         struct ipath_portdata *pd)
1065 {
1066         struct ipath_devdata *dd = pd->port_dd;
1067         size_t total_size;
1068         int ret;
1069
1070         /*
1071          * kmalloc'ed memory, physically contiguous; this is from
1072          * spi_rcvhdr_base; we allow user to map read-write so they can
1073          * write hdrq entries to allow protocol code to directly poll
1074          * whether a hdrq entry has been written.
1075          */
1076         total_size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1077                            sizeof(u32), PAGE_SIZE);
1078         if ((vma->vm_end - vma->vm_start) > total_size) {
1079                 dev_info(&dd->pcidev->dev,
1080                          "FAIL on rcvhdrq: reqlen %lx > actual %lx\n",
1081                          vma->vm_end - vma->vm_start,
1082                          (unsigned long) total_size);
1083                 ret = -EFAULT;
1084                 goto bail;
1085         }
1086
1087         ret = remap_pfn_range(vma, vma->vm_start,
1088                               pd->port_rcvhdrq_phys >> PAGE_SHIFT,
1089                               vma->vm_end - vma->vm_start,
1090                               vma->vm_page_prot);
1091 bail:
1092         return ret;
1093 }
1094
1095 static int mmap_pioavailregs(struct vm_area_struct *vma,
1096                              struct ipath_portdata *pd)
1097 {
1098         struct ipath_devdata *dd = pd->port_dd;
1099         int ret;
1100
1101         /*
1102          * when we map the PIO bufferavail registers, we want to map them as
1103          * readonly, no write possible.
1104          *
1105          * kmalloc'ed memory, physically contiguous, one page only, readonly
1106          */
1107
1108         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1109                 dev_info(&dd->pcidev->dev, "FAIL on pioavailregs_dma: "
1110                          "reqlen %lx > actual %lx\n",
1111                          vma->vm_end - vma->vm_start,
1112                          (unsigned long) PAGE_SIZE);
1113                 ret = -EFAULT;
1114                 goto bail;
1115         }
1116
1117         if (vma->vm_flags & VM_WRITE) {
1118                 dev_info(&dd->pcidev->dev,
1119                          "Can't map pioavailregs as writable (flags=%lx)\n",
1120                          vma->vm_flags);
1121                 ret = -EPERM;
1122                 goto bail;
1123         }
1124
1125         /* don't allow them to later change with mprotect */
1126         vma->vm_flags &= ~VM_MAYWRITE;
1127
1128         ret = remap_pfn_range(vma, vma->vm_start,
1129                               dd->ipath_pioavailregs_phys >> PAGE_SHIFT,
1130                               PAGE_SIZE, vma->vm_page_prot);
1131 bail:
1132         return ret;
1133 }
1134
1135 /**
1136  * ipath_mmap - mmap various structures into user space
1137  * @fp: the file pointer
1138  * @vma: the VM area
1139  *
1140  * We use this to have a shared buffer between the kernel and the user code
1141  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1142  * buffers in the chip.  We have the open and close entries so we can bump
1143  * the ref count and keep the driver from being unloaded while still mapped.
1144  */
1145 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1146 {
1147         struct ipath_portdata *pd;
1148         struct ipath_devdata *dd;
1149         u64 pgaddr, ureg;
1150         int ret;
1151
1152         pd = port_fp(fp);
1153         dd = pd->port_dd;
1154         /*
1155          * This is the ipath_do_user_init() code, mapping the shared buffers
1156          * into the user process. The address referred to by vm_pgoff is the
1157          * virtual, not physical, address; we only do one mmap for each
1158          * space mapped.
1159          */
1160         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1161
1162         /*
1163          * note that ureg does *NOT* have the kregvirt as part of it, to be
1164          * sure that for 32 bit programs, we don't end up trying to map a >
1165          * 44 address.  Has to match ipath_get_base_info() code that sets
1166          * __spi_uregbase
1167          */
1168
1169         ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1170
1171         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u\n",
1172                    (unsigned long long) pgaddr, vma->vm_start,
1173                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1174                    pd->port_port);
1175
1176         if (pgaddr == ureg)
1177                 ret = mmap_ureg(vma, dd, ureg);
1178         else if (pgaddr == pd->port_piobufs)
1179                 ret = mmap_piobufs(vma, dd, pd);
1180         else if (pgaddr == (u64) pd->port_rcvegr_phys)
1181                 ret = mmap_rcvegrbufs(vma, pd);
1182         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1183                 ret = mmap_rcvhdrq(vma, pd);
1184         else if (pgaddr == dd->ipath_pioavailregs_phys)
1185                 ret = mmap_pioavailregs(vma, pd);
1186         else
1187                 ret = -EINVAL;
1188
1189         vma->vm_private_data = NULL;
1190
1191         if (ret < 0)
1192                 dev_info(&dd->pcidev->dev,
1193                          "Failure %d on addr %lx, off %lx\n",
1194                          -ret, vma->vm_start, vma->vm_pgoff);
1195
1196         return ret;
1197 }
1198
1199 static unsigned int ipath_poll(struct file *fp,
1200                                struct poll_table_struct *pt)
1201 {
1202         struct ipath_portdata *pd;
1203         u32 head, tail;
1204         int bit;
1205         struct ipath_devdata *dd;
1206
1207         pd = port_fp(fp);
1208         dd = pd->port_dd;
1209
1210         bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
1211         set_bit(bit, &dd->ipath_rcvctrl);
1212
1213         /*
1214          * Before blocking, make sure that head is still == tail,
1215          * reading from the chip, so we can be sure the interrupt
1216          * enable has made it to the chip.  If not equal, disable
1217          * interrupt again and return immediately.  This avoids races,
1218          * and the overhead of the chip read doesn't matter much at
1219          * this point, since we are waiting for something anyway.
1220          */
1221
1222         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1223                          dd->ipath_rcvctrl);
1224
1225         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1226         tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1227
1228         if (tail == head) {
1229                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1230                 if(dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1231                         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1232                                                dd->ipath_rhdrhead_intr_off
1233                                                | head, pd->port_port);
1234                 poll_wait(fp, &pd->port_wait, pt);
1235
1236                 if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
1237                         /* timed out, no packets received */
1238                         clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1239                         pd->port_rcvwait_to++;
1240                 }
1241         }
1242         else {
1243                 /* it's already happened; don't do wait_event overhead */
1244                 pd->port_rcvnowait++;
1245         }
1246
1247         clear_bit(bit, &dd->ipath_rcvctrl);
1248         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1249                          dd->ipath_rcvctrl);
1250
1251         return 0;
1252 }
1253
1254 static int try_alloc_port(struct ipath_devdata *dd, int port,
1255                           struct file *fp)
1256 {
1257         int ret;
1258
1259         if (!dd->ipath_pd[port]) {
1260                 void *p, *ptmp;
1261
1262                 p = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1263
1264                 /*
1265                  * Allocate memory for use in ipath_tid_update() just once
1266                  * at open, not per call.  Reduces cost of expected send
1267                  * setup.
1268                  */
1269                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1270                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1271                                GFP_KERNEL);
1272                 if (!p || !ptmp) {
1273                         ipath_dev_err(dd, "Unable to allocate portdata "
1274                                       "memory, failing open\n");
1275                         ret = -ENOMEM;
1276                         kfree(p);
1277                         kfree(ptmp);
1278                         goto bail;
1279                 }
1280                 dd->ipath_pd[port] = p;
1281                 dd->ipath_pd[port]->port_port = port;
1282                 dd->ipath_pd[port]->port_dd = dd;
1283                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1284                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1285         }
1286         if (!dd->ipath_pd[port]->port_cnt) {
1287                 dd->ipath_pd[port]->port_cnt = 1;
1288                 fp->private_data = (void *) dd->ipath_pd[port];
1289                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1290                            current->comm, current->pid, dd->ipath_unit,
1291                            port);
1292                 dd->ipath_pd[port]->port_pid = current->pid;
1293                 strncpy(dd->ipath_pd[port]->port_comm, current->comm,
1294                         sizeof(dd->ipath_pd[port]->port_comm));
1295                 ipath_stats.sps_ports++;
1296                 ret = 0;
1297                 goto bail;
1298         }
1299         ret = -EBUSY;
1300
1301 bail:
1302         return ret;
1303 }
1304
1305 static inline int usable(struct ipath_devdata *dd)
1306 {
1307         return dd &&
1308                 (dd->ipath_flags & IPATH_PRESENT) &&
1309                 dd->ipath_kregbase &&
1310                 dd->ipath_lid &&
1311                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1312                                      | IPATH_LINKUNK));
1313 }
1314
1315 static int find_free_port(int unit, struct file *fp)
1316 {
1317         struct ipath_devdata *dd = ipath_lookup(unit);
1318         int ret, i;
1319
1320         if (!dd) {
1321                 ret = -ENODEV;
1322                 goto bail;
1323         }
1324
1325         if (!usable(dd)) {
1326                 ret = -ENETDOWN;
1327                 goto bail;
1328         }
1329
1330         for (i = 0; i < dd->ipath_cfgports; i++) {
1331                 ret = try_alloc_port(dd, i, fp);
1332                 if (ret != -EBUSY)
1333                         goto bail;
1334         }
1335         ret = -EBUSY;
1336
1337 bail:
1338         return ret;
1339 }
1340
1341 static int find_best_unit(struct file *fp)
1342 {
1343         int ret = 0, i, prefunit = -1, devmax;
1344         int maxofallports, npresent, nup;
1345         int ndev;
1346
1347         (void) ipath_count_units(&npresent, &nup, &maxofallports);
1348
1349         /*
1350          * This code is present to allow a knowledgeable person to
1351          * specify the layout of processes to processors before opening
1352          * this driver, and then we'll assign the process to the "closest"
1353          * HT-400 to that processor (we assume reasonable connectivity,
1354          * for now).  This code assumes that if affinity has been set
1355          * before this point, that at most one cpu is set; for now this
1356          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1357          * in case some kernel variant sets none of the bits when no
1358          * affinity is set.  2.6.11 and 12 kernels have all present
1359          * cpus set.  Some day we'll have to fix it up further to handle
1360          * a cpu subset.  This algorithm fails for two HT-400's connected
1361          * in tunnel fashion.  Eventually this needs real topology
1362          * information.  There may be some issues with dual core numbering
1363          * as well.  This needs more work prior to release.
1364          */
1365         if (!cpus_empty(current->cpus_allowed) &&
1366             !cpus_full(current->cpus_allowed)) {
1367                 int ncpus = num_online_cpus(), curcpu = -1;
1368                 for (i = 0; i < ncpus; i++)
1369                         if (cpu_isset(i, current->cpus_allowed)) {
1370                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1371                                            "cpu %d\n", current->comm,
1372                                            current->pid, i);
1373                                 curcpu = i;
1374                         }
1375                 if (curcpu != -1) {
1376                         if (npresent) {
1377                                 prefunit = curcpu / (ncpus / npresent);
1378                                 ipath_dbg("%s[%u] %d chips, %d cpus, "
1379                                           "%d cpus/chip, select unit %d\n",
1380                                           current->comm, current->pid,
1381                                           npresent, ncpus, ncpus / npresent,
1382                                           prefunit);
1383                         }
1384                 }
1385         }
1386
1387         /*
1388          * user ports start at 1, kernel port is 0
1389          * For now, we do round-robin access across all chips
1390          */
1391
1392         if (prefunit != -1)
1393                 devmax = prefunit + 1;
1394         else
1395                 devmax = ipath_count_units(NULL, NULL, NULL);
1396 recheck:
1397         for (i = 1; i < maxofallports; i++) {
1398                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1399                      ndev++) {
1400                         struct ipath_devdata *dd = ipath_lookup(ndev);
1401
1402                         if (!usable(dd))
1403                                 continue; /* can't use this unit */
1404                         if (i >= dd->ipath_cfgports)
1405                                 /*
1406                                  * Maxed out on users of this unit. Try
1407                                  * next.
1408                                  */
1409                                 continue;
1410                         ret = try_alloc_port(dd, i, fp);
1411                         if (!ret)
1412                                 goto done;
1413                 }
1414         }
1415
1416         if (npresent) {
1417                 if (nup == 0) {
1418                         ret = -ENETDOWN;
1419                         ipath_dbg("No ports available (none initialized "
1420                                   "and ready)\n");
1421                 } else {
1422                         if (prefunit > 0) {
1423                                 /* if started above 0, retry from 0 */
1424                                 ipath_cdbg(PROC,
1425                                            "%s[%u] no ports on prefunit "
1426                                            "%d, clear and re-check\n",
1427                                            current->comm, current->pid,
1428                                            prefunit);
1429                                 devmax = ipath_count_units(NULL, NULL,
1430                                                            NULL);
1431                                 prefunit = -1;
1432                                 goto recheck;
1433                         }
1434                         ret = -EBUSY;
1435                         ipath_dbg("No ports available\n");
1436                 }
1437         } else {
1438                 ret = -ENXIO;
1439                 ipath_dbg("No boards found\n");
1440         }
1441
1442 done:
1443         return ret;
1444 }
1445
1446 static int ipath_open(struct inode *in, struct file *fp)
1447 {
1448         int ret, user_minor;
1449
1450         mutex_lock(&ipath_mutex);
1451
1452         user_minor = iminor(in) - IPATH_USER_MINOR_BASE;
1453         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1454                    (long)in->i_rdev, user_minor);
1455
1456         if (user_minor)
1457                 ret = find_free_port(user_minor - 1, fp);
1458         else
1459                 ret = find_best_unit(fp);
1460
1461         mutex_unlock(&ipath_mutex);
1462         return ret;
1463 }
1464
1465 /**
1466  * unlock_exptid - unlock any expected TID entries port still had in use
1467  * @pd: port
1468  *
1469  * We don't actually update the chip here, because we do a bulk update
1470  * below, using ipath_f_clear_tids.
1471  */
1472 static void unlock_expected_tids(struct ipath_portdata *pd)
1473 {
1474         struct ipath_devdata *dd = pd->port_dd;
1475         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1476         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1477
1478         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1479                    pd->port_port);
1480         for (i = port_tidbase; i < maxtid; i++) {
1481                 if (!dd->ipath_pageshadow[i])
1482                         continue;
1483
1484                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1485                                                   1);
1486                 dd->ipath_pageshadow[i] = NULL;
1487                 cnt++;
1488                 ipath_stats.sps_pageunlocks++;
1489         }
1490         if (cnt)
1491                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1492                            pd->port_port, cnt);
1493
1494         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1495                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1496                            (unsigned long long) ipath_stats.sps_pagelocks,
1497                            (unsigned long long)
1498                            ipath_stats.sps_pageunlocks);
1499 }
1500
1501 static int ipath_close(struct inode *in, struct file *fp)
1502 {
1503         int ret = 0;
1504         struct ipath_portdata *pd;
1505         struct ipath_devdata *dd;
1506         unsigned port;
1507
1508         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1509                    (long)in->i_rdev, fp->private_data);
1510
1511         mutex_lock(&ipath_mutex);
1512
1513         pd = port_fp(fp);
1514         port = pd->port_port;
1515         fp->private_data = NULL;
1516         dd = pd->port_dd;
1517
1518         if (pd->port_hdrqfull) {
1519                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
1520                            "during run\n", pd->port_comm, pd->port_pid,
1521                            pd->port_hdrqfull);
1522                 pd->port_hdrqfull = 0;
1523         }
1524
1525         if (pd->port_rcvwait_to || pd->port_piowait_to
1526             || pd->port_rcvnowait || pd->port_pionowait) {
1527                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
1528                            "%u rcv %u, pio already\n",
1529                            pd->port_port, pd->port_rcvwait_to,
1530                            pd->port_piowait_to, pd->port_rcvnowait,
1531                            pd->port_pionowait);
1532                 pd->port_rcvwait_to = pd->port_piowait_to =
1533                         pd->port_rcvnowait = pd->port_pionowait = 0;
1534         }
1535         if (pd->port_flag) {
1536                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
1537                           pd->port_port, pd->port_flag);
1538                 pd->port_flag = 0;
1539         }
1540
1541         if (dd->ipath_kregbase) {
1542                 if (pd->port_rcvhdrtail_uaddr) {
1543                         pd->port_rcvhdrtail_uaddr = 0;
1544                         pd->port_rcvhdrtail_kvaddr = NULL;
1545                         ipath_release_user_pages_on_close(
1546                                 &pd->port_rcvhdrtail_pagep, 1);
1547                         pd->port_rcvhdrtail_pagep = NULL;
1548                         ipath_stats.sps_pageunlocks++;
1549                 }
1550                 ipath_write_kreg_port(
1551                         dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1552                         port, 0ULL);
1553                 ipath_write_kreg_port(
1554                         dd, dd->ipath_kregs->kr_rcvhdraddr,
1555                         pd->port_port, 0);
1556
1557                 /* clean up the pkeys for this port user */
1558                 ipath_clean_part_key(pd, dd);
1559
1560                 if (port < dd->ipath_cfgports) {
1561                         int i = dd->ipath_pbufsport * (port - 1);
1562                         ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
1563
1564                         /* atomically clear receive enable port. */
1565                         clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
1566                                   &dd->ipath_rcvctrl);
1567                         ipath_write_kreg(
1568                                 dd,
1569                                 dd->ipath_kregs->kr_rcvctrl,
1570                                 dd->ipath_rcvctrl);
1571
1572                         if (dd->ipath_pageshadow)
1573                                 unlock_expected_tids(pd);
1574                         ipath_stats.sps_ports--;
1575                         ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
1576                                    pd->port_comm, pd->port_pid,
1577                                    dd->ipath_unit, port);
1578                 }
1579         }
1580
1581         pd->port_cnt = 0;
1582         pd->port_pid = 0;
1583
1584         dd->ipath_f_clear_tids(dd, pd->port_port);
1585
1586         ipath_free_pddata(dd, pd->port_port, 0);
1587
1588         mutex_unlock(&ipath_mutex);
1589
1590         return ret;
1591 }
1592
1593 static int ipath_port_info(struct ipath_portdata *pd,
1594                            struct ipath_port_info __user *uinfo)
1595 {
1596         struct ipath_port_info info;
1597         int nup;
1598         int ret;
1599
1600         (void) ipath_count_units(NULL, &nup, NULL);
1601         info.num_active = nup;
1602         info.unit = pd->port_dd->ipath_unit;
1603         info.port = pd->port_port;
1604
1605         if (copy_to_user(uinfo, &info, sizeof(info))) {
1606                 ret = -EFAULT;
1607                 goto bail;
1608         }
1609         ret = 0;
1610
1611 bail:
1612         return ret;
1613 }
1614
1615 static ssize_t ipath_write(struct file *fp, const char __user *data,
1616                            size_t count, loff_t *off)
1617 {
1618         const struct ipath_cmd __user *ucmd;
1619         struct ipath_portdata *pd;
1620         const void __user *src;
1621         size_t consumed, copy;
1622         struct ipath_cmd cmd;
1623         ssize_t ret = 0;
1624         void *dest;
1625
1626         if (count < sizeof(cmd.type)) {
1627                 ret = -EINVAL;
1628                 goto bail;
1629         }
1630
1631         ucmd = (const struct ipath_cmd __user *) data;
1632
1633         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
1634                 ret = -EFAULT;
1635                 goto bail;
1636         }
1637
1638         consumed = sizeof(cmd.type);
1639
1640         switch (cmd.type) {
1641         case IPATH_CMD_USER_INIT:
1642                 copy = sizeof(cmd.cmd.user_info);
1643                 dest = &cmd.cmd.user_info;
1644                 src = &ucmd->cmd.user_info;
1645                 break;
1646         case IPATH_CMD_RECV_CTRL:
1647                 copy = sizeof(cmd.cmd.recv_ctrl);
1648                 dest = &cmd.cmd.recv_ctrl;
1649                 src = &ucmd->cmd.recv_ctrl;
1650                 break;
1651         case IPATH_CMD_PORT_INFO:
1652                 copy = sizeof(cmd.cmd.port_info);
1653                 dest = &cmd.cmd.port_info;
1654                 src = &ucmd->cmd.port_info;
1655                 break;
1656         case IPATH_CMD_TID_UPDATE:
1657         case IPATH_CMD_TID_FREE:
1658                 copy = sizeof(cmd.cmd.tid_info);
1659                 dest = &cmd.cmd.tid_info;
1660                 src = &ucmd->cmd.tid_info;
1661                 break;
1662         case IPATH_CMD_SET_PART_KEY:
1663                 copy = sizeof(cmd.cmd.part_key);
1664                 dest = &cmd.cmd.part_key;
1665                 src = &ucmd->cmd.part_key;
1666                 break;
1667         default:
1668                 ret = -EINVAL;
1669                 goto bail;
1670         }
1671
1672         if ((count - consumed) < copy) {
1673                 ret = -EINVAL;
1674                 goto bail;
1675         }
1676
1677         if (copy_from_user(dest, src, copy)) {
1678                 ret = -EFAULT;
1679                 goto bail;
1680         }
1681
1682         consumed += copy;
1683         pd = port_fp(fp);
1684
1685         switch (cmd.type) {
1686         case IPATH_CMD_USER_INIT:
1687                 ret = ipath_do_user_init(pd, &cmd.cmd.user_info);
1688                 if (ret < 0)
1689                         goto bail;
1690                 ret = ipath_get_base_info(
1691                         pd, (void __user *) (unsigned long)
1692                         cmd.cmd.user_info.spu_base_info,
1693                         cmd.cmd.user_info.spu_base_info_size);
1694                 break;
1695         case IPATH_CMD_RECV_CTRL:
1696                 ret = ipath_manage_rcvq(pd, cmd.cmd.recv_ctrl);
1697                 break;
1698         case IPATH_CMD_PORT_INFO:
1699                 ret = ipath_port_info(pd,
1700                                       (struct ipath_port_info __user *)
1701                                       (unsigned long) cmd.cmd.port_info);
1702                 break;
1703         case IPATH_CMD_TID_UPDATE:
1704                 ret = ipath_tid_update(pd, &cmd.cmd.tid_info);
1705                 break;
1706         case IPATH_CMD_TID_FREE:
1707                 ret = ipath_tid_free(pd, &cmd.cmd.tid_info);
1708                 break;
1709         case IPATH_CMD_SET_PART_KEY:
1710                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
1711                 break;
1712         }
1713
1714         if (ret >= 0)
1715                 ret = consumed;
1716
1717 bail:
1718         return ret;
1719 }
1720
1721 static struct class *ipath_class;
1722
1723 static int init_cdev(int minor, char *name, struct file_operations *fops,
1724                      struct cdev **cdevp, struct class_device **class_devp)
1725 {
1726         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
1727         struct cdev *cdev = NULL;
1728         struct class_device *class_dev = NULL;
1729         int ret;
1730
1731         cdev = cdev_alloc();
1732         if (!cdev) {
1733                 printk(KERN_ERR IPATH_DRV_NAME
1734                        ": Could not allocate cdev for minor %d, %s\n",
1735                        minor, name);
1736                 ret = -ENOMEM;
1737                 goto done;
1738         }
1739
1740         cdev->owner = THIS_MODULE;
1741         cdev->ops = fops;
1742         kobject_set_name(&cdev->kobj, name);
1743
1744         ret = cdev_add(cdev, dev, 1);
1745         if (ret < 0) {
1746                 printk(KERN_ERR IPATH_DRV_NAME
1747                        ": Could not add cdev for minor %d, %s (err %d)\n",
1748                        minor, name, -ret);
1749                 goto err_cdev;
1750         }
1751
1752         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
1753
1754         if (IS_ERR(class_dev)) {
1755                 ret = PTR_ERR(class_dev);
1756                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
1757                        "class_dev for minor %d, %s (err %d)\n",
1758                        minor, name, -ret);
1759                 goto err_cdev;
1760         }
1761
1762         goto done;
1763
1764 err_cdev:
1765         cdev_del(cdev);
1766         cdev = NULL;
1767
1768 done:
1769         if (ret >= 0) {
1770                 *cdevp = cdev;
1771                 *class_devp = class_dev;
1772         } else {
1773                 *cdevp = NULL;
1774                 *class_devp = NULL;
1775         }
1776
1777         return ret;
1778 }
1779
1780 int ipath_cdev_init(int minor, char *name, struct file_operations *fops,
1781                     struct cdev **cdevp, struct class_device **class_devp)
1782 {
1783         return init_cdev(minor, name, fops, cdevp, class_devp);
1784 }
1785
1786 static void cleanup_cdev(struct cdev **cdevp,
1787                          struct class_device **class_devp)
1788 {
1789         struct class_device *class_dev = *class_devp;
1790
1791         if (class_dev) {
1792                 class_device_unregister(class_dev);
1793                 *class_devp = NULL;
1794         }
1795
1796         if (*cdevp) {
1797                 cdev_del(*cdevp);
1798                 *cdevp = NULL;
1799         }
1800 }
1801
1802 void ipath_cdev_cleanup(struct cdev **cdevp,
1803                         struct class_device **class_devp)
1804 {
1805         cleanup_cdev(cdevp, class_devp);
1806 }
1807
1808 static struct cdev *wildcard_cdev;
1809 static struct class_device *wildcard_class_dev;
1810
1811 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
1812
1813 static int user_init(void)
1814 {
1815         int ret;
1816
1817         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
1818         if (ret < 0) {
1819                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
1820                        "chrdev region (err %d)\n", -ret);
1821                 goto done;
1822         }
1823
1824         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
1825
1826         if (IS_ERR(ipath_class)) {
1827                 ret = PTR_ERR(ipath_class);
1828                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
1829                        "device class (err %d)\n", -ret);
1830                 goto bail;
1831         }
1832
1833         goto done;
1834 bail:
1835         unregister_chrdev_region(dev, IPATH_NMINORS);
1836 done:
1837         return ret;
1838 }
1839
1840 static void user_cleanup(void)
1841 {
1842         if (ipath_class) {
1843                 class_destroy(ipath_class);
1844                 ipath_class = NULL;
1845         }
1846
1847         unregister_chrdev_region(dev, IPATH_NMINORS);
1848 }
1849
1850 static atomic_t user_count = ATOMIC_INIT(0);
1851 static atomic_t user_setup = ATOMIC_INIT(0);
1852
1853 int ipath_user_add(struct ipath_devdata *dd)
1854 {
1855         char name[10];
1856         int ret;
1857
1858         if (atomic_inc_return(&user_count) == 1) {
1859                 ret = user_init();
1860                 if (ret < 0) {
1861                         ipath_dev_err(dd, "Unable to set up user support: "
1862                                       "error %d\n", -ret);
1863                         goto bail;
1864                 }
1865                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
1866                                 &wildcard_class_dev);
1867                 if (ret < 0) {
1868                         ipath_dev_err(dd, "Could not create wildcard "
1869                                       "minor: error %d\n", -ret);
1870                         goto bail_sma;
1871                 }
1872
1873                 atomic_set(&user_setup, 1);
1874         }
1875
1876         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
1877
1878         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
1879                         &dd->user_cdev, &dd->user_class_dev);
1880         if (ret < 0)
1881                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
1882                               dd->ipath_unit + 1, name);
1883
1884         goto bail;
1885
1886 bail_sma:
1887         user_cleanup();
1888 bail:
1889         return ret;
1890 }
1891
1892 void ipath_user_remove(struct ipath_devdata *dd)
1893 {
1894         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
1895
1896         if (atomic_dec_return(&user_count) == 0) {
1897                 if (atomic_read(&user_setup) == 0)
1898                         goto bail;
1899
1900                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
1901                 user_cleanup();
1902
1903                 atomic_set(&user_setup, 0);
1904         }
1905 bail:
1906         return;
1907 }